1 | /* Statement Analysis and Transformation for Vectorization |
2 | Copyright (C) 2003-2023 Free Software Foundation, Inc. |
3 | Contributed by Dorit Naishlos <dorit@il.ibm.com> |
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" |
25 | #include "backend.h" |
26 | #include "target.h" |
27 | #include "rtl.h" |
28 | #include "tree.h" |
29 | #include "gimple.h" |
30 | #include "ssa.h" |
31 | #include "optabs-tree.h" |
32 | #include "insn-config.h" |
33 | #include "recog.h" /* FIXME: for insn_data */ |
34 | #include "cgraph.h" |
35 | #include "dumpfile.h" |
36 | #include "alias.h" |
37 | #include "fold-const.h" |
38 | #include "stor-layout.h" |
39 | #include "tree-eh.h" |
40 | #include "gimplify.h" |
41 | #include "gimple-iterator.h" |
42 | #include "gimplify-me.h" |
43 | #include "tree-cfg.h" |
44 | #include "tree-ssa-loop-manip.h" |
45 | #include "cfgloop.h" |
46 | #include "explow.h" |
47 | #include "tree-ssa-loop.h" |
48 | #include "tree-scalar-evolution.h" |
49 | #include "tree-vectorizer.h" |
50 | #include "builtins.h" |
51 | #include "internal-fn.h" |
52 | #include "tree-vector-builder.h" |
53 | #include "vec-perm-indices.h" |
54 | #include "gimple-range.h" |
55 | #include "tree-ssa-loop-niter.h" |
56 | #include "gimple-fold.h" |
57 | #include "regs.h" |
58 | #include "attribs.h" |
59 | #include "optabs-libfuncs.h" |
60 | |
61 | /* For lang_hooks.types.type_for_mode. */ |
62 | #include "langhooks.h" |
63 | |
64 | /* Return the vectorized type for the given statement. */ |
65 | |
66 | tree |
67 | stmt_vectype (class _stmt_vec_info *stmt_info) |
68 | { |
69 | return STMT_VINFO_VECTYPE (stmt_info); |
70 | } |
71 | |
72 | /* Return TRUE iff the given statement is in an inner loop relative to |
73 | the loop being vectorized. */ |
74 | bool |
75 | stmt_in_inner_loop_p (vec_info *vinfo, class _stmt_vec_info *stmt_info) |
76 | { |
77 | gimple *stmt = STMT_VINFO_STMT (stmt_info); |
78 | basic_block bb = gimple_bb (g: stmt); |
79 | loop_vec_info loop_vinfo = dyn_cast <loop_vec_info> (p: vinfo); |
80 | class loop* loop; |
81 | |
82 | if (!loop_vinfo) |
83 | return false; |
84 | |
85 | loop = LOOP_VINFO_LOOP (loop_vinfo); |
86 | |
87 | return (bb->loop_father == loop->inner); |
88 | } |
89 | |
90 | /* Record the cost of a statement, either by directly informing the |
91 | target model or by saving it in a vector for later processing. |
92 | Return a preliminary estimate of the statement's cost. */ |
93 | |
94 | static unsigned |
95 | record_stmt_cost (stmt_vector_for_cost *body_cost_vec, int count, |
96 | enum vect_cost_for_stmt kind, |
97 | stmt_vec_info stmt_info, slp_tree node, |
98 | tree vectype, int misalign, |
99 | enum vect_cost_model_location where) |
100 | { |
101 | if ((kind == vector_load || kind == unaligned_load) |
102 | && (stmt_info && STMT_VINFO_GATHER_SCATTER_P (stmt_info))) |
103 | kind = vector_gather_load; |
104 | if ((kind == vector_store || kind == unaligned_store) |
105 | && (stmt_info && STMT_VINFO_GATHER_SCATTER_P (stmt_info))) |
106 | kind = vector_scatter_store; |
107 | |
108 | stmt_info_for_cost si |
109 | = { .count: count, .kind: kind, .where: where, .stmt_info: stmt_info, .node: node, .vectype: vectype, .misalign: misalign }; |
110 | body_cost_vec->safe_push (obj: si); |
111 | |
112 | return (unsigned) |
113 | (builtin_vectorization_cost (type_of_cost: kind, vectype, misalign) * count); |
114 | } |
115 | |
116 | unsigned |
117 | record_stmt_cost (stmt_vector_for_cost *body_cost_vec, int count, |
118 | enum vect_cost_for_stmt kind, stmt_vec_info stmt_info, |
119 | tree vectype, int misalign, |
120 | enum vect_cost_model_location where) |
121 | { |
122 | return record_stmt_cost (body_cost_vec, count, kind, stmt_info, NULL, |
123 | vectype, misalign, where); |
124 | } |
125 | |
126 | unsigned |
127 | record_stmt_cost (stmt_vector_for_cost *body_cost_vec, int count, |
128 | enum vect_cost_for_stmt kind, slp_tree node, |
129 | tree vectype, int misalign, |
130 | enum vect_cost_model_location where) |
131 | { |
132 | return record_stmt_cost (body_cost_vec, count, kind, NULL, node, |
133 | vectype, misalign, where); |
134 | } |
135 | |
136 | unsigned |
137 | record_stmt_cost (stmt_vector_for_cost *body_cost_vec, int count, |
138 | enum vect_cost_for_stmt kind, |
139 | enum vect_cost_model_location where) |
140 | { |
141 | gcc_assert (kind == cond_branch_taken || kind == cond_branch_not_taken |
142 | || kind == scalar_stmt); |
143 | return record_stmt_cost (body_cost_vec, count, kind, NULL, NULL, |
144 | NULL_TREE, misalign: 0, where); |
145 | } |
146 | |
147 | /* Return a variable of type ELEM_TYPE[NELEMS]. */ |
148 | |
149 | static tree |
150 | create_vector_array (tree elem_type, unsigned HOST_WIDE_INT nelems) |
151 | { |
152 | return create_tmp_var (build_array_type_nelts (elem_type, nelems), |
153 | "vect_array" ); |
154 | } |
155 | |
156 | /* ARRAY is an array of vectors created by create_vector_array. |
157 | Return an SSA_NAME for the vector in index N. The reference |
158 | is part of the vectorization of STMT_INFO and the vector is associated |
159 | with scalar destination SCALAR_DEST. */ |
160 | |
161 | static tree |
162 | read_vector_array (vec_info *vinfo, |
163 | stmt_vec_info stmt_info, gimple_stmt_iterator *gsi, |
164 | tree scalar_dest, tree array, unsigned HOST_WIDE_INT n) |
165 | { |
166 | tree vect_type, vect, vect_name, array_ref; |
167 | gimple *new_stmt; |
168 | |
169 | gcc_assert (TREE_CODE (TREE_TYPE (array)) == ARRAY_TYPE); |
170 | vect_type = TREE_TYPE (TREE_TYPE (array)); |
171 | vect = vect_create_destination_var (scalar_dest, vect_type); |
172 | array_ref = build4 (ARRAY_REF, vect_type, array, |
173 | build_int_cst (size_type_node, n), |
174 | NULL_TREE, NULL_TREE); |
175 | |
176 | new_stmt = gimple_build_assign (vect, array_ref); |
177 | vect_name = make_ssa_name (var: vect, stmt: new_stmt); |
178 | gimple_assign_set_lhs (gs: new_stmt, lhs: vect_name); |
179 | vect_finish_stmt_generation (vinfo, stmt_info, new_stmt, gsi); |
180 | |
181 | return vect_name; |
182 | } |
183 | |
184 | /* ARRAY is an array of vectors created by create_vector_array. |
185 | Emit code to store SSA_NAME VECT in index N of the array. |
186 | The store is part of the vectorization of STMT_INFO. */ |
187 | |
188 | static void |
189 | write_vector_array (vec_info *vinfo, |
190 | stmt_vec_info stmt_info, gimple_stmt_iterator *gsi, |
191 | tree vect, tree array, unsigned HOST_WIDE_INT n) |
192 | { |
193 | tree array_ref; |
194 | gimple *new_stmt; |
195 | |
196 | array_ref = build4 (ARRAY_REF, TREE_TYPE (vect), array, |
197 | build_int_cst (size_type_node, n), |
198 | NULL_TREE, NULL_TREE); |
199 | |
200 | new_stmt = gimple_build_assign (array_ref, vect); |
201 | vect_finish_stmt_generation (vinfo, stmt_info, new_stmt, gsi); |
202 | } |
203 | |
204 | /* PTR is a pointer to an array of type TYPE. Return a representation |
205 | of *PTR. The memory reference replaces those in FIRST_DR |
206 | (and its group). */ |
207 | |
208 | static tree |
209 | create_array_ref (tree type, tree ptr, tree alias_ptr_type) |
210 | { |
211 | tree mem_ref; |
212 | |
213 | mem_ref = build2 (MEM_REF, type, ptr, build_int_cst (alias_ptr_type, 0)); |
214 | /* Arrays have the same alignment as their type. */ |
215 | set_ptr_info_alignment (get_ptr_info (ptr), TYPE_ALIGN_UNIT (type), 0); |
216 | return mem_ref; |
217 | } |
218 | |
219 | /* Add a clobber of variable VAR to the vectorization of STMT_INFO. |
220 | Emit the clobber before *GSI. */ |
221 | |
222 | static void |
223 | vect_clobber_variable (vec_info *vinfo, stmt_vec_info stmt_info, |
224 | gimple_stmt_iterator *gsi, tree var) |
225 | { |
226 | tree clobber = build_clobber (TREE_TYPE (var)); |
227 | gimple *new_stmt = gimple_build_assign (var, clobber); |
228 | vect_finish_stmt_generation (vinfo, stmt_info, new_stmt, gsi); |
229 | } |
230 | |
231 | /* Utility functions used by vect_mark_stmts_to_be_vectorized. */ |
232 | |
233 | /* Function vect_mark_relevant. |
234 | |
235 | Mark STMT_INFO as "relevant for vectorization" and add it to WORKLIST. */ |
236 | |
237 | static void |
238 | vect_mark_relevant (vec<stmt_vec_info> *worklist, stmt_vec_info stmt_info, |
239 | enum vect_relevant relevant, bool live_p) |
240 | { |
241 | enum vect_relevant save_relevant = STMT_VINFO_RELEVANT (stmt_info); |
242 | bool save_live_p = STMT_VINFO_LIVE_P (stmt_info); |
243 | |
244 | if (dump_enabled_p ()) |
245 | dump_printf_loc (MSG_NOTE, vect_location, |
246 | "mark relevant %d, live %d: %G" , relevant, live_p, |
247 | stmt_info->stmt); |
248 | |
249 | /* If this stmt is an original stmt in a pattern, we might need to mark its |
250 | related pattern stmt instead of the original stmt. However, such stmts |
251 | may have their own uses that are not in any pattern, in such cases the |
252 | stmt itself should be marked. */ |
253 | if (STMT_VINFO_IN_PATTERN_P (stmt_info)) |
254 | { |
255 | /* This is the last stmt in a sequence that was detected as a |
256 | pattern that can potentially be vectorized. Don't mark the stmt |
257 | as relevant/live because it's not going to be vectorized. |
258 | Instead mark the pattern-stmt that replaces it. */ |
259 | |
260 | if (dump_enabled_p ()) |
261 | dump_printf_loc (MSG_NOTE, vect_location, |
262 | "last stmt in pattern. don't mark" |
263 | " relevant/live.\n" ); |
264 | |
265 | stmt_vec_info old_stmt_info = stmt_info; |
266 | stmt_info = STMT_VINFO_RELATED_STMT (stmt_info); |
267 | gcc_assert (STMT_VINFO_RELATED_STMT (stmt_info) == old_stmt_info); |
268 | save_relevant = STMT_VINFO_RELEVANT (stmt_info); |
269 | save_live_p = STMT_VINFO_LIVE_P (stmt_info); |
270 | |
271 | if (live_p && relevant == vect_unused_in_scope) |
272 | { |
273 | if (dump_enabled_p ()) |
274 | dump_printf_loc (MSG_NOTE, vect_location, |
275 | "vec_stmt_relevant_p: forcing live pattern stmt " |
276 | "relevant.\n" ); |
277 | relevant = vect_used_only_live; |
278 | } |
279 | |
280 | if (dump_enabled_p ()) |
281 | dump_printf_loc (MSG_NOTE, vect_location, |
282 | "mark relevant %d, live %d: %G" , relevant, live_p, |
283 | stmt_info->stmt); |
284 | } |
285 | |
286 | STMT_VINFO_LIVE_P (stmt_info) |= live_p; |
287 | if (relevant > STMT_VINFO_RELEVANT (stmt_info)) |
288 | STMT_VINFO_RELEVANT (stmt_info) = relevant; |
289 | |
290 | if (STMT_VINFO_RELEVANT (stmt_info) == save_relevant |
291 | && STMT_VINFO_LIVE_P (stmt_info) == save_live_p) |
292 | { |
293 | if (dump_enabled_p ()) |
294 | dump_printf_loc (MSG_NOTE, vect_location, |
295 | "already marked relevant/live.\n" ); |
296 | return; |
297 | } |
298 | |
299 | worklist->safe_push (obj: stmt_info); |
300 | } |
301 | |
302 | |
303 | /* Function is_simple_and_all_uses_invariant |
304 | |
305 | Return true if STMT_INFO is simple and all uses of it are invariant. */ |
306 | |
307 | bool |
308 | is_simple_and_all_uses_invariant (stmt_vec_info stmt_info, |
309 | loop_vec_info loop_vinfo) |
310 | { |
311 | tree op; |
312 | ssa_op_iter iter; |
313 | |
314 | gassign *stmt = dyn_cast <gassign *> (p: stmt_info->stmt); |
315 | if (!stmt) |
316 | return false; |
317 | |
318 | FOR_EACH_SSA_TREE_OPERAND (op, stmt, iter, SSA_OP_USE) |
319 | { |
320 | enum vect_def_type dt = vect_uninitialized_def; |
321 | |
322 | if (!vect_is_simple_use (op, loop_vinfo, &dt)) |
323 | { |
324 | if (dump_enabled_p ()) |
325 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
326 | "use not simple.\n" ); |
327 | return false; |
328 | } |
329 | |
330 | if (dt != vect_external_def && dt != vect_constant_def) |
331 | return false; |
332 | } |
333 | return true; |
334 | } |
335 | |
336 | /* Function vect_stmt_relevant_p. |
337 | |
338 | Return true if STMT_INFO, in the loop that is represented by LOOP_VINFO, |
339 | is "relevant for vectorization". |
340 | |
341 | A stmt is considered "relevant for vectorization" if: |
342 | - it has uses outside the loop. |
343 | - it has vdefs (it alters memory). |
344 | - control stmts in the loop (except for the exit condition). |
345 | |
346 | CHECKME: what other side effects would the vectorizer allow? */ |
347 | |
348 | static bool |
349 | vect_stmt_relevant_p (stmt_vec_info stmt_info, loop_vec_info loop_vinfo, |
350 | enum vect_relevant *relevant, bool *live_p) |
351 | { |
352 | class loop *loop = LOOP_VINFO_LOOP (loop_vinfo); |
353 | ssa_op_iter op_iter; |
354 | imm_use_iterator imm_iter; |
355 | use_operand_p use_p; |
356 | def_operand_p def_p; |
357 | |
358 | *relevant = vect_unused_in_scope; |
359 | *live_p = false; |
360 | |
361 | /* cond stmt other than loop exit cond. */ |
362 | if (is_ctrl_stmt (stmt_info->stmt) |
363 | && STMT_VINFO_TYPE (stmt_info) != loop_exit_ctrl_vec_info_type) |
364 | *relevant = vect_used_in_scope; |
365 | |
366 | /* changing memory. */ |
367 | if (gimple_code (g: stmt_info->stmt) != GIMPLE_PHI) |
368 | if (gimple_vdef (g: stmt_info->stmt) |
369 | && !gimple_clobber_p (s: stmt_info->stmt)) |
370 | { |
371 | if (dump_enabled_p ()) |
372 | dump_printf_loc (MSG_NOTE, vect_location, |
373 | "vec_stmt_relevant_p: stmt has vdefs.\n" ); |
374 | *relevant = vect_used_in_scope; |
375 | } |
376 | |
377 | /* uses outside the loop. */ |
378 | FOR_EACH_PHI_OR_STMT_DEF (def_p, stmt_info->stmt, op_iter, SSA_OP_DEF) |
379 | { |
380 | FOR_EACH_IMM_USE_FAST (use_p, imm_iter, DEF_FROM_PTR (def_p)) |
381 | { |
382 | basic_block bb = gimple_bb (USE_STMT (use_p)); |
383 | if (!flow_bb_inside_loop_p (loop, bb)) |
384 | { |
385 | if (is_gimple_debug (USE_STMT (use_p))) |
386 | continue; |
387 | |
388 | if (dump_enabled_p ()) |
389 | dump_printf_loc (MSG_NOTE, vect_location, |
390 | "vec_stmt_relevant_p: used out of loop.\n" ); |
391 | |
392 | /* We expect all such uses to be in the loop exit phis |
393 | (because of loop closed form) */ |
394 | gcc_assert (gimple_code (USE_STMT (use_p)) == GIMPLE_PHI); |
395 | gcc_assert (bb == single_exit (loop)->dest); |
396 | |
397 | *live_p = true; |
398 | } |
399 | } |
400 | } |
401 | |
402 | if (*live_p && *relevant == vect_unused_in_scope |
403 | && !is_simple_and_all_uses_invariant (stmt_info, loop_vinfo)) |
404 | { |
405 | if (dump_enabled_p ()) |
406 | dump_printf_loc (MSG_NOTE, vect_location, |
407 | "vec_stmt_relevant_p: stmt live but not relevant.\n" ); |
408 | *relevant = vect_used_only_live; |
409 | } |
410 | |
411 | return (*live_p || *relevant); |
412 | } |
413 | |
414 | |
415 | /* Function exist_non_indexing_operands_for_use_p |
416 | |
417 | USE is one of the uses attached to STMT_INFO. Check if USE is |
418 | used in STMT_INFO for anything other than indexing an array. */ |
419 | |
420 | static bool |
421 | exist_non_indexing_operands_for_use_p (tree use, stmt_vec_info stmt_info) |
422 | { |
423 | tree operand; |
424 | |
425 | /* USE corresponds to some operand in STMT. If there is no data |
426 | reference in STMT, then any operand that corresponds to USE |
427 | is not indexing an array. */ |
428 | if (!STMT_VINFO_DATA_REF (stmt_info)) |
429 | return true; |
430 | |
431 | /* STMT has a data_ref. FORNOW this means that its of one of |
432 | the following forms: |
433 | -1- ARRAY_REF = var |
434 | -2- var = ARRAY_REF |
435 | (This should have been verified in analyze_data_refs). |
436 | |
437 | 'var' in the second case corresponds to a def, not a use, |
438 | so USE cannot correspond to any operands that are not used |
439 | for array indexing. |
440 | |
441 | Therefore, all we need to check is if STMT falls into the |
442 | first case, and whether var corresponds to USE. */ |
443 | |
444 | gassign *assign = dyn_cast <gassign *> (p: stmt_info->stmt); |
445 | if (!assign || !gimple_assign_copy_p (assign)) |
446 | { |
447 | gcall *call = dyn_cast <gcall *> (p: stmt_info->stmt); |
448 | if (call && gimple_call_internal_p (gs: call)) |
449 | { |
450 | internal_fn ifn = gimple_call_internal_fn (gs: call); |
451 | int mask_index = internal_fn_mask_index (ifn); |
452 | if (mask_index >= 0 |
453 | && use == gimple_call_arg (gs: call, index: mask_index)) |
454 | return true; |
455 | int stored_value_index = internal_fn_stored_value_index (ifn); |
456 | if (stored_value_index >= 0 |
457 | && use == gimple_call_arg (gs: call, index: stored_value_index)) |
458 | return true; |
459 | if (internal_gather_scatter_fn_p (ifn) |
460 | && use == gimple_call_arg (gs: call, index: 1)) |
461 | return true; |
462 | } |
463 | return false; |
464 | } |
465 | |
466 | if (TREE_CODE (gimple_assign_lhs (assign)) == SSA_NAME) |
467 | return false; |
468 | operand = gimple_assign_rhs1 (gs: assign); |
469 | if (TREE_CODE (operand) != SSA_NAME) |
470 | return false; |
471 | |
472 | if (operand == use) |
473 | return true; |
474 | |
475 | return false; |
476 | } |
477 | |
478 | |
479 | /* |
480 | Function process_use. |
481 | |
482 | Inputs: |
483 | - a USE in STMT_VINFO in a loop represented by LOOP_VINFO |
484 | - RELEVANT - enum value to be set in the STMT_VINFO of the stmt |
485 | that defined USE. This is done by calling mark_relevant and passing it |
486 | the WORKLIST (to add DEF_STMT to the WORKLIST in case it is relevant). |
487 | - FORCE is true if exist_non_indexing_operands_for_use_p check shouldn't |
488 | be performed. |
489 | |
490 | Outputs: |
491 | Generally, LIVE_P and RELEVANT are used to define the liveness and |
492 | relevance info of the DEF_STMT of this USE: |
493 | STMT_VINFO_LIVE_P (DEF_stmt_vinfo) <-- live_p |
494 | STMT_VINFO_RELEVANT (DEF_stmt_vinfo) <-- relevant |
495 | Exceptions: |
496 | - case 1: If USE is used only for address computations (e.g. array indexing), |
497 | which does not need to be directly vectorized, then the liveness/relevance |
498 | of the respective DEF_STMT is left unchanged. |
499 | - case 2: If STMT_VINFO is a reduction phi and DEF_STMT is a reduction stmt, |
500 | we skip DEF_STMT cause it had already been processed. |
501 | - case 3: If DEF_STMT and STMT_VINFO are in different nests, then |
502 | "relevant" will be modified accordingly. |
503 | |
504 | Return true if everything is as expected. Return false otherwise. */ |
505 | |
506 | static opt_result |
507 | process_use (stmt_vec_info stmt_vinfo, tree use, loop_vec_info loop_vinfo, |
508 | enum vect_relevant relevant, vec<stmt_vec_info> *worklist, |
509 | bool force) |
510 | { |
511 | stmt_vec_info dstmt_vinfo; |
512 | enum vect_def_type dt; |
513 | |
514 | /* case 1: we are only interested in uses that need to be vectorized. Uses |
515 | that are used for address computation are not considered relevant. */ |
516 | if (!force && !exist_non_indexing_operands_for_use_p (use, stmt_info: stmt_vinfo)) |
517 | return opt_result::success (); |
518 | |
519 | if (!vect_is_simple_use (use, loop_vinfo, &dt, &dstmt_vinfo)) |
520 | return opt_result::failure_at (loc: stmt_vinfo->stmt, |
521 | fmt: "not vectorized:" |
522 | " unsupported use in stmt.\n" ); |
523 | |
524 | if (!dstmt_vinfo) |
525 | return opt_result::success (); |
526 | |
527 | basic_block def_bb = gimple_bb (g: dstmt_vinfo->stmt); |
528 | basic_block bb = gimple_bb (g: stmt_vinfo->stmt); |
529 | |
530 | /* case 2: A reduction phi (STMT) defined by a reduction stmt (DSTMT_VINFO). |
531 | We have to force the stmt live since the epilogue loop needs it to |
532 | continue computing the reduction. */ |
533 | if (gimple_code (g: stmt_vinfo->stmt) == GIMPLE_PHI |
534 | && STMT_VINFO_DEF_TYPE (stmt_vinfo) == vect_reduction_def |
535 | && gimple_code (g: dstmt_vinfo->stmt) != GIMPLE_PHI |
536 | && STMT_VINFO_DEF_TYPE (dstmt_vinfo) == vect_reduction_def |
537 | && bb->loop_father == def_bb->loop_father) |
538 | { |
539 | if (dump_enabled_p ()) |
540 | dump_printf_loc (MSG_NOTE, vect_location, |
541 | "reduc-stmt defining reduc-phi in the same nest.\n" ); |
542 | vect_mark_relevant (worklist, stmt_info: dstmt_vinfo, relevant, live_p: true); |
543 | return opt_result::success (); |
544 | } |
545 | |
546 | /* case 3a: outer-loop stmt defining an inner-loop stmt: |
547 | outer-loop-header-bb: |
548 | d = dstmt_vinfo |
549 | inner-loop: |
550 | stmt # use (d) |
551 | outer-loop-tail-bb: |
552 | ... */ |
553 | if (flow_loop_nested_p (def_bb->loop_father, bb->loop_father)) |
554 | { |
555 | if (dump_enabled_p ()) |
556 | dump_printf_loc (MSG_NOTE, vect_location, |
557 | "outer-loop def-stmt defining inner-loop stmt.\n" ); |
558 | |
559 | switch (relevant) |
560 | { |
561 | case vect_unused_in_scope: |
562 | relevant = (STMT_VINFO_DEF_TYPE (stmt_vinfo) == vect_nested_cycle) ? |
563 | vect_used_in_scope : vect_unused_in_scope; |
564 | break; |
565 | |
566 | case vect_used_in_outer_by_reduction: |
567 | gcc_assert (STMT_VINFO_DEF_TYPE (stmt_vinfo) != vect_reduction_def); |
568 | relevant = vect_used_by_reduction; |
569 | break; |
570 | |
571 | case vect_used_in_outer: |
572 | gcc_assert (STMT_VINFO_DEF_TYPE (stmt_vinfo) != vect_reduction_def); |
573 | relevant = vect_used_in_scope; |
574 | break; |
575 | |
576 | case vect_used_in_scope: |
577 | break; |
578 | |
579 | default: |
580 | gcc_unreachable (); |
581 | } |
582 | } |
583 | |
584 | /* case 3b: inner-loop stmt defining an outer-loop stmt: |
585 | outer-loop-header-bb: |
586 | ... |
587 | inner-loop: |
588 | d = dstmt_vinfo |
589 | outer-loop-tail-bb (or outer-loop-exit-bb in double reduction): |
590 | stmt # use (d) */ |
591 | else if (flow_loop_nested_p (bb->loop_father, def_bb->loop_father)) |
592 | { |
593 | if (dump_enabled_p ()) |
594 | dump_printf_loc (MSG_NOTE, vect_location, |
595 | "inner-loop def-stmt defining outer-loop stmt.\n" ); |
596 | |
597 | switch (relevant) |
598 | { |
599 | case vect_unused_in_scope: |
600 | relevant = (STMT_VINFO_DEF_TYPE (stmt_vinfo) == vect_reduction_def |
601 | || STMT_VINFO_DEF_TYPE (stmt_vinfo) == vect_double_reduction_def) ? |
602 | vect_used_in_outer_by_reduction : vect_unused_in_scope; |
603 | break; |
604 | |
605 | case vect_used_by_reduction: |
606 | case vect_used_only_live: |
607 | relevant = vect_used_in_outer_by_reduction; |
608 | break; |
609 | |
610 | case vect_used_in_scope: |
611 | relevant = vect_used_in_outer; |
612 | break; |
613 | |
614 | default: |
615 | gcc_unreachable (); |
616 | } |
617 | } |
618 | /* We are also not interested in uses on loop PHI backedges that are |
619 | inductions. Otherwise we'll needlessly vectorize the IV increment |
620 | and cause hybrid SLP for SLP inductions. Unless the PHI is live |
621 | of course. */ |
622 | else if (gimple_code (g: stmt_vinfo->stmt) == GIMPLE_PHI |
623 | && STMT_VINFO_DEF_TYPE (stmt_vinfo) == vect_induction_def |
624 | && ! STMT_VINFO_LIVE_P (stmt_vinfo) |
625 | && (PHI_ARG_DEF_FROM_EDGE (stmt_vinfo->stmt, |
626 | loop_latch_edge (bb->loop_father)) |
627 | == use)) |
628 | { |
629 | if (dump_enabled_p ()) |
630 | dump_printf_loc (MSG_NOTE, vect_location, |
631 | "induction value on backedge.\n" ); |
632 | return opt_result::success (); |
633 | } |
634 | |
635 | |
636 | vect_mark_relevant (worklist, stmt_info: dstmt_vinfo, relevant, live_p: false); |
637 | return opt_result::success (); |
638 | } |
639 | |
640 | |
641 | /* Function vect_mark_stmts_to_be_vectorized. |
642 | |
643 | Not all stmts in the loop need to be vectorized. For example: |
644 | |
645 | for i... |
646 | for j... |
647 | 1. T0 = i + j |
648 | 2. T1 = a[T0] |
649 | |
650 | 3. j = j + 1 |
651 | |
652 | Stmt 1 and 3 do not need to be vectorized, because loop control and |
653 | addressing of vectorized data-refs are handled differently. |
654 | |
655 | This pass detects such stmts. */ |
656 | |
657 | opt_result |
658 | vect_mark_stmts_to_be_vectorized (loop_vec_info loop_vinfo, bool *fatal) |
659 | { |
660 | class loop *loop = LOOP_VINFO_LOOP (loop_vinfo); |
661 | basic_block *bbs = LOOP_VINFO_BBS (loop_vinfo); |
662 | unsigned int nbbs = loop->num_nodes; |
663 | gimple_stmt_iterator si; |
664 | unsigned int i; |
665 | basic_block bb; |
666 | bool live_p; |
667 | enum vect_relevant relevant; |
668 | |
669 | DUMP_VECT_SCOPE ("vect_mark_stmts_to_be_vectorized" ); |
670 | |
671 | auto_vec<stmt_vec_info, 64> worklist; |
672 | |
673 | /* 1. Init worklist. */ |
674 | for (i = 0; i < nbbs; i++) |
675 | { |
676 | bb = bbs[i]; |
677 | for (si = gsi_start_phis (bb); !gsi_end_p (i: si); gsi_next (i: &si)) |
678 | { |
679 | stmt_vec_info phi_info = loop_vinfo->lookup_stmt (gsi_stmt (i: si)); |
680 | if (dump_enabled_p ()) |
681 | dump_printf_loc (MSG_NOTE, vect_location, "init: phi relevant? %G" , |
682 | phi_info->stmt); |
683 | |
684 | if (vect_stmt_relevant_p (stmt_info: phi_info, loop_vinfo, relevant: &relevant, live_p: &live_p)) |
685 | vect_mark_relevant (worklist: &worklist, stmt_info: phi_info, relevant, live_p); |
686 | } |
687 | for (si = gsi_start_bb (bb); !gsi_end_p (i: si); gsi_next (i: &si)) |
688 | { |
689 | if (is_gimple_debug (gs: gsi_stmt (i: si))) |
690 | continue; |
691 | stmt_vec_info stmt_info = loop_vinfo->lookup_stmt (gsi_stmt (i: si)); |
692 | if (dump_enabled_p ()) |
693 | dump_printf_loc (MSG_NOTE, vect_location, |
694 | "init: stmt relevant? %G" , stmt_info->stmt); |
695 | |
696 | if (vect_stmt_relevant_p (stmt_info, loop_vinfo, relevant: &relevant, live_p: &live_p)) |
697 | vect_mark_relevant (worklist: &worklist, stmt_info, relevant, live_p); |
698 | } |
699 | } |
700 | |
701 | /* 2. Process_worklist */ |
702 | while (worklist.length () > 0) |
703 | { |
704 | use_operand_p use_p; |
705 | ssa_op_iter iter; |
706 | |
707 | stmt_vec_info stmt_vinfo = worklist.pop (); |
708 | if (dump_enabled_p ()) |
709 | dump_printf_loc (MSG_NOTE, vect_location, |
710 | "worklist: examine stmt: %G" , stmt_vinfo->stmt); |
711 | |
712 | /* Examine the USEs of STMT. For each USE, mark the stmt that defines it |
713 | (DEF_STMT) as relevant/irrelevant according to the relevance property |
714 | of STMT. */ |
715 | relevant = STMT_VINFO_RELEVANT (stmt_vinfo); |
716 | |
717 | /* Generally, the relevance property of STMT (in STMT_VINFO_RELEVANT) is |
718 | propagated as is to the DEF_STMTs of its USEs. |
719 | |
720 | One exception is when STMT has been identified as defining a reduction |
721 | variable; in this case we set the relevance to vect_used_by_reduction. |
722 | This is because we distinguish between two kinds of relevant stmts - |
723 | those that are used by a reduction computation, and those that are |
724 | (also) used by a regular computation. This allows us later on to |
725 | identify stmts that are used solely by a reduction, and therefore the |
726 | order of the results that they produce does not have to be kept. */ |
727 | |
728 | switch (STMT_VINFO_DEF_TYPE (stmt_vinfo)) |
729 | { |
730 | case vect_reduction_def: |
731 | gcc_assert (relevant != vect_unused_in_scope); |
732 | if (relevant != vect_unused_in_scope |
733 | && relevant != vect_used_in_scope |
734 | && relevant != vect_used_by_reduction |
735 | && relevant != vect_used_only_live) |
736 | return opt_result::failure_at |
737 | (loc: stmt_vinfo->stmt, fmt: "unsupported use of reduction.\n" ); |
738 | break; |
739 | |
740 | case vect_nested_cycle: |
741 | if (relevant != vect_unused_in_scope |
742 | && relevant != vect_used_in_outer_by_reduction |
743 | && relevant != vect_used_in_outer) |
744 | return opt_result::failure_at |
745 | (loc: stmt_vinfo->stmt, fmt: "unsupported use of nested cycle.\n" ); |
746 | break; |
747 | |
748 | case vect_double_reduction_def: |
749 | if (relevant != vect_unused_in_scope |
750 | && relevant != vect_used_by_reduction |
751 | && relevant != vect_used_only_live) |
752 | return opt_result::failure_at |
753 | (loc: stmt_vinfo->stmt, fmt: "unsupported use of double reduction.\n" ); |
754 | break; |
755 | |
756 | default: |
757 | break; |
758 | } |
759 | |
760 | if (is_pattern_stmt_p (stmt_info: stmt_vinfo)) |
761 | { |
762 | /* Pattern statements are not inserted into the code, so |
763 | FOR_EACH_PHI_OR_STMT_USE optimizes their operands out, and we |
764 | have to scan the RHS or function arguments instead. */ |
765 | if (gassign *assign = dyn_cast <gassign *> (p: stmt_vinfo->stmt)) |
766 | { |
767 | enum tree_code rhs_code = gimple_assign_rhs_code (gs: assign); |
768 | tree op = gimple_assign_rhs1 (gs: assign); |
769 | |
770 | i = 1; |
771 | if (rhs_code == COND_EXPR && COMPARISON_CLASS_P (op)) |
772 | { |
773 | opt_result res |
774 | = process_use (stmt_vinfo, TREE_OPERAND (op, 0), |
775 | loop_vinfo, relevant, worklist: &worklist, force: false); |
776 | if (!res) |
777 | return res; |
778 | res = process_use (stmt_vinfo, TREE_OPERAND (op, 1), |
779 | loop_vinfo, relevant, worklist: &worklist, force: false); |
780 | if (!res) |
781 | return res; |
782 | i = 2; |
783 | } |
784 | for (; i < gimple_num_ops (gs: assign); i++) |
785 | { |
786 | op = gimple_op (gs: assign, i); |
787 | if (TREE_CODE (op) == SSA_NAME) |
788 | { |
789 | opt_result res |
790 | = process_use (stmt_vinfo, use: op, loop_vinfo, relevant, |
791 | worklist: &worklist, force: false); |
792 | if (!res) |
793 | return res; |
794 | } |
795 | } |
796 | } |
797 | else if (gcall *call = dyn_cast <gcall *> (p: stmt_vinfo->stmt)) |
798 | { |
799 | for (i = 0; i < gimple_call_num_args (gs: call); i++) |
800 | { |
801 | tree arg = gimple_call_arg (gs: call, index: i); |
802 | opt_result res |
803 | = process_use (stmt_vinfo, use: arg, loop_vinfo, relevant, |
804 | worklist: &worklist, force: false); |
805 | if (!res) |
806 | return res; |
807 | } |
808 | } |
809 | } |
810 | else |
811 | FOR_EACH_PHI_OR_STMT_USE (use_p, stmt_vinfo->stmt, iter, SSA_OP_USE) |
812 | { |
813 | tree op = USE_FROM_PTR (use_p); |
814 | opt_result res |
815 | = process_use (stmt_vinfo, use: op, loop_vinfo, relevant, |
816 | worklist: &worklist, force: false); |
817 | if (!res) |
818 | return res; |
819 | } |
820 | |
821 | if (STMT_VINFO_GATHER_SCATTER_P (stmt_vinfo)) |
822 | { |
823 | gather_scatter_info gs_info; |
824 | if (!vect_check_gather_scatter (stmt_vinfo, loop_vinfo, &gs_info)) |
825 | gcc_unreachable (); |
826 | opt_result res |
827 | = process_use (stmt_vinfo, use: gs_info.offset, loop_vinfo, relevant, |
828 | worklist: &worklist, force: true); |
829 | if (!res) |
830 | { |
831 | if (fatal) |
832 | *fatal = false; |
833 | return res; |
834 | } |
835 | } |
836 | } /* while worklist */ |
837 | |
838 | return opt_result::success (); |
839 | } |
840 | |
841 | /* Function vect_model_simple_cost. |
842 | |
843 | Models cost for simple operations, i.e. those that only emit ncopies of a |
844 | single op. Right now, this does not account for multiple insns that could |
845 | be generated for the single vector op. We will handle that shortly. */ |
846 | |
847 | static void |
848 | vect_model_simple_cost (vec_info *, |
849 | stmt_vec_info stmt_info, int ncopies, |
850 | enum vect_def_type *dt, |
851 | int ndts, |
852 | slp_tree node, |
853 | stmt_vector_for_cost *cost_vec, |
854 | vect_cost_for_stmt kind = vector_stmt) |
855 | { |
856 | int inside_cost = 0, prologue_cost = 0; |
857 | |
858 | gcc_assert (cost_vec != NULL); |
859 | |
860 | /* ??? Somehow we need to fix this at the callers. */ |
861 | if (node) |
862 | ncopies = SLP_TREE_NUMBER_OF_VEC_STMTS (node); |
863 | |
864 | if (!node) |
865 | /* Cost the "broadcast" of a scalar operand in to a vector operand. |
866 | Use scalar_to_vec to cost the broadcast, as elsewhere in the vector |
867 | cost model. */ |
868 | for (int i = 0; i < ndts; i++) |
869 | if (dt[i] == vect_constant_def || dt[i] == vect_external_def) |
870 | prologue_cost += record_stmt_cost (body_cost_vec: cost_vec, count: 1, kind: scalar_to_vec, |
871 | stmt_info, misalign: 0, where: vect_prologue); |
872 | |
873 | /* Pass the inside-of-loop statements to the target-specific cost model. */ |
874 | inside_cost += record_stmt_cost (body_cost_vec: cost_vec, count: ncopies, kind, |
875 | stmt_info, misalign: 0, where: vect_body); |
876 | |
877 | if (dump_enabled_p ()) |
878 | dump_printf_loc (MSG_NOTE, vect_location, |
879 | "vect_model_simple_cost: inside_cost = %d, " |
880 | "prologue_cost = %d .\n" , inside_cost, prologue_cost); |
881 | } |
882 | |
883 | |
884 | /* Model cost for type demotion and promotion operations. PWR is |
885 | normally zero for single-step promotions and demotions. It will be |
886 | one if two-step promotion/demotion is required, and so on. NCOPIES |
887 | is the number of vector results (and thus number of instructions) |
888 | for the narrowest end of the operation chain. Each additional |
889 | step doubles the number of instructions required. If WIDEN_ARITH |
890 | is true the stmt is doing widening arithmetic. */ |
891 | |
892 | static void |
893 | vect_model_promotion_demotion_cost (stmt_vec_info stmt_info, |
894 | enum vect_def_type *dt, |
895 | unsigned int ncopies, int pwr, |
896 | stmt_vector_for_cost *cost_vec, |
897 | bool widen_arith) |
898 | { |
899 | int i; |
900 | int inside_cost = 0, prologue_cost = 0; |
901 | |
902 | for (i = 0; i < pwr + 1; i++) |
903 | { |
904 | inside_cost += record_stmt_cost (body_cost_vec: cost_vec, count: ncopies, |
905 | kind: widen_arith |
906 | ? vector_stmt : vec_promote_demote, |
907 | stmt_info, misalign: 0, where: vect_body); |
908 | ncopies *= 2; |
909 | } |
910 | |
911 | /* FORNOW: Assuming maximum 2 args per stmts. */ |
912 | for (i = 0; i < 2; i++) |
913 | if (dt[i] == vect_constant_def || dt[i] == vect_external_def) |
914 | prologue_cost += record_stmt_cost (body_cost_vec: cost_vec, count: 1, kind: vector_stmt, |
915 | stmt_info, misalign: 0, where: vect_prologue); |
916 | |
917 | if (dump_enabled_p ()) |
918 | dump_printf_loc (MSG_NOTE, vect_location, |
919 | "vect_model_promotion_demotion_cost: inside_cost = %d, " |
920 | "prologue_cost = %d .\n" , inside_cost, prologue_cost); |
921 | } |
922 | |
923 | /* Returns true if the current function returns DECL. */ |
924 | |
925 | static bool |
926 | cfun_returns (tree decl) |
927 | { |
928 | edge_iterator ei; |
929 | edge e; |
930 | FOR_EACH_EDGE (e, ei, EXIT_BLOCK_PTR_FOR_FN (cfun)->preds) |
931 | { |
932 | greturn *ret = safe_dyn_cast <greturn *> (p: *gsi_last_bb (bb: e->src)); |
933 | if (!ret) |
934 | continue; |
935 | if (gimple_return_retval (gs: ret) == decl) |
936 | return true; |
937 | /* We often end up with an aggregate copy to the result decl, |
938 | handle that case as well. First skip intermediate clobbers |
939 | though. */ |
940 | gimple *def = ret; |
941 | do |
942 | { |
943 | def = SSA_NAME_DEF_STMT (gimple_vuse (def)); |
944 | } |
945 | while (gimple_clobber_p (s: def)); |
946 | if (is_a <gassign *> (p: def) |
947 | && gimple_assign_lhs (gs: def) == gimple_return_retval (gs: ret) |
948 | && gimple_assign_rhs1 (gs: def) == decl) |
949 | return true; |
950 | } |
951 | return false; |
952 | } |
953 | |
954 | /* Calculate cost of DR's memory access. */ |
955 | void |
956 | vect_get_store_cost (vec_info *, stmt_vec_info stmt_info, int ncopies, |
957 | dr_alignment_support alignment_support_scheme, |
958 | int misalignment, |
959 | unsigned int *inside_cost, |
960 | stmt_vector_for_cost *body_cost_vec) |
961 | { |
962 | switch (alignment_support_scheme) |
963 | { |
964 | case dr_aligned: |
965 | { |
966 | *inside_cost += record_stmt_cost (body_cost_vec, count: ncopies, |
967 | kind: vector_store, stmt_info, misalign: 0, |
968 | where: vect_body); |
969 | |
970 | if (dump_enabled_p ()) |
971 | dump_printf_loc (MSG_NOTE, vect_location, |
972 | "vect_model_store_cost: aligned.\n" ); |
973 | break; |
974 | } |
975 | |
976 | case dr_unaligned_supported: |
977 | { |
978 | /* Here, we assign an additional cost for the unaligned store. */ |
979 | *inside_cost += record_stmt_cost (body_cost_vec, count: ncopies, |
980 | kind: unaligned_store, stmt_info, |
981 | misalign: misalignment, where: vect_body); |
982 | if (dump_enabled_p ()) |
983 | dump_printf_loc (MSG_NOTE, vect_location, |
984 | "vect_model_store_cost: unaligned supported by " |
985 | "hardware.\n" ); |
986 | break; |
987 | } |
988 | |
989 | case dr_unaligned_unsupported: |
990 | { |
991 | *inside_cost = VECT_MAX_COST; |
992 | |
993 | if (dump_enabled_p ()) |
994 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
995 | "vect_model_store_cost: unsupported access.\n" ); |
996 | break; |
997 | } |
998 | |
999 | default: |
1000 | gcc_unreachable (); |
1001 | } |
1002 | } |
1003 | |
1004 | /* Calculate cost of DR's memory access. */ |
1005 | void |
1006 | vect_get_load_cost (vec_info *, stmt_vec_info stmt_info, int ncopies, |
1007 | dr_alignment_support alignment_support_scheme, |
1008 | int misalignment, |
1009 | bool add_realign_cost, unsigned int *inside_cost, |
1010 | unsigned int *prologue_cost, |
1011 | stmt_vector_for_cost *prologue_cost_vec, |
1012 | stmt_vector_for_cost *body_cost_vec, |
1013 | bool record_prologue_costs) |
1014 | { |
1015 | switch (alignment_support_scheme) |
1016 | { |
1017 | case dr_aligned: |
1018 | { |
1019 | *inside_cost += record_stmt_cost (body_cost_vec, count: ncopies, kind: vector_load, |
1020 | stmt_info, misalign: 0, where: vect_body); |
1021 | |
1022 | if (dump_enabled_p ()) |
1023 | dump_printf_loc (MSG_NOTE, vect_location, |
1024 | "vect_model_load_cost: aligned.\n" ); |
1025 | |
1026 | break; |
1027 | } |
1028 | case dr_unaligned_supported: |
1029 | { |
1030 | /* Here, we assign an additional cost for the unaligned load. */ |
1031 | *inside_cost += record_stmt_cost (body_cost_vec, count: ncopies, |
1032 | kind: unaligned_load, stmt_info, |
1033 | misalign: misalignment, where: vect_body); |
1034 | |
1035 | if (dump_enabled_p ()) |
1036 | dump_printf_loc (MSG_NOTE, vect_location, |
1037 | "vect_model_load_cost: unaligned supported by " |
1038 | "hardware.\n" ); |
1039 | |
1040 | break; |
1041 | } |
1042 | case dr_explicit_realign: |
1043 | { |
1044 | *inside_cost += record_stmt_cost (body_cost_vec, count: ncopies * 2, |
1045 | kind: vector_load, stmt_info, misalign: 0, where: vect_body); |
1046 | *inside_cost += record_stmt_cost (body_cost_vec, count: ncopies, |
1047 | kind: vec_perm, stmt_info, misalign: 0, where: vect_body); |
1048 | |
1049 | /* FIXME: If the misalignment remains fixed across the iterations of |
1050 | the containing loop, the following cost should be added to the |
1051 | prologue costs. */ |
1052 | if (targetm.vectorize.builtin_mask_for_load) |
1053 | *inside_cost += record_stmt_cost (body_cost_vec, count: 1, kind: vector_stmt, |
1054 | stmt_info, misalign: 0, where: vect_body); |
1055 | |
1056 | if (dump_enabled_p ()) |
1057 | dump_printf_loc (MSG_NOTE, vect_location, |
1058 | "vect_model_load_cost: explicit realign\n" ); |
1059 | |
1060 | break; |
1061 | } |
1062 | case dr_explicit_realign_optimized: |
1063 | { |
1064 | if (dump_enabled_p ()) |
1065 | dump_printf_loc (MSG_NOTE, vect_location, |
1066 | "vect_model_load_cost: unaligned software " |
1067 | "pipelined.\n" ); |
1068 | |
1069 | /* Unaligned software pipeline has a load of an address, an initial |
1070 | load, and possibly a mask operation to "prime" the loop. However, |
1071 | if this is an access in a group of loads, which provide grouped |
1072 | access, then the above cost should only be considered for one |
1073 | access in the group. Inside the loop, there is a load op |
1074 | and a realignment op. */ |
1075 | |
1076 | if (add_realign_cost && record_prologue_costs) |
1077 | { |
1078 | *prologue_cost += record_stmt_cost (body_cost_vec: prologue_cost_vec, count: 2, |
1079 | kind: vector_stmt, stmt_info, |
1080 | misalign: 0, where: vect_prologue); |
1081 | if (targetm.vectorize.builtin_mask_for_load) |
1082 | *prologue_cost += record_stmt_cost (body_cost_vec: prologue_cost_vec, count: 1, |
1083 | kind: vector_stmt, stmt_info, |
1084 | misalign: 0, where: vect_prologue); |
1085 | } |
1086 | |
1087 | *inside_cost += record_stmt_cost (body_cost_vec, count: ncopies, kind: vector_load, |
1088 | stmt_info, misalign: 0, where: vect_body); |
1089 | *inside_cost += record_stmt_cost (body_cost_vec, count: ncopies, kind: vec_perm, |
1090 | stmt_info, misalign: 0, where: vect_body); |
1091 | |
1092 | if (dump_enabled_p ()) |
1093 | dump_printf_loc (MSG_NOTE, vect_location, |
1094 | "vect_model_load_cost: explicit realign optimized" |
1095 | "\n" ); |
1096 | |
1097 | break; |
1098 | } |
1099 | |
1100 | case dr_unaligned_unsupported: |
1101 | { |
1102 | *inside_cost = VECT_MAX_COST; |
1103 | |
1104 | if (dump_enabled_p ()) |
1105 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
1106 | "vect_model_load_cost: unsupported access.\n" ); |
1107 | break; |
1108 | } |
1109 | |
1110 | default: |
1111 | gcc_unreachable (); |
1112 | } |
1113 | } |
1114 | |
1115 | /* Insert the new stmt NEW_STMT at *GSI or at the appropriate place in |
1116 | the loop preheader for the vectorized stmt STMT_VINFO. */ |
1117 | |
1118 | static void |
1119 | vect_init_vector_1 (vec_info *vinfo, stmt_vec_info stmt_vinfo, gimple *new_stmt, |
1120 | gimple_stmt_iterator *gsi) |
1121 | { |
1122 | if (gsi) |
1123 | vect_finish_stmt_generation (vinfo, stmt_vinfo, new_stmt, gsi); |
1124 | else |
1125 | vinfo->insert_on_entry (stmt_vinfo, new_stmt); |
1126 | |
1127 | if (dump_enabled_p ()) |
1128 | dump_printf_loc (MSG_NOTE, vect_location, |
1129 | "created new init_stmt: %G" , new_stmt); |
1130 | } |
1131 | |
1132 | /* Function vect_init_vector. |
1133 | |
1134 | Insert a new stmt (INIT_STMT) that initializes a new variable of type |
1135 | TYPE with the value VAL. If TYPE is a vector type and VAL does not have |
1136 | vector type a vector with all elements equal to VAL is created first. |
1137 | Place the initialization at GSI if it is not NULL. Otherwise, place the |
1138 | initialization at the loop preheader. |
1139 | Return the DEF of INIT_STMT. |
1140 | It will be used in the vectorization of STMT_INFO. */ |
1141 | |
1142 | tree |
1143 | vect_init_vector (vec_info *vinfo, stmt_vec_info stmt_info, tree val, tree type, |
1144 | gimple_stmt_iterator *gsi) |
1145 | { |
1146 | gimple *init_stmt; |
1147 | tree new_temp; |
1148 | |
1149 | /* We abuse this function to push sth to a SSA name with initial 'val'. */ |
1150 | if (! useless_type_conversion_p (type, TREE_TYPE (val))) |
1151 | { |
1152 | gcc_assert (VECTOR_TYPE_P (type)); |
1153 | if (! types_compatible_p (TREE_TYPE (type), TREE_TYPE (val))) |
1154 | { |
1155 | /* Scalar boolean value should be transformed into |
1156 | all zeros or all ones value before building a vector. */ |
1157 | if (VECTOR_BOOLEAN_TYPE_P (type)) |
1158 | { |
1159 | tree true_val = build_all_ones_cst (TREE_TYPE (type)); |
1160 | tree false_val = build_zero_cst (TREE_TYPE (type)); |
1161 | |
1162 | if (CONSTANT_CLASS_P (val)) |
1163 | val = integer_zerop (val) ? false_val : true_val; |
1164 | else |
1165 | { |
1166 | new_temp = make_ssa_name (TREE_TYPE (type)); |
1167 | init_stmt = gimple_build_assign (new_temp, COND_EXPR, |
1168 | val, true_val, false_val); |
1169 | vect_init_vector_1 (vinfo, stmt_vinfo: stmt_info, new_stmt: init_stmt, gsi); |
1170 | val = new_temp; |
1171 | } |
1172 | } |
1173 | else |
1174 | { |
1175 | gimple_seq stmts = NULL; |
1176 | if (! INTEGRAL_TYPE_P (TREE_TYPE (val))) |
1177 | val = gimple_build (seq: &stmts, code: VIEW_CONVERT_EXPR, |
1178 | TREE_TYPE (type), ops: val); |
1179 | else |
1180 | /* ??? Condition vectorization expects us to do |
1181 | promotion of invariant/external defs. */ |
1182 | val = gimple_convert (seq: &stmts, TREE_TYPE (type), op: val); |
1183 | for (gimple_stmt_iterator gsi2 = gsi_start (seq&: stmts); |
1184 | !gsi_end_p (i: gsi2); ) |
1185 | { |
1186 | init_stmt = gsi_stmt (i: gsi2); |
1187 | gsi_remove (&gsi2, false); |
1188 | vect_init_vector_1 (vinfo, stmt_vinfo: stmt_info, new_stmt: init_stmt, gsi); |
1189 | } |
1190 | } |
1191 | } |
1192 | val = build_vector_from_val (type, val); |
1193 | } |
1194 | |
1195 | new_temp = vect_get_new_ssa_name (type, vect_simple_var, "cst_" ); |
1196 | init_stmt = gimple_build_assign (new_temp, val); |
1197 | vect_init_vector_1 (vinfo, stmt_vinfo: stmt_info, new_stmt: init_stmt, gsi); |
1198 | return new_temp; |
1199 | } |
1200 | |
1201 | |
1202 | /* Function vect_get_vec_defs_for_operand. |
1203 | |
1204 | OP is an operand in STMT_VINFO. This function returns a vector of |
1205 | NCOPIES defs that will be used in the vectorized stmts for STMT_VINFO. |
1206 | |
1207 | In the case that OP is an SSA_NAME which is defined in the loop, then |
1208 | STMT_VINFO_VEC_STMTS of the defining stmt holds the relevant defs. |
1209 | |
1210 | In case OP is an invariant or constant, a new stmt that creates a vector def |
1211 | needs to be introduced. VECTYPE may be used to specify a required type for |
1212 | vector invariant. */ |
1213 | |
1214 | void |
1215 | vect_get_vec_defs_for_operand (vec_info *vinfo, stmt_vec_info stmt_vinfo, |
1216 | unsigned ncopies, |
1217 | tree op, vec<tree> *vec_oprnds, tree vectype) |
1218 | { |
1219 | gimple *def_stmt; |
1220 | enum vect_def_type dt; |
1221 | bool is_simple_use; |
1222 | loop_vec_info loop_vinfo = dyn_cast <loop_vec_info> (p: vinfo); |
1223 | |
1224 | if (dump_enabled_p ()) |
1225 | dump_printf_loc (MSG_NOTE, vect_location, |
1226 | "vect_get_vec_defs_for_operand: %T\n" , op); |
1227 | |
1228 | stmt_vec_info def_stmt_info; |
1229 | is_simple_use = vect_is_simple_use (op, loop_vinfo, &dt, |
1230 | &def_stmt_info, &def_stmt); |
1231 | gcc_assert (is_simple_use); |
1232 | if (def_stmt && dump_enabled_p ()) |
1233 | dump_printf_loc (MSG_NOTE, vect_location, " def_stmt = %G" , def_stmt); |
1234 | |
1235 | vec_oprnds->create (nelems: ncopies); |
1236 | if (dt == vect_constant_def || dt == vect_external_def) |
1237 | { |
1238 | tree stmt_vectype = STMT_VINFO_VECTYPE (stmt_vinfo); |
1239 | tree vector_type; |
1240 | |
1241 | if (vectype) |
1242 | vector_type = vectype; |
1243 | else if (VECT_SCALAR_BOOLEAN_TYPE_P (TREE_TYPE (op)) |
1244 | && VECTOR_BOOLEAN_TYPE_P (stmt_vectype)) |
1245 | vector_type = truth_type_for (stmt_vectype); |
1246 | else |
1247 | vector_type = get_vectype_for_scalar_type (loop_vinfo, TREE_TYPE (op)); |
1248 | |
1249 | gcc_assert (vector_type); |
1250 | tree vop = vect_init_vector (vinfo, stmt_info: stmt_vinfo, val: op, type: vector_type, NULL); |
1251 | while (ncopies--) |
1252 | vec_oprnds->quick_push (obj: vop); |
1253 | } |
1254 | else |
1255 | { |
1256 | def_stmt_info = vect_stmt_to_vectorize (stmt_info: def_stmt_info); |
1257 | gcc_assert (STMT_VINFO_VEC_STMTS (def_stmt_info).length () == ncopies); |
1258 | for (unsigned i = 0; i < ncopies; ++i) |
1259 | vec_oprnds->quick_push (obj: gimple_get_lhs |
1260 | (STMT_VINFO_VEC_STMTS (def_stmt_info)[i])); |
1261 | } |
1262 | } |
1263 | |
1264 | |
1265 | /* Get vectorized definitions for OP0 and OP1. */ |
1266 | |
1267 | void |
1268 | vect_get_vec_defs (vec_info *vinfo, stmt_vec_info stmt_info, slp_tree slp_node, |
1269 | unsigned ncopies, |
1270 | tree op0, vec<tree> *vec_oprnds0, tree vectype0, |
1271 | tree op1, vec<tree> *vec_oprnds1, tree vectype1, |
1272 | tree op2, vec<tree> *vec_oprnds2, tree vectype2, |
1273 | tree op3, vec<tree> *vec_oprnds3, tree vectype3) |
1274 | { |
1275 | if (slp_node) |
1276 | { |
1277 | if (op0) |
1278 | vect_get_slp_defs (SLP_TREE_CHILDREN (slp_node)[0], vec_oprnds0); |
1279 | if (op1) |
1280 | vect_get_slp_defs (SLP_TREE_CHILDREN (slp_node)[1], vec_oprnds1); |
1281 | if (op2) |
1282 | vect_get_slp_defs (SLP_TREE_CHILDREN (slp_node)[2], vec_oprnds2); |
1283 | if (op3) |
1284 | vect_get_slp_defs (SLP_TREE_CHILDREN (slp_node)[3], vec_oprnds3); |
1285 | } |
1286 | else |
1287 | { |
1288 | if (op0) |
1289 | vect_get_vec_defs_for_operand (vinfo, stmt_vinfo: stmt_info, ncopies, |
1290 | op: op0, vec_oprnds: vec_oprnds0, vectype: vectype0); |
1291 | if (op1) |
1292 | vect_get_vec_defs_for_operand (vinfo, stmt_vinfo: stmt_info, ncopies, |
1293 | op: op1, vec_oprnds: vec_oprnds1, vectype: vectype1); |
1294 | if (op2) |
1295 | vect_get_vec_defs_for_operand (vinfo, stmt_vinfo: stmt_info, ncopies, |
1296 | op: op2, vec_oprnds: vec_oprnds2, vectype: vectype2); |
1297 | if (op3) |
1298 | vect_get_vec_defs_for_operand (vinfo, stmt_vinfo: stmt_info, ncopies, |
1299 | op: op3, vec_oprnds: vec_oprnds3, vectype: vectype3); |
1300 | } |
1301 | } |
1302 | |
1303 | void |
1304 | vect_get_vec_defs (vec_info *vinfo, stmt_vec_info stmt_info, slp_tree slp_node, |
1305 | unsigned ncopies, |
1306 | tree op0, vec<tree> *vec_oprnds0, |
1307 | tree op1, vec<tree> *vec_oprnds1, |
1308 | tree op2, vec<tree> *vec_oprnds2, |
1309 | tree op3, vec<tree> *vec_oprnds3) |
1310 | { |
1311 | vect_get_vec_defs (vinfo, stmt_info, slp_node, ncopies, |
1312 | op0, vec_oprnds0, NULL_TREE, |
1313 | op1, vec_oprnds1, NULL_TREE, |
1314 | op2, vec_oprnds2, NULL_TREE, |
1315 | op3, vec_oprnds3, NULL_TREE); |
1316 | } |
1317 | |
1318 | /* Helper function called by vect_finish_replace_stmt and |
1319 | vect_finish_stmt_generation. Set the location of the new |
1320 | statement and create and return a stmt_vec_info for it. */ |
1321 | |
1322 | static void |
1323 | vect_finish_stmt_generation_1 (vec_info *, |
1324 | stmt_vec_info stmt_info, gimple *vec_stmt) |
1325 | { |
1326 | if (dump_enabled_p ()) |
1327 | dump_printf_loc (MSG_NOTE, vect_location, "add new stmt: %G" , vec_stmt); |
1328 | |
1329 | if (stmt_info) |
1330 | { |
1331 | gimple_set_location (g: vec_stmt, location: gimple_location (g: stmt_info->stmt)); |
1332 | |
1333 | /* While EH edges will generally prevent vectorization, stmt might |
1334 | e.g. be in a must-not-throw region. Ensure newly created stmts |
1335 | that could throw are part of the same region. */ |
1336 | int lp_nr = lookup_stmt_eh_lp (stmt_info->stmt); |
1337 | if (lp_nr != 0 && stmt_could_throw_p (cfun, vec_stmt)) |
1338 | add_stmt_to_eh_lp (vec_stmt, lp_nr); |
1339 | } |
1340 | else |
1341 | gcc_assert (!stmt_could_throw_p (cfun, vec_stmt)); |
1342 | } |
1343 | |
1344 | /* Replace the scalar statement STMT_INFO with a new vector statement VEC_STMT, |
1345 | which sets the same scalar result as STMT_INFO did. Create and return a |
1346 | stmt_vec_info for VEC_STMT. */ |
1347 | |
1348 | void |
1349 | vect_finish_replace_stmt (vec_info *vinfo, |
1350 | stmt_vec_info stmt_info, gimple *vec_stmt) |
1351 | { |
1352 | gimple *scalar_stmt = vect_orig_stmt (stmt_info)->stmt; |
1353 | gcc_assert (gimple_get_lhs (scalar_stmt) == gimple_get_lhs (vec_stmt)); |
1354 | |
1355 | gimple_stmt_iterator gsi = gsi_for_stmt (scalar_stmt); |
1356 | gsi_replace (&gsi, vec_stmt, true); |
1357 | |
1358 | vect_finish_stmt_generation_1 (vinfo, stmt_info, vec_stmt); |
1359 | } |
1360 | |
1361 | /* Add VEC_STMT to the vectorized implementation of STMT_INFO and insert it |
1362 | before *GSI. Create and return a stmt_vec_info for VEC_STMT. */ |
1363 | |
1364 | void |
1365 | vect_finish_stmt_generation (vec_info *vinfo, |
1366 | stmt_vec_info stmt_info, gimple *vec_stmt, |
1367 | gimple_stmt_iterator *gsi) |
1368 | { |
1369 | gcc_assert (!stmt_info || gimple_code (stmt_info->stmt) != GIMPLE_LABEL); |
1370 | |
1371 | if (!gsi_end_p (i: *gsi) |
1372 | && gimple_has_mem_ops (g: vec_stmt)) |
1373 | { |
1374 | gimple *at_stmt = gsi_stmt (i: *gsi); |
1375 | tree vuse = gimple_vuse (g: at_stmt); |
1376 | if (vuse && TREE_CODE (vuse) == SSA_NAME) |
1377 | { |
1378 | tree vdef = gimple_vdef (g: at_stmt); |
1379 | gimple_set_vuse (g: vec_stmt, vuse: gimple_vuse (g: at_stmt)); |
1380 | gimple_set_modified (s: vec_stmt, modifiedp: true); |
1381 | /* If we have an SSA vuse and insert a store, update virtual |
1382 | SSA form to avoid triggering the renamer. Do so only |
1383 | if we can easily see all uses - which is what almost always |
1384 | happens with the way vectorized stmts are inserted. */ |
1385 | if ((vdef && TREE_CODE (vdef) == SSA_NAME) |
1386 | && ((is_gimple_assign (gs: vec_stmt) |
1387 | && !is_gimple_reg (gimple_assign_lhs (gs: vec_stmt))) |
1388 | || (is_gimple_call (gs: vec_stmt) |
1389 | && (!(gimple_call_flags (vec_stmt) |
1390 | & (ECF_CONST|ECF_PURE|ECF_NOVOPS)) |
1391 | || (gimple_call_lhs (gs: vec_stmt) |
1392 | && !is_gimple_reg (gimple_call_lhs (gs: vec_stmt))))))) |
1393 | { |
1394 | tree new_vdef = copy_ssa_name (var: vuse, stmt: vec_stmt); |
1395 | gimple_set_vdef (g: vec_stmt, vdef: new_vdef); |
1396 | SET_USE (gimple_vuse_op (at_stmt), new_vdef); |
1397 | } |
1398 | } |
1399 | } |
1400 | gsi_insert_before (gsi, vec_stmt, GSI_SAME_STMT); |
1401 | vect_finish_stmt_generation_1 (vinfo, stmt_info, vec_stmt); |
1402 | } |
1403 | |
1404 | /* We want to vectorize a call to combined function CFN with function |
1405 | decl FNDECL, using VECTYPE_OUT as the type of the output and VECTYPE_IN |
1406 | as the types of all inputs. Check whether this is possible using |
1407 | an internal function, returning its code if so or IFN_LAST if not. */ |
1408 | |
1409 | static internal_fn |
1410 | vectorizable_internal_function (combined_fn cfn, tree fndecl, |
1411 | tree vectype_out, tree vectype_in) |
1412 | { |
1413 | internal_fn ifn; |
1414 | if (internal_fn_p (code: cfn)) |
1415 | ifn = as_internal_fn (code: cfn); |
1416 | else |
1417 | ifn = associated_internal_fn (fndecl); |
1418 | if (ifn != IFN_LAST && direct_internal_fn_p (fn: ifn)) |
1419 | { |
1420 | const direct_internal_fn_info &info = direct_internal_fn (fn: ifn); |
1421 | if (info.vectorizable) |
1422 | { |
1423 | bool same_size_p = TYPE_SIZE (vectype_in) == TYPE_SIZE (vectype_out); |
1424 | tree type0 = (info.type0 < 0 ? vectype_out : vectype_in); |
1425 | tree type1 = (info.type1 < 0 ? vectype_out : vectype_in); |
1426 | |
1427 | /* The type size of both the vectype_in and vectype_out should be |
1428 | exactly the same when vectype_out isn't participating the optab. |
1429 | While there is no restriction for type size when vectype_out |
1430 | is part of the optab query. */ |
1431 | if (type0 != vectype_out && type1 != vectype_out && !same_size_p) |
1432 | return IFN_LAST; |
1433 | |
1434 | if (direct_internal_fn_supported_p (ifn, tree_pair (type0, type1), |
1435 | OPTIMIZE_FOR_SPEED)) |
1436 | return ifn; |
1437 | } |
1438 | } |
1439 | return IFN_LAST; |
1440 | } |
1441 | |
1442 | |
1443 | static tree permute_vec_elements (vec_info *, tree, tree, tree, stmt_vec_info, |
1444 | gimple_stmt_iterator *); |
1445 | |
1446 | /* Check whether a load or store statement in the loop described by |
1447 | LOOP_VINFO is possible in a loop using partial vectors. This is |
1448 | testing whether the vectorizer pass has the appropriate support, |
1449 | as well as whether the target does. |
1450 | |
1451 | VLS_TYPE says whether the statement is a load or store and VECTYPE |
1452 | is the type of the vector being loaded or stored. SLP_NODE is the SLP |
1453 | node that contains the statement, or null if none. MEMORY_ACCESS_TYPE |
1454 | says how the load or store is going to be implemented and GROUP_SIZE |
1455 | is the number of load or store statements in the containing group. |
1456 | If the access is a gather load or scatter store, GS_INFO describes |
1457 | its arguments. If the load or store is conditional, SCALAR_MASK is the |
1458 | condition under which it occurs. |
1459 | |
1460 | Clear LOOP_VINFO_CAN_USE_PARTIAL_VECTORS_P if a loop using partial |
1461 | vectors is not supported, otherwise record the required rgroup control |
1462 | types. */ |
1463 | |
1464 | static void |
1465 | check_load_store_for_partial_vectors (loop_vec_info loop_vinfo, tree vectype, |
1466 | slp_tree slp_node, |
1467 | vec_load_store_type vls_type, |
1468 | int group_size, |
1469 | vect_memory_access_type |
1470 | memory_access_type, |
1471 | gather_scatter_info *gs_info, |
1472 | tree scalar_mask) |
1473 | { |
1474 | /* Invariant loads need no special support. */ |
1475 | if (memory_access_type == VMAT_INVARIANT) |
1476 | return; |
1477 | |
1478 | unsigned int nvectors; |
1479 | if (slp_node) |
1480 | nvectors = SLP_TREE_NUMBER_OF_VEC_STMTS (slp_node); |
1481 | else |
1482 | nvectors = vect_get_num_copies (loop_vinfo, vectype); |
1483 | |
1484 | vec_loop_masks *masks = &LOOP_VINFO_MASKS (loop_vinfo); |
1485 | vec_loop_lens *lens = &LOOP_VINFO_LENS (loop_vinfo); |
1486 | machine_mode vecmode = TYPE_MODE (vectype); |
1487 | bool is_load = (vls_type == VLS_LOAD); |
1488 | if (memory_access_type == VMAT_LOAD_STORE_LANES) |
1489 | { |
1490 | internal_fn ifn |
1491 | = (is_load ? vect_load_lanes_supported (vectype, group_size, true) |
1492 | : vect_store_lanes_supported (vectype, group_size, true)); |
1493 | if (ifn == IFN_MASK_LEN_LOAD_LANES || ifn == IFN_MASK_LEN_STORE_LANES) |
1494 | vect_record_loop_len (loop_vinfo, lens, nvectors, vectype, 1); |
1495 | else if (ifn == IFN_MASK_LOAD_LANES || ifn == IFN_MASK_STORE_LANES) |
1496 | vect_record_loop_mask (loop_vinfo, masks, nvectors, vectype, |
1497 | scalar_mask); |
1498 | else |
1499 | { |
1500 | if (dump_enabled_p ()) |
1501 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
1502 | "can't operate on partial vectors because" |
1503 | " the target doesn't have an appropriate" |
1504 | " load/store-lanes instruction.\n" ); |
1505 | LOOP_VINFO_CAN_USE_PARTIAL_VECTORS_P (loop_vinfo) = false; |
1506 | } |
1507 | return; |
1508 | } |
1509 | |
1510 | if (memory_access_type == VMAT_GATHER_SCATTER) |
1511 | { |
1512 | internal_fn ifn = (is_load |
1513 | ? IFN_MASK_GATHER_LOAD |
1514 | : IFN_MASK_SCATTER_STORE); |
1515 | internal_fn len_ifn = (is_load |
1516 | ? IFN_MASK_LEN_GATHER_LOAD |
1517 | : IFN_MASK_LEN_SCATTER_STORE); |
1518 | if (internal_gather_scatter_fn_supported_p (len_ifn, vectype, |
1519 | gs_info->memory_type, |
1520 | gs_info->offset_vectype, |
1521 | gs_info->scale)) |
1522 | vect_record_loop_len (loop_vinfo, lens, nvectors, vectype, 1); |
1523 | else if (internal_gather_scatter_fn_supported_p (ifn, vectype, |
1524 | gs_info->memory_type, |
1525 | gs_info->offset_vectype, |
1526 | gs_info->scale)) |
1527 | vect_record_loop_mask (loop_vinfo, masks, nvectors, vectype, |
1528 | scalar_mask); |
1529 | else |
1530 | { |
1531 | if (dump_enabled_p ()) |
1532 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
1533 | "can't operate on partial vectors because" |
1534 | " the target doesn't have an appropriate" |
1535 | " gather load or scatter store instruction.\n" ); |
1536 | LOOP_VINFO_CAN_USE_PARTIAL_VECTORS_P (loop_vinfo) = false; |
1537 | } |
1538 | return; |
1539 | } |
1540 | |
1541 | if (memory_access_type != VMAT_CONTIGUOUS |
1542 | && memory_access_type != VMAT_CONTIGUOUS_PERMUTE) |
1543 | { |
1544 | /* Element X of the data must come from iteration i * VF + X of the |
1545 | scalar loop. We need more work to support other mappings. */ |
1546 | if (dump_enabled_p ()) |
1547 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
1548 | "can't operate on partial vectors because an" |
1549 | " access isn't contiguous.\n" ); |
1550 | LOOP_VINFO_CAN_USE_PARTIAL_VECTORS_P (loop_vinfo) = false; |
1551 | return; |
1552 | } |
1553 | |
1554 | if (!VECTOR_MODE_P (vecmode)) |
1555 | { |
1556 | if (dump_enabled_p ()) |
1557 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
1558 | "can't operate on partial vectors when emulating" |
1559 | " vector operations.\n" ); |
1560 | LOOP_VINFO_CAN_USE_PARTIAL_VECTORS_P (loop_vinfo) = false; |
1561 | return; |
1562 | } |
1563 | |
1564 | /* We might load more scalars than we need for permuting SLP loads. |
1565 | We checked in get_group_load_store_type that the extra elements |
1566 | don't leak into a new vector. */ |
1567 | auto group_memory_nvectors = [](poly_uint64 size, poly_uint64 nunits) |
1568 | { |
1569 | unsigned int nvectors; |
1570 | if (can_div_away_from_zero_p (a: size, b: nunits, quotient: &nvectors)) |
1571 | return nvectors; |
1572 | gcc_unreachable (); |
1573 | }; |
1574 | |
1575 | poly_uint64 nunits = TYPE_VECTOR_SUBPARTS (node: vectype); |
1576 | poly_uint64 vf = LOOP_VINFO_VECT_FACTOR (loop_vinfo); |
1577 | machine_mode mask_mode; |
1578 | machine_mode vmode; |
1579 | bool using_partial_vectors_p = false; |
1580 | if (get_len_load_store_mode (vecmode, is_load).exists (mode: &vmode)) |
1581 | { |
1582 | nvectors = group_memory_nvectors (group_size * vf, nunits); |
1583 | unsigned factor = (vecmode == vmode) ? 1 : GET_MODE_UNIT_SIZE (vecmode); |
1584 | vect_record_loop_len (loop_vinfo, lens, nvectors, vectype, factor); |
1585 | using_partial_vectors_p = true; |
1586 | } |
1587 | else if (targetm.vectorize.get_mask_mode (vecmode).exists (mode: &mask_mode) |
1588 | && can_vec_mask_load_store_p (vecmode, mask_mode, is_load)) |
1589 | { |
1590 | nvectors = group_memory_nvectors (group_size * vf, nunits); |
1591 | vect_record_loop_mask (loop_vinfo, masks, nvectors, vectype, scalar_mask); |
1592 | using_partial_vectors_p = true; |
1593 | } |
1594 | |
1595 | if (!using_partial_vectors_p) |
1596 | { |
1597 | if (dump_enabled_p ()) |
1598 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
1599 | "can't operate on partial vectors because the" |
1600 | " target doesn't have the appropriate partial" |
1601 | " vectorization load or store.\n" ); |
1602 | LOOP_VINFO_CAN_USE_PARTIAL_VECTORS_P (loop_vinfo) = false; |
1603 | } |
1604 | } |
1605 | |
1606 | /* Return the mask input to a masked load or store. VEC_MASK is the vectorized |
1607 | form of the scalar mask condition and LOOP_MASK, if nonnull, is the mask |
1608 | that needs to be applied to all loads and stores in a vectorized loop. |
1609 | Return VEC_MASK if LOOP_MASK is null or if VEC_MASK is already masked, |
1610 | otherwise return VEC_MASK & LOOP_MASK. |
1611 | |
1612 | MASK_TYPE is the type of both masks. If new statements are needed, |
1613 | insert them before GSI. */ |
1614 | |
1615 | static tree |
1616 | prepare_vec_mask (loop_vec_info loop_vinfo, tree mask_type, tree loop_mask, |
1617 | tree vec_mask, gimple_stmt_iterator *gsi) |
1618 | { |
1619 | gcc_assert (useless_type_conversion_p (mask_type, TREE_TYPE (vec_mask))); |
1620 | if (!loop_mask) |
1621 | return vec_mask; |
1622 | |
1623 | gcc_assert (TREE_TYPE (loop_mask) == mask_type); |
1624 | |
1625 | if (loop_vinfo->vec_cond_masked_set.contains (k: { vec_mask, loop_mask })) |
1626 | return vec_mask; |
1627 | |
1628 | tree and_res = make_temp_ssa_name (type: mask_type, NULL, name: "vec_mask_and" ); |
1629 | gimple *and_stmt = gimple_build_assign (and_res, BIT_AND_EXPR, |
1630 | vec_mask, loop_mask); |
1631 | |
1632 | gsi_insert_before (gsi, and_stmt, GSI_SAME_STMT); |
1633 | return and_res; |
1634 | } |
1635 | |
1636 | /* Determine whether we can use a gather load or scatter store to vectorize |
1637 | strided load or store STMT_INFO by truncating the current offset to a |
1638 | smaller width. We need to be able to construct an offset vector: |
1639 | |
1640 | { 0, X, X*2, X*3, ... } |
1641 | |
1642 | without loss of precision, where X is STMT_INFO's DR_STEP. |
1643 | |
1644 | Return true if this is possible, describing the gather load or scatter |
1645 | store in GS_INFO. MASKED_P is true if the load or store is conditional. */ |
1646 | |
1647 | static bool |
1648 | vect_truncate_gather_scatter_offset (stmt_vec_info stmt_info, |
1649 | loop_vec_info loop_vinfo, bool masked_p, |
1650 | gather_scatter_info *gs_info) |
1651 | { |
1652 | dr_vec_info *dr_info = STMT_VINFO_DR_INFO (stmt_info); |
1653 | data_reference *dr = dr_info->dr; |
1654 | tree step = DR_STEP (dr); |
1655 | if (TREE_CODE (step) != INTEGER_CST) |
1656 | { |
1657 | /* ??? Perhaps we could use range information here? */ |
1658 | if (dump_enabled_p ()) |
1659 | dump_printf_loc (MSG_NOTE, vect_location, |
1660 | "cannot truncate variable step.\n" ); |
1661 | return false; |
1662 | } |
1663 | |
1664 | /* Get the number of bits in an element. */ |
1665 | tree vectype = STMT_VINFO_VECTYPE (stmt_info); |
1666 | scalar_mode element_mode = SCALAR_TYPE_MODE (TREE_TYPE (vectype)); |
1667 | unsigned int element_bits = GET_MODE_BITSIZE (mode: element_mode); |
1668 | |
1669 | /* Set COUNT to the upper limit on the number of elements - 1. |
1670 | Start with the maximum vectorization factor. */ |
1671 | unsigned HOST_WIDE_INT count = vect_max_vf (loop_vinfo) - 1; |
1672 | |
1673 | /* Try lowering COUNT to the number of scalar latch iterations. */ |
1674 | class loop *loop = LOOP_VINFO_LOOP (loop_vinfo); |
1675 | widest_int max_iters; |
1676 | if (max_loop_iterations (loop, &max_iters) |
1677 | && max_iters < count) |
1678 | count = max_iters.to_shwi (); |
1679 | |
1680 | /* Try scales of 1 and the element size. */ |
1681 | int scales[] = { 1, vect_get_scalar_dr_size (dr_info) }; |
1682 | wi::overflow_type overflow = wi::OVF_NONE; |
1683 | for (int i = 0; i < 2; ++i) |
1684 | { |
1685 | int scale = scales[i]; |
1686 | widest_int factor; |
1687 | if (!wi::multiple_of_p (x: wi::to_widest (t: step), y: scale, sgn: SIGNED, res: &factor)) |
1688 | continue; |
1689 | |
1690 | /* Determine the minimum precision of (COUNT - 1) * STEP / SCALE. */ |
1691 | widest_int range = wi::mul (x: count, y: factor, sgn: SIGNED, overflow: &overflow); |
1692 | if (overflow) |
1693 | continue; |
1694 | signop sign = range >= 0 ? UNSIGNED : SIGNED; |
1695 | unsigned int min_offset_bits = wi::min_precision (x: range, sgn: sign); |
1696 | |
1697 | /* Find the narrowest viable offset type. */ |
1698 | unsigned int offset_bits = 1U << ceil_log2 (x: min_offset_bits); |
1699 | tree offset_type = build_nonstandard_integer_type (offset_bits, |
1700 | sign == UNSIGNED); |
1701 | |
1702 | /* See whether the target supports the operation with an offset |
1703 | no narrower than OFFSET_TYPE. */ |
1704 | tree memory_type = TREE_TYPE (DR_REF (dr)); |
1705 | if (!vect_gather_scatter_fn_p (loop_vinfo, DR_IS_READ (dr), masked_p, |
1706 | vectype, memory_type, offset_type, scale, |
1707 | &gs_info->ifn, &gs_info->offset_vectype) |
1708 | || gs_info->ifn == IFN_LAST) |
1709 | continue; |
1710 | |
1711 | gs_info->decl = NULL_TREE; |
1712 | /* Logically the sum of DR_BASE_ADDRESS, DR_INIT and DR_OFFSET, |
1713 | but we don't need to store that here. */ |
1714 | gs_info->base = NULL_TREE; |
1715 | gs_info->element_type = TREE_TYPE (vectype); |
1716 | gs_info->offset = fold_convert (offset_type, step); |
1717 | gs_info->offset_dt = vect_constant_def; |
1718 | gs_info->scale = scale; |
1719 | gs_info->memory_type = memory_type; |
1720 | return true; |
1721 | } |
1722 | |
1723 | if (overflow && dump_enabled_p ()) |
1724 | dump_printf_loc (MSG_NOTE, vect_location, |
1725 | "truncating gather/scatter offset to %d bits" |
1726 | " might change its value.\n" , element_bits); |
1727 | |
1728 | return false; |
1729 | } |
1730 | |
1731 | /* Return true if we can use gather/scatter internal functions to |
1732 | vectorize STMT_INFO, which is a grouped or strided load or store. |
1733 | MASKED_P is true if load or store is conditional. When returning |
1734 | true, fill in GS_INFO with the information required to perform the |
1735 | operation. */ |
1736 | |
1737 | static bool |
1738 | vect_use_strided_gather_scatters_p (stmt_vec_info stmt_info, |
1739 | loop_vec_info loop_vinfo, bool masked_p, |
1740 | gather_scatter_info *gs_info) |
1741 | { |
1742 | if (!vect_check_gather_scatter (stmt_info, loop_vinfo, gs_info) |
1743 | || gs_info->ifn == IFN_LAST) |
1744 | return vect_truncate_gather_scatter_offset (stmt_info, loop_vinfo, |
1745 | masked_p, gs_info); |
1746 | |
1747 | tree old_offset_type = TREE_TYPE (gs_info->offset); |
1748 | tree new_offset_type = TREE_TYPE (gs_info->offset_vectype); |
1749 | |
1750 | gcc_assert (TYPE_PRECISION (new_offset_type) |
1751 | >= TYPE_PRECISION (old_offset_type)); |
1752 | gs_info->offset = fold_convert (new_offset_type, gs_info->offset); |
1753 | |
1754 | if (dump_enabled_p ()) |
1755 | dump_printf_loc (MSG_NOTE, vect_location, |
1756 | "using gather/scatter for strided/grouped access," |
1757 | " scale = %d\n" , gs_info->scale); |
1758 | |
1759 | return true; |
1760 | } |
1761 | |
1762 | /* STMT_INFO is a non-strided load or store, meaning that it accesses |
1763 | elements with a known constant step. Return -1 if that step |
1764 | is negative, 0 if it is zero, and 1 if it is greater than zero. */ |
1765 | |
1766 | static int |
1767 | compare_step_with_zero (vec_info *vinfo, stmt_vec_info stmt_info) |
1768 | { |
1769 | dr_vec_info *dr_info = STMT_VINFO_DR_INFO (stmt_info); |
1770 | return tree_int_cst_compare (t1: vect_dr_behavior (vinfo, dr_info)->step, |
1771 | size_zero_node); |
1772 | } |
1773 | |
1774 | /* If the target supports a permute mask that reverses the elements in |
1775 | a vector of type VECTYPE, return that mask, otherwise return null. */ |
1776 | |
1777 | static tree |
1778 | perm_mask_for_reverse (tree vectype) |
1779 | { |
1780 | poly_uint64 nunits = TYPE_VECTOR_SUBPARTS (node: vectype); |
1781 | |
1782 | /* The encoding has a single stepped pattern. */ |
1783 | vec_perm_builder sel (nunits, 1, 3); |
1784 | for (int i = 0; i < 3; ++i) |
1785 | sel.quick_push (obj: nunits - 1 - i); |
1786 | |
1787 | vec_perm_indices indices (sel, 1, nunits); |
1788 | if (!can_vec_perm_const_p (TYPE_MODE (vectype), TYPE_MODE (vectype), |
1789 | indices)) |
1790 | return NULL_TREE; |
1791 | return vect_gen_perm_mask_checked (vectype, indices); |
1792 | } |
1793 | |
1794 | /* A subroutine of get_load_store_type, with a subset of the same |
1795 | arguments. Handle the case where STMT_INFO is a load or store that |
1796 | accesses consecutive elements with a negative step. Sets *POFFSET |
1797 | to the offset to be applied to the DR for the first access. */ |
1798 | |
1799 | static vect_memory_access_type |
1800 | get_negative_load_store_type (vec_info *vinfo, |
1801 | stmt_vec_info stmt_info, tree vectype, |
1802 | vec_load_store_type vls_type, |
1803 | unsigned int ncopies, poly_int64 *poffset) |
1804 | { |
1805 | dr_vec_info *dr_info = STMT_VINFO_DR_INFO (stmt_info); |
1806 | dr_alignment_support alignment_support_scheme; |
1807 | |
1808 | if (ncopies > 1) |
1809 | { |
1810 | if (dump_enabled_p ()) |
1811 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
1812 | "multiple types with negative step.\n" ); |
1813 | return VMAT_ELEMENTWISE; |
1814 | } |
1815 | |
1816 | /* For backward running DRs the first access in vectype actually is |
1817 | N-1 elements before the address of the DR. */ |
1818 | *poffset = ((-TYPE_VECTOR_SUBPARTS (node: vectype) + 1) |
1819 | * TREE_INT_CST_LOW (TYPE_SIZE_UNIT (TREE_TYPE (vectype)))); |
1820 | |
1821 | int misalignment = dr_misalignment (dr_info, vectype, offset: *poffset); |
1822 | alignment_support_scheme |
1823 | = vect_supportable_dr_alignment (vinfo, dr_info, vectype, misalignment); |
1824 | if (alignment_support_scheme != dr_aligned |
1825 | && alignment_support_scheme != dr_unaligned_supported) |
1826 | { |
1827 | if (dump_enabled_p ()) |
1828 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
1829 | "negative step but alignment required.\n" ); |
1830 | *poffset = 0; |
1831 | return VMAT_ELEMENTWISE; |
1832 | } |
1833 | |
1834 | if (vls_type == VLS_STORE_INVARIANT) |
1835 | { |
1836 | if (dump_enabled_p ()) |
1837 | dump_printf_loc (MSG_NOTE, vect_location, |
1838 | "negative step with invariant source;" |
1839 | " no permute needed.\n" ); |
1840 | return VMAT_CONTIGUOUS_DOWN; |
1841 | } |
1842 | |
1843 | if (!perm_mask_for_reverse (vectype)) |
1844 | { |
1845 | if (dump_enabled_p ()) |
1846 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
1847 | "negative step and reversing not supported.\n" ); |
1848 | *poffset = 0; |
1849 | return VMAT_ELEMENTWISE; |
1850 | } |
1851 | |
1852 | return VMAT_CONTIGUOUS_REVERSE; |
1853 | } |
1854 | |
1855 | /* STMT_INFO is either a masked or unconditional store. Return the value |
1856 | being stored. */ |
1857 | |
1858 | tree |
1859 | vect_get_store_rhs (stmt_vec_info stmt_info) |
1860 | { |
1861 | if (gassign *assign = dyn_cast <gassign *> (p: stmt_info->stmt)) |
1862 | { |
1863 | gcc_assert (gimple_assign_single_p (assign)); |
1864 | return gimple_assign_rhs1 (gs: assign); |
1865 | } |
1866 | if (gcall *call = dyn_cast <gcall *> (p: stmt_info->stmt)) |
1867 | { |
1868 | internal_fn ifn = gimple_call_internal_fn (gs: call); |
1869 | int index = internal_fn_stored_value_index (ifn); |
1870 | gcc_assert (index >= 0); |
1871 | return gimple_call_arg (gs: call, index); |
1872 | } |
1873 | gcc_unreachable (); |
1874 | } |
1875 | |
1876 | /* Function VECTOR_VECTOR_COMPOSITION_TYPE |
1877 | |
1878 | This function returns a vector type which can be composed with NETLS pieces, |
1879 | whose type is recorded in PTYPE. VTYPE should be a vector type, and has the |
1880 | same vector size as the return vector. It checks target whether supports |
1881 | pieces-size vector mode for construction firstly, if target fails to, check |
1882 | pieces-size scalar mode for construction further. It returns NULL_TREE if |
1883 | fails to find the available composition. |
1884 | |
1885 | For example, for (vtype=V16QI, nelts=4), we can probably get: |
1886 | - V16QI with PTYPE V4QI. |
1887 | - V4SI with PTYPE SI. |
1888 | - NULL_TREE. */ |
1889 | |
1890 | static tree |
1891 | vector_vector_composition_type (tree vtype, poly_uint64 nelts, tree *ptype) |
1892 | { |
1893 | gcc_assert (VECTOR_TYPE_P (vtype)); |
1894 | gcc_assert (known_gt (nelts, 0U)); |
1895 | |
1896 | machine_mode vmode = TYPE_MODE (vtype); |
1897 | if (!VECTOR_MODE_P (vmode)) |
1898 | return NULL_TREE; |
1899 | |
1900 | /* When we are asked to compose the vector from its components let |
1901 | that happen directly. */ |
1902 | if (known_eq (TYPE_VECTOR_SUBPARTS (vtype), nelts)) |
1903 | { |
1904 | *ptype = TREE_TYPE (vtype); |
1905 | return vtype; |
1906 | } |
1907 | |
1908 | poly_uint64 vbsize = GET_MODE_BITSIZE (mode: vmode); |
1909 | unsigned int pbsize; |
1910 | if (constant_multiple_p (a: vbsize, b: nelts, multiple: &pbsize)) |
1911 | { |
1912 | /* First check if vec_init optab supports construction from |
1913 | vector pieces directly. */ |
1914 | scalar_mode elmode = SCALAR_TYPE_MODE (TREE_TYPE (vtype)); |
1915 | poly_uint64 inelts = pbsize / GET_MODE_BITSIZE (mode: elmode); |
1916 | machine_mode rmode; |
1917 | if (related_vector_mode (vmode, elmode, inelts).exists (mode: &rmode) |
1918 | && (convert_optab_handler (op: vec_init_optab, to_mode: vmode, from_mode: rmode) |
1919 | != CODE_FOR_nothing)) |
1920 | { |
1921 | *ptype = build_vector_type (TREE_TYPE (vtype), inelts); |
1922 | return vtype; |
1923 | } |
1924 | |
1925 | /* Otherwise check if exists an integer type of the same piece size and |
1926 | if vec_init optab supports construction from it directly. */ |
1927 | if (int_mode_for_size (size: pbsize, limit: 0).exists (mode: &elmode) |
1928 | && related_vector_mode (vmode, elmode, nelts).exists (mode: &rmode) |
1929 | && (convert_optab_handler (op: vec_init_optab, to_mode: rmode, from_mode: elmode) |
1930 | != CODE_FOR_nothing)) |
1931 | { |
1932 | *ptype = build_nonstandard_integer_type (pbsize, 1); |
1933 | return build_vector_type (*ptype, nelts); |
1934 | } |
1935 | } |
1936 | |
1937 | return NULL_TREE; |
1938 | } |
1939 | |
1940 | /* A subroutine of get_load_store_type, with a subset of the same |
1941 | arguments. Handle the case where STMT_INFO is part of a grouped load |
1942 | or store. |
1943 | |
1944 | For stores, the statements in the group are all consecutive |
1945 | and there is no gap at the end. For loads, the statements in the |
1946 | group might not be consecutive; there can be gaps between statements |
1947 | as well as at the end. */ |
1948 | |
1949 | static bool |
1950 | get_group_load_store_type (vec_info *vinfo, stmt_vec_info stmt_info, |
1951 | tree vectype, slp_tree slp_node, |
1952 | bool masked_p, vec_load_store_type vls_type, |
1953 | vect_memory_access_type *memory_access_type, |
1954 | poly_int64 *poffset, |
1955 | dr_alignment_support *alignment_support_scheme, |
1956 | int *misalignment, |
1957 | gather_scatter_info *gs_info, |
1958 | internal_fn *lanes_ifn) |
1959 | { |
1960 | loop_vec_info loop_vinfo = dyn_cast <loop_vec_info> (p: vinfo); |
1961 | class loop *loop = loop_vinfo ? LOOP_VINFO_LOOP (loop_vinfo) : NULL; |
1962 | stmt_vec_info first_stmt_info; |
1963 | unsigned int group_size; |
1964 | unsigned HOST_WIDE_INT gap; |
1965 | if (STMT_VINFO_GROUPED_ACCESS (stmt_info)) |
1966 | { |
1967 | first_stmt_info = DR_GROUP_FIRST_ELEMENT (stmt_info); |
1968 | group_size = DR_GROUP_SIZE (first_stmt_info); |
1969 | gap = DR_GROUP_GAP (first_stmt_info); |
1970 | } |
1971 | else |
1972 | { |
1973 | first_stmt_info = stmt_info; |
1974 | group_size = 1; |
1975 | gap = 0; |
1976 | } |
1977 | dr_vec_info *first_dr_info = STMT_VINFO_DR_INFO (first_stmt_info); |
1978 | bool single_element_p = (stmt_info == first_stmt_info |
1979 | && !DR_GROUP_NEXT_ELEMENT (stmt_info)); |
1980 | poly_uint64 nunits = TYPE_VECTOR_SUBPARTS (node: vectype); |
1981 | |
1982 | /* True if the vectorized statements would access beyond the last |
1983 | statement in the group. */ |
1984 | bool overrun_p = false; |
1985 | |
1986 | /* True if we can cope with such overrun by peeling for gaps, so that |
1987 | there is at least one final scalar iteration after the vector loop. */ |
1988 | bool can_overrun_p = (!masked_p |
1989 | && vls_type == VLS_LOAD |
1990 | && loop_vinfo |
1991 | && !loop->inner); |
1992 | |
1993 | /* There can only be a gap at the end of the group if the stride is |
1994 | known at compile time. */ |
1995 | gcc_assert (!STMT_VINFO_STRIDED_P (first_stmt_info) || gap == 0); |
1996 | |
1997 | /* Stores can't yet have gaps. */ |
1998 | gcc_assert (slp_node || vls_type == VLS_LOAD || gap == 0); |
1999 | |
2000 | if (slp_node) |
2001 | { |
2002 | /* For SLP vectorization we directly vectorize a subchain |
2003 | without permutation. */ |
2004 | if (! SLP_TREE_LOAD_PERMUTATION (slp_node).exists ()) |
2005 | first_dr_info |
2006 | = STMT_VINFO_DR_INFO (SLP_TREE_SCALAR_STMTS (slp_node)[0]); |
2007 | if (STMT_VINFO_STRIDED_P (first_stmt_info)) |
2008 | { |
2009 | /* Try to use consecutive accesses of DR_GROUP_SIZE elements, |
2010 | separated by the stride, until we have a complete vector. |
2011 | Fall back to scalar accesses if that isn't possible. */ |
2012 | if (multiple_p (a: nunits, b: group_size)) |
2013 | *memory_access_type = VMAT_STRIDED_SLP; |
2014 | else |
2015 | *memory_access_type = VMAT_ELEMENTWISE; |
2016 | } |
2017 | else |
2018 | { |
2019 | overrun_p = loop_vinfo && gap != 0; |
2020 | if (overrun_p && vls_type != VLS_LOAD) |
2021 | { |
2022 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
2023 | "Grouped store with gaps requires" |
2024 | " non-consecutive accesses\n" ); |
2025 | return false; |
2026 | } |
2027 | /* An overrun is fine if the trailing elements are smaller |
2028 | than the alignment boundary B. Every vector access will |
2029 | be a multiple of B and so we are guaranteed to access a |
2030 | non-gap element in the same B-sized block. */ |
2031 | if (overrun_p |
2032 | && gap < (vect_known_alignment_in_bytes (dr_info: first_dr_info, |
2033 | vectype) |
2034 | / vect_get_scalar_dr_size (dr_info: first_dr_info))) |
2035 | overrun_p = false; |
2036 | |
2037 | /* If the gap splits the vector in half and the target |
2038 | can do half-vector operations avoid the epilogue peeling |
2039 | by simply loading half of the vector only. Usually |
2040 | the construction with an upper zero half will be elided. */ |
2041 | dr_alignment_support alss; |
2042 | int misalign = dr_misalignment (dr_info: first_dr_info, vectype); |
2043 | tree half_vtype; |
2044 | if (overrun_p |
2045 | && !masked_p |
2046 | && (((alss = vect_supportable_dr_alignment (vinfo, first_dr_info, |
2047 | vectype, misalign))) |
2048 | == dr_aligned |
2049 | || alss == dr_unaligned_supported) |
2050 | && known_eq (nunits, (group_size - gap) * 2) |
2051 | && known_eq (nunits, group_size) |
2052 | && (vector_vector_composition_type (vtype: vectype, nelts: 2, ptype: &half_vtype) |
2053 | != NULL_TREE)) |
2054 | overrun_p = false; |
2055 | |
2056 | if (overrun_p && !can_overrun_p) |
2057 | { |
2058 | if (dump_enabled_p ()) |
2059 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
2060 | "Peeling for outer loop is not supported\n" ); |
2061 | return false; |
2062 | } |
2063 | int cmp = compare_step_with_zero (vinfo, stmt_info); |
2064 | if (cmp < 0) |
2065 | { |
2066 | if (single_element_p) |
2067 | /* ??? The VMAT_CONTIGUOUS_REVERSE code generation is |
2068 | only correct for single element "interleaving" SLP. */ |
2069 | *memory_access_type = get_negative_load_store_type |
2070 | (vinfo, stmt_info, vectype, vls_type, ncopies: 1, poffset); |
2071 | else |
2072 | { |
2073 | /* Try to use consecutive accesses of DR_GROUP_SIZE elements, |
2074 | separated by the stride, until we have a complete vector. |
2075 | Fall back to scalar accesses if that isn't possible. */ |
2076 | if (multiple_p (a: nunits, b: group_size)) |
2077 | *memory_access_type = VMAT_STRIDED_SLP; |
2078 | else |
2079 | *memory_access_type = VMAT_ELEMENTWISE; |
2080 | } |
2081 | } |
2082 | else if (cmp == 0 && loop_vinfo) |
2083 | { |
2084 | gcc_assert (vls_type == VLS_LOAD); |
2085 | *memory_access_type = VMAT_INVARIANT; |
2086 | /* Invariant accesses perform only component accesses, alignment |
2087 | is irrelevant for them. */ |
2088 | *alignment_support_scheme = dr_unaligned_supported; |
2089 | } |
2090 | else |
2091 | *memory_access_type = VMAT_CONTIGUOUS; |
2092 | |
2093 | /* When we have a contiguous access across loop iterations |
2094 | but the access in the loop doesn't cover the full vector |
2095 | we can end up with no gap recorded but still excess |
2096 | elements accessed, see PR103116. Make sure we peel for |
2097 | gaps if necessary and sufficient and give up if not. |
2098 | |
2099 | If there is a combination of the access not covering the full |
2100 | vector and a gap recorded then we may need to peel twice. */ |
2101 | if (loop_vinfo |
2102 | && *memory_access_type == VMAT_CONTIGUOUS |
2103 | && SLP_TREE_LOAD_PERMUTATION (slp_node).exists () |
2104 | && !multiple_p (a: group_size * LOOP_VINFO_VECT_FACTOR (loop_vinfo), |
2105 | b: nunits)) |
2106 | { |
2107 | unsigned HOST_WIDE_INT cnunits, cvf; |
2108 | if (!can_overrun_p |
2109 | || !nunits.is_constant (const_value: &cnunits) |
2110 | || !LOOP_VINFO_VECT_FACTOR (loop_vinfo).is_constant (const_value: &cvf) |
2111 | /* Peeling for gaps assumes that a single scalar iteration |
2112 | is enough to make sure the last vector iteration doesn't |
2113 | access excess elements. |
2114 | ??? Enhancements include peeling multiple iterations |
2115 | or using masked loads with a static mask. */ |
2116 | || (group_size * cvf) % cnunits + group_size - gap < cnunits) |
2117 | { |
2118 | if (dump_enabled_p ()) |
2119 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
2120 | "peeling for gaps insufficient for " |
2121 | "access\n" ); |
2122 | return false; |
2123 | } |
2124 | overrun_p = true; |
2125 | } |
2126 | } |
2127 | } |
2128 | else |
2129 | { |
2130 | /* We can always handle this case using elementwise accesses, |
2131 | but see if something more efficient is available. */ |
2132 | *memory_access_type = VMAT_ELEMENTWISE; |
2133 | |
2134 | /* If there is a gap at the end of the group then these optimizations |
2135 | would access excess elements in the last iteration. */ |
2136 | bool would_overrun_p = (gap != 0); |
2137 | /* An overrun is fine if the trailing elements are smaller than the |
2138 | alignment boundary B. Every vector access will be a multiple of B |
2139 | and so we are guaranteed to access a non-gap element in the |
2140 | same B-sized block. */ |
2141 | if (would_overrun_p |
2142 | && !masked_p |
2143 | && gap < (vect_known_alignment_in_bytes (dr_info: first_dr_info, vectype) |
2144 | / vect_get_scalar_dr_size (dr_info: first_dr_info))) |
2145 | would_overrun_p = false; |
2146 | |
2147 | if (!STMT_VINFO_STRIDED_P (first_stmt_info) |
2148 | && (can_overrun_p || !would_overrun_p) |
2149 | && compare_step_with_zero (vinfo, stmt_info) > 0) |
2150 | { |
2151 | /* First cope with the degenerate case of a single-element |
2152 | vector. */ |
2153 | if (known_eq (TYPE_VECTOR_SUBPARTS (vectype), 1U)) |
2154 | ; |
2155 | |
2156 | else |
2157 | { |
2158 | /* Otherwise try using LOAD/STORE_LANES. */ |
2159 | *lanes_ifn |
2160 | = vls_type == VLS_LOAD |
2161 | ? vect_load_lanes_supported (vectype, group_size, masked_p) |
2162 | : vect_store_lanes_supported (vectype, group_size, |
2163 | masked_p); |
2164 | if (*lanes_ifn != IFN_LAST) |
2165 | { |
2166 | *memory_access_type = VMAT_LOAD_STORE_LANES; |
2167 | overrun_p = would_overrun_p; |
2168 | } |
2169 | |
2170 | /* If that fails, try using permuting loads. */ |
2171 | else if (vls_type == VLS_LOAD |
2172 | ? vect_grouped_load_supported (vectype, |
2173 | single_element_p, |
2174 | group_size) |
2175 | : vect_grouped_store_supported (vectype, group_size)) |
2176 | { |
2177 | *memory_access_type = VMAT_CONTIGUOUS_PERMUTE; |
2178 | overrun_p = would_overrun_p; |
2179 | } |
2180 | } |
2181 | } |
2182 | |
2183 | /* As a last resort, trying using a gather load or scatter store. |
2184 | |
2185 | ??? Although the code can handle all group sizes correctly, |
2186 | it probably isn't a win to use separate strided accesses based |
2187 | on nearby locations. Or, even if it's a win over scalar code, |
2188 | it might not be a win over vectorizing at a lower VF, if that |
2189 | allows us to use contiguous accesses. */ |
2190 | if (*memory_access_type == VMAT_ELEMENTWISE |
2191 | && single_element_p |
2192 | && loop_vinfo |
2193 | && vect_use_strided_gather_scatters_p (stmt_info, loop_vinfo, |
2194 | masked_p, gs_info)) |
2195 | *memory_access_type = VMAT_GATHER_SCATTER; |
2196 | } |
2197 | |
2198 | if (*memory_access_type == VMAT_GATHER_SCATTER |
2199 | || *memory_access_type == VMAT_ELEMENTWISE) |
2200 | { |
2201 | *alignment_support_scheme = dr_unaligned_supported; |
2202 | *misalignment = DR_MISALIGNMENT_UNKNOWN; |
2203 | } |
2204 | else |
2205 | { |
2206 | *misalignment = dr_misalignment (dr_info: first_dr_info, vectype, offset: *poffset); |
2207 | *alignment_support_scheme |
2208 | = vect_supportable_dr_alignment (vinfo, first_dr_info, vectype, |
2209 | *misalignment); |
2210 | } |
2211 | |
2212 | if (vls_type != VLS_LOAD && first_stmt_info == stmt_info) |
2213 | { |
2214 | /* STMT is the leader of the group. Check the operands of all the |
2215 | stmts of the group. */ |
2216 | stmt_vec_info next_stmt_info = DR_GROUP_NEXT_ELEMENT (stmt_info); |
2217 | while (next_stmt_info) |
2218 | { |
2219 | tree op = vect_get_store_rhs (stmt_info: next_stmt_info); |
2220 | enum vect_def_type dt; |
2221 | if (!vect_is_simple_use (op, vinfo, &dt)) |
2222 | { |
2223 | if (dump_enabled_p ()) |
2224 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
2225 | "use not simple.\n" ); |
2226 | return false; |
2227 | } |
2228 | next_stmt_info = DR_GROUP_NEXT_ELEMENT (next_stmt_info); |
2229 | } |
2230 | } |
2231 | |
2232 | if (overrun_p) |
2233 | { |
2234 | gcc_assert (can_overrun_p); |
2235 | if (dump_enabled_p ()) |
2236 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
2237 | "Data access with gaps requires scalar " |
2238 | "epilogue loop\n" ); |
2239 | LOOP_VINFO_PEELING_FOR_GAPS (loop_vinfo) = true; |
2240 | } |
2241 | |
2242 | return true; |
2243 | } |
2244 | |
2245 | /* Analyze load or store statement STMT_INFO of type VLS_TYPE. Return true |
2246 | if there is a memory access type that the vectorized form can use, |
2247 | storing it in *MEMORY_ACCESS_TYPE if so. If we decide to use gathers |
2248 | or scatters, fill in GS_INFO accordingly. In addition |
2249 | *ALIGNMENT_SUPPORT_SCHEME is filled out and false is returned if |
2250 | the target does not support the alignment scheme. *MISALIGNMENT |
2251 | is set according to the alignment of the access (including |
2252 | DR_MISALIGNMENT_UNKNOWN when it is unknown). |
2253 | |
2254 | SLP says whether we're performing SLP rather than loop vectorization. |
2255 | MASKED_P is true if the statement is conditional on a vectorized mask. |
2256 | VECTYPE is the vector type that the vectorized statements will use. |
2257 | NCOPIES is the number of vector statements that will be needed. */ |
2258 | |
2259 | static bool |
2260 | get_load_store_type (vec_info *vinfo, stmt_vec_info stmt_info, |
2261 | tree vectype, slp_tree slp_node, |
2262 | bool masked_p, vec_load_store_type vls_type, |
2263 | unsigned int ncopies, |
2264 | vect_memory_access_type *memory_access_type, |
2265 | poly_int64 *poffset, |
2266 | dr_alignment_support *alignment_support_scheme, |
2267 | int *misalignment, |
2268 | gather_scatter_info *gs_info, |
2269 | internal_fn *lanes_ifn) |
2270 | { |
2271 | loop_vec_info loop_vinfo = dyn_cast <loop_vec_info> (p: vinfo); |
2272 | poly_uint64 nunits = TYPE_VECTOR_SUBPARTS (node: vectype); |
2273 | *misalignment = DR_MISALIGNMENT_UNKNOWN; |
2274 | *poffset = 0; |
2275 | if (STMT_VINFO_GATHER_SCATTER_P (stmt_info)) |
2276 | { |
2277 | *memory_access_type = VMAT_GATHER_SCATTER; |
2278 | if (!vect_check_gather_scatter (stmt_info, loop_vinfo, gs_info)) |
2279 | gcc_unreachable (); |
2280 | /* When using internal functions, we rely on pattern recognition |
2281 | to convert the type of the offset to the type that the target |
2282 | requires, with the result being a call to an internal function. |
2283 | If that failed for some reason (e.g. because another pattern |
2284 | took priority), just handle cases in which the offset already |
2285 | has the right type. */ |
2286 | else if (gs_info->ifn != IFN_LAST |
2287 | && !is_gimple_call (gs: stmt_info->stmt) |
2288 | && !tree_nop_conversion_p (TREE_TYPE (gs_info->offset), |
2289 | TREE_TYPE (gs_info->offset_vectype))) |
2290 | { |
2291 | if (dump_enabled_p ()) |
2292 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
2293 | "%s offset requires a conversion\n" , |
2294 | vls_type == VLS_LOAD ? "gather" : "scatter" ); |
2295 | return false; |
2296 | } |
2297 | else if (!vect_is_simple_use (gs_info->offset, vinfo, |
2298 | &gs_info->offset_dt, |
2299 | &gs_info->offset_vectype)) |
2300 | { |
2301 | if (dump_enabled_p ()) |
2302 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
2303 | "%s index use not simple.\n" , |
2304 | vls_type == VLS_LOAD ? "gather" : "scatter" ); |
2305 | return false; |
2306 | } |
2307 | else if (gs_info->ifn == IFN_LAST && !gs_info->decl) |
2308 | { |
2309 | if (!TYPE_VECTOR_SUBPARTS (node: vectype).is_constant () |
2310 | || !TYPE_VECTOR_SUBPARTS (node: gs_info->offset_vectype).is_constant () |
2311 | || !constant_multiple_p (a: TYPE_VECTOR_SUBPARTS |
2312 | (node: gs_info->offset_vectype), |
2313 | b: TYPE_VECTOR_SUBPARTS (node: vectype))) |
2314 | { |
2315 | if (dump_enabled_p ()) |
2316 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
2317 | "unsupported vector types for emulated " |
2318 | "gather.\n" ); |
2319 | return false; |
2320 | } |
2321 | } |
2322 | /* Gather-scatter accesses perform only component accesses, alignment |
2323 | is irrelevant for them. */ |
2324 | *alignment_support_scheme = dr_unaligned_supported; |
2325 | } |
2326 | else if (STMT_VINFO_GROUPED_ACCESS (stmt_info) || slp_node) |
2327 | { |
2328 | if (!get_group_load_store_type (vinfo, stmt_info, vectype, slp_node, |
2329 | masked_p, |
2330 | vls_type, memory_access_type, poffset, |
2331 | alignment_support_scheme, |
2332 | misalignment, gs_info, lanes_ifn)) |
2333 | return false; |
2334 | } |
2335 | else if (STMT_VINFO_STRIDED_P (stmt_info)) |
2336 | { |
2337 | gcc_assert (!slp_node); |
2338 | if (loop_vinfo |
2339 | && vect_use_strided_gather_scatters_p (stmt_info, loop_vinfo, |
2340 | masked_p, gs_info)) |
2341 | *memory_access_type = VMAT_GATHER_SCATTER; |
2342 | else |
2343 | *memory_access_type = VMAT_ELEMENTWISE; |
2344 | /* Alignment is irrelevant here. */ |
2345 | *alignment_support_scheme = dr_unaligned_supported; |
2346 | } |
2347 | else |
2348 | { |
2349 | int cmp = compare_step_with_zero (vinfo, stmt_info); |
2350 | if (cmp == 0) |
2351 | { |
2352 | gcc_assert (vls_type == VLS_LOAD); |
2353 | *memory_access_type = VMAT_INVARIANT; |
2354 | /* Invariant accesses perform only component accesses, alignment |
2355 | is irrelevant for them. */ |
2356 | *alignment_support_scheme = dr_unaligned_supported; |
2357 | } |
2358 | else |
2359 | { |
2360 | if (cmp < 0) |
2361 | *memory_access_type = get_negative_load_store_type |
2362 | (vinfo, stmt_info, vectype, vls_type, ncopies, poffset); |
2363 | else |
2364 | *memory_access_type = VMAT_CONTIGUOUS; |
2365 | *misalignment = dr_misalignment (STMT_VINFO_DR_INFO (stmt_info), |
2366 | vectype, offset: *poffset); |
2367 | *alignment_support_scheme |
2368 | = vect_supportable_dr_alignment (vinfo, |
2369 | STMT_VINFO_DR_INFO (stmt_info), |
2370 | vectype, *misalignment); |
2371 | } |
2372 | } |
2373 | |
2374 | if ((*memory_access_type == VMAT_ELEMENTWISE |
2375 | || *memory_access_type == VMAT_STRIDED_SLP) |
2376 | && !nunits.is_constant ()) |
2377 | { |
2378 | if (dump_enabled_p ()) |
2379 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
2380 | "Not using elementwise accesses due to variable " |
2381 | "vectorization factor.\n" ); |
2382 | return false; |
2383 | } |
2384 | |
2385 | if (*alignment_support_scheme == dr_unaligned_unsupported) |
2386 | { |
2387 | if (dump_enabled_p ()) |
2388 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
2389 | "unsupported unaligned access\n" ); |
2390 | return false; |
2391 | } |
2392 | |
2393 | /* FIXME: At the moment the cost model seems to underestimate the |
2394 | cost of using elementwise accesses. This check preserves the |
2395 | traditional behavior until that can be fixed. */ |
2396 | stmt_vec_info first_stmt_info = DR_GROUP_FIRST_ELEMENT (stmt_info); |
2397 | if (!first_stmt_info) |
2398 | first_stmt_info = stmt_info; |
2399 | if (*memory_access_type == VMAT_ELEMENTWISE |
2400 | && !STMT_VINFO_STRIDED_P (first_stmt_info) |
2401 | && !(stmt_info == DR_GROUP_FIRST_ELEMENT (stmt_info) |
2402 | && !DR_GROUP_NEXT_ELEMENT (stmt_info) |
2403 | && !pow2p_hwi (DR_GROUP_SIZE (stmt_info)))) |
2404 | { |
2405 | if (dump_enabled_p ()) |
2406 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
2407 | "not falling back to elementwise accesses\n" ); |
2408 | return false; |
2409 | } |
2410 | return true; |
2411 | } |
2412 | |
2413 | /* Return true if boolean argument at MASK_INDEX is suitable for vectorizing |
2414 | conditional operation STMT_INFO. When returning true, store the mask |
2415 | in *MASK, the type of its definition in *MASK_DT_OUT, the type of the |
2416 | vectorized mask in *MASK_VECTYPE_OUT and the SLP node corresponding |
2417 | to the mask in *MASK_NODE if MASK_NODE is not NULL. */ |
2418 | |
2419 | static bool |
2420 | vect_check_scalar_mask (vec_info *vinfo, stmt_vec_info stmt_info, |
2421 | slp_tree slp_node, unsigned mask_index, |
2422 | tree *mask, slp_tree *mask_node, |
2423 | vect_def_type *mask_dt_out, tree *mask_vectype_out) |
2424 | { |
2425 | enum vect_def_type mask_dt; |
2426 | tree mask_vectype; |
2427 | slp_tree mask_node_1; |
2428 | if (!vect_is_simple_use (vinfo, stmt_info, slp_node, mask_index, |
2429 | mask, &mask_node_1, &mask_dt, &mask_vectype)) |
2430 | { |
2431 | if (dump_enabled_p ()) |
2432 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
2433 | "mask use not simple.\n" ); |
2434 | return false; |
2435 | } |
2436 | |
2437 | if (!VECT_SCALAR_BOOLEAN_TYPE_P (TREE_TYPE (*mask))) |
2438 | { |
2439 | if (dump_enabled_p ()) |
2440 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
2441 | "mask argument is not a boolean.\n" ); |
2442 | return false; |
2443 | } |
2444 | |
2445 | /* If the caller is not prepared for adjusting an external/constant |
2446 | SLP mask vector type fail. */ |
2447 | if (slp_node |
2448 | && !mask_node |
2449 | && SLP_TREE_DEF_TYPE (mask_node_1) != vect_internal_def) |
2450 | { |
2451 | if (dump_enabled_p ()) |
2452 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
2453 | "SLP mask argument is not vectorized.\n" ); |
2454 | return false; |
2455 | } |
2456 | |
2457 | tree vectype = STMT_VINFO_VECTYPE (stmt_info); |
2458 | if (!mask_vectype) |
2459 | mask_vectype = get_mask_type_for_scalar_type (vinfo, TREE_TYPE (vectype), |
2460 | mask_node_1); |
2461 | |
2462 | if (!mask_vectype || !VECTOR_BOOLEAN_TYPE_P (mask_vectype)) |
2463 | { |
2464 | if (dump_enabled_p ()) |
2465 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
2466 | "could not find an appropriate vector mask type.\n" ); |
2467 | return false; |
2468 | } |
2469 | |
2470 | if (maybe_ne (a: TYPE_VECTOR_SUBPARTS (node: mask_vectype), |
2471 | b: TYPE_VECTOR_SUBPARTS (node: vectype))) |
2472 | { |
2473 | if (dump_enabled_p ()) |
2474 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
2475 | "vector mask type %T" |
2476 | " does not match vector data type %T.\n" , |
2477 | mask_vectype, vectype); |
2478 | |
2479 | return false; |
2480 | } |
2481 | |
2482 | *mask_dt_out = mask_dt; |
2483 | *mask_vectype_out = mask_vectype; |
2484 | if (mask_node) |
2485 | *mask_node = mask_node_1; |
2486 | return true; |
2487 | } |
2488 | |
2489 | /* Return true if stored value is suitable for vectorizing store |
2490 | statement STMT_INFO. When returning true, store the scalar stored |
2491 | in *RHS and *RHS_NODE, the type of the definition in *RHS_DT_OUT, |
2492 | the type of the vectorized store value in |
2493 | *RHS_VECTYPE_OUT and the type of the store in *VLS_TYPE_OUT. */ |
2494 | |
2495 | static bool |
2496 | vect_check_store_rhs (vec_info *vinfo, stmt_vec_info stmt_info, |
2497 | slp_tree slp_node, tree *rhs, slp_tree *rhs_node, |
2498 | vect_def_type *rhs_dt_out, tree *rhs_vectype_out, |
2499 | vec_load_store_type *vls_type_out) |
2500 | { |
2501 | int op_no = 0; |
2502 | if (gcall *call = dyn_cast <gcall *> (p: stmt_info->stmt)) |
2503 | { |
2504 | if (gimple_call_internal_p (gs: call) |
2505 | && internal_store_fn_p (gimple_call_internal_fn (gs: call))) |
2506 | op_no = internal_fn_stored_value_index (gimple_call_internal_fn (gs: call)); |
2507 | } |
2508 | if (slp_node) |
2509 | op_no = vect_slp_child_index_for_operand |
2510 | (stmt_info->stmt, op: op_no, STMT_VINFO_GATHER_SCATTER_P (stmt_info)); |
2511 | |
2512 | enum vect_def_type rhs_dt; |
2513 | tree rhs_vectype; |
2514 | if (!vect_is_simple_use (vinfo, stmt_info, slp_node, op_no, |
2515 | rhs, rhs_node, &rhs_dt, &rhs_vectype)) |
2516 | { |
2517 | if (dump_enabled_p ()) |
2518 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
2519 | "use not simple.\n" ); |
2520 | return false; |
2521 | } |
2522 | |
2523 | /* In the case this is a store from a constant make sure |
2524 | native_encode_expr can handle it. */ |
2525 | if (CONSTANT_CLASS_P (*rhs) && native_encode_expr (*rhs, NULL, 64) == 0) |
2526 | { |
2527 | if (dump_enabled_p ()) |
2528 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
2529 | "cannot encode constant as a byte sequence.\n" ); |
2530 | return false; |
2531 | } |
2532 | |
2533 | tree vectype = STMT_VINFO_VECTYPE (stmt_info); |
2534 | if (rhs_vectype && !useless_type_conversion_p (vectype, rhs_vectype)) |
2535 | { |
2536 | if (dump_enabled_p ()) |
2537 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
2538 | "incompatible vector types.\n" ); |
2539 | return false; |
2540 | } |
2541 | |
2542 | *rhs_dt_out = rhs_dt; |
2543 | *rhs_vectype_out = rhs_vectype; |
2544 | if (rhs_dt == vect_constant_def || rhs_dt == vect_external_def) |
2545 | *vls_type_out = VLS_STORE_INVARIANT; |
2546 | else |
2547 | *vls_type_out = VLS_STORE; |
2548 | return true; |
2549 | } |
2550 | |
2551 | /* Build an all-ones vector mask of type MASKTYPE while vectorizing STMT_INFO. |
2552 | Note that we support masks with floating-point type, in which case the |
2553 | floats are interpreted as a bitmask. */ |
2554 | |
2555 | static tree |
2556 | vect_build_all_ones_mask (vec_info *vinfo, |
2557 | stmt_vec_info stmt_info, tree masktype) |
2558 | { |
2559 | if (TREE_CODE (masktype) == INTEGER_TYPE) |
2560 | return build_int_cst (masktype, -1); |
2561 | else if (VECTOR_BOOLEAN_TYPE_P (masktype) |
2562 | || TREE_CODE (TREE_TYPE (masktype)) == INTEGER_TYPE) |
2563 | { |
2564 | tree mask = build_int_cst (TREE_TYPE (masktype), -1); |
2565 | mask = build_vector_from_val (masktype, mask); |
2566 | return vect_init_vector (vinfo, stmt_info, val: mask, type: masktype, NULL); |
2567 | } |
2568 | else if (SCALAR_FLOAT_TYPE_P (TREE_TYPE (masktype))) |
2569 | { |
2570 | REAL_VALUE_TYPE r; |
2571 | long tmp[6]; |
2572 | for (int j = 0; j < 6; ++j) |
2573 | tmp[j] = -1; |
2574 | real_from_target (&r, tmp, TYPE_MODE (TREE_TYPE (masktype))); |
2575 | tree mask = build_real (TREE_TYPE (masktype), r); |
2576 | mask = build_vector_from_val (masktype, mask); |
2577 | return vect_init_vector (vinfo, stmt_info, val: mask, type: masktype, NULL); |
2578 | } |
2579 | gcc_unreachable (); |
2580 | } |
2581 | |
2582 | /* Build an all-zero merge value of type VECTYPE while vectorizing |
2583 | STMT_INFO as a gather load. */ |
2584 | |
2585 | static tree |
2586 | vect_build_zero_merge_argument (vec_info *vinfo, |
2587 | stmt_vec_info stmt_info, tree vectype) |
2588 | { |
2589 | tree merge; |
2590 | if (TREE_CODE (TREE_TYPE (vectype)) == INTEGER_TYPE) |
2591 | merge = build_int_cst (TREE_TYPE (vectype), 0); |
2592 | else if (SCALAR_FLOAT_TYPE_P (TREE_TYPE (vectype))) |
2593 | { |
2594 | REAL_VALUE_TYPE r; |
2595 | long tmp[6]; |
2596 | for (int j = 0; j < 6; ++j) |
2597 | tmp[j] = 0; |
2598 | real_from_target (&r, tmp, TYPE_MODE (TREE_TYPE (vectype))); |
2599 | merge = build_real (TREE_TYPE (vectype), r); |
2600 | } |
2601 | else |
2602 | gcc_unreachable (); |
2603 | merge = build_vector_from_val (vectype, merge); |
2604 | return vect_init_vector (vinfo, stmt_info, val: merge, type: vectype, NULL); |
2605 | } |
2606 | |
2607 | /* Build a gather load call while vectorizing STMT_INFO. Insert new |
2608 | instructions before GSI and add them to VEC_STMT. GS_INFO describes |
2609 | the gather load operation. If the load is conditional, MASK is the |
2610 | vectorized condition, otherwise MASK is null. PTR is the base |
2611 | pointer and OFFSET is the vectorized offset. */ |
2612 | |
2613 | static gimple * |
2614 | vect_build_one_gather_load_call (vec_info *vinfo, stmt_vec_info stmt_info, |
2615 | gimple_stmt_iterator *gsi, |
2616 | gather_scatter_info *gs_info, |
2617 | tree ptr, tree offset, tree mask) |
2618 | { |
2619 | tree vectype = STMT_VINFO_VECTYPE (stmt_info); |
2620 | tree arglist = TYPE_ARG_TYPES (TREE_TYPE (gs_info->decl)); |
2621 | tree rettype = TREE_TYPE (TREE_TYPE (gs_info->decl)); |
2622 | tree srctype = TREE_VALUE (arglist); arglist = TREE_CHAIN (arglist); |
2623 | /* ptrtype */ arglist = TREE_CHAIN (arglist); |
2624 | tree idxtype = TREE_VALUE (arglist); arglist = TREE_CHAIN (arglist); |
2625 | tree masktype = TREE_VALUE (arglist); arglist = TREE_CHAIN (arglist); |
2626 | tree scaletype = TREE_VALUE (arglist); |
2627 | tree var; |
2628 | gcc_checking_assert (types_compatible_p (srctype, rettype) |
2629 | && (!mask |
2630 | || TREE_CODE (masktype) == INTEGER_TYPE |
2631 | || types_compatible_p (srctype, masktype))); |
2632 | |
2633 | tree op = offset; |
2634 | if (!useless_type_conversion_p (idxtype, TREE_TYPE (op))) |
2635 | { |
2636 | gcc_assert (known_eq (TYPE_VECTOR_SUBPARTS (TREE_TYPE (op)), |
2637 | TYPE_VECTOR_SUBPARTS (idxtype))); |
2638 | var = vect_get_new_ssa_name (idxtype, vect_simple_var); |
2639 | op = build1 (VIEW_CONVERT_EXPR, idxtype, op); |
2640 | gassign *new_stmt = gimple_build_assign (var, VIEW_CONVERT_EXPR, op); |
2641 | vect_finish_stmt_generation (vinfo, stmt_info, vec_stmt: new_stmt, gsi); |
2642 | op = var; |
2643 | } |
2644 | |
2645 | tree src_op = NULL_TREE; |
2646 | tree mask_op = NULL_TREE; |
2647 | if (mask) |
2648 | { |
2649 | if (!useless_type_conversion_p (masktype, TREE_TYPE (mask))) |
2650 | { |
2651 | tree utype, optype = TREE_TYPE (mask); |
2652 | if (VECTOR_TYPE_P (masktype) |
2653 | || TYPE_MODE (masktype) == TYPE_MODE (optype)) |
2654 | utype = masktype; |
2655 | else |
2656 | utype = lang_hooks.types.type_for_mode (TYPE_MODE (optype), 1); |
2657 | var = vect_get_new_ssa_name (utype, vect_scalar_var); |
2658 | tree mask_arg = build1 (VIEW_CONVERT_EXPR, utype, mask); |
2659 | gassign *new_stmt |
2660 | = gimple_build_assign (var, VIEW_CONVERT_EXPR, mask_arg); |
2661 | vect_finish_stmt_generation (vinfo, stmt_info, vec_stmt: new_stmt, gsi); |
2662 | mask_arg = var; |
2663 | if (!useless_type_conversion_p (masktype, utype)) |
2664 | { |
2665 | gcc_assert (TYPE_PRECISION (utype) |
2666 | <= TYPE_PRECISION (masktype)); |
2667 | var = vect_get_new_ssa_name (masktype, vect_scalar_var); |
2668 | new_stmt = gimple_build_assign (var, NOP_EXPR, mask_arg); |
2669 | vect_finish_stmt_generation (vinfo, stmt_info, vec_stmt: new_stmt, gsi); |
2670 | mask_arg = var; |
2671 | } |
2672 | src_op = build_zero_cst (srctype); |
2673 | mask_op = mask_arg; |
2674 | } |
2675 | else |
2676 | { |
2677 | src_op = mask; |
2678 | mask_op = mask; |
2679 | } |
2680 | } |
2681 | else |
2682 | { |
2683 | src_op = vect_build_zero_merge_argument (vinfo, stmt_info, vectype: rettype); |
2684 | mask_op = vect_build_all_ones_mask (vinfo, stmt_info, masktype); |
2685 | } |
2686 | |
2687 | tree scale = build_int_cst (scaletype, gs_info->scale); |
2688 | gimple *new_stmt = gimple_build_call (gs_info->decl, 5, src_op, ptr, op, |
2689 | mask_op, scale); |
2690 | |
2691 | if (!useless_type_conversion_p (vectype, rettype)) |
2692 | { |
2693 | gcc_assert (known_eq (TYPE_VECTOR_SUBPARTS (vectype), |
2694 | TYPE_VECTOR_SUBPARTS (rettype))); |
2695 | op = vect_get_new_ssa_name (rettype, vect_simple_var); |
2696 | gimple_call_set_lhs (gs: new_stmt, lhs: op); |
2697 | vect_finish_stmt_generation (vinfo, stmt_info, vec_stmt: new_stmt, gsi); |
2698 | op = build1 (VIEW_CONVERT_EXPR, vectype, op); |
2699 | new_stmt = gimple_build_assign (NULL_TREE, VIEW_CONVERT_EXPR, op); |
2700 | } |
2701 | |
2702 | return new_stmt; |
2703 | } |
2704 | |
2705 | /* Build a scatter store call while vectorizing STMT_INFO. Insert new |
2706 | instructions before GSI. GS_INFO describes the scatter store operation. |
2707 | PTR is the base pointer, OFFSET the vectorized offsets and OPRND the |
2708 | vectorized data to store. |
2709 | If the store is conditional, MASK is the vectorized condition, otherwise |
2710 | MASK is null. */ |
2711 | |
2712 | static gimple * |
2713 | vect_build_one_scatter_store_call (vec_info *vinfo, stmt_vec_info stmt_info, |
2714 | gimple_stmt_iterator *gsi, |
2715 | gather_scatter_info *gs_info, |
2716 | tree ptr, tree offset, tree oprnd, tree mask) |
2717 | { |
2718 | tree rettype = TREE_TYPE (TREE_TYPE (gs_info->decl)); |
2719 | tree arglist = TYPE_ARG_TYPES (TREE_TYPE (gs_info->decl)); |
2720 | /* tree ptrtype = TREE_VALUE (arglist); */ arglist = TREE_CHAIN (arglist); |
2721 | tree masktype = TREE_VALUE (arglist); arglist = TREE_CHAIN (arglist); |
2722 | tree idxtype = TREE_VALUE (arglist); arglist = TREE_CHAIN (arglist); |
2723 | tree srctype = TREE_VALUE (arglist); arglist = TREE_CHAIN (arglist); |
2724 | tree scaletype = TREE_VALUE (arglist); |
2725 | gcc_checking_assert (TREE_CODE (masktype) == INTEGER_TYPE |
2726 | && TREE_CODE (rettype) == VOID_TYPE); |
2727 | |
2728 | tree mask_arg = NULL_TREE; |
2729 | if (mask) |
2730 | { |
2731 | mask_arg = mask; |
2732 | tree optype = TREE_TYPE (mask_arg); |
2733 | tree utype; |
2734 | if (TYPE_MODE (masktype) == TYPE_MODE (optype)) |
2735 | utype = masktype; |
2736 | else |
2737 | utype = lang_hooks.types.type_for_mode (TYPE_MODE (optype), 1); |
2738 | tree var = vect_get_new_ssa_name (utype, vect_scalar_var); |
2739 | mask_arg = build1 (VIEW_CONVERT_EXPR, utype, mask_arg); |
2740 | gassign *new_stmt |
2741 | = gimple_build_assign (var, VIEW_CONVERT_EXPR, mask_arg); |
2742 | vect_finish_stmt_generation (vinfo, stmt_info, vec_stmt: new_stmt, gsi); |
2743 | mask_arg = var; |
2744 | if (!useless_type_conversion_p (masktype, utype)) |
2745 | { |
2746 | gcc_assert (TYPE_PRECISION (utype) <= TYPE_PRECISION (masktype)); |
2747 | tree var = vect_get_new_ssa_name (masktype, vect_scalar_var); |
2748 | new_stmt = gimple_build_assign (var, NOP_EXPR, mask_arg); |
2749 | vect_finish_stmt_generation (vinfo, stmt_info, vec_stmt: new_stmt, gsi); |
2750 | mask_arg = var; |
2751 | } |
2752 | } |
2753 | else |
2754 | { |
2755 | mask_arg = build_int_cst (masktype, -1); |
2756 | mask_arg = vect_init_vector (vinfo, stmt_info, val: mask_arg, type: masktype, NULL); |
2757 | } |
2758 | |
2759 | tree src = oprnd; |
2760 | if (!useless_type_conversion_p (srctype, TREE_TYPE (src))) |
2761 | { |
2762 | gcc_assert (known_eq (TYPE_VECTOR_SUBPARTS (TREE_TYPE (src)), |
2763 | TYPE_VECTOR_SUBPARTS (srctype))); |
2764 | tree var = vect_get_new_ssa_name (srctype, vect_simple_var); |
2765 | src = build1 (VIEW_CONVERT_EXPR, srctype, src); |
2766 | gassign *new_stmt = gimple_build_assign (var, VIEW_CONVERT_EXPR, src); |
2767 | vect_finish_stmt_generation (vinfo, stmt_info, vec_stmt: new_stmt, gsi); |
2768 | src = var; |
2769 | } |
2770 | |
2771 | tree op = offset; |
2772 | if (!useless_type_conversion_p (idxtype, TREE_TYPE (op))) |
2773 | { |
2774 | gcc_assert (known_eq (TYPE_VECTOR_SUBPARTS (TREE_TYPE (op)), |
2775 | TYPE_VECTOR_SUBPARTS (idxtype))); |
2776 | tree var = vect_get_new_ssa_name (idxtype, vect_simple_var); |
2777 | op = build1 (VIEW_CONVERT_EXPR, idxtype, op); |
2778 | gassign *new_stmt = gimple_build_assign (var, VIEW_CONVERT_EXPR, op); |
2779 | vect_finish_stmt_generation (vinfo, stmt_info, vec_stmt: new_stmt, gsi); |
2780 | op = var; |
2781 | } |
2782 | |
2783 | tree scale = build_int_cst (scaletype, gs_info->scale); |
2784 | gcall *new_stmt |
2785 | = gimple_build_call (gs_info->decl, 5, ptr, mask_arg, op, src, scale); |
2786 | return new_stmt; |
2787 | } |
2788 | |
2789 | /* Prepare the base and offset in GS_INFO for vectorization. |
2790 | Set *DATAREF_PTR to the loop-invariant base address and *VEC_OFFSET |
2791 | to the vectorized offset argument for the first copy of STMT_INFO. |
2792 | STMT_INFO is the statement described by GS_INFO and LOOP is the |
2793 | containing loop. */ |
2794 | |
2795 | static void |
2796 | vect_get_gather_scatter_ops (loop_vec_info loop_vinfo, |
2797 | class loop *loop, stmt_vec_info stmt_info, |
2798 | slp_tree slp_node, gather_scatter_info *gs_info, |
2799 | tree *dataref_ptr, vec<tree> *vec_offset) |
2800 | { |
2801 | gimple_seq stmts = NULL; |
2802 | *dataref_ptr = force_gimple_operand (gs_info->base, &stmts, true, NULL_TREE); |
2803 | if (stmts != NULL) |
2804 | { |
2805 | basic_block new_bb; |
2806 | edge pe = loop_preheader_edge (loop); |
2807 | new_bb = gsi_insert_seq_on_edge_immediate (pe, stmts); |
2808 | gcc_assert (!new_bb); |
2809 | } |
2810 | if (slp_node) |
2811 | vect_get_slp_defs (SLP_TREE_CHILDREN (slp_node)[0], vec_offset); |
2812 | else |
2813 | { |
2814 | unsigned ncopies |
2815 | = vect_get_num_copies (loop_vinfo, vectype: gs_info->offset_vectype); |
2816 | vect_get_vec_defs_for_operand (vinfo: loop_vinfo, stmt_vinfo: stmt_info, ncopies, |
2817 | op: gs_info->offset, vec_oprnds: vec_offset, |
2818 | vectype: gs_info->offset_vectype); |
2819 | } |
2820 | } |
2821 | |
2822 | /* Prepare to implement a grouped or strided load or store using |
2823 | the gather load or scatter store operation described by GS_INFO. |
2824 | STMT_INFO is the load or store statement. |
2825 | |
2826 | Set *DATAREF_BUMP to the amount that should be added to the base |
2827 | address after each copy of the vectorized statement. Set *VEC_OFFSET |
2828 | to an invariant offset vector in which element I has the value |
2829 | I * DR_STEP / SCALE. */ |
2830 | |
2831 | static void |
2832 | vect_get_strided_load_store_ops (stmt_vec_info stmt_info, |
2833 | loop_vec_info loop_vinfo, |
2834 | gimple_stmt_iterator *gsi, |
2835 | gather_scatter_info *gs_info, |
2836 | tree *dataref_bump, tree *vec_offset, |
2837 | vec_loop_lens *loop_lens) |
2838 | { |
2839 | struct data_reference *dr = STMT_VINFO_DATA_REF (stmt_info); |
2840 | tree vectype = STMT_VINFO_VECTYPE (stmt_info); |
2841 | |
2842 | if (LOOP_VINFO_USING_SELECT_VL_P (loop_vinfo)) |
2843 | { |
2844 | /* _31 = .SELECT_VL (ivtmp_29, POLY_INT_CST [4, 4]); |
2845 | ivtmp_8 = _31 * 16 (step in bytes); |
2846 | .MASK_LEN_SCATTER_STORE (vectp_a.9_7, ... ); |
2847 | vectp_a.9_26 = vectp_a.9_7 + ivtmp_8; */ |
2848 | tree loop_len |
2849 | = vect_get_loop_len (loop_vinfo, gsi, loop_lens, 1, vectype, 0, 0); |
2850 | tree tmp |
2851 | = fold_build2 (MULT_EXPR, sizetype, |
2852 | fold_convert (sizetype, unshare_expr (DR_STEP (dr))), |
2853 | loop_len); |
2854 | *dataref_bump = force_gimple_operand_gsi (gsi, tmp, true, NULL_TREE, true, |
2855 | GSI_SAME_STMT); |
2856 | } |
2857 | else |
2858 | { |
2859 | tree bump |
2860 | = size_binop (MULT_EXPR, |
2861 | fold_convert (sizetype, unshare_expr (DR_STEP (dr))), |
2862 | size_int (TYPE_VECTOR_SUBPARTS (vectype))); |
2863 | *dataref_bump = cse_and_gimplify_to_preheader (loop_vinfo, bump); |
2864 | } |
2865 | |
2866 | /* The offset given in GS_INFO can have pointer type, so use the element |
2867 | type of the vector instead. */ |
2868 | tree offset_type = TREE_TYPE (gs_info->offset_vectype); |
2869 | |
2870 | /* Calculate X = DR_STEP / SCALE and convert it to the appropriate type. */ |
2871 | tree step = size_binop (EXACT_DIV_EXPR, unshare_expr (DR_STEP (dr)), |
2872 | ssize_int (gs_info->scale)); |
2873 | step = fold_convert (offset_type, step); |
2874 | |
2875 | /* Create {0, X, X*2, X*3, ...}. */ |
2876 | tree offset = fold_build2 (VEC_SERIES_EXPR, gs_info->offset_vectype, |
2877 | build_zero_cst (offset_type), step); |
2878 | *vec_offset = cse_and_gimplify_to_preheader (loop_vinfo, offset); |
2879 | } |
2880 | |
2881 | /* Prepare the pointer IVs which needs to be updated by a variable amount. |
2882 | Such variable amount is the outcome of .SELECT_VL. In this case, we can |
2883 | allow each iteration process the flexible number of elements as long as |
2884 | the number <= vf elments. |
2885 | |
2886 | Return data reference according to SELECT_VL. |
2887 | If new statements are needed, insert them before GSI. */ |
2888 | |
2889 | static tree |
2890 | vect_get_loop_variant_data_ptr_increment ( |
2891 | vec_info *vinfo, tree aggr_type, gimple_stmt_iterator *gsi, |
2892 | vec_loop_lens *loop_lens, dr_vec_info *dr_info, |
2893 | vect_memory_access_type memory_access_type) |
2894 | { |
2895 | loop_vec_info loop_vinfo = dyn_cast<loop_vec_info> (p: vinfo); |
2896 | tree step = vect_dr_behavior (vinfo, dr_info)->step; |
2897 | |
2898 | /* gather/scatter never reach here. */ |
2899 | gcc_assert (memory_access_type != VMAT_GATHER_SCATTER); |
2900 | |
2901 | /* When we support SELECT_VL pattern, we dynamic adjust |
2902 | the memory address by .SELECT_VL result. |
2903 | |
2904 | The result of .SELECT_VL is the number of elements to |
2905 | be processed of each iteration. So the memory address |
2906 | adjustment operation should be: |
2907 | |
2908 | addr = addr + .SELECT_VL (ARG..) * step; |
2909 | */ |
2910 | tree loop_len |
2911 | = vect_get_loop_len (loop_vinfo, gsi, loop_lens, 1, aggr_type, 0, 0); |
2912 | tree len_type = TREE_TYPE (loop_len); |
2913 | /* Since the outcome of .SELECT_VL is element size, we should adjust |
2914 | it into bytesize so that it can be used in address pointer variable |
2915 | amount IVs adjustment. */ |
2916 | tree tmp = fold_build2 (MULT_EXPR, len_type, loop_len, |
2917 | wide_int_to_tree (len_type, wi::to_widest (step))); |
2918 | tree bump = make_temp_ssa_name (type: len_type, NULL, name: "ivtmp" ); |
2919 | gassign *assign = gimple_build_assign (bump, tmp); |
2920 | gsi_insert_before (gsi, assign, GSI_SAME_STMT); |
2921 | return bump; |
2922 | } |
2923 | |
2924 | /* Return the amount that should be added to a vector pointer to move |
2925 | to the next or previous copy of AGGR_TYPE. DR_INFO is the data reference |
2926 | being vectorized and MEMORY_ACCESS_TYPE describes the type of |
2927 | vectorization. */ |
2928 | |
2929 | static tree |
2930 | vect_get_data_ptr_increment (vec_info *vinfo, gimple_stmt_iterator *gsi, |
2931 | dr_vec_info *dr_info, tree aggr_type, |
2932 | vect_memory_access_type memory_access_type, |
2933 | vec_loop_lens *loop_lens = nullptr) |
2934 | { |
2935 | if (memory_access_type == VMAT_INVARIANT) |
2936 | return size_zero_node; |
2937 | |
2938 | loop_vec_info loop_vinfo = dyn_cast<loop_vec_info> (p: vinfo); |
2939 | if (loop_vinfo && LOOP_VINFO_USING_SELECT_VL_P (loop_vinfo)) |
2940 | return vect_get_loop_variant_data_ptr_increment (vinfo, aggr_type, gsi, |
2941 | loop_lens, dr_info, |
2942 | memory_access_type); |
2943 | |
2944 | tree iv_step = TYPE_SIZE_UNIT (aggr_type); |
2945 | tree step = vect_dr_behavior (vinfo, dr_info)->step; |
2946 | if (tree_int_cst_sgn (step) == -1) |
2947 | iv_step = fold_build1 (NEGATE_EXPR, TREE_TYPE (iv_step), iv_step); |
2948 | return iv_step; |
2949 | } |
2950 | |
2951 | /* Check and perform vectorization of BUILT_IN_BSWAP{16,32,64,128}. */ |
2952 | |
2953 | static bool |
2954 | vectorizable_bswap (vec_info *vinfo, |
2955 | stmt_vec_info stmt_info, gimple_stmt_iterator *gsi, |
2956 | gimple **vec_stmt, slp_tree slp_node, |
2957 | slp_tree *slp_op, |
2958 | tree vectype_in, stmt_vector_for_cost *cost_vec) |
2959 | { |
2960 | tree op, vectype; |
2961 | gcall *stmt = as_a <gcall *> (p: stmt_info->stmt); |
2962 | loop_vec_info loop_vinfo = dyn_cast <loop_vec_info> (p: vinfo); |
2963 | unsigned ncopies; |
2964 | |
2965 | op = gimple_call_arg (gs: stmt, index: 0); |
2966 | vectype = STMT_VINFO_VECTYPE (stmt_info); |
2967 | poly_uint64 nunits = TYPE_VECTOR_SUBPARTS (node: vectype); |
2968 | |
2969 | /* Multiple types in SLP are handled by creating the appropriate number of |
2970 | vectorized stmts for each SLP node. Hence, NCOPIES is always 1 in |
2971 | case of SLP. */ |
2972 | if (slp_node) |
2973 | ncopies = 1; |
2974 | else |
2975 | ncopies = vect_get_num_copies (loop_vinfo, vectype); |
2976 | |
2977 | gcc_assert (ncopies >= 1); |
2978 | |
2979 | tree char_vectype = get_same_sized_vectype (char_type_node, vectype_in); |
2980 | if (! char_vectype) |
2981 | return false; |
2982 | |
2983 | poly_uint64 num_bytes = TYPE_VECTOR_SUBPARTS (node: char_vectype); |
2984 | unsigned word_bytes; |
2985 | if (!constant_multiple_p (a: num_bytes, b: nunits, multiple: &word_bytes)) |
2986 | return false; |
2987 | |
2988 | /* The encoding uses one stepped pattern for each byte in the word. */ |
2989 | vec_perm_builder elts (num_bytes, word_bytes, 3); |
2990 | for (unsigned i = 0; i < 3; ++i) |
2991 | for (unsigned j = 0; j < word_bytes; ++j) |
2992 | elts.quick_push (obj: (i + 1) * word_bytes - j - 1); |
2993 | |
2994 | vec_perm_indices indices (elts, 1, num_bytes); |
2995 | machine_mode vmode = TYPE_MODE (char_vectype); |
2996 | if (!can_vec_perm_const_p (vmode, vmode, indices)) |
2997 | return false; |
2998 | |
2999 | if (! vec_stmt) |
3000 | { |
3001 | if (slp_node |
3002 | && !vect_maybe_update_slp_op_vectype (slp_op[0], vectype_in)) |
3003 | { |
3004 | if (dump_enabled_p ()) |
3005 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
3006 | "incompatible vector types for invariants\n" ); |
3007 | return false; |
3008 | } |
3009 | |
3010 | STMT_VINFO_TYPE (stmt_info) = call_vec_info_type; |
3011 | DUMP_VECT_SCOPE ("vectorizable_bswap" ); |
3012 | record_stmt_cost (body_cost_vec: cost_vec, |
3013 | count: 1, kind: vector_stmt, stmt_info, misalign: 0, where: vect_prologue); |
3014 | record_stmt_cost (body_cost_vec: cost_vec, |
3015 | count: slp_node |
3016 | ? SLP_TREE_NUMBER_OF_VEC_STMTS (slp_node) : ncopies, |
3017 | kind: vec_perm, stmt_info, misalign: 0, where: vect_body); |
3018 | return true; |
3019 | } |
3020 | |
3021 | tree bswap_vconst = vec_perm_indices_to_tree (char_vectype, indices); |
3022 | |
3023 | /* Transform. */ |
3024 | vec<tree> vec_oprnds = vNULL; |
3025 | vect_get_vec_defs (vinfo, stmt_info, slp_node, ncopies, |
3026 | op0: op, vec_oprnds0: &vec_oprnds); |
3027 | /* Arguments are ready. create the new vector stmt. */ |
3028 | unsigned i; |
3029 | tree vop; |
3030 | FOR_EACH_VEC_ELT (vec_oprnds, i, vop) |
3031 | { |
3032 | gimple *new_stmt; |
3033 | tree tem = make_ssa_name (var: char_vectype); |
3034 | new_stmt = gimple_build_assign (tem, build1 (VIEW_CONVERT_EXPR, |
3035 | char_vectype, vop)); |
3036 | vect_finish_stmt_generation (vinfo, stmt_info, vec_stmt: new_stmt, gsi); |
3037 | tree tem2 = make_ssa_name (var: char_vectype); |
3038 | new_stmt = gimple_build_assign (tem2, VEC_PERM_EXPR, |
3039 | tem, tem, bswap_vconst); |
3040 | vect_finish_stmt_generation (vinfo, stmt_info, vec_stmt: new_stmt, gsi); |
3041 | tem = make_ssa_name (var: vectype); |
3042 | new_stmt = gimple_build_assign (tem, build1 (VIEW_CONVERT_EXPR, |
3043 | vectype, tem2)); |
3044 | vect_finish_stmt_generation (vinfo, stmt_info, vec_stmt: new_stmt, gsi); |
3045 | if (slp_node) |
3046 | slp_node->push_vec_def (def: new_stmt); |
3047 | else |
3048 | STMT_VINFO_VEC_STMTS (stmt_info).safe_push (obj: new_stmt); |
3049 | } |
3050 | |
3051 | if (!slp_node) |
3052 | *vec_stmt = STMT_VINFO_VEC_STMTS (stmt_info)[0]; |
3053 | |
3054 | vec_oprnds.release (); |
3055 | return true; |
3056 | } |
3057 | |
3058 | /* Return true if vector types VECTYPE_IN and VECTYPE_OUT have |
3059 | integer elements and if we can narrow VECTYPE_IN to VECTYPE_OUT |
3060 | in a single step. On success, store the binary pack code in |
3061 | *CONVERT_CODE. */ |
3062 | |
3063 | static bool |
3064 | simple_integer_narrowing (tree vectype_out, tree vectype_in, |
3065 | code_helper *convert_code) |
3066 | { |
3067 | if (!INTEGRAL_TYPE_P (TREE_TYPE (vectype_out)) |
3068 | || !INTEGRAL_TYPE_P (TREE_TYPE (vectype_in))) |
3069 | return false; |
3070 | |
3071 | code_helper code; |
3072 | int multi_step_cvt = 0; |
3073 | auto_vec <tree, 8> interm_types; |
3074 | if (!supportable_narrowing_operation (NOP_EXPR, vectype_out, vectype_in, |
3075 | &code, &multi_step_cvt, &interm_types) |
3076 | || multi_step_cvt) |
3077 | return false; |
3078 | |
3079 | *convert_code = code; |
3080 | return true; |
3081 | } |
3082 | |
3083 | /* Function vectorizable_call. |
3084 | |
3085 | Check if STMT_INFO performs a function call that can be vectorized. |
3086 | If VEC_STMT is also passed, vectorize STMT_INFO: create a vectorized |
3087 | stmt to replace it, put it in VEC_STMT, and insert it at GSI. |
3088 | Return true if STMT_INFO is vectorizable in this way. */ |
3089 | |
3090 | static bool |
3091 | vectorizable_call (vec_info *vinfo, |
3092 | stmt_vec_info stmt_info, gimple_stmt_iterator *gsi, |
3093 | gimple **vec_stmt, slp_tree slp_node, |
3094 | stmt_vector_for_cost *cost_vec) |
3095 | { |
3096 | gcall *stmt; |
3097 | tree vec_dest; |
3098 | tree scalar_dest; |
3099 | tree op; |
3100 | tree vec_oprnd0 = NULL_TREE, vec_oprnd1 = NULL_TREE; |
3101 | tree vectype_out, vectype_in; |
3102 | poly_uint64 nunits_in; |
3103 | poly_uint64 nunits_out; |
3104 | loop_vec_info loop_vinfo = dyn_cast <loop_vec_info> (p: vinfo); |
3105 | bb_vec_info bb_vinfo = dyn_cast <bb_vec_info> (p: vinfo); |
3106 | tree fndecl, new_temp, rhs_type; |
3107 | enum vect_def_type dt[4] |
3108 | = { vect_unknown_def_type, vect_unknown_def_type, vect_unknown_def_type, |
3109 | vect_unknown_def_type }; |
3110 | tree vectypes[ARRAY_SIZE (dt)] = {}; |
3111 | slp_tree slp_op[ARRAY_SIZE (dt)] = {}; |
3112 | int ndts = ARRAY_SIZE (dt); |
3113 | int ncopies, j; |
3114 | auto_vec<tree, 8> vargs; |
3115 | enum { NARROW, NONE, WIDEN } modifier; |
3116 | size_t i, nargs; |
3117 | tree lhs; |
3118 | |
3119 | if (!STMT_VINFO_RELEVANT_P (stmt_info) && !bb_vinfo) |
3120 | return false; |
3121 | |
3122 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def |
3123 | && ! vec_stmt) |
3124 | return false; |
3125 | |
3126 | /* Is STMT_INFO a vectorizable call? */ |
3127 | stmt = dyn_cast <gcall *> (p: stmt_info->stmt); |
3128 | if (!stmt) |
3129 | return false; |
3130 | |
3131 | if (gimple_call_internal_p (gs: stmt) |
3132 | && (internal_load_fn_p (gimple_call_internal_fn (gs: stmt)) |
3133 | || internal_store_fn_p (gimple_call_internal_fn (gs: stmt)))) |
3134 | /* Handled by vectorizable_load and vectorizable_store. */ |
3135 | return false; |
3136 | |
3137 | if (gimple_call_lhs (gs: stmt) == NULL_TREE |
3138 | || TREE_CODE (gimple_call_lhs (stmt)) != SSA_NAME) |
3139 | return false; |
3140 | |
3141 | gcc_checking_assert (!stmt_can_throw_internal (cfun, stmt)); |
3142 | |
3143 | vectype_out = STMT_VINFO_VECTYPE (stmt_info); |
3144 | |
3145 | /* Process function arguments. */ |
3146 | rhs_type = NULL_TREE; |
3147 | vectype_in = NULL_TREE; |
3148 | nargs = gimple_call_num_args (gs: stmt); |
3149 | |
3150 | /* Bail out if the function has more than four arguments, we do not have |
3151 | interesting builtin functions to vectorize with more than two arguments |
3152 | except for fma. No arguments is also not good. */ |
3153 | if (nargs == 0 || nargs > 4) |
3154 | return false; |
3155 | |
3156 | /* Ignore the arguments of IFN_GOMP_SIMD_LANE, they are magic. */ |
3157 | combined_fn cfn = gimple_call_combined_fn (stmt); |
3158 | if (cfn == CFN_GOMP_SIMD_LANE) |
3159 | { |
3160 | nargs = 0; |
3161 | rhs_type = unsigned_type_node; |
3162 | } |
3163 | |
3164 | int mask_opno = -1; |
3165 | if (internal_fn_p (code: cfn)) |
3166 | mask_opno = internal_fn_mask_index (as_internal_fn (code: cfn)); |
3167 | |
3168 | for (i = 0; i < nargs; i++) |
3169 | { |
3170 | if ((int) i == mask_opno) |
3171 | { |
3172 | if (!vect_check_scalar_mask (vinfo, stmt_info, slp_node, mask_index: mask_opno, |
3173 | mask: &op, mask_node: &slp_op[i], mask_dt_out: &dt[i], mask_vectype_out: &vectypes[i])) |
3174 | return false; |
3175 | continue; |
3176 | } |
3177 | |
3178 | if (!vect_is_simple_use (vinfo, stmt_info, slp_node, |
3179 | i, &op, &slp_op[i], &dt[i], &vectypes[i])) |
3180 | { |
3181 | if (dump_enabled_p ()) |
3182 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
3183 | "use not simple.\n" ); |
3184 | return false; |
3185 | } |
3186 | |
3187 | /* We can only handle calls with arguments of the same type. */ |
3188 | if (rhs_type |
3189 | && !types_compatible_p (type1: rhs_type, TREE_TYPE (op))) |
3190 | { |
3191 | if (dump_enabled_p ()) |
3192 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
3193 | "argument types differ.\n" ); |
3194 | return false; |
3195 | } |
3196 | if (!rhs_type) |
3197 | rhs_type = TREE_TYPE (op); |
3198 | |
3199 | if (!vectype_in) |
3200 | vectype_in = vectypes[i]; |
3201 | else if (vectypes[i] |
3202 | && !types_compatible_p (type1: vectypes[i], type2: vectype_in)) |
3203 | { |
3204 | if (dump_enabled_p ()) |
3205 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
3206 | "argument vector types differ.\n" ); |
3207 | return false; |
3208 | } |
3209 | } |
3210 | /* If all arguments are external or constant defs, infer the vector type |
3211 | from the scalar type. */ |
3212 | if (!vectype_in) |
3213 | vectype_in = get_vectype_for_scalar_type (vinfo, rhs_type, slp_node); |
3214 | if (vec_stmt) |
3215 | gcc_assert (vectype_in); |
3216 | if (!vectype_in) |
3217 | { |
3218 | if (dump_enabled_p ()) |
3219 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
3220 | "no vectype for scalar type %T\n" , rhs_type); |
3221 | |
3222 | return false; |
3223 | } |
3224 | |
3225 | if (VECTOR_BOOLEAN_TYPE_P (vectype_out) |
3226 | != VECTOR_BOOLEAN_TYPE_P (vectype_in)) |
3227 | { |
3228 | if (dump_enabled_p ()) |
3229 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
3230 | "mixed mask and nonmask vector types\n" ); |
3231 | return false; |
3232 | } |
3233 | |
3234 | if (vect_emulated_vector_p (vectype_in) || vect_emulated_vector_p (vectype_out)) |
3235 | { |
3236 | if (dump_enabled_p ()) |
3237 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
3238 | "use emulated vector type for call\n" ); |
3239 | return false; |
3240 | } |
3241 | |
3242 | /* FORNOW */ |
3243 | nunits_in = TYPE_VECTOR_SUBPARTS (node: vectype_in); |
3244 | nunits_out = TYPE_VECTOR_SUBPARTS (node: vectype_out); |
3245 | if (known_eq (nunits_in * 2, nunits_out)) |
3246 | modifier = NARROW; |
3247 | else if (known_eq (nunits_out, nunits_in)) |
3248 | modifier = NONE; |
3249 | else if (known_eq (nunits_out * 2, nunits_in)) |
3250 | modifier = WIDEN; |
3251 | else |
3252 | return false; |
3253 | |
3254 | /* We only handle functions that do not read or clobber memory. */ |
3255 | if (gimple_vuse (g: stmt)) |
3256 | { |
3257 | if (dump_enabled_p ()) |
3258 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
3259 | "function reads from or writes to memory.\n" ); |
3260 | return false; |
3261 | } |
3262 | |
3263 | /* For now, we only vectorize functions if a target specific builtin |
3264 | is available. TODO -- in some cases, it might be profitable to |
3265 | insert the calls for pieces of the vector, in order to be able |
3266 | to vectorize other operations in the loop. */ |
3267 | fndecl = NULL_TREE; |
3268 | internal_fn ifn = IFN_LAST; |
3269 | tree callee = gimple_call_fndecl (gs: stmt); |
3270 | |
3271 | /* First try using an internal function. */ |
3272 | code_helper convert_code = MAX_TREE_CODES; |
3273 | if (cfn != CFN_LAST |
3274 | && (modifier == NONE |
3275 | || (modifier == NARROW |
3276 | && simple_integer_narrowing (vectype_out, vectype_in, |
3277 | convert_code: &convert_code)))) |
3278 | ifn = vectorizable_internal_function (cfn, fndecl: callee, vectype_out, |
3279 | vectype_in); |
3280 | |
3281 | /* If that fails, try asking for a target-specific built-in function. */ |
3282 | if (ifn == IFN_LAST) |
3283 | { |
3284 | if (cfn != CFN_LAST) |
3285 | fndecl = targetm.vectorize.builtin_vectorized_function |
3286 | (cfn, vectype_out, vectype_in); |
3287 | else if (callee && fndecl_built_in_p (node: callee, klass: BUILT_IN_MD)) |
3288 | fndecl = targetm.vectorize.builtin_md_vectorized_function |
3289 | (callee, vectype_out, vectype_in); |
3290 | } |
3291 | |
3292 | if (ifn == IFN_LAST && !fndecl) |
3293 | { |
3294 | if (cfn == CFN_GOMP_SIMD_LANE |
3295 | && !slp_node |
3296 | && loop_vinfo |
3297 | && LOOP_VINFO_LOOP (loop_vinfo)->simduid |
3298 | && TREE_CODE (gimple_call_arg (stmt, 0)) == SSA_NAME |
3299 | && LOOP_VINFO_LOOP (loop_vinfo)->simduid |
3300 | == SSA_NAME_VAR (gimple_call_arg (stmt, 0))) |
3301 | { |
3302 | /* We can handle IFN_GOMP_SIMD_LANE by returning a |
3303 | { 0, 1, 2, ... vf - 1 } vector. */ |
3304 | gcc_assert (nargs == 0); |
3305 | } |
3306 | else if (modifier == NONE |
3307 | && (gimple_call_builtin_p (stmt, BUILT_IN_BSWAP16) |
3308 | || gimple_call_builtin_p (stmt, BUILT_IN_BSWAP32) |
3309 | || gimple_call_builtin_p (stmt, BUILT_IN_BSWAP64) |
3310 | || gimple_call_builtin_p (stmt, BUILT_IN_BSWAP128))) |
3311 | return vectorizable_bswap (vinfo, stmt_info, gsi, vec_stmt, slp_node, |
3312 | slp_op, vectype_in, cost_vec); |
3313 | else |
3314 | { |
3315 | if (dump_enabled_p ()) |
3316 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
3317 | "function is not vectorizable.\n" ); |
3318 | return false; |
3319 | } |
3320 | } |
3321 | |
3322 | if (slp_node) |
3323 | ncopies = 1; |
3324 | else if (modifier == NARROW && ifn == IFN_LAST) |
3325 | ncopies = vect_get_num_copies (loop_vinfo, vectype: vectype_out); |
3326 | else |
3327 | ncopies = vect_get_num_copies (loop_vinfo, vectype: vectype_in); |
3328 | |
3329 | /* Sanity check: make sure that at least one copy of the vectorized stmt |
3330 | needs to be generated. */ |
3331 | gcc_assert (ncopies >= 1); |
3332 | |
3333 | int reduc_idx = STMT_VINFO_REDUC_IDX (stmt_info); |
3334 | internal_fn cond_fn = get_conditional_internal_fn (ifn); |
3335 | internal_fn cond_len_fn = get_len_internal_fn (ifn); |
3336 | int len_opno = internal_fn_len_index (cond_len_fn); |
3337 | vec_loop_masks *masks = (loop_vinfo ? &LOOP_VINFO_MASKS (loop_vinfo) : NULL); |
3338 | vec_loop_lens *lens = (loop_vinfo ? &LOOP_VINFO_LENS (loop_vinfo) : NULL); |
3339 | if (!vec_stmt) /* transformation not required. */ |
3340 | { |
3341 | if (slp_node) |
3342 | for (i = 0; i < nargs; ++i) |
3343 | if (!vect_maybe_update_slp_op_vectype (slp_op[i], |
3344 | vectypes[i] |
3345 | ? vectypes[i] : vectype_in)) |
3346 | { |
3347 | if (dump_enabled_p ()) |
3348 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
3349 | "incompatible vector types for invariants\n" ); |
3350 | return false; |
3351 | } |
3352 | STMT_VINFO_TYPE (stmt_info) = call_vec_info_type; |
3353 | DUMP_VECT_SCOPE ("vectorizable_call" ); |
3354 | vect_model_simple_cost (vinfo, stmt_info, |
3355 | ncopies, dt, ndts, node: slp_node, cost_vec); |
3356 | if (ifn != IFN_LAST && modifier == NARROW && !slp_node) |
3357 | record_stmt_cost (body_cost_vec: cost_vec, count: ncopies / 2, |
3358 | kind: vec_promote_demote, stmt_info, misalign: 0, where: vect_body); |
3359 | |
3360 | if (loop_vinfo |
3361 | && LOOP_VINFO_CAN_USE_PARTIAL_VECTORS_P (loop_vinfo) |
3362 | && (reduc_idx >= 0 || mask_opno >= 0)) |
3363 | { |
3364 | if (reduc_idx >= 0 |
3365 | && (cond_fn == IFN_LAST |
3366 | || !direct_internal_fn_supported_p (cond_fn, vectype_out, |
3367 | OPTIMIZE_FOR_SPEED)) |
3368 | && (cond_len_fn == IFN_LAST |
3369 | || !direct_internal_fn_supported_p (cond_len_fn, vectype_out, |
3370 | OPTIMIZE_FOR_SPEED))) |
3371 | { |
3372 | if (dump_enabled_p ()) |
3373 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
3374 | "can't use a fully-masked loop because no" |
3375 | " conditional operation is available.\n" ); |
3376 | LOOP_VINFO_CAN_USE_PARTIAL_VECTORS_P (loop_vinfo) = false; |
3377 | } |
3378 | else |
3379 | { |
3380 | unsigned int nvectors |
3381 | = (slp_node |
3382 | ? SLP_TREE_NUMBER_OF_VEC_STMTS (slp_node) |
3383 | : ncopies); |
3384 | tree scalar_mask = NULL_TREE; |
3385 | if (mask_opno >= 0) |
3386 | scalar_mask = gimple_call_arg (gs: stmt_info->stmt, index: mask_opno); |
3387 | if (cond_len_fn != IFN_LAST |
3388 | && direct_internal_fn_supported_p (cond_len_fn, vectype_out, |
3389 | OPTIMIZE_FOR_SPEED)) |
3390 | vect_record_loop_len (loop_vinfo, lens, nvectors, vectype_out, |
3391 | 1); |
3392 | else |
3393 | vect_record_loop_mask (loop_vinfo, masks, nvectors, vectype_out, |
3394 | scalar_mask); |
3395 | } |
3396 | } |
3397 | return true; |
3398 | } |
3399 | |
3400 | /* Transform. */ |
3401 | |
3402 | if (dump_enabled_p ()) |
3403 | dump_printf_loc (MSG_NOTE, vect_location, "transform call.\n" ); |
3404 | |
3405 | /* Handle def. */ |
3406 | scalar_dest = gimple_call_lhs (gs: stmt); |
3407 | vec_dest = vect_create_destination_var (scalar_dest, vectype_out); |
3408 | |
3409 | bool masked_loop_p = loop_vinfo && LOOP_VINFO_FULLY_MASKED_P (loop_vinfo); |
3410 | bool len_loop_p = loop_vinfo && LOOP_VINFO_FULLY_WITH_LENGTH_P (loop_vinfo); |
3411 | unsigned int vect_nargs = nargs; |
3412 | if (len_loop_p) |
3413 | { |
3414 | if (len_opno >= 0) |
3415 | { |
3416 | ifn = cond_len_fn; |
3417 | /* COND_* -> COND_LEN_* takes 2 extra arguments:LEN,BIAS. */ |
3418 | vect_nargs += 2; |
3419 | } |
3420 | else if (reduc_idx >= 0) |
3421 | gcc_unreachable (); |
3422 | } |
3423 | else if (masked_loop_p && reduc_idx >= 0) |
3424 | { |
3425 | ifn = cond_fn; |
3426 | vect_nargs += 2; |
3427 | } |
3428 | |
3429 | if (modifier == NONE || ifn != IFN_LAST) |
3430 | { |
3431 | tree prev_res = NULL_TREE; |
3432 | vargs.safe_grow (len: vect_nargs, exact: true); |
3433 | auto_vec<vec<tree> > vec_defs (nargs); |
3434 | for (j = 0; j < ncopies; ++j) |
3435 | { |
3436 | /* Build argument list for the vectorized call. */ |
3437 | if (slp_node) |
3438 | { |
3439 | vec<tree> vec_oprnds0; |
3440 | |
3441 | vect_get_slp_defs (vinfo, slp_node, &vec_defs); |
3442 | vec_oprnds0 = vec_defs[0]; |
3443 | |
3444 | /* Arguments are ready. Create the new vector stmt. */ |
3445 | FOR_EACH_VEC_ELT (vec_oprnds0, i, vec_oprnd0) |
3446 | { |
3447 | int varg = 0; |
3448 | if (masked_loop_p && reduc_idx >= 0) |
3449 | { |
3450 | unsigned int vec_num = vec_oprnds0.length (); |
3451 | /* Always true for SLP. */ |
3452 | gcc_assert (ncopies == 1); |
3453 | vargs[varg++] = vect_get_loop_mask (loop_vinfo, |
3454 | gsi, masks, vec_num, |
3455 | vectype_out, i); |
3456 | } |
3457 | size_t k; |
3458 | for (k = 0; k < nargs; k++) |
3459 | { |
3460 | vec<tree> vec_oprndsk = vec_defs[k]; |
3461 | vargs[varg++] = vec_oprndsk[i]; |
3462 | } |
3463 | if (masked_loop_p && reduc_idx >= 0) |
3464 | vargs[varg++] = vargs[reduc_idx + 1]; |
3465 | gimple *new_stmt; |
3466 | if (modifier == NARROW) |
3467 | { |
3468 | /* We don't define any narrowing conditional functions |
3469 | at present. */ |
3470 | gcc_assert (mask_opno < 0); |
3471 | tree half_res = make_ssa_name (var: vectype_in); |
3472 | gcall *call |
3473 | = gimple_build_call_internal_vec (ifn, vargs); |
3474 | gimple_call_set_lhs (gs: call, lhs: half_res); |
3475 | gimple_call_set_nothrow (s: call, nothrow_p: true); |
3476 | vect_finish_stmt_generation (vinfo, stmt_info, vec_stmt: call, gsi); |
3477 | if ((i & 1) == 0) |
3478 | { |
3479 | prev_res = half_res; |
3480 | continue; |
3481 | } |
3482 | new_temp = make_ssa_name (var: vec_dest); |
3483 | new_stmt = vect_gimple_build (new_temp, convert_code, |
3484 | prev_res, half_res); |
3485 | vect_finish_stmt_generation (vinfo, stmt_info, |
3486 | vec_stmt: new_stmt, gsi); |
3487 | } |
3488 | else |
3489 | { |
3490 | if (len_opno >= 0 && len_loop_p) |
3491 | { |
3492 | unsigned int vec_num = vec_oprnds0.length (); |
3493 | /* Always true for SLP. */ |
3494 | gcc_assert (ncopies == 1); |
3495 | tree len |
3496 | = vect_get_loop_len (loop_vinfo, gsi, lens, vec_num, |
3497 | vectype_out, i, 1); |
3498 | signed char biasval |
3499 | = LOOP_VINFO_PARTIAL_LOAD_STORE_BIAS (loop_vinfo); |
3500 | tree bias = build_int_cst (intQI_type_node, biasval); |
3501 | vargs[len_opno] = len; |
3502 | vargs[len_opno + 1] = bias; |
3503 | } |
3504 | else if (mask_opno >= 0 && masked_loop_p) |
3505 | { |
3506 | unsigned int vec_num = vec_oprnds0.length (); |
3507 | /* Always true for SLP. */ |
3508 | gcc_assert (ncopies == 1); |
3509 | tree mask = vect_get_loop_mask (loop_vinfo, |
3510 | gsi, masks, vec_num, |
3511 | vectype_out, i); |
3512 | vargs[mask_opno] = prepare_vec_mask |
3513 | (loop_vinfo, TREE_TYPE (mask), loop_mask: mask, |
3514 | vec_mask: vargs[mask_opno], gsi); |
3515 | } |
3516 | |
3517 | gcall *call; |
3518 | if (ifn != IFN_LAST) |
3519 | call = gimple_build_call_internal_vec (ifn, vargs); |
3520 | else |
3521 | call = gimple_build_call_vec (fndecl, vargs); |
3522 | new_temp = make_ssa_name (var: vec_dest, stmt: call); |
3523 | gimple_call_set_lhs (gs: call, lhs: new_temp); |
3524 | gimple_call_set_nothrow (s: call, nothrow_p: true); |
3525 | vect_finish_stmt_generation (vinfo, stmt_info, vec_stmt: call, gsi); |
3526 | new_stmt = call; |
3527 | } |
3528 | slp_node->push_vec_def (def: new_stmt); |
3529 | } |
3530 | continue; |
3531 | } |
3532 | |
3533 | int varg = 0; |
3534 | if (masked_loop_p && reduc_idx >= 0) |
3535 | vargs[varg++] = vect_get_loop_mask (loop_vinfo, gsi, masks, ncopies, |
3536 | vectype_out, j); |
3537 | for (i = 0; i < nargs; i++) |
3538 | { |
3539 | op = gimple_call_arg (gs: stmt, index: i); |
3540 | if (j == 0) |
3541 | { |
3542 | vec_defs.quick_push (obj: vNULL); |
3543 | vect_get_vec_defs_for_operand (vinfo, stmt_vinfo: stmt_info, ncopies, |
3544 | op, vec_oprnds: &vec_defs[i], |
3545 | vectype: vectypes[i]); |
3546 | } |
3547 | vargs[varg++] = vec_defs[i][j]; |
3548 | } |
3549 | if (masked_loop_p && reduc_idx >= 0) |
3550 | vargs[varg++] = vargs[reduc_idx + 1]; |
3551 | |
3552 | if (len_opno >= 0 && len_loop_p) |
3553 | { |
3554 | tree len = vect_get_loop_len (loop_vinfo, gsi, lens, ncopies, |
3555 | vectype_out, j, 1); |
3556 | signed char biasval |
3557 | = LOOP_VINFO_PARTIAL_LOAD_STORE_BIAS (loop_vinfo); |
3558 | tree bias = build_int_cst (intQI_type_node, biasval); |
3559 | vargs[len_opno] = len; |
3560 | vargs[len_opno + 1] = bias; |
3561 | } |
3562 | else if (mask_opno >= 0 && masked_loop_p) |
3563 | { |
3564 | tree mask = vect_get_loop_mask (loop_vinfo, gsi, masks, ncopies, |
3565 | vectype_out, j); |
3566 | vargs[mask_opno] |
3567 | = prepare_vec_mask (loop_vinfo, TREE_TYPE (mask), loop_mask: mask, |
3568 | vec_mask: vargs[mask_opno], gsi); |
3569 | } |
3570 | |
3571 | gimple *new_stmt; |
3572 | if (cfn == CFN_GOMP_SIMD_LANE) |
3573 | { |
3574 | tree cst = build_index_vector (vectype_out, j * nunits_out, 1); |
3575 | tree new_var |
3576 | = vect_get_new_ssa_name (vectype_out, vect_simple_var, "cst_" ); |
3577 | gimple *init_stmt = gimple_build_assign (new_var, cst); |
3578 | vect_init_vector_1 (vinfo, stmt_vinfo: stmt_info, new_stmt: init_stmt, NULL); |
3579 | new_temp = make_ssa_name (var: vec_dest); |
3580 | new_stmt = gimple_build_assign (new_temp, new_var); |
3581 | vect_finish_stmt_generation (vinfo, stmt_info, vec_stmt: new_stmt, gsi); |
3582 | } |
3583 | else if (modifier == NARROW) |
3584 | { |
3585 | /* We don't define any narrowing conditional functions at |
3586 | present. */ |
3587 | gcc_assert (mask_opno < 0); |
3588 | tree half_res = make_ssa_name (var: vectype_in); |
3589 | gcall *call = gimple_build_call_internal_vec (ifn, vargs); |
3590 | gimple_call_set_lhs (gs: call, lhs: half_res); |
3591 | gimple_call_set_nothrow (s: call, nothrow_p: true); |
3592 | vect_finish_stmt_generation (vinfo, stmt_info, vec_stmt: call, gsi); |
3593 | if ((j & 1) == 0) |
3594 | { |
3595 | prev_res = half_res; |
3596 | continue; |
3597 | } |
3598 | new_temp = make_ssa_name (var: vec_dest); |
3599 | new_stmt = vect_gimple_build (new_temp, convert_code, prev_res, |
3600 | half_res); |
3601 | vect_finish_stmt_generation (vinfo, stmt_info, vec_stmt: new_stmt, gsi); |
3602 | } |
3603 | else |
3604 | { |
3605 | gcall *call; |
3606 | if (ifn != IFN_LAST) |
3607 | call = gimple_build_call_internal_vec (ifn, vargs); |
3608 | else |
3609 | call = gimple_build_call_vec (fndecl, vargs); |
3610 | new_temp = make_ssa_name (var: vec_dest, stmt: call); |
3611 | gimple_call_set_lhs (gs: call, lhs: new_temp); |
3612 | gimple_call_set_nothrow (s: call, nothrow_p: true); |
3613 | vect_finish_stmt_generation (vinfo, stmt_info, vec_stmt: call, gsi); |
3614 | new_stmt = call; |
3615 | } |
3616 | |
3617 | if (j == (modifier == NARROW ? 1 : 0)) |
3618 | *vec_stmt = new_stmt; |
3619 | STMT_VINFO_VEC_STMTS (stmt_info).safe_push (obj: new_stmt); |
3620 | } |
3621 | for (i = 0; i < nargs; i++) |
3622 | { |
3623 | vec<tree> vec_oprndsi = vec_defs[i]; |
3624 | vec_oprndsi.release (); |
3625 | } |
3626 | } |
3627 | else if (modifier == NARROW) |
3628 | { |
3629 | auto_vec<vec<tree> > vec_defs (nargs); |
3630 | /* We don't define any narrowing conditional functions at present. */ |
3631 | gcc_assert (mask_opno < 0); |
3632 | for (j = 0; j < ncopies; ++j) |
3633 | { |
3634 | /* Build argument list for the vectorized call. */ |
3635 | if (j == 0) |
3636 | vargs.create (nelems: nargs * 2); |
3637 | else |
3638 | vargs.truncate (size: 0); |
3639 | |
3640 | if (slp_node) |
3641 | { |
3642 | vec<tree> vec_oprnds0; |
3643 | |
3644 | vect_get_slp_defs (vinfo, slp_node, &vec_defs); |
3645 | vec_oprnds0 = vec_defs[0]; |
3646 | |
3647 | /* Arguments are ready. Create the new vector stmt. */ |
3648 | for (i = 0; vec_oprnds0.iterate (ix: i, ptr: &vec_oprnd0); i += 2) |
3649 | { |
3650 | size_t k; |
3651 | vargs.truncate (size: 0); |
3652 | for (k = 0; k < nargs; k++) |
3653 | { |
3654 | vec<tree> vec_oprndsk = vec_defs[k]; |
3655 | vargs.quick_push (obj: vec_oprndsk[i]); |
3656 | vargs.quick_push (obj: vec_oprndsk[i + 1]); |
3657 | } |
3658 | gcall *call; |
3659 | if (ifn != IFN_LAST) |
3660 | call = gimple_build_call_internal_vec (ifn, vargs); |
3661 | else |
3662 | call = gimple_build_call_vec (fndecl, vargs); |
3663 | new_temp = make_ssa_name (var: vec_dest, stmt: call); |
3664 | gimple_call_set_lhs (gs: call, lhs: new_temp); |
3665 | gimple_call_set_nothrow (s: call, nothrow_p: true); |
3666 | vect_finish_stmt_generation (vinfo, stmt_info, vec_stmt: call, gsi); |
3667 | slp_node->push_vec_def (def: call); |
3668 | } |
3669 | continue; |
3670 | } |
3671 | |
3672 | for (i = 0; i < nargs; i++) |
3673 | { |
3674 | op = gimple_call_arg (gs: stmt, index: i); |
3675 | if (j == 0) |
3676 | { |
3677 | vec_defs.quick_push (obj: vNULL); |
3678 | vect_get_vec_defs_for_operand (vinfo, stmt_vinfo: stmt_info, ncopies: 2 * ncopies, |
3679 | op, vec_oprnds: &vec_defs[i], vectype: vectypes[i]); |
3680 | } |
3681 | vec_oprnd0 = vec_defs[i][2*j]; |
3682 | vec_oprnd1 = vec_defs[i][2*j+1]; |
3683 | |
3684 | vargs.quick_push (obj: vec_oprnd0); |
3685 | vargs.quick_push (obj: vec_oprnd1); |
3686 | } |
3687 | |
3688 | gcall *new_stmt = gimple_build_call_vec (fndecl, vargs); |
3689 | new_temp = make_ssa_name (var: vec_dest, stmt: new_stmt); |
3690 | gimple_call_set_lhs (gs: new_stmt, lhs: new_temp); |
3691 | vect_finish_stmt_generation (vinfo, stmt_info, vec_stmt: new_stmt, gsi); |
3692 | |
3693 | STMT_VINFO_VEC_STMTS (stmt_info).safe_push (obj: new_stmt); |
3694 | } |
3695 | |
3696 | if (!slp_node) |
3697 | *vec_stmt = STMT_VINFO_VEC_STMTS (stmt_info)[0]; |
3698 | |
3699 | for (i = 0; i < nargs; i++) |
3700 | { |
3701 | vec<tree> vec_oprndsi = vec_defs[i]; |
3702 | vec_oprndsi.release (); |
3703 | } |
3704 | } |
3705 | else |
3706 | /* No current target implements this case. */ |
3707 | return false; |
3708 | |
3709 | vargs.release (); |
3710 | |
3711 | /* The call in STMT might prevent it from being removed in dce. |
3712 | We however cannot remove it here, due to the way the ssa name |
3713 | it defines is mapped to the new definition. So just replace |
3714 | rhs of the statement with something harmless. */ |
3715 | |
3716 | if (slp_node) |
3717 | return true; |
3718 | |
3719 | stmt_info = vect_orig_stmt (stmt_info); |
3720 | lhs = gimple_get_lhs (stmt_info->stmt); |
3721 | |
3722 | gassign *new_stmt |
3723 | = gimple_build_assign (lhs, build_zero_cst (TREE_TYPE (lhs))); |
3724 | vinfo->replace_stmt (gsi, stmt_info, new_stmt); |
3725 | |
3726 | return true; |
3727 | } |
3728 | |
3729 | |
3730 | struct simd_call_arg_info |
3731 | { |
3732 | tree vectype; |
3733 | tree op; |
3734 | HOST_WIDE_INT linear_step; |
3735 | enum vect_def_type dt; |
3736 | unsigned int align; |
3737 | bool simd_lane_linear; |
3738 | }; |
3739 | |
3740 | /* Helper function of vectorizable_simd_clone_call. If OP, an SSA_NAME, |
3741 | is linear within simd lane (but not within whole loop), note it in |
3742 | *ARGINFO. */ |
3743 | |
3744 | static void |
3745 | vect_simd_lane_linear (tree op, class loop *loop, |
3746 | struct simd_call_arg_info *arginfo) |
3747 | { |
3748 | gimple *def_stmt = SSA_NAME_DEF_STMT (op); |
3749 | |
3750 | if (!is_gimple_assign (gs: def_stmt) |
3751 | || gimple_assign_rhs_code (gs: def_stmt) != POINTER_PLUS_EXPR |
3752 | || !is_gimple_min_invariant (gimple_assign_rhs1 (gs: def_stmt))) |
3753 | return; |
3754 | |
3755 | tree base = gimple_assign_rhs1 (gs: def_stmt); |
3756 | HOST_WIDE_INT linear_step = 0; |
3757 | tree v = gimple_assign_rhs2 (gs: def_stmt); |
3758 | while (TREE_CODE (v) == SSA_NAME) |
3759 | { |
3760 | tree t; |
3761 | def_stmt = SSA_NAME_DEF_STMT (v); |
3762 | if (is_gimple_assign (gs: def_stmt)) |
3763 | switch (gimple_assign_rhs_code (gs: def_stmt)) |
3764 | { |
3765 | case PLUS_EXPR: |
3766 | t = gimple_assign_rhs2 (gs: def_stmt); |
3767 | if (linear_step || TREE_CODE (t) != INTEGER_CST) |
3768 | return; |
3769 | base = fold_build2 (POINTER_PLUS_EXPR, TREE_TYPE (base), base, t); |
3770 | v = gimple_assign_rhs1 (gs: def_stmt); |
3771 | continue; |
3772 | case MULT_EXPR: |
3773 | t = gimple_assign_rhs2 (gs: def_stmt); |
3774 | if (linear_step || !tree_fits_shwi_p (t) || integer_zerop (t)) |
3775 | return; |
3776 | linear_step = tree_to_shwi (t); |
3777 | v = gimple_assign_rhs1 (gs: def_stmt); |
3778 | continue; |
3779 | CASE_CONVERT: |
3780 | t = gimple_assign_rhs1 (gs: def_stmt); |
3781 | if (TREE_CODE (TREE_TYPE (t)) != INTEGER_TYPE |
3782 | || (TYPE_PRECISION (TREE_TYPE (v)) |
3783 | < TYPE_PRECISION (TREE_TYPE (t)))) |
3784 | return; |
3785 | if (!linear_step) |
3786 | linear_step = 1; |
3787 | v = t; |
3788 | continue; |
3789 | default: |
3790 | return; |
3791 | } |
3792 | else if (gimple_call_internal_p (gs: def_stmt, fn: IFN_GOMP_SIMD_LANE) |
3793 | && loop->simduid |
3794 | && TREE_CODE (gimple_call_arg (def_stmt, 0)) == SSA_NAME |
3795 | && (SSA_NAME_VAR (gimple_call_arg (def_stmt, 0)) |
3796 | == loop->simduid)) |
3797 | { |
3798 | if (!linear_step) |
3799 | linear_step = 1; |
3800 | arginfo->linear_step = linear_step; |
3801 | arginfo->op = base; |
3802 | arginfo->simd_lane_linear = true; |
3803 | return; |
3804 | } |
3805 | } |
3806 | } |
3807 | |
3808 | /* Function vectorizable_simd_clone_call. |
3809 | |
3810 | Check if STMT_INFO performs a function call that can be vectorized |
3811 | by calling a simd clone of the function. |
3812 | If VEC_STMT is also passed, vectorize STMT_INFO: create a vectorized |
3813 | stmt to replace it, put it in VEC_STMT, and insert it at GSI. |
3814 | Return true if STMT_INFO is vectorizable in this way. */ |
3815 | |
3816 | static bool |
3817 | vectorizable_simd_clone_call (vec_info *vinfo, stmt_vec_info stmt_info, |
3818 | gimple_stmt_iterator *gsi, |
3819 | gimple **vec_stmt, slp_tree slp_node, |
3820 | stmt_vector_for_cost *) |
3821 | { |
3822 | tree vec_dest; |
3823 | tree scalar_dest; |
3824 | tree op, type; |
3825 | tree vec_oprnd0 = NULL_TREE; |
3826 | tree vectype; |
3827 | poly_uint64 nunits; |
3828 | loop_vec_info loop_vinfo = dyn_cast <loop_vec_info> (p: vinfo); |
3829 | bb_vec_info bb_vinfo = dyn_cast <bb_vec_info> (p: vinfo); |
3830 | class loop *loop = loop_vinfo ? LOOP_VINFO_LOOP (loop_vinfo) : NULL; |
3831 | tree fndecl, new_temp; |
3832 | int ncopies, j; |
3833 | auto_vec<simd_call_arg_info> arginfo; |
3834 | vec<tree> vargs = vNULL; |
3835 | size_t i, nargs; |
3836 | tree lhs, rtype, ratype; |
3837 | vec<constructor_elt, va_gc> *ret_ctor_elts = NULL; |
3838 | int masked_call_offset = 0; |
3839 | |
3840 | /* Is STMT a vectorizable call? */ |
3841 | gcall *stmt = dyn_cast <gcall *> (p: stmt_info->stmt); |
3842 | if (!stmt) |
3843 | return false; |
3844 | |
3845 | fndecl = gimple_call_fndecl (gs: stmt); |
3846 | if (fndecl == NULL_TREE |
3847 | && gimple_call_internal_p (gs: stmt, fn: IFN_MASK_CALL)) |
3848 | { |
3849 | fndecl = gimple_call_arg (gs: stmt, index: 0); |
3850 | gcc_checking_assert (TREE_CODE (fndecl) == ADDR_EXPR); |
3851 | fndecl = TREE_OPERAND (fndecl, 0); |
3852 | gcc_checking_assert (TREE_CODE (fndecl) == FUNCTION_DECL); |
3853 | masked_call_offset = 1; |
3854 | } |
3855 | if (fndecl == NULL_TREE) |
3856 | return false; |
3857 | |
3858 | struct cgraph_node *node = cgraph_node::get (decl: fndecl); |
3859 | if (node == NULL || node->simd_clones == NULL) |
3860 | return false; |
3861 | |
3862 | if (!STMT_VINFO_RELEVANT_P (stmt_info) && !bb_vinfo) |
3863 | return false; |
3864 | |
3865 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def |
3866 | && ! vec_stmt) |
3867 | return false; |
3868 | |
3869 | if (gimple_call_lhs (gs: stmt) |
3870 | && TREE_CODE (gimple_call_lhs (stmt)) != SSA_NAME) |
3871 | return false; |
3872 | |
3873 | gcc_checking_assert (!stmt_can_throw_internal (cfun, stmt)); |
3874 | |
3875 | vectype = STMT_VINFO_VECTYPE (stmt_info); |
3876 | |
3877 | if (loop_vinfo && nested_in_vect_loop_p (loop, stmt_info)) |
3878 | return false; |
3879 | |
3880 | /* Process function arguments. */ |
3881 | nargs = gimple_call_num_args (gs: stmt) - masked_call_offset; |
3882 | |
3883 | /* Bail out if the function has zero arguments. */ |
3884 | if (nargs == 0) |
3885 | return false; |
3886 | |
3887 | vec<tree>& simd_clone_info = (slp_node ? SLP_TREE_SIMD_CLONE_INFO (slp_node) |
3888 | : STMT_VINFO_SIMD_CLONE_INFO (stmt_info)); |
3889 | arginfo.reserve (nelems: nargs, exact: true); |
3890 | auto_vec<slp_tree> slp_op; |
3891 | slp_op.safe_grow_cleared (len: nargs); |
3892 | |
3893 | for (i = 0; i < nargs; i++) |
3894 | { |
3895 | simd_call_arg_info thisarginfo; |
3896 | affine_iv iv; |
3897 | |
3898 | thisarginfo.linear_step = 0; |
3899 | thisarginfo.align = 0; |
3900 | thisarginfo.op = NULL_TREE; |
3901 | thisarginfo.simd_lane_linear = false; |
3902 | |
3903 | int op_no = i + masked_call_offset; |
3904 | if (slp_node) |
3905 | op_no = vect_slp_child_index_for_operand (stmt, op: op_no, false); |
3906 | if (!vect_is_simple_use (vinfo, stmt_info, slp_node, |
3907 | op_no, &op, &slp_op[i], |
3908 | &thisarginfo.dt, &thisarginfo.vectype) |
3909 | || thisarginfo.dt == vect_uninitialized_def) |
3910 | { |
3911 | if (dump_enabled_p ()) |
3912 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
3913 | "use not simple.\n" ); |
3914 | return false; |
3915 | } |
3916 | |
3917 | if (thisarginfo.dt == vect_constant_def |
3918 | || thisarginfo.dt == vect_external_def) |
3919 | { |
3920 | /* With SLP we determine the vector type of constants/externals |
3921 | at analysis time, handling conflicts via |
3922 | vect_maybe_update_slp_op_vectype. At transform time |
3923 | we have a vector type recorded for SLP. */ |
3924 | gcc_assert (!vec_stmt |
3925 | || !slp_node |
3926 | || thisarginfo.vectype != NULL_TREE); |
3927 | if (!vec_stmt) |
3928 | thisarginfo.vectype = get_vectype_for_scalar_type (vinfo, |
3929 | TREE_TYPE (op), |
3930 | slp_node); |
3931 | } |
3932 | else |
3933 | gcc_assert (thisarginfo.vectype != NULL_TREE); |
3934 | |
3935 | /* For linear arguments, the analyze phase should have saved |
3936 | the base and step in {STMT_VINFO,SLP_TREE}_SIMD_CLONE_INFO. */ |
3937 | if (i * 3 + 4 <= simd_clone_info.length () |
3938 | && simd_clone_info[i * 3 + 2]) |
3939 | { |
3940 | gcc_assert (vec_stmt); |
3941 | thisarginfo.linear_step = tree_to_shwi (simd_clone_info[i * 3 + 2]); |
3942 | thisarginfo.op = simd_clone_info[i * 3 + 1]; |
3943 | thisarginfo.simd_lane_linear |
3944 | = (simd_clone_info[i * 3 + 3] == boolean_true_node); |
3945 | /* If loop has been peeled for alignment, we need to adjust it. */ |
3946 | tree n1 = LOOP_VINFO_NITERS_UNCHANGED (loop_vinfo); |
3947 | tree n2 = LOOP_VINFO_NITERS (loop_vinfo); |
3948 | if (n1 != n2 && !thisarginfo.simd_lane_linear) |
3949 | { |
3950 | tree bias = fold_build2 (MINUS_EXPR, TREE_TYPE (n1), n1, n2); |
3951 | tree step = simd_clone_info[i * 3 + 2]; |
3952 | tree opt = TREE_TYPE (thisarginfo.op); |
3953 | bias = fold_convert (TREE_TYPE (step), bias); |
3954 | bias = fold_build2 (MULT_EXPR, TREE_TYPE (step), bias, step); |
3955 | thisarginfo.op |
3956 | = fold_build2 (POINTER_TYPE_P (opt) |
3957 | ? POINTER_PLUS_EXPR : PLUS_EXPR, opt, |
3958 | thisarginfo.op, bias); |
3959 | } |
3960 | } |
3961 | else if (!vec_stmt |
3962 | && thisarginfo.dt != vect_constant_def |
3963 | && thisarginfo.dt != vect_external_def |
3964 | && loop_vinfo |
3965 | && TREE_CODE (op) == SSA_NAME |
3966 | && simple_iv (loop, loop_containing_stmt (stmt), op, |
3967 | &iv, false) |
3968 | && tree_fits_shwi_p (iv.step)) |
3969 | { |
3970 | thisarginfo.linear_step = tree_to_shwi (iv.step); |
3971 | thisarginfo.op = iv.base; |
3972 | } |
3973 | else if ((thisarginfo.dt == vect_constant_def |
3974 | || thisarginfo.dt == vect_external_def) |
3975 | && POINTER_TYPE_P (TREE_TYPE (op))) |
3976 | thisarginfo.align = get_pointer_alignment (op) / BITS_PER_UNIT; |
3977 | /* Addresses of array elements indexed by GOMP_SIMD_LANE are |
3978 | linear too. */ |
3979 | if (POINTER_TYPE_P (TREE_TYPE (op)) |
3980 | && !thisarginfo.linear_step |
3981 | && !vec_stmt |
3982 | && thisarginfo.dt != vect_constant_def |
3983 | && thisarginfo.dt != vect_external_def |
3984 | && loop_vinfo |
3985 | && TREE_CODE (op) == SSA_NAME) |
3986 | vect_simd_lane_linear (op, loop, arginfo: &thisarginfo); |
3987 | |
3988 | arginfo.quick_push (obj: thisarginfo); |
3989 | } |
3990 | |
3991 | poly_uint64 vf = loop_vinfo ? LOOP_VINFO_VECT_FACTOR (loop_vinfo) : 1; |
3992 | unsigned group_size = slp_node ? SLP_TREE_LANES (slp_node) : 1; |
3993 | unsigned int badness = 0; |
3994 | struct cgraph_node *bestn = NULL; |
3995 | if (simd_clone_info.exists ()) |
3996 | bestn = cgraph_node::get (decl: simd_clone_info[0]); |
3997 | else |
3998 | for (struct cgraph_node *n = node->simd_clones; n != NULL; |
3999 | n = n->simdclone->next_clone) |
4000 | { |
4001 | unsigned int this_badness = 0; |
4002 | unsigned int num_calls; |
4003 | /* The number of arguments in the call and the number of parameters in |
4004 | the simdclone should match. However, when the simdclone is |
4005 | 'inbranch', it could have one more paramater than nargs when using |
4006 | an inbranch simdclone to call a non-inbranch call, either in a |
4007 | non-masked loop using a all true constant mask, or inside a masked |
4008 | loop using it's mask. */ |
4009 | size_t simd_nargs = n->simdclone->nargs; |
4010 | if (!masked_call_offset && n->simdclone->inbranch) |
4011 | simd_nargs--; |
4012 | if (!constant_multiple_p (a: vf * group_size, b: n->simdclone->simdlen, |
4013 | multiple: &num_calls) |
4014 | || (!n->simdclone->inbranch && (masked_call_offset > 0)) |
4015 | || (nargs != simd_nargs)) |
4016 | continue; |
4017 | if (num_calls != 1) |
4018 | this_badness += exact_log2 (x: num_calls) * 4096; |
4019 | if (n->simdclone->inbranch) |
4020 | this_badness += 8192; |
4021 | int target_badness = targetm.simd_clone.usable (n); |
4022 | if (target_badness < 0) |
4023 | continue; |
4024 | this_badness += target_badness * 512; |
4025 | for (i = 0; i < nargs; i++) |
4026 | { |
4027 | switch (n->simdclone->args[i].arg_type) |
4028 | { |
4029 | case SIMD_CLONE_ARG_TYPE_VECTOR: |
4030 | if (!useless_type_conversion_p |
4031 | (n->simdclone->args[i].orig_type, |
4032 | TREE_TYPE (gimple_call_arg (stmt, |
4033 | i + masked_call_offset)))) |
4034 | i = -1; |
4035 | else if (arginfo[i].dt == vect_constant_def |
4036 | || arginfo[i].dt == vect_external_def |
4037 | || arginfo[i].linear_step) |
4038 | this_badness += 64; |
4039 | break; |
4040 | case SIMD_CLONE_ARG_TYPE_UNIFORM: |
4041 | if (arginfo[i].dt != vect_constant_def |
4042 | && arginfo[i].dt != vect_external_def) |
4043 | i = -1; |
4044 | break; |
4045 | case SIMD_CLONE_ARG_TYPE_LINEAR_CONSTANT_STEP: |
4046 | case SIMD_CLONE_ARG_TYPE_LINEAR_REF_CONSTANT_STEP: |
4047 | if (arginfo[i].dt == vect_constant_def |
4048 | || arginfo[i].dt == vect_external_def |
4049 | || (arginfo[i].linear_step |
4050 | != n->simdclone->args[i].linear_step)) |
4051 | i = -1; |
4052 | break; |
4053 | case SIMD_CLONE_ARG_TYPE_LINEAR_VARIABLE_STEP: |
4054 | case SIMD_CLONE_ARG_TYPE_LINEAR_VAL_CONSTANT_STEP: |
4055 | case SIMD_CLONE_ARG_TYPE_LINEAR_UVAL_CONSTANT_STEP: |
4056 | case SIMD_CLONE_ARG_TYPE_LINEAR_REF_VARIABLE_STEP: |
4057 | case SIMD_CLONE_ARG_TYPE_LINEAR_VAL_VARIABLE_STEP: |
4058 | case SIMD_CLONE_ARG_TYPE_LINEAR_UVAL_VARIABLE_STEP: |
4059 | /* FORNOW */ |
4060 | i = -1; |
4061 | break; |
4062 | case SIMD_CLONE_ARG_TYPE_MASK: |
4063 | /* While we can create a traditional data vector from |
4064 | an incoming integer mode mask we have no good way to |
4065 | force generate an integer mode mask from a traditional |
4066 | boolean vector input. */ |
4067 | if (SCALAR_INT_MODE_P (n->simdclone->mask_mode) |
4068 | && !SCALAR_INT_MODE_P (TYPE_MODE (arginfo[i].vectype))) |
4069 | i = -1; |
4070 | else if (!SCALAR_INT_MODE_P (n->simdclone->mask_mode) |
4071 | && SCALAR_INT_MODE_P (TYPE_MODE (arginfo[i].vectype))) |
4072 | this_badness += 2048; |
4073 | break; |
4074 | } |
4075 | if (i == (size_t) -1) |
4076 | break; |
4077 | if (n->simdclone->args[i].alignment > arginfo[i].align) |
4078 | { |
4079 | i = -1; |
4080 | break; |
4081 | } |
4082 | if (arginfo[i].align) |
4083 | this_badness += (exact_log2 (x: arginfo[i].align) |
4084 | - exact_log2 (x: n->simdclone->args[i].alignment)); |
4085 | } |
4086 | if (i == (size_t) -1) |
4087 | continue; |
4088 | if (masked_call_offset == 0 |
4089 | && n->simdclone->inbranch |
4090 | && n->simdclone->nargs > nargs) |
4091 | { |
4092 | gcc_assert (n->simdclone->args[n->simdclone->nargs - 1].arg_type == |
4093 | SIMD_CLONE_ARG_TYPE_MASK); |
4094 | /* Penalize using a masked SIMD clone in a non-masked loop, that is |
4095 | not in a branch, as we'd have to construct an all-true mask. */ |
4096 | if (!loop_vinfo || !LOOP_VINFO_FULLY_MASKED_P (loop_vinfo)) |
4097 | this_badness += 64; |
4098 | } |
4099 | if (bestn == NULL || this_badness < badness) |
4100 | { |
4101 | bestn = n; |
4102 | badness = this_badness; |
4103 | } |
4104 | } |
4105 | |
4106 | if (bestn == NULL) |
4107 | return false; |
4108 | |
4109 | unsigned int num_mask_args = 0; |
4110 | if (SCALAR_INT_MODE_P (bestn->simdclone->mask_mode)) |
4111 | for (i = 0; i < nargs; i++) |
4112 | if (bestn->simdclone->args[i].arg_type == SIMD_CLONE_ARG_TYPE_MASK) |
4113 | num_mask_args++; |
4114 | |
4115 | for (i = 0; i < nargs; i++) |
4116 | { |
4117 | if ((arginfo[i].dt == vect_constant_def |
4118 | || arginfo[i].dt == vect_external_def) |
4119 | && bestn->simdclone->args[i].arg_type == SIMD_CLONE_ARG_TYPE_VECTOR) |
4120 | { |
4121 | tree arg_type = TREE_TYPE (gimple_call_arg (stmt, |
4122 | i + masked_call_offset)); |
4123 | arginfo[i].vectype = get_vectype_for_scalar_type (vinfo, arg_type, |
4124 | slp_node); |
4125 | if (arginfo[i].vectype == NULL |
4126 | || !constant_multiple_p (a: bestn->simdclone->simdlen, |
4127 | b: TYPE_VECTOR_SUBPARTS (node: arginfo[i].vectype))) |
4128 | return false; |
4129 | } |
4130 | |
4131 | if (bestn->simdclone->args[i].arg_type == SIMD_CLONE_ARG_TYPE_VECTOR |
4132 | && VECTOR_BOOLEAN_TYPE_P (bestn->simdclone->args[i].vector_type)) |
4133 | { |
4134 | if (dump_enabled_p ()) |
4135 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
4136 | "vector mask arguments are not supported.\n" ); |
4137 | return false; |
4138 | } |
4139 | |
4140 | if (bestn->simdclone->args[i].arg_type == SIMD_CLONE_ARG_TYPE_MASK) |
4141 | { |
4142 | tree clone_arg_vectype = bestn->simdclone->args[i].vector_type; |
4143 | if (bestn->simdclone->mask_mode == VOIDmode) |
4144 | { |
4145 | if (maybe_ne (a: TYPE_VECTOR_SUBPARTS (node: clone_arg_vectype), |
4146 | b: TYPE_VECTOR_SUBPARTS (node: arginfo[i].vectype))) |
4147 | { |
4148 | /* FORNOW we only have partial support for vector-type masks |
4149 | that can't hold all of simdlen. */ |
4150 | if (dump_enabled_p ()) |
4151 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, |
4152 | vect_location, |
4153 | "in-branch vector clones are not yet" |
4154 | " supported for mismatched vector sizes.\n" ); |
4155 | return false; |
4156 | } |
4157 | } |
4158 | else if (SCALAR_INT_MODE_P (bestn->simdclone->mask_mode)) |
4159 | { |
4160 | if (!SCALAR_INT_MODE_P (TYPE_MODE (arginfo[i].vectype)) |
4161 | || maybe_ne (a: exact_div (a: bestn->simdclone->simdlen, |
4162 | b: num_mask_args), |
4163 | b: TYPE_VECTOR_SUBPARTS (node: arginfo[i].vectype))) |
4164 | { |
4165 | /* FORNOW we only have partial support for integer-type masks |
4166 | that represent the same number of lanes as the |
4167 | vectorized mask inputs. */ |
4168 | if (dump_enabled_p ()) |
4169 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, |
4170 | vect_location, |
4171 | "in-branch vector clones are not yet " |
4172 | "supported for mismatched vector sizes.\n" ); |
4173 | return false; |
4174 | } |
4175 | } |
4176 | else |
4177 | { |
4178 | if (dump_enabled_p ()) |
4179 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, |
4180 | vect_location, |
4181 | "in-branch vector clones not supported" |
4182 | " on this target.\n" ); |
4183 | return false; |
4184 | } |
4185 | } |
4186 | } |
4187 | |
4188 | fndecl = bestn->decl; |
4189 | nunits = bestn->simdclone->simdlen; |
4190 | if (slp_node) |
4191 | ncopies = vector_unroll_factor (vf * group_size, nunits); |
4192 | else |
4193 | ncopies = vector_unroll_factor (vf, nunits); |
4194 | |
4195 | /* If the function isn't const, only allow it in simd loops where user |
4196 | has asserted that at least nunits consecutive iterations can be |
4197 | performed using SIMD instructions. */ |
4198 | if ((loop == NULL || maybe_lt (a: (unsigned) loop->safelen, b: nunits)) |
4199 | && gimple_vuse (g: stmt)) |
4200 | return false; |
4201 | |
4202 | /* Sanity check: make sure that at least one copy of the vectorized stmt |
4203 | needs to be generated. */ |
4204 | gcc_assert (ncopies >= 1); |
4205 | |
4206 | if (!vec_stmt) /* transformation not required. */ |
4207 | { |
4208 | if (slp_node) |
4209 | for (unsigned i = 0; i < nargs; ++i) |
4210 | if (!vect_maybe_update_slp_op_vectype (slp_op[i], arginfo[i].vectype)) |
4211 | { |
4212 | if (dump_enabled_p ()) |
4213 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
4214 | "incompatible vector types for invariants\n" ); |
4215 | return false; |
4216 | } |
4217 | /* When the original call is pure or const but the SIMD ABI dictates |
4218 | an aggregate return we will have to use a virtual definition and |
4219 | in a loop eventually even need to add a virtual PHI. That's |
4220 | not straight-forward so allow to fix this up via renaming. */ |
4221 | if (gimple_call_lhs (gs: stmt) |
4222 | && !gimple_vdef (g: stmt) |
4223 | && TREE_CODE (TREE_TYPE (TREE_TYPE (bestn->decl))) == ARRAY_TYPE) |
4224 | vinfo->any_known_not_updated_vssa = true; |
4225 | /* ??? For SLP code-gen we end up inserting after the last |
4226 | vector argument def rather than at the original call position |
4227 | so automagic virtual operand updating doesn't work. */ |
4228 | if (gimple_vuse (g: stmt) && slp_node) |
4229 | vinfo->any_known_not_updated_vssa = true; |
4230 | simd_clone_info.safe_push (obj: bestn->decl); |
4231 | for (i = 0; i < bestn->simdclone->nargs; i++) |
4232 | { |
4233 | switch (bestn->simdclone->args[i].arg_type) |
4234 | { |
4235 | default: |
4236 | continue; |
4237 | case SIMD_CLONE_ARG_TYPE_LINEAR_CONSTANT_STEP: |
4238 | case SIMD_CLONE_ARG_TYPE_LINEAR_REF_CONSTANT_STEP: |
4239 | { |
4240 | simd_clone_info.safe_grow_cleared (len: i * 3 + 1, exact: true); |
4241 | simd_clone_info.safe_push (obj: arginfo[i].op); |
4242 | tree lst = POINTER_TYPE_P (TREE_TYPE (arginfo[i].op)) |
4243 | ? size_type_node : TREE_TYPE (arginfo[i].op); |
4244 | tree ls = build_int_cst (lst, arginfo[i].linear_step); |
4245 | simd_clone_info.safe_push (obj: ls); |
4246 | tree sll = arginfo[i].simd_lane_linear |
4247 | ? boolean_true_node : boolean_false_node; |
4248 | simd_clone_info.safe_push (obj: sll); |
4249 | } |
4250 | break; |
4251 | case SIMD_CLONE_ARG_TYPE_MASK: |
4252 | if (loop_vinfo |
4253 | && LOOP_VINFO_CAN_USE_PARTIAL_VECTORS_P (loop_vinfo)) |
4254 | vect_record_loop_mask (loop_vinfo, |
4255 | &LOOP_VINFO_MASKS (loop_vinfo), |
4256 | ncopies, vectype, op); |
4257 | |
4258 | break; |
4259 | } |
4260 | } |
4261 | |
4262 | if (!bestn->simdclone->inbranch && loop_vinfo) |
4263 | { |
4264 | if (dump_enabled_p () |
4265 | && LOOP_VINFO_CAN_USE_PARTIAL_VECTORS_P (loop_vinfo)) |
4266 | dump_printf_loc (MSG_NOTE, vect_location, |
4267 | "can't use a fully-masked loop because a" |
4268 | " non-masked simd clone was selected.\n" ); |
4269 | LOOP_VINFO_CAN_USE_PARTIAL_VECTORS_P (loop_vinfo) = false; |
4270 | } |
4271 | |
4272 | STMT_VINFO_TYPE (stmt_info) = call_simd_clone_vec_info_type; |
4273 | DUMP_VECT_SCOPE ("vectorizable_simd_clone_call" ); |
4274 | /* vect_model_simple_cost (vinfo, stmt_info, ncopies, |
4275 | dt, slp_node, cost_vec); */ |
4276 | return true; |
4277 | } |
4278 | |
4279 | /* Transform. */ |
4280 | |
4281 | if (dump_enabled_p ()) |
4282 | dump_printf_loc (MSG_NOTE, vect_location, "transform call.\n" ); |
4283 | |
4284 | /* Handle def. */ |
4285 | scalar_dest = gimple_call_lhs (gs: stmt); |
4286 | vec_dest = NULL_TREE; |
4287 | rtype = NULL_TREE; |
4288 | ratype = NULL_TREE; |
4289 | if (scalar_dest) |
4290 | { |
4291 | vec_dest = vect_create_destination_var (scalar_dest, vectype); |
4292 | rtype = TREE_TYPE (TREE_TYPE (fndecl)); |
4293 | if (TREE_CODE (rtype) == ARRAY_TYPE) |
4294 | { |
4295 | ratype = rtype; |
4296 | rtype = TREE_TYPE (ratype); |
4297 | } |
4298 | } |
4299 | |
4300 | auto_vec<vec<tree> > vec_oprnds; |
4301 | auto_vec<unsigned> vec_oprnds_i; |
4302 | vec_oprnds_i.safe_grow_cleared (len: nargs, exact: true); |
4303 | if (slp_node) |
4304 | { |
4305 | vec_oprnds.reserve_exact (nelems: nargs); |
4306 | vect_get_slp_defs (vinfo, slp_node, &vec_oprnds); |
4307 | } |
4308 | else |
4309 | vec_oprnds.safe_grow_cleared (len: nargs, exact: true); |
4310 | for (j = 0; j < ncopies; ++j) |
4311 | { |
4312 | poly_uint64 callee_nelements; |
4313 | poly_uint64 caller_nelements; |
4314 | /* Build argument list for the vectorized call. */ |
4315 | if (j == 0) |
4316 | vargs.create (nelems: nargs); |
4317 | else |
4318 | vargs.truncate (size: 0); |
4319 | |
4320 | for (i = 0; i < nargs; i++) |
4321 | { |
4322 | unsigned int k, l, m, o; |
4323 | tree atype; |
4324 | op = gimple_call_arg (gs: stmt, index: i + masked_call_offset); |
4325 | switch (bestn->simdclone->args[i].arg_type) |
4326 | { |
4327 | case SIMD_CLONE_ARG_TYPE_VECTOR: |
4328 | atype = bestn->simdclone->args[i].vector_type; |
4329 | caller_nelements = TYPE_VECTOR_SUBPARTS (node: arginfo[i].vectype); |
4330 | callee_nelements = TYPE_VECTOR_SUBPARTS (node: atype); |
4331 | o = vector_unroll_factor (nunits, callee_nelements); |
4332 | for (m = j * o; m < (j + 1) * o; m++) |
4333 | { |
4334 | if (known_lt (callee_nelements, caller_nelements)) |
4335 | { |
4336 | poly_uint64 prec = GET_MODE_BITSIZE (TYPE_MODE (atype)); |
4337 | if (!constant_multiple_p (a: caller_nelements, |
4338 | b: callee_nelements, multiple: &k)) |
4339 | gcc_unreachable (); |
4340 | |
4341 | gcc_assert ((k & (k - 1)) == 0); |
4342 | if (m == 0) |
4343 | { |
4344 | if (!slp_node) |
4345 | vect_get_vec_defs_for_operand (vinfo, stmt_vinfo: stmt_info, |
4346 | ncopies: ncopies * o / k, op, |
4347 | vec_oprnds: &vec_oprnds[i]); |
4348 | vec_oprnds_i[i] = 0; |
4349 | vec_oprnd0 = vec_oprnds[i][vec_oprnds_i[i]++]; |
4350 | } |
4351 | else |
4352 | { |
4353 | vec_oprnd0 = arginfo[i].op; |
4354 | if ((m & (k - 1)) == 0) |
4355 | vec_oprnd0 = vec_oprnds[i][vec_oprnds_i[i]++]; |
4356 | } |
4357 | arginfo[i].op = vec_oprnd0; |
4358 | vec_oprnd0 |
4359 | = build3 (BIT_FIELD_REF, atype, vec_oprnd0, |
4360 | bitsize_int (prec), |
4361 | bitsize_int ((m & (k - 1)) * prec)); |
4362 | gassign *new_stmt |
4363 | = gimple_build_assign (make_ssa_name (var: atype), |
4364 | vec_oprnd0); |
4365 | vect_finish_stmt_generation (vinfo, stmt_info, |
4366 | vec_stmt: new_stmt, gsi); |
4367 | vargs.safe_push (obj: gimple_assign_lhs (gs: new_stmt)); |
4368 | } |
4369 | else |
4370 | { |
4371 | if (!constant_multiple_p (a: callee_nelements, |
4372 | b: caller_nelements, multiple: &k)) |
4373 | gcc_unreachable (); |
4374 | gcc_assert ((k & (k - 1)) == 0); |
4375 | vec<constructor_elt, va_gc> *ctor_elts; |
4376 | if (k != 1) |
4377 | vec_alloc (v&: ctor_elts, nelems: k); |
4378 | else |
4379 | ctor_elts = NULL; |
4380 | for (l = 0; l < k; l++) |
4381 | { |
4382 | if (m == 0 && l == 0) |
4383 | { |
4384 | if (!slp_node) |
4385 | vect_get_vec_defs_for_operand (vinfo, stmt_vinfo: stmt_info, |
4386 | ncopies: k * o * ncopies, |
4387 | op, |
4388 | vec_oprnds: &vec_oprnds[i]); |
4389 | vec_oprnds_i[i] = 0; |
4390 | vec_oprnd0 = vec_oprnds[i][vec_oprnds_i[i]++]; |
4391 | } |
4392 | else |
4393 | vec_oprnd0 = vec_oprnds[i][vec_oprnds_i[i]++]; |
4394 | arginfo[i].op = vec_oprnd0; |
4395 | if (k == 1) |
4396 | break; |
4397 | CONSTRUCTOR_APPEND_ELT (ctor_elts, NULL_TREE, |
4398 | vec_oprnd0); |
4399 | } |
4400 | if (k == 1) |
4401 | if (!useless_type_conversion_p (TREE_TYPE (vec_oprnd0), |
4402 | atype)) |
4403 | { |
4404 | vec_oprnd0 = build1 (VIEW_CONVERT_EXPR, atype, |
4405 | vec_oprnd0); |
4406 | gassign *new_stmt |
4407 | = gimple_build_assign (make_ssa_name (var: atype), |
4408 | vec_oprnd0); |
4409 | vect_finish_stmt_generation (vinfo, stmt_info, |
4410 | vec_stmt: new_stmt, gsi); |
4411 | vargs.safe_push (obj: gimple_get_lhs (new_stmt)); |
4412 | } |
4413 | else |
4414 | vargs.safe_push (obj: vec_oprnd0); |
4415 | else |
4416 | { |
4417 | vec_oprnd0 = build_constructor (atype, ctor_elts); |
4418 | gassign *new_stmt |
4419 | = gimple_build_assign (make_ssa_name (var: atype), |
4420 | vec_oprnd0); |
4421 | vect_finish_stmt_generation (vinfo, stmt_info, |
4422 | vec_stmt: new_stmt, gsi); |
4423 | vargs.safe_push (obj: gimple_assign_lhs (gs: new_stmt)); |
4424 | } |
4425 | } |
4426 | } |
4427 | break; |
4428 | case SIMD_CLONE_ARG_TYPE_MASK: |
4429 | if (bestn->simdclone->mask_mode == VOIDmode) |
4430 | { |
4431 | atype = bestn->simdclone->args[i].vector_type; |
4432 | tree elt_type = TREE_TYPE (atype); |
4433 | tree one = fold_convert (elt_type, integer_one_node); |
4434 | tree zero = fold_convert (elt_type, integer_zero_node); |
4435 | callee_nelements = TYPE_VECTOR_SUBPARTS (node: atype); |
4436 | caller_nelements = TYPE_VECTOR_SUBPARTS (node: arginfo[i].vectype); |
4437 | o = vector_unroll_factor (nunits, callee_nelements); |
4438 | for (m = j * o; m < (j + 1) * o; m++) |
4439 | { |
4440 | if (maybe_lt (a: callee_nelements, b: caller_nelements)) |
4441 | { |
4442 | /* The mask type has fewer elements than simdlen. */ |
4443 | |
4444 | /* FORNOW */ |
4445 | gcc_unreachable (); |
4446 | } |
4447 | else if (known_eq (callee_nelements, caller_nelements)) |
4448 | { |
4449 | /* The SIMD clone function has the same number of |
4450 | elements as the current function. */ |
4451 | if (m == 0) |
4452 | { |
4453 | if (!slp_node) |
4454 | vect_get_vec_defs_for_operand (vinfo, stmt_vinfo: stmt_info, |
4455 | ncopies: o * ncopies, |
4456 | op, |
4457 | vec_oprnds: &vec_oprnds[i]); |
4458 | vec_oprnds_i[i] = 0; |
4459 | } |
4460 | vec_oprnd0 = vec_oprnds[i][vec_oprnds_i[i]++]; |
4461 | if (loop_vinfo |
4462 | && LOOP_VINFO_FULLY_MASKED_P (loop_vinfo)) |
4463 | { |
4464 | vec_loop_masks *loop_masks |
4465 | = &LOOP_VINFO_MASKS (loop_vinfo); |
4466 | tree loop_mask |
4467 | = vect_get_loop_mask (loop_vinfo, gsi, |
4468 | loop_masks, ncopies, |
4469 | vectype, j); |
4470 | vec_oprnd0 |
4471 | = prepare_vec_mask (loop_vinfo, |
4472 | TREE_TYPE (loop_mask), |
4473 | loop_mask, vec_mask: vec_oprnd0, |
4474 | gsi); |
4475 | loop_vinfo->vec_cond_masked_set.add (k: { vec_oprnd0, |
4476 | loop_mask }); |
4477 | |
4478 | } |
4479 | vec_oprnd0 |
4480 | = build3 (VEC_COND_EXPR, atype, vec_oprnd0, |
4481 | build_vector_from_val (atype, one), |
4482 | build_vector_from_val (atype, zero)); |
4483 | gassign *new_stmt |
4484 | = gimple_build_assign (make_ssa_name (var: atype), |
4485 | vec_oprnd0); |
4486 | vect_finish_stmt_generation (vinfo, stmt_info, |
4487 | vec_stmt: new_stmt, gsi); |
4488 | vargs.safe_push (obj: gimple_assign_lhs (gs: new_stmt)); |
4489 | } |
4490 | else |
4491 | { |
4492 | /* The mask type has more elements than simdlen. */ |
4493 | |
4494 | /* FORNOW */ |
4495 | gcc_unreachable (); |
4496 | } |
4497 | } |
4498 | } |
4499 | else if (SCALAR_INT_MODE_P (bestn->simdclone->mask_mode)) |
4500 | { |
4501 | atype = bestn->simdclone->args[i].vector_type; |
4502 | /* Guess the number of lanes represented by atype. */ |
4503 | poly_uint64 atype_subparts |
4504 | = exact_div (a: bestn->simdclone->simdlen, |
4505 | b: num_mask_args); |
4506 | o = vector_unroll_factor (nunits, atype_subparts); |
4507 | for (m = j * o; m < (j + 1) * o; m++) |
4508 | { |
4509 | if (m == 0) |
4510 | { |
4511 | if (!slp_node) |
4512 | vect_get_vec_defs_for_operand (vinfo, stmt_vinfo: stmt_info, |
4513 | ncopies: o * ncopies, |
4514 | op, |
4515 | vec_oprnds: &vec_oprnds[i]); |
4516 | vec_oprnds_i[i] = 0; |
4517 | } |
4518 | if (maybe_lt (a: atype_subparts, |
4519 | b: TYPE_VECTOR_SUBPARTS (node: arginfo[i].vectype))) |
4520 | { |
4521 | /* The mask argument has fewer elements than the |
4522 | input vector. */ |
4523 | /* FORNOW */ |
4524 | gcc_unreachable (); |
4525 | } |
4526 | else if (known_eq (atype_subparts, |
4527 | TYPE_VECTOR_SUBPARTS (arginfo[i].vectype))) |
4528 | { |
4529 | /* The vector mask argument matches the input |
4530 | in the number of lanes, but not necessarily |
4531 | in the mode. */ |
4532 | vec_oprnd0 = vec_oprnds[i][vec_oprnds_i[i]++]; |
4533 | tree st = lang_hooks.types.type_for_mode |
4534 | (TYPE_MODE (TREE_TYPE (vec_oprnd0)), 1); |
4535 | vec_oprnd0 = build1 (VIEW_CONVERT_EXPR, st, |
4536 | vec_oprnd0); |
4537 | gassign *new_stmt |
4538 | = gimple_build_assign (make_ssa_name (var: st), |
4539 | vec_oprnd0); |
4540 | vect_finish_stmt_generation (vinfo, stmt_info, |
4541 | vec_stmt: new_stmt, gsi); |
4542 | if (!types_compatible_p (type1: atype, type2: st)) |
4543 | { |
4544 | new_stmt |
4545 | = gimple_build_assign (make_ssa_name (var: atype), |
4546 | NOP_EXPR, |
4547 | gimple_assign_lhs |
4548 | (gs: new_stmt)); |
4549 | vect_finish_stmt_generation (vinfo, stmt_info, |
4550 | vec_stmt: new_stmt, gsi); |
4551 | } |
4552 | vargs.safe_push (obj: gimple_assign_lhs (gs: new_stmt)); |
4553 | } |
4554 | else |
4555 | { |
4556 | /* The mask argument has more elements than the |
4557 | input vector. */ |
4558 | /* FORNOW */ |
4559 | gcc_unreachable (); |
4560 | } |
4561 | } |
4562 | } |
4563 | else |
4564 | gcc_unreachable (); |
4565 | break; |
4566 | case SIMD_CLONE_ARG_TYPE_UNIFORM: |
4567 | vargs.safe_push (obj: op); |
4568 | break; |
4569 | case SIMD_CLONE_ARG_TYPE_LINEAR_CONSTANT_STEP: |
4570 | case SIMD_CLONE_ARG_TYPE_LINEAR_REF_CONSTANT_STEP: |
4571 | if (j == 0) |
4572 | { |
4573 | gimple_seq stmts; |
4574 | arginfo[i].op |
4575 | = force_gimple_operand (unshare_expr (arginfo[i].op), |
4576 | &stmts, true, NULL_TREE); |
4577 | if (stmts != NULL) |
4578 | { |
4579 | basic_block new_bb; |
4580 | edge pe = loop_preheader_edge (loop); |
4581 | new_bb = gsi_insert_seq_on_edge_immediate (pe, stmts); |
4582 | gcc_assert (!new_bb); |
4583 | } |
4584 | if (arginfo[i].simd_lane_linear) |
4585 | { |
4586 | vargs.safe_push (obj: arginfo[i].op); |
4587 | break; |
4588 | } |
4589 | tree phi_res = copy_ssa_name (var: op); |
4590 | gphi *new_phi = create_phi_node (phi_res, loop->header); |
4591 | add_phi_arg (new_phi, arginfo[i].op, |
4592 | loop_preheader_edge (loop), UNKNOWN_LOCATION); |
4593 | enum tree_code code |
4594 | = POINTER_TYPE_P (TREE_TYPE (op)) |
4595 | ? POINTER_PLUS_EXPR : PLUS_EXPR; |
4596 | tree type = POINTER_TYPE_P (TREE_TYPE (op)) |
4597 | ? sizetype : TREE_TYPE (op); |
4598 | poly_widest_int cst |
4599 | = wi::mul (a: bestn->simdclone->args[i].linear_step, |
4600 | b: ncopies * nunits); |
4601 | tree tcst = wide_int_to_tree (type, cst); |
4602 | tree phi_arg = copy_ssa_name (var: op); |
4603 | gassign *new_stmt |
4604 | = gimple_build_assign (phi_arg, code, phi_res, tcst); |
4605 | gimple_stmt_iterator si = gsi_after_labels (bb: loop->header); |
4606 | gsi_insert_after (&si, new_stmt, GSI_NEW_STMT); |
4607 | add_phi_arg (new_phi, phi_arg, loop_latch_edge (loop), |
4608 | UNKNOWN_LOCATION); |
4609 | arginfo[i].op = phi_res; |
4610 | vargs.safe_push (obj: phi_res); |
4611 | } |
4612 | else |
4613 | { |
4614 | enum tree_code code |
4615 | = POINTER_TYPE_P (TREE_TYPE (op)) |
4616 | ? POINTER_PLUS_EXPR : PLUS_EXPR; |
4617 | tree type = POINTER_TYPE_P (TREE_TYPE (op)) |
4618 | ? sizetype : TREE_TYPE (op); |
4619 | poly_widest_int cst |
4620 | = wi::mul (a: bestn->simdclone->args[i].linear_step, |
4621 | b: j * nunits); |
4622 | tree tcst = wide_int_to_tree (type, cst); |
4623 | new_temp = make_ssa_name (TREE_TYPE (op)); |
4624 | gassign *new_stmt |
4625 | = gimple_build_assign (new_temp, code, |
4626 | arginfo[i].op, tcst); |
4627 | vect_finish_stmt_generation (vinfo, stmt_info, vec_stmt: new_stmt, gsi); |
4628 | vargs.safe_push (obj: new_temp); |
4629 | } |
4630 | break; |
4631 | case SIMD_CLONE_ARG_TYPE_LINEAR_VAL_CONSTANT_STEP: |
4632 | case SIMD_CLONE_ARG_TYPE_LINEAR_UVAL_CONSTANT_STEP: |
4633 | case SIMD_CLONE_ARG_TYPE_LINEAR_VARIABLE_STEP: |
4634 | case SIMD_CLONE_ARG_TYPE_LINEAR_REF_VARIABLE_STEP: |
4635 | case SIMD_CLONE_ARG_TYPE_LINEAR_VAL_VARIABLE_STEP: |
4636 | case SIMD_CLONE_ARG_TYPE_LINEAR_UVAL_VARIABLE_STEP: |
4637 | default: |
4638 | gcc_unreachable (); |
4639 | } |
4640 | } |
4641 | |
4642 | if (masked_call_offset == 0 |
4643 | && bestn->simdclone->inbranch |
4644 | && bestn->simdclone->nargs > nargs) |
4645 | { |
4646 | unsigned long m, o; |
4647 | size_t mask_i = bestn->simdclone->nargs - 1; |
4648 | tree mask; |
4649 | gcc_assert (bestn->simdclone->args[mask_i].arg_type == |
4650 | SIMD_CLONE_ARG_TYPE_MASK); |
4651 | |
4652 | tree masktype = bestn->simdclone->args[mask_i].vector_type; |
4653 | callee_nelements = TYPE_VECTOR_SUBPARTS (node: masktype); |
4654 | o = vector_unroll_factor (nunits, callee_nelements); |
4655 | for (m = j * o; m < (j + 1) * o; m++) |
4656 | { |
4657 | if (loop_vinfo && LOOP_VINFO_FULLY_MASKED_P (loop_vinfo)) |
4658 | { |
4659 | vec_loop_masks *loop_masks = &LOOP_VINFO_MASKS (loop_vinfo); |
4660 | mask = vect_get_loop_mask (loop_vinfo, gsi, loop_masks, |
4661 | ncopies, vectype, j); |
4662 | } |
4663 | else |
4664 | mask = vect_build_all_ones_mask (vinfo, stmt_info, masktype); |
4665 | |
4666 | gassign *new_stmt; |
4667 | if (SCALAR_INT_MODE_P (bestn->simdclone->mask_mode)) |
4668 | { |
4669 | /* This means we are dealing with integer mask modes. |
4670 | First convert to an integer type with the same size as |
4671 | the current vector type. */ |
4672 | unsigned HOST_WIDE_INT intermediate_size |
4673 | = tree_to_uhwi (TYPE_SIZE (TREE_TYPE (mask))); |
4674 | tree mid_int_type = |
4675 | build_nonstandard_integer_type (intermediate_size, 1); |
4676 | mask = build1 (VIEW_CONVERT_EXPR, mid_int_type, mask); |
4677 | new_stmt |
4678 | = gimple_build_assign (make_ssa_name (var: mid_int_type), |
4679 | mask); |
4680 | gsi_insert_before (gsi, new_stmt, GSI_SAME_STMT); |
4681 | /* Then zero-extend to the mask mode. */ |
4682 | mask = fold_build1 (NOP_EXPR, masktype, |
4683 | gimple_get_lhs (new_stmt)); |
4684 | } |
4685 | else if (bestn->simdclone->mask_mode == VOIDmode) |
4686 | { |
4687 | tree one = fold_convert (TREE_TYPE (masktype), |
4688 | integer_one_node); |
4689 | tree zero = fold_convert (TREE_TYPE (masktype), |
4690 | integer_zero_node); |
4691 | mask = build3 (VEC_COND_EXPR, masktype, mask, |
4692 | build_vector_from_val (masktype, one), |
4693 | build_vector_from_val (masktype, zero)); |
4694 | } |
4695 | else |
4696 | gcc_unreachable (); |
4697 | |
4698 | new_stmt = gimple_build_assign (make_ssa_name (var: masktype), mask); |
4699 | vect_finish_stmt_generation (vinfo, stmt_info, |
4700 | vec_stmt: new_stmt, gsi); |
4701 | mask = gimple_assign_lhs (gs: new_stmt); |
4702 | vargs.safe_push (obj: mask); |
4703 | } |
4704 | } |
4705 | |
4706 | gcall *new_call = gimple_build_call_vec (fndecl, vargs); |
4707 | if (vec_dest) |
4708 | { |
4709 | gcc_assert (ratype |
4710 | || known_eq (TYPE_VECTOR_SUBPARTS (rtype), nunits)); |
4711 | if (ratype) |
4712 | new_temp = create_tmp_var (ratype); |
4713 | else if (useless_type_conversion_p (vectype, rtype)) |
4714 | new_temp = make_ssa_name (var: vec_dest, stmt: new_call); |
4715 | else |
4716 | new_temp = make_ssa_name (var: rtype, stmt: new_call); |
4717 | gimple_call_set_lhs (gs: new_call, lhs: new_temp); |
4718 | } |
4719 | vect_finish_stmt_generation (vinfo, stmt_info, vec_stmt: new_call, gsi); |
4720 | gimple *new_stmt = new_call; |
4721 | |
4722 | if (vec_dest) |
4723 | { |
4724 | if (!multiple_p (a: TYPE_VECTOR_SUBPARTS (node: vectype), b: nunits)) |
4725 | { |
4726 | unsigned int k, l; |
4727 | poly_uint64 prec = GET_MODE_BITSIZE (TYPE_MODE (vectype)); |
4728 | poly_uint64 bytes = GET_MODE_SIZE (TYPE_MODE (vectype)); |
4729 | k = vector_unroll_factor (nunits, |
4730 | TYPE_VECTOR_SUBPARTS (vectype)); |
4731 | gcc_assert ((k & (k - 1)) == 0); |
4732 | for (l = 0; l < k; l++) |
4733 | { |
4734 | tree t; |
4735 | if (ratype) |
4736 | { |
4737 | t = build_fold_addr_expr (new_temp); |
4738 | t = build2 (MEM_REF, vectype, t, |
4739 | build_int_cst (TREE_TYPE (t), l * bytes)); |
4740 | } |
4741 | else |
4742 | t = build3 (BIT_FIELD_REF, vectype, new_temp, |
4743 | bitsize_int (prec), bitsize_int (l * prec)); |
4744 | new_stmt = gimple_build_assign (make_ssa_name (var: vectype), t); |
4745 | vect_finish_stmt_generation (vinfo, stmt_info, vec_stmt: new_stmt, gsi); |
4746 | |
4747 | if (j == 0 && l == 0) |
4748 | *vec_stmt = new_stmt; |
4749 | if (slp_node) |
4750 | SLP_TREE_VEC_DEFS (slp_node) |
4751 | .quick_push (obj: gimple_assign_lhs (gs: new_stmt)); |
4752 | else |
4753 | STMT_VINFO_VEC_STMTS (stmt_info).safe_push (obj: new_stmt); |
4754 | } |
4755 | |
4756 | if (ratype) |
4757 | vect_clobber_variable (vinfo, stmt_info, gsi, var: new_temp); |
4758 | continue; |
4759 | } |
4760 | else if (!multiple_p (a: nunits, b: TYPE_VECTOR_SUBPARTS (node: vectype))) |
4761 | { |
4762 | unsigned int k; |
4763 | if (!constant_multiple_p (a: TYPE_VECTOR_SUBPARTS (node: vectype), |
4764 | b: TYPE_VECTOR_SUBPARTS (node: rtype), multiple: &k)) |
4765 | gcc_unreachable (); |
4766 | gcc_assert ((k & (k - 1)) == 0); |
4767 | if ((j & (k - 1)) == 0) |
4768 | vec_alloc (v&: ret_ctor_elts, nelems: k); |
4769 | if (ratype) |
4770 | { |
4771 | unsigned int m, o; |
4772 | o = vector_unroll_factor (nunits, |
4773 | TYPE_VECTOR_SUBPARTS (rtype)); |
4774 | for (m = 0; m < o; m++) |
4775 | { |
4776 | tree tem = build4 (ARRAY_REF, rtype, new_temp, |
4777 | size_int (m), NULL_TREE, NULL_TREE); |
4778 | new_stmt = gimple_build_assign (make_ssa_name (var: rtype), |
4779 | tem); |
4780 | vect_finish_stmt_generation (vinfo, stmt_info, |
4781 | vec_stmt: new_stmt, gsi); |
4782 | CONSTRUCTOR_APPEND_ELT (ret_ctor_elts, NULL_TREE, |
4783 | gimple_assign_lhs (new_stmt)); |
4784 | } |
4785 | vect_clobber_variable (vinfo, stmt_info, gsi, var: new_temp); |
4786 | } |
4787 | else |
4788 | CONSTRUCTOR_APPEND_ELT (ret_ctor_elts, NULL_TREE, new_temp); |
4789 | if ((j & (k - 1)) != k - 1) |
4790 | continue; |
4791 | vec_oprnd0 = build_constructor (vectype, ret_ctor_elts); |
4792 | new_stmt |
4793 | = gimple_build_assign (make_ssa_name (var: vec_dest), vec_oprnd0); |
4794 | vect_finish_stmt_generation (vinfo, stmt_info, vec_stmt: new_stmt, gsi); |
4795 | |
4796 | if ((unsigned) j == k - 1) |
4797 | *vec_stmt = new_stmt; |
4798 | if (slp_node) |
4799 | SLP_TREE_VEC_DEFS (slp_node) |
4800 | .quick_push (obj: gimple_assign_lhs (gs: new_stmt)); |
4801 | else |
4802 | STMT_VINFO_VEC_STMTS (stmt_info).safe_push (obj: new_stmt); |
4803 | continue; |
4804 | } |
4805 | else if (ratype) |
4806 | { |
4807 | tree t = build_fold_addr_expr (new_temp); |
4808 | t = build2 (MEM_REF, vectype, t, |
4809 | build_int_cst (TREE_TYPE (t), 0)); |
4810 | new_stmt = gimple_build_assign (make_ssa_name (var: vec_dest), t); |
4811 | vect_finish_stmt_generation (vinfo, stmt_info, vec_stmt: new_stmt, gsi); |
4812 | vect_clobber_variable (vinfo, stmt_info, gsi, var: new_temp); |
4813 | } |
4814 | else if (!useless_type_conversion_p (vectype, rtype)) |
4815 | { |
4816 | vec_oprnd0 = build1 (VIEW_CONVERT_EXPR, vectype, new_temp); |
4817 | new_stmt |
4818 | = gimple_build_assign (make_ssa_name (var: vec_dest), vec_oprnd0); |
4819 | vect_finish_stmt_generation (vinfo, stmt_info, vec_stmt: new_stmt, gsi); |
4820 | } |
4821 | } |
4822 | |
4823 | if (j == 0) |
4824 | *vec_stmt = new_stmt; |
4825 | if (slp_node) |
4826 | SLP_TREE_VEC_DEFS (slp_node).quick_push (obj: gimple_get_lhs (new_stmt)); |
4827 | else |
4828 | STMT_VINFO_VEC_STMTS (stmt_info).safe_push (obj: new_stmt); |
4829 | } |
4830 | |
4831 | for (i = 0; i < nargs; ++i) |
4832 | { |
4833 | vec<tree> oprndsi = vec_oprnds[i]; |
4834 | oprndsi.release (); |
4835 | } |
4836 | vargs.release (); |
4837 | |
4838 | /* Mark the clone as no longer being a candidate for GC. */ |
4839 | bestn->gc_candidate = false; |
4840 | |
4841 | /* The call in STMT might prevent it from being removed in dce. |
4842 | We however cannot remove it here, due to the way the ssa name |
4843 | it defines is mapped to the new definition. So just replace |
4844 | rhs of the statement with something harmless. */ |
4845 | |
4846 | if (slp_node) |
4847 | return true; |
4848 | |
4849 | gimple *new_stmt; |
4850 | if (scalar_dest) |
4851 | { |
4852 | type = TREE_TYPE (scalar_dest); |
4853 | lhs = gimple_call_lhs (gs: vect_orig_stmt (stmt_info)->stmt); |
4854 | new_stmt = gimple_build_assign (lhs, build_zero_cst (type)); |
4855 | } |
4856 | else |
4857 | new_stmt = gimple_build_nop (); |
4858 | vinfo->replace_stmt (gsi, vect_orig_stmt (stmt_info), new_stmt); |
4859 | unlink_stmt_vdef (stmt); |
4860 | |
4861 | return true; |
4862 | } |
4863 | |
4864 | |
4865 | /* Function vect_gen_widened_results_half |
4866 | |
4867 | Create a vector stmt whose code, type, number of arguments, and result |
4868 | variable are CODE, OP_TYPE, and VEC_DEST, and its arguments are |
4869 | VEC_OPRND0 and VEC_OPRND1. The new vector stmt is to be inserted at GSI. |
4870 | In the case that CODE is a CALL_EXPR, this means that a call to DECL |
4871 | needs to be created (DECL is a function-decl of a target-builtin). |
4872 | STMT_INFO is the original scalar stmt that we are vectorizing. */ |
4873 | |
4874 | static gimple * |
4875 | vect_gen_widened_results_half (vec_info *vinfo, code_helper ch, |
4876 | tree vec_oprnd0, tree vec_oprnd1, int op_type, |
4877 | tree vec_dest, gimple_stmt_iterator *gsi, |
4878 | stmt_vec_info stmt_info) |
4879 | { |
4880 | gimple *new_stmt; |
4881 | tree new_temp; |
4882 | |
4883 | /* Generate half of the widened result: */ |
4884 | if (op_type != binary_op) |
4885 | vec_oprnd1 = NULL; |
4886 | new_stmt = vect_gimple_build (vec_dest, ch, vec_oprnd0, vec_oprnd1); |
4887 | new_temp = make_ssa_name (var: vec_dest, stmt: new_stmt); |
4888 | gimple_set_lhs (new_stmt, new_temp); |
4889 | vect_finish_stmt_generation (vinfo, stmt_info, vec_stmt: new_stmt, gsi); |
4890 | |
4891 | return new_stmt; |
4892 | } |
4893 | |
4894 | |
4895 | /* Create vectorized demotion statements for vector operands from VEC_OPRNDS. |
4896 | For multi-step conversions store the resulting vectors and call the function |
4897 | recursively. When NARROW_SRC_P is true, there's still a conversion after |
4898 | narrowing, don't store the vectors in the SLP_NODE or in vector info of |
4899 | the scalar statement(or in STMT_VINFO_RELATED_STMT chain). */ |
4900 | |
4901 | static void |
4902 | vect_create_vectorized_demotion_stmts (vec_info *vinfo, vec<tree> *vec_oprnds, |
4903 | int multi_step_cvt, |
4904 | stmt_vec_info stmt_info, |
4905 | vec<tree> &vec_dsts, |
4906 | gimple_stmt_iterator *gsi, |
4907 | slp_tree slp_node, code_helper code, |
4908 | bool narrow_src_p) |
4909 | { |
4910 | unsigned int i; |
4911 | tree vop0, vop1, new_tmp, vec_dest; |
4912 | |
4913 | vec_dest = vec_dsts.pop (); |
4914 | |
4915 | for (i = 0; i < vec_oprnds->length (); i += 2) |
4916 | { |
4917 | /* Create demotion operation. */ |
4918 | vop0 = (*vec_oprnds)[i]; |
4919 | vop1 = (*vec_oprnds)[i + 1]; |
4920 | gimple *new_stmt = vect_gimple_build (vec_dest, code, vop0, vop1); |
4921 | new_tmp = make_ssa_name (var: vec_dest, stmt: new_stmt); |
4922 | gimple_set_lhs (new_stmt, new_tmp); |
4923 | vect_finish_stmt_generation (vinfo, stmt_info, vec_stmt: new_stmt, gsi); |
4924 | if (multi_step_cvt || narrow_src_p) |
4925 | /* Store the resulting vector for next recursive call, |
4926 | or return the resulting vector_tmp for NARROW FLOAT_EXPR. */ |
4927 | (*vec_oprnds)[i/2] = new_tmp; |
4928 | else |
4929 | { |
4930 | /* This is the last step of the conversion sequence. Store the |
4931 | vectors in SLP_NODE or in vector info of the scalar statement |
4932 | (or in STMT_VINFO_RELATED_STMT chain). */ |
4933 | if (slp_node) |
4934 | slp_node->push_vec_def (def: new_stmt); |
4935 | else |
4936 | STMT_VINFO_VEC_STMTS (stmt_info).safe_push (obj: new_stmt); |
4937 | } |
4938 | } |
4939 | |
4940 | /* For multi-step demotion operations we first generate demotion operations |
4941 | from the source type to the intermediate types, and then combine the |
4942 | results (stored in VEC_OPRNDS) in demotion operation to the destination |
4943 | type. */ |
4944 | if (multi_step_cvt) |
4945 | { |
4946 | /* At each level of recursion we have half of the operands we had at the |
4947 | previous level. */ |
4948 | vec_oprnds->truncate (size: (i+1)/2); |
4949 | vect_create_vectorized_demotion_stmts (vinfo, vec_oprnds, |
4950 | multi_step_cvt: multi_step_cvt - 1, |
4951 | stmt_info, vec_dsts, gsi, |
4952 | slp_node, code: VEC_PACK_TRUNC_EXPR, |
4953 | narrow_src_p); |
4954 | } |
4955 | |
4956 | vec_dsts.quick_push (obj: vec_dest); |
4957 | } |
4958 | |
4959 | |
4960 | /* Create vectorized promotion statements for vector operands from VEC_OPRNDS0 |
4961 | and VEC_OPRNDS1, for a binary operation associated with scalar statement |
4962 | STMT_INFO. For multi-step conversions store the resulting vectors and |
4963 | call the function recursively. */ |
4964 | |
4965 | static void |
4966 | vect_create_vectorized_promotion_stmts (vec_info *vinfo, |
4967 | vec<tree> *vec_oprnds0, |
4968 | vec<tree> *vec_oprnds1, |
4969 | stmt_vec_info stmt_info, tree vec_dest, |
4970 | gimple_stmt_iterator *gsi, |
4971 | code_helper ch1, |
4972 | code_helper ch2, int op_type) |
4973 | { |
4974 | int i; |
4975 | tree vop0, vop1, new_tmp1, new_tmp2; |
4976 | gimple *new_stmt1, *new_stmt2; |
4977 | vec<tree> vec_tmp = vNULL; |
4978 | |
4979 | vec_tmp.create (nelems: vec_oprnds0->length () * 2); |
4980 | FOR_EACH_VEC_ELT (*vec_oprnds0, i, vop0) |
4981 | { |
4982 | if (op_type == binary_op) |
4983 | vop1 = (*vec_oprnds1)[i]; |
4984 | else |
4985 | vop1 = NULL_TREE; |
4986 | |
4987 | /* Generate the two halves of promotion operation. */ |
4988 | new_stmt1 = vect_gen_widened_results_half (vinfo, ch: ch1, vec_oprnd0: vop0, vec_oprnd1: vop1, |
4989 | op_type, vec_dest, gsi, |
4990 | stmt_info); |
4991 | new_stmt2 = vect_gen_widened_results_half (vinfo, ch: ch2, vec_oprnd0: vop0, vec_oprnd1: vop1, |
4992 | op_type, vec_dest, gsi, |
4993 | stmt_info); |
4994 | if (is_gimple_call (gs: new_stmt1)) |
4995 | { |
4996 | new_tmp1 = gimple_call_lhs (gs: new_stmt1); |
4997 | new_tmp2 = gimple_call_lhs (gs: new_stmt2); |
4998 | } |
4999 | else |
5000 | { |
5001 | new_tmp1 = gimple_assign_lhs (gs: new_stmt1); |
5002 | new_tmp2 = gimple_assign_lhs (gs: new_stmt2); |
5003 | } |
5004 | |
5005 | /* Store the results for the next step. */ |
5006 | vec_tmp.quick_push (obj: new_tmp1); |
5007 | vec_tmp.quick_push (obj: new_tmp2); |
5008 | } |
5009 | |
5010 | vec_oprnds0->release (); |
5011 | *vec_oprnds0 = vec_tmp; |
5012 | } |
5013 | |
5014 | /* Create vectorized promotion stmts for widening stmts using only half the |
5015 | potential vector size for input. */ |
5016 | static void |
5017 | vect_create_half_widening_stmts (vec_info *vinfo, |
5018 | vec<tree> *vec_oprnds0, |
5019 | vec<tree> *vec_oprnds1, |
5020 | stmt_vec_info stmt_info, tree vec_dest, |
5021 | gimple_stmt_iterator *gsi, |
5022 | code_helper code1, |
5023 | int op_type) |
5024 | { |
5025 | int i; |
5026 | tree vop0, vop1; |
5027 | gimple *new_stmt1; |
5028 | gimple *new_stmt2; |
5029 | gimple *new_stmt3; |
5030 | vec<tree> vec_tmp = vNULL; |
5031 | |
5032 | vec_tmp.create (nelems: vec_oprnds0->length ()); |
5033 | FOR_EACH_VEC_ELT (*vec_oprnds0, i, vop0) |
5034 | { |
5035 | tree new_tmp1, new_tmp2, new_tmp3, out_type; |
5036 | |
5037 | gcc_assert (op_type == binary_op); |
5038 | vop1 = (*vec_oprnds1)[i]; |
5039 | |
5040 | /* Widen the first vector input. */ |
5041 | out_type = TREE_TYPE (vec_dest); |
5042 | new_tmp1 = make_ssa_name (var: out_type); |
5043 | new_stmt1 = gimple_build_assign (new_tmp1, NOP_EXPR, vop0); |
5044 | vect_finish_stmt_generation (vinfo, stmt_info, vec_stmt: new_stmt1, gsi); |
5045 | if (VECTOR_TYPE_P (TREE_TYPE (vop1))) |
5046 | { |
5047 | /* Widen the second vector input. */ |
5048 | new_tmp2 = make_ssa_name (var: out_type); |
5049 | new_stmt2 = gimple_build_assign (new_tmp2, NOP_EXPR, vop1); |
5050 | vect_finish_stmt_generation (vinfo, stmt_info, vec_stmt: new_stmt2, gsi); |
5051 | /* Perform the operation. With both vector inputs widened. */ |
5052 | new_stmt3 = vect_gimple_build (vec_dest, code1, new_tmp1, new_tmp2); |
5053 | } |
5054 | else |
5055 | { |
5056 | /* Perform the operation. With the single vector input widened. */ |
5057 | new_stmt3 = vect_gimple_build (vec_dest, code1, new_tmp1, vop1); |
5058 | } |
5059 | |
5060 | new_tmp3 = make_ssa_name (var: vec_dest, stmt: new_stmt3); |
5061 | gimple_assign_set_lhs (gs: new_stmt3, lhs: new_tmp3); |
5062 | vect_finish_stmt_generation (vinfo, stmt_info, vec_stmt: new_stmt3, gsi); |
5063 | |
5064 | /* Store the results for the next step. */ |
5065 | vec_tmp.quick_push (obj: new_tmp3); |
5066 | } |
5067 | |
5068 | vec_oprnds0->release (); |
5069 | *vec_oprnds0 = vec_tmp; |
5070 | } |
5071 | |
5072 | |
5073 | /* Check if STMT_INFO performs a conversion operation that can be vectorized. |
5074 | If VEC_STMT is also passed, vectorize STMT_INFO: create a vectorized |
5075 | stmt to replace it, put it in VEC_STMT, and insert it at GSI. |
5076 | Return true if STMT_INFO is vectorizable in this way. */ |
5077 | |
5078 | static bool |
5079 | vectorizable_conversion (vec_info *vinfo, |
5080 | stmt_vec_info stmt_info, gimple_stmt_iterator *gsi, |
5081 | gimple **vec_stmt, slp_tree slp_node, |
5082 | stmt_vector_for_cost *cost_vec) |
5083 | { |
5084 | tree vec_dest, cvt_op = NULL_TREE; |
5085 | tree scalar_dest; |
5086 | tree op0, op1 = NULL_TREE; |
5087 | loop_vec_info loop_vinfo = dyn_cast <loop_vec_info> (p: vinfo); |
5088 | tree_code tc1, tc2; |
5089 | code_helper code, code1, code2; |
5090 | code_helper codecvt1 = ERROR_MARK, codecvt2 = ERROR_MARK; |
5091 | tree new_temp; |
5092 | enum vect_def_type dt[2] = {vect_unknown_def_type, vect_unknown_def_type}; |
5093 | int ndts = 2; |
5094 | poly_uint64 nunits_in; |
5095 | poly_uint64 nunits_out; |
5096 | tree vectype_out, vectype_in; |
5097 | int ncopies, i; |
5098 | tree lhs_type, rhs_type; |
5099 | /* For conversions between floating point and integer, there're 2 NARROW |
5100 | cases. NARROW_SRC is for FLOAT_EXPR, means |
5101 | integer --DEMOTION--> integer --FLOAT_EXPR--> floating point. |
5102 | This is safe when the range of the source integer can fit into the lower |
5103 | precision. NARROW_DST is for FIX_TRUNC_EXPR, means |
5104 | floating point --FIX_TRUNC_EXPR--> integer --DEMOTION--> INTEGER. |
5105 | For other conversions, when there's narrowing, NARROW_DST is used as |
5106 | default. */ |
5107 | enum { NARROW_SRC, NARROW_DST, NONE, WIDEN } modifier; |
5108 | vec<tree> vec_oprnds0 = vNULL; |
5109 | vec<tree> vec_oprnds1 = vNULL; |
5110 | tree vop0; |
5111 | bb_vec_info bb_vinfo = dyn_cast <bb_vec_info> (p: vinfo); |
5112 | int multi_step_cvt = 0; |
5113 | vec<tree> interm_types = vNULL; |
5114 | tree intermediate_type, cvt_type = NULL_TREE; |
5115 | int op_type; |
5116 | unsigned short fltsz; |
5117 | |
5118 | /* Is STMT a vectorizable conversion? */ |
5119 | |
5120 | if (!STMT_VINFO_RELEVANT_P (stmt_info) && !bb_vinfo) |
5121 | return false; |
5122 | |
5123 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def |
5124 | && ! vec_stmt) |
5125 | return false; |
5126 | |
5127 | gimple* stmt = stmt_info->stmt; |
5128 | if (!(is_gimple_assign (gs: stmt) || is_gimple_call (gs: stmt))) |
5129 | return false; |
5130 | |
5131 | if (gimple_get_lhs (stmt) == NULL_TREE |
5132 | || TREE_CODE (gimple_get_lhs (stmt)) != SSA_NAME) |
5133 | return false; |
5134 | |
5135 | if (TREE_CODE (gimple_get_lhs (stmt)) != SSA_NAME) |
5136 | return false; |
5137 | |
5138 | if (is_gimple_assign (gs: stmt)) |
5139 | { |
5140 | code = gimple_assign_rhs_code (gs: stmt); |
5141 | op_type = TREE_CODE_LENGTH ((tree_code) code); |
5142 | } |
5143 | else if (gimple_call_internal_p (gs: stmt)) |
5144 | { |
5145 | code = gimple_call_internal_fn (gs: stmt); |
5146 | op_type = gimple_call_num_args (gs: stmt); |
5147 | } |
5148 | else |
5149 | return false; |
5150 | |
5151 | bool widen_arith = (code == WIDEN_MULT_EXPR |
5152 | || code == WIDEN_LSHIFT_EXPR |
5153 | || widening_fn_p (code)); |
5154 | |
5155 | if (!widen_arith |
5156 | && !CONVERT_EXPR_CODE_P (code) |
5157 | && code != FIX_TRUNC_EXPR |
5158 | && code != FLOAT_EXPR) |
5159 | return false; |
5160 | |
5161 | /* Check types of lhs and rhs. */ |
5162 | scalar_dest = gimple_get_lhs (stmt); |
5163 | lhs_type = TREE_TYPE (scalar_dest); |
5164 | vectype_out = STMT_VINFO_VECTYPE (stmt_info); |
5165 | |
5166 | /* Check the operands of the operation. */ |
5167 | slp_tree slp_op0, slp_op1 = NULL; |
5168 | if (!vect_is_simple_use (vinfo, stmt_info, slp_node, |
5169 | 0, &op0, &slp_op0, &dt[0], &vectype_in)) |
5170 | { |
5171 | if (dump_enabled_p ()) |
5172 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
5173 | "use not simple.\n" ); |
5174 | return false; |
5175 | } |
5176 | |
5177 | rhs_type = TREE_TYPE (op0); |
5178 | if ((code != FIX_TRUNC_EXPR && code != FLOAT_EXPR) |
5179 | && !((INTEGRAL_TYPE_P (lhs_type) |
5180 | && INTEGRAL_TYPE_P (rhs_type)) |
5181 | || (SCALAR_FLOAT_TYPE_P (lhs_type) |
5182 | && SCALAR_FLOAT_TYPE_P (rhs_type)))) |
5183 | return false; |
5184 | |
5185 | if (!VECTOR_BOOLEAN_TYPE_P (vectype_out) |
5186 | && ((INTEGRAL_TYPE_P (lhs_type) |
5187 | && !type_has_mode_precision_p (t: lhs_type)) |
5188 | || (INTEGRAL_TYPE_P (rhs_type) |
5189 | && !type_has_mode_precision_p (t: rhs_type)))) |
5190 | { |
5191 | if (dump_enabled_p ()) |
5192 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
5193 | "type conversion to/from bit-precision unsupported." |
5194 | "\n" ); |
5195 | return false; |
5196 | } |
5197 | |
5198 | if (op_type == binary_op) |
5199 | { |
5200 | gcc_assert (code == WIDEN_MULT_EXPR |
5201 | || code == WIDEN_LSHIFT_EXPR |
5202 | || widening_fn_p (code)); |
5203 | |
5204 | op1 = is_gimple_assign (gs: stmt) ? gimple_assign_rhs2 (gs: stmt) : |
5205 | gimple_call_arg (gs: stmt, index: 0); |
5206 | tree vectype1_in; |
5207 | if (!vect_is_simple_use (vinfo, stmt_info, slp_node, 1, |
5208 | &op1, &slp_op1, &dt[1], &vectype1_in)) |
5209 | { |
5210 | if (dump_enabled_p ()) |
5211 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
5212 | "use not simple.\n" ); |
5213 | return false; |
5214 | } |
5215 | /* For WIDEN_MULT_EXPR, if OP0 is a constant, use the type of |
5216 | OP1. */ |
5217 | if (!vectype_in) |
5218 | vectype_in = vectype1_in; |
5219 | } |
5220 | |
5221 | /* If op0 is an external or constant def, infer the vector type |
5222 | from the scalar type. */ |
5223 | if (!vectype_in) |
5224 | vectype_in = get_vectype_for_scalar_type (vinfo, rhs_type, slp_node); |
5225 | if (vec_stmt) |
5226 | gcc_assert (vectype_in); |
5227 | if (!vectype_in) |
5228 | { |
5229 | if (dump_enabled_p ()) |
5230 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
5231 | "no vectype for scalar type %T\n" , rhs_type); |
5232 | |
5233 | return false; |
5234 | } |
5235 | |
5236 | if (VECTOR_BOOLEAN_TYPE_P (vectype_out) |
5237 | && !VECTOR_BOOLEAN_TYPE_P (vectype_in)) |
5238 | { |
5239 | if (dump_enabled_p ()) |
5240 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
5241 | "can't convert between boolean and non " |
5242 | "boolean vectors %T\n" , rhs_type); |
5243 | |
5244 | return false; |
5245 | } |
5246 | |
5247 | nunits_in = TYPE_VECTOR_SUBPARTS (node: vectype_in); |
5248 | nunits_out = TYPE_VECTOR_SUBPARTS (node: vectype_out); |
5249 | if (known_eq (nunits_out, nunits_in)) |
5250 | if (widen_arith) |
5251 | modifier = WIDEN; |
5252 | else |
5253 | modifier = NONE; |
5254 | else if (multiple_p (a: nunits_out, b: nunits_in)) |
5255 | modifier = NARROW_DST; |
5256 | else |
5257 | { |
5258 | gcc_checking_assert (multiple_p (nunits_in, nunits_out)); |
5259 | modifier = WIDEN; |
5260 | } |
5261 | |
5262 | /* Multiple types in SLP are handled by creating the appropriate number of |
5263 | vectorized stmts for each SLP node. Hence, NCOPIES is always 1 in |
5264 | case of SLP. */ |
5265 | if (slp_node) |
5266 | ncopies = 1; |
5267 | else if (modifier == NARROW_DST) |
5268 | ncopies = vect_get_num_copies (loop_vinfo, vectype: vectype_out); |
5269 | else |
5270 | ncopies = vect_get_num_copies (loop_vinfo, vectype: vectype_in); |
5271 | |
5272 | /* Sanity check: make sure that at least one copy of the vectorized stmt |
5273 | needs to be generated. */ |
5274 | gcc_assert (ncopies >= 1); |
5275 | |
5276 | bool found_mode = false; |
5277 | scalar_mode lhs_mode = SCALAR_TYPE_MODE (lhs_type); |
5278 | scalar_mode rhs_mode = SCALAR_TYPE_MODE (rhs_type); |
5279 | opt_scalar_mode rhs_mode_iter; |
5280 | |
5281 | /* Supportable by target? */ |
5282 | switch (modifier) |
5283 | { |
5284 | case NONE: |
5285 | if (code != FIX_TRUNC_EXPR |
5286 | && code != FLOAT_EXPR |
5287 | && !CONVERT_EXPR_CODE_P (code)) |
5288 | return false; |
5289 | gcc_assert (code.is_tree_code ()); |
5290 | if (supportable_convert_operation ((tree_code) code, vectype_out, |
5291 | vectype_in, &tc1)) |
5292 | { |
5293 | code1 = tc1; |
5294 | break; |
5295 | } |
5296 | |
5297 | /* For conversions between float and integer types try whether |
5298 | we can use intermediate signed integer types to support the |
5299 | conversion. */ |
5300 | if (GET_MODE_SIZE (mode: lhs_mode) != GET_MODE_SIZE (mode: rhs_mode) |
5301 | && (code == FLOAT_EXPR || |
5302 | (code == FIX_TRUNC_EXPR && !flag_trapping_math))) |
5303 | { |
5304 | bool demotion = GET_MODE_SIZE (mode: rhs_mode) > GET_MODE_SIZE (mode: lhs_mode); |
5305 | bool float_expr_p = code == FLOAT_EXPR; |
5306 | unsigned short target_size; |
5307 | scalar_mode intermediate_mode; |
5308 | if (demotion) |
5309 | { |
5310 | intermediate_mode = lhs_mode; |
5311 | target_size = GET_MODE_SIZE (mode: rhs_mode); |
5312 | } |
5313 | else |
5314 | { |
5315 | target_size = GET_MODE_SIZE (mode: lhs_mode); |
5316 | if (!int_mode_for_size |
5317 | (size: GET_MODE_BITSIZE (mode: rhs_mode), limit: 0).exists (mode: &intermediate_mode)) |
5318 | goto unsupported; |
5319 | } |
5320 | code1 = float_expr_p ? code : NOP_EXPR; |
5321 | codecvt1 = float_expr_p ? NOP_EXPR : code; |
5322 | opt_scalar_mode mode_iter; |
5323 | FOR_EACH_2XWIDER_MODE (mode_iter, intermediate_mode) |
5324 | { |
5325 | intermediate_mode = mode_iter.require (); |
5326 | |
5327 | if (GET_MODE_SIZE (mode: intermediate_mode) > target_size) |
5328 | break; |
5329 | |
5330 | scalar_mode cvt_mode; |
5331 | if (!int_mode_for_size |
5332 | (size: GET_MODE_BITSIZE (mode: intermediate_mode), limit: 0).exists (mode: &cvt_mode)) |
5333 | break; |
5334 | |
5335 | cvt_type = build_nonstandard_integer_type |
5336 | (GET_MODE_BITSIZE (mode: cvt_mode), 0); |
5337 | |
5338 | /* Check if the intermediate type can hold OP0's range. |
5339 | When converting from float to integer this is not necessary |
5340 | because values that do not fit the (smaller) target type are |
5341 | unspecified anyway. */ |
5342 | if (demotion && float_expr_p) |
5343 | { |
5344 | wide_int op_min_value, op_max_value; |
5345 | if (!vect_get_range_info (op0, &op_min_value, &op_max_value)) |
5346 | break; |
5347 | |
5348 | if (cvt_type == NULL_TREE |
5349 | || (wi::min_precision (x: op_max_value, sgn: SIGNED) |
5350 | > TYPE_PRECISION (cvt_type)) |
5351 | || (wi::min_precision (x: op_min_value, sgn: SIGNED) |
5352 | > TYPE_PRECISION (cvt_type))) |
5353 | continue; |
5354 | } |
5355 | |
5356 | cvt_type = get_vectype_for_scalar_type (vinfo, cvt_type, slp_node); |
5357 | /* This should only happened for SLP as long as loop vectorizer |
5358 | only supports same-sized vector. */ |
5359 | if (cvt_type == NULL_TREE |
5360 | || maybe_ne (a: TYPE_VECTOR_SUBPARTS (node: cvt_type), b: nunits_in) |
5361 | || !supportable_convert_operation ((tree_code) code1, |
5362 | vectype_out, |
5363 | cvt_type, &tc1) |
5364 | || !supportable_convert_operation ((tree_code) codecvt1, |
5365 | cvt_type, |
5366 | vectype_in, &tc2)) |
5367 | continue; |
5368 | |
5369 | found_mode = true; |
5370 | break; |
5371 | } |
5372 | |
5373 | if (found_mode) |
5374 | { |
5375 | multi_step_cvt++; |
5376 | interm_types.safe_push (obj: cvt_type); |
5377 | cvt_type = NULL_TREE; |
5378 | code1 = tc1; |
5379 | codecvt1 = tc2; |
5380 | break; |
5381 | } |
5382 | } |
5383 | /* FALLTHRU */ |
5384 | unsupported: |
5385 | if (dump_enabled_p ()) |
5386 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
5387 | "conversion not supported by target.\n" ); |
5388 | return false; |
5389 | |
5390 | case WIDEN: |
5391 | if (known_eq (nunits_in, nunits_out)) |
5392 | { |
5393 | if (!(code.is_tree_code () |
5394 | && supportable_half_widening_operation ((tree_code) code, |
5395 | vectype_out, vectype_in, |
5396 | &tc1))) |
5397 | goto unsupported; |
5398 | code1 = tc1; |
5399 | gcc_assert (!(multi_step_cvt && op_type == binary_op)); |
5400 | break; |
5401 | } |
5402 | if (supportable_widening_operation (vinfo, code, stmt_info, |
5403 | vectype_out, vectype_in, &code1, |
5404 | &code2, &multi_step_cvt, |
5405 | &interm_types)) |
5406 | { |
5407 | /* Binary widening operation can only be supported directly by the |
5408 | architecture. */ |
5409 | gcc_assert (!(multi_step_cvt && op_type == binary_op)); |
5410 | break; |
5411 | } |
5412 | |
5413 | if (code != FLOAT_EXPR |
5414 | || GET_MODE_SIZE (mode: lhs_mode) <= GET_MODE_SIZE (mode: rhs_mode)) |
5415 | goto unsupported; |
5416 | |
5417 | fltsz = GET_MODE_SIZE (mode: lhs_mode); |
5418 | FOR_EACH_2XWIDER_MODE (rhs_mode_iter, rhs_mode) |
5419 | { |
5420 | rhs_mode = rhs_mode_iter.require (); |
5421 | if (GET_MODE_SIZE (mode: rhs_mode) > fltsz) |
5422 | break; |
5423 | |
5424 | cvt_type |
5425 | = build_nonstandard_integer_type (GET_MODE_BITSIZE (mode: rhs_mode), 0); |
5426 | cvt_type = get_same_sized_vectype (cvt_type, vectype_in); |
5427 | if (cvt_type == NULL_TREE) |
5428 | goto unsupported; |
5429 | |
5430 | if (GET_MODE_SIZE (mode: rhs_mode) == fltsz) |
5431 | { |
5432 | tc1 = ERROR_MARK; |
5433 | gcc_assert (code.is_tree_code ()); |
5434 | if (!supportable_convert_operation ((tree_code) code, vectype_out, |
5435 | cvt_type, &tc1)) |
5436 | goto unsupported; |
5437 | codecvt1 = tc1; |
5438 | } |
5439 | else if (!supportable_widening_operation (vinfo, code, |
5440 | stmt_info, vectype_out, |
5441 | cvt_type, &codecvt1, |
5442 | &codecvt2, &multi_step_cvt, |
5443 | &interm_types)) |
5444 | continue; |
5445 | else |
5446 | gcc_assert (multi_step_cvt == 0); |
5447 | |
5448 | if (supportable_widening_operation (vinfo, NOP_EXPR, stmt_info, |
5449 | cvt_type, |
5450 | vectype_in, &code1, |
5451 | &code2, &multi_step_cvt, |
5452 | &interm_types)) |
5453 | { |
5454 | found_mode = true; |
5455 | break; |
5456 | } |
5457 | } |
5458 | |
5459 | if (!found_mode) |
5460 | goto unsupported; |
5461 | |
5462 | if (GET_MODE_SIZE (mode: rhs_mode) == fltsz) |
5463 | codecvt2 = ERROR_MARK; |
5464 | else |
5465 | { |
5466 | multi_step_cvt++; |
5467 | interm_types.safe_push (obj: cvt_type); |
5468 | cvt_type = NULL_TREE; |
5469 | } |
5470 | break; |
5471 | |
5472 | case NARROW_DST: |
5473 | gcc_assert (op_type == unary_op); |
5474 | if (supportable_narrowing_operation (code, vectype_out, vectype_in, |
5475 | &code1, &multi_step_cvt, |
5476 | &interm_types)) |
5477 | break; |
5478 | |
5479 | if (GET_MODE_SIZE (mode: lhs_mode) >= GET_MODE_SIZE (mode: rhs_mode)) |
5480 | goto unsupported; |
5481 | |
5482 | if (code == FIX_TRUNC_EXPR) |
5483 | { |
5484 | cvt_type |
5485 | = build_nonstandard_integer_type (GET_MODE_BITSIZE (mode: rhs_mode), 0); |
5486 | cvt_type = get_same_sized_vectype (cvt_type, vectype_in); |
5487 | if (cvt_type == NULL_TREE) |
5488 | goto unsupported; |
5489 | if (supportable_convert_operation ((tree_code) code, cvt_type, vectype_in, |
5490 | &tc1)) |
5491 | codecvt1 = tc1; |
5492 | else |
5493 | goto unsupported; |
5494 | if (supportable_narrowing_operation (NOP_EXPR, vectype_out, cvt_type, |
5495 | &code1, &multi_step_cvt, |
5496 | &interm_types)) |
5497 | break; |
5498 | } |
5499 | /* If op0 can be represented with low precision integer, |
5500 | truncate it to cvt_type and the do FLOAT_EXPR. */ |
5501 | else if (code == FLOAT_EXPR) |
5502 | { |
5503 | wide_int op_min_value, op_max_value; |
5504 | if (!vect_get_range_info (op0, &op_min_value, &op_max_value)) |
5505 | goto unsupported; |
5506 | |
5507 | cvt_type |
5508 | = build_nonstandard_integer_type (GET_MODE_BITSIZE (mode: lhs_mode), 0); |
5509 | if (cvt_type == NULL_TREE |
5510 | || (wi::min_precision (x: op_max_value, sgn: SIGNED) |
5511 | > TYPE_PRECISION (cvt_type)) |
5512 | || (wi::min_precision (x: op_min_value, sgn: SIGNED) |
5513 | > TYPE_PRECISION (cvt_type))) |
5514 | goto unsupported; |
5515 | |
5516 | cvt_type = get_same_sized_vectype (cvt_type, vectype_out); |
5517 | if (cvt_type == NULL_TREE) |
5518 | goto unsupported; |
5519 | if (!supportable_narrowing_operation (NOP_EXPR, cvt_type, vectype_in, |
5520 | &code1, &multi_step_cvt, |
5521 | &interm_types)) |
5522 | goto unsupported; |
5523 | if (supportable_convert_operation ((tree_code) code, vectype_out, |
5524 | cvt_type, &tc1)) |
5525 | { |
5526 | codecvt1 = tc1; |
5527 | modifier = NARROW_SRC; |
5528 | break; |
5529 | } |
5530 | } |
5531 | |
5532 | goto unsupported; |
5533 | |
5534 | default: |
5535 | gcc_unreachable (); |
5536 | } |
5537 | |
5538 | if (!vec_stmt) /* transformation not required. */ |
5539 | { |
5540 | if (slp_node |
5541 | && (!vect_maybe_update_slp_op_vectype (slp_op0, vectype_in) |
5542 | || !vect_maybe_update_slp_op_vectype (slp_op1, vectype_in))) |
5543 | { |
5544 | if (dump_enabled_p ()) |
5545 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
5546 | "incompatible vector types for invariants\n" ); |
5547 | return false; |
5548 | } |
5549 | DUMP_VECT_SCOPE ("vectorizable_conversion" ); |
5550 | if (modifier == NONE) |
5551 | { |
5552 | STMT_VINFO_TYPE (stmt_info) = type_conversion_vec_info_type; |
5553 | vect_model_simple_cost (vinfo, stmt_info, |
5554 | ncopies: ncopies * (1 + multi_step_cvt), |
5555 | dt, ndts, node: slp_node, cost_vec); |
5556 | } |
5557 | else if (modifier == NARROW_SRC || modifier == NARROW_DST) |
5558 | { |
5559 | STMT_VINFO_TYPE (stmt_info) = type_demotion_vec_info_type; |
5560 | /* The final packing step produces one vector result per copy. */ |
5561 | unsigned int nvectors |
5562 | = (slp_node ? SLP_TREE_NUMBER_OF_VEC_STMTS (slp_node) : ncopies); |
5563 | vect_model_promotion_demotion_cost (stmt_info, dt, ncopies: nvectors, |
5564 | pwr: multi_step_cvt, cost_vec, |
5565 | widen_arith); |
5566 | } |
5567 | else |
5568 | { |
5569 | STMT_VINFO_TYPE (stmt_info) = type_promotion_vec_info_type; |
5570 | /* The initial unpacking step produces two vector results |
5571 | per copy. MULTI_STEP_CVT is 0 for a single conversion, |
5572 | so >> MULTI_STEP_CVT divides by 2^(number of steps - 1). */ |
5573 | unsigned int nvectors |
5574 | = (slp_node |
5575 | ? SLP_TREE_NUMBER_OF_VEC_STMTS (slp_node) >> multi_step_cvt |
5576 | : ncopies * 2); |
5577 | vect_model_promotion_demotion_cost (stmt_info, dt, ncopies: nvectors, |
5578 | pwr: multi_step_cvt, cost_vec, |
5579 | widen_arith); |
5580 | } |
5581 | interm_types.release (); |
5582 | return true; |
5583 | } |
5584 | |
5585 | /* Transform. */ |
5586 | if (dump_enabled_p ()) |
5587 | dump_printf_loc (MSG_NOTE, vect_location, |
5588 | "transform conversion. ncopies = %d.\n" , ncopies); |
5589 | |
5590 | if (op_type == binary_op) |
5591 | { |
5592 | if (CONSTANT_CLASS_P (op0)) |
5593 | op0 = fold_convert (TREE_TYPE (op1), op0); |
5594 | else if (CONSTANT_CLASS_P (op1)) |
5595 | op1 = fold_convert (TREE_TYPE (op0), op1); |
5596 | } |
5597 | |
5598 | /* In case of multi-step conversion, we first generate conversion operations |
5599 | to the intermediate types, and then from that types to the final one. |
5600 | We create vector destinations for the intermediate type (TYPES) received |
5601 | from supportable_*_operation, and store them in the correct order |
5602 | for future use in vect_create_vectorized_*_stmts (). */ |
5603 | auto_vec<tree> vec_dsts (multi_step_cvt + 1); |
5604 | bool widen_or_narrow_float_p |
5605 | = cvt_type && (modifier == WIDEN || modifier == NARROW_SRC); |
5606 | vec_dest = vect_create_destination_var (scalar_dest, |
5607 | widen_or_narrow_float_p |
5608 | ? cvt_type : vectype_out); |
5609 | vec_dsts.quick_push (obj: vec_dest); |
5610 | |
5611 | if (multi_step_cvt) |
5612 | { |
5613 | for (i = interm_types.length () - 1; |
5614 | interm_types.iterate (ix: i, ptr: &intermediate_type); i--) |
5615 | { |
5616 | vec_dest = vect_create_destination_var (scalar_dest, |
5617 | intermediate_type); |
5618 | vec_dsts.quick_push (obj: vec_dest); |
5619 | } |
5620 | } |
5621 | |
5622 | if (cvt_type) |
5623 | vec_dest = vect_create_destination_var (scalar_dest, |
5624 | widen_or_narrow_float_p |
5625 | ? vectype_out : cvt_type); |
5626 | |
5627 | int ninputs = 1; |
5628 | if (!slp_node) |
5629 | { |
5630 | if (modifier == WIDEN) |
5631 | ; |
5632 | else if (modifier == NARROW_SRC || modifier == NARROW_DST) |
5633 | { |
5634 | if (multi_step_cvt) |
5635 | ninputs = vect_pow2 (x: multi_step_cvt); |
5636 | ninputs *= 2; |
5637 | } |
5638 | } |
5639 | |
5640 | switch (modifier) |
5641 | { |
5642 | case NONE: |
5643 | vect_get_vec_defs (vinfo, stmt_info, slp_node, ncopies, |
5644 | op0, vec_oprnds0: &vec_oprnds0); |
5645 | /* vec_dest is intermediate type operand when multi_step_cvt. */ |
5646 | if (multi_step_cvt) |
5647 | { |
5648 | cvt_op = vec_dest; |
5649 | vec_dest = vec_dsts[0]; |
5650 | } |
5651 | |
5652 | FOR_EACH_VEC_ELT (vec_oprnds0, i, vop0) |
5653 | { |
5654 | /* Arguments are ready, create the new vector stmt. */ |
5655 | gimple* new_stmt; |
5656 | if (multi_step_cvt) |
5657 | { |
5658 | gcc_assert (multi_step_cvt == 1); |
5659 | new_stmt = vect_gimple_build (cvt_op, codecvt1, vop0); |
5660 | new_temp = make_ssa_name (var: cvt_op, stmt: new_stmt); |
5661 | gimple_assign_set_lhs (gs: new_stmt, lhs: new_temp); |
5662 | vect_finish_stmt_generation (vinfo, stmt_info, vec_stmt: new_stmt, gsi); |
5663 | vop0 = new_temp; |
5664 | } |
5665 | new_stmt = vect_gimple_build (vec_dest, code1, vop0); |
5666 | new_temp = make_ssa_name (var: vec_dest, stmt: new_stmt); |
5667 | gimple_set_lhs (new_stmt, new_temp); |
5668 | vect_finish_stmt_generation (vinfo, stmt_info, vec_stmt: new_stmt, gsi); |
5669 | |
5670 | if (slp_node) |
5671 | slp_node->push_vec_def (def: new_stmt); |
5672 | else |
5673 | STMT_VINFO_VEC_STMTS (stmt_info).safe_push (obj: new_stmt); |
5674 | } |
5675 | break; |
5676 | |
5677 | case WIDEN: |
5678 | /* In case the vectorization factor (VF) is bigger than the number |
5679 | of elements that we can fit in a vectype (nunits), we have to |
5680 | generate more than one vector stmt - i.e - we need to "unroll" |
5681 | the vector stmt by a factor VF/nunits. */ |
5682 | vect_get_vec_defs (vinfo, stmt_info, slp_node, ncopies: ncopies * ninputs, |
5683 | op0, vec_oprnds0: &vec_oprnds0, |
5684 | op1: code == WIDEN_LSHIFT_EXPR ? NULL_TREE : op1, |
5685 | vec_oprnds1: &vec_oprnds1); |
5686 | if (code == WIDEN_LSHIFT_EXPR) |
5687 | { |
5688 | int oprnds_size = vec_oprnds0.length (); |
5689 | vec_oprnds1.create (nelems: oprnds_size); |
5690 | for (i = 0; i < oprnds_size; ++i) |
5691 | vec_oprnds1.quick_push (obj: op1); |
5692 | } |
5693 | /* Arguments are ready. Create the new vector stmts. */ |
5694 | for (i = multi_step_cvt; i >= 0; i--) |
5695 | { |
5696 | tree this_dest = vec_dsts[i]; |
5697 | code_helper c1 = code1, c2 = code2; |
5698 | if (i == 0 && codecvt2 != ERROR_MARK) |
5699 | { |
5700 | c1 = codecvt1; |
5701 | c2 = codecvt2; |
5702 | } |
5703 | if (known_eq (nunits_out, nunits_in)) |
5704 | vect_create_half_widening_stmts (vinfo, vec_oprnds0: &vec_oprnds0, vec_oprnds1: &vec_oprnds1, |
5705 | stmt_info, vec_dest: this_dest, gsi, code1: c1, |
5706 | op_type); |
5707 | else |
5708 | vect_create_vectorized_promotion_stmts (vinfo, vec_oprnds0: &vec_oprnds0, |
5709 | vec_oprnds1: &vec_oprnds1, stmt_info, |
5710 | vec_dest: this_dest, gsi, |
5711 | ch1: c1, ch2: c2, op_type); |
5712 | } |
5713 | |
5714 | FOR_EACH_VEC_ELT (vec_oprnds0, i, vop0) |
5715 | { |
5716 | gimple *new_stmt; |
5717 | if (cvt_type) |
5718 | { |
5719 | new_temp = make_ssa_name (var: vec_dest); |
5720 | new_stmt = vect_gimple_build (new_temp, codecvt1, vop0); |
5721 | vect_finish_stmt_generation (vinfo, stmt_info, vec_stmt: new_stmt, gsi); |
5722 | } |
5723 | else |
5724 | new_stmt = SSA_NAME_DEF_STMT (vop0); |
5725 | |
5726 | if (slp_node) |
5727 | slp_node->push_vec_def (def: new_stmt); |
5728 | else |
5729 | STMT_VINFO_VEC_STMTS (stmt_info).safe_push (obj: new_stmt); |
5730 | } |
5731 | break; |
5732 | |
5733 | case NARROW_SRC: |
5734 | case NARROW_DST: |
5735 | /* In case the vectorization factor (VF) is bigger than the number |
5736 | of elements that we can fit in a vectype (nunits), we have to |
5737 | generate more than one vector stmt - i.e - we need to "unroll" |
5738 | the vector stmt by a factor VF/nunits. */ |
5739 | vect_get_vec_defs (vinfo, stmt_info, slp_node, ncopies: ncopies * ninputs, |
5740 | op0, vec_oprnds0: &vec_oprnds0); |
5741 | /* Arguments are ready. Create the new vector stmts. */ |
5742 | if (cvt_type && modifier == NARROW_DST) |
5743 | FOR_EACH_VEC_ELT (vec_oprnds0, i, vop0) |
5744 | { |
5745 | new_temp = make_ssa_name (var: vec_dest); |
5746 | gimple *new_stmt = vect_gimple_build (new_temp, codecvt1, vop0); |
5747 | vect_finish_stmt_generation (vinfo, stmt_info, vec_stmt: new_stmt, gsi); |
5748 | vec_oprnds0[i] = new_temp; |
5749 | } |
5750 | |
5751 | vect_create_vectorized_demotion_stmts (vinfo, vec_oprnds: &vec_oprnds0, |
5752 | multi_step_cvt, |
5753 | stmt_info, vec_dsts, gsi, |
5754 | slp_node, code: code1, |
5755 | narrow_src_p: modifier == NARROW_SRC); |
5756 | /* After demoting op0 to cvt_type, convert it to dest. */ |
5757 | if (cvt_type && code == FLOAT_EXPR) |
5758 | { |
5759 | for (unsigned int i = 0; i != vec_oprnds0.length() / 2; i++) |
5760 | { |
5761 | /* Arguments are ready, create the new vector stmt. */ |
5762 | gcc_assert (TREE_CODE_LENGTH ((tree_code) codecvt1) == unary_op); |
5763 | gimple *new_stmt |
5764 | = vect_gimple_build (vec_dest, codecvt1, vec_oprnds0[i]); |
5765 | new_temp = make_ssa_name (var: vec_dest, stmt: new_stmt); |
5766 | gimple_set_lhs (new_stmt, new_temp); |
5767 | vect_finish_stmt_generation (vinfo, stmt_info, vec_stmt: new_stmt, gsi); |
5768 | |
5769 | /* This is the last step of the conversion sequence. Store the |
5770 | vectors in SLP_NODE or in vector info of the scalar statement |
5771 | (or in STMT_VINFO_RELATED_STMT chain). */ |
5772 | if (slp_node) |
5773 | slp_node->push_vec_def (def: new_stmt); |
5774 | else |
5775 | STMT_VINFO_VEC_STMTS (stmt_info).safe_push (obj: new_stmt); |
5776 | } |
5777 | } |
5778 | break; |
5779 | } |
5780 | if (!slp_node) |
5781 | *vec_stmt = STMT_VINFO_VEC_STMTS (stmt_info)[0]; |
5782 | |
5783 | vec_oprnds0.release (); |
5784 | vec_oprnds1.release (); |
5785 | interm_types.release (); |
5786 | |
5787 | return true; |
5788 | } |
5789 | |
5790 | /* Return true if we can assume from the scalar form of STMT_INFO that |
5791 | neither the scalar nor the vector forms will generate code. STMT_INFO |
5792 | is known not to involve a data reference. */ |
5793 | |
5794 | bool |
5795 | vect_nop_conversion_p (stmt_vec_info stmt_info) |
5796 | { |
5797 | gassign *stmt = dyn_cast <gassign *> (p: stmt_info->stmt); |
5798 | if (!stmt) |
5799 | return false; |
5800 | |
5801 | tree lhs = gimple_assign_lhs (gs: stmt); |
5802 | tree_code code = gimple_assign_rhs_code (gs: stmt); |
5803 | tree rhs = gimple_assign_rhs1 (gs: stmt); |
5804 | |
5805 | if (code == SSA_NAME || code == VIEW_CONVERT_EXPR) |
5806 | return true; |
5807 | |
5808 | if (CONVERT_EXPR_CODE_P (code)) |
5809 | return tree_nop_conversion_p (TREE_TYPE (lhs), TREE_TYPE (rhs)); |
5810 | |
5811 | return false; |
5812 | } |
5813 | |
5814 | /* Function vectorizable_assignment. |
5815 | |
5816 | Check if STMT_INFO performs an assignment (copy) that can be vectorized. |
5817 | If VEC_STMT is also passed, vectorize the STMT_INFO: create a vectorized |
5818 | stmt to replace it, put it in VEC_STMT, and insert it at GSI. |
5819 | Return true if STMT_INFO is vectorizable in this way. */ |
5820 | |
5821 | static bool |
5822 | vectorizable_assignment (vec_info *vinfo, |
5823 | stmt_vec_info stmt_info, gimple_stmt_iterator *gsi, |
5824 | gimple **vec_stmt, slp_tree slp_node, |
5825 | stmt_vector_for_cost *cost_vec) |
5826 | { |
5827 | tree vec_dest; |
5828 | tree scalar_dest; |
5829 | tree op; |
5830 | loop_vec_info loop_vinfo = dyn_cast <loop_vec_info> (p: vinfo); |
5831 | tree new_temp; |
5832 | enum vect_def_type dt[1] = {vect_unknown_def_type}; |
5833 | int ndts = 1; |
5834 | int ncopies; |
5835 | int i; |
5836 | vec<tree> vec_oprnds = vNULL; |
5837 | tree vop; |
5838 | bb_vec_info bb_vinfo = dyn_cast <bb_vec_info> (p: vinfo); |
5839 | enum tree_code code; |
5840 | tree vectype_in; |
5841 | |
5842 | if (!STMT_VINFO_RELEVANT_P (stmt_info) && !bb_vinfo) |
5843 | return false; |
5844 | |
5845 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def |
5846 | && ! vec_stmt) |
5847 | return false; |
5848 | |
5849 | /* Is vectorizable assignment? */ |
5850 | gassign *stmt = dyn_cast <gassign *> (p: stmt_info->stmt); |
5851 | if (!stmt) |
5852 | return false; |
5853 | |
5854 | scalar_dest = gimple_assign_lhs (gs: stmt); |
5855 | if (TREE_CODE (scalar_dest) != SSA_NAME) |
5856 | return false; |
5857 | |
5858 | if (STMT_VINFO_DATA_REF (stmt_info)) |
5859 | return false; |
5860 | |
5861 | code = gimple_assign_rhs_code (gs: stmt); |
5862 | if (!(gimple_assign_single_p (gs: stmt) |
5863 | || code == PAREN_EXPR |
5864 | || CONVERT_EXPR_CODE_P (code))) |
5865 | return false; |
5866 | |
5867 | tree vectype = STMT_VINFO_VECTYPE (stmt_info); |
5868 | poly_uint64 nunits = TYPE_VECTOR_SUBPARTS (node: vectype); |
5869 | |
5870 | /* Multiple types in SLP are handled by creating the appropriate number of |
5871 | vectorized stmts for each SLP node. Hence, NCOPIES is always 1 in |
5872 | case of SLP. */ |
5873 | if (slp_node) |
5874 | ncopies = 1; |
5875 | else |
5876 | ncopies = vect_get_num_copies (loop_vinfo, vectype); |
5877 | |
5878 | gcc_assert (ncopies >= 1); |
5879 | |
5880 | slp_tree slp_op; |
5881 | if (!vect_is_simple_use (vinfo, stmt_info, slp_node, 0, &op, &slp_op, |
5882 | &dt[0], &vectype_in)) |
5883 | { |
5884 | if (dump_enabled_p ()) |
5885 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
5886 | "use not simple.\n" ); |
5887 | return false; |
5888 | } |
5889 | if (!vectype_in) |
5890 | vectype_in = get_vectype_for_scalar_type (vinfo, TREE_TYPE (op), slp_node); |
5891 | |
5892 | /* We can handle NOP_EXPR conversions that do not change the number |
5893 | of elements or the vector size. */ |
5894 | if ((CONVERT_EXPR_CODE_P (code) |
5895 | || code == VIEW_CONVERT_EXPR) |
5896 | && (!vectype_in |
5897 | || maybe_ne (a: TYPE_VECTOR_SUBPARTS (node: vectype_in), b: nunits) |
5898 | || maybe_ne (a: GET_MODE_SIZE (TYPE_MODE (vectype)), |
5899 | b: GET_MODE_SIZE (TYPE_MODE (vectype_in))))) |
5900 | return false; |
5901 | |
5902 | if (VECTOR_BOOLEAN_TYPE_P (vectype) != VECTOR_BOOLEAN_TYPE_P (vectype_in)) |
5903 | { |
5904 | if (dump_enabled_p ()) |
5905 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
5906 | "can't convert between boolean and non " |
5907 | "boolean vectors %T\n" , TREE_TYPE (op)); |
5908 | |
5909 | return false; |
5910 | } |
5911 | |
5912 | /* We do not handle bit-precision changes. */ |
5913 | if ((CONVERT_EXPR_CODE_P (code) |
5914 | || code == VIEW_CONVERT_EXPR) |
5915 | && ((INTEGRAL_TYPE_P (TREE_TYPE (scalar_dest)) |
5916 | && !type_has_mode_precision_p (TREE_TYPE (scalar_dest))) |
5917 | || (INTEGRAL_TYPE_P (TREE_TYPE (op)) |
5918 | && !type_has_mode_precision_p (TREE_TYPE (op)))) |
5919 | /* But a conversion that does not change the bit-pattern is ok. */ |
5920 | && !(INTEGRAL_TYPE_P (TREE_TYPE (scalar_dest)) |
5921 | && INTEGRAL_TYPE_P (TREE_TYPE (op)) |
5922 | && (((TYPE_PRECISION (TREE_TYPE (scalar_dest)) |
5923 | > TYPE_PRECISION (TREE_TYPE (op))) |
5924 | && TYPE_UNSIGNED (TREE_TYPE (op))) |
5925 | || (TYPE_PRECISION (TREE_TYPE (scalar_dest)) |
5926 | == TYPE_PRECISION (TREE_TYPE (op)))))) |
5927 | { |
5928 | if (dump_enabled_p ()) |
5929 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
5930 | "type conversion to/from bit-precision " |
5931 | "unsupported.\n" ); |
5932 | return false; |
5933 | } |
5934 | |
5935 | if (!vec_stmt) /* transformation not required. */ |
5936 | { |
5937 | if (slp_node |
5938 | && !vect_maybe_update_slp_op_vectype (slp_op, vectype_in)) |
5939 | { |
5940 | if (dump_enabled_p ()) |
5941 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
5942 | "incompatible vector types for invariants\n" ); |
5943 | return false; |
5944 | } |
5945 | STMT_VINFO_TYPE (stmt_info) = assignment_vec_info_type; |
5946 | DUMP_VECT_SCOPE ("vectorizable_assignment" ); |
5947 | if (!vect_nop_conversion_p (stmt_info)) |
5948 | vect_model_simple_cost (vinfo, stmt_info, ncopies, dt, ndts, node: slp_node, |
5949 | cost_vec); |
5950 | return true; |
5951 | } |
5952 | |
5953 | /* Transform. */ |
5954 | if (dump_enabled_p ()) |
5955 | dump_printf_loc (MSG_NOTE, vect_location, "transform assignment.\n" ); |
5956 | |
5957 | /* Handle def. */ |
5958 | vec_dest = vect_create_destination_var (scalar_dest, vectype); |
5959 | |
5960 | /* Handle use. */ |
5961 | vect_get_vec_defs (vinfo, stmt_info, slp_node, ncopies, op0: op, vec_oprnds0: &vec_oprnds); |
5962 | |
5963 | /* Arguments are ready. create the new vector stmt. */ |
5964 | FOR_EACH_VEC_ELT (vec_oprnds, i, vop) |
5965 | { |
5966 | if (CONVERT_EXPR_CODE_P (code) |
5967 | || code == VIEW_CONVERT_EXPR) |
5968 | vop = build1 (VIEW_CONVERT_EXPR, vectype, vop); |
5969 | gassign *new_stmt = gimple_build_assign (vec_dest, vop); |
5970 | new_temp = make_ssa_name (var: vec_dest, stmt: new_stmt); |
5971 | gimple_assign_set_lhs (gs: new_stmt, lhs: new_temp); |
5972 | vect_finish_stmt_generation (vinfo, stmt_info, vec_stmt: new_stmt, gsi); |
5973 | if (slp_node) |
5974 | slp_node->push_vec_def (def: new_stmt); |
5975 | else |
5976 | STMT_VINFO_VEC_STMTS (stmt_info).safe_push (obj: new_stmt); |
5977 | } |
5978 | if (!slp_node) |
5979 | *vec_stmt = STMT_VINFO_VEC_STMTS (stmt_info)[0]; |
5980 | |
5981 | vec_oprnds.release (); |
5982 | return true; |
5983 | } |
5984 | |
5985 | |
5986 | /* Return TRUE if CODE (a shift operation) is supported for SCALAR_TYPE |
5987 | either as shift by a scalar or by a vector. */ |
5988 | |
5989 | bool |
5990 | vect_supportable_shift (vec_info *vinfo, enum tree_code code, tree scalar_type) |
5991 | { |
5992 | |
5993 | machine_mode vec_mode; |
5994 | optab optab; |
5995 | int icode; |
5996 | tree vectype; |
5997 | |
5998 | vectype = get_vectype_for_scalar_type (vinfo, scalar_type); |
5999 | if (!vectype) |
6000 | return false; |
6001 | |
6002 | optab = optab_for_tree_code (code, vectype, optab_scalar); |
6003 | if (!optab |
6004 | || optab_handler (op: optab, TYPE_MODE (vectype)) == CODE_FOR_nothing) |
6005 | { |
6006 | optab = optab_for_tree_code (code, vectype, optab_vector); |
6007 | if (!optab |
6008 | || (optab_handler (op: optab, TYPE_MODE (vectype)) |
6009 | == CODE_FOR_nothing)) |
6010 | return false; |
6011 | } |
6012 | |
6013 | vec_mode = TYPE_MODE (vectype); |
6014 | icode = (int) optab_handler (op: optab, mode: vec_mode); |
6015 | if (icode == CODE_FOR_nothing) |
6016 | return false; |
6017 | |
6018 | return true; |
6019 | } |
6020 | |
6021 | |
6022 | /* Function vectorizable_shift. |
6023 | |
6024 | Check if STMT_INFO performs a shift operation that can be vectorized. |
6025 | If VEC_STMT is also passed, vectorize the STMT_INFO: create a vectorized |
6026 | stmt to replace it, put it in VEC_STMT, and insert it at GSI. |
6027 | Return true if STMT_INFO is vectorizable in this way. */ |
6028 | |
6029 | static bool |
6030 | vectorizable_shift (vec_info *vinfo, |
6031 | stmt_vec_info stmt_info, gimple_stmt_iterator *gsi, |
6032 | gimple **vec_stmt, slp_tree slp_node, |
6033 | stmt_vector_for_cost *cost_vec) |
6034 | { |
6035 | tree vec_dest; |
6036 | tree scalar_dest; |
6037 | tree op0, op1 = NULL; |
6038 | tree vec_oprnd1 = NULL_TREE; |
6039 | tree vectype; |
6040 | loop_vec_info loop_vinfo = dyn_cast <loop_vec_info> (p: vinfo); |
6041 | enum tree_code code; |
6042 | machine_mode vec_mode; |
6043 | tree new_temp; |
6044 | optab optab; |
6045 | int icode; |
6046 | machine_mode optab_op2_mode; |
6047 | enum vect_def_type dt[2] = {vect_unknown_def_type, vect_unknown_def_type}; |
6048 | int ndts = 2; |
6049 | poly_uint64 nunits_in; |
6050 | poly_uint64 nunits_out; |
6051 | tree vectype_out; |
6052 | tree op1_vectype; |
6053 | int ncopies; |
6054 | int i; |
6055 | vec<tree> vec_oprnds0 = vNULL; |
6056 | vec<tree> vec_oprnds1 = vNULL; |
6057 | tree vop0, vop1; |
6058 | unsigned int k; |
6059 | bool scalar_shift_arg = true; |
6060 | bb_vec_info bb_vinfo = dyn_cast <bb_vec_info> (p: vinfo); |
6061 | bool incompatible_op1_vectype_p = false; |
6062 | |
6063 | if (!STMT_VINFO_RELEVANT_P (stmt_info) && !bb_vinfo) |
6064 | return false; |
6065 | |
6066 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def |
6067 | && STMT_VINFO_DEF_TYPE (stmt_info) != vect_nested_cycle |
6068 | && ! vec_stmt) |
6069 | return false; |
6070 | |
6071 | /* Is STMT a vectorizable binary/unary operation? */ |
6072 | gassign *stmt = dyn_cast <gassign *> (p: stmt_info->stmt); |
6073 | if (!stmt) |
6074 | return false; |
6075 | |
6076 | if (TREE_CODE (gimple_assign_lhs (stmt)) != SSA_NAME) |
6077 | return false; |
6078 | |
6079 | code = gimple_assign_rhs_code (gs: stmt); |
6080 | |
6081 | if (!(code == LSHIFT_EXPR || code == RSHIFT_EXPR || code == LROTATE_EXPR |
6082 | || code == RROTATE_EXPR)) |
6083 | return false; |
6084 | |
6085 | scalar_dest = gimple_assign_lhs (gs: stmt); |
6086 | vectype_out = STMT_VINFO_VECTYPE (stmt_info); |
6087 | if (!type_has_mode_precision_p (TREE_TYPE (scalar_dest))) |
6088 | { |
6089 | if (dump_enabled_p ()) |
6090 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
6091 | "bit-precision shifts not supported.\n" ); |
6092 | return false; |
6093 | } |
6094 | |
6095 | slp_tree slp_op0; |
6096 | if (!vect_is_simple_use (vinfo, stmt_info, slp_node, |
6097 | 0, &op0, &slp_op0, &dt[0], &vectype)) |
6098 | { |
6099 | if (dump_enabled_p ()) |
6100 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
6101 | "use not simple.\n" ); |
6102 | return false; |
6103 | } |
6104 | /* If op0 is an external or constant def, infer the vector type |
6105 | from the scalar type. */ |
6106 | if (!vectype) |
6107 | vectype = get_vectype_for_scalar_type (vinfo, TREE_TYPE (op0), slp_node); |
6108 | if (vec_stmt) |
6109 | gcc_assert (vectype); |
6110 | if (!vectype) |
6111 | { |
6112 | if (dump_enabled_p ()) |
6113 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
6114 | "no vectype for scalar type\n" ); |
6115 | return false; |
6116 | } |
6117 | |
6118 | nunits_out = TYPE_VECTOR_SUBPARTS (node: vectype_out); |
6119 | nunits_in = TYPE_VECTOR_SUBPARTS (node: vectype); |
6120 | if (maybe_ne (a: nunits_out, b: nunits_in)) |
6121 | return false; |
6122 | |
6123 | stmt_vec_info op1_def_stmt_info; |
6124 | slp_tree slp_op1; |
6125 | if (!vect_is_simple_use (vinfo, stmt_info, slp_node, 1, &op1, &slp_op1, |
6126 | &dt[1], &op1_vectype, &op1_def_stmt_info)) |
6127 | { |
6128 | if (dump_enabled_p ()) |
6129 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
6130 | "use not simple.\n" ); |
6131 | return false; |
6132 | } |
6133 | |
6134 | /* Multiple types in SLP are handled by creating the appropriate number of |
6135 | vectorized stmts for each SLP node. Hence, NCOPIES is always 1 in |
6136 | case of SLP. */ |
6137 | if (slp_node) |
6138 | ncopies = 1; |
6139 | else |
6140 | ncopies = vect_get_num_copies (loop_vinfo, vectype); |
6141 | |
6142 | gcc_assert (ncopies >= 1); |
6143 | |
6144 | /* Determine whether the shift amount is a vector, or scalar. If the |
6145 | shift/rotate amount is a vector, use the vector/vector shift optabs. */ |
6146 | |
6147 | if ((dt[1] == vect_internal_def |
6148 | || dt[1] == vect_induction_def |
6149 | || dt[1] == vect_nested_cycle) |
6150 | && !slp_node) |
6151 | scalar_shift_arg = false; |
6152 | else if (dt[1] == vect_constant_def |
6153 | || dt[1] == vect_external_def |
6154 | || dt[1] == vect_internal_def) |
6155 | { |
6156 | /* In SLP, need to check whether the shift count is the same, |
6157 | in loops if it is a constant or invariant, it is always |
6158 | a scalar shift. */ |
6159 | if (slp_node) |
6160 | { |
6161 | vec<stmt_vec_info> stmts = SLP_TREE_SCALAR_STMTS (slp_node); |
6162 | stmt_vec_info slpstmt_info; |
6163 | |
6164 | FOR_EACH_VEC_ELT (stmts, k, slpstmt_info) |
6165 | { |
6166 | gassign *slpstmt = as_a <gassign *> (p: slpstmt_info->stmt); |
6167 | if (!operand_equal_p (gimple_assign_rhs2 (gs: slpstmt), op1, flags: 0)) |
6168 | scalar_shift_arg = false; |
6169 | } |
6170 | |
6171 | /* For internal SLP defs we have to make sure we see scalar stmts |
6172 | for all vector elements. |
6173 | ??? For different vectors we could resort to a different |
6174 | scalar shift operand but code-generation below simply always |
6175 | takes the first. */ |
6176 | if (dt[1] == vect_internal_def |
6177 | && maybe_ne (a: nunits_out * SLP_TREE_NUMBER_OF_VEC_STMTS (slp_node), |
6178 | b: stmts.length ())) |
6179 | scalar_shift_arg = false; |
6180 | } |
6181 | |
6182 | /* If the shift amount is computed by a pattern stmt we cannot |
6183 | use the scalar amount directly thus give up and use a vector |
6184 | shift. */ |
6185 | if (op1_def_stmt_info && is_pattern_stmt_p (stmt_info: op1_def_stmt_info)) |
6186 | scalar_shift_arg = false; |
6187 | } |
6188 | else |
6189 | { |
6190 | if (dump_enabled_p ()) |
6191 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
6192 | "operand mode requires invariant argument.\n" ); |
6193 | return false; |
6194 | } |
6195 | |
6196 | /* Vector shifted by vector. */ |
6197 | bool was_scalar_shift_arg = scalar_shift_arg; |
6198 | if (!scalar_shift_arg) |
6199 | { |
6200 | optab = optab_for_tree_code (code, vectype, optab_vector); |
6201 | if (dump_enabled_p ()) |
6202 | dump_printf_loc (MSG_NOTE, vect_location, |
6203 | "vector/vector shift/rotate found.\n" ); |
6204 | |
6205 | if (!op1_vectype) |
6206 | op1_vectype = get_vectype_for_scalar_type (vinfo, TREE_TYPE (op1), |
6207 | slp_op1); |
6208 | incompatible_op1_vectype_p |
6209 | = (op1_vectype == NULL_TREE |
6210 | || maybe_ne (a: TYPE_VECTOR_SUBPARTS (node: op1_vectype), |
6211 | b: TYPE_VECTOR_SUBPARTS (node: vectype)) |
6212 | || TYPE_MODE (op1_vectype) != TYPE_MODE (vectype)); |
6213 | if (incompatible_op1_vectype_p |
6214 | && (!slp_node |
6215 | || SLP_TREE_DEF_TYPE (slp_op1) != vect_constant_def |
6216 | || slp_op1->refcnt != 1)) |
6217 | { |
6218 | if (dump_enabled_p ()) |
6219 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
6220 | "unusable type for last operand in" |
6221 | " vector/vector shift/rotate.\n" ); |
6222 | return false; |
6223 | } |
6224 | } |
6225 | /* See if the machine has a vector shifted by scalar insn and if not |
6226 | then see if it has a vector shifted by vector insn. */ |
6227 | else |
6228 | { |
6229 | optab = optab_for_tree_code (code, vectype, optab_scalar); |
6230 | if (optab |
6231 | && optab_handler (op: optab, TYPE_MODE (vectype)) != CODE_FOR_nothing) |
6232 | { |
6233 | if (dump_enabled_p ()) |
6234 | dump_printf_loc (MSG_NOTE, vect_location, |
6235 | "vector/scalar shift/rotate found.\n" ); |
6236 | } |
6237 | else |
6238 | { |
6239 | optab = optab_for_tree_code (code, vectype, optab_vector); |
6240 | if (optab |
6241 | && (optab_handler (op: optab, TYPE_MODE (vectype)) |
6242 | != CODE_FOR_nothing)) |
6243 | { |
6244 | scalar_shift_arg = false; |
6245 | |
6246 | if (dump_enabled_p ()) |
6247 | dump_printf_loc (MSG_NOTE, vect_location, |
6248 | "vector/vector shift/rotate found.\n" ); |
6249 | |
6250 | if (!op1_vectype) |
6251 | op1_vectype = get_vectype_for_scalar_type (vinfo, |
6252 | TREE_TYPE (op1), |
6253 | slp_op1); |
6254 | |
6255 | /* Unlike the other binary operators, shifts/rotates have |
6256 | the rhs being int, instead of the same type as the lhs, |
6257 | so make sure the scalar is the right type if we are |
6258 | dealing with vectors of long long/long/short/char. */ |
6259 | incompatible_op1_vectype_p |
6260 | = (!op1_vectype |
6261 | || !tree_nop_conversion_p (TREE_TYPE (vectype), |
6262 | TREE_TYPE (op1))); |
6263 | if (incompatible_op1_vectype_p |
6264 | && dt[1] == vect_internal_def) |
6265 | { |
6266 | if (dump_enabled_p ()) |
6267 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
6268 | "unusable type for last operand in" |
6269 | " vector/vector shift/rotate.\n" ); |
6270 | return false; |
6271 | } |
6272 | } |
6273 | } |
6274 | } |
6275 | |
6276 | /* Supportable by target? */ |
6277 | if (!optab) |
6278 | { |
6279 | if (dump_enabled_p ()) |
6280 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
6281 | "no optab.\n" ); |
6282 | return false; |
6283 | } |
6284 | vec_mode = TYPE_MODE (vectype); |
6285 | icode = (int) optab_handler (op: optab, mode: vec_mode); |
6286 | if (icode == CODE_FOR_nothing) |
6287 | { |
6288 | if (dump_enabled_p ()) |
6289 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
6290 | "op not supported by target.\n" ); |
6291 | return false; |
6292 | } |
6293 | /* vector lowering cannot optimize vector shifts using word arithmetic. */ |
6294 | if (vect_emulated_vector_p (vectype)) |
6295 | return false; |
6296 | |
6297 | if (!vec_stmt) /* transformation not required. */ |
6298 | { |
6299 | if (slp_node |
6300 | && (!vect_maybe_update_slp_op_vectype (slp_op0, vectype) |
6301 | || ((!scalar_shift_arg || dt[1] == vect_internal_def) |
6302 | && (!incompatible_op1_vectype_p |
6303 | || dt[1] == vect_constant_def) |
6304 | && !vect_maybe_update_slp_op_vectype |
6305 | (slp_op1, |
6306 | incompatible_op1_vectype_p ? vectype : op1_vectype)))) |
6307 | { |
6308 | if (dump_enabled_p ()) |
6309 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
6310 | "incompatible vector types for invariants\n" ); |
6311 | return false; |
6312 | } |
6313 | /* Now adjust the constant shift amount in place. */ |
6314 | if (slp_node |
6315 | && incompatible_op1_vectype_p |
6316 | && dt[1] == vect_constant_def) |
6317 | { |
6318 | for (unsigned i = 0; |
6319 | i < SLP_TREE_SCALAR_OPS (slp_op1).length (); ++i) |
6320 | { |
6321 | SLP_TREE_SCALAR_OPS (slp_op1)[i] |
6322 | = fold_convert (TREE_TYPE (vectype), |
6323 | SLP_TREE_SCALAR_OPS (slp_op1)[i]); |
6324 | gcc_assert ((TREE_CODE (SLP_TREE_SCALAR_OPS (slp_op1)[i]) |
6325 | == INTEGER_CST)); |
6326 | } |
6327 | } |
6328 | STMT_VINFO_TYPE (stmt_info) = shift_vec_info_type; |
6329 | DUMP_VECT_SCOPE ("vectorizable_shift" ); |
6330 | vect_model_simple_cost (vinfo, stmt_info, ncopies, dt, |
6331 | ndts: scalar_shift_arg ? 1 : ndts, node: slp_node, cost_vec); |
6332 | return true; |
6333 | } |
6334 | |
6335 | /* Transform. */ |
6336 | |
6337 | if (dump_enabled_p ()) |
6338 | dump_printf_loc (MSG_NOTE, vect_location, |
6339 | "transform binary/unary operation.\n" ); |
6340 | |
6341 | if (incompatible_op1_vectype_p && !slp_node) |
6342 | { |
6343 | gcc_assert (!scalar_shift_arg && was_scalar_shift_arg); |
6344 | op1 = fold_convert (TREE_TYPE (vectype), op1); |
6345 | if (dt[1] != vect_constant_def) |
6346 | op1 = vect_init_vector (vinfo, stmt_info, val: op1, |
6347 | TREE_TYPE (vectype), NULL); |
6348 | } |
6349 | |
6350 | /* Handle def. */ |
6351 | vec_dest = vect_create_destination_var (scalar_dest, vectype); |
6352 | |
6353 | if (scalar_shift_arg && dt[1] != vect_internal_def) |
6354 | { |
6355 | /* Vector shl and shr insn patterns can be defined with scalar |
6356 | operand 2 (shift operand). In this case, use constant or loop |
6357 | invariant op1 directly, without extending it to vector mode |
6358 | first. */ |
6359 | optab_op2_mode = insn_data[icode].operand[2].mode; |
6360 | if (!VECTOR_MODE_P (optab_op2_mode)) |
6361 | { |
6362 | if (dump_enabled_p ()) |
6363 | dump_printf_loc (MSG_NOTE, vect_location, |
6364 | "operand 1 using scalar mode.\n" ); |
6365 | vec_oprnd1 = op1; |
6366 | vec_oprnds1.create (nelems: slp_node ? slp_node->vec_stmts_size : ncopies); |
6367 | vec_oprnds1.quick_push (obj: vec_oprnd1); |
6368 | /* Store vec_oprnd1 for every vector stmt to be created. |
6369 | We check during the analysis that all the shift arguments |
6370 | are the same. |
6371 | TODO: Allow different constants for different vector |
6372 | stmts generated for an SLP instance. */ |
6373 | for (k = 0; |
6374 | k < (slp_node ? slp_node->vec_stmts_size - 1 : ncopies - 1); k++) |
6375 | vec_oprnds1.quick_push (obj: vec_oprnd1); |
6376 | } |
6377 | } |
6378 | else if (!scalar_shift_arg && slp_node && incompatible_op1_vectype_p) |
6379 | { |
6380 | if (was_scalar_shift_arg) |
6381 | { |
6382 | /* If the argument was the same in all lanes create |
6383 | the correctly typed vector shift amount directly. */ |
6384 | op1 = fold_convert (TREE_TYPE (vectype), op1); |
6385 | op1 = vect_init_vector (vinfo, stmt_info, val: op1, TREE_TYPE (vectype), |
6386 | gsi: !loop_vinfo ? gsi : NULL); |
6387 | vec_oprnd1 = vect_init_vector (vinfo, stmt_info, val: op1, type: vectype, |
6388 | gsi: !loop_vinfo ? gsi : NULL); |
6389 | vec_oprnds1.create (nelems: slp_node->vec_stmts_size); |
6390 | for (k = 0; k < slp_node->vec_stmts_size; k++) |
6391 | vec_oprnds1.quick_push (obj: vec_oprnd1); |
6392 | } |
6393 | else if (dt[1] == vect_constant_def) |
6394 | /* The constant shift amount has been adjusted in place. */ |
6395 | ; |
6396 | else |
6397 | gcc_assert (TYPE_MODE (op1_vectype) == TYPE_MODE (vectype)); |
6398 | } |
6399 | |
6400 | /* vec_oprnd1 is available if operand 1 should be of a scalar-type |
6401 | (a special case for certain kind of vector shifts); otherwise, |
6402 | operand 1 should be of a vector type (the usual case). */ |
6403 | vect_get_vec_defs (vinfo, stmt_info, slp_node, ncopies, |
6404 | op0, vec_oprnds0: &vec_oprnds0, |
6405 | op1: vec_oprnd1 ? NULL_TREE : op1, vec_oprnds1: &vec_oprnds1); |
6406 | |
6407 | /* Arguments are ready. Create the new vector stmt. */ |
6408 | FOR_EACH_VEC_ELT (vec_oprnds0, i, vop0) |
6409 | { |
6410 | /* For internal defs where we need to use a scalar shift arg |
6411 | extract the first lane. */ |
6412 | if (scalar_shift_arg && dt[1] == vect_internal_def) |
6413 | { |
6414 | vop1 = vec_oprnds1[0]; |
6415 | new_temp = make_ssa_name (TREE_TYPE (TREE_TYPE (vop1))); |
6416 | gassign *new_stmt |
6417 | = gimple_build_assign (new_temp, |
6418 | build3 (BIT_FIELD_REF, TREE_TYPE (new_temp), |
6419 | vop1, |
6420 | TYPE_SIZE (TREE_TYPE (new_temp)), |
6421 | bitsize_zero_node)); |
6422 | vect_finish_stmt_generation (vinfo, stmt_info, vec_stmt: new_stmt, gsi); |
6423 | vop1 = new_temp; |
6424 | } |
6425 | else |
6426 | vop1 = vec_oprnds1[i]; |
6427 | gassign *new_stmt = gimple_build_assign (vec_dest, code, vop0, vop1); |
6428 | new_temp = make_ssa_name (var: vec_dest, stmt: new_stmt); |
6429 | gimple_assign_set_lhs (gs: new_stmt, lhs: new_temp); |
6430 | vect_finish_stmt_generation (vinfo, stmt_info, vec_stmt: new_stmt, gsi); |
6431 | if (slp_node) |
6432 | slp_node->push_vec_def (def: new_stmt); |
6433 | else |
6434 | STMT_VINFO_VEC_STMTS (stmt_info).safe_push (obj: new_stmt); |
6435 | } |
6436 | |
6437 | if (!slp_node) |
6438 | *vec_stmt = STMT_VINFO_VEC_STMTS (stmt_info)[0]; |
6439 | |
6440 | vec_oprnds0.release (); |
6441 | vec_oprnds1.release (); |
6442 | |
6443 | return true; |
6444 | } |
6445 | |
6446 | /* Function vectorizable_operation. |
6447 | |
6448 | Check if STMT_INFO performs a binary, unary or ternary operation that can |
6449 | be vectorized. |
6450 | If VEC_STMT is also passed, vectorize STMT_INFO: create a vectorized |
6451 | stmt to replace it, put it in VEC_STMT, and insert it at GSI. |
6452 | Return true if STMT_INFO is vectorizable in this way. */ |
6453 | |
6454 | static bool |
6455 | vectorizable_operation (vec_info *vinfo, |
6456 | stmt_vec_info stmt_info, gimple_stmt_iterator *gsi, |
6457 | gimple **vec_stmt, slp_tree slp_node, |
6458 | stmt_vector_for_cost *cost_vec) |
6459 | { |
6460 | tree vec_dest; |
6461 | tree scalar_dest; |
6462 | tree op0, op1 = NULL_TREE, op2 = NULL_TREE; |
6463 | tree vectype; |
6464 | loop_vec_info loop_vinfo = dyn_cast <loop_vec_info> (p: vinfo); |
6465 | enum tree_code code, orig_code; |
6466 | machine_mode vec_mode; |
6467 | tree new_temp; |
6468 | int op_type; |
6469 | optab optab; |
6470 | bool target_support_p; |
6471 | enum vect_def_type dt[3] |
6472 | = {vect_unknown_def_type, vect_unknown_def_type, vect_unknown_def_type}; |
6473 | int ndts = 3; |
6474 | poly_uint64 nunits_in; |
6475 | poly_uint64 nunits_out; |
6476 | tree vectype_out; |
6477 | int ncopies, vec_num; |
6478 | int i; |
6479 | vec<tree> vec_oprnds0 = vNULL; |
6480 | vec<tree> vec_oprnds1 = vNULL; |
6481 | vec<tree> vec_oprnds2 = vNULL; |
6482 | tree vop0, vop1, vop2; |
6483 | bb_vec_info bb_vinfo = dyn_cast <bb_vec_info> (p: vinfo); |
6484 | |
6485 | if (!STMT_VINFO_RELEVANT_P (stmt_info) && !bb_vinfo) |
6486 | return false; |
6487 | |
6488 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def |
6489 | && ! vec_stmt) |
6490 | return false; |
6491 | |
6492 | /* Is STMT a vectorizable binary/unary operation? */ |
6493 | gassign *stmt = dyn_cast <gassign *> (p: stmt_info->stmt); |
6494 | if (!stmt) |
6495 | return false; |
6496 | |
6497 | /* Loads and stores are handled in vectorizable_{load,store}. */ |
6498 | if (STMT_VINFO_DATA_REF (stmt_info)) |
6499 | return false; |
6500 | |
6501 | orig_code = code = gimple_assign_rhs_code (gs: stmt); |
6502 | |
6503 | /* Shifts are handled in vectorizable_shift. */ |
6504 | if (code == LSHIFT_EXPR |
6505 | || code == RSHIFT_EXPR |
6506 | || code == LROTATE_EXPR |
6507 | || code == RROTATE_EXPR) |
6508 | return false; |
6509 | |
6510 | /* Comparisons are handled in vectorizable_comparison. */ |
6511 | if (TREE_CODE_CLASS (code) == tcc_comparison) |
6512 | return false; |
6513 | |
6514 | /* Conditions are handled in vectorizable_condition. */ |
6515 | if (code == COND_EXPR) |
6516 | return false; |
6517 | |
6518 | /* For pointer addition and subtraction, we should use the normal |
6519 | plus and minus for the vector operation. */ |
6520 | if (code == POINTER_PLUS_EXPR) |
6521 | code = PLUS_EXPR; |
6522 | if (code == POINTER_DIFF_EXPR) |
6523 | code = MINUS_EXPR; |
6524 | |
6525 | /* Support only unary or binary operations. */ |
6526 | op_type = TREE_CODE_LENGTH (code); |
6527 | if (op_type != unary_op && op_type != binary_op && op_type != ternary_op) |
6528 | { |
6529 | if (dump_enabled_p ()) |
6530 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
6531 | "num. args = %d (not unary/binary/ternary op).\n" , |
6532 | op_type); |
6533 | return false; |
6534 | } |
6535 | |
6536 | scalar_dest = gimple_assign_lhs (gs: stmt); |
6537 | vectype_out = STMT_VINFO_VECTYPE (stmt_info); |
6538 | |
6539 | /* Most operations cannot handle bit-precision types without extra |
6540 | truncations. */ |
6541 | bool mask_op_p = VECTOR_BOOLEAN_TYPE_P (vectype_out); |
6542 | if (!mask_op_p |
6543 | && !type_has_mode_precision_p (TREE_TYPE (scalar_dest)) |
6544 | /* Exception are bitwise binary operations. */ |
6545 | && code != BIT_IOR_EXPR |
6546 | && code != BIT_XOR_EXPR |
6547 | && code != BIT_AND_EXPR) |
6548 | { |
6549 | if (dump_enabled_p ()) |
6550 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
6551 | "bit-precision arithmetic not supported.\n" ); |
6552 | return false; |
6553 | } |
6554 | |
6555 | slp_tree slp_op0; |
6556 | if (!vect_is_simple_use (vinfo, stmt_info, slp_node, |
6557 | 0, &op0, &slp_op0, &dt[0], &vectype)) |
6558 | { |
6559 | if (dump_enabled_p ()) |
6560 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
6561 | "use not simple.\n" ); |
6562 | return false; |
6563 | } |
6564 | bool is_invariant = (dt[0] == vect_external_def |
6565 | || dt[0] == vect_constant_def); |
6566 | /* If op0 is an external or constant def, infer the vector type |
6567 | from the scalar type. */ |
6568 | if (!vectype) |
6569 | { |
6570 | /* For boolean type we cannot determine vectype by |
6571 | invariant value (don't know whether it is a vector |
6572 | of booleans or vector of integers). We use output |
6573 | vectype because operations on boolean don't change |
6574 | type. */ |
6575 | if (VECT_SCALAR_BOOLEAN_TYPE_P (TREE_TYPE (op0))) |
6576 | { |
6577 | if (!VECT_SCALAR_BOOLEAN_TYPE_P (TREE_TYPE (scalar_dest))) |
6578 | { |
6579 | if (dump_enabled_p ()) |
6580 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
6581 | "not supported operation on bool value.\n" ); |
6582 | return false; |
6583 | } |
6584 | vectype = vectype_out; |
6585 | } |
6586 | else |
6587 | vectype = get_vectype_for_scalar_type (vinfo, TREE_TYPE (op0), |
6588 | slp_node); |
6589 | } |
6590 | if (vec_stmt) |
6591 | gcc_assert (vectype); |
6592 | if (!vectype) |
6593 | { |
6594 | if (dump_enabled_p ()) |
6595 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
6596 | "no vectype for scalar type %T\n" , |
6597 | TREE_TYPE (op0)); |
6598 | |
6599 | return false; |
6600 | } |
6601 | |
6602 | nunits_out = TYPE_VECTOR_SUBPARTS (node: vectype_out); |
6603 | nunits_in = TYPE_VECTOR_SUBPARTS (node: vectype); |
6604 | if (maybe_ne (a: nunits_out, b: nunits_in)) |
6605 | return false; |
6606 | |
6607 | tree vectype2 = NULL_TREE, vectype3 = NULL_TREE; |
6608 | slp_tree slp_op1 = NULL, slp_op2 = NULL; |
6609 | if (op_type == binary_op || op_type == ternary_op) |
6610 | { |
6611 | if (!vect_is_simple_use (vinfo, stmt_info, slp_node, |
6612 | 1, &op1, &slp_op1, &dt[1], &vectype2)) |
6613 | { |
6614 | if (dump_enabled_p ()) |
6615 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
6616 | "use not simple.\n" ); |
6617 | return false; |
6618 | } |
6619 | is_invariant &= (dt[1] == vect_external_def |
6620 | || dt[1] == vect_constant_def); |
6621 | if (vectype2 |
6622 | && maybe_ne (a: nunits_out, b: TYPE_VECTOR_SUBPARTS (node: vectype2))) |
6623 | return false; |
6624 | } |
6625 | if (op_type == ternary_op) |
6626 | { |
6627 | if (!vect_is_simple_use (vinfo, stmt_info, slp_node, |
6628 | 2, &op2, &slp_op2, &dt[2], &vectype3)) |
6629 | { |
6630 | if (dump_enabled_p ()) |
6631 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
6632 | "use not simple.\n" ); |
6633 | return false; |
6634 | } |
6635 | is_invariant &= (dt[2] == vect_external_def |
6636 | || dt[2] == vect_constant_def); |
6637 | if (vectype3 |
6638 | && maybe_ne (a: nunits_out, b: TYPE_VECTOR_SUBPARTS (node: vectype3))) |
6639 | return false; |
6640 | } |
6641 | |
6642 | /* Multiple types in SLP are handled by creating the appropriate number of |
6643 | vectorized stmts for each SLP node. Hence, NCOPIES is always 1 in |
6644 | case of SLP. */ |
6645 | if (slp_node) |
6646 | { |
6647 | ncopies = 1; |
6648 | vec_num = SLP_TREE_NUMBER_OF_VEC_STMTS (slp_node); |
6649 | } |
6650 | else |
6651 | { |
6652 | ncopies = vect_get_num_copies (loop_vinfo, vectype); |
6653 | vec_num = 1; |
6654 | } |
6655 | |
6656 | gcc_assert (ncopies >= 1); |
6657 | |
6658 | /* Reject attempts to combine mask types with nonmask types, e.g. if |
6659 | we have an AND between a (nonmask) boolean loaded from memory and |
6660 | a (mask) boolean result of a comparison. |
6661 | |
6662 | TODO: We could easily fix these cases up using pattern statements. */ |
6663 | if (VECTOR_BOOLEAN_TYPE_P (vectype) != mask_op_p |
6664 | || (vectype2 && VECTOR_BOOLEAN_TYPE_P (vectype2) != mask_op_p) |
6665 | || (vectype3 && VECTOR_BOOLEAN_TYPE_P (vectype3) != mask_op_p)) |
6666 | { |
6667 | if (dump_enabled_p ()) |
6668 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
6669 | "mixed mask and nonmask vector types\n" ); |
6670 | return false; |
6671 | } |
6672 | |
6673 | /* Supportable by target? */ |
6674 | |
6675 | vec_mode = TYPE_MODE (vectype); |
6676 | if (code == MULT_HIGHPART_EXPR) |
6677 | target_support_p = can_mult_highpart_p (vec_mode, TYPE_UNSIGNED (vectype)); |
6678 | else |
6679 | { |
6680 | optab = optab_for_tree_code (code, vectype, optab_default); |
6681 | if (!optab) |
6682 | { |
6683 | if (dump_enabled_p ()) |
6684 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
6685 | "no optab.\n" ); |
6686 | return false; |
6687 | } |
6688 | target_support_p = (optab_handler (op: optab, mode: vec_mode) != CODE_FOR_nothing |
6689 | || optab_libfunc (optab, vec_mode)); |
6690 | } |
6691 | |
6692 | bool using_emulated_vectors_p = vect_emulated_vector_p (vectype); |
6693 | if (!target_support_p || using_emulated_vectors_p) |
6694 | { |
6695 | if (dump_enabled_p ()) |
6696 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
6697 | "op not supported by target.\n" ); |
6698 | /* When vec_mode is not a vector mode and we verified ops we |
6699 | do not have to lower like AND are natively supported let |
6700 | those through even when the mode isn't word_mode. For |
6701 | ops we have to lower the lowering code assumes we are |
6702 | dealing with word_mode. */ |
6703 | if ((((code == PLUS_EXPR || code == MINUS_EXPR || code == NEGATE_EXPR) |
6704 | || !target_support_p) |
6705 | && maybe_ne (a: GET_MODE_SIZE (mode: vec_mode), UNITS_PER_WORD)) |
6706 | /* Check only during analysis. */ |
6707 | || (!vec_stmt && !vect_can_vectorize_without_simd_p (code))) |
6708 | { |
6709 | if (dump_enabled_p ()) |
6710 | dump_printf (MSG_NOTE, "using word mode not possible.\n" ); |
6711 | return false; |
6712 | } |
6713 | if (dump_enabled_p ()) |
6714 | dump_printf_loc (MSG_NOTE, vect_location, |
6715 | "proceeding using word mode.\n" ); |
6716 | using_emulated_vectors_p = true; |
6717 | } |
6718 | |
6719 | int reduc_idx = STMT_VINFO_REDUC_IDX (stmt_info); |
6720 | vec_loop_masks *masks = (loop_vinfo ? &LOOP_VINFO_MASKS (loop_vinfo) : NULL); |
6721 | vec_loop_lens *lens = (loop_vinfo ? &LOOP_VINFO_LENS (loop_vinfo) : NULL); |
6722 | internal_fn cond_fn = get_conditional_internal_fn (code); |
6723 | internal_fn cond_len_fn = get_conditional_len_internal_fn (code); |
6724 | |
6725 | /* If operating on inactive elements could generate spurious traps, |
6726 | we need to restrict the operation to active lanes. Note that this |
6727 | specifically doesn't apply to unhoisted invariants, since they |
6728 | operate on the same value for every lane. |
6729 | |
6730 | Similarly, if this operation is part of a reduction, a fully-masked |
6731 | loop should only change the active lanes of the reduction chain, |
6732 | keeping the inactive lanes as-is. */ |
6733 | bool mask_out_inactive = ((!is_invariant && gimple_could_trap_p (stmt)) |
6734 | || reduc_idx >= 0); |
6735 | |
6736 | if (!vec_stmt) /* transformation not required. */ |
6737 | { |
6738 | if (loop_vinfo |
6739 | && LOOP_VINFO_CAN_USE_PARTIAL_VECTORS_P (loop_vinfo) |
6740 | && mask_out_inactive) |
6741 | { |
6742 | if (cond_len_fn != IFN_LAST |
6743 | && direct_internal_fn_supported_p (cond_len_fn, vectype, |
6744 | OPTIMIZE_FOR_SPEED)) |
6745 | vect_record_loop_len (loop_vinfo, lens, ncopies * vec_num, vectype, |
6746 | 1); |
6747 | else if (cond_fn != IFN_LAST |
6748 | && direct_internal_fn_supported_p (cond_fn, vectype, |
6749 | OPTIMIZE_FOR_SPEED)) |
6750 | vect_record_loop_mask (loop_vinfo, masks, ncopies * vec_num, |
6751 | vectype, NULL); |
6752 | else |
6753 | { |
6754 | if (dump_enabled_p ()) |
6755 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
6756 | "can't use a fully-masked loop because no" |
6757 | " conditional operation is available.\n" ); |
6758 | LOOP_VINFO_CAN_USE_PARTIAL_VECTORS_P (loop_vinfo) = false; |
6759 | } |
6760 | } |
6761 | |
6762 | /* Put types on constant and invariant SLP children. */ |
6763 | if (slp_node |
6764 | && (!vect_maybe_update_slp_op_vectype (slp_op0, vectype) |
6765 | || !vect_maybe_update_slp_op_vectype (slp_op1, vectype) |
6766 | || !vect_maybe_update_slp_op_vectype (slp_op2, vectype))) |
6767 | { |
6768 | if (dump_enabled_p ()) |
6769 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
6770 | "incompatible vector types for invariants\n" ); |
6771 | return false; |
6772 | } |
6773 | |
6774 | STMT_VINFO_TYPE (stmt_info) = op_vec_info_type; |
6775 | DUMP_VECT_SCOPE ("vectorizable_operation" ); |
6776 | vect_model_simple_cost (vinfo, stmt_info, |
6777 | ncopies, dt, ndts, node: slp_node, cost_vec); |
6778 | if (using_emulated_vectors_p) |
6779 | { |
6780 | /* The above vect_model_simple_cost call handles constants |
6781 | in the prologue and (mis-)costs one of the stmts as |
6782 | vector stmt. See below for the actual lowering that will |
6783 | be applied. */ |
6784 | unsigned n |
6785 | = slp_node ? SLP_TREE_NUMBER_OF_VEC_STMTS (slp_node) : ncopies; |
6786 | switch (code) |
6787 | { |
6788 | case PLUS_EXPR: |
6789 | n *= 5; |
6790 | break; |
6791 | case MINUS_EXPR: |
6792 | n *= 6; |
6793 | break; |
6794 | case NEGATE_EXPR: |
6795 | n *= 4; |
6796 | break; |
6797 | default: |
6798 | /* Bit operations do not have extra cost and are accounted |
6799 | as vector stmt by vect_model_simple_cost. */ |
6800 | n = 0; |
6801 | break; |
6802 | } |
6803 | if (n != 0) |
6804 | { |
6805 | /* We also need to materialize two large constants. */ |
6806 | record_stmt_cost (body_cost_vec: cost_vec, count: 2, kind: scalar_stmt, stmt_info, |
6807 | misalign: 0, where: vect_prologue); |
6808 | record_stmt_cost (body_cost_vec: cost_vec, count: n, kind: scalar_stmt, stmt_info, |
6809 | misalign: 0, where: vect_body); |
6810 | } |
6811 | } |
6812 | return true; |
6813 | } |
6814 | |
6815 | /* Transform. */ |
6816 | |
6817 | if (dump_enabled_p ()) |
6818 | dump_printf_loc (MSG_NOTE, vect_location, |
6819 | "transform binary/unary operation.\n" ); |
6820 | |
6821 | bool masked_loop_p = loop_vinfo && LOOP_VINFO_FULLY_MASKED_P (loop_vinfo); |
6822 | bool len_loop_p = loop_vinfo && LOOP_VINFO_FULLY_WITH_LENGTH_P (loop_vinfo); |
6823 | |
6824 | /* POINTER_DIFF_EXPR has pointer arguments which are vectorized as |
6825 | vectors with unsigned elements, but the result is signed. So, we |
6826 | need to compute the MINUS_EXPR into vectype temporary and |
6827 | VIEW_CONVERT_EXPR it into the final vectype_out result. */ |
6828 | tree vec_cvt_dest = NULL_TREE; |
6829 | if (orig_code == POINTER_DIFF_EXPR) |
6830 | { |
6831 | vec_dest = vect_create_destination_var (scalar_dest, vectype); |
6832 | vec_cvt_dest = vect_create_destination_var (scalar_dest, vectype_out); |
6833 | } |
6834 | /* Handle def. */ |
6835 | else |
6836 | vec_dest = vect_create_destination_var (scalar_dest, vectype_out); |
6837 | |
6838 | /* In case the vectorization factor (VF) is bigger than the number |
6839 | of elements that we can fit in a vectype (nunits), we have to generate |
6840 | more than one vector stmt - i.e - we need to "unroll" the |
6841 | vector stmt by a factor VF/nunits. In doing so, we record a pointer |
6842 | from one copy of the vector stmt to the next, in the field |
6843 | STMT_VINFO_RELATED_STMT. This is necessary in order to allow following |
6844 | stages to find the correct vector defs to be used when vectorizing |
6845 | stmts that use the defs of the current stmt. The example below |
6846 | illustrates the vectorization process when VF=16 and nunits=4 (i.e., |
6847 | we need to create 4 vectorized stmts): |
6848 | |
6849 | before vectorization: |
6850 | RELATED_STMT VEC_STMT |
6851 | S1: x = memref - - |
6852 | S2: z = x + 1 - - |
6853 | |
6854 | step 1: vectorize stmt S1 (done in vectorizable_load. See more details |
6855 | there): |
6856 | RELATED_STMT VEC_STMT |
6857 | VS1_0: vx0 = memref0 VS1_1 - |
6858 | VS1_1: vx1 = memref1 VS1_2 - |
6859 | VS1_2: vx2 = memref2 VS1_3 - |
6860 | VS1_3: vx3 = memref3 - - |
6861 | S1: x = load - VS1_0 |
6862 | S2: z = x + 1 - - |
6863 | |
6864 | step2: vectorize stmt S2 (done here): |
6865 | To vectorize stmt S2 we first need to find the relevant vector |
6866 | def for the first operand 'x'. This is, as usual, obtained from |
6867 | the vector stmt recorded in the STMT_VINFO_VEC_STMT of the stmt |
6868 | that defines 'x' (S1). This way we find the stmt VS1_0, and the |
6869 | relevant vector def 'vx0'. Having found 'vx0' we can generate |
6870 | the vector stmt VS2_0, and as usual, record it in the |
6871 | STMT_VINFO_VEC_STMT of stmt S2. |
6872 | When creating the second copy (VS2_1), we obtain the relevant vector |
6873 | def from the vector stmt recorded in the STMT_VINFO_RELATED_STMT of |
6874 | stmt VS1_0. This way we find the stmt VS1_1 and the relevant |
6875 | vector def 'vx1'. Using 'vx1' we create stmt VS2_1 and record a |
6876 | pointer to it in the STMT_VINFO_RELATED_STMT of the vector stmt VS2_0. |
6877 | Similarly when creating stmts VS2_2 and VS2_3. This is the resulting |
6878 | chain of stmts and pointers: |
6879 | RELATED_STMT VEC_STMT |
6880 | VS1_0: vx0 = memref0 VS1_1 - |
6881 | VS1_1: vx1 = memref1 VS1_2 - |
6882 | VS1_2: vx2 = memref2 VS1_3 - |
6883 | VS1_3: vx3 = memref3 - - |
6884 | S1: x = load - VS1_0 |
6885 | VS2_0: vz0 = vx0 + v1 VS2_1 - |
6886 | VS2_1: vz1 = vx1 + v1 VS2_2 - |
6887 | VS2_2: vz2 = vx2 + v1 VS2_3 - |
6888 | VS2_3: vz3 = vx3 + v1 - - |
6889 | S2: z = x + 1 - VS2_0 */ |
6890 | |
6891 | vect_get_vec_defs (vinfo, stmt_info, slp_node, ncopies, |
6892 | op0, vec_oprnds0: &vec_oprnds0, op1, vec_oprnds1: &vec_oprnds1, op2, vec_oprnds2: &vec_oprnds2); |
6893 | /* Arguments are ready. Create the new vector stmt. */ |
6894 | FOR_EACH_VEC_ELT (vec_oprnds0, i, vop0) |
6895 | { |
6896 | gimple *new_stmt = NULL; |
6897 | vop1 = ((op_type == binary_op || op_type == ternary_op) |
6898 | ? vec_oprnds1[i] : NULL_TREE); |
6899 | vop2 = ((op_type == ternary_op) ? vec_oprnds2[i] : NULL_TREE); |
6900 | if (using_emulated_vectors_p |
6901 | && (code == PLUS_EXPR || code == MINUS_EXPR || code == NEGATE_EXPR)) |
6902 | { |
6903 | /* Lower the operation. This follows vector lowering. */ |
6904 | unsigned int width = vector_element_bits (vectype); |
6905 | tree inner_type = TREE_TYPE (vectype); |
6906 | tree word_type |
6907 | = build_nonstandard_integer_type (GET_MODE_BITSIZE (mode: word_mode), 1); |
6908 | HOST_WIDE_INT max = GET_MODE_MASK (TYPE_MODE (inner_type)); |
6909 | tree low_bits = build_replicated_int_cst (word_type, width, max >> 1); |
6910 | tree high_bits |
6911 | = build_replicated_int_cst (word_type, width, max & ~(max >> 1)); |
6912 | tree wvop0 = make_ssa_name (var: word_type); |
6913 | new_stmt = gimple_build_assign (wvop0, VIEW_CONVERT_EXPR, |
6914 | build1 (VIEW_CONVERT_EXPR, |
6915 | word_type, vop0)); |
6916 | vect_finish_stmt_generation (vinfo, stmt_info, vec_stmt: new_stmt, gsi); |
6917 | tree result_low, signs; |
6918 | if (code == PLUS_EXPR || code == MINUS_EXPR) |
6919 | { |
6920 | tree wvop1 = make_ssa_name (var: word_type); |
6921 | new_stmt = gimple_build_assign (wvop1, VIEW_CONVERT_EXPR, |
6922 | build1 (VIEW_CONVERT_EXPR, |
6923 | word_type, vop1)); |
6924 | vect_finish_stmt_generation (vinfo, stmt_info, vec_stmt: new_stmt, gsi); |
6925 | signs = make_ssa_name (var: word_type); |
6926 | new_stmt = gimple_build_assign (signs, |
6927 | BIT_XOR_EXPR, wvop0, wvop1); |
6928 | vect_finish_stmt_generation (vinfo, stmt_info, vec_stmt: new_stmt, gsi); |
6929 | tree b_low = make_ssa_name (var: word_type); |
6930 | new_stmt = gimple_build_assign (b_low, |
6931 | BIT_AND_EXPR, wvop1, low_bits); |
6932 | vect_finish_stmt_generation (vinfo, stmt_info, vec_stmt: new_stmt, gsi); |
6933 | tree a_low = make_ssa_name (var: word_type); |
6934 | if (code == PLUS_EXPR) |
6935 | new_stmt = gimple_build_assign (a_low, |
6936 | BIT_AND_EXPR, wvop0, low_bits); |
6937 | else |
6938 | new_stmt = gimple_build_assign (a_low, |
6939 | BIT_IOR_EXPR, wvop0, high_bits); |
6940 | vect_finish_stmt_generation (vinfo, stmt_info, vec_stmt: new_stmt, gsi); |
6941 | if (code == MINUS_EXPR) |
6942 | { |
6943 | new_stmt = gimple_build_assign (NULL_TREE, |
6944 | BIT_NOT_EXPR, signs); |
6945 | signs = make_ssa_name (var: word_type); |
6946 | gimple_assign_set_lhs (gs: new_stmt, lhs: signs); |
6947 | vect_finish_stmt_generation (vinfo, stmt_info, vec_stmt: new_stmt, gsi); |
6948 | } |
6949 | new_stmt = gimple_build_assign (NULL_TREE, |
6950 | BIT_AND_EXPR, signs, high_bits); |
6951 | signs = make_ssa_name (var: word_type); |
6952 | gimple_assign_set_lhs (gs: new_stmt, lhs: signs); |
6953 | vect_finish_stmt_generation (vinfo, stmt_info, vec_stmt: new_stmt, gsi); |
6954 | result_low = make_ssa_name (var: word_type); |
6955 | new_stmt = gimple_build_assign (result_low, code, a_low, b_low); |
6956 | vect_finish_stmt_generation (vinfo, stmt_info, vec_stmt: new_stmt, gsi); |
6957 | } |
6958 | else |
6959 | { |
6960 | tree a_low = make_ssa_name (var: word_type); |
6961 | new_stmt = gimple_build_assign (a_low, |
6962 | BIT_AND_EXPR, wvop0, low_bits); |
6963 | vect_finish_stmt_generation (vinfo, stmt_info, vec_stmt: new_stmt, gsi); |
6964 | signs = make_ssa_name (var: word_type); |
6965 | new_stmt = gimple_build_assign (signs, BIT_NOT_EXPR, wvop0); |
6966 | vect_finish_stmt_generation (vinfo, stmt_info, vec_stmt: new_stmt, gsi); |
6967 | new_stmt = gimple_build_assign (NULL_TREE, |
6968 | BIT_AND_EXPR, signs, high_bits); |
6969 | signs = make_ssa_name (var: word_type); |
6970 | gimple_assign_set_lhs (gs: new_stmt, lhs: signs); |
6971 | vect_finish_stmt_generation (vinfo, stmt_info, vec_stmt: new_stmt, gsi); |
6972 | result_low = make_ssa_name (var: word_type); |
6973 | new_stmt = gimple_build_assign (result_low, |
6974 | MINUS_EXPR, high_bits, a_low); |
6975 | vect_finish_stmt_generation (vinfo, stmt_info, vec_stmt: new_stmt, gsi); |
6976 | } |
6977 | new_stmt = gimple_build_assign (NULL_TREE, BIT_XOR_EXPR, result_low, |
6978 | signs); |
6979 | result_low = make_ssa_name (var: word_type); |
6980 | gimple_assign_set_lhs (gs: new_stmt, lhs: result_low); |
6981 | vect_finish_stmt_generation (vinfo, stmt_info, vec_stmt: new_stmt, gsi); |
6982 | new_stmt = gimple_build_assign (NULL_TREE, VIEW_CONVERT_EXPR, |
6983 | build1 (VIEW_CONVERT_EXPR, |
6984 | vectype, result_low)); |
6985 | new_temp = make_ssa_name (var: vectype); |
6986 | gimple_assign_set_lhs (gs: new_stmt, lhs: new_temp); |
6987 | vect_finish_stmt_generation (vinfo, stmt_info, vec_stmt: new_stmt, gsi); |
6988 | } |
6989 | else if ((masked_loop_p || len_loop_p) && mask_out_inactive) |
6990 | { |
6991 | tree mask; |
6992 | if (masked_loop_p) |
6993 | mask = vect_get_loop_mask (loop_vinfo, gsi, masks, |
6994 | vec_num * ncopies, vectype, i); |
6995 | else |
6996 | /* Dummy mask. */ |
6997 | mask = build_minus_one_cst (truth_type_for (vectype)); |
6998 | auto_vec<tree> vops (6); |
6999 | vops.quick_push (obj: mask); |
7000 | vops.quick_push (obj: vop0); |
7001 | if (vop1) |
7002 | vops.quick_push (obj: vop1); |
7003 | if (vop2) |
7004 | vops.quick_push (obj: vop2); |
7005 | if (reduc_idx >= 0) |
7006 | { |
7007 | /* Perform the operation on active elements only and take |
7008 | inactive elements from the reduction chain input. */ |
7009 | gcc_assert (!vop2); |
7010 | vops.quick_push (obj: reduc_idx == 1 ? vop1 : vop0); |
7011 | } |
7012 | else |
7013 | { |
7014 | auto else_value = targetm.preferred_else_value |
7015 | (cond_fn, vectype, vops.length () - 1, &vops[1]); |
7016 | vops.quick_push (obj: else_value); |
7017 | } |
7018 | if (len_loop_p) |
7019 | { |
7020 | tree len = vect_get_loop_len (loop_vinfo, gsi, lens, |
7021 | vec_num * ncopies, vectype, i, 1); |
7022 | signed char biasval |
7023 | = LOOP_VINFO_PARTIAL_LOAD_STORE_BIAS (loop_vinfo); |
7024 | tree bias = build_int_cst (intQI_type_node, biasval); |
7025 | vops.quick_push (obj: len); |
7026 | vops.quick_push (obj: bias); |
7027 | } |
7028 | gcall *call |
7029 | = gimple_build_call_internal_vec (masked_loop_p ? cond_fn |
7030 | : cond_len_fn, |
7031 | vops); |
7032 | new_temp = make_ssa_name (var: vec_dest, stmt: call); |
7033 | gimple_call_set_lhs (gs: call, lhs: new_temp); |
7034 | gimple_call_set_nothrow (s: call, nothrow_p: true); |
7035 | vect_finish_stmt_generation (vinfo, stmt_info, vec_stmt: call, gsi); |
7036 | new_stmt = call; |
7037 | } |
7038 | else |
7039 | { |
7040 | tree mask = NULL_TREE; |
7041 | /* When combining two masks check if either of them is elsewhere |
7042 | combined with a loop mask, if that's the case we can mark that the |
7043 | new combined mask doesn't need to be combined with a loop mask. */ |
7044 | if (masked_loop_p |
7045 | && code == BIT_AND_EXPR |
7046 | && VECTOR_BOOLEAN_TYPE_P (vectype)) |
7047 | { |
7048 | if (loop_vinfo->scalar_cond_masked_set.contains (k: { op0, |
7049 | ncopies})) |
7050 | { |
7051 | mask = vect_get_loop_mask (loop_vinfo, gsi, masks, |
7052 | vec_num * ncopies, vectype, i); |
7053 | |
7054 | vop0 = prepare_vec_mask (loop_vinfo, TREE_TYPE (mask), loop_mask: mask, |
7055 | vec_mask: vop0, gsi); |
7056 | } |
7057 | |
7058 | if (loop_vinfo->scalar_cond_masked_set.contains (k: { op1, |
7059 | ncopies })) |
7060 | { |
7061 | mask = vect_get_loop_mask (loop_vinfo, gsi, masks, |
7062 | vec_num * ncopies, vectype, i); |
7063 | |
7064 | vop1 = prepare_vec_mask (loop_vinfo, TREE_TYPE (mask), loop_mask: mask, |
7065 | vec_mask: vop1, gsi); |
7066 | } |
7067 | } |
7068 | |
7069 | new_stmt = gimple_build_assign (vec_dest, code, vop0, vop1, vop2); |
7070 | new_temp = make_ssa_name (var: vec_dest, stmt: new_stmt); |
7071 | gimple_assign_set_lhs (gs: new_stmt, lhs: new_temp); |
7072 | vect_finish_stmt_generation (vinfo, stmt_info, vec_stmt: new_stmt, gsi); |
7073 | if (using_emulated_vectors_p) |
7074 | suppress_warning (new_stmt, OPT_Wvector_operation_performance); |
7075 | |
7076 | /* Enter the combined value into the vector cond hash so we don't |
7077 | AND it with a loop mask again. */ |
7078 | if (mask) |
7079 | loop_vinfo->vec_cond_masked_set.add (k: { new_temp, mask }); |
7080 | } |
7081 | |
7082 | if (vec_cvt_dest) |
7083 | { |
7084 | new_temp = build1 (VIEW_CONVERT_EXPR, vectype_out, new_temp); |
7085 | new_stmt = gimple_build_assign (vec_cvt_dest, VIEW_CONVERT_EXPR, |
7086 | new_temp); |
7087 | new_temp = make_ssa_name (var: vec_cvt_dest, stmt: new_stmt); |
7088 | gimple_assign_set_lhs (gs: new_stmt, lhs: new_temp); |
7089 | vect_finish_stmt_generation (vinfo, stmt_info, |
7090 | vec_stmt: new_stmt, gsi); |
7091 | } |
7092 | |
7093 | if (slp_node) |
7094 | slp_node->push_vec_def (def: new_stmt); |
7095 | else |
7096 | STMT_VINFO_VEC_STMTS (stmt_info).safe_push (obj: new_stmt); |
7097 | } |
7098 | |
7099 | if (!slp_node) |
7100 | *vec_stmt = STMT_VINFO_VEC_STMTS (stmt_info)[0]; |
7101 | |
7102 | vec_oprnds0.release (); |
7103 | vec_oprnds1.release (); |
7104 | vec_oprnds2.release (); |
7105 | |
7106 | return true; |
7107 | } |
7108 | |
7109 | /* A helper function to ensure data reference DR_INFO's base alignment. */ |
7110 | |
7111 | static void |
7112 | ensure_base_align (dr_vec_info *dr_info) |
7113 | { |
7114 | /* Alignment is only analyzed for the first element of a DR group, |
7115 | use that to look at base alignment we need to enforce. */ |
7116 | if (STMT_VINFO_GROUPED_ACCESS (dr_info->stmt)) |
7117 | dr_info = STMT_VINFO_DR_INFO (DR_GROUP_FIRST_ELEMENT (dr_info->stmt)); |
7118 | |
7119 | gcc_assert (dr_info->misalignment != DR_MISALIGNMENT_UNINITIALIZED); |
7120 | |
7121 | if (dr_info->base_misaligned) |
7122 | { |
7123 | tree base_decl = dr_info->base_decl; |
7124 | |
7125 | // We should only be able to increase the alignment of a base object if |
7126 | // we know what its new alignment should be at compile time. |
7127 | unsigned HOST_WIDE_INT align_base_to = |
7128 | DR_TARGET_ALIGNMENT (dr_info).to_constant () * BITS_PER_UNIT; |
7129 | |
7130 | if (decl_in_symtab_p (decl: base_decl)) |
7131 | symtab_node::get (decl: base_decl)->increase_alignment (align: align_base_to); |
7132 | else if (DECL_ALIGN (base_decl) < align_base_to) |
7133 | { |
7134 | SET_DECL_ALIGN (base_decl, align_base_to); |
7135 | DECL_USER_ALIGN (base_decl) = 1; |
7136 | } |
7137 | dr_info->base_misaligned = false; |
7138 | } |
7139 | } |
7140 | |
7141 | |
7142 | /* Function get_group_alias_ptr_type. |
7143 | |
7144 | Return the alias type for the group starting at FIRST_STMT_INFO. */ |
7145 | |
7146 | static tree |
7147 | get_group_alias_ptr_type (stmt_vec_info first_stmt_info) |
7148 | { |
7149 | struct data_reference *first_dr, *next_dr; |
7150 | |
7151 | first_dr = STMT_VINFO_DATA_REF (first_stmt_info); |
7152 | stmt_vec_info next_stmt_info = DR_GROUP_NEXT_ELEMENT (first_stmt_info); |
7153 | while (next_stmt_info) |
7154 | { |
7155 | next_dr = STMT_VINFO_DATA_REF (next_stmt_info); |
7156 | if (get_alias_set (DR_REF (first_dr)) |
7157 | != get_alias_set (DR_REF (next_dr))) |
7158 | { |
7159 | if (dump_enabled_p ()) |
7160 | dump_printf_loc (MSG_NOTE, vect_location, |
7161 | "conflicting alias set types.\n" ); |
7162 | return ptr_type_node; |
7163 | } |
7164 | next_stmt_info = DR_GROUP_NEXT_ELEMENT (next_stmt_info); |
7165 | } |
7166 | return reference_alias_ptr_type (DR_REF (first_dr)); |
7167 | } |
7168 | |
7169 | |
7170 | /* Function scan_operand_equal_p. |
7171 | |
7172 | Helper function for check_scan_store. Compare two references |
7173 | with .GOMP_SIMD_LANE bases. */ |
7174 | |
7175 | static bool |
7176 | scan_operand_equal_p (tree ref1, tree ref2) |
7177 | { |
7178 | tree ref[2] = { ref1, ref2 }; |
7179 | poly_int64 bitsize[2], bitpos[2]; |
7180 | tree offset[2], base[2]; |
7181 | for (int i = 0; i < 2; ++i) |
7182 | { |
7183 | machine_mode mode; |
7184 | int unsignedp, reversep, volatilep = 0; |
7185 | base[i] = get_inner_reference (ref[i], &bitsize[i], &bitpos[i], |
7186 | &offset[i], &mode, &unsignedp, |
7187 | &reversep, &volatilep); |
7188 | if (reversep || volatilep || maybe_ne (a: bitpos[i], b: 0)) |
7189 | return false; |
7190 | if (TREE_CODE (base[i]) == MEM_REF |
7191 | && offset[i] == NULL_TREE |
7192 | && TREE_CODE (TREE_OPERAND (base[i], 0)) == SSA_NAME) |
7193 | { |
7194 | gimple *def_stmt = SSA_NAME_DEF_STMT (TREE_OPERAND (base[i], 0)); |
7195 | if (is_gimple_assign (gs: def_stmt) |
7196 | && gimple_assign_rhs_code (gs: def_stmt) == POINTER_PLUS_EXPR |
7197 | && TREE_CODE (gimple_assign_rhs1 (def_stmt)) == ADDR_EXPR |
7198 | && TREE_CODE (gimple_assign_rhs2 (def_stmt)) == SSA_NAME) |
7199 | { |
7200 | if (maybe_ne (a: mem_ref_offset (base[i]), b: 0)) |
7201 | return false; |
7202 | base[i] = TREE_OPERAND (gimple_assign_rhs1 (def_stmt), 0); |
7203 | offset[i] = gimple_assign_rhs2 (gs: def_stmt); |
7204 | } |
7205 | } |
7206 | } |
7207 | |
7208 | if (!operand_equal_p (base[0], base[1], flags: 0)) |
7209 | return false; |
7210 | if (maybe_ne (a: bitsize[0], b: bitsize[1])) |
7211 | return false; |
7212 | if (offset[0] != offset[1]) |
7213 | { |
7214 | if (!offset[0] || !offset[1]) |
7215 | return false; |
7216 | if (!operand_equal_p (offset[0], offset[1], flags: 0)) |
7217 | { |
7218 | tree step[2]; |
7219 | for (int i = 0; i < 2; ++i) |
7220 | { |
7221 | step[i] = integer_one_node; |
7222 | if (TREE_CODE (offset[i]) == SSA_NAME) |
7223 | { |
7224 | gimple *def_stmt = SSA_NAME_DEF_STMT (offset[i]); |
7225 | if (is_gimple_assign (gs: def_stmt) |
7226 | && gimple_assign_rhs_code (gs: def_stmt) == MULT_EXPR |
7227 | && (TREE_CODE (gimple_assign_rhs2 (def_stmt)) |
7228 | == INTEGER_CST)) |
7229 | { |
7230 | step[i] = gimple_assign_rhs2 (gs: def_stmt); |
7231 | offset[i] = gimple_assign_rhs1 (gs: def_stmt); |
7232 | } |
7233 | } |
7234 | else if (TREE_CODE (offset[i]) == MULT_EXPR) |
7235 | { |
7236 | step[i] = TREE_OPERAND (offset[i], 1); |
7237 | offset[i] = TREE_OPERAND (offset[i], 0); |
7238 | } |
7239 | tree rhs1 = NULL_TREE; |
7240 | if (TREE_CODE (offset[i]) == SSA_NAME) |
7241 | { |
7242 | gimple *def_stmt = SSA_NAME_DEF_STMT (offset[i]); |
7243 | if (gimple_assign_cast_p (s: def_stmt)) |
7244 | rhs1 = gimple_assign_rhs1 (gs: def_stmt); |
7245 | } |
7246 | else if (CONVERT_EXPR_P (offset[i])) |
7247 | rhs1 = TREE_OPERAND (offset[i], 0); |
7248 | if (rhs1 |
7249 | && INTEGRAL_TYPE_P (TREE_TYPE (rhs1)) |
7250 | && INTEGRAL_TYPE_P (TREE_TYPE (offset[i])) |
7251 | && (TYPE_PRECISION (TREE_TYPE (offset[i])) |
7252 | >= TYPE_PRECISION (TREE_TYPE (rhs1)))) |
7253 | offset[i] = rhs1; |
7254 | } |
7255 | if (!operand_equal_p (offset[0], offset[1], flags: 0) |
7256 | || !operand_equal_p (step[0], step[1], flags: 0)) |
7257 | return false; |
7258 | } |
7259 | } |
7260 | return true; |
7261 | } |
7262 | |
7263 | |
7264 | enum scan_store_kind { |
7265 | /* Normal permutation. */ |
7266 | scan_store_kind_perm, |
7267 | |
7268 | /* Whole vector left shift permutation with zero init. */ |
7269 | scan_store_kind_lshift_zero, |
7270 | |
7271 | /* Whole vector left shift permutation and VEC_COND_EXPR. */ |
7272 | scan_store_kind_lshift_cond |
7273 | }; |
7274 | |
7275 | /* Function check_scan_store. |
7276 | |
7277 | Verify if we can perform the needed permutations or whole vector shifts. |
7278 | Return -1 on failure, otherwise exact log2 of vectype's nunits. |
7279 | USE_WHOLE_VECTOR is a vector of enum scan_store_kind which operation |
7280 | to do at each step. */ |
7281 | |
7282 | static int |
7283 | scan_store_can_perm_p (tree vectype, tree init, |
7284 | vec<enum scan_store_kind> *use_whole_vector = NULL) |
7285 | { |
7286 | enum machine_mode vec_mode = TYPE_MODE (vectype); |
7287 | unsigned HOST_WIDE_INT nunits; |
7288 | if (!TYPE_VECTOR_SUBPARTS (node: vectype).is_constant (const_value: &nunits)) |
7289 | return -1; |
7290 | int units_log2 = exact_log2 (x: nunits); |
7291 | if (units_log2 <= 0) |
7292 | return -1; |
7293 | |
7294 | int i; |
7295 | enum scan_store_kind whole_vector_shift_kind = scan_store_kind_perm; |
7296 | for (i = 0; i <= units_log2; ++i) |
7297 | { |
7298 | unsigned HOST_WIDE_INT j, k; |
7299 | enum scan_store_kind kind = scan_store_kind_perm; |
7300 | vec_perm_builder sel (nunits, nunits, 1); |
7301 | sel.quick_grow (len: nunits); |
7302 | if (i == units_log2) |
7303 | { |
7304 | for (j = 0; j < nunits; ++j) |
7305 | sel[j] = nunits - 1; |
7306 | } |
7307 | else |
7308 | { |
7309 | for (j = 0; j < (HOST_WIDE_INT_1U << i); ++j) |
7310 | sel[j] = j; |
7311 | for (k = 0; j < nunits; ++j, ++k) |
7312 | sel[j] = nunits + k; |
7313 | } |
7314 | vec_perm_indices indices (sel, i == units_log2 ? 1 : 2, nunits); |
7315 | if (!can_vec_perm_const_p (vec_mode, vec_mode, indices)) |
7316 | { |
7317 | if (i == units_log2) |
7318 | return -1; |
7319 | |
7320 | if (whole_vector_shift_kind == scan_store_kind_perm) |
7321 | { |
7322 | if (optab_handler (op: vec_shl_optab, mode: vec_mode) == CODE_FOR_nothing) |
7323 | return -1; |
7324 | whole_vector_shift_kind = scan_store_kind_lshift_zero; |
7325 | /* Whole vector shifts shift in zeros, so if init is all zero |
7326 | constant, there is no need to do anything further. */ |
7327 | if ((TREE_CODE (init) != INTEGER_CST |
7328 | && TREE_CODE (init) != REAL_CST) |
7329 | || !initializer_zerop (init)) |
7330 | { |
7331 | tree masktype = truth_type_for (vectype); |
7332 | if (!expand_vec_cond_expr_p (vectype, masktype, VECTOR_CST)) |
7333 | return -1; |
7334 | whole_vector_shift_kind = scan_store_kind_lshift_cond; |
7335 | } |
7336 | } |
7337 | kind = whole_vector_shift_kind; |
7338 | } |
7339 | if (use_whole_vector) |
7340 | { |
7341 | if (kind != scan_store_kind_perm && use_whole_vector->is_empty ()) |
7342 | use_whole_vector->safe_grow_cleared (len: i, exact: true); |
7343 | if (kind != scan_store_kind_perm || !use_whole_vector->is_empty ()) |
7344 | use_whole_vector->safe_push (obj: kind); |
7345 | } |
7346 | } |
7347 | |
7348 | return units_log2; |
7349 | } |
7350 | |
7351 | |
7352 | /* Function check_scan_store. |
7353 | |
7354 | Check magic stores for #pragma omp scan {in,ex}clusive reductions. */ |
7355 | |
7356 | static bool |
7357 | check_scan_store (vec_info *vinfo, stmt_vec_info stmt_info, tree vectype, |
7358 | enum vect_def_type rhs_dt, bool slp, tree mask, |
7359 | vect_memory_access_type memory_access_type) |
7360 | { |
7361 | loop_vec_info loop_vinfo = dyn_cast <loop_vec_info> (p: vinfo); |
7362 | dr_vec_info *dr_info = STMT_VINFO_DR_INFO (stmt_info); |
7363 | tree ref_type; |
7364 | |
7365 | gcc_assert (STMT_VINFO_SIMD_LANE_ACCESS_P (stmt_info) > 1); |
7366 | if (slp |
7367 | || mask |
7368 | || memory_access_type != VMAT_CONTIGUOUS |
7369 | || TREE_CODE (DR_BASE_ADDRESS (dr_info->dr)) != ADDR_EXPR |
7370 | || !VAR_P (TREE_OPERAND (DR_BASE_ADDRESS (dr_info->dr), 0)) |
7371 | || loop_vinfo == NULL |
7372 | || LOOP_VINFO_FULLY_MASKED_P (loop_vinfo) |
7373 | || STMT_VINFO_GROUPED_ACCESS (stmt_info) |
7374 | || !integer_zerop (get_dr_vinfo_offset (vinfo, dr_info)) |
7375 | || !integer_zerop (DR_INIT (dr_info->dr)) |
7376 | || !(ref_type = reference_alias_ptr_type (DR_REF (dr_info->dr))) |
7377 | || !alias_sets_conflict_p (get_alias_set (vectype), |
7378 | get_alias_set (TREE_TYPE (ref_type)))) |
7379 | { |
7380 | if (dump_enabled_p ()) |
7381 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
7382 | "unsupported OpenMP scan store.\n" ); |
7383 | return false; |
7384 | } |
7385 | |
7386 | /* We need to pattern match code built by OpenMP lowering and simplified |
7387 | by following optimizations into something we can handle. |
7388 | #pragma omp simd reduction(inscan,+:r) |
7389 | for (...) |
7390 | { |
7391 | r += something (); |
7392 | #pragma omp scan inclusive (r) |
7393 | use (r); |
7394 | } |
7395 | shall have body with: |
7396 | // Initialization for input phase, store the reduction initializer: |
7397 | _20 = .GOMP_SIMD_LANE (simduid.3_14(D), 0); |
7398 | _21 = .GOMP_SIMD_LANE (simduid.3_14(D), 1); |
7399 | D.2042[_21] = 0; |
7400 | // Actual input phase: |
7401 | ... |
7402 | r.0_5 = D.2042[_20]; |
7403 | _6 = _4 + r.0_5; |
7404 | D.2042[_20] = _6; |
7405 | // Initialization for scan phase: |
7406 | _25 = .GOMP_SIMD_LANE (simduid.3_14(D), 2); |
7407 | _26 = D.2043[_25]; |
7408 | _27 = D.2042[_25]; |
7409 | _28 = _26 + _27; |
7410 | D.2043[_25] = _28; |
7411 | D.2042[_25] = _28; |
7412 | // Actual scan phase: |
7413 | ... |
7414 | r.1_8 = D.2042[_20]; |
7415 | ... |
7416 | The "omp simd array" variable D.2042 holds the privatized copy used |
7417 | inside of the loop and D.2043 is another one that holds copies of |
7418 | the current original list item. The separate GOMP_SIMD_LANE ifn |
7419 | kinds are there in order to allow optimizing the initializer store |
7420 | and combiner sequence, e.g. if it is originally some C++ish user |
7421 | defined reduction, but allow the vectorizer to pattern recognize it |
7422 | and turn into the appropriate vectorized scan. |
7423 | |
7424 | For exclusive scan, this is slightly different: |
7425 | #pragma omp simd reduction(inscan,+:r) |
7426 | for (...) |
7427 | { |
7428 | use (r); |
7429 | #pragma omp scan exclusive (r) |
7430 | r += something (); |
7431 | } |
7432 | shall have body with: |
7433 | // Initialization for input phase, store the reduction initializer: |
7434 | _20 = .GOMP_SIMD_LANE (simduid.3_14(D), 0); |
7435 | _21 = .GOMP_SIMD_LANE (simduid.3_14(D), 1); |
7436 | D.2042[_21] = 0; |
7437 | // Actual input phase: |
7438 | ... |
7439 | r.0_5 = D.2042[_20]; |
7440 | _6 = _4 + r.0_5; |
7441 | D.2042[_20] = _6; |
7442 | // Initialization for scan phase: |
7443 | _25 = .GOMP_SIMD_LANE (simduid.3_14(D), 3); |
7444 | _26 = D.2043[_25]; |
7445 | D.2044[_25] = _26; |
7446 | _27 = D.2042[_25]; |
7447 | _28 = _26 + _27; |
7448 | D.2043[_25] = _28; |
7449 | // Actual scan phase: |
7450 | ... |
7451 | r.1_8 = D.2044[_20]; |
7452 | ... */ |
7453 | |
7454 | if (STMT_VINFO_SIMD_LANE_ACCESS_P (stmt_info) == 2) |
7455 | { |
7456 | /* Match the D.2042[_21] = 0; store above. Just require that |
7457 | it is a constant or external definition store. */ |
7458 | if (rhs_dt != vect_constant_def && rhs_dt != vect_external_def) |
7459 | { |
7460 | fail_init: |
7461 | if (dump_enabled_p ()) |
7462 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
7463 | "unsupported OpenMP scan initializer store.\n" ); |
7464 | return false; |
7465 | } |
7466 | |
7467 | if (! loop_vinfo->scan_map) |
7468 | loop_vinfo->scan_map = new hash_map<tree, tree>; |
7469 | tree var = TREE_OPERAND (DR_BASE_ADDRESS (dr_info->dr), 0); |
7470 | tree &cached = loop_vinfo->scan_map->get_or_insert (k: var); |
7471 | if (cached) |
7472 | goto fail_init; |
7473 | cached = gimple_assign_rhs1 (STMT_VINFO_STMT (stmt_info)); |
7474 | |
7475 | /* These stores can be vectorized normally. */ |
7476 | return true; |
7477 | } |
7478 | |
7479 | if (rhs_dt != vect_internal_def) |
7480 | { |
7481 | fail: |
7482 | if (dump_enabled_p ()) |
7483 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
7484 | "unsupported OpenMP scan combiner pattern.\n" ); |
7485 | return false; |
7486 | } |
7487 | |
7488 | gimple *stmt = STMT_VINFO_STMT (stmt_info); |
7489 | tree rhs = gimple_assign_rhs1 (gs: stmt); |
7490 | if (TREE_CODE (rhs) != SSA_NAME) |
7491 | goto fail; |
7492 | |
7493 | gimple *other_store_stmt = NULL; |
7494 | tree var = TREE_OPERAND (DR_BASE_ADDRESS (dr_info->dr), 0); |
7495 | bool inscan_var_store |
7496 | = lookup_attribute (attr_name: "omp simd inscan" , DECL_ATTRIBUTES (var)) != NULL; |
7497 | |
7498 | if (STMT_VINFO_SIMD_LANE_ACCESS_P (stmt_info) == 4) |
7499 | { |
7500 | if (!inscan_var_store) |
7501 | { |
7502 | use_operand_p use_p; |
7503 | imm_use_iterator iter; |
7504 | FOR_EACH_IMM_USE_FAST (use_p, iter, rhs) |
7505 | { |
7506 | gimple *use_stmt = USE_STMT (use_p); |
7507 | if (use_stmt == stmt || is_gimple_debug (gs: use_stmt)) |
7508 | continue; |
7509 | if (gimple_bb (g: use_stmt) != gimple_bb (g: stmt) |
7510 | || !is_gimple_assign (gs: use_stmt) |
7511 | || gimple_assign_rhs_class (gs: use_stmt) != GIMPLE_BINARY_RHS |
7512 | || other_store_stmt |
7513 | || TREE_CODE (gimple_assign_lhs (use_stmt)) != SSA_NAME) |
7514 | goto fail; |
7515 | other_store_stmt = use_stmt; |
7516 | } |
7517 | if (other_store_stmt == NULL) |
7518 | goto fail; |
7519 | rhs = gimple_assign_lhs (gs: other_store_stmt); |
7520 | if (!single_imm_use (var: rhs, use_p: &use_p, stmt: &other_store_stmt)) |
7521 | goto fail; |
7522 | } |
7523 | } |
7524 | else if (STMT_VINFO_SIMD_LANE_ACCESS_P (stmt_info) == 3) |
7525 | { |
7526 | use_operand_p use_p; |
7527 | imm_use_iterator iter; |
7528 | FOR_EACH_IMM_USE_FAST (use_p, iter, rhs) |
7529 | { |
7530 | gimple *use_stmt = USE_STMT (use_p); |
7531 | if (use_stmt == stmt || is_gimple_debug (gs: use_stmt)) |
7532 | continue; |
7533 | if (other_store_stmt) |
7534 | goto fail; |
7535 | other_store_stmt = use_stmt; |
7536 | } |
7537 | } |
7538 | else |
7539 | goto fail; |
7540 | |
7541 | gimple *def_stmt = SSA_NAME_DEF_STMT (rhs); |
7542 | if (gimple_bb (g: def_stmt) != gimple_bb (g: stmt) |
7543 | || !is_gimple_assign (gs: def_stmt) |
7544 | || gimple_assign_rhs_class (gs: def_stmt) != GIMPLE_BINARY_RHS) |
7545 | goto fail; |
7546 | |
7547 | enum tree_code code = gimple_assign_rhs_code (gs: def_stmt); |
7548 | /* For pointer addition, we should use the normal plus for the vector |
7549 | operation. */ |
7550 | switch (code) |
7551 | { |
7552 | case POINTER_PLUS_EXPR: |
7553 | code = PLUS_EXPR; |
7554 | break; |
7555 | case MULT_HIGHPART_EXPR: |
7556 | goto fail; |
7557 | default: |
7558 | break; |
7559 | } |
7560 | if (TREE_CODE_LENGTH (code) != binary_op || !commutative_tree_code (code)) |
7561 | goto fail; |
7562 | |
7563 | tree rhs1 = gimple_assign_rhs1 (gs: def_stmt); |
7564 | tree rhs2 = gimple_assign_rhs2 (gs: def_stmt); |
7565 | if (TREE_CODE (rhs1) != SSA_NAME || TREE_CODE (rhs2) != SSA_NAME) |
7566 | goto fail; |
7567 | |
7568 | gimple *load1_stmt = SSA_NAME_DEF_STMT (rhs1); |
7569 | gimple *load2_stmt = SSA_NAME_DEF_STMT (rhs2); |
7570 | if (gimple_bb (g: load1_stmt) != gimple_bb (g: stmt) |
7571 | || !gimple_assign_load_p (load1_stmt) |
7572 | || gimple_bb (g: load2_stmt) != gimple_bb (g: stmt) |
7573 | || !gimple_assign_load_p (load2_stmt)) |
7574 | goto fail; |
7575 | |
7576 | stmt_vec_info load1_stmt_info = loop_vinfo->lookup_stmt (load1_stmt); |
7577 | stmt_vec_info load2_stmt_info = loop_vinfo->lookup_stmt (load2_stmt); |
7578 | if (load1_stmt_info == NULL |
7579 | || load2_stmt_info == NULL |
7580 | || (STMT_VINFO_SIMD_LANE_ACCESS_P (load1_stmt_info) |
7581 | != STMT_VINFO_SIMD_LANE_ACCESS_P (stmt_info)) |
7582 | || (STMT_VINFO_SIMD_LANE_ACCESS_P (load2_stmt_info) |
7583 | != STMT_VINFO_SIMD_LANE_ACCESS_P (stmt_info))) |
7584 | goto fail; |
7585 | |
7586 | if (STMT_VINFO_SIMD_LANE_ACCESS_P (stmt_info) == 4 && inscan_var_store) |
7587 | { |
7588 | dr_vec_info *load1_dr_info = STMT_VINFO_DR_INFO (load1_stmt_info); |
7589 | if (TREE_CODE (DR_BASE_ADDRESS (load1_dr_info->dr)) != ADDR_EXPR |
7590 | || !VAR_P (TREE_OPERAND (DR_BASE_ADDRESS (load1_dr_info->dr), 0))) |
7591 | goto fail; |
7592 | tree var1 = TREE_OPERAND (DR_BASE_ADDRESS (load1_dr_info->dr), 0); |
7593 | tree lrhs; |
7594 | if (lookup_attribute (attr_name: "omp simd inscan" , DECL_ATTRIBUTES (var1))) |
7595 | lrhs = rhs1; |
7596 | else |
7597 | lrhs = rhs2; |
7598 | use_operand_p use_p; |
7599 | imm_use_iterator iter; |
7600 | FOR_EACH_IMM_USE_FAST (use_p, iter, lrhs) |
7601 | { |
7602 | gimple *use_stmt = USE_STMT (use_p); |
7603 | if (use_stmt == def_stmt || is_gimple_debug (gs: use_stmt)) |
7604 | continue; |
7605 | if (other_store_stmt) |
7606 | goto fail; |
7607 | other_store_stmt = use_stmt; |
7608 | } |
7609 | } |
7610 | |
7611 | if (other_store_stmt == NULL) |
7612 | goto fail; |
7613 | if (gimple_bb (g: other_store_stmt) != gimple_bb (g: stmt) |
7614 | || !gimple_store_p (gs: other_store_stmt)) |
7615 | goto fail; |
7616 | |
7617 | stmt_vec_info other_store_stmt_info |
7618 | = loop_vinfo->lookup_stmt (other_store_stmt); |
7619 | if (other_store_stmt_info == NULL |
7620 | || (STMT_VINFO_SIMD_LANE_ACCESS_P (other_store_stmt_info) |
7621 | != STMT_VINFO_SIMD_LANE_ACCESS_P (stmt_info))) |
7622 | goto fail; |
7623 | |
7624 | gimple *stmt1 = stmt; |
7625 | gimple *stmt2 = other_store_stmt; |
7626 | if (STMT_VINFO_SIMD_LANE_ACCESS_P (stmt_info) == 4 && !inscan_var_store) |
7627 | std::swap (a&: stmt1, b&: stmt2); |
7628 | if (scan_operand_equal_p (ref1: gimple_assign_lhs (gs: stmt1), |
7629 | ref2: gimple_assign_rhs1 (gs: load2_stmt))) |
7630 | { |
7631 | std::swap (a&: rhs1, b&: rhs2); |
7632 | std::swap (a&: load1_stmt, b&: load2_stmt); |
7633 | std::swap (a&: load1_stmt_info, b&: load2_stmt_info); |
7634 | } |
7635 | if (!scan_operand_equal_p (ref1: gimple_assign_lhs (gs: stmt1), |
7636 | ref2: gimple_assign_rhs1 (gs: load1_stmt))) |
7637 | goto fail; |
7638 | |
7639 | tree var3 = NULL_TREE; |
7640 | if (STMT_VINFO_SIMD_LANE_ACCESS_P (stmt_info) == 3 |
7641 | && !scan_operand_equal_p (ref1: gimple_assign_lhs (gs: stmt2), |
7642 | ref2: gimple_assign_rhs1 (gs: load2_stmt))) |
7643 | goto fail; |
7644 | else if (STMT_VINFO_SIMD_LANE_ACCESS_P (stmt_info) == 4) |
7645 | { |
7646 | dr_vec_info *load2_dr_info = STMT_VINFO_DR_INFO (load2_stmt_info); |
7647 | if (TREE_CODE (DR_BASE_ADDRESS (load2_dr_info->dr)) != ADDR_EXPR |
7648 | || !VAR_P (TREE_OPERAND (DR_BASE_ADDRESS (load2_dr_info->dr), 0))) |
7649 | goto fail; |
7650 | var3 = TREE_OPERAND (DR_BASE_ADDRESS (load2_dr_info->dr), 0); |
7651 | if (!lookup_attribute (attr_name: "omp simd array" , DECL_ATTRIBUTES (var3)) |
7652 | || lookup_attribute (attr_name: "omp simd inscan" , DECL_ATTRIBUTES (var3)) |
7653 | || lookup_attribute (attr_name: "omp simd inscan exclusive" , |
7654 | DECL_ATTRIBUTES (var3))) |
7655 | goto fail; |
7656 | } |
7657 | |
7658 | dr_vec_info *other_dr_info = STMT_VINFO_DR_INFO (other_store_stmt_info); |
7659 | if (TREE_CODE (DR_BASE_ADDRESS (other_dr_info->dr)) != ADDR_EXPR |
7660 | || !VAR_P (TREE_OPERAND (DR_BASE_ADDRESS (other_dr_info->dr), 0))) |
7661 | goto fail; |
7662 | |
7663 | tree var1 = TREE_OPERAND (DR_BASE_ADDRESS (dr_info->dr), 0); |
7664 | tree var2 = TREE_OPERAND (DR_BASE_ADDRESS (other_dr_info->dr), 0); |
7665 | if (!lookup_attribute (attr_name: "omp simd array" , DECL_ATTRIBUTES (var1)) |
7666 | || !lookup_attribute (attr_name: "omp simd array" , DECL_ATTRIBUTES (var2)) |
7667 | || (!lookup_attribute (attr_name: "omp simd inscan" , DECL_ATTRIBUTES (var1))) |
7668 | == (!lookup_attribute (attr_name: "omp simd inscan" , DECL_ATTRIBUTES (var2)))) |
7669 | goto fail; |
7670 | |
7671 | if (lookup_attribute (attr_name: "omp simd inscan" , DECL_ATTRIBUTES (var1))) |
7672 | std::swap (a&: var1, b&: var2); |
7673 | |
7674 | if (STMT_VINFO_SIMD_LANE_ACCESS_P (stmt_info) == 4) |
7675 | { |
7676 | if (!lookup_attribute (attr_name: "omp simd inscan exclusive" , |
7677 | DECL_ATTRIBUTES (var1))) |
7678 | goto fail; |
7679 | var1 = var3; |
7680 | } |
7681 | |
7682 | if (loop_vinfo->scan_map == NULL) |
7683 | goto fail; |
7684 | tree *init = loop_vinfo->scan_map->get (k: var1); |
7685 | if (init == NULL) |
7686 | goto fail; |
7687 | |
7688 | /* The IL is as expected, now check if we can actually vectorize it. |
7689 | Inclusive scan: |
7690 | _26 = D.2043[_25]; |
7691 | _27 = D.2042[_25]; |
7692 | _28 = _26 + _27; |
7693 | D.2043[_25] = _28; |
7694 | D.2042[_25] = _28; |
7695 | should be vectorized as (where _40 is the vectorized rhs |
7696 | from the D.2042[_21] = 0; store): |
7697 | _30 = MEM <vector(8) int> [(int *)&D.2043]; |
7698 | _31 = MEM <vector(8) int> [(int *)&D.2042]; |
7699 | _32 = VEC_PERM_EXPR <_40, _31, { 0, 8, 9, 10, 11, 12, 13, 14 }>; |
7700 | _33 = _31 + _32; |
7701 | // _33 = { _31[0], _31[0]+_31[1], _31[1]+_31[2], ..., _31[6]+_31[7] }; |
7702 | _34 = VEC_PERM_EXPR <_40, _33, { 0, 1, 8, 9, 10, 11, 12, 13 }>; |
7703 | _35 = _33 + _34; |
7704 | // _35 = { _31[0], _31[0]+_31[1], _31[0]+.._31[2], _31[0]+.._31[3], |
7705 | // _31[1]+.._31[4], ... _31[4]+.._31[7] }; |
7706 | _36 = VEC_PERM_EXPR <_40, _35, { 0, 1, 2, 3, 8, 9, 10, 11 }>; |
7707 | _37 = _35 + _36; |
7708 | // _37 = { _31[0], _31[0]+_31[1], _31[0]+.._31[2], _31[0]+.._31[3], |
7709 | // _31[0]+.._31[4], ... _31[0]+.._31[7] }; |
7710 | _38 = _30 + _37; |
7711 | _39 = VEC_PERM_EXPR <_38, _38, { 7, 7, 7, 7, 7, 7, 7, 7 }>; |
7712 | MEM <vector(8) int> [(int *)&D.2043] = _39; |
7713 | MEM <vector(8) int> [(int *)&D.2042] = _38; |
7714 | Exclusive scan: |
7715 | _26 = D.2043[_25]; |
7716 | D.2044[_25] = _26; |
7717 | _27 = D.2042[_25]; |
7718 | _28 = _26 + _27; |
7719 | D.2043[_25] = _28; |
7720 | should be vectorized as (where _40 is the vectorized rhs |
7721 | from the D.2042[_21] = 0; store): |
7722 | _30 = MEM <vector(8) int> [(int *)&D.2043]; |
7723 | _31 = MEM <vector(8) int> [(int *)&D.2042]; |
7724 | _32 = VEC_PERM_EXPR <_40, _31, { 0, 8, 9, 10, 11, 12, 13, 14 }>; |
7725 | _33 = VEC_PERM_EXPR <_40, _32, { 0, 8, 9, 10, 11, 12, 13, 14 }>; |
7726 | _34 = _32 + _33; |
7727 | // _34 = { 0, _31[0], _31[0]+_31[1], _31[1]+_31[2], _31[2]+_31[3], |
7728 | // _31[3]+_31[4], ... _31[5]+.._31[6] }; |
7729 | _35 = VEC_PERM_EXPR <_40, _34, { 0, 1, 8, 9, 10, 11, 12, 13 }>; |
7730 | _36 = _34 + _35; |
7731 | // _36 = { 0, _31[0], _31[0]+_31[1], _31[0]+.._31[2], _31[0]+.._31[3], |
7732 | // _31[1]+.._31[4], ... _31[3]+.._31[6] }; |
7733 | _37 = VEC_PERM_EXPR <_40, _36, { 0, 1, 2, 3, 8, 9, 10, 11 }>; |
7734 | _38 = _36 + _37; |
7735 | // _38 = { 0, _31[0], _31[0]+_31[1], _31[0]+.._31[2], _31[0]+.._31[3], |
7736 | // _31[0]+.._31[4], ... _31[0]+.._31[6] }; |
7737 | _39 = _30 + _38; |
7738 | _50 = _31 + _39; |
7739 | _51 = VEC_PERM_EXPR <_50, _50, { 7, 7, 7, 7, 7, 7, 7, 7 }>; |
7740 | MEM <vector(8) int> [(int *)&D.2044] = _39; |
7741 | MEM <vector(8) int> [(int *)&D.2042] = _51; */ |
7742 | enum machine_mode vec_mode = TYPE_MODE (vectype); |
7743 | optab optab = optab_for_tree_code (code, vectype, optab_default); |
7744 | if (!optab || optab_handler (op: optab, mode: vec_mode) == CODE_FOR_nothing) |
7745 | goto fail; |
7746 | |
7747 | int units_log2 = scan_store_can_perm_p (vectype, init: *init); |
7748 | if (units_log2 == -1) |
7749 | goto fail; |
7750 | |
7751 | return true; |
7752 | } |
7753 | |
7754 | |
7755 | /* Function vectorizable_scan_store. |
7756 | |
7757 | Helper of vectorizable_score, arguments like on vectorizable_store. |
7758 | Handle only the transformation, checking is done in check_scan_store. */ |
7759 | |
7760 | static bool |
7761 | vectorizable_scan_store (vec_info *vinfo, |
7762 | stmt_vec_info stmt_info, gimple_stmt_iterator *gsi, |
7763 | gimple **vec_stmt, int ncopies) |
7764 | { |
7765 | loop_vec_info loop_vinfo = dyn_cast <loop_vec_info> (p: vinfo); |
7766 | dr_vec_info *dr_info = STMT_VINFO_DR_INFO (stmt_info); |
7767 | tree ref_type = reference_alias_ptr_type (DR_REF (dr_info->dr)); |
7768 | tree vectype = STMT_VINFO_VECTYPE (stmt_info); |
7769 | |
7770 | if (dump_enabled_p ()) |
7771 | dump_printf_loc (MSG_NOTE, vect_location, |
7772 | "transform scan store. ncopies = %d\n" , ncopies); |
7773 | |
7774 | gimple *stmt = STMT_VINFO_STMT (stmt_info); |
7775 | tree rhs = gimple_assign_rhs1 (gs: stmt); |
7776 | gcc_assert (TREE_CODE (rhs) == SSA_NAME); |
7777 | |
7778 | tree var = TREE_OPERAND (DR_BASE_ADDRESS (dr_info->dr), 0); |
7779 | bool inscan_var_store |
7780 | = lookup_attribute (attr_name: "omp simd inscan" , DECL_ATTRIBUTES (var)) != NULL; |
7781 | |
7782 | if (STMT_VINFO_SIMD_LANE_ACCESS_P (stmt_info) == 4 && !inscan_var_store) |
7783 | { |
7784 | use_operand_p use_p; |
7785 | imm_use_iterator iter; |
7786 | FOR_EACH_IMM_USE_FAST (use_p, iter, rhs) |
7787 | { |
7788 | gimple *use_stmt = USE_STMT (use_p); |
7789 | if (use_stmt == stmt || is_gimple_debug (gs: use_stmt)) |
7790 | continue; |
7791 | rhs = gimple_assign_lhs (gs: use_stmt); |
7792 | break; |
7793 | } |
7794 | } |
7795 | |
7796 | gimple *def_stmt = SSA_NAME_DEF_STMT (rhs); |
7797 | enum tree_code code = gimple_assign_rhs_code (gs: def_stmt); |
7798 | if (code == POINTER_PLUS_EXPR) |
7799 | code = PLUS_EXPR; |
7800 | gcc_assert (TREE_CODE_LENGTH (code) == binary_op |
7801 | && commutative_tree_code (code)); |
7802 | tree rhs1 = gimple_assign_rhs1 (gs: def_stmt); |
7803 | tree rhs2 = gimple_assign_rhs2 (gs: def_stmt); |
7804 | gcc_assert (TREE_CODE (rhs1) == SSA_NAME && TREE_CODE (rhs2) == SSA_NAME); |
7805 | gimple *load1_stmt = SSA_NAME_DEF_STMT (rhs1); |
7806 | gimple *load2_stmt = SSA_NAME_DEF_STMT (rhs2); |
7807 | stmt_vec_info load1_stmt_info = loop_vinfo->lookup_stmt (load1_stmt); |
7808 | stmt_vec_info load2_stmt_info = loop_vinfo->lookup_stmt (load2_stmt); |
7809 | dr_vec_info *load1_dr_info = STMT_VINFO_DR_INFO (load1_stmt_info); |
7810 | dr_vec_info *load2_dr_info = STMT_VINFO_DR_INFO (load2_stmt_info); |
7811 | tree var1 = TREE_OPERAND (DR_BASE_ADDRESS (load1_dr_info->dr), 0); |
7812 | tree var2 = TREE_OPERAND (DR_BASE_ADDRESS (load2_dr_info->dr), 0); |
7813 | |
7814 | if (lookup_attribute (attr_name: "omp simd inscan" , DECL_ATTRIBUTES (var1))) |
7815 | { |
7816 | std::swap (a&: rhs1, b&: rhs2); |
7817 | std::swap (a&: var1, b&: var2); |
7818 | std::swap (a&: load1_dr_info, b&: load2_dr_info); |
7819 | } |
7820 | |
7821 | tree *init = loop_vinfo->scan_map->get (k: var1); |
7822 | gcc_assert (init); |
7823 | |
7824 | unsigned HOST_WIDE_INT nunits; |
7825 | if (!TYPE_VECTOR_SUBPARTS (node: vectype).is_constant (const_value: &nunits)) |
7826 | gcc_unreachable (); |
7827 | auto_vec<enum scan_store_kind, 16> use_whole_vector; |
7828 | int units_log2 = scan_store_can_perm_p (vectype, init: *init, use_whole_vector: &use_whole_vector); |
7829 | gcc_assert (units_log2 > 0); |
7830 | auto_vec<tree, 16> perms; |
7831 | perms.quick_grow (len: units_log2 + 1); |
7832 | tree zero_vec = NULL_TREE, masktype = NULL_TREE; |
7833 | for (int i = 0; i <= units_log2; ++i) |
7834 | { |
7835 | unsigned HOST_WIDE_INT j, k; |
7836 | vec_perm_builder sel (nunits, nunits, 1); |
7837 | sel.quick_grow (len: nunits); |
7838 | if (i == units_log2) |
7839 | for (j = 0; j < nunits; ++j) |
7840 | sel[j] = nunits - 1; |
7841 | else |
7842 | { |
7843 | for (j = 0; j < (HOST_WIDE_INT_1U << i); ++j) |
7844 | sel[j] = j; |
7845 | for (k = 0; j < nunits; ++j, ++k) |
7846 | sel[j] = nunits + k; |
7847 | } |
7848 | vec_perm_indices indices (sel, i == units_log2 ? 1 : 2, nunits); |
7849 | if (!use_whole_vector.is_empty () |
7850 | && use_whole_vector[i] != scan_store_kind_perm) |
7851 | { |
7852 | if (zero_vec == NULL_TREE) |
7853 | zero_vec = build_zero_cst (vectype); |
7854 | if (masktype == NULL_TREE |
7855 | && use_whole_vector[i] == scan_store_kind_lshift_cond) |
7856 | masktype = truth_type_for (vectype); |
7857 | perms[i] = vect_gen_perm_mask_any (vectype, indices); |
7858 | } |
7859 | else |
7860 | perms[i] = vect_gen_perm_mask_checked (vectype, indices); |
7861 | } |
7862 | |
7863 | tree vec_oprnd1 = NULL_TREE; |
7864 | tree vec_oprnd2 = NULL_TREE; |
7865 | tree vec_oprnd3 = NULL_TREE; |
7866 | tree dataref_ptr = DR_BASE_ADDRESS (dr_info->dr); |
7867 | tree dataref_offset = build_int_cst (ref_type, 0); |
7868 | tree bump = vect_get_data_ptr_increment (vinfo, gsi, dr_info, |
7869 | aggr_type: vectype, memory_access_type: VMAT_CONTIGUOUS); |
7870 | tree ldataref_ptr = NULL_TREE; |
7871 | tree orig = NULL_TREE; |
7872 | if (STMT_VINFO_SIMD_LANE_ACCESS_P (stmt_info) == 4 && !inscan_var_store) |
7873 | ldataref_ptr = DR_BASE_ADDRESS (load1_dr_info->dr); |
7874 | auto_vec<tree> vec_oprnds1; |
7875 | auto_vec<tree> vec_oprnds2; |
7876 | auto_vec<tree> vec_oprnds3; |
7877 | vect_get_vec_defs (vinfo, stmt_info, NULL, ncopies, |
7878 | op0: *init, vec_oprnds0: &vec_oprnds1, |
7879 | op1: ldataref_ptr == NULL ? rhs1 : NULL, vec_oprnds1: &vec_oprnds2, |
7880 | op2: rhs2, vec_oprnds2: &vec_oprnds3); |
7881 | for (int j = 0; j < ncopies; j++) |
7882 | { |
7883 | vec_oprnd1 = vec_oprnds1[j]; |
7884 | if (ldataref_ptr == NULL) |
7885 | vec_oprnd2 = vec_oprnds2[j]; |
7886 | vec_oprnd3 = vec_oprnds3[j]; |
7887 | if (j == 0) |
7888 | orig = vec_oprnd3; |
7889 | else if (!inscan_var_store) |
7890 | dataref_offset = int_const_binop (PLUS_EXPR, dataref_offset, bump); |
7891 | |
7892 | if (ldataref_ptr) |
7893 | { |
7894 | vec_oprnd2 = make_ssa_name (var: vectype); |
7895 | tree data_ref = fold_build2 (MEM_REF, vectype, |
7896 | unshare_expr (ldataref_ptr), |
7897 | dataref_offset); |
7898 | vect_copy_ref_info (data_ref, DR_REF (load1_dr_info->dr)); |
7899 | gimple *g = gimple_build_assign (vec_oprnd2, data_ref); |
7900 | vect_finish_stmt_generation (vinfo, stmt_info, vec_stmt: g, gsi); |
7901 | STMT_VINFO_VEC_STMTS (stmt_info).safe_push (obj: g); |
7902 | *vec_stmt = STMT_VINFO_VEC_STMTS (stmt_info)[0]; |
7903 | } |
7904 | |
7905 | tree v = vec_oprnd2; |
7906 | for (int i = 0; i < units_log2; ++i) |
7907 | { |
7908 | tree new_temp = make_ssa_name (var: vectype); |
7909 | gimple *g = gimple_build_assign (new_temp, VEC_PERM_EXPR, |
7910 | (zero_vec |
7911 | && (use_whole_vector[i] |
7912 | != scan_store_kind_perm)) |
7913 | ? zero_vec : vec_oprnd1, v, |
7914 | perms[i]); |
7915 | vect_finish_stmt_generation (vinfo, stmt_info, vec_stmt: g, gsi); |
7916 | STMT_VINFO_VEC_STMTS (stmt_info).safe_push (obj: g); |
7917 | *vec_stmt = STMT_VINFO_VEC_STMTS (stmt_info)[0]; |
7918 | |
7919 | if (zero_vec && use_whole_vector[i] == scan_store_kind_lshift_cond) |
7920 | { |
7921 | /* Whole vector shift shifted in zero bits, but if *init |
7922 | is not initializer_zerop, we need to replace those elements |
7923 | with elements from vec_oprnd1. */ |
7924 | tree_vector_builder vb (masktype, nunits, 1); |
7925 | for (unsigned HOST_WIDE_INT k = 0; k < nunits; ++k) |
7926 | vb.quick_push (obj: k < (HOST_WIDE_INT_1U << i) |
7927 | ? boolean_false_node : boolean_true_node); |
7928 | |
7929 | tree new_temp2 = make_ssa_name (var: vectype); |
7930 | g = gimple_build_assign (new_temp2, VEC_COND_EXPR, vb.build (), |
7931 | new_temp, vec_oprnd1); |
7932 | vect_finish_stmt_generation (vinfo, stmt_info, |
7933 | vec_stmt: g, gsi); |
7934 | STMT_VINFO_VEC_STMTS (stmt_info).safe_push (obj: g); |
7935 | new_temp = new_temp2; |
7936 | } |
7937 | |
7938 | /* For exclusive scan, perform the perms[i] permutation once |
7939 | more. */ |
7940 | if (i == 0 |
7941 | && STMT_VINFO_SIMD_LANE_ACCESS_P (stmt_info) == 4 |
7942 | && v == vec_oprnd2) |
7943 | { |
7944 | v = new_temp; |
7945 | --i; |
7946 | continue; |
7947 | } |
7948 | |
7949 | tree new_temp2 = make_ssa_name (var: vectype); |
7950 | g = gimple_build_assign (new_temp2, code, v, new_temp); |
7951 | vect_finish_stmt_generation (vinfo, stmt_info, vec_stmt: g, gsi); |
7952 | STMT_VINFO_VEC_STMTS (stmt_info).safe_push (obj: g); |
7953 | |
7954 | v = new_temp2; |
7955 | } |
7956 | |
7957 | tree new_temp = make_ssa_name (var: vectype); |
7958 | gimple *g = gimple_build_assign (new_temp, code, orig, v); |
7959 | vect_finish_stmt_generation (vinfo, stmt_info, vec_stmt: g, gsi); |
7960 | STMT_VINFO_VEC_STMTS (stmt_info).safe_push (obj: g); |
7961 | |
7962 | tree last_perm_arg = new_temp; |
7963 | /* For exclusive scan, new_temp computed above is the exclusive scan |
7964 | prefix sum. Turn it into inclusive prefix sum for the broadcast |
7965 | of the last element into orig. */ |
7966 | if (STMT_VINFO_SIMD_LANE_ACCESS_P (stmt_info) == 4) |
7967 | { |
7968 | last_perm_arg = make_ssa_name (var: vectype); |
7969 | g = gimple_build_assign (last_perm_arg, code, new_temp, vec_oprnd2); |
7970 | vect_finish_stmt_generation (vinfo, stmt_info, vec_stmt: g, gsi); |
7971 | STMT_VINFO_VEC_STMTS (stmt_info).safe_push (obj: g); |
7972 | } |
7973 | |
7974 | orig = make_ssa_name (var: vectype); |
7975 | g = gimple_build_assign (orig, VEC_PERM_EXPR, last_perm_arg, |
7976 | last_perm_arg, perms[units_log2]); |
7977 | vect_finish_stmt_generation (vinfo, stmt_info, vec_stmt: g, gsi); |
7978 | STMT_VINFO_VEC_STMTS (stmt_info).safe_push (obj: g); |
7979 | |
7980 | if (!inscan_var_store) |
7981 | { |
7982 | tree data_ref = fold_build2 (MEM_REF, vectype, |
7983 | unshare_expr (dataref_ptr), |
7984 | dataref_offset); |
7985 | vect_copy_ref_info (data_ref, DR_REF (dr_info->dr)); |
7986 | g = gimple_build_assign (data_ref, new_temp); |
7987 | vect_finish_stmt_generation (vinfo, stmt_info, vec_stmt: g, gsi); |
7988 | STMT_VINFO_VEC_STMTS (stmt_info).safe_push (obj: g); |
7989 | } |
7990 | } |
7991 | |
7992 | if (inscan_var_store) |
7993 | for (int j = 0; j < ncopies; j++) |
7994 | { |
7995 | if (j != 0) |
7996 | dataref_offset = int_const_binop (PLUS_EXPR, dataref_offset, bump); |
7997 | |
7998 | tree data_ref = fold_build2 (MEM_REF, vectype, |
7999 | unshare_expr (dataref_ptr), |
8000 | dataref_offset); |
8001 | vect_copy_ref_info (data_ref, DR_REF (dr_info->dr)); |
8002 | gimple *g = gimple_build_assign (data_ref, orig); |
8003 | vect_finish_stmt_generation (vinfo, stmt_info, vec_stmt: g, gsi); |
8004 | STMT_VINFO_VEC_STMTS (stmt_info).safe_push (obj: g); |
8005 | } |
8006 | return true; |
8007 | } |
8008 | |
8009 | |
8010 | /* Function vectorizable_store. |
8011 | |
8012 | Check if STMT_INFO defines a non scalar data-ref (array/pointer/structure) |
8013 | that can be vectorized. |
8014 | If VEC_STMT is also passed, vectorize STMT_INFO: create a vectorized |
8015 | stmt to replace it, put it in VEC_STMT, and insert it at GSI. |
8016 | Return true if STMT_INFO is vectorizable in this way. */ |
8017 | |
8018 | static bool |
8019 | vectorizable_store (vec_info *vinfo, |
8020 | stmt_vec_info stmt_info, gimple_stmt_iterator *gsi, |
8021 | gimple **vec_stmt, slp_tree slp_node, |
8022 | stmt_vector_for_cost *cost_vec) |
8023 | { |
8024 | tree data_ref; |
8025 | tree vec_oprnd = NULL_TREE; |
8026 | tree elem_type; |
8027 | loop_vec_info loop_vinfo = dyn_cast <loop_vec_info> (p: vinfo); |
8028 | class loop *loop = NULL; |
8029 | machine_mode vec_mode; |
8030 | tree dummy; |
8031 | enum vect_def_type rhs_dt = vect_unknown_def_type; |
8032 | enum vect_def_type mask_dt = vect_unknown_def_type; |
8033 | tree dataref_ptr = NULL_TREE; |
8034 | tree dataref_offset = NULL_TREE; |
8035 | gimple *ptr_incr = NULL; |
8036 | int ncopies; |
8037 | int j; |
8038 | stmt_vec_info first_stmt_info; |
8039 | bool grouped_store; |
8040 | unsigned int group_size, i; |
8041 | bool slp = (slp_node != NULL); |
8042 | unsigned int vec_num; |
8043 | bb_vec_info bb_vinfo = dyn_cast <bb_vec_info> (p: vinfo); |
8044 | tree aggr_type; |
8045 | gather_scatter_info gs_info; |
8046 | poly_uint64 vf; |
8047 | vec_load_store_type vls_type; |
8048 | tree ref_type; |
8049 | |
8050 | if (!STMT_VINFO_RELEVANT_P (stmt_info) && !bb_vinfo) |
8051 | return false; |
8052 | |
8053 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def |
8054 | && ! vec_stmt) |
8055 | return false; |
8056 | |
8057 | /* Is vectorizable store? */ |
8058 | |
8059 | tree mask = NULL_TREE, mask_vectype = NULL_TREE; |
8060 | slp_tree mask_node = NULL; |
8061 | if (gassign *assign = dyn_cast <gassign *> (p: stmt_info->stmt)) |
8062 | { |
8063 | tree scalar_dest = gimple_assign_lhs (gs: assign); |
8064 | if (TREE_CODE (scalar_dest) == VIEW_CONVERT_EXPR |
8065 | && is_pattern_stmt_p (stmt_info)) |
8066 | scalar_dest = TREE_OPERAND (scalar_dest, 0); |
8067 | if (TREE_CODE (scalar_dest) != ARRAY_REF |
8068 | && TREE_CODE (scalar_dest) != BIT_FIELD_REF |
8069 | && TREE_CODE (scalar_dest) != INDIRECT_REF |
8070 | && TREE_CODE (scalar_dest) != COMPONENT_REF |
8071 | && TREE_CODE (scalar_dest) != IMAGPART_EXPR |
8072 | && TREE_CODE (scalar_dest) != REALPART_EXPR |
8073 | && TREE_CODE (scalar_dest) != MEM_REF) |
8074 | return false; |
8075 | } |
8076 | else |
8077 | { |
8078 | gcall *call = dyn_cast <gcall *> (p: stmt_info->stmt); |
8079 | if (!call || !gimple_call_internal_p (gs: call)) |
8080 | return false; |
8081 | |
8082 | internal_fn ifn = gimple_call_internal_fn (gs: call); |
8083 | if (!internal_store_fn_p (ifn)) |
8084 | return false; |
8085 | |
8086 | int mask_index = internal_fn_mask_index (ifn); |
8087 | if (mask_index >= 0 && slp_node) |
8088 | mask_index = vect_slp_child_index_for_operand |
8089 | (call, op: mask_index, STMT_VINFO_GATHER_SCATTER_P (stmt_info)); |
8090 | if (mask_index >= 0 |
8091 | && !vect_check_scalar_mask (vinfo, stmt_info, slp_node, mask_index, |
8092 | mask: &mask, mask_node: &mask_node, mask_dt_out: &mask_dt, |
8093 | mask_vectype_out: &mask_vectype)) |
8094 | return false; |
8095 | } |
8096 | |
8097 | /* Cannot have hybrid store SLP -- that would mean storing to the |
8098 | same location twice. */ |
8099 | gcc_assert (slp == PURE_SLP_STMT (stmt_info)); |
8100 | |
8101 | tree vectype = STMT_VINFO_VECTYPE (stmt_info), rhs_vectype = NULL_TREE; |
8102 | poly_uint64 nunits = TYPE_VECTOR_SUBPARTS (node: vectype); |
8103 | |
8104 | if (loop_vinfo) |
8105 | { |
8106 | loop = LOOP_VINFO_LOOP (loop_vinfo); |
8107 | vf = LOOP_VINFO_VECT_FACTOR (loop_vinfo); |
8108 | } |
8109 | else |
8110 | vf = 1; |
8111 | |
8112 | /* Multiple types in SLP are handled by creating the appropriate number of |
8113 | vectorized stmts for each SLP node. Hence, NCOPIES is always 1 in |
8114 | case of SLP. */ |
8115 | if (slp) |
8116 | ncopies = 1; |
8117 | else |
8118 | ncopies = vect_get_num_copies (loop_vinfo, vectype); |
8119 | |
8120 | gcc_assert (ncopies >= 1); |
8121 | |
8122 | /* FORNOW. This restriction should be relaxed. */ |
8123 | if (loop && nested_in_vect_loop_p (loop, stmt_info) && ncopies > 1) |
8124 | { |
8125 | if (dump_enabled_p ()) |
8126 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
8127 | "multiple types in nested loop.\n" ); |
8128 | return false; |
8129 | } |
8130 | |
8131 | tree op; |
8132 | slp_tree op_node; |
8133 | if (!vect_check_store_rhs (vinfo, stmt_info, slp_node, |
8134 | rhs: &op, rhs_node: &op_node, rhs_dt_out: &rhs_dt, rhs_vectype_out: &rhs_vectype, vls_type_out: &vls_type)) |
8135 | return false; |
8136 | |
8137 | elem_type = TREE_TYPE (vectype); |
8138 | vec_mode = TYPE_MODE (vectype); |
8139 | |
8140 | if (!STMT_VINFO_DATA_REF (stmt_info)) |
8141 | return false; |
8142 | |
8143 | vect_memory_access_type memory_access_type; |
8144 | enum dr_alignment_support alignment_support_scheme; |
8145 | int misalignment; |
8146 | poly_int64 poffset; |
8147 | internal_fn lanes_ifn; |
8148 | if (!get_load_store_type (vinfo, stmt_info, vectype, slp_node, masked_p: mask, vls_type, |
8149 | ncopies, memory_access_type: &memory_access_type, poffset: &poffset, |
8150 | alignment_support_scheme: &alignment_support_scheme, misalignment: &misalignment, gs_info: &gs_info, |
8151 | lanes_ifn: &lanes_ifn)) |
8152 | return false; |
8153 | |
8154 | if (mask) |
8155 | { |
8156 | if (memory_access_type == VMAT_CONTIGUOUS) |
8157 | { |
8158 | if (!VECTOR_MODE_P (vec_mode) |
8159 | || !can_vec_mask_load_store_p (vec_mode, |
8160 | TYPE_MODE (mask_vectype), false)) |
8161 | return false; |
8162 | } |
8163 | else if (memory_access_type != VMAT_LOAD_STORE_LANES |
8164 | && (memory_access_type != VMAT_GATHER_SCATTER |
8165 | || (gs_info.decl && !VECTOR_BOOLEAN_TYPE_P (mask_vectype)))) |
8166 | { |
8167 | if (dump_enabled_p ()) |
8168 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
8169 | "unsupported access type for masked store.\n" ); |
8170 | return false; |
8171 | } |
8172 | else if (memory_access_type == VMAT_GATHER_SCATTER |
8173 | && gs_info.ifn == IFN_LAST |
8174 | && !gs_info.decl) |
8175 | { |
8176 | if (dump_enabled_p ()) |
8177 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
8178 | "unsupported masked emulated scatter.\n" ); |
8179 | return false; |
8180 | } |
8181 | } |
8182 | else |
8183 | { |
8184 | /* FORNOW. In some cases can vectorize even if data-type not supported |
8185 | (e.g. - array initialization with 0). */ |
8186 | if (optab_handler (op: mov_optab, mode: vec_mode) == CODE_FOR_nothing) |
8187 | return false; |
8188 | } |
8189 | |
8190 | dr_vec_info *dr_info = STMT_VINFO_DR_INFO (stmt_info), *first_dr_info = NULL; |
8191 | grouped_store = (STMT_VINFO_GROUPED_ACCESS (stmt_info) |
8192 | && memory_access_type != VMAT_GATHER_SCATTER |
8193 | && (slp || memory_access_type != VMAT_CONTIGUOUS)); |
8194 | if (grouped_store) |
8195 | { |
8196 | first_stmt_info = DR_GROUP_FIRST_ELEMENT (stmt_info); |
8197 | first_dr_info = STMT_VINFO_DR_INFO (first_stmt_info); |
8198 | group_size = DR_GROUP_SIZE (first_stmt_info); |
8199 | } |
8200 | else |
8201 | { |
8202 | first_stmt_info = stmt_info; |
8203 | first_dr_info = dr_info; |
8204 | group_size = vec_num = 1; |
8205 | } |
8206 | |
8207 | if (STMT_VINFO_SIMD_LANE_ACCESS_P (stmt_info) > 1 && !vec_stmt) |
8208 | { |
8209 | if (!check_scan_store (vinfo, stmt_info, vectype, rhs_dt, slp, mask, |
8210 | memory_access_type)) |
8211 | return false; |
8212 | } |
8213 | |
8214 | bool costing_p = !vec_stmt; |
8215 | if (costing_p) /* transformation not required. */ |
8216 | { |
8217 | STMT_VINFO_MEMORY_ACCESS_TYPE (stmt_info) = memory_access_type; |
8218 | |
8219 | if (loop_vinfo |
8220 | && LOOP_VINFO_CAN_USE_PARTIAL_VECTORS_P (loop_vinfo)) |
8221 | check_load_store_for_partial_vectors (loop_vinfo, vectype, slp_node, |
8222 | vls_type, group_size, |
8223 | memory_access_type, gs_info: &gs_info, |
8224 | scalar_mask: mask); |
8225 | |
8226 | if (slp_node |
8227 | && (!vect_maybe_update_slp_op_vectype (op_node, vectype) |
8228 | || (mask |
8229 | && !vect_maybe_update_slp_op_vectype (mask_node, |
8230 | mask_vectype)))) |
8231 | { |
8232 | if (dump_enabled_p ()) |
8233 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
8234 | "incompatible vector types for invariants\n" ); |
8235 | return false; |
8236 | } |
8237 | |
8238 | if (dump_enabled_p () |
8239 | && memory_access_type != VMAT_ELEMENTWISE |
8240 | && memory_access_type != VMAT_GATHER_SCATTER |
8241 | && alignment_support_scheme != dr_aligned) |
8242 | dump_printf_loc (MSG_NOTE, vect_location, |
8243 | "Vectorizing an unaligned access.\n" ); |
8244 | |
8245 | STMT_VINFO_TYPE (stmt_info) = store_vec_info_type; |
8246 | |
8247 | /* As function vect_transform_stmt shows, for interleaving stores |
8248 | the whole chain is vectorized when the last store in the chain |
8249 | is reached, the other stores in the group are skipped. So we |
8250 | want to only cost the last one here, but it's not trivial to |
8251 | get the last, as it's equivalent to use the first one for |
8252 | costing, use the first one instead. */ |
8253 | if (grouped_store |
8254 | && !slp |
8255 | && first_stmt_info != stmt_info) |
8256 | return true; |
8257 | } |
8258 | gcc_assert (memory_access_type == STMT_VINFO_MEMORY_ACCESS_TYPE (stmt_info)); |
8259 | |
8260 | /* Transform. */ |
8261 | |
8262 | ensure_base_align (dr_info); |
8263 | |
8264 | if (STMT_VINFO_SIMD_LANE_ACCESS_P (stmt_info) >= 3) |
8265 | { |
8266 | gcc_assert (memory_access_type == VMAT_CONTIGUOUS); |
8267 | gcc_assert (!slp); |
8268 | if (costing_p) |
8269 | { |
8270 | unsigned int inside_cost = 0, prologue_cost = 0; |
8271 | if (vls_type == VLS_STORE_INVARIANT) |
8272 | prologue_cost += record_stmt_cost (body_cost_vec: cost_vec, count: 1, kind: scalar_to_vec, |
8273 | stmt_info, misalign: 0, where: vect_prologue); |
8274 | vect_get_store_cost (vinfo, stmt_info, ncopies, |
8275 | alignment_support_scheme, misalignment, |
8276 | inside_cost: &inside_cost, body_cost_vec: cost_vec); |
8277 | |
8278 | if (dump_enabled_p ()) |
8279 | dump_printf_loc (MSG_NOTE, vect_location, |
8280 | "vect_model_store_cost: inside_cost = %d, " |
8281 | "prologue_cost = %d .\n" , |
8282 | inside_cost, prologue_cost); |
8283 | |
8284 | return true; |
8285 | } |
8286 | return vectorizable_scan_store (vinfo, stmt_info, gsi, vec_stmt, ncopies); |
8287 | } |
8288 | |
8289 | if (grouped_store) |
8290 | { |
8291 | /* FORNOW */ |
8292 | gcc_assert (!loop || !nested_in_vect_loop_p (loop, stmt_info)); |
8293 | |
8294 | if (slp) |
8295 | { |
8296 | grouped_store = false; |
8297 | /* VEC_NUM is the number of vect stmts to be created for this |
8298 | group. */ |
8299 | vec_num = SLP_TREE_NUMBER_OF_VEC_STMTS (slp_node); |
8300 | first_stmt_info = SLP_TREE_SCALAR_STMTS (slp_node)[0]; |
8301 | gcc_assert (DR_GROUP_FIRST_ELEMENT (first_stmt_info) |
8302 | == first_stmt_info); |
8303 | first_dr_info = STMT_VINFO_DR_INFO (first_stmt_info); |
8304 | op = vect_get_store_rhs (stmt_info: first_stmt_info); |
8305 | } |
8306 | else |
8307 | /* VEC_NUM is the number of vect stmts to be created for this |
8308 | group. */ |
8309 | vec_num = group_size; |
8310 | |
8311 | ref_type = get_group_alias_ptr_type (first_stmt_info); |
8312 | } |
8313 | else |
8314 | ref_type = reference_alias_ptr_type (DR_REF (first_dr_info->dr)); |
8315 | |
8316 | if (!costing_p && dump_enabled_p ()) |
8317 | dump_printf_loc (MSG_NOTE, vect_location, "transform store. ncopies = %d\n" , |
8318 | ncopies); |
8319 | |
8320 | /* Check if we need to update prologue cost for invariant, |
8321 | and update it accordingly if so. If it's not for |
8322 | interleaving store, we can just check vls_type; but if |
8323 | it's for interleaving store, need to check the def_type |
8324 | of the stored value since the current vls_type is just |
8325 | for first_stmt_info. */ |
8326 | auto update_prologue_cost = [&](unsigned *prologue_cost, tree store_rhs) |
8327 | { |
8328 | gcc_assert (costing_p); |
8329 | if (slp) |
8330 | return; |
8331 | if (grouped_store) |
8332 | { |
8333 | gcc_assert (store_rhs); |
8334 | enum vect_def_type cdt; |
8335 | gcc_assert (vect_is_simple_use (store_rhs, vinfo, &cdt)); |
8336 | if (cdt != vect_constant_def && cdt != vect_external_def) |
8337 | return; |
8338 | } |
8339 | else if (vls_type != VLS_STORE_INVARIANT) |
8340 | return; |
8341 | *prologue_cost += record_stmt_cost (body_cost_vec: cost_vec, count: 1, kind: scalar_to_vec, stmt_info, |
8342 | misalign: 0, where: vect_prologue); |
8343 | }; |
8344 | |
8345 | if (memory_access_type == VMAT_ELEMENTWISE |
8346 | || memory_access_type == VMAT_STRIDED_SLP) |
8347 | { |
8348 | unsigned inside_cost = 0, prologue_cost = 0; |
8349 | gimple_stmt_iterator incr_gsi; |
8350 | bool insert_after; |
8351 | gimple *incr; |
8352 | tree offvar; |
8353 | tree ivstep; |
8354 | tree running_off; |
8355 | tree stride_base, stride_step, alias_off; |
8356 | tree vec_oprnd = NULL_TREE; |
8357 | tree dr_offset; |
8358 | unsigned int g; |
8359 | /* Checked by get_load_store_type. */ |
8360 | unsigned int const_nunits = nunits.to_constant (); |
8361 | |
8362 | gcc_assert (!LOOP_VINFO_FULLY_MASKED_P (loop_vinfo)); |
8363 | gcc_assert (!nested_in_vect_loop_p (loop, stmt_info)); |
8364 | |
8365 | dr_offset = get_dr_vinfo_offset (vinfo, dr_info: first_dr_info); |
8366 | stride_base |
8367 | = fold_build_pointer_plus |
8368 | (DR_BASE_ADDRESS (first_dr_info->dr), |
8369 | size_binop (PLUS_EXPR, |
8370 | convert_to_ptrofftype (dr_offset), |
8371 | convert_to_ptrofftype (DR_INIT (first_dr_info->dr)))); |
8372 | stride_step = fold_convert (sizetype, DR_STEP (first_dr_info->dr)); |
8373 | |
8374 | /* For a store with loop-invariant (but other than power-of-2) |
8375 | stride (i.e. not a grouped access) like so: |
8376 | |
8377 | for (i = 0; i < n; i += stride) |
8378 | array[i] = ...; |
8379 | |
8380 | we generate a new induction variable and new stores from |
8381 | the components of the (vectorized) rhs: |
8382 | |
8383 | for (j = 0; ; j += VF*stride) |
8384 | vectemp = ...; |
8385 | tmp1 = vectemp[0]; |
8386 | array[j] = tmp1; |
8387 | tmp2 = vectemp[1]; |
8388 | array[j + stride] = tmp2; |
8389 | ... |
8390 | */ |
8391 | |
8392 | unsigned nstores = const_nunits; |
8393 | unsigned lnel = 1; |
8394 | tree ltype = elem_type; |
8395 | tree lvectype = vectype; |
8396 | if (slp) |
8397 | { |
8398 | if (group_size < const_nunits |
8399 | && const_nunits % group_size == 0) |
8400 | { |
8401 | nstores = const_nunits / group_size; |
8402 | lnel = group_size; |
8403 | ltype = build_vector_type (elem_type, group_size); |
8404 | lvectype = vectype; |
8405 | |
8406 | /* First check if vec_extract optab doesn't support extraction |
8407 | of vector elts directly. */ |
8408 | scalar_mode elmode = SCALAR_TYPE_MODE (elem_type); |
8409 | machine_mode vmode; |
8410 | if (!VECTOR_MODE_P (TYPE_MODE (vectype)) |
8411 | || !related_vector_mode (TYPE_MODE (vectype), elmode, |
8412 | group_size).exists (mode: &vmode) |
8413 | || (convert_optab_handler (op: vec_extract_optab, |
8414 | TYPE_MODE (vectype), from_mode: vmode) |
8415 | == CODE_FOR_nothing)) |
8416 | { |
8417 | /* Try to avoid emitting an extract of vector elements |
8418 | by performing the extracts using an integer type of the |
8419 | same size, extracting from a vector of those and then |
8420 | re-interpreting it as the original vector type if |
8421 | supported. */ |
8422 | unsigned lsize |
8423 | = group_size * GET_MODE_BITSIZE (mode: elmode); |
8424 | unsigned int lnunits = const_nunits / group_size; |
8425 | /* If we can't construct such a vector fall back to |
8426 | element extracts from the original vector type and |
8427 | element size stores. */ |
8428 | if (int_mode_for_size (size: lsize, limit: 0).exists (mode: &elmode) |
8429 | && VECTOR_MODE_P (TYPE_MODE (vectype)) |
8430 | && related_vector_mode (TYPE_MODE (vectype), elmode, |
8431 | lnunits).exists (mode: &vmode) |
8432 | && (convert_optab_handler (op: vec_extract_optab, |
8433 | to_mode: vmode, from_mode: elmode) |
8434 | != CODE_FOR_nothing)) |
8435 | { |
8436 | nstores = lnunits; |
8437 | lnel = group_size; |
8438 | ltype = build_nonstandard_integer_type (lsize, 1); |
8439 | lvectype = build_vector_type (ltype, nstores); |
8440 | } |
8441 | /* Else fall back to vector extraction anyway. |
8442 | Fewer stores are more important than avoiding spilling |
8443 | of the vector we extract from. Compared to the |
8444 | construction case in vectorizable_load no store-forwarding |
8445 | issue exists here for reasonable archs. */ |
8446 | } |
8447 | } |
8448 | else if (group_size >= const_nunits |
8449 | && group_size % const_nunits == 0) |
8450 | { |
8451 | int mis_align = dr_misalignment (dr_info: first_dr_info, vectype); |
8452 | dr_alignment_support dr_align |
8453 | = vect_supportable_dr_alignment (vinfo, dr_info, vectype, |
8454 | mis_align); |
8455 | if (dr_align == dr_aligned |
8456 | || dr_align == dr_unaligned_supported) |
8457 | { |
8458 | nstores = 1; |
8459 | lnel = const_nunits; |
8460 | ltype = vectype; |
8461 | lvectype = vectype; |
8462 | alignment_support_scheme = dr_align; |
8463 | misalignment = mis_align; |
8464 | } |
8465 | } |
8466 | ltype = build_aligned_type (ltype, TYPE_ALIGN (elem_type)); |
8467 | ncopies = SLP_TREE_NUMBER_OF_VEC_STMTS (slp_node); |
8468 | } |
8469 | |
8470 | if (!costing_p) |
8471 | { |
8472 | ivstep = stride_step; |
8473 | ivstep = fold_build2 (MULT_EXPR, TREE_TYPE (ivstep), ivstep, |
8474 | build_int_cst (TREE_TYPE (ivstep), vf)); |
8475 | |
8476 | standard_iv_increment_position (loop, &incr_gsi, &insert_after); |
8477 | |
8478 | stride_base = cse_and_gimplify_to_preheader (loop_vinfo, stride_base); |
8479 | ivstep = cse_and_gimplify_to_preheader (loop_vinfo, ivstep); |
8480 | create_iv (stride_base, PLUS_EXPR, ivstep, NULL, loop, &incr_gsi, |
8481 | insert_after, &offvar, NULL); |
8482 | incr = gsi_stmt (i: incr_gsi); |
8483 | |
8484 | stride_step = cse_and_gimplify_to_preheader (loop_vinfo, stride_step); |
8485 | } |
8486 | |
8487 | alias_off = build_int_cst (ref_type, 0); |
8488 | stmt_vec_info next_stmt_info = first_stmt_info; |
8489 | auto_vec<tree> vec_oprnds (ncopies); |
8490 | /* For costing some adjacent vector stores, we'd like to cost with |
8491 | the total number of them once instead of cost each one by one. */ |
8492 | unsigned int n_adjacent_stores = 0; |
8493 | for (g = 0; g < group_size; g++) |
8494 | { |
8495 | running_off = offvar; |
8496 | if (!costing_p) |
8497 | { |
8498 | if (g) |
8499 | { |
8500 | tree size = TYPE_SIZE_UNIT (ltype); |
8501 | tree pos |
8502 | = fold_build2 (MULT_EXPR, sizetype, size_int (g), size); |
8503 | tree newoff = copy_ssa_name (var: running_off, NULL); |
8504 | incr = gimple_build_assign (newoff, POINTER_PLUS_EXPR, |
8505 | running_off, pos); |
8506 | vect_finish_stmt_generation (vinfo, stmt_info, vec_stmt: incr, gsi); |
8507 | running_off = newoff; |
8508 | } |
8509 | } |
8510 | if (!slp) |
8511 | op = vect_get_store_rhs (stmt_info: next_stmt_info); |
8512 | if (!costing_p) |
8513 | vect_get_vec_defs (vinfo, stmt_info: next_stmt_info, slp_node, ncopies, op0: op, |
8514 | vec_oprnds0: &vec_oprnds); |
8515 | else |
8516 | update_prologue_cost (&prologue_cost, op); |
8517 | unsigned int group_el = 0; |
8518 | unsigned HOST_WIDE_INT |
8519 | elsz = tree_to_uhwi (TYPE_SIZE_UNIT (TREE_TYPE (vectype))); |
8520 | for (j = 0; j < ncopies; j++) |
8521 | { |
8522 | if (!costing_p) |
8523 | { |
8524 | vec_oprnd = vec_oprnds[j]; |
8525 | /* Pun the vector to extract from if necessary. */ |
8526 | if (lvectype != vectype) |
8527 | { |
8528 | tree tem = make_ssa_name (var: lvectype); |
8529 | tree cvt |
8530 | = build1 (VIEW_CONVERT_EXPR, lvectype, vec_oprnd); |
8531 | gimple *pun = gimple_build_assign (tem, cvt); |
8532 | vect_finish_stmt_generation (vinfo, stmt_info, vec_stmt: pun, gsi); |
8533 | vec_oprnd = tem; |
8534 | } |
8535 | } |
8536 | for (i = 0; i < nstores; i++) |
8537 | { |
8538 | if (costing_p) |
8539 | { |
8540 | /* Only need vector extracting when there are more |
8541 | than one stores. */ |
8542 | if (nstores > 1) |
8543 | inside_cost |
8544 | += record_stmt_cost (body_cost_vec: cost_vec, count: 1, kind: vec_to_scalar, |
8545 | stmt_info, misalign: 0, where: vect_body); |
8546 | /* Take a single lane vector type store as scalar |
8547 | store to avoid ICE like 110776. */ |
8548 | if (VECTOR_TYPE_P (ltype) |
8549 | && known_ne (TYPE_VECTOR_SUBPARTS (ltype), 1U)) |
8550 | n_adjacent_stores++; |
8551 | else |
8552 | inside_cost |
8553 | += record_stmt_cost (body_cost_vec: cost_vec, count: 1, kind: scalar_store, |
8554 | stmt_info, misalign: 0, where: vect_body); |
8555 | continue; |
8556 | } |
8557 | tree newref, newoff; |
8558 | gimple *incr, *assign; |
8559 | tree size = TYPE_SIZE (ltype); |
8560 | /* Extract the i'th component. */ |
8561 | tree pos = fold_build2 (MULT_EXPR, bitsizetype, |
8562 | bitsize_int (i), size); |
8563 | tree elem = fold_build3 (BIT_FIELD_REF, ltype, vec_oprnd, |
8564 | size, pos); |
8565 | |
8566 | elem = force_gimple_operand_gsi (gsi, elem, true, |
8567 | NULL_TREE, true, |
8568 | GSI_SAME_STMT); |
8569 | |
8570 | tree this_off = build_int_cst (TREE_TYPE (alias_off), |
8571 | group_el * elsz); |
8572 | newref = build2 (MEM_REF, ltype, |
8573 | running_off, this_off); |
8574 | vect_copy_ref_info (newref, DR_REF (first_dr_info->dr)); |
8575 | |
8576 | /* And store it to *running_off. */ |
8577 | assign = gimple_build_assign (newref, elem); |
8578 | vect_finish_stmt_generation (vinfo, stmt_info, vec_stmt: assign, gsi); |
8579 | |
8580 | group_el += lnel; |
8581 | if (! slp |
8582 | || group_el == group_size) |
8583 | { |
8584 | newoff = copy_ssa_name (var: running_off, NULL); |
8585 | incr = gimple_build_assign (newoff, POINTER_PLUS_EXPR, |
8586 | running_off, stride_step); |
8587 | vect_finish_stmt_generation (vinfo, stmt_info, vec_stmt: incr, gsi); |
8588 | |
8589 | running_off = newoff; |
8590 | group_el = 0; |
8591 | } |
8592 | if (g == group_size - 1 |
8593 | && !slp) |
8594 | { |
8595 | if (j == 0 && i == 0) |
8596 | *vec_stmt = assign; |
8597 | STMT_VINFO_VEC_STMTS (stmt_info).safe_push (obj: assign); |
8598 | } |
8599 | } |
8600 | } |
8601 | next_stmt_info = DR_GROUP_NEXT_ELEMENT (next_stmt_info); |
8602 | vec_oprnds.truncate(size: 0); |
8603 | if (slp) |
8604 | break; |
8605 | } |
8606 | |
8607 | if (costing_p) |
8608 | { |
8609 | if (n_adjacent_stores > 0) |
8610 | vect_get_store_cost (vinfo, stmt_info, ncopies: n_adjacent_stores, |
8611 | alignment_support_scheme, misalignment, |
8612 | inside_cost: &inside_cost, body_cost_vec: cost_vec); |
8613 | if (dump_enabled_p ()) |
8614 | dump_printf_loc (MSG_NOTE, vect_location, |
8615 | "vect_model_store_cost: inside_cost = %d, " |
8616 | "prologue_cost = %d .\n" , |
8617 | inside_cost, prologue_cost); |
8618 | } |
8619 | |
8620 | return true; |
8621 | } |
8622 | |
8623 | gcc_assert (alignment_support_scheme); |
8624 | vec_loop_masks *loop_masks |
8625 | = (loop_vinfo && LOOP_VINFO_FULLY_MASKED_P (loop_vinfo) |
8626 | ? &LOOP_VINFO_MASKS (loop_vinfo) |
8627 | : NULL); |
8628 | vec_loop_lens *loop_lens |
8629 | = (loop_vinfo && LOOP_VINFO_FULLY_WITH_LENGTH_P (loop_vinfo) |
8630 | ? &LOOP_VINFO_LENS (loop_vinfo) |
8631 | : NULL); |
8632 | |
8633 | /* Shouldn't go with length-based approach if fully masked. */ |
8634 | gcc_assert (!loop_lens || !loop_masks); |
8635 | |
8636 | /* Targets with store-lane instructions must not require explicit |
8637 | realignment. vect_supportable_dr_alignment always returns either |
8638 | dr_aligned or dr_unaligned_supported for masked operations. */ |
8639 | gcc_assert ((memory_access_type != VMAT_LOAD_STORE_LANES |
8640 | && !mask |
8641 | && !loop_masks) |
8642 | || alignment_support_scheme == dr_aligned |
8643 | || alignment_support_scheme == dr_unaligned_supported); |
8644 | |
8645 | tree offset = NULL_TREE; |
8646 | if (!known_eq (poffset, 0)) |
8647 | offset = size_int (poffset); |
8648 | |
8649 | tree bump; |
8650 | tree vec_offset = NULL_TREE; |
8651 | if (STMT_VINFO_GATHER_SCATTER_P (stmt_info)) |
8652 | { |
8653 | aggr_type = NULL_TREE; |
8654 | bump = NULL_TREE; |
8655 | } |
8656 | else if (memory_access_type == VMAT_GATHER_SCATTER) |
8657 | { |
8658 | aggr_type = elem_type; |
8659 | if (!costing_p) |
8660 | vect_get_strided_load_store_ops (stmt_info, loop_vinfo, gsi, gs_info: &gs_info, |
8661 | dataref_bump: &bump, vec_offset: &vec_offset, loop_lens); |
8662 | } |
8663 | else |
8664 | { |
8665 | if (memory_access_type == VMAT_LOAD_STORE_LANES) |
8666 | aggr_type = build_array_type_nelts (elem_type, vec_num * nunits); |
8667 | else |
8668 | aggr_type = vectype; |
8669 | bump = vect_get_data_ptr_increment (vinfo, gsi, dr_info, aggr_type, |
8670 | memory_access_type, loop_lens); |
8671 | } |
8672 | |
8673 | if (mask && !costing_p) |
8674 | LOOP_VINFO_HAS_MASK_STORE (loop_vinfo) = true; |
8675 | |
8676 | /* In case the vectorization factor (VF) is bigger than the number |
8677 | of elements that we can fit in a vectype (nunits), we have to generate |
8678 | more than one vector stmt - i.e - we need to "unroll" the |
8679 | vector stmt by a factor VF/nunits. */ |
8680 | |
8681 | /* In case of interleaving (non-unit grouped access): |
8682 | |
8683 | S1: &base + 2 = x2 |
8684 | S2: &base = x0 |
8685 | S3: &base + 1 = x1 |
8686 | S4: &base + 3 = x3 |
8687 | |
8688 | We create vectorized stores starting from base address (the access of the |
8689 | first stmt in the chain (S2 in the above example), when the last store stmt |
8690 | of the chain (S4) is reached: |
8691 | |
8692 | VS1: &base = vx2 |
8693 | VS2: &base + vec_size*1 = vx0 |
8694 | VS3: &base + vec_size*2 = vx1 |
8695 | VS4: &base + vec_size*3 = vx3 |
8696 | |
8697 | Then permutation statements are generated: |
8698 | |
8699 | VS5: vx5 = VEC_PERM_EXPR < vx0, vx3, {0, 8, 1, 9, 2, 10, 3, 11} > |
8700 | VS6: vx6 = VEC_PERM_EXPR < vx0, vx3, {4, 12, 5, 13, 6, 14, 7, 15} > |
8701 | ... |
8702 | |
8703 | And they are put in STMT_VINFO_VEC_STMT of the corresponding scalar stmts |
8704 | (the order of the data-refs in the output of vect_permute_store_chain |
8705 | corresponds to the order of scalar stmts in the interleaving chain - see |
8706 | the documentation of vect_permute_store_chain()). |
8707 | |
8708 | In case of both multiple types and interleaving, above vector stores and |
8709 | permutation stmts are created for every copy. The result vector stmts are |
8710 | put in STMT_VINFO_VEC_STMT for the first copy and in the corresponding |
8711 | STMT_VINFO_RELATED_STMT for the next copies. |
8712 | */ |
8713 | |
8714 | auto_vec<tree> dr_chain (group_size); |
8715 | auto_vec<tree> vec_masks; |
8716 | tree vec_mask = NULL; |
8717 | auto_delete_vec<auto_vec<tree>> gvec_oprnds (group_size); |
8718 | for (i = 0; i < group_size; i++) |
8719 | gvec_oprnds.quick_push (obj: new auto_vec<tree> (ncopies)); |
8720 | |
8721 | if (memory_access_type == VMAT_LOAD_STORE_LANES) |
8722 | { |
8723 | gcc_assert (!slp && grouped_store); |
8724 | unsigned inside_cost = 0, prologue_cost = 0; |
8725 | /* For costing some adjacent vector stores, we'd like to cost with |
8726 | the total number of them once instead of cost each one by one. */ |
8727 | unsigned int n_adjacent_stores = 0; |
8728 | for (j = 0; j < ncopies; j++) |
8729 | { |
8730 | gimple *new_stmt; |
8731 | if (j == 0) |
8732 | { |
8733 | /* For interleaved stores we collect vectorized defs for all |
8734 | the stores in the group in DR_CHAIN. DR_CHAIN is then used |
8735 | as an input to vect_permute_store_chain(). */ |
8736 | stmt_vec_info next_stmt_info = first_stmt_info; |
8737 | for (i = 0; i < group_size; i++) |
8738 | { |
8739 | /* Since gaps are not supported for interleaved stores, |
8740 | DR_GROUP_SIZE is the exact number of stmts in the |
8741 | chain. Therefore, NEXT_STMT_INFO can't be NULL_TREE. */ |
8742 | op = vect_get_store_rhs (stmt_info: next_stmt_info); |
8743 | if (costing_p) |
8744 | update_prologue_cost (&prologue_cost, op); |
8745 | else |
8746 | { |
8747 | vect_get_vec_defs_for_operand (vinfo, stmt_vinfo: next_stmt_info, |
8748 | ncopies, op, |
8749 | vec_oprnds: gvec_oprnds[i]); |
8750 | vec_oprnd = (*gvec_oprnds[i])[0]; |
8751 | dr_chain.quick_push (obj: vec_oprnd); |
8752 | } |
8753 | next_stmt_info = DR_GROUP_NEXT_ELEMENT (next_stmt_info); |
8754 | } |
8755 | |
8756 | if (!costing_p) |
8757 | { |
8758 | if (mask) |
8759 | { |
8760 | vect_get_vec_defs_for_operand (vinfo, stmt_vinfo: stmt_info, ncopies, |
8761 | op: mask, vec_oprnds: &vec_masks, |
8762 | vectype: mask_vectype); |
8763 | vec_mask = vec_masks[0]; |
8764 | } |
8765 | |
8766 | /* We should have catched mismatched types earlier. */ |
8767 | gcc_assert ( |
8768 | useless_type_conversion_p (vectype, TREE_TYPE (vec_oprnd))); |
8769 | dataref_ptr |
8770 | = vect_create_data_ref_ptr (vinfo, first_stmt_info, |
8771 | aggr_type, NULL, offset, &dummy, |
8772 | gsi, &ptr_incr, false, bump); |
8773 | } |
8774 | } |
8775 | else if (!costing_p) |
8776 | { |
8777 | gcc_assert (!LOOP_VINFO_USING_SELECT_VL_P (loop_vinfo)); |
8778 | /* DR_CHAIN is then used as an input to |
8779 | vect_permute_store_chain(). */ |
8780 | for (i = 0; i < group_size; i++) |
8781 | { |
8782 | vec_oprnd = (*gvec_oprnds[i])[j]; |
8783 | dr_chain[i] = vec_oprnd; |
8784 | } |
8785 | if (mask) |
8786 | vec_mask = vec_masks[j]; |
8787 | dataref_ptr = bump_vector_ptr (vinfo, dataref_ptr, ptr_incr, gsi, |
8788 | stmt_info, bump); |
8789 | } |
8790 | |
8791 | if (costing_p) |
8792 | { |
8793 | n_adjacent_stores += vec_num; |
8794 | continue; |
8795 | } |
8796 | |
8797 | /* Get an array into which we can store the individual vectors. */ |
8798 | tree vec_array = create_vector_array (elem_type: vectype, nelems: vec_num); |
8799 | |
8800 | /* Invalidate the current contents of VEC_ARRAY. This should |
8801 | become an RTL clobber too, which prevents the vector registers |
8802 | from being upward-exposed. */ |
8803 | vect_clobber_variable (vinfo, stmt_info, gsi, var: vec_array); |
8804 | |
8805 | /* Store the individual vectors into the array. */ |
8806 | for (i = 0; i < vec_num; i++) |
8807 | { |
8808 | vec_oprnd = dr_chain[i]; |
8809 | write_vector_array (vinfo, stmt_info, gsi, vect: vec_oprnd, array: vec_array, |
8810 | n: i); |
8811 | } |
8812 | |
8813 | tree final_mask = NULL; |
8814 | tree final_len = NULL; |
8815 | tree bias = NULL; |
8816 | if (loop_masks) |
8817 | final_mask = vect_get_loop_mask (loop_vinfo, gsi, loop_masks, |
8818 | ncopies, vectype, j); |
8819 | if (vec_mask) |
8820 | final_mask = prepare_vec_mask (loop_vinfo, mask_type: mask_vectype, loop_mask: final_mask, |
8821 | vec_mask, gsi); |
8822 | |
8823 | if (lanes_ifn == IFN_MASK_LEN_STORE_LANES) |
8824 | { |
8825 | if (loop_lens) |
8826 | final_len = vect_get_loop_len (loop_vinfo, gsi, loop_lens, |
8827 | ncopies, vectype, j, 1); |
8828 | else |
8829 | final_len = size_int (TYPE_VECTOR_SUBPARTS (vectype)); |
8830 | signed char biasval |
8831 | = LOOP_VINFO_PARTIAL_LOAD_STORE_BIAS (loop_vinfo); |
8832 | bias = build_int_cst (intQI_type_node, biasval); |
8833 | if (!final_mask) |
8834 | { |
8835 | mask_vectype = truth_type_for (vectype); |
8836 | final_mask = build_minus_one_cst (mask_vectype); |
8837 | } |
8838 | } |
8839 | |
8840 | gcall *call; |
8841 | if (final_len && final_mask) |
8842 | { |
8843 | /* Emit: |
8844 | MASK_LEN_STORE_LANES (DATAREF_PTR, ALIAS_PTR, VEC_MASK, |
8845 | LEN, BIAS, VEC_ARRAY). */ |
8846 | unsigned int align = TYPE_ALIGN (TREE_TYPE (vectype)); |
8847 | tree alias_ptr = build_int_cst (ref_type, align); |
8848 | call = gimple_build_call_internal (IFN_MASK_LEN_STORE_LANES, 6, |
8849 | dataref_ptr, alias_ptr, |
8850 | final_mask, final_len, bias, |
8851 | vec_array); |
8852 | } |
8853 | else if (final_mask) |
8854 | { |
8855 | /* Emit: |
8856 | MASK_STORE_LANES (DATAREF_PTR, ALIAS_PTR, VEC_MASK, |
8857 | VEC_ARRAY). */ |
8858 | unsigned int align = TYPE_ALIGN (TREE_TYPE (vectype)); |
8859 | tree alias_ptr = build_int_cst (ref_type, align); |
8860 | call = gimple_build_call_internal (IFN_MASK_STORE_LANES, 4, |
8861 | dataref_ptr, alias_ptr, |
8862 | final_mask, vec_array); |
8863 | } |
8864 | else |
8865 | { |
8866 | /* Emit: |
8867 | MEM_REF[...all elements...] = STORE_LANES (VEC_ARRAY). */ |
8868 | data_ref = create_array_ref (type: aggr_type, ptr: dataref_ptr, alias_ptr_type: ref_type); |
8869 | call = gimple_build_call_internal (IFN_STORE_LANES, 1, vec_array); |
8870 | gimple_call_set_lhs (gs: call, lhs: data_ref); |
8871 | } |
8872 | gimple_call_set_nothrow (s: call, nothrow_p: true); |
8873 | vect_finish_stmt_generation (vinfo, stmt_info, vec_stmt: call, gsi); |
8874 | new_stmt = call; |
8875 | |
8876 | /* Record that VEC_ARRAY is now dead. */ |
8877 | vect_clobber_variable (vinfo, stmt_info, gsi, var: vec_array); |
8878 | if (j == 0) |
8879 | *vec_stmt = new_stmt; |
8880 | STMT_VINFO_VEC_STMTS (stmt_info).safe_push (obj: new_stmt); |
8881 | } |
8882 | |
8883 | if (costing_p) |
8884 | { |
8885 | if (n_adjacent_stores > 0) |
8886 | vect_get_store_cost (vinfo, stmt_info, ncopies: n_adjacent_stores, |
8887 | alignment_support_scheme, misalignment, |
8888 | inside_cost: &inside_cost, body_cost_vec: cost_vec); |
8889 | if (dump_enabled_p ()) |
8890 | dump_printf_loc (MSG_NOTE, vect_location, |
8891 | "vect_model_store_cost: inside_cost = %d, " |
8892 | "prologue_cost = %d .\n" , |
8893 | inside_cost, prologue_cost); |
8894 | } |
8895 | |
8896 | return true; |
8897 | } |
8898 | |
8899 | if (memory_access_type == VMAT_GATHER_SCATTER) |
8900 | { |
8901 | gcc_assert (!grouped_store); |
8902 | auto_vec<tree> vec_offsets; |
8903 | unsigned int inside_cost = 0, prologue_cost = 0; |
8904 | for (j = 0; j < ncopies; j++) |
8905 | { |
8906 | gimple *new_stmt; |
8907 | if (j == 0) |
8908 | { |
8909 | if (costing_p && vls_type == VLS_STORE_INVARIANT) |
8910 | prologue_cost += record_stmt_cost (body_cost_vec: cost_vec, count: 1, kind: scalar_to_vec, |
8911 | stmt_info, misalign: 0, where: vect_prologue); |
8912 | else if (!costing_p) |
8913 | { |
8914 | /* Since the store is not grouped, DR_GROUP_SIZE is 1, and |
8915 | DR_CHAIN is of size 1. */ |
8916 | gcc_assert (group_size == 1); |
8917 | if (slp_node) |
8918 | vect_get_slp_defs (op_node, gvec_oprnds[0]); |
8919 | else |
8920 | vect_get_vec_defs_for_operand (vinfo, stmt_vinfo: first_stmt_info, |
8921 | ncopies, op, vec_oprnds: gvec_oprnds[0]); |
8922 | if (mask) |
8923 | { |
8924 | if (slp_node) |
8925 | vect_get_slp_defs (mask_node, &vec_masks); |
8926 | else |
8927 | vect_get_vec_defs_for_operand (vinfo, stmt_vinfo: stmt_info, |
8928 | ncopies, |
8929 | op: mask, vec_oprnds: &vec_masks, |
8930 | vectype: mask_vectype); |
8931 | } |
8932 | |
8933 | if (STMT_VINFO_GATHER_SCATTER_P (stmt_info)) |
8934 | vect_get_gather_scatter_ops (loop_vinfo, loop, stmt_info, |
8935 | slp_node, gs_info: &gs_info, |
8936 | dataref_ptr: &dataref_ptr, vec_offset: &vec_offsets); |
8937 | else |
8938 | dataref_ptr |
8939 | = vect_create_data_ref_ptr (vinfo, first_stmt_info, |
8940 | aggr_type, NULL, offset, |
8941 | &dummy, gsi, &ptr_incr, false, |
8942 | bump); |
8943 | } |
8944 | } |
8945 | else if (!costing_p) |
8946 | { |
8947 | gcc_assert (!LOOP_VINFO_USING_SELECT_VL_P (loop_vinfo)); |
8948 | if (!STMT_VINFO_GATHER_SCATTER_P (stmt_info)) |
8949 | dataref_ptr = bump_vector_ptr (vinfo, dataref_ptr, ptr_incr, |
8950 | gsi, stmt_info, bump); |
8951 | } |
8952 | |
8953 | new_stmt = NULL; |
8954 | for (i = 0; i < vec_num; ++i) |
8955 | { |
8956 | if (!costing_p) |
8957 | { |
8958 | vec_oprnd = (*gvec_oprnds[0])[vec_num * j + i]; |
8959 | if (mask) |
8960 | vec_mask = vec_masks[vec_num * j + i]; |
8961 | /* We should have catched mismatched types earlier. */ |
8962 | gcc_assert (useless_type_conversion_p (vectype, |
8963 | TREE_TYPE (vec_oprnd))); |
8964 | } |
8965 | unsigned HOST_WIDE_INT align; |
8966 | tree final_mask = NULL_TREE; |
8967 | tree final_len = NULL_TREE; |
8968 | tree bias = NULL_TREE; |
8969 | if (!costing_p) |
8970 | { |
8971 | if (loop_masks) |
8972 | final_mask = vect_get_loop_mask (loop_vinfo, gsi, |
8973 | loop_masks, ncopies, |
8974 | vectype, j); |
8975 | if (vec_mask) |
8976 | final_mask = prepare_vec_mask (loop_vinfo, mask_type: mask_vectype, |
8977 | loop_mask: final_mask, vec_mask, gsi); |
8978 | } |
8979 | |
8980 | if (gs_info.ifn != IFN_LAST) |
8981 | { |
8982 | if (costing_p) |
8983 | { |
8984 | unsigned int cnunits = vect_nunits_for_cost (vec_type: vectype); |
8985 | inside_cost |
8986 | += record_stmt_cost (body_cost_vec: cost_vec, count: cnunits, kind: scalar_store, |
8987 | stmt_info, misalign: 0, where: vect_body); |
8988 | continue; |
8989 | } |
8990 | |
8991 | if (STMT_VINFO_GATHER_SCATTER_P (stmt_info)) |
8992 | vec_offset = vec_offsets[vec_num * j + i]; |
8993 | tree scale = size_int (gs_info.scale); |
8994 | |
8995 | if (gs_info.ifn == IFN_MASK_LEN_SCATTER_STORE) |
8996 | { |
8997 | if (loop_lens) |
8998 | final_len = vect_get_loop_len (loop_vinfo, gsi, |
8999 | loop_lens, ncopies, |
9000 | vectype, j, 1); |
9001 | else |
9002 | final_len = size_int (TYPE_VECTOR_SUBPARTS (vectype)); |
9003 | signed char biasval |
9004 | = LOOP_VINFO_PARTIAL_LOAD_STORE_BIAS (loop_vinfo); |
9005 | bias = build_int_cst (intQI_type_node, biasval); |
9006 | if (!final_mask) |
9007 | { |
9008 | mask_vectype = truth_type_for (vectype); |
9009 | final_mask = build_minus_one_cst (mask_vectype); |
9010 | } |
9011 | } |
9012 | |
9013 | gcall *call; |
9014 | if (final_len && final_mask) |
9015 | call = gimple_build_call_internal |
9016 | (IFN_MASK_LEN_SCATTER_STORE, 7, dataref_ptr, |
9017 | vec_offset, scale, vec_oprnd, final_mask, |
9018 | final_len, bias); |
9019 | else if (final_mask) |
9020 | call = gimple_build_call_internal |
9021 | (IFN_MASK_SCATTER_STORE, 5, dataref_ptr, |
9022 | vec_offset, scale, vec_oprnd, final_mask); |
9023 | else |
9024 | call = gimple_build_call_internal (IFN_SCATTER_STORE, 4, |
9025 | dataref_ptr, vec_offset, |
9026 | scale, vec_oprnd); |
9027 | gimple_call_set_nothrow (s: call, nothrow_p: true); |
9028 | vect_finish_stmt_generation (vinfo, stmt_info, vec_stmt: call, gsi); |
9029 | new_stmt = call; |
9030 | } |
9031 | else if (gs_info.decl) |
9032 | { |
9033 | /* The builtin decls path for scatter is legacy, x86 only. */ |
9034 | gcc_assert (nunits.is_constant () |
9035 | && (!final_mask |
9036 | || SCALAR_INT_MODE_P |
9037 | (TYPE_MODE (TREE_TYPE (final_mask))))); |
9038 | if (costing_p) |
9039 | { |
9040 | unsigned int cnunits = vect_nunits_for_cost (vec_type: vectype); |
9041 | inside_cost |
9042 | += record_stmt_cost (body_cost_vec: cost_vec, count: cnunits, kind: scalar_store, |
9043 | stmt_info, misalign: 0, where: vect_body); |
9044 | continue; |
9045 | } |
9046 | poly_uint64 offset_nunits |
9047 | = TYPE_VECTOR_SUBPARTS (node: gs_info.offset_vectype); |
9048 | if (known_eq (nunits, offset_nunits)) |
9049 | { |
9050 | new_stmt = vect_build_one_scatter_store_call |
9051 | (vinfo, stmt_info, gsi, gs_info: &gs_info, |
9052 | ptr: dataref_ptr, offset: vec_offsets[vec_num * j + i], |
9053 | oprnd: vec_oprnd, mask: final_mask); |
9054 | vect_finish_stmt_generation (vinfo, stmt_info, |
9055 | vec_stmt: new_stmt, gsi); |
9056 | } |
9057 | else if (known_eq (nunits, offset_nunits * 2)) |
9058 | { |
9059 | /* We have a offset vector with half the number of |
9060 | lanes but the builtins will store full vectype |
9061 | data from the lower lanes. */ |
9062 | new_stmt = vect_build_one_scatter_store_call |
9063 | (vinfo, stmt_info, gsi, gs_info: &gs_info, |
9064 | ptr: dataref_ptr, |
9065 | offset: vec_offsets[2 * vec_num * j + 2 * i], |
9066 | oprnd: vec_oprnd, mask: final_mask); |
9067 | vect_finish_stmt_generation (vinfo, stmt_info, |
9068 | vec_stmt: new_stmt, gsi); |
9069 | int count = nunits.to_constant (); |
9070 | vec_perm_builder sel (count, count, 1); |
9071 | sel.quick_grow (len: count); |
9072 | for (int i = 0; i < count; ++i) |
9073 | sel[i] = i | (count / 2); |
9074 | vec_perm_indices indices (sel, 2, count); |
9075 | tree perm_mask |
9076 | = vect_gen_perm_mask_checked (vectype, indices); |
9077 | new_stmt = gimple_build_assign (NULL_TREE, VEC_PERM_EXPR, |
9078 | vec_oprnd, vec_oprnd, |
9079 | perm_mask); |
9080 | vec_oprnd = make_ssa_name (var: vectype); |
9081 | gimple_set_lhs (new_stmt, vec_oprnd); |
9082 | vect_finish_stmt_generation (vinfo, stmt_info, |
9083 | vec_stmt: new_stmt, gsi); |
9084 | if (final_mask) |
9085 | { |
9086 | new_stmt = gimple_build_assign (NULL_TREE, |
9087 | VEC_UNPACK_HI_EXPR, |
9088 | final_mask); |
9089 | final_mask = make_ssa_name |
9090 | (var: truth_type_for (gs_info.offset_vectype)); |
9091 | gimple_set_lhs (new_stmt, final_mask); |
9092 | vect_finish_stmt_generation (vinfo, stmt_info, |
9093 | vec_stmt: new_stmt, gsi); |
9094 | } |
9095 | new_stmt = vect_build_one_scatter_store_call |
9096 | (vinfo, stmt_info, gsi, gs_info: &gs_info, |
9097 | ptr: dataref_ptr, |
9098 | offset: vec_offsets[2 * vec_num * j + 2 * i + 1], |
9099 | oprnd: vec_oprnd, mask: final_mask); |
9100 | vect_finish_stmt_generation (vinfo, stmt_info, |
9101 | vec_stmt: new_stmt, gsi); |
9102 | } |
9103 | else if (known_eq (nunits * 2, offset_nunits)) |
9104 | { |
9105 | /* We have a offset vector with double the number of |
9106 | lanes. Select the low/high part accordingly. */ |
9107 | vec_offset = vec_offsets[(vec_num * j + i) / 2]; |
9108 | if ((vec_num * j + i) & 1) |
9109 | { |
9110 | int count = offset_nunits.to_constant (); |
9111 | vec_perm_builder sel (count, count, 1); |
9112 | sel.quick_grow (len: count); |
9113 | for (int i = 0; i < count; ++i) |
9114 | sel[i] = i | (count / 2); |
9115 | vec_perm_indices indices (sel, 2, count); |
9116 | tree perm_mask = vect_gen_perm_mask_checked |
9117 | (TREE_TYPE (vec_offset), indices); |
9118 | new_stmt = gimple_build_assign (NULL_TREE, |
9119 | VEC_PERM_EXPR, |
9120 | vec_offset, |
9121 | vec_offset, |
9122 | perm_mask); |
9123 | vec_offset = make_ssa_name (TREE_TYPE (vec_offset)); |
9124 | gimple_set_lhs (new_stmt, vec_offset); |
9125 | vect_finish_stmt_generation (vinfo, stmt_info, |
9126 | vec_stmt: new_stmt, gsi); |
9127 | } |
9128 | new_stmt = vect_build_one_scatter_store_call |
9129 | (vinfo, stmt_info, gsi, gs_info: &gs_info, |
9130 | ptr: dataref_ptr, offset: vec_offset, |
9131 | oprnd: vec_oprnd, mask: final_mask); |
9132 | vect_finish_stmt_generation (vinfo, stmt_info, |
9133 | vec_stmt: new_stmt, gsi); |
9134 | } |
9135 | else |
9136 | gcc_unreachable (); |
9137 | } |
9138 | else |
9139 | { |
9140 | /* Emulated scatter. */ |
9141 | gcc_assert (!final_mask); |
9142 | if (costing_p) |
9143 | { |
9144 | unsigned int cnunits = vect_nunits_for_cost (vec_type: vectype); |
9145 | /* For emulated scatter N offset vector element extracts |
9146 | (we assume the scalar scaling and ptr + offset add is |
9147 | consumed by the load). */ |
9148 | inside_cost |
9149 | += record_stmt_cost (body_cost_vec: cost_vec, count: cnunits, kind: vec_to_scalar, |
9150 | stmt_info, misalign: 0, where: vect_body); |
9151 | /* N scalar stores plus extracting the elements. */ |
9152 | inside_cost |
9153 | += record_stmt_cost (body_cost_vec: cost_vec, count: cnunits, kind: vec_to_scalar, |
9154 | stmt_info, misalign: 0, where: vect_body); |
9155 | inside_cost |
9156 | += record_stmt_cost (body_cost_vec: cost_vec, count: cnunits, kind: scalar_store, |
9157 | stmt_info, misalign: 0, where: vect_body); |
9158 | continue; |
9159 | } |
9160 | |
9161 | unsigned HOST_WIDE_INT const_nunits = nunits.to_constant (); |
9162 | unsigned HOST_WIDE_INT const_offset_nunits |
9163 | = TYPE_VECTOR_SUBPARTS (node: gs_info.offset_vectype).to_constant (); |
9164 | vec<constructor_elt, va_gc> *ctor_elts; |
9165 | vec_alloc (v&: ctor_elts, nelems: const_nunits); |
9166 | gimple_seq stmts = NULL; |
9167 | tree elt_type = TREE_TYPE (vectype); |
9168 | unsigned HOST_WIDE_INT elt_size |
9169 | = tree_to_uhwi (TYPE_SIZE (elt_type)); |
9170 | /* We support offset vectors with more elements |
9171 | than the data vector for now. */ |
9172 | unsigned HOST_WIDE_INT factor |
9173 | = const_offset_nunits / const_nunits; |
9174 | vec_offset = vec_offsets[(vec_num * j + i) / factor]; |
9175 | unsigned elt_offset = (j % factor) * const_nunits; |
9176 | tree idx_type = TREE_TYPE (TREE_TYPE (vec_offset)); |
9177 | tree scale = size_int (gs_info.scale); |
9178 | align = get_object_alignment (DR_REF (first_dr_info->dr)); |
9179 | tree ltype = build_aligned_type (TREE_TYPE (vectype), align); |
9180 | for (unsigned k = 0; k < const_nunits; ++k) |
9181 | { |
9182 | /* Compute the offsetted pointer. */ |
9183 | tree boff = size_binop (MULT_EXPR, TYPE_SIZE (idx_type), |
9184 | bitsize_int (k + elt_offset)); |
9185 | tree idx |
9186 | = gimple_build (seq: &stmts, code: BIT_FIELD_REF, type: idx_type, |
9187 | ops: vec_offset, TYPE_SIZE (idx_type), ops: boff); |
9188 | idx = gimple_convert (seq: &stmts, sizetype, op: idx); |
9189 | idx = gimple_build (seq: &stmts, code: MULT_EXPR, sizetype, |
9190 | ops: idx, ops: scale); |
9191 | tree ptr |
9192 | = gimple_build (seq: &stmts, code: PLUS_EXPR, |
9193 | TREE_TYPE (dataref_ptr), |
9194 | ops: dataref_ptr, ops: idx); |
9195 | ptr = gimple_convert (seq: &stmts, ptr_type_node, op: ptr); |
9196 | /* Extract the element to be stored. */ |
9197 | tree elt |
9198 | = gimple_build (seq: &stmts, code: BIT_FIELD_REF, |
9199 | TREE_TYPE (vectype), |
9200 | ops: vec_oprnd, TYPE_SIZE (elt_type), |
9201 | bitsize_int (k * elt_size)); |
9202 | gsi_insert_seq_before (gsi, stmts, GSI_SAME_STMT); |
9203 | stmts = NULL; |
9204 | tree ref |
9205 | = build2 (MEM_REF, ltype, ptr, |
9206 | build_int_cst (ref_type, 0)); |
9207 | new_stmt = gimple_build_assign (ref, elt); |
9208 | vect_finish_stmt_generation (vinfo, stmt_info, |
9209 | vec_stmt: new_stmt, gsi); |
9210 | } |
9211 | if (slp) |
9212 | slp_node->push_vec_def (def: new_stmt); |
9213 | } |
9214 | } |
9215 | if (!slp && !costing_p) |
9216 | STMT_VINFO_VEC_STMTS (stmt_info).safe_push (obj: new_stmt); |
9217 | } |
9218 | |
9219 | if (!slp && !costing_p) |
9220 | *vec_stmt = STMT_VINFO_VEC_STMTS (stmt_info)[0]; |
9221 | |
9222 | if (costing_p && dump_enabled_p ()) |
9223 | dump_printf_loc (MSG_NOTE, vect_location, |
9224 | "vect_model_store_cost: inside_cost = %d, " |
9225 | "prologue_cost = %d .\n" , |
9226 | inside_cost, prologue_cost); |
9227 | |
9228 | return true; |
9229 | } |
9230 | |
9231 | gcc_assert (memory_access_type == VMAT_CONTIGUOUS |
9232 | || memory_access_type == VMAT_CONTIGUOUS_DOWN |
9233 | || memory_access_type == VMAT_CONTIGUOUS_PERMUTE |
9234 | || memory_access_type == VMAT_CONTIGUOUS_REVERSE); |
9235 | |
9236 | unsigned inside_cost = 0, prologue_cost = 0; |
9237 | /* For costing some adjacent vector stores, we'd like to cost with |
9238 | the total number of them once instead of cost each one by one. */ |
9239 | unsigned int n_adjacent_stores = 0; |
9240 | auto_vec<tree> result_chain (group_size); |
9241 | auto_vec<tree, 1> vec_oprnds; |
9242 | for (j = 0; j < ncopies; j++) |
9243 | { |
9244 | gimple *new_stmt; |
9245 | if (j == 0) |
9246 | { |
9247 | if (slp && !costing_p) |
9248 | { |
9249 | /* Get vectorized arguments for SLP_NODE. */ |
9250 | vect_get_vec_defs (vinfo, stmt_info, slp_node, ncopies: 1, op0: op, |
9251 | vec_oprnds0: &vec_oprnds, op1: mask, vec_oprnds1: &vec_masks); |
9252 | vec_oprnd = vec_oprnds[0]; |
9253 | if (mask) |
9254 | vec_mask = vec_masks[0]; |
9255 | } |
9256 | else |
9257 | { |
9258 | /* For interleaved stores we collect vectorized defs for all the |
9259 | stores in the group in DR_CHAIN. DR_CHAIN is then used as an |
9260 | input to vect_permute_store_chain(). |
9261 | |
9262 | If the store is not grouped, DR_GROUP_SIZE is 1, and DR_CHAIN |
9263 | is of size 1. */ |
9264 | stmt_vec_info next_stmt_info = first_stmt_info; |
9265 | for (i = 0; i < group_size; i++) |
9266 | { |
9267 | /* Since gaps are not supported for interleaved stores, |
9268 | DR_GROUP_SIZE is the exact number of stmts in the chain. |
9269 | Therefore, NEXT_STMT_INFO can't be NULL_TREE. In case |
9270 | that there is no interleaving, DR_GROUP_SIZE is 1, |
9271 | and only one iteration of the loop will be executed. */ |
9272 | op = vect_get_store_rhs (stmt_info: next_stmt_info); |
9273 | if (costing_p) |
9274 | update_prologue_cost (&prologue_cost, op); |
9275 | else |
9276 | { |
9277 | vect_get_vec_defs_for_operand (vinfo, stmt_vinfo: next_stmt_info, |
9278 | ncopies, op, |
9279 | vec_oprnds: gvec_oprnds[i]); |
9280 | vec_oprnd = (*gvec_oprnds[i])[0]; |
9281 | dr_chain.quick_push (obj: vec_oprnd); |
9282 | } |
9283 | next_stmt_info = DR_GROUP_NEXT_ELEMENT (next_stmt_info); |
9284 | } |
9285 | if (mask && !costing_p) |
9286 | { |
9287 | vect_get_vec_defs_for_operand (vinfo, stmt_vinfo: stmt_info, ncopies, |
9288 | op: mask, vec_oprnds: &vec_masks, |
9289 | vectype: mask_vectype); |
9290 | vec_mask = vec_masks[0]; |
9291 | } |
9292 | } |
9293 | |
9294 | /* We should have catched mismatched types earlier. */ |
9295 | gcc_assert (costing_p |
9296 | || useless_type_conversion_p (vectype, |
9297 | TREE_TYPE (vec_oprnd))); |
9298 | bool simd_lane_access_p |
9299 | = STMT_VINFO_SIMD_LANE_ACCESS_P (stmt_info) != 0; |
9300 | if (!costing_p |
9301 | && simd_lane_access_p |
9302 | && !loop_masks |
9303 | && TREE_CODE (DR_BASE_ADDRESS (first_dr_info->dr)) == ADDR_EXPR |
9304 | && VAR_P (TREE_OPERAND (DR_BASE_ADDRESS (first_dr_info->dr), 0)) |
9305 | && integer_zerop (get_dr_vinfo_offset (vinfo, dr_info: first_dr_info)) |
9306 | && integer_zerop (DR_INIT (first_dr_info->dr)) |
9307 | && alias_sets_conflict_p (get_alias_set (aggr_type), |
9308 | get_alias_set (TREE_TYPE (ref_type)))) |
9309 | { |
9310 | dataref_ptr = unshare_expr (DR_BASE_ADDRESS (first_dr_info->dr)); |
9311 | dataref_offset = build_int_cst (ref_type, 0); |
9312 | } |
9313 | else if (!costing_p) |
9314 | dataref_ptr |
9315 | = vect_create_data_ref_ptr (vinfo, first_stmt_info, aggr_type, |
9316 | simd_lane_access_p ? loop : NULL, |
9317 | offset, &dummy, gsi, &ptr_incr, |
9318 | simd_lane_access_p, bump); |
9319 | } |
9320 | else if (!costing_p) |
9321 | { |
9322 | gcc_assert (!LOOP_VINFO_USING_SELECT_VL_P (loop_vinfo)); |
9323 | /* DR_CHAIN is then used as an input to vect_permute_store_chain(). |
9324 | If the store is not grouped, DR_GROUP_SIZE is 1, and DR_CHAIN is |
9325 | of size 1. */ |
9326 | for (i = 0; i < group_size; i++) |
9327 | { |
9328 | vec_oprnd = (*gvec_oprnds[i])[j]; |
9329 | dr_chain[i] = vec_oprnd; |
9330 | } |
9331 | if (mask) |
9332 | vec_mask = vec_masks[j]; |
9333 | if (dataref_offset) |
9334 | dataref_offset = int_const_binop (PLUS_EXPR, dataref_offset, bump); |
9335 | else |
9336 | dataref_ptr = bump_vector_ptr (vinfo, dataref_ptr, ptr_incr, gsi, |
9337 | stmt_info, bump); |
9338 | } |
9339 | |
9340 | new_stmt = NULL; |
9341 | if (grouped_store) |
9342 | { |
9343 | /* Permute. */ |
9344 | gcc_assert (memory_access_type == VMAT_CONTIGUOUS_PERMUTE); |
9345 | if (costing_p) |
9346 | { |
9347 | int group_size = DR_GROUP_SIZE (first_stmt_info); |
9348 | int nstmts = ceil_log2 (x: group_size) * group_size; |
9349 | inside_cost += record_stmt_cost (body_cost_vec: cost_vec, count: nstmts, kind: vec_perm, |
9350 | stmt_info, misalign: 0, where: vect_body); |
9351 | if (dump_enabled_p ()) |
9352 | dump_printf_loc (MSG_NOTE, vect_location, |
9353 | "vect_model_store_cost: " |
9354 | "strided group_size = %d .\n" , |
9355 | group_size); |
9356 | } |
9357 | else |
9358 | vect_permute_store_chain (vinfo, dr_chain, group_size, stmt_info, |
9359 | gsi, &result_chain); |
9360 | } |
9361 | |
9362 | stmt_vec_info next_stmt_info = first_stmt_info; |
9363 | for (i = 0; i < vec_num; i++) |
9364 | { |
9365 | if (!costing_p) |
9366 | { |
9367 | if (slp) |
9368 | vec_oprnd = vec_oprnds[i]; |
9369 | else if (grouped_store) |
9370 | /* For grouped stores vectorized defs are interleaved in |
9371 | vect_permute_store_chain(). */ |
9372 | vec_oprnd = result_chain[i]; |
9373 | } |
9374 | |
9375 | if (memory_access_type == VMAT_CONTIGUOUS_REVERSE) |
9376 | { |
9377 | if (costing_p) |
9378 | inside_cost += record_stmt_cost (body_cost_vec: cost_vec, count: 1, kind: vec_perm, |
9379 | stmt_info, misalign: 0, where: vect_body); |
9380 | else |
9381 | { |
9382 | tree perm_mask = perm_mask_for_reverse (vectype); |
9383 | tree perm_dest = vect_create_destination_var ( |
9384 | vect_get_store_rhs (stmt_info), vectype); |
9385 | tree new_temp = make_ssa_name (var: perm_dest); |
9386 | |
9387 | /* Generate the permute statement. */ |
9388 | gimple *perm_stmt |
9389 | = gimple_build_assign (new_temp, VEC_PERM_EXPR, vec_oprnd, |
9390 | vec_oprnd, perm_mask); |
9391 | vect_finish_stmt_generation (vinfo, stmt_info, vec_stmt: perm_stmt, |
9392 | gsi); |
9393 | |
9394 | perm_stmt = SSA_NAME_DEF_STMT (new_temp); |
9395 | vec_oprnd = new_temp; |
9396 | } |
9397 | } |
9398 | |
9399 | if (costing_p) |
9400 | { |
9401 | n_adjacent_stores++; |
9402 | |
9403 | if (!slp) |
9404 | { |
9405 | next_stmt_info = DR_GROUP_NEXT_ELEMENT (next_stmt_info); |
9406 | if (!next_stmt_info) |
9407 | break; |
9408 | } |
9409 | |
9410 | continue; |
9411 | } |
9412 | |
9413 | tree final_mask = NULL_TREE; |
9414 | tree final_len = NULL_TREE; |
9415 | tree bias = NULL_TREE; |
9416 | if (loop_masks) |
9417 | final_mask = vect_get_loop_mask (loop_vinfo, gsi, loop_masks, |
9418 | vec_num * ncopies, vectype, |
9419 | vec_num * j + i); |
9420 | if (slp && vec_mask) |
9421 | vec_mask = vec_masks[i]; |
9422 | if (vec_mask) |
9423 | final_mask = prepare_vec_mask (loop_vinfo, mask_type: mask_vectype, loop_mask: final_mask, |
9424 | vec_mask, gsi); |
9425 | |
9426 | if (i > 0) |
9427 | /* Bump the vector pointer. */ |
9428 | dataref_ptr = bump_vector_ptr (vinfo, dataref_ptr, ptr_incr, gsi, |
9429 | stmt_info, bump); |
9430 | |
9431 | unsigned misalign; |
9432 | unsigned HOST_WIDE_INT align; |
9433 | align = known_alignment (DR_TARGET_ALIGNMENT (first_dr_info)); |
9434 | if (alignment_support_scheme == dr_aligned) |
9435 | misalign = 0; |
9436 | else if (misalignment == DR_MISALIGNMENT_UNKNOWN) |
9437 | { |
9438 | align = dr_alignment (vect_dr_behavior (vinfo, dr_info: first_dr_info)); |
9439 | misalign = 0; |
9440 | } |
9441 | else |
9442 | misalign = misalignment; |
9443 | if (dataref_offset == NULL_TREE |
9444 | && TREE_CODE (dataref_ptr) == SSA_NAME) |
9445 | set_ptr_info_alignment (get_ptr_info (dataref_ptr), align, |
9446 | misalign); |
9447 | align = least_bit_hwi (x: misalign | align); |
9448 | |
9449 | /* Compute IFN when LOOP_LENS or final_mask valid. */ |
9450 | machine_mode vmode = TYPE_MODE (vectype); |
9451 | machine_mode new_vmode = vmode; |
9452 | internal_fn partial_ifn = IFN_LAST; |
9453 | if (loop_lens) |
9454 | { |
9455 | opt_machine_mode new_ovmode |
9456 | = get_len_load_store_mode (vmode, false, &partial_ifn); |
9457 | new_vmode = new_ovmode.require (); |
9458 | unsigned factor |
9459 | = (new_ovmode == vmode) ? 1 : GET_MODE_UNIT_SIZE (vmode); |
9460 | final_len = vect_get_loop_len (loop_vinfo, gsi, loop_lens, |
9461 | vec_num * ncopies, vectype, |
9462 | vec_num * j + i, factor); |
9463 | } |
9464 | else if (final_mask) |
9465 | { |
9466 | if (!can_vec_mask_load_store_p ( |
9467 | vmode, TYPE_MODE (TREE_TYPE (final_mask)), false, |
9468 | &partial_ifn)) |
9469 | gcc_unreachable (); |
9470 | } |
9471 | |
9472 | if (partial_ifn == IFN_MASK_LEN_STORE) |
9473 | { |
9474 | if (!final_len) |
9475 | { |
9476 | /* Pass VF value to 'len' argument of |
9477 | MASK_LEN_STORE if LOOP_LENS is invalid. */ |
9478 | final_len = size_int (TYPE_VECTOR_SUBPARTS (vectype)); |
9479 | } |
9480 | if (!final_mask) |
9481 | { |
9482 | /* Pass all ones value to 'mask' argument of |
9483 | MASK_LEN_STORE if final_mask is invalid. */ |
9484 | mask_vectype = truth_type_for (vectype); |
9485 | final_mask = build_minus_one_cst (mask_vectype); |
9486 | } |
9487 | } |
9488 | if (final_len) |
9489 | { |
9490 | signed char biasval |
9491 | = LOOP_VINFO_PARTIAL_LOAD_STORE_BIAS (loop_vinfo); |
9492 | |
9493 | bias = build_int_cst (intQI_type_node, biasval); |
9494 | } |
9495 | |
9496 | /* Arguments are ready. Create the new vector stmt. */ |
9497 | if (final_len) |
9498 | { |
9499 | gcall *call; |
9500 | tree ptr = build_int_cst (ref_type, align * BITS_PER_UNIT); |
9501 | /* Need conversion if it's wrapped with VnQI. */ |
9502 | if (vmode != new_vmode) |
9503 | { |
9504 | tree new_vtype |
9505 | = build_vector_type_for_mode (unsigned_intQI_type_node, |
9506 | new_vmode); |
9507 | tree var = vect_get_new_ssa_name (new_vtype, vect_simple_var); |
9508 | vec_oprnd = build1 (VIEW_CONVERT_EXPR, new_vtype, vec_oprnd); |
9509 | gassign *new_stmt |
9510 | = gimple_build_assign (var, VIEW_CONVERT_EXPR, vec_oprnd); |
9511 | vect_finish_stmt_generation (vinfo, stmt_info, vec_stmt: new_stmt, gsi); |
9512 | vec_oprnd = var; |
9513 | } |
9514 | |
9515 | if (partial_ifn == IFN_MASK_LEN_STORE) |
9516 | call = gimple_build_call_internal (IFN_MASK_LEN_STORE, 6, |
9517 | dataref_ptr, ptr, final_mask, |
9518 | final_len, bias, vec_oprnd); |
9519 | else |
9520 | call = gimple_build_call_internal (IFN_LEN_STORE, 5, |
9521 | dataref_ptr, ptr, final_len, |
9522 | bias, vec_oprnd); |
9523 | gimple_call_set_nothrow (s: call, nothrow_p: true); |
9524 | vect_finish_stmt_generation (vinfo, stmt_info, vec_stmt: call, gsi); |
9525 | new_stmt = call; |
9526 | } |
9527 | else if (final_mask) |
9528 | { |
9529 | tree ptr = build_int_cst (ref_type, align * BITS_PER_UNIT); |
9530 | gcall *call |
9531 | = gimple_build_call_internal (IFN_MASK_STORE, 4, dataref_ptr, |
9532 | ptr, final_mask, vec_oprnd); |
9533 | gimple_call_set_nothrow (s: call, nothrow_p: true); |
9534 | vect_finish_stmt_generation (vinfo, stmt_info, vec_stmt: call, gsi); |
9535 | new_stmt = call; |
9536 | } |
9537 | else |
9538 | { |
9539 | data_ref |
9540 | = fold_build2 (MEM_REF, vectype, dataref_ptr, |
9541 | dataref_offset ? dataref_offset |
9542 | : build_int_cst (ref_type, 0)); |
9543 | if (alignment_support_scheme == dr_aligned) |
9544 | ; |
9545 | else |
9546 | TREE_TYPE (data_ref) |
9547 | = build_aligned_type (TREE_TYPE (data_ref), |
9548 | align * BITS_PER_UNIT); |
9549 | vect_copy_ref_info (data_ref, DR_REF (first_dr_info->dr)); |
9550 | new_stmt = gimple_build_assign (data_ref, vec_oprnd); |
9551 | vect_finish_stmt_generation (vinfo, stmt_info, vec_stmt: new_stmt, gsi); |
9552 | } |
9553 | |
9554 | if (slp) |
9555 | continue; |
9556 | |
9557 | next_stmt_info = DR_GROUP_NEXT_ELEMENT (next_stmt_info); |
9558 | if (!next_stmt_info) |
9559 | break; |
9560 | } |
9561 | if (!slp && !costing_p) |
9562 | { |
9563 | if (j == 0) |
9564 | *vec_stmt = new_stmt; |
9565 | STMT_VINFO_VEC_STMTS (stmt_info).safe_push (obj: new_stmt); |
9566 | } |
9567 | } |
9568 | |
9569 | if (costing_p) |
9570 | { |
9571 | if (n_adjacent_stores > 0) |
9572 | vect_get_store_cost (vinfo, stmt_info, ncopies: n_adjacent_stores, |
9573 | alignment_support_scheme, misalignment, |
9574 | inside_cost: &inside_cost, body_cost_vec: cost_vec); |
9575 | |
9576 | /* When vectorizing a store into the function result assign |
9577 | a penalty if the function returns in a multi-register location. |
9578 | In this case we assume we'll end up with having to spill the |
9579 | vector result and do piecewise loads as a conservative estimate. */ |
9580 | tree base = get_base_address (STMT_VINFO_DATA_REF (stmt_info)->ref); |
9581 | if (base |
9582 | && (TREE_CODE (base) == RESULT_DECL |
9583 | || (DECL_P (base) && cfun_returns (decl: base))) |
9584 | && !aggregate_value_p (base, cfun->decl)) |
9585 | { |
9586 | rtx reg = hard_function_value (TREE_TYPE (base), cfun->decl, 0, 1); |
9587 | /* ??? Handle PARALLEL in some way. */ |
9588 | if (REG_P (reg)) |
9589 | { |
9590 | int nregs = hard_regno_nregs (REGNO (reg), GET_MODE (reg)); |
9591 | /* Assume that a single reg-reg move is possible and cheap, |
9592 | do not account for vector to gp register move cost. */ |
9593 | if (nregs > 1) |
9594 | { |
9595 | /* Spill. */ |
9596 | prologue_cost |
9597 | += record_stmt_cost (body_cost_vec: cost_vec, count: ncopies, kind: vector_store, |
9598 | stmt_info, misalign: 0, where: vect_epilogue); |
9599 | /* Loads. */ |
9600 | prologue_cost |
9601 | += record_stmt_cost (body_cost_vec: cost_vec, count: ncopies * nregs, kind: scalar_load, |
9602 | stmt_info, misalign: 0, where: vect_epilogue); |
9603 | } |
9604 | } |
9605 | } |
9606 | if (dump_enabled_p ()) |
9607 | dump_printf_loc (MSG_NOTE, vect_location, |
9608 | "vect_model_store_cost: inside_cost = %d, " |
9609 | "prologue_cost = %d .\n" , |
9610 | inside_cost, prologue_cost); |
9611 | } |
9612 | |
9613 | return true; |
9614 | } |
9615 | |
9616 | /* Given a vector type VECTYPE, turns permutation SEL into the equivalent |
9617 | VECTOR_CST mask. No checks are made that the target platform supports the |
9618 | mask, so callers may wish to test can_vec_perm_const_p separately, or use |
9619 | vect_gen_perm_mask_checked. */ |
9620 | |
9621 | tree |
9622 | vect_gen_perm_mask_any (tree vectype, const vec_perm_indices &sel) |
9623 | { |
9624 | tree mask_type; |
9625 | |
9626 | poly_uint64 nunits = sel.length (); |
9627 | gcc_assert (known_eq (nunits, TYPE_VECTOR_SUBPARTS (vectype))); |
9628 | |
9629 | mask_type = build_vector_type (ssizetype, nunits); |
9630 | return vec_perm_indices_to_tree (mask_type, sel); |
9631 | } |
9632 | |
9633 | /* Checked version of vect_gen_perm_mask_any. Asserts can_vec_perm_const_p, |
9634 | i.e. that the target supports the pattern _for arbitrary input vectors_. */ |
9635 | |
9636 | tree |
9637 | vect_gen_perm_mask_checked (tree vectype, const vec_perm_indices &sel) |
9638 | { |
9639 | machine_mode vmode = TYPE_MODE (vectype); |
9640 | gcc_assert (can_vec_perm_const_p (vmode, vmode, sel)); |
9641 | return vect_gen_perm_mask_any (vectype, sel); |
9642 | } |
9643 | |
9644 | /* Given a vector variable X and Y, that was generated for the scalar |
9645 | STMT_INFO, generate instructions to permute the vector elements of X and Y |
9646 | using permutation mask MASK_VEC, insert them at *GSI and return the |
9647 | permuted vector variable. */ |
9648 | |
9649 | static tree |
9650 | permute_vec_elements (vec_info *vinfo, |
9651 | tree x, tree y, tree mask_vec, stmt_vec_info stmt_info, |
9652 | gimple_stmt_iterator *gsi) |
9653 | { |
9654 | tree vectype = TREE_TYPE (x); |
9655 | tree perm_dest, data_ref; |
9656 | gimple *perm_stmt; |
9657 | |
9658 | tree scalar_dest = gimple_get_lhs (stmt_info->stmt); |
9659 | if (scalar_dest && TREE_CODE (scalar_dest) == SSA_NAME) |
9660 | perm_dest = vect_create_destination_var (scalar_dest, vectype); |
9661 | else |
9662 | perm_dest = vect_get_new_vect_var (vectype, vect_simple_var, NULL); |
9663 | data_ref = make_ssa_name (var: perm_dest); |
9664 | |
9665 | /* Generate the permute statement. */ |
9666 | perm_stmt = gimple_build_assign (data_ref, VEC_PERM_EXPR, x, y, mask_vec); |
9667 | vect_finish_stmt_generation (vinfo, stmt_info, vec_stmt: perm_stmt, gsi); |
9668 | |
9669 | return data_ref; |
9670 | } |
9671 | |
9672 | /* Hoist the definitions of all SSA uses on STMT_INFO out of the loop LOOP, |
9673 | inserting them on the loops preheader edge. Returns true if we |
9674 | were successful in doing so (and thus STMT_INFO can be moved then), |
9675 | otherwise returns false. HOIST_P indicates if we want to hoist the |
9676 | definitions of all SSA uses, it would be false when we are costing. */ |
9677 | |
9678 | static bool |
9679 | hoist_defs_of_uses (stmt_vec_info stmt_info, class loop *loop, bool hoist_p) |
9680 | { |
9681 | ssa_op_iter i; |
9682 | tree op; |
9683 | bool any = false; |
9684 | |
9685 | FOR_EACH_SSA_TREE_OPERAND (op, stmt_info->stmt, i, SSA_OP_USE) |
9686 | { |
9687 | gimple *def_stmt = SSA_NAME_DEF_STMT (op); |
9688 | if (!gimple_nop_p (g: def_stmt) |
9689 | && flow_bb_inside_loop_p (loop, gimple_bb (g: def_stmt))) |
9690 | { |
9691 | /* Make sure we don't need to recurse. While we could do |
9692 | so in simple cases when there are more complex use webs |
9693 | we don't have an easy way to preserve stmt order to fulfil |
9694 | dependencies within them. */ |
9695 | tree op2; |
9696 | ssa_op_iter i2; |
9697 | if (gimple_code (g: def_stmt) == GIMPLE_PHI) |
9698 | return false; |
9699 | FOR_EACH_SSA_TREE_OPERAND (op2, def_stmt, i2, SSA_OP_USE) |
9700 | { |
9701 | gimple *def_stmt2 = SSA_NAME_DEF_STMT (op2); |
9702 | if (!gimple_nop_p (g: def_stmt2) |
9703 | && flow_bb_inside_loop_p (loop, gimple_bb (g: def_stmt2))) |
9704 | return false; |
9705 | } |
9706 | any = true; |
9707 | } |
9708 | } |
9709 | |
9710 | if (!any) |
9711 | return true; |
9712 | |
9713 | if (!hoist_p) |
9714 | return true; |
9715 | |
9716 | FOR_EACH_SSA_TREE_OPERAND (op, stmt_info->stmt, i, SSA_OP_USE) |
9717 | { |
9718 | gimple *def_stmt = SSA_NAME_DEF_STMT (op); |
9719 | if (!gimple_nop_p (g: def_stmt) |
9720 | && flow_bb_inside_loop_p (loop, gimple_bb (g: def_stmt))) |
9721 | { |
9722 | gimple_stmt_iterator gsi = gsi_for_stmt (def_stmt); |
9723 | gsi_remove (&gsi, false); |
9724 | gsi_insert_on_edge_immediate (loop_preheader_edge (loop), def_stmt); |
9725 | } |
9726 | } |
9727 | |
9728 | return true; |
9729 | } |
9730 | |
9731 | /* vectorizable_load. |
9732 | |
9733 | Check if STMT_INFO reads a non scalar data-ref (array/pointer/structure) |
9734 | that can be vectorized. |
9735 | If VEC_STMT is also passed, vectorize STMT_INFO: create a vectorized |
9736 | stmt to replace it, put it in VEC_STMT, and insert it at GSI. |
9737 | Return true if STMT_INFO is vectorizable in this way. */ |
9738 | |
9739 | static bool |
9740 | vectorizable_load (vec_info *vinfo, |
9741 | stmt_vec_info stmt_info, gimple_stmt_iterator *gsi, |
9742 | gimple **vec_stmt, slp_tree slp_node, |
9743 | stmt_vector_for_cost *cost_vec) |
9744 | { |
9745 | tree scalar_dest; |
9746 | tree vec_dest = NULL; |
9747 | tree data_ref = NULL; |
9748 | loop_vec_info loop_vinfo = dyn_cast <loop_vec_info> (p: vinfo); |
9749 | class loop *loop = NULL; |
9750 | class loop *containing_loop = gimple_bb (g: stmt_info->stmt)->loop_father; |
9751 | bool nested_in_vect_loop = false; |
9752 | tree elem_type; |
9753 | /* Avoid false positive uninitialized warning, see PR110652. */ |
9754 | tree new_temp = NULL_TREE; |
9755 | machine_mode mode; |
9756 | tree dummy; |
9757 | tree dataref_ptr = NULL_TREE; |
9758 | tree dataref_offset = NULL_TREE; |
9759 | gimple *ptr_incr = NULL; |
9760 | int ncopies; |
9761 | int i, j; |
9762 | unsigned int group_size; |
9763 | poly_uint64 group_gap_adj; |
9764 | tree msq = NULL_TREE, lsq; |
9765 | tree realignment_token = NULL_TREE; |
9766 | gphi *phi = NULL; |
9767 | vec<tree> dr_chain = vNULL; |
9768 | bool grouped_load = false; |
9769 | stmt_vec_info first_stmt_info; |
9770 | stmt_vec_info first_stmt_info_for_drptr = NULL; |
9771 | bool compute_in_loop = false; |
9772 | class loop *at_loop; |
9773 | int vec_num; |
9774 | bool slp = (slp_node != NULL); |
9775 | bool slp_perm = false; |
9776 | bb_vec_info bb_vinfo = dyn_cast <bb_vec_info> (p: vinfo); |
9777 | poly_uint64 vf; |
9778 | tree aggr_type; |
9779 | gather_scatter_info gs_info; |
9780 | tree ref_type; |
9781 | enum vect_def_type mask_dt = vect_unknown_def_type; |
9782 | |
9783 | if (!STMT_VINFO_RELEVANT_P (stmt_info) && !bb_vinfo) |
9784 | return false; |
9785 | |
9786 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def |
9787 | && ! vec_stmt) |
9788 | return false; |
9789 | |
9790 | if (!STMT_VINFO_DATA_REF (stmt_info)) |
9791 | return false; |
9792 | |
9793 | tree mask = NULL_TREE, mask_vectype = NULL_TREE; |
9794 | int mask_index = -1; |
9795 | slp_tree slp_op = NULL; |
9796 | if (gassign *assign = dyn_cast <gassign *> (p: stmt_info->stmt)) |
9797 | { |
9798 | scalar_dest = gimple_assign_lhs (gs: assign); |
9799 | if (TREE_CODE (scalar_dest) != SSA_NAME) |
9800 | return false; |
9801 | |
9802 | tree_code code = gimple_assign_rhs_code (gs: assign); |
9803 | if (code != ARRAY_REF |
9804 | && code != BIT_FIELD_REF |
9805 | && code != INDIRECT_REF |
9806 | && code != COMPONENT_REF |
9807 | && code != IMAGPART_EXPR |
9808 | && code != REALPART_EXPR |
9809 | && code != MEM_REF |
9810 | && TREE_CODE_CLASS (code) != tcc_declaration) |
9811 | return false; |
9812 | } |
9813 | else |
9814 | { |
9815 | gcall *call = dyn_cast <gcall *> (p: stmt_info->stmt); |
9816 | if (!call || !gimple_call_internal_p (gs: call)) |
9817 | return false; |
9818 | |
9819 | internal_fn ifn = gimple_call_internal_fn (gs: call); |
9820 | if (!internal_load_fn_p (ifn)) |
9821 | return false; |
9822 | |
9823 | scalar_dest = gimple_call_lhs (gs: call); |
9824 | if (!scalar_dest) |
9825 | return false; |
9826 | |
9827 | mask_index = internal_fn_mask_index (ifn); |
9828 | if (mask_index >= 0 && slp_node) |
9829 | mask_index = vect_slp_child_index_for_operand |
9830 | (call, op: mask_index, STMT_VINFO_GATHER_SCATTER_P (stmt_info)); |
9831 | if (mask_index >= 0 |
9832 | && !vect_check_scalar_mask (vinfo, stmt_info, slp_node, mask_index, |
9833 | mask: &mask, mask_node: &slp_op, mask_dt_out: &mask_dt, mask_vectype_out: &mask_vectype)) |
9834 | return false; |
9835 | } |
9836 | |
9837 | tree vectype = STMT_VINFO_VECTYPE (stmt_info); |
9838 | poly_uint64 nunits = TYPE_VECTOR_SUBPARTS (node: vectype); |
9839 | |
9840 | if (loop_vinfo) |
9841 | { |
9842 | loop = LOOP_VINFO_LOOP (loop_vinfo); |
9843 | nested_in_vect_loop = nested_in_vect_loop_p (loop, stmt_info); |
9844 | vf = LOOP_VINFO_VECT_FACTOR (loop_vinfo); |
9845 | } |
9846 | else |
9847 | vf = 1; |
9848 | |
9849 | /* Multiple types in SLP are handled by creating the appropriate number of |
9850 | vectorized stmts for each SLP node. Hence, NCOPIES is always 1 in |
9851 | case of SLP. */ |
9852 | if (slp) |
9853 | ncopies = 1; |
9854 | else |
9855 | ncopies = vect_get_num_copies (loop_vinfo, vectype); |
9856 | |
9857 | gcc_assert (ncopies >= 1); |
9858 | |
9859 | /* FORNOW. This restriction should be relaxed. */ |
9860 | if (nested_in_vect_loop && ncopies > 1) |
9861 | { |
9862 | if (dump_enabled_p ()) |
9863 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
9864 | "multiple types in nested loop.\n" ); |
9865 | return false; |
9866 | } |
9867 | |
9868 | /* Invalidate assumptions made by dependence analysis when vectorization |
9869 | on the unrolled body effectively re-orders stmts. */ |
9870 | if (ncopies > 1 |
9871 | && STMT_VINFO_MIN_NEG_DIST (stmt_info) != 0 |
9872 | && maybe_gt (LOOP_VINFO_VECT_FACTOR (loop_vinfo), |
9873 | STMT_VINFO_MIN_NEG_DIST (stmt_info))) |
9874 | { |
9875 | if (dump_enabled_p ()) |
9876 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
9877 | "cannot perform implicit CSE when unrolling " |
9878 | "with negative dependence distance\n" ); |
9879 | return false; |
9880 | } |
9881 | |
9882 | elem_type = TREE_TYPE (vectype); |
9883 | mode = TYPE_MODE (vectype); |
9884 | |
9885 | /* FORNOW. In some cases can vectorize even if data-type not supported |
9886 | (e.g. - data copies). */ |
9887 | if (optab_handler (op: mov_optab, mode) == CODE_FOR_nothing) |
9888 | { |
9889 | if (dump_enabled_p ()) |
9890 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
9891 | "Aligned load, but unsupported type.\n" ); |
9892 | return false; |
9893 | } |
9894 | |
9895 | /* Check if the load is a part of an interleaving chain. */ |
9896 | if (STMT_VINFO_GROUPED_ACCESS (stmt_info)) |
9897 | { |
9898 | grouped_load = true; |
9899 | /* FORNOW */ |
9900 | gcc_assert (!nested_in_vect_loop); |
9901 | gcc_assert (!STMT_VINFO_GATHER_SCATTER_P (stmt_info)); |
9902 | |
9903 | first_stmt_info = DR_GROUP_FIRST_ELEMENT (stmt_info); |
9904 | group_size = DR_GROUP_SIZE (first_stmt_info); |
9905 | |
9906 | /* Refuse non-SLP vectorization of SLP-only groups. */ |
9907 | if (!slp && STMT_VINFO_SLP_VECT_ONLY (first_stmt_info)) |
9908 | { |
9909 | if (dump_enabled_p ()) |
9910 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
9911 | "cannot vectorize load in non-SLP mode.\n" ); |
9912 | return false; |
9913 | } |
9914 | |
9915 | /* Invalidate assumptions made by dependence analysis when vectorization |
9916 | on the unrolled body effectively re-orders stmts. */ |
9917 | if (!PURE_SLP_STMT (stmt_info) |
9918 | && STMT_VINFO_MIN_NEG_DIST (stmt_info) != 0 |
9919 | && maybe_gt (LOOP_VINFO_VECT_FACTOR (loop_vinfo), |
9920 | STMT_VINFO_MIN_NEG_DIST (stmt_info))) |
9921 | { |
9922 | if (dump_enabled_p ()) |
9923 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
9924 | "cannot perform implicit CSE when performing " |
9925 | "group loads with negative dependence distance\n" ); |
9926 | return false; |
9927 | } |
9928 | } |
9929 | else |
9930 | group_size = 1; |
9931 | |
9932 | if (slp && SLP_TREE_LOAD_PERMUTATION (slp_node).exists ()) |
9933 | { |
9934 | slp_perm = true; |
9935 | |
9936 | if (!loop_vinfo) |
9937 | { |
9938 | /* In BB vectorization we may not actually use a loaded vector |
9939 | accessing elements in excess of DR_GROUP_SIZE. */ |
9940 | stmt_vec_info group_info = SLP_TREE_SCALAR_STMTS (slp_node)[0]; |
9941 | group_info = DR_GROUP_FIRST_ELEMENT (group_info); |
9942 | unsigned HOST_WIDE_INT nunits; |
9943 | unsigned j, k, maxk = 0; |
9944 | FOR_EACH_VEC_ELT (SLP_TREE_LOAD_PERMUTATION (slp_node), j, k) |
9945 | if (k > maxk) |
9946 | maxk = k; |
9947 | tree vectype = SLP_TREE_VECTYPE (slp_node); |
9948 | if (!TYPE_VECTOR_SUBPARTS (node: vectype).is_constant (const_value: &nunits) |
9949 | || maxk >= (DR_GROUP_SIZE (group_info) & ~(nunits - 1))) |
9950 | { |
9951 | if (dump_enabled_p ()) |
9952 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
9953 | "BB vectorization with gaps at the end of " |
9954 | "a load is not supported\n" ); |
9955 | return false; |
9956 | } |
9957 | } |
9958 | |
9959 | auto_vec<tree> tem; |
9960 | unsigned n_perms; |
9961 | if (!vect_transform_slp_perm_load (vinfo, slp_node, tem, NULL, vf, |
9962 | true, &n_perms)) |
9963 | { |
9964 | if (dump_enabled_p ()) |
9965 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, |
9966 | vect_location, |
9967 | "unsupported load permutation\n" ); |
9968 | return false; |
9969 | } |
9970 | } |
9971 | |
9972 | vect_memory_access_type memory_access_type; |
9973 | enum dr_alignment_support alignment_support_scheme; |
9974 | int misalignment; |
9975 | poly_int64 poffset; |
9976 | internal_fn lanes_ifn; |
9977 | if (!get_load_store_type (vinfo, stmt_info, vectype, slp_node, masked_p: mask, vls_type: VLS_LOAD, |
9978 | ncopies, memory_access_type: &memory_access_type, poffset: &poffset, |
9979 | alignment_support_scheme: &alignment_support_scheme, misalignment: &misalignment, gs_info: &gs_info, |
9980 | lanes_ifn: &lanes_ifn)) |
9981 | return false; |
9982 | |
9983 | if (mask) |
9984 | { |
9985 | if (memory_access_type == VMAT_CONTIGUOUS) |
9986 | { |
9987 | machine_mode vec_mode = TYPE_MODE (vectype); |
9988 | if (!VECTOR_MODE_P (vec_mode) |
9989 | || !can_vec_mask_load_store_p (vec_mode, |
9990 | TYPE_MODE (mask_vectype), true)) |
9991 | return false; |
9992 | } |
9993 | else if (memory_access_type != VMAT_LOAD_STORE_LANES |
9994 | && memory_access_type != VMAT_GATHER_SCATTER) |
9995 | { |
9996 | if (dump_enabled_p ()) |
9997 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
9998 | "unsupported access type for masked load.\n" ); |
9999 | return false; |
10000 | } |
10001 | else if (memory_access_type == VMAT_GATHER_SCATTER |
10002 | && gs_info.ifn == IFN_LAST |
10003 | && !gs_info.decl) |
10004 | { |
10005 | if (dump_enabled_p ()) |
10006 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
10007 | "unsupported masked emulated gather.\n" ); |
10008 | return false; |
10009 | } |
10010 | } |
10011 | |
10012 | bool costing_p = !vec_stmt; |
10013 | |
10014 | if (costing_p) /* transformation not required. */ |
10015 | { |
10016 | if (slp_node |
10017 | && mask |
10018 | && !vect_maybe_update_slp_op_vectype (slp_op, |
10019 | mask_vectype)) |
10020 | { |
10021 | if (dump_enabled_p ()) |
10022 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
10023 | "incompatible vector types for invariants\n" ); |
10024 | return false; |
10025 | } |
10026 | |
10027 | if (!slp) |
10028 | STMT_VINFO_MEMORY_ACCESS_TYPE (stmt_info) = memory_access_type; |
10029 | |
10030 | if (loop_vinfo |
10031 | && LOOP_VINFO_CAN_USE_PARTIAL_VECTORS_P (loop_vinfo)) |
10032 | check_load_store_for_partial_vectors (loop_vinfo, vectype, slp_node, |
10033 | vls_type: VLS_LOAD, group_size, |
10034 | memory_access_type, gs_info: &gs_info, |
10035 | scalar_mask: mask); |
10036 | |
10037 | if (dump_enabled_p () |
10038 | && memory_access_type != VMAT_ELEMENTWISE |
10039 | && memory_access_type != VMAT_GATHER_SCATTER |
10040 | && alignment_support_scheme != dr_aligned) |
10041 | dump_printf_loc (MSG_NOTE, vect_location, |
10042 | "Vectorizing an unaligned access.\n" ); |
10043 | |
10044 | if (memory_access_type == VMAT_LOAD_STORE_LANES) |
10045 | vinfo->any_known_not_updated_vssa = true; |
10046 | |
10047 | STMT_VINFO_TYPE (stmt_info) = load_vec_info_type; |
10048 | } |
10049 | |
10050 | if (!slp) |
10051 | gcc_assert (memory_access_type |
10052 | == STMT_VINFO_MEMORY_ACCESS_TYPE (stmt_info)); |
10053 | |
10054 | if (dump_enabled_p () && !costing_p) |
10055 | dump_printf_loc (MSG_NOTE, vect_location, |
10056 | "transform load. ncopies = %d\n" , ncopies); |
10057 | |
10058 | /* Transform. */ |
10059 | |
10060 | dr_vec_info *dr_info = STMT_VINFO_DR_INFO (stmt_info), *first_dr_info = NULL; |
10061 | ensure_base_align (dr_info); |
10062 | |
10063 | if (memory_access_type == VMAT_INVARIANT) |
10064 | { |
10065 | gcc_assert (!grouped_load && !mask && !bb_vinfo); |
10066 | /* If we have versioned for aliasing or the loop doesn't |
10067 | have any data dependencies that would preclude this, |
10068 | then we are sure this is a loop invariant load and |
10069 | thus we can insert it on the preheader edge. |
10070 | TODO: hoist_defs_of_uses should ideally be computed |
10071 | once at analysis time, remembered and used in the |
10072 | transform time. */ |
10073 | bool hoist_p = (LOOP_VINFO_NO_DATA_DEPENDENCIES (loop_vinfo) |
10074 | && !nested_in_vect_loop |
10075 | && hoist_defs_of_uses (stmt_info, loop, hoist_p: !costing_p)); |
10076 | if (costing_p) |
10077 | { |
10078 | enum vect_cost_model_location cost_loc |
10079 | = hoist_p ? vect_prologue : vect_body; |
10080 | unsigned int cost = record_stmt_cost (body_cost_vec: cost_vec, count: 1, kind: scalar_load, |
10081 | stmt_info, misalign: 0, where: cost_loc); |
10082 | cost += record_stmt_cost (body_cost_vec: cost_vec, count: 1, kind: scalar_to_vec, stmt_info, misalign: 0, |
10083 | where: cost_loc); |
10084 | unsigned int prologue_cost = hoist_p ? cost : 0; |
10085 | unsigned int inside_cost = hoist_p ? 0 : cost; |
10086 | if (dump_enabled_p ()) |
10087 | dump_printf_loc (MSG_NOTE, vect_location, |
10088 | "vect_model_load_cost: inside_cost = %d, " |
10089 | "prologue_cost = %d .\n" , |
10090 | inside_cost, prologue_cost); |
10091 | return true; |
10092 | } |
10093 | if (hoist_p) |
10094 | { |
10095 | gassign *stmt = as_a <gassign *> (p: stmt_info->stmt); |
10096 | if (dump_enabled_p ()) |
10097 | dump_printf_loc (MSG_NOTE, vect_location, |
10098 | "hoisting out of the vectorized loop: %G" , |
10099 | (gimple *) stmt); |
10100 | scalar_dest = copy_ssa_name (var: scalar_dest); |
10101 | tree rhs = unshare_expr (gimple_assign_rhs1 (gs: stmt)); |
10102 | edge pe = loop_preheader_edge (loop); |
10103 | gphi *vphi = get_virtual_phi (loop->header); |
10104 | tree vuse; |
10105 | if (vphi) |
10106 | vuse = PHI_ARG_DEF_FROM_EDGE (vphi, pe); |
10107 | else |
10108 | vuse = gimple_vuse (g: gsi_stmt (i: *gsi)); |
10109 | gimple *new_stmt = gimple_build_assign (scalar_dest, rhs); |
10110 | gimple_set_vuse (g: new_stmt, vuse); |
10111 | gsi_insert_on_edge_immediate (pe, new_stmt); |
10112 | } |
10113 | /* These copies are all equivalent. */ |
10114 | if (hoist_p) |
10115 | new_temp = vect_init_vector (vinfo, stmt_info, val: scalar_dest, |
10116 | type: vectype, NULL); |
10117 | else |
10118 | { |
10119 | gimple_stmt_iterator gsi2 = *gsi; |
10120 | gsi_next (i: &gsi2); |
10121 | new_temp = vect_init_vector (vinfo, stmt_info, val: scalar_dest, |
10122 | type: vectype, gsi: &gsi2); |
10123 | } |
10124 | gimple *new_stmt = SSA_NAME_DEF_STMT (new_temp); |
10125 | if (slp) |
10126 | for (j = 0; j < (int) SLP_TREE_NUMBER_OF_VEC_STMTS (slp_node); ++j) |
10127 | slp_node->push_vec_def (def: new_stmt); |
10128 | else |
10129 | { |
10130 | for (j = 0; j < ncopies; ++j) |
10131 | STMT_VINFO_VEC_STMTS (stmt_info).safe_push (obj: new_stmt); |
10132 | *vec_stmt = new_stmt; |
10133 | } |
10134 | return true; |
10135 | } |
10136 | |
10137 | if (memory_access_type == VMAT_ELEMENTWISE |
10138 | || memory_access_type == VMAT_STRIDED_SLP) |
10139 | { |
10140 | gimple_stmt_iterator incr_gsi; |
10141 | bool insert_after; |
10142 | tree offvar; |
10143 | tree ivstep; |
10144 | tree running_off; |
10145 | vec<constructor_elt, va_gc> *v = NULL; |
10146 | tree stride_base, stride_step, alias_off; |
10147 | /* Checked by get_load_store_type. */ |
10148 | unsigned int const_nunits = nunits.to_constant (); |
10149 | unsigned HOST_WIDE_INT cst_offset = 0; |
10150 | tree dr_offset; |
10151 | unsigned int inside_cost = 0; |
10152 | |
10153 | gcc_assert (!LOOP_VINFO_USING_PARTIAL_VECTORS_P (loop_vinfo)); |
10154 | gcc_assert (!nested_in_vect_loop); |
10155 | |
10156 | if (grouped_load) |
10157 | { |
10158 | first_stmt_info = DR_GROUP_FIRST_ELEMENT (stmt_info); |
10159 | first_dr_info = STMT_VINFO_DR_INFO (first_stmt_info); |
10160 | } |
10161 | else |
10162 | { |
10163 | first_stmt_info = stmt_info; |
10164 | first_dr_info = dr_info; |
10165 | } |
10166 | |
10167 | if (slp && grouped_load) |
10168 | { |
10169 | group_size = DR_GROUP_SIZE (first_stmt_info); |
10170 | ref_type = get_group_alias_ptr_type (first_stmt_info); |
10171 | } |
10172 | else |
10173 | { |
10174 | if (grouped_load) |
10175 | cst_offset |
10176 | = (tree_to_uhwi (TYPE_SIZE_UNIT (TREE_TYPE (vectype))) |
10177 | * vect_get_place_in_interleaving_chain (stmt_info, |
10178 | first_stmt_info)); |
10179 | group_size = 1; |
10180 | ref_type = reference_alias_ptr_type (DR_REF (dr_info->dr)); |
10181 | } |
10182 | |
10183 | if (!costing_p) |
10184 | { |
10185 | dr_offset = get_dr_vinfo_offset (vinfo, dr_info: first_dr_info); |
10186 | stride_base = fold_build_pointer_plus ( |
10187 | DR_BASE_ADDRESS (first_dr_info->dr), |
10188 | size_binop (PLUS_EXPR, convert_to_ptrofftype (dr_offset), |
10189 | convert_to_ptrofftype (DR_INIT (first_dr_info->dr)))); |
10190 | stride_step = fold_convert (sizetype, DR_STEP (first_dr_info->dr)); |
10191 | |
10192 | /* For a load with loop-invariant (but other than power-of-2) |
10193 | stride (i.e. not a grouped access) like so: |
10194 | |
10195 | for (i = 0; i < n; i += stride) |
10196 | ... = array[i]; |
10197 | |
10198 | we generate a new induction variable and new accesses to |
10199 | form a new vector (or vectors, depending on ncopies): |
10200 | |
10201 | for (j = 0; ; j += VF*stride) |
10202 | tmp1 = array[j]; |
10203 | tmp2 = array[j + stride]; |
10204 | ... |
10205 | vectemp = {tmp1, tmp2, ...} |
10206 | */ |
10207 | |
10208 | ivstep = fold_build2 (MULT_EXPR, TREE_TYPE (stride_step), stride_step, |
10209 | build_int_cst (TREE_TYPE (stride_step), vf)); |
10210 | |
10211 | standard_iv_increment_position (loop, &incr_gsi, &insert_after); |
10212 | |
10213 | stride_base = cse_and_gimplify_to_preheader (loop_vinfo, stride_base); |
10214 | ivstep = cse_and_gimplify_to_preheader (loop_vinfo, ivstep); |
10215 | create_iv (stride_base, PLUS_EXPR, ivstep, NULL, |
10216 | loop, &incr_gsi, insert_after, |
10217 | &offvar, NULL); |
10218 | |
10219 | stride_step = cse_and_gimplify_to_preheader (loop_vinfo, stride_step); |
10220 | } |
10221 | |
10222 | running_off = offvar; |
10223 | alias_off = build_int_cst (ref_type, 0); |
10224 | int nloads = const_nunits; |
10225 | int lnel = 1; |
10226 | tree ltype = TREE_TYPE (vectype); |
10227 | tree lvectype = vectype; |
10228 | auto_vec<tree> dr_chain; |
10229 | if (memory_access_type == VMAT_STRIDED_SLP) |
10230 | { |
10231 | if (group_size < const_nunits) |
10232 | { |
10233 | /* First check if vec_init optab supports construction from vector |
10234 | elts directly. Otherwise avoid emitting a constructor of |
10235 | vector elements by performing the loads using an integer type |
10236 | of the same size, constructing a vector of those and then |
10237 | re-interpreting it as the original vector type. This avoids a |
10238 | huge runtime penalty due to the general inability to perform |
10239 | store forwarding from smaller stores to a larger load. */ |
10240 | tree ptype; |
10241 | tree vtype |
10242 | = vector_vector_composition_type (vtype: vectype, |
10243 | nelts: const_nunits / group_size, |
10244 | ptype: &ptype); |
10245 | if (vtype != NULL_TREE) |
10246 | { |
10247 | nloads = const_nunits / group_size; |
10248 | lnel = group_size; |
10249 | lvectype = vtype; |
10250 | ltype = ptype; |
10251 | } |
10252 | } |
10253 | else |
10254 | { |
10255 | nloads = 1; |
10256 | lnel = const_nunits; |
10257 | ltype = vectype; |
10258 | } |
10259 | ltype = build_aligned_type (ltype, TYPE_ALIGN (TREE_TYPE (vectype))); |
10260 | } |
10261 | /* Load vector(1) scalar_type if it's 1 element-wise vectype. */ |
10262 | else if (nloads == 1) |
10263 | ltype = vectype; |
10264 | |
10265 | if (slp) |
10266 | { |
10267 | /* For SLP permutation support we need to load the whole group, |
10268 | not only the number of vector stmts the permutation result |
10269 | fits in. */ |
10270 | if (slp_perm) |
10271 | { |
10272 | /* We don't yet generate SLP_TREE_LOAD_PERMUTATIONs for |
10273 | variable VF. */ |
10274 | unsigned int const_vf = vf.to_constant (); |
10275 | ncopies = CEIL (group_size * const_vf, const_nunits); |
10276 | dr_chain.create (nelems: ncopies); |
10277 | } |
10278 | else |
10279 | ncopies = SLP_TREE_NUMBER_OF_VEC_STMTS (slp_node); |
10280 | } |
10281 | unsigned int group_el = 0; |
10282 | unsigned HOST_WIDE_INT |
10283 | elsz = tree_to_uhwi (TYPE_SIZE_UNIT (TREE_TYPE (vectype))); |
10284 | unsigned int n_groups = 0; |
10285 | /* For costing some adjacent vector loads, we'd like to cost with |
10286 | the total number of them once instead of cost each one by one. */ |
10287 | unsigned int n_adjacent_loads = 0; |
10288 | for (j = 0; j < ncopies; j++) |
10289 | { |
10290 | if (nloads > 1 && !costing_p) |
10291 | vec_alloc (v, nelems: nloads); |
10292 | gimple *new_stmt = NULL; |
10293 | for (i = 0; i < nloads; i++) |
10294 | { |
10295 | if (costing_p) |
10296 | { |
10297 | /* For VMAT_ELEMENTWISE, just cost it as scalar_load to |
10298 | avoid ICE, see PR110776. */ |
10299 | if (VECTOR_TYPE_P (ltype) |
10300 | && memory_access_type != VMAT_ELEMENTWISE) |
10301 | n_adjacent_loads++; |
10302 | else |
10303 | inside_cost += record_stmt_cost (body_cost_vec: cost_vec, count: 1, kind: scalar_load, |
10304 | stmt_info, misalign: 0, where: vect_body); |
10305 | continue; |
10306 | } |
10307 | tree this_off = build_int_cst (TREE_TYPE (alias_off), |
10308 | group_el * elsz + cst_offset); |
10309 | tree data_ref = build2 (MEM_REF, ltype, running_off, this_off); |
10310 | vect_copy_ref_info (data_ref, DR_REF (first_dr_info->dr)); |
10311 | new_stmt = gimple_build_assign (make_ssa_name (var: ltype), data_ref); |
10312 | vect_finish_stmt_generation (vinfo, stmt_info, vec_stmt: new_stmt, gsi); |
10313 | if (nloads > 1) |
10314 | CONSTRUCTOR_APPEND_ELT (v, NULL_TREE, |
10315 | gimple_assign_lhs (new_stmt)); |
10316 | |
10317 | group_el += lnel; |
10318 | if (! slp |
10319 | || group_el == group_size) |
10320 | { |
10321 | n_groups++; |
10322 | /* When doing SLP make sure to not load elements from |
10323 | the next vector iteration, those will not be accessed |
10324 | so just use the last element again. See PR107451. */ |
10325 | if (!slp || known_lt (n_groups, vf)) |
10326 | { |
10327 | tree newoff = copy_ssa_name (var: running_off); |
10328 | gimple *incr |
10329 | = gimple_build_assign (newoff, POINTER_PLUS_EXPR, |
10330 | running_off, stride_step); |
10331 | vect_finish_stmt_generation (vinfo, stmt_info, vec_stmt: incr, gsi); |
10332 | running_off = newoff; |
10333 | } |
10334 | group_el = 0; |
10335 | } |
10336 | } |
10337 | |
10338 | if (nloads > 1) |
10339 | { |
10340 | if (costing_p) |
10341 | inside_cost += record_stmt_cost (body_cost_vec: cost_vec, count: 1, kind: vec_construct, |
10342 | stmt_info, misalign: 0, where: vect_body); |
10343 | else |
10344 | { |
10345 | tree vec_inv = build_constructor (lvectype, v); |
10346 | new_temp = vect_init_vector (vinfo, stmt_info, val: vec_inv, |
10347 | type: lvectype, gsi); |
10348 | new_stmt = SSA_NAME_DEF_STMT (new_temp); |
10349 | if (lvectype != vectype) |
10350 | { |
10351 | new_stmt |
10352 | = gimple_build_assign (make_ssa_name (var: vectype), |
10353 | VIEW_CONVERT_EXPR, |
10354 | build1 (VIEW_CONVERT_EXPR, |
10355 | vectype, new_temp)); |
10356 | vect_finish_stmt_generation (vinfo, stmt_info, vec_stmt: new_stmt, |
10357 | gsi); |
10358 | } |
10359 | } |
10360 | } |
10361 | |
10362 | if (!costing_p) |
10363 | { |
10364 | if (slp) |
10365 | { |
10366 | if (slp_perm) |
10367 | dr_chain.quick_push (obj: gimple_assign_lhs (gs: new_stmt)); |
10368 | else |
10369 | slp_node->push_vec_def (def: new_stmt); |
10370 | } |
10371 | else |
10372 | { |
10373 | if (j == 0) |
10374 | *vec_stmt = new_stmt; |
10375 | STMT_VINFO_VEC_STMTS (stmt_info).safe_push (obj: new_stmt); |
10376 | } |
10377 | } |
10378 | } |
10379 | if (slp_perm) |
10380 | { |
10381 | unsigned n_perms; |
10382 | if (costing_p) |
10383 | { |
10384 | unsigned n_loads; |
10385 | vect_transform_slp_perm_load (vinfo, slp_node, vNULL, NULL, vf, |
10386 | true, &n_perms, &n_loads); |
10387 | inside_cost += record_stmt_cost (body_cost_vec: cost_vec, count: n_perms, kind: vec_perm, |
10388 | stmt_info: first_stmt_info, misalign: 0, where: vect_body); |
10389 | } |
10390 | else |
10391 | vect_transform_slp_perm_load (vinfo, slp_node, dr_chain, gsi, vf, |
10392 | false, &n_perms); |
10393 | } |
10394 | |
10395 | if (costing_p) |
10396 | { |
10397 | if (n_adjacent_loads > 0) |
10398 | vect_get_load_cost (vinfo, stmt_info, ncopies: n_adjacent_loads, |
10399 | alignment_support_scheme, misalignment, add_realign_cost: false, |
10400 | inside_cost: &inside_cost, prologue_cost: nullptr, prologue_cost_vec: cost_vec, body_cost_vec: cost_vec, |
10401 | record_prologue_costs: true); |
10402 | if (dump_enabled_p ()) |
10403 | dump_printf_loc (MSG_NOTE, vect_location, |
10404 | "vect_model_load_cost: inside_cost = %u, " |
10405 | "prologue_cost = 0 .\n" , |
10406 | inside_cost); |
10407 | } |
10408 | |
10409 | return true; |
10410 | } |
10411 | |
10412 | if (memory_access_type == VMAT_GATHER_SCATTER |
10413 | || (!slp && memory_access_type == VMAT_CONTIGUOUS)) |
10414 | grouped_load = false; |
10415 | |
10416 | if (grouped_load |
10417 | || (slp && SLP_TREE_LOAD_PERMUTATION (slp_node).exists ())) |
10418 | { |
10419 | if (grouped_load) |
10420 | { |
10421 | first_stmt_info = DR_GROUP_FIRST_ELEMENT (stmt_info); |
10422 | group_size = DR_GROUP_SIZE (first_stmt_info); |
10423 | } |
10424 | else |
10425 | { |
10426 | first_stmt_info = stmt_info; |
10427 | group_size = 1; |
10428 | } |
10429 | /* For SLP vectorization we directly vectorize a subchain |
10430 | without permutation. */ |
10431 | if (slp && ! SLP_TREE_LOAD_PERMUTATION (slp_node).exists ()) |
10432 | first_stmt_info = SLP_TREE_SCALAR_STMTS (slp_node)[0]; |
10433 | /* For BB vectorization always use the first stmt to base |
10434 | the data ref pointer on. */ |
10435 | if (bb_vinfo) |
10436 | first_stmt_info_for_drptr |
10437 | = vect_find_first_scalar_stmt_in_slp (slp_node); |
10438 | |
10439 | /* Check if the chain of loads is already vectorized. */ |
10440 | if (STMT_VINFO_VEC_STMTS (first_stmt_info).exists () |
10441 | /* For SLP we would need to copy over SLP_TREE_VEC_DEFS. |
10442 | ??? But we can only do so if there is exactly one |
10443 | as we have no way to get at the rest. Leave the CSE |
10444 | opportunity alone. |
10445 | ??? With the group load eventually participating |
10446 | in multiple different permutations (having multiple |
10447 | slp nodes which refer to the same group) the CSE |
10448 | is even wrong code. See PR56270. */ |
10449 | && !slp) |
10450 | { |
10451 | *vec_stmt = STMT_VINFO_VEC_STMTS (stmt_info)[0]; |
10452 | return true; |
10453 | } |
10454 | first_dr_info = STMT_VINFO_DR_INFO (first_stmt_info); |
10455 | group_gap_adj = 0; |
10456 | |
10457 | /* VEC_NUM is the number of vect stmts to be created for this group. */ |
10458 | if (slp) |
10459 | { |
10460 | grouped_load = false; |
10461 | /* If an SLP permutation is from N elements to N elements, |
10462 | and if one vector holds a whole number of N, we can load |
10463 | the inputs to the permutation in the same way as an |
10464 | unpermuted sequence. In other cases we need to load the |
10465 | whole group, not only the number of vector stmts the |
10466 | permutation result fits in. */ |
10467 | unsigned scalar_lanes = SLP_TREE_LANES (slp_node); |
10468 | if (slp_perm |
10469 | && (group_size != scalar_lanes |
10470 | || !multiple_p (a: nunits, b: group_size))) |
10471 | { |
10472 | /* We don't yet generate such SLP_TREE_LOAD_PERMUTATIONs for |
10473 | variable VF; see vect_transform_slp_perm_load. */ |
10474 | unsigned int const_vf = vf.to_constant (); |
10475 | unsigned int const_nunits = nunits.to_constant (); |
10476 | vec_num = CEIL (group_size * const_vf, const_nunits); |
10477 | group_gap_adj = vf * group_size - nunits * vec_num; |
10478 | } |
10479 | else |
10480 | { |
10481 | vec_num = SLP_TREE_NUMBER_OF_VEC_STMTS (slp_node); |
10482 | group_gap_adj |
10483 | = group_size - scalar_lanes; |
10484 | } |
10485 | } |
10486 | else |
10487 | vec_num = group_size; |
10488 | |
10489 | ref_type = get_group_alias_ptr_type (first_stmt_info); |
10490 | } |
10491 | else |
10492 | { |
10493 | first_stmt_info = stmt_info; |
10494 | first_dr_info = dr_info; |
10495 | group_size = vec_num = 1; |
10496 | group_gap_adj = 0; |
10497 | ref_type = reference_alias_ptr_type (DR_REF (first_dr_info->dr)); |
10498 | if (slp) |
10499 | vec_num = SLP_TREE_NUMBER_OF_VEC_STMTS (slp_node); |
10500 | } |
10501 | |
10502 | gcc_assert (alignment_support_scheme); |
10503 | vec_loop_masks *loop_masks |
10504 | = (loop_vinfo && LOOP_VINFO_FULLY_MASKED_P (loop_vinfo) |
10505 | ? &LOOP_VINFO_MASKS (loop_vinfo) |
10506 | : NULL); |
10507 | vec_loop_lens *loop_lens |
10508 | = (loop_vinfo && LOOP_VINFO_FULLY_WITH_LENGTH_P (loop_vinfo) |
10509 | ? &LOOP_VINFO_LENS (loop_vinfo) |
10510 | : NULL); |
10511 | |
10512 | /* Shouldn't go with length-based approach if fully masked. */ |
10513 | gcc_assert (!loop_lens || !loop_masks); |
10514 | |
10515 | /* Targets with store-lane instructions must not require explicit |
10516 | realignment. vect_supportable_dr_alignment always returns either |
10517 | dr_aligned or dr_unaligned_supported for masked operations. */ |
10518 | gcc_assert ((memory_access_type != VMAT_LOAD_STORE_LANES |
10519 | && !mask |
10520 | && !loop_masks) |
10521 | || alignment_support_scheme == dr_aligned |
10522 | || alignment_support_scheme == dr_unaligned_supported); |
10523 | |
10524 | /* In case the vectorization factor (VF) is bigger than the number |
10525 | of elements that we can fit in a vectype (nunits), we have to generate |
10526 | more than one vector stmt - i.e - we need to "unroll" the |
10527 | vector stmt by a factor VF/nunits. In doing so, we record a pointer |
10528 | from one copy of the vector stmt to the next, in the field |
10529 | STMT_VINFO_RELATED_STMT. This is necessary in order to allow following |
10530 | stages to find the correct vector defs to be used when vectorizing |
10531 | stmts that use the defs of the current stmt. The example below |
10532 | illustrates the vectorization process when VF=16 and nunits=4 (i.e., we |
10533 | need to create 4 vectorized stmts): |
10534 | |
10535 | before vectorization: |
10536 | RELATED_STMT VEC_STMT |
10537 | S1: x = memref - - |
10538 | S2: z = x + 1 - - |
10539 | |
10540 | step 1: vectorize stmt S1: |
10541 | We first create the vector stmt VS1_0, and, as usual, record a |
10542 | pointer to it in the STMT_VINFO_VEC_STMT of the scalar stmt S1. |
10543 | Next, we create the vector stmt VS1_1, and record a pointer to |
10544 | it in the STMT_VINFO_RELATED_STMT of the vector stmt VS1_0. |
10545 | Similarly, for VS1_2 and VS1_3. This is the resulting chain of |
10546 | stmts and pointers: |
10547 | RELATED_STMT VEC_STMT |
10548 | VS1_0: vx0 = memref0 VS1_1 - |
10549 | VS1_1: vx1 = memref1 VS1_2 - |
10550 | VS1_2: vx2 = memref2 VS1_3 - |
10551 | VS1_3: vx3 = memref3 - - |
10552 | S1: x = load - VS1_0 |
10553 | S2: z = x + 1 - - |
10554 | */ |
10555 | |
10556 | /* In case of interleaving (non-unit grouped access): |
10557 | |
10558 | S1: x2 = &base + 2 |
10559 | S2: x0 = &base |
10560 | S3: x1 = &base + 1 |
10561 | S4: x3 = &base + 3 |
10562 | |
10563 | Vectorized loads are created in the order of memory accesses |
10564 | starting from the access of the first stmt of the chain: |
10565 | |
10566 | VS1: vx0 = &base |
10567 | VS2: vx1 = &base + vec_size*1 |
10568 | VS3: vx3 = &base + vec_size*2 |
10569 | VS4: vx4 = &base + vec_size*3 |
10570 | |
10571 | Then permutation statements are generated: |
10572 | |
10573 | VS5: vx5 = VEC_PERM_EXPR < vx0, vx1, { 0, 2, ..., i*2 } > |
10574 | VS6: vx6 = VEC_PERM_EXPR < vx0, vx1, { 1, 3, ..., i*2+1 } > |
10575 | ... |
10576 | |
10577 | And they are put in STMT_VINFO_VEC_STMT of the corresponding scalar stmts |
10578 | (the order of the data-refs in the output of vect_permute_load_chain |
10579 | corresponds to the order of scalar stmts in the interleaving chain - see |
10580 | the documentation of vect_permute_load_chain()). |
10581 | The generation of permutation stmts and recording them in |
10582 | STMT_VINFO_VEC_STMT is done in vect_transform_grouped_load(). |
10583 | |
10584 | In case of both multiple types and interleaving, the vector loads and |
10585 | permutation stmts above are created for every copy. The result vector |
10586 | stmts are put in STMT_VINFO_VEC_STMT for the first copy and in the |
10587 | corresponding STMT_VINFO_RELATED_STMT for the next copies. */ |
10588 | |
10589 | /* If the data reference is aligned (dr_aligned) or potentially unaligned |
10590 | on a target that supports unaligned accesses (dr_unaligned_supported) |
10591 | we generate the following code: |
10592 | p = initial_addr; |
10593 | indx = 0; |
10594 | loop { |
10595 | p = p + indx * vectype_size; |
10596 | vec_dest = *(p); |
10597 | indx = indx + 1; |
10598 | } |
10599 | |
10600 | Otherwise, the data reference is potentially unaligned on a target that |
10601 | does not support unaligned accesses (dr_explicit_realign_optimized) - |
10602 | then generate the following code, in which the data in each iteration is |
10603 | obtained by two vector loads, one from the previous iteration, and one |
10604 | from the current iteration: |
10605 | p1 = initial_addr; |
10606 | msq_init = *(floor(p1)) |
10607 | p2 = initial_addr + VS - 1; |
10608 | realignment_token = call target_builtin; |
10609 | indx = 0; |
10610 | loop { |
10611 | p2 = p2 + indx * vectype_size |
10612 | lsq = *(floor(p2)) |
10613 | vec_dest = realign_load (msq, lsq, realignment_token) |
10614 | indx = indx + 1; |
10615 | msq = lsq; |
10616 | } */ |
10617 | |
10618 | /* If the misalignment remains the same throughout the execution of the |
10619 | loop, we can create the init_addr and permutation mask at the loop |
10620 | preheader. Otherwise, it needs to be created inside the loop. |
10621 | This can only occur when vectorizing memory accesses in the inner-loop |
10622 | nested within an outer-loop that is being vectorized. */ |
10623 | |
10624 | if (nested_in_vect_loop |
10625 | && !multiple_p (DR_STEP_ALIGNMENT (dr_info->dr), |
10626 | b: GET_MODE_SIZE (TYPE_MODE (vectype)))) |
10627 | { |
10628 | gcc_assert (alignment_support_scheme != dr_explicit_realign_optimized); |
10629 | compute_in_loop = true; |
10630 | } |
10631 | |
10632 | bool diff_first_stmt_info |
10633 | = first_stmt_info_for_drptr && first_stmt_info != first_stmt_info_for_drptr; |
10634 | |
10635 | tree offset = NULL_TREE; |
10636 | if ((alignment_support_scheme == dr_explicit_realign_optimized |
10637 | || alignment_support_scheme == dr_explicit_realign) |
10638 | && !compute_in_loop) |
10639 | { |
10640 | /* If we have different first_stmt_info, we can't set up realignment |
10641 | here, since we can't guarantee first_stmt_info DR has been |
10642 | initialized yet, use first_stmt_info_for_drptr DR by bumping the |
10643 | distance from first_stmt_info DR instead as below. */ |
10644 | if (!costing_p) |
10645 | { |
10646 | if (!diff_first_stmt_info) |
10647 | msq = vect_setup_realignment (vinfo, first_stmt_info, gsi, |
10648 | &realignment_token, |
10649 | alignment_support_scheme, NULL_TREE, |
10650 | &at_loop); |
10651 | if (alignment_support_scheme == dr_explicit_realign_optimized) |
10652 | { |
10653 | phi = as_a<gphi *> (SSA_NAME_DEF_STMT (msq)); |
10654 | offset = size_binop (MINUS_EXPR, TYPE_SIZE_UNIT (vectype), |
10655 | size_one_node); |
10656 | gcc_assert (!first_stmt_info_for_drptr); |
10657 | } |
10658 | } |
10659 | } |
10660 | else |
10661 | at_loop = loop; |
10662 | |
10663 | if (!known_eq (poffset, 0)) |
10664 | offset = (offset |
10665 | ? size_binop (PLUS_EXPR, offset, size_int (poffset)) |
10666 | : size_int (poffset)); |
10667 | |
10668 | tree bump; |
10669 | tree vec_offset = NULL_TREE; |
10670 | if (STMT_VINFO_GATHER_SCATTER_P (stmt_info)) |
10671 | { |
10672 | aggr_type = NULL_TREE; |
10673 | bump = NULL_TREE; |
10674 | } |
10675 | else if (memory_access_type == VMAT_GATHER_SCATTER) |
10676 | { |
10677 | aggr_type = elem_type; |
10678 | if (!costing_p) |
10679 | vect_get_strided_load_store_ops (stmt_info, loop_vinfo, gsi, gs_info: &gs_info, |
10680 | dataref_bump: &bump, vec_offset: &vec_offset, loop_lens); |
10681 | } |
10682 | else |
10683 | { |
10684 | if (memory_access_type == VMAT_LOAD_STORE_LANES) |
10685 | aggr_type = build_array_type_nelts (elem_type, vec_num * nunits); |
10686 | else |
10687 | aggr_type = vectype; |
10688 | bump = vect_get_data_ptr_increment (vinfo, gsi, dr_info, aggr_type, |
10689 | memory_access_type, loop_lens); |
10690 | } |
10691 | |
10692 | auto_vec<tree> vec_offsets; |
10693 | auto_vec<tree> vec_masks; |
10694 | if (mask && !costing_p) |
10695 | { |
10696 | if (slp_node) |
10697 | vect_get_slp_defs (SLP_TREE_CHILDREN (slp_node)[mask_index], |
10698 | &vec_masks); |
10699 | else |
10700 | vect_get_vec_defs_for_operand (vinfo, stmt_vinfo: stmt_info, ncopies, op: mask, |
10701 | vec_oprnds: &vec_masks, vectype: mask_vectype); |
10702 | } |
10703 | |
10704 | tree vec_mask = NULL_TREE; |
10705 | if (memory_access_type == VMAT_LOAD_STORE_LANES) |
10706 | { |
10707 | gcc_assert (alignment_support_scheme == dr_aligned |
10708 | || alignment_support_scheme == dr_unaligned_supported); |
10709 | gcc_assert (grouped_load && !slp); |
10710 | |
10711 | unsigned int inside_cost = 0, prologue_cost = 0; |
10712 | /* For costing some adjacent vector loads, we'd like to cost with |
10713 | the total number of them once instead of cost each one by one. */ |
10714 | unsigned int n_adjacent_loads = 0; |
10715 | for (j = 0; j < ncopies; j++) |
10716 | { |
10717 | if (costing_p) |
10718 | { |
10719 | /* An IFN_LOAD_LANES will load all its vector results, |
10720 | regardless of which ones we actually need. Account |
10721 | for the cost of unused results. */ |
10722 | if (first_stmt_info == stmt_info) |
10723 | { |
10724 | unsigned int gaps = DR_GROUP_SIZE (first_stmt_info); |
10725 | stmt_vec_info next_stmt_info = first_stmt_info; |
10726 | do |
10727 | { |
10728 | gaps -= 1; |
10729 | next_stmt_info = DR_GROUP_NEXT_ELEMENT (next_stmt_info); |
10730 | } |
10731 | while (next_stmt_info); |
10732 | if (gaps) |
10733 | { |
10734 | if (dump_enabled_p ()) |
10735 | dump_printf_loc (MSG_NOTE, vect_location, |
10736 | "vect_model_load_cost: %d " |
10737 | "unused vectors.\n" , |
10738 | gaps); |
10739 | vect_get_load_cost (vinfo, stmt_info, ncopies: gaps, |
10740 | alignment_support_scheme, |
10741 | misalignment, add_realign_cost: false, inside_cost: &inside_cost, |
10742 | prologue_cost: &prologue_cost, prologue_cost_vec: cost_vec, body_cost_vec: cost_vec, |
10743 | record_prologue_costs: true); |
10744 | } |
10745 | } |
10746 | n_adjacent_loads++; |
10747 | continue; |
10748 | } |
10749 | |
10750 | /* 1. Create the vector or array pointer update chain. */ |
10751 | if (j == 0) |
10752 | dataref_ptr |
10753 | = vect_create_data_ref_ptr (vinfo, first_stmt_info, aggr_type, |
10754 | at_loop, offset, &dummy, gsi, |
10755 | &ptr_incr, false, bump); |
10756 | else |
10757 | { |
10758 | gcc_assert (!LOOP_VINFO_USING_SELECT_VL_P (loop_vinfo)); |
10759 | dataref_ptr = bump_vector_ptr (vinfo, dataref_ptr, ptr_incr, gsi, |
10760 | stmt_info, bump); |
10761 | } |
10762 | if (mask) |
10763 | vec_mask = vec_masks[j]; |
10764 | |
10765 | tree vec_array = create_vector_array (elem_type: vectype, nelems: vec_num); |
10766 | |
10767 | tree final_mask = NULL_TREE; |
10768 | tree final_len = NULL_TREE; |
10769 | tree bias = NULL_TREE; |
10770 | if (loop_masks) |
10771 | final_mask = vect_get_loop_mask (loop_vinfo, gsi, loop_masks, |
10772 | ncopies, vectype, j); |
10773 | if (vec_mask) |
10774 | final_mask = prepare_vec_mask (loop_vinfo, mask_type: mask_vectype, loop_mask: final_mask, |
10775 | vec_mask, gsi); |
10776 | |
10777 | if (lanes_ifn == IFN_MASK_LEN_LOAD_LANES) |
10778 | { |
10779 | if (loop_lens) |
10780 | final_len = vect_get_loop_len (loop_vinfo, gsi, loop_lens, |
10781 | ncopies, vectype, j, 1); |
10782 | else |
10783 | final_len = size_int (TYPE_VECTOR_SUBPARTS (vectype)); |
10784 | signed char biasval |
10785 | = LOOP_VINFO_PARTIAL_LOAD_STORE_BIAS (loop_vinfo); |
10786 | bias = build_int_cst (intQI_type_node, biasval); |
10787 | if (!final_mask) |
10788 | { |
10789 | mask_vectype = truth_type_for (vectype); |
10790 | final_mask = build_minus_one_cst (mask_vectype); |
10791 | } |
10792 | } |
10793 | |
10794 | gcall *call; |
10795 | if (final_len && final_mask) |
10796 | { |
10797 | /* Emit: |
10798 | VEC_ARRAY = MASK_LEN_LOAD_LANES (DATAREF_PTR, ALIAS_PTR, |
10799 | VEC_MASK, LEN, BIAS). */ |
10800 | unsigned int align = TYPE_ALIGN (TREE_TYPE (vectype)); |
10801 | tree alias_ptr = build_int_cst (ref_type, align); |
10802 | call = gimple_build_call_internal (IFN_MASK_LEN_LOAD_LANES, 5, |
10803 | dataref_ptr, alias_ptr, |
10804 | final_mask, final_len, bias); |
10805 | } |
10806 | else if (final_mask) |
10807 | { |
10808 | /* Emit: |
10809 | VEC_ARRAY = MASK_LOAD_LANES (DATAREF_PTR, ALIAS_PTR, |
10810 | VEC_MASK). */ |
10811 | unsigned int align = TYPE_ALIGN (TREE_TYPE (vectype)); |
10812 | tree alias_ptr = build_int_cst (ref_type, align); |
10813 | call = gimple_build_call_internal (IFN_MASK_LOAD_LANES, 3, |
10814 | dataref_ptr, alias_ptr, |
10815 | final_mask); |
10816 | } |
10817 | else |
10818 | { |
10819 | /* Emit: |
10820 | VEC_ARRAY = LOAD_LANES (MEM_REF[...all elements...]). */ |
10821 | data_ref = create_array_ref (type: aggr_type, ptr: dataref_ptr, alias_ptr_type: ref_type); |
10822 | call = gimple_build_call_internal (IFN_LOAD_LANES, 1, data_ref); |
10823 | } |
10824 | gimple_call_set_lhs (gs: call, lhs: vec_array); |
10825 | gimple_call_set_nothrow (s: call, nothrow_p: true); |
10826 | vect_finish_stmt_generation (vinfo, stmt_info, vec_stmt: call, gsi); |
10827 | |
10828 | dr_chain.create (nelems: vec_num); |
10829 | /* Extract each vector into an SSA_NAME. */ |
10830 | for (i = 0; i < vec_num; i++) |
10831 | { |
10832 | new_temp = read_vector_array (vinfo, stmt_info, gsi, scalar_dest, |
10833 | array: vec_array, n: i); |
10834 | dr_chain.quick_push (obj: new_temp); |
10835 | } |
10836 | |
10837 | /* Record the mapping between SSA_NAMEs and statements. */ |
10838 | vect_record_grouped_load_vectors (vinfo, stmt_info, dr_chain); |
10839 | |
10840 | /* Record that VEC_ARRAY is now dead. */ |
10841 | vect_clobber_variable (vinfo, stmt_info, gsi, var: vec_array); |
10842 | |
10843 | dr_chain.release (); |
10844 | |
10845 | *vec_stmt = STMT_VINFO_VEC_STMTS (stmt_info)[0]; |
10846 | } |
10847 | |
10848 | if (costing_p) |
10849 | { |
10850 | if (n_adjacent_loads > 0) |
10851 | vect_get_load_cost (vinfo, stmt_info, ncopies: n_adjacent_loads, |
10852 | alignment_support_scheme, misalignment, add_realign_cost: false, |
10853 | inside_cost: &inside_cost, prologue_cost: &prologue_cost, prologue_cost_vec: cost_vec, |
10854 | body_cost_vec: cost_vec, record_prologue_costs: true); |
10855 | if (dump_enabled_p ()) |
10856 | dump_printf_loc (MSG_NOTE, vect_location, |
10857 | "vect_model_load_cost: inside_cost = %u, " |
10858 | "prologue_cost = %u .\n" , |
10859 | inside_cost, prologue_cost); |
10860 | } |
10861 | |
10862 | return true; |
10863 | } |
10864 | |
10865 | if (memory_access_type == VMAT_GATHER_SCATTER) |
10866 | { |
10867 | gcc_assert (alignment_support_scheme == dr_aligned |
10868 | || alignment_support_scheme == dr_unaligned_supported); |
10869 | gcc_assert (!grouped_load && !slp_perm); |
10870 | |
10871 | unsigned int inside_cost = 0, prologue_cost = 0; |
10872 | for (j = 0; j < ncopies; j++) |
10873 | { |
10874 | /* 1. Create the vector or array pointer update chain. */ |
10875 | if (j == 0 && !costing_p) |
10876 | { |
10877 | if (STMT_VINFO_GATHER_SCATTER_P (stmt_info)) |
10878 | vect_get_gather_scatter_ops (loop_vinfo, loop, stmt_info, |
10879 | slp_node, gs_info: &gs_info, dataref_ptr: &dataref_ptr, |
10880 | vec_offset: &vec_offsets); |
10881 | else |
10882 | dataref_ptr |
10883 | = vect_create_data_ref_ptr (vinfo, first_stmt_info, aggr_type, |
10884 | at_loop, offset, &dummy, gsi, |
10885 | &ptr_incr, false, bump); |
10886 | } |
10887 | else if (!costing_p) |
10888 | { |
10889 | gcc_assert (!LOOP_VINFO_USING_SELECT_VL_P (loop_vinfo)); |
10890 | if (!STMT_VINFO_GATHER_SCATTER_P (stmt_info)) |
10891 | dataref_ptr = bump_vector_ptr (vinfo, dataref_ptr, ptr_incr, |
10892 | gsi, stmt_info, bump); |
10893 | } |
10894 | |
10895 | gimple *new_stmt = NULL; |
10896 | for (i = 0; i < vec_num; i++) |
10897 | { |
10898 | tree final_mask = NULL_TREE; |
10899 | tree final_len = NULL_TREE; |
10900 | tree bias = NULL_TREE; |
10901 | if (!costing_p) |
10902 | { |
10903 | if (mask) |
10904 | vec_mask = vec_masks[vec_num * j + i]; |
10905 | if (loop_masks) |
10906 | final_mask |
10907 | = vect_get_loop_mask (loop_vinfo, gsi, loop_masks, |
10908 | vec_num * ncopies, vectype, |
10909 | vec_num * j + i); |
10910 | if (vec_mask) |
10911 | final_mask = prepare_vec_mask (loop_vinfo, mask_type: mask_vectype, |
10912 | loop_mask: final_mask, vec_mask, gsi); |
10913 | |
10914 | if (i > 0 && !STMT_VINFO_GATHER_SCATTER_P (stmt_info)) |
10915 | dataref_ptr = bump_vector_ptr (vinfo, dataref_ptr, ptr_incr, |
10916 | gsi, stmt_info, bump); |
10917 | } |
10918 | |
10919 | /* 2. Create the vector-load in the loop. */ |
10920 | unsigned HOST_WIDE_INT align; |
10921 | if (gs_info.ifn != IFN_LAST) |
10922 | { |
10923 | if (costing_p) |
10924 | { |
10925 | unsigned int cnunits = vect_nunits_for_cost (vec_type: vectype); |
10926 | inside_cost |
10927 | = record_stmt_cost (body_cost_vec: cost_vec, count: cnunits, kind: scalar_load, |
10928 | stmt_info, misalign: 0, where: vect_body); |
10929 | continue; |
10930 | } |
10931 | if (STMT_VINFO_GATHER_SCATTER_P (stmt_info)) |
10932 | vec_offset = vec_offsets[vec_num * j + i]; |
10933 | tree zero = build_zero_cst (vectype); |
10934 | tree scale = size_int (gs_info.scale); |
10935 | |
10936 | if (gs_info.ifn == IFN_MASK_LEN_GATHER_LOAD) |
10937 | { |
10938 | if (loop_lens) |
10939 | final_len |
10940 | = vect_get_loop_len (loop_vinfo, gsi, loop_lens, |
10941 | vec_num * ncopies, vectype, |
10942 | vec_num * j + i, 1); |
10943 | else |
10944 | final_len |
10945 | = build_int_cst (sizetype, |
10946 | TYPE_VECTOR_SUBPARTS (node: vectype)); |
10947 | signed char biasval |
10948 | = LOOP_VINFO_PARTIAL_LOAD_STORE_BIAS (loop_vinfo); |
10949 | bias = build_int_cst (intQI_type_node, biasval); |
10950 | if (!final_mask) |
10951 | { |
10952 | mask_vectype = truth_type_for (vectype); |
10953 | final_mask = build_minus_one_cst (mask_vectype); |
10954 | } |
10955 | } |
10956 | |
10957 | gcall *call; |
10958 | if (final_len && final_mask) |
10959 | call |
10960 | = gimple_build_call_internal (IFN_MASK_LEN_GATHER_LOAD, 7, |
10961 | dataref_ptr, vec_offset, |
10962 | scale, zero, final_mask, |
10963 | final_len, bias); |
10964 | else if (final_mask) |
10965 | call = gimple_build_call_internal (IFN_MASK_GATHER_LOAD, 5, |
10966 | dataref_ptr, vec_offset, |
10967 | scale, zero, final_mask); |
10968 | else |
10969 | call = gimple_build_call_internal (IFN_GATHER_LOAD, 4, |
10970 | dataref_ptr, vec_offset, |
10971 | scale, zero); |
10972 | gimple_call_set_nothrow (s: call, nothrow_p: true); |
10973 | new_stmt = call; |
10974 | data_ref = NULL_TREE; |
10975 | } |
10976 | else if (gs_info.decl) |
10977 | { |
10978 | /* The builtin decls path for gather is legacy, x86 only. */ |
10979 | gcc_assert (!final_len && nunits.is_constant ()); |
10980 | if (costing_p) |
10981 | { |
10982 | unsigned int cnunits = vect_nunits_for_cost (vec_type: vectype); |
10983 | inside_cost |
10984 | = record_stmt_cost (body_cost_vec: cost_vec, count: cnunits, kind: scalar_load, |
10985 | stmt_info, misalign: 0, where: vect_body); |
10986 | continue; |
10987 | } |
10988 | poly_uint64 offset_nunits |
10989 | = TYPE_VECTOR_SUBPARTS (node: gs_info.offset_vectype); |
10990 | if (known_eq (nunits, offset_nunits)) |
10991 | { |
10992 | new_stmt = vect_build_one_gather_load_call |
10993 | (vinfo, stmt_info, gsi, gs_info: &gs_info, |
10994 | ptr: dataref_ptr, offset: vec_offsets[vec_num * j + i], |
10995 | mask: final_mask); |
10996 | data_ref = NULL_TREE; |
10997 | } |
10998 | else if (known_eq (nunits, offset_nunits * 2)) |
10999 | { |
11000 | /* We have a offset vector with half the number of |
11001 | lanes but the builtins will produce full vectype |
11002 | data with just the lower lanes filled. */ |
11003 | new_stmt = vect_build_one_gather_load_call |
11004 | (vinfo, stmt_info, gsi, gs_info: &gs_info, |
11005 | ptr: dataref_ptr, offset: vec_offsets[2 * vec_num * j + 2 * i], |
11006 | mask: final_mask); |
11007 | tree low = make_ssa_name (var: vectype); |
11008 | gimple_set_lhs (new_stmt, low); |
11009 | vect_finish_stmt_generation (vinfo, stmt_info, |
11010 | vec_stmt: new_stmt, gsi); |
11011 | |
11012 | /* now put upper half of final_mask in final_mask low. */ |
11013 | if (final_mask |
11014 | && !SCALAR_INT_MODE_P |
11015 | (TYPE_MODE (TREE_TYPE (final_mask)))) |
11016 | { |
11017 | int count = nunits.to_constant (); |
11018 | vec_perm_builder sel (count, count, 1); |
11019 | sel.quick_grow (len: count); |
11020 | for (int i = 0; i < count; ++i) |
11021 | sel[i] = i | (count / 2); |
11022 | vec_perm_indices indices (sel, 2, count); |
11023 | tree perm_mask = vect_gen_perm_mask_checked |
11024 | (TREE_TYPE (final_mask), sel: indices); |
11025 | new_stmt = gimple_build_assign (NULL_TREE, |
11026 | VEC_PERM_EXPR, |
11027 | final_mask, |
11028 | final_mask, |
11029 | perm_mask); |
11030 | final_mask = make_ssa_name (TREE_TYPE (final_mask)); |
11031 | gimple_set_lhs (new_stmt, final_mask); |
11032 | vect_finish_stmt_generation (vinfo, stmt_info, |
11033 | vec_stmt: new_stmt, gsi); |
11034 | } |
11035 | else if (final_mask) |
11036 | { |
11037 | new_stmt = gimple_build_assign (NULL_TREE, |
11038 | VEC_UNPACK_HI_EXPR, |
11039 | final_mask); |
11040 | final_mask = make_ssa_name |
11041 | (var: truth_type_for (gs_info.offset_vectype)); |
11042 | gimple_set_lhs (new_stmt, final_mask); |
11043 | vect_finish_stmt_generation (vinfo, stmt_info, |
11044 | vec_stmt: new_stmt, gsi); |
11045 | } |
11046 | |
11047 | new_stmt = vect_build_one_gather_load_call |
11048 | (vinfo, stmt_info, gsi, gs_info: &gs_info, |
11049 | ptr: dataref_ptr, |
11050 | offset: vec_offsets[2 * vec_num * j + 2 * i + 1], |
11051 | mask: final_mask); |
11052 | tree high = make_ssa_name (var: vectype); |
11053 | gimple_set_lhs (new_stmt, high); |
11054 | vect_finish_stmt_generation (vinfo, stmt_info, |
11055 | vec_stmt: new_stmt, gsi); |
11056 | |
11057 | /* compose low + high. */ |
11058 | int count = nunits.to_constant (); |
11059 | vec_perm_builder sel (count, count, 1); |
11060 | sel.quick_grow (len: count); |
11061 | for (int i = 0; i < count; ++i) |
11062 | sel[i] = i < count / 2 ? i : i + count / 2; |
11063 | vec_perm_indices indices (sel, 2, count); |
11064 | tree perm_mask |
11065 | = vect_gen_perm_mask_checked (vectype, sel: indices); |
11066 | new_stmt = gimple_build_assign (NULL_TREE, |
11067 | VEC_PERM_EXPR, |
11068 | low, high, perm_mask); |
11069 | data_ref = NULL_TREE; |
11070 | } |
11071 | else if (known_eq (nunits * 2, offset_nunits)) |
11072 | { |
11073 | /* We have a offset vector with double the number of |
11074 | lanes. Select the low/high part accordingly. */ |
11075 | vec_offset = vec_offsets[(vec_num * j + i) / 2]; |
11076 | if ((vec_num * j + i) & 1) |
11077 | { |
11078 | int count = offset_nunits.to_constant (); |
11079 | vec_perm_builder sel (count, count, 1); |
11080 | sel.quick_grow (len: count); |
11081 | for (int i = 0; i < count; ++i) |
11082 | sel[i] = i | (count / 2); |
11083 | vec_perm_indices indices (sel, 2, count); |
11084 | tree perm_mask = vect_gen_perm_mask_checked |
11085 | (TREE_TYPE (vec_offset), sel: indices); |
11086 | new_stmt = gimple_build_assign (NULL_TREE, |
11087 | VEC_PERM_EXPR, |
11088 | vec_offset, |
11089 | vec_offset, |
11090 | perm_mask); |
11091 | vec_offset = make_ssa_name (TREE_TYPE (vec_offset)); |
11092 | gimple_set_lhs (new_stmt, vec_offset); |
11093 | vect_finish_stmt_generation (vinfo, stmt_info, |
11094 | vec_stmt: new_stmt, gsi); |
11095 | } |
11096 | new_stmt = vect_build_one_gather_load_call |
11097 | (vinfo, stmt_info, gsi, gs_info: &gs_info, |
11098 | ptr: dataref_ptr, offset: vec_offset, mask: final_mask); |
11099 | data_ref = NULL_TREE; |
11100 | } |
11101 | else |
11102 | gcc_unreachable (); |
11103 | } |
11104 | else |
11105 | { |
11106 | /* Emulated gather-scatter. */ |
11107 | gcc_assert (!final_mask); |
11108 | unsigned HOST_WIDE_INT const_nunits = nunits.to_constant (); |
11109 | if (costing_p) |
11110 | { |
11111 | /* For emulated gathers N offset vector element |
11112 | offset add is consumed by the load). */ |
11113 | inside_cost = record_stmt_cost (body_cost_vec: cost_vec, count: const_nunits, |
11114 | kind: vec_to_scalar, stmt_info, |
11115 | misalign: 0, where: vect_body); |
11116 | /* N scalar loads plus gathering them into a |
11117 | vector. */ |
11118 | inside_cost |
11119 | = record_stmt_cost (body_cost_vec: cost_vec, count: const_nunits, kind: scalar_load, |
11120 | stmt_info, misalign: 0, where: vect_body); |
11121 | inside_cost |
11122 | = record_stmt_cost (body_cost_vec: cost_vec, count: 1, kind: vec_construct, |
11123 | stmt_info, misalign: 0, where: vect_body); |
11124 | continue; |
11125 | } |
11126 | unsigned HOST_WIDE_INT const_offset_nunits |
11127 | = TYPE_VECTOR_SUBPARTS (node: gs_info.offset_vectype) |
11128 | .to_constant (); |
11129 | vec<constructor_elt, va_gc> *ctor_elts; |
11130 | vec_alloc (v&: ctor_elts, nelems: const_nunits); |
11131 | gimple_seq stmts = NULL; |
11132 | /* We support offset vectors with more elements |
11133 | than the data vector for now. */ |
11134 | unsigned HOST_WIDE_INT factor |
11135 | = const_offset_nunits / const_nunits; |
11136 | vec_offset = vec_offsets[(vec_num * j + i) / factor]; |
11137 | unsigned elt_offset = (j % factor) * const_nunits; |
11138 | tree idx_type = TREE_TYPE (TREE_TYPE (vec_offset)); |
11139 | tree scale = size_int (gs_info.scale); |
11140 | align = get_object_alignment (DR_REF (first_dr_info->dr)); |
11141 | tree ltype = build_aligned_type (TREE_TYPE (vectype), align); |
11142 | for (unsigned k = 0; k < const_nunits; ++k) |
11143 | { |
11144 | tree boff = size_binop (MULT_EXPR, TYPE_SIZE (idx_type), |
11145 | bitsize_int (k + elt_offset)); |
11146 | tree idx |
11147 | = gimple_build (seq: &stmts, code: BIT_FIELD_REF, type: idx_type, |
11148 | ops: vec_offset, TYPE_SIZE (idx_type), ops: boff); |
11149 | idx = gimple_convert (seq: &stmts, sizetype, op: idx); |
11150 | idx = gimple_build (seq: &stmts, code: MULT_EXPR, sizetype, ops: idx, |
11151 | ops: scale); |
11152 | tree ptr = gimple_build (seq: &stmts, code: PLUS_EXPR, |
11153 | TREE_TYPE (dataref_ptr), |
11154 | ops: dataref_ptr, ops: idx); |
11155 | ptr = gimple_convert (seq: &stmts, ptr_type_node, op: ptr); |
11156 | tree elt = make_ssa_name (TREE_TYPE (vectype)); |
11157 | tree ref = build2 (MEM_REF, ltype, ptr, |
11158 | build_int_cst (ref_type, 0)); |
11159 | new_stmt = gimple_build_assign (elt, ref); |
11160 | gimple_set_vuse (g: new_stmt, vuse: gimple_vuse (g: gsi_stmt (i: *gsi))); |
11161 | gimple_seq_add_stmt (&stmts, new_stmt); |
11162 | CONSTRUCTOR_APPEND_ELT (ctor_elts, NULL_TREE, elt); |
11163 | } |
11164 | gsi_insert_seq_before (gsi, stmts, GSI_SAME_STMT); |
11165 | new_stmt = gimple_build_assign ( |
11166 | NULL_TREE, build_constructor (vectype, ctor_elts)); |
11167 | data_ref = NULL_TREE; |
11168 | } |
11169 | |
11170 | vec_dest = vect_create_destination_var (scalar_dest, vectype); |
11171 | /* DATA_REF is null if we've already built the statement. */ |
11172 | if (data_ref) |
11173 | { |
11174 | vect_copy_ref_info (data_ref, DR_REF (first_dr_info->dr)); |
11175 | new_stmt = gimple_build_assign (vec_dest, data_ref); |
11176 | } |
11177 | new_temp = make_ssa_name (var: vec_dest, stmt: new_stmt); |
11178 | gimple_set_lhs (new_stmt, new_temp); |
11179 | vect_finish_stmt_generation (vinfo, stmt_info, vec_stmt: new_stmt, gsi); |
11180 | |
11181 | /* Store vector loads in the corresponding SLP_NODE. */ |
11182 | if (slp) |
11183 | slp_node->push_vec_def (def: new_stmt); |
11184 | } |
11185 | |
11186 | if (!slp && !costing_p) |
11187 | STMT_VINFO_VEC_STMTS (stmt_info).safe_push (obj: new_stmt); |
11188 | } |
11189 | |
11190 | if (!slp && !costing_p) |
11191 | *vec_stmt = STMT_VINFO_VEC_STMTS (stmt_info)[0]; |
11192 | |
11193 | if (costing_p && dump_enabled_p ()) |
11194 | dump_printf_loc (MSG_NOTE, vect_location, |
11195 | "vect_model_load_cost: inside_cost = %u, " |
11196 | "prologue_cost = %u .\n" , |
11197 | inside_cost, prologue_cost); |
11198 | return true; |
11199 | } |
11200 | |
11201 | poly_uint64 group_elt = 0; |
11202 | unsigned int inside_cost = 0, prologue_cost = 0; |
11203 | /* For costing some adjacent vector loads, we'd like to cost with |
11204 | the total number of them once instead of cost each one by one. */ |
11205 | unsigned int n_adjacent_loads = 0; |
11206 | for (j = 0; j < ncopies; j++) |
11207 | { |
11208 | /* 1. Create the vector or array pointer update chain. */ |
11209 | if (j == 0 && !costing_p) |
11210 | { |
11211 | bool simd_lane_access_p |
11212 | = STMT_VINFO_SIMD_LANE_ACCESS_P (stmt_info) != 0; |
11213 | if (simd_lane_access_p |
11214 | && TREE_CODE (DR_BASE_ADDRESS (first_dr_info->dr)) == ADDR_EXPR |
11215 | && VAR_P (TREE_OPERAND (DR_BASE_ADDRESS (first_dr_info->dr), 0)) |
11216 | && integer_zerop (get_dr_vinfo_offset (vinfo, dr_info: first_dr_info)) |
11217 | && integer_zerop (DR_INIT (first_dr_info->dr)) |
11218 | && alias_sets_conflict_p (get_alias_set (aggr_type), |
11219 | get_alias_set (TREE_TYPE (ref_type))) |
11220 | && (alignment_support_scheme == dr_aligned |
11221 | || alignment_support_scheme == dr_unaligned_supported)) |
11222 | { |
11223 | dataref_ptr = unshare_expr (DR_BASE_ADDRESS (first_dr_info->dr)); |
11224 | dataref_offset = build_int_cst (ref_type, 0); |
11225 | } |
11226 | else if (diff_first_stmt_info) |
11227 | { |
11228 | dataref_ptr |
11229 | = vect_create_data_ref_ptr (vinfo, first_stmt_info_for_drptr, |
11230 | aggr_type, at_loop, offset, &dummy, |
11231 | gsi, &ptr_incr, simd_lane_access_p, |
11232 | bump); |
11233 | /* Adjust the pointer by the difference to first_stmt. */ |
11234 | data_reference_p ptrdr |
11235 | = STMT_VINFO_DATA_REF (first_stmt_info_for_drptr); |
11236 | tree diff |
11237 | = fold_convert (sizetype, |
11238 | size_binop (MINUS_EXPR, |
11239 | DR_INIT (first_dr_info->dr), |
11240 | DR_INIT (ptrdr))); |
11241 | dataref_ptr = bump_vector_ptr (vinfo, dataref_ptr, ptr_incr, gsi, |
11242 | stmt_info, diff); |
11243 | if (alignment_support_scheme == dr_explicit_realign) |
11244 | { |
11245 | msq = vect_setup_realignment (vinfo, |
11246 | first_stmt_info_for_drptr, gsi, |
11247 | &realignment_token, |
11248 | alignment_support_scheme, |
11249 | dataref_ptr, &at_loop); |
11250 | gcc_assert (!compute_in_loop); |
11251 | } |
11252 | } |
11253 | else |
11254 | dataref_ptr |
11255 | = vect_create_data_ref_ptr (vinfo, first_stmt_info, aggr_type, |
11256 | at_loop, |
11257 | offset, &dummy, gsi, &ptr_incr, |
11258 | simd_lane_access_p, bump); |
11259 | } |
11260 | else if (!costing_p) |
11261 | { |
11262 | gcc_assert (!LOOP_VINFO_USING_SELECT_VL_P (loop_vinfo)); |
11263 | if (dataref_offset) |
11264 | dataref_offset = int_const_binop (PLUS_EXPR, dataref_offset, |
11265 | bump); |
11266 | else |
11267 | dataref_ptr = bump_vector_ptr (vinfo, dataref_ptr, ptr_incr, gsi, |
11268 | stmt_info, bump); |
11269 | } |
11270 | |
11271 | if (grouped_load || slp_perm) |
11272 | dr_chain.create (nelems: vec_num); |
11273 | |
11274 | gimple *new_stmt = NULL; |
11275 | for (i = 0; i < vec_num; i++) |
11276 | { |
11277 | tree final_mask = NULL_TREE; |
11278 | tree final_len = NULL_TREE; |
11279 | tree bias = NULL_TREE; |
11280 | if (!costing_p) |
11281 | { |
11282 | if (mask) |
11283 | vec_mask = vec_masks[vec_num * j + i]; |
11284 | if (loop_masks) |
11285 | final_mask = vect_get_loop_mask (loop_vinfo, gsi, loop_masks, |
11286 | vec_num * ncopies, vectype, |
11287 | vec_num * j + i); |
11288 | if (vec_mask) |
11289 | final_mask = prepare_vec_mask (loop_vinfo, mask_type: mask_vectype, |
11290 | loop_mask: final_mask, vec_mask, gsi); |
11291 | |
11292 | if (i > 0) |
11293 | dataref_ptr = bump_vector_ptr (vinfo, dataref_ptr, ptr_incr, |
11294 | gsi, stmt_info, bump); |
11295 | } |
11296 | |
11297 | /* 2. Create the vector-load in the loop. */ |
11298 | switch (alignment_support_scheme) |
11299 | { |
11300 | case dr_aligned: |
11301 | case dr_unaligned_supported: |
11302 | { |
11303 | if (costing_p) |
11304 | break; |
11305 | |
11306 | unsigned int misalign; |
11307 | unsigned HOST_WIDE_INT align; |
11308 | align = known_alignment (DR_TARGET_ALIGNMENT (first_dr_info)); |
11309 | if (alignment_support_scheme == dr_aligned) |
11310 | misalign = 0; |
11311 | else if (misalignment == DR_MISALIGNMENT_UNKNOWN) |
11312 | { |
11313 | align |
11314 | = dr_alignment (vect_dr_behavior (vinfo, dr_info: first_dr_info)); |
11315 | misalign = 0; |
11316 | } |
11317 | else |
11318 | misalign = misalignment; |
11319 | if (dataref_offset == NULL_TREE |
11320 | && TREE_CODE (dataref_ptr) == SSA_NAME) |
11321 | set_ptr_info_alignment (get_ptr_info (dataref_ptr), align, |
11322 | misalign); |
11323 | align = least_bit_hwi (x: misalign | align); |
11324 | |
11325 | /* Compute IFN when LOOP_LENS or final_mask valid. */ |
11326 | machine_mode vmode = TYPE_MODE (vectype); |
11327 | machine_mode new_vmode = vmode; |
11328 | internal_fn partial_ifn = IFN_LAST; |
11329 | if (loop_lens) |
11330 | { |
11331 | opt_machine_mode new_ovmode |
11332 | = get_len_load_store_mode (vmode, true, &partial_ifn); |
11333 | new_vmode = new_ovmode.require (); |
11334 | unsigned factor |
11335 | = (new_ovmode == vmode) ? 1 : GET_MODE_UNIT_SIZE (vmode); |
11336 | final_len = vect_get_loop_len (loop_vinfo, gsi, loop_lens, |
11337 | vec_num * ncopies, vectype, |
11338 | vec_num * j + i, factor); |
11339 | } |
11340 | else if (final_mask) |
11341 | { |
11342 | if (!can_vec_mask_load_store_p ( |
11343 | vmode, TYPE_MODE (TREE_TYPE (final_mask)), true, |
11344 | &partial_ifn)) |
11345 | gcc_unreachable (); |
11346 | } |
11347 | |
11348 | if (partial_ifn == IFN_MASK_LEN_LOAD) |
11349 | { |
11350 | if (!final_len) |
11351 | { |
11352 | /* Pass VF value to 'len' argument of |
11353 | MASK_LEN_LOAD if LOOP_LENS is invalid. */ |
11354 | final_len = size_int (TYPE_VECTOR_SUBPARTS (vectype)); |
11355 | } |
11356 | if (!final_mask) |
11357 | { |
11358 | /* Pass all ones value to 'mask' argument of |
11359 | MASK_LEN_LOAD if final_mask is invalid. */ |
11360 | mask_vectype = truth_type_for (vectype); |
11361 | final_mask = build_minus_one_cst (mask_vectype); |
11362 | } |
11363 | } |
11364 | if (final_len) |
11365 | { |
11366 | signed char biasval |
11367 | = LOOP_VINFO_PARTIAL_LOAD_STORE_BIAS (loop_vinfo); |
11368 | |
11369 | bias = build_int_cst (intQI_type_node, biasval); |
11370 | } |
11371 | |
11372 | if (final_len) |
11373 | { |
11374 | tree ptr = build_int_cst (ref_type, align * BITS_PER_UNIT); |
11375 | gcall *call; |
11376 | if (partial_ifn == IFN_MASK_LEN_LOAD) |
11377 | call = gimple_build_call_internal (IFN_MASK_LEN_LOAD, 5, |
11378 | dataref_ptr, ptr, |
11379 | final_mask, final_len, |
11380 | bias); |
11381 | else |
11382 | call = gimple_build_call_internal (IFN_LEN_LOAD, 4, |
11383 | dataref_ptr, ptr, |
11384 | final_len, bias); |
11385 | gimple_call_set_nothrow (s: call, nothrow_p: true); |
11386 | new_stmt = call; |
11387 | data_ref = NULL_TREE; |
11388 | |
11389 | /* Need conversion if it's wrapped with VnQI. */ |
11390 | if (vmode != new_vmode) |
11391 | { |
11392 | tree new_vtype = build_vector_type_for_mode ( |
11393 | unsigned_intQI_type_node, new_vmode); |
11394 | tree var |
11395 | = vect_get_new_ssa_name (new_vtype, vect_simple_var); |
11396 | gimple_set_lhs (call, var); |
11397 | vect_finish_stmt_generation (vinfo, stmt_info, vec_stmt: call, |
11398 | gsi); |
11399 | tree op = build1 (VIEW_CONVERT_EXPR, vectype, var); |
11400 | new_stmt = gimple_build_assign (vec_dest, |
11401 | VIEW_CONVERT_EXPR, op); |
11402 | } |
11403 | } |
11404 | else if (final_mask) |
11405 | { |
11406 | tree ptr = build_int_cst (ref_type, align * BITS_PER_UNIT); |
11407 | gcall *call = gimple_build_call_internal (IFN_MASK_LOAD, 3, |
11408 | dataref_ptr, ptr, |
11409 | final_mask); |
11410 | gimple_call_set_nothrow (s: call, nothrow_p: true); |
11411 | new_stmt = call; |
11412 | data_ref = NULL_TREE; |
11413 | } |
11414 | else |
11415 | { |
11416 | tree ltype = vectype; |
11417 | tree new_vtype = NULL_TREE; |
11418 | unsigned HOST_WIDE_INT gap = DR_GROUP_GAP (first_stmt_info); |
11419 | unsigned int vect_align |
11420 | = vect_known_alignment_in_bytes (dr_info: first_dr_info, vectype); |
11421 | unsigned int scalar_dr_size |
11422 | = vect_get_scalar_dr_size (dr_info: first_dr_info); |
11423 | /* If there's no peeling for gaps but we have a gap |
11424 | with slp loads then load the lower half of the |
11425 | vector only. See get_group_load_store_type for |
11426 | when we apply this optimization. */ |
11427 | if (slp |
11428 | && loop_vinfo |
11429 | && !LOOP_VINFO_PEELING_FOR_GAPS (loop_vinfo) && gap != 0 |
11430 | && known_eq (nunits, (group_size - gap) * 2) |
11431 | && known_eq (nunits, group_size) |
11432 | && gap >= (vect_align / scalar_dr_size)) |
11433 | { |
11434 | tree half_vtype; |
11435 | new_vtype |
11436 | = vector_vector_composition_type (vtype: vectype, nelts: 2, |
11437 | ptype: &half_vtype); |
11438 | if (new_vtype != NULL_TREE) |
11439 | ltype = half_vtype; |
11440 | } |
11441 | tree offset |
11442 | = (dataref_offset ? dataref_offset |
11443 | : build_int_cst (ref_type, 0)); |
11444 | if (ltype != vectype |
11445 | && memory_access_type == VMAT_CONTIGUOUS_REVERSE) |
11446 | { |
11447 | unsigned HOST_WIDE_INT gap_offset |
11448 | = gap * tree_to_uhwi (TYPE_SIZE_UNIT (elem_type)); |
11449 | tree gapcst = build_int_cst (ref_type, gap_offset); |
11450 | offset = size_binop (PLUS_EXPR, offset, gapcst); |
11451 | } |
11452 | data_ref |
11453 | = fold_build2 (MEM_REF, ltype, dataref_ptr, offset); |
11454 | if (alignment_support_scheme == dr_aligned) |
11455 | ; |
11456 | else |
11457 | TREE_TYPE (data_ref) |
11458 | = build_aligned_type (TREE_TYPE (data_ref), |
11459 | align * BITS_PER_UNIT); |
11460 | if (ltype != vectype) |
11461 | { |
11462 | vect_copy_ref_info (data_ref, |
11463 | DR_REF (first_dr_info->dr)); |
11464 | tree tem = make_ssa_name (var: ltype); |
11465 | new_stmt = gimple_build_assign (tem, data_ref); |
11466 | vect_finish_stmt_generation (vinfo, stmt_info, vec_stmt: new_stmt, |
11467 | gsi); |
11468 | data_ref = NULL; |
11469 | vec<constructor_elt, va_gc> *v; |
11470 | vec_alloc (v, nelems: 2); |
11471 | if (memory_access_type == VMAT_CONTIGUOUS_REVERSE) |
11472 | { |
11473 | CONSTRUCTOR_APPEND_ELT (v, NULL_TREE, |
11474 | build_zero_cst (ltype)); |
11475 | CONSTRUCTOR_APPEND_ELT (v, NULL_TREE, tem); |
11476 | } |
11477 | else |
11478 | { |
11479 | CONSTRUCTOR_APPEND_ELT (v, NULL_TREE, tem); |
11480 | CONSTRUCTOR_APPEND_ELT (v, NULL_TREE, |
11481 | build_zero_cst (ltype)); |
11482 | } |
11483 | gcc_assert (new_vtype != NULL_TREE); |
11484 | if (new_vtype == vectype) |
11485 | new_stmt = gimple_build_assign ( |
11486 | vec_dest, build_constructor (vectype, v)); |
11487 | else |
11488 | { |
11489 | tree new_vname = make_ssa_name (var: new_vtype); |
11490 | new_stmt = gimple_build_assign ( |
11491 | new_vname, build_constructor (new_vtype, v)); |
11492 | vect_finish_stmt_generation (vinfo, stmt_info, |
11493 | vec_stmt: new_stmt, gsi); |
11494 | new_stmt = gimple_build_assign ( |
11495 | vec_dest, |
11496 | build1 (VIEW_CONVERT_EXPR, vectype, new_vname)); |
11497 | } |
11498 | } |
11499 | } |
11500 | break; |
11501 | } |
11502 | case dr_explicit_realign: |
11503 | { |
11504 | if (costing_p) |
11505 | break; |
11506 | tree ptr, bump; |
11507 | |
11508 | tree vs = size_int (TYPE_VECTOR_SUBPARTS (vectype)); |
11509 | |
11510 | if (compute_in_loop) |
11511 | msq = vect_setup_realignment (vinfo, first_stmt_info, gsi, |
11512 | &realignment_token, |
11513 | dr_explicit_realign, |
11514 | dataref_ptr, NULL); |
11515 | |
11516 | if (TREE_CODE (dataref_ptr) == SSA_NAME) |
11517 | ptr = copy_ssa_name (var: dataref_ptr); |
11518 | else |
11519 | ptr = make_ssa_name (TREE_TYPE (dataref_ptr)); |
11520 | // For explicit realign the target alignment should be |
11521 | // known at compile time. |
11522 | unsigned HOST_WIDE_INT align |
11523 | = DR_TARGET_ALIGNMENT (first_dr_info).to_constant (); |
11524 | new_stmt = gimple_build_assign ( |
11525 | ptr, BIT_AND_EXPR, dataref_ptr, |
11526 | build_int_cst (TREE_TYPE (dataref_ptr), |
11527 | -(HOST_WIDE_INT) align)); |
11528 | vect_finish_stmt_generation (vinfo, stmt_info, vec_stmt: new_stmt, gsi); |
11529 | data_ref |
11530 | = build2 (MEM_REF, vectype, ptr, build_int_cst (ref_type, 0)); |
11531 | vect_copy_ref_info (data_ref, DR_REF (first_dr_info->dr)); |
11532 | vec_dest = vect_create_destination_var (scalar_dest, vectype); |
11533 | new_stmt = gimple_build_assign (vec_dest, data_ref); |
11534 | new_temp = make_ssa_name (var: vec_dest, stmt: new_stmt); |
11535 | gimple_assign_set_lhs (gs: new_stmt, lhs: new_temp); |
11536 | gimple_move_vops (new_stmt, stmt_info->stmt); |
11537 | vect_finish_stmt_generation (vinfo, stmt_info, vec_stmt: new_stmt, gsi); |
11538 | msq = new_temp; |
11539 | |
11540 | bump = size_binop (MULT_EXPR, vs, TYPE_SIZE_UNIT (elem_type)); |
11541 | bump = size_binop (MINUS_EXPR, bump, size_one_node); |
11542 | ptr = bump_vector_ptr (vinfo, dataref_ptr, NULL, gsi, stmt_info, |
11543 | bump); |
11544 | new_stmt = gimple_build_assign ( |
11545 | NULL_TREE, BIT_AND_EXPR, ptr, |
11546 | build_int_cst (TREE_TYPE (ptr), -(HOST_WIDE_INT) align)); |
11547 | if (TREE_CODE (ptr) == SSA_NAME) |
11548 | ptr = copy_ssa_name (var: ptr, stmt: new_stmt); |
11549 | else |
11550 | ptr = make_ssa_name (TREE_TYPE (ptr), stmt: new_stmt); |
11551 | gimple_assign_set_lhs (gs: new_stmt, lhs: ptr); |
11552 | vect_finish_stmt_generation (vinfo, stmt_info, vec_stmt: new_stmt, gsi); |
11553 | data_ref |
11554 | = build2 (MEM_REF, vectype, ptr, build_int_cst (ref_type, 0)); |
11555 | break; |
11556 | } |
11557 | case dr_explicit_realign_optimized: |
11558 | { |
11559 | if (costing_p) |
11560 | break; |
11561 | if (TREE_CODE (dataref_ptr) == SSA_NAME) |
11562 | new_temp = copy_ssa_name (var: dataref_ptr); |
11563 | else |
11564 | new_temp = make_ssa_name (TREE_TYPE (dataref_ptr)); |
11565 | // We should only be doing this if we know the target |
11566 | // alignment at compile time. |
11567 | unsigned HOST_WIDE_INT align |
11568 | = DR_TARGET_ALIGNMENT (first_dr_info).to_constant (); |
11569 | new_stmt = gimple_build_assign ( |
11570 | new_temp, BIT_AND_EXPR, dataref_ptr, |
11571 | build_int_cst (TREE_TYPE (dataref_ptr), |
11572 | -(HOST_WIDE_INT) align)); |
11573 | vect_finish_stmt_generation (vinfo, stmt_info, vec_stmt: new_stmt, gsi); |
11574 | data_ref = build2 (MEM_REF, vectype, new_temp, |
11575 | build_int_cst (ref_type, 0)); |
11576 | break; |
11577 | } |
11578 | default: |
11579 | gcc_unreachable (); |
11580 | } |
11581 | |
11582 | /* One common place to cost the above vect load for different |
11583 | alignment support schemes. */ |
11584 | if (costing_p) |
11585 | { |
11586 | /* For VMAT_CONTIGUOUS_PERMUTE if it's grouped load, we |
11587 | only need to take care of the first stmt, whose |
11588 | stmt_info is first_stmt_info, vec_num iterating on it |
11589 | will cover the cost for the remaining, it's consistent |
11590 | with transforming. For the prologue cost for realign, |
11591 | we only need to count it once for the whole group. */ |
11592 | bool first_stmt_info_p = first_stmt_info == stmt_info; |
11593 | bool add_realign_cost = first_stmt_info_p && i == 0; |
11594 | if (memory_access_type == VMAT_CONTIGUOUS |
11595 | || memory_access_type == VMAT_CONTIGUOUS_REVERSE |
11596 | || (memory_access_type == VMAT_CONTIGUOUS_PERMUTE |
11597 | && (!grouped_load || first_stmt_info_p))) |
11598 | { |
11599 | /* Leave realign cases alone to keep them simple. */ |
11600 | if (alignment_support_scheme == dr_explicit_realign_optimized |
11601 | || alignment_support_scheme == dr_explicit_realign) |
11602 | vect_get_load_cost (vinfo, stmt_info, ncopies: 1, |
11603 | alignment_support_scheme, misalignment, |
11604 | add_realign_cost, inside_cost: &inside_cost, |
11605 | prologue_cost: &prologue_cost, prologue_cost_vec: cost_vec, body_cost_vec: cost_vec, |
11606 | record_prologue_costs: true); |
11607 | else |
11608 | n_adjacent_loads++; |
11609 | } |
11610 | } |
11611 | else |
11612 | { |
11613 | vec_dest = vect_create_destination_var (scalar_dest, vectype); |
11614 | /* DATA_REF is null if we've already built the statement. */ |
11615 | if (data_ref) |
11616 | { |
11617 | vect_copy_ref_info (data_ref, DR_REF (first_dr_info->dr)); |
11618 | new_stmt = gimple_build_assign (vec_dest, data_ref); |
11619 | } |
11620 | new_temp = make_ssa_name (var: vec_dest, stmt: new_stmt); |
11621 | gimple_set_lhs (new_stmt, new_temp); |
11622 | vect_finish_stmt_generation (vinfo, stmt_info, vec_stmt: new_stmt, gsi); |
11623 | } |
11624 | |
11625 | /* 3. Handle explicit realignment if necessary/supported. |
11626 | Create in loop: |
11627 | vec_dest = realign_load (msq, lsq, realignment_token) */ |
11628 | if (!costing_p |
11629 | && (alignment_support_scheme == dr_explicit_realign_optimized |
11630 | || alignment_support_scheme == dr_explicit_realign)) |
11631 | { |
11632 | lsq = gimple_assign_lhs (gs: new_stmt); |
11633 | if (!realignment_token) |
11634 | realignment_token = dataref_ptr; |
11635 | vec_dest = vect_create_destination_var (scalar_dest, vectype); |
11636 | new_stmt = gimple_build_assign (vec_dest, REALIGN_LOAD_EXPR, msq, |
11637 | lsq, realignment_token); |
11638 | new_temp = make_ssa_name (var: vec_dest, stmt: new_stmt); |
11639 | gimple_assign_set_lhs (gs: new_stmt, lhs: new_temp); |
11640 | vect_finish_stmt_generation (vinfo, stmt_info, vec_stmt: new_stmt, gsi); |
11641 | |
11642 | if (alignment_support_scheme == dr_explicit_realign_optimized) |
11643 | { |
11644 | gcc_assert (phi); |
11645 | if (i == vec_num - 1 && j == ncopies - 1) |
11646 | add_phi_arg (phi, lsq, loop_latch_edge (containing_loop), |
11647 | UNKNOWN_LOCATION); |
11648 | msq = lsq; |
11649 | } |
11650 | } |
11651 | |
11652 | if (memory_access_type == VMAT_CONTIGUOUS_REVERSE) |
11653 | { |
11654 | if (costing_p) |
11655 | inside_cost = record_stmt_cost (body_cost_vec: cost_vec, count: 1, kind: vec_perm, |
11656 | stmt_info, misalign: 0, where: vect_body); |
11657 | else |
11658 | { |
11659 | tree perm_mask = perm_mask_for_reverse (vectype); |
11660 | new_temp = permute_vec_elements (vinfo, x: new_temp, y: new_temp, |
11661 | mask_vec: perm_mask, stmt_info, gsi); |
11662 | new_stmt = SSA_NAME_DEF_STMT (new_temp); |
11663 | } |
11664 | } |
11665 | |
11666 | /* Collect vector loads and later create their permutation in |
11667 | vect_transform_grouped_load (). */ |
11668 | if (!costing_p && (grouped_load || slp_perm)) |
11669 | dr_chain.quick_push (obj: new_temp); |
11670 | |
11671 | /* Store vector loads in the corresponding SLP_NODE. */ |
11672 | if (!costing_p && slp && !slp_perm) |
11673 | slp_node->push_vec_def (def: new_stmt); |
11674 | |
11675 | /* With SLP permutation we load the gaps as well, without |
11676 | we need to skip the gaps after we manage to fully load |
11677 | all elements. group_gap_adj is DR_GROUP_SIZE here. */ |
11678 | group_elt += nunits; |
11679 | if (!costing_p |
11680 | && maybe_ne (a: group_gap_adj, b: 0U) |
11681 | && !slp_perm |
11682 | && known_eq (group_elt, group_size - group_gap_adj)) |
11683 | { |
11684 | poly_wide_int bump_val |
11685 | = (wi::to_wide (TYPE_SIZE_UNIT (elem_type)) * group_gap_adj); |
11686 | if (tree_int_cst_sgn (vect_dr_behavior (vinfo, dr_info)->step) |
11687 | == -1) |
11688 | bump_val = -bump_val; |
11689 | tree bump = wide_int_to_tree (sizetype, cst: bump_val); |
11690 | dataref_ptr = bump_vector_ptr (vinfo, dataref_ptr, ptr_incr, gsi, |
11691 | stmt_info, bump); |
11692 | group_elt = 0; |
11693 | } |
11694 | } |
11695 | /* Bump the vector pointer to account for a gap or for excess |
11696 | elements loaded for a permuted SLP load. */ |
11697 | if (!costing_p |
11698 | && maybe_ne (a: group_gap_adj, b: 0U) |
11699 | && slp_perm) |
11700 | { |
11701 | poly_wide_int bump_val |
11702 | = (wi::to_wide (TYPE_SIZE_UNIT (elem_type)) * group_gap_adj); |
11703 | if (tree_int_cst_sgn (vect_dr_behavior (vinfo, dr_info)->step) == -1) |
11704 | bump_val = -bump_val; |
11705 | tree bump = wide_int_to_tree (sizetype, cst: bump_val); |
11706 | dataref_ptr = bump_vector_ptr (vinfo, dataref_ptr, ptr_incr, gsi, |
11707 | stmt_info, bump); |
11708 | } |
11709 | |
11710 | if (slp && !slp_perm) |
11711 | continue; |
11712 | |
11713 | if (slp_perm) |
11714 | { |
11715 | unsigned n_perms; |
11716 | /* For SLP we know we've seen all possible uses of dr_chain so |
11717 | direct vect_transform_slp_perm_load to DCE the unused parts. |
11718 | ??? This is a hack to prevent compile-time issues as seen |
11719 | in PR101120 and friends. */ |
11720 | if (costing_p) |
11721 | { |
11722 | vect_transform_slp_perm_load (vinfo, slp_node, vNULL, nullptr, vf, |
11723 | true, &n_perms, nullptr); |
11724 | inside_cost = record_stmt_cost (body_cost_vec: cost_vec, count: n_perms, kind: vec_perm, |
11725 | stmt_info, misalign: 0, where: vect_body); |
11726 | } |
11727 | else |
11728 | { |
11729 | bool ok = vect_transform_slp_perm_load (vinfo, slp_node, dr_chain, |
11730 | gsi, vf, false, &n_perms, |
11731 | nullptr, true); |
11732 | gcc_assert (ok); |
11733 | } |
11734 | } |
11735 | else |
11736 | { |
11737 | if (grouped_load) |
11738 | { |
11739 | gcc_assert (memory_access_type == VMAT_CONTIGUOUS_PERMUTE); |
11740 | /* We assume that the cost of a single load-lanes instruction |
11741 | is equivalent to the cost of DR_GROUP_SIZE separate loads. |
11742 | If a grouped access is instead being provided by a |
11743 | load-and-permute operation, include the cost of the |
11744 | permutes. */ |
11745 | if (costing_p && first_stmt_info == stmt_info) |
11746 | { |
11747 | /* Uses an even and odd extract operations or shuffle |
11748 | operations for each needed permute. */ |
11749 | int group_size = DR_GROUP_SIZE (first_stmt_info); |
11750 | int nstmts = ceil_log2 (x: group_size) * group_size; |
11751 | inside_cost += record_stmt_cost (body_cost_vec: cost_vec, count: nstmts, kind: vec_perm, |
11752 | stmt_info, misalign: 0, where: vect_body); |
11753 | |
11754 | if (dump_enabled_p ()) |
11755 | dump_printf_loc (MSG_NOTE, vect_location, |
11756 | "vect_model_load_cost:" |
11757 | "strided group_size = %d .\n" , |
11758 | group_size); |
11759 | } |
11760 | else if (!costing_p) |
11761 | { |
11762 | vect_transform_grouped_load (vinfo, stmt_info, dr_chain, |
11763 | group_size, gsi); |
11764 | *vec_stmt = STMT_VINFO_VEC_STMTS (stmt_info)[0]; |
11765 | } |
11766 | } |
11767 | else if (!costing_p) |
11768 | STMT_VINFO_VEC_STMTS (stmt_info).safe_push (obj: new_stmt); |
11769 | } |
11770 | dr_chain.release (); |
11771 | } |
11772 | if (!slp && !costing_p) |
11773 | *vec_stmt = STMT_VINFO_VEC_STMTS (stmt_info)[0]; |
11774 | |
11775 | if (costing_p) |
11776 | { |
11777 | gcc_assert (memory_access_type == VMAT_CONTIGUOUS |
11778 | || memory_access_type == VMAT_CONTIGUOUS_REVERSE |
11779 | || memory_access_type == VMAT_CONTIGUOUS_PERMUTE); |
11780 | if (n_adjacent_loads > 0) |
11781 | vect_get_load_cost (vinfo, stmt_info, ncopies: n_adjacent_loads, |
11782 | alignment_support_scheme, misalignment, add_realign_cost: false, |
11783 | inside_cost: &inside_cost, prologue_cost: &prologue_cost, prologue_cost_vec: cost_vec, body_cost_vec: cost_vec, |
11784 | record_prologue_costs: true); |
11785 | if (dump_enabled_p ()) |
11786 | dump_printf_loc (MSG_NOTE, vect_location, |
11787 | "vect_model_load_cost: inside_cost = %u, " |
11788 | "prologue_cost = %u .\n" , |
11789 | inside_cost, prologue_cost); |
11790 | } |
11791 | |
11792 | return true; |
11793 | } |
11794 | |
11795 | /* Function vect_is_simple_cond. |
11796 | |
11797 | Input: |
11798 | LOOP - the loop that is being vectorized. |
11799 | COND - Condition that is checked for simple use. |
11800 | |
11801 | Output: |
11802 | *COMP_VECTYPE - the vector type for the comparison. |
11803 | *DTS - The def types for the arguments of the comparison |
11804 | |
11805 | Returns whether a COND can be vectorized. Checks whether |
11806 | condition operands are supportable using vec_is_simple_use. */ |
11807 | |
11808 | static bool |
11809 | vect_is_simple_cond (tree cond, vec_info *vinfo, stmt_vec_info stmt_info, |
11810 | slp_tree slp_node, tree *comp_vectype, |
11811 | enum vect_def_type *dts, tree vectype) |
11812 | { |
11813 | tree lhs, rhs; |
11814 | tree vectype1 = NULL_TREE, vectype2 = NULL_TREE; |
11815 | slp_tree slp_op; |
11816 | |
11817 | /* Mask case. */ |
11818 | if (TREE_CODE (cond) == SSA_NAME |
11819 | && VECT_SCALAR_BOOLEAN_TYPE_P (TREE_TYPE (cond))) |
11820 | { |
11821 | if (!vect_is_simple_use (vinfo, stmt_info, slp_node, 0, &cond, |
11822 | &slp_op, &dts[0], comp_vectype) |
11823 | || !*comp_vectype |
11824 | || !VECTOR_BOOLEAN_TYPE_P (*comp_vectype)) |
11825 | return false; |
11826 | return true; |
11827 | } |
11828 | |
11829 | if (!COMPARISON_CLASS_P (cond)) |
11830 | return false; |
11831 | |
11832 | lhs = TREE_OPERAND (cond, 0); |
11833 | rhs = TREE_OPERAND (cond, 1); |
11834 | |
11835 | if (TREE_CODE (lhs) == SSA_NAME) |
11836 | { |
11837 | if (!vect_is_simple_use (vinfo, stmt_info, slp_node, 0, |
11838 | &lhs, &slp_op, &dts[0], &vectype1)) |
11839 | return false; |
11840 | } |
11841 | else if (TREE_CODE (lhs) == INTEGER_CST || TREE_CODE (lhs) == REAL_CST |
11842 | || TREE_CODE (lhs) == FIXED_CST) |
11843 | dts[0] = vect_constant_def; |
11844 | else |
11845 | return false; |
11846 | |
11847 | if (TREE_CODE (rhs) == SSA_NAME) |
11848 | { |
11849 | if (!vect_is_simple_use (vinfo, stmt_info, slp_node, 1, |
11850 | &rhs, &slp_op, &dts[1], &vectype2)) |
11851 | return false; |
11852 | } |
11853 | else if (TREE_CODE (rhs) == INTEGER_CST || TREE_CODE (rhs) == REAL_CST |
11854 | || TREE_CODE (rhs) == FIXED_CST) |
11855 | dts[1] = vect_constant_def; |
11856 | else |
11857 | return false; |
11858 | |
11859 | if (vectype1 && vectype2 |
11860 | && maybe_ne (a: TYPE_VECTOR_SUBPARTS (node: vectype1), |
11861 | b: TYPE_VECTOR_SUBPARTS (node: vectype2))) |
11862 | return false; |
11863 | |
11864 | *comp_vectype = vectype1 ? vectype1 : vectype2; |
11865 | /* Invariant comparison. */ |
11866 | if (! *comp_vectype) |
11867 | { |
11868 | tree scalar_type = TREE_TYPE (lhs); |
11869 | if (VECT_SCALAR_BOOLEAN_TYPE_P (scalar_type)) |
11870 | *comp_vectype = truth_type_for (vectype); |
11871 | else |
11872 | { |
11873 | /* If we can widen the comparison to match vectype do so. */ |
11874 | if (INTEGRAL_TYPE_P (scalar_type) |
11875 | && !slp_node |
11876 | && tree_int_cst_lt (TYPE_SIZE (scalar_type), |
11877 | TYPE_SIZE (TREE_TYPE (vectype)))) |
11878 | scalar_type = build_nonstandard_integer_type |
11879 | (vector_element_bits (vectype), TYPE_UNSIGNED (scalar_type)); |
11880 | *comp_vectype = get_vectype_for_scalar_type (vinfo, scalar_type, |
11881 | slp_node); |
11882 | } |
11883 | } |
11884 | |
11885 | return true; |
11886 | } |
11887 | |
11888 | /* vectorizable_condition. |
11889 | |
11890 | Check if STMT_INFO is conditional modify expression that can be vectorized. |
11891 | If VEC_STMT is also passed, vectorize STMT_INFO: create a vectorized |
11892 | stmt using VEC_COND_EXPR to replace it, put it in VEC_STMT, and insert it |
11893 | at GSI. |
11894 | |
11895 | When STMT_INFO is vectorized as a nested cycle, for_reduction is true. |
11896 | |
11897 | Return true if STMT_INFO is vectorizable in this way. */ |
11898 | |
11899 | static bool |
11900 | vectorizable_condition (vec_info *vinfo, |
11901 | stmt_vec_info stmt_info, gimple_stmt_iterator *gsi, |
11902 | gimple **vec_stmt, |
11903 | slp_tree slp_node, stmt_vector_for_cost *cost_vec) |
11904 | { |
11905 | tree scalar_dest = NULL_TREE; |
11906 | tree vec_dest = NULL_TREE; |
11907 | tree cond_expr, cond_expr0 = NULL_TREE, cond_expr1 = NULL_TREE; |
11908 | tree then_clause, else_clause; |
11909 | tree comp_vectype = NULL_TREE; |
11910 | tree vec_cond_lhs = NULL_TREE, vec_cond_rhs = NULL_TREE; |
11911 | tree vec_then_clause = NULL_TREE, vec_else_clause = NULL_TREE; |
11912 | tree vec_compare; |
11913 | tree new_temp; |
11914 | loop_vec_info loop_vinfo = dyn_cast <loop_vec_info> (p: vinfo); |
11915 | enum vect_def_type dts[4] |
11916 | = {vect_unknown_def_type, vect_unknown_def_type, |
11917 | vect_unknown_def_type, vect_unknown_def_type}; |
11918 | int ndts = 4; |
11919 | int ncopies; |
11920 | int vec_num; |
11921 | enum tree_code code, cond_code, bitop1 = NOP_EXPR, bitop2 = NOP_EXPR; |
11922 | int i; |
11923 | bb_vec_info bb_vinfo = dyn_cast <bb_vec_info> (p: vinfo); |
11924 | vec<tree> vec_oprnds0 = vNULL; |
11925 | vec<tree> vec_oprnds1 = vNULL; |
11926 | vec<tree> vec_oprnds2 = vNULL; |
11927 | vec<tree> vec_oprnds3 = vNULL; |
11928 | tree vec_cmp_type; |
11929 | bool masked = false; |
11930 | |
11931 | if (!STMT_VINFO_RELEVANT_P (stmt_info) && !bb_vinfo) |
11932 | return false; |
11933 | |
11934 | /* Is vectorizable conditional operation? */ |
11935 | gassign *stmt = dyn_cast <gassign *> (p: stmt_info->stmt); |
11936 | if (!stmt) |
11937 | return false; |
11938 | |
11939 | code = gimple_assign_rhs_code (gs: stmt); |
11940 | if (code != COND_EXPR) |
11941 | return false; |
11942 | |
11943 | stmt_vec_info reduc_info = NULL; |
11944 | int reduc_index = -1; |
11945 | vect_reduction_type reduction_type = TREE_CODE_REDUCTION; |
11946 | bool for_reduction |
11947 | = STMT_VINFO_REDUC_DEF (vect_orig_stmt (stmt_info)) != NULL; |
11948 | if (for_reduction) |
11949 | { |
11950 | if (slp_node) |
11951 | return false; |
11952 | reduc_info = info_for_reduction (vinfo, stmt_info); |
11953 | reduction_type = STMT_VINFO_REDUC_TYPE (reduc_info); |
11954 | reduc_index = STMT_VINFO_REDUC_IDX (stmt_info); |
11955 | gcc_assert (reduction_type != EXTRACT_LAST_REDUCTION |
11956 | || reduc_index != -1); |
11957 | } |
11958 | else |
11959 | { |
11960 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def) |
11961 | return false; |
11962 | } |
11963 | |
11964 | tree vectype = STMT_VINFO_VECTYPE (stmt_info); |
11965 | tree vectype1 = NULL_TREE, vectype2 = NULL_TREE; |
11966 | |
11967 | if (slp_node) |
11968 | { |
11969 | ncopies = 1; |
11970 | vec_num = SLP_TREE_NUMBER_OF_VEC_STMTS (slp_node); |
11971 | } |
11972 | else |
11973 | { |
11974 | ncopies = vect_get_num_copies (loop_vinfo, vectype); |
11975 | vec_num = 1; |
11976 | } |
11977 | |
11978 | gcc_assert (ncopies >= 1); |
11979 | if (for_reduction && ncopies > 1) |
11980 | return false; /* FORNOW */ |
11981 | |
11982 | cond_expr = gimple_assign_rhs1 (gs: stmt); |
11983 | |
11984 | if (!vect_is_simple_cond (cond: cond_expr, vinfo, stmt_info, slp_node, |
11985 | comp_vectype: &comp_vectype, dts: &dts[0], vectype) |
11986 | || !comp_vectype) |
11987 | return false; |
11988 | |
11989 | unsigned op_adjust = COMPARISON_CLASS_P (cond_expr) ? 1 : 0; |
11990 | slp_tree then_slp_node, else_slp_node; |
11991 | if (!vect_is_simple_use (vinfo, stmt_info, slp_node, 1 + op_adjust, |
11992 | &then_clause, &then_slp_node, &dts[2], &vectype1)) |
11993 | return false; |
11994 | if (!vect_is_simple_use (vinfo, stmt_info, slp_node, 2 + op_adjust, |
11995 | &else_clause, &else_slp_node, &dts[3], &vectype2)) |
11996 | return false; |
11997 | |
11998 | if (vectype1 && !useless_type_conversion_p (vectype, vectype1)) |
11999 | return false; |
12000 | |
12001 | if (vectype2 && !useless_type_conversion_p (vectype, vectype2)) |
12002 | return false; |
12003 | |
12004 | masked = !COMPARISON_CLASS_P (cond_expr); |
12005 | vec_cmp_type = truth_type_for (comp_vectype); |
12006 | |
12007 | if (vec_cmp_type == NULL_TREE) |
12008 | return false; |
12009 | |
12010 | cond_code = TREE_CODE (cond_expr); |
12011 | if (!masked) |
12012 | { |
12013 | cond_expr0 = TREE_OPERAND (cond_expr, 0); |
12014 | cond_expr1 = TREE_OPERAND (cond_expr, 1); |
12015 | } |
12016 | |
12017 | /* For conditional reductions, the "then" value needs to be the candidate |
12018 | value calculated by this iteration while the "else" value needs to be |
12019 | the result carried over from previous iterations. If the COND_EXPR |
12020 | is the other way around, we need to swap it. */ |
12021 | bool must_invert_cmp_result = false; |
12022 | if (reduction_type == EXTRACT_LAST_REDUCTION && reduc_index == 1) |
12023 | { |
12024 | if (masked) |
12025 | must_invert_cmp_result = true; |
12026 | else |
12027 | { |
12028 | bool honor_nans = HONOR_NANS (TREE_TYPE (cond_expr0)); |
12029 | tree_code new_code = invert_tree_comparison (cond_code, honor_nans); |
12030 | if (new_code == ERROR_MARK) |
12031 | must_invert_cmp_result = true; |
12032 | else |
12033 | { |
12034 | cond_code = new_code; |
12035 | /* Make sure we don't accidentally use the old condition. */ |
12036 | cond_expr = NULL_TREE; |
12037 | } |
12038 | } |
12039 | std::swap (a&: then_clause, b&: else_clause); |
12040 | } |
12041 | |
12042 | if (!masked && VECTOR_BOOLEAN_TYPE_P (comp_vectype)) |
12043 | { |
12044 | /* Boolean values may have another representation in vectors |
12045 | and therefore we prefer bit operations over comparison for |
12046 | them (which also works for scalar masks). We store opcodes |
12047 | to use in bitop1 and bitop2. Statement is vectorized as |
12048 | BITOP2 (rhs1 BITOP1 rhs2) or rhs1 BITOP2 (BITOP1 rhs2) |
12049 | depending on bitop1 and bitop2 arity. */ |
12050 | switch (cond_code) |
12051 | { |
12052 | case GT_EXPR: |
12053 | bitop1 = BIT_NOT_EXPR; |
12054 | bitop2 = BIT_AND_EXPR; |
12055 | break; |
12056 | case GE_EXPR: |
12057 | bitop1 = BIT_NOT_EXPR; |
12058 | bitop2 = BIT_IOR_EXPR; |
12059 | break; |
12060 | case LT_EXPR: |
12061 | bitop1 = BIT_NOT_EXPR; |
12062 | bitop2 = BIT_AND_EXPR; |
12063 | std::swap (a&: cond_expr0, b&: cond_expr1); |
12064 | break; |
12065 | case LE_EXPR: |
12066 | bitop1 = BIT_NOT_EXPR; |
12067 | bitop2 = BIT_IOR_EXPR; |
12068 | std::swap (a&: cond_expr0, b&: cond_expr1); |
12069 | break; |
12070 | case NE_EXPR: |
12071 | bitop1 = BIT_XOR_EXPR; |
12072 | break; |
12073 | case EQ_EXPR: |
12074 | bitop1 = BIT_XOR_EXPR; |
12075 | bitop2 = BIT_NOT_EXPR; |
12076 | break; |
12077 | default: |
12078 | return false; |
12079 | } |
12080 | cond_code = SSA_NAME; |
12081 | } |
12082 | |
12083 | if (TREE_CODE_CLASS (cond_code) == tcc_comparison |
12084 | && reduction_type == EXTRACT_LAST_REDUCTION |
12085 | && !expand_vec_cmp_expr_p (comp_vectype, vec_cmp_type, cond_code)) |
12086 | { |
12087 | if (dump_enabled_p ()) |
12088 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
12089 | "reduction comparison operation not supported.\n" ); |
12090 | return false; |
12091 | } |
12092 | |
12093 | if (!vec_stmt) |
12094 | { |
12095 | if (bitop1 != NOP_EXPR) |
12096 | { |
12097 | machine_mode mode = TYPE_MODE (comp_vectype); |
12098 | optab optab; |
12099 | |
12100 | optab = optab_for_tree_code (bitop1, comp_vectype, optab_default); |
12101 | if (!optab || optab_handler (op: optab, mode) == CODE_FOR_nothing) |
12102 | return false; |
12103 | |
12104 | if (bitop2 != NOP_EXPR) |
12105 | { |
12106 | optab = optab_for_tree_code (bitop2, comp_vectype, |
12107 | optab_default); |
12108 | if (!optab || optab_handler (op: optab, mode) == CODE_FOR_nothing) |
12109 | return false; |
12110 | } |
12111 | } |
12112 | |
12113 | vect_cost_for_stmt kind = vector_stmt; |
12114 | if (reduction_type == EXTRACT_LAST_REDUCTION) |
12115 | /* Count one reduction-like operation per vector. */ |
12116 | kind = vec_to_scalar; |
12117 | else if (!expand_vec_cond_expr_p (vectype, comp_vectype, cond_code) |
12118 | && (masked |
12119 | || (!expand_vec_cmp_expr_p (comp_vectype, vec_cmp_type, |
12120 | cond_code) |
12121 | || !expand_vec_cond_expr_p (vectype, vec_cmp_type, |
12122 | ERROR_MARK)))) |
12123 | return false; |
12124 | |
12125 | if (slp_node |
12126 | && (!vect_maybe_update_slp_op_vectype |
12127 | (SLP_TREE_CHILDREN (slp_node)[0], comp_vectype) |
12128 | || (op_adjust == 1 |
12129 | && !vect_maybe_update_slp_op_vectype |
12130 | (SLP_TREE_CHILDREN (slp_node)[1], comp_vectype)) |
12131 | || !vect_maybe_update_slp_op_vectype (then_slp_node, vectype) |
12132 | || !vect_maybe_update_slp_op_vectype (else_slp_node, vectype))) |
12133 | { |
12134 | if (dump_enabled_p ()) |
12135 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
12136 | "incompatible vector types for invariants\n" ); |
12137 | return false; |
12138 | } |
12139 | |
12140 | if (loop_vinfo && for_reduction |
12141 | && LOOP_VINFO_CAN_USE_PARTIAL_VECTORS_P (loop_vinfo)) |
12142 | { |
12143 | if (reduction_type == EXTRACT_LAST_REDUCTION) |
12144 | { |
12145 | if (direct_internal_fn_supported_p (IFN_LEN_FOLD_EXTRACT_LAST, |
12146 | vectype, OPTIMIZE_FOR_SPEED)) |
12147 | vect_record_loop_len (loop_vinfo, |
12148 | &LOOP_VINFO_LENS (loop_vinfo), |
12149 | ncopies * vec_num, vectype, 1); |
12150 | else |
12151 | vect_record_loop_mask (loop_vinfo, |
12152 | &LOOP_VINFO_MASKS (loop_vinfo), |
12153 | ncopies * vec_num, vectype, NULL); |
12154 | } |
12155 | /* Extra inactive lanes should be safe for vect_nested_cycle. */ |
12156 | else if (STMT_VINFO_DEF_TYPE (reduc_info) != vect_nested_cycle) |
12157 | { |
12158 | if (dump_enabled_p ()) |
12159 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
12160 | "conditional reduction prevents the use" |
12161 | " of partial vectors.\n" ); |
12162 | LOOP_VINFO_CAN_USE_PARTIAL_VECTORS_P (loop_vinfo) = false; |
12163 | } |
12164 | } |
12165 | |
12166 | STMT_VINFO_TYPE (stmt_info) = condition_vec_info_type; |
12167 | vect_model_simple_cost (vinfo, stmt_info, ncopies, dt: dts, ndts, node: slp_node, |
12168 | cost_vec, kind); |
12169 | return true; |
12170 | } |
12171 | |
12172 | /* Transform. */ |
12173 | |
12174 | /* Handle def. */ |
12175 | scalar_dest = gimple_assign_lhs (gs: stmt); |
12176 | if (reduction_type != EXTRACT_LAST_REDUCTION) |
12177 | vec_dest = vect_create_destination_var (scalar_dest, vectype); |
12178 | |
12179 | bool swap_cond_operands = false; |
12180 | |
12181 | /* See whether another part of the vectorized code applies a loop |
12182 | mask to the condition, or to its inverse. */ |
12183 | |
12184 | vec_loop_masks *masks = NULL; |
12185 | vec_loop_lens *lens = NULL; |
12186 | if (loop_vinfo && LOOP_VINFO_FULLY_WITH_LENGTH_P (loop_vinfo)) |
12187 | { |
12188 | if (reduction_type == EXTRACT_LAST_REDUCTION) |
12189 | lens = &LOOP_VINFO_LENS (loop_vinfo); |
12190 | } |
12191 | else if (loop_vinfo && LOOP_VINFO_FULLY_MASKED_P (loop_vinfo)) |
12192 | { |
12193 | if (reduction_type == EXTRACT_LAST_REDUCTION) |
12194 | masks = &LOOP_VINFO_MASKS (loop_vinfo); |
12195 | else |
12196 | { |
12197 | scalar_cond_masked_key cond (cond_expr, ncopies); |
12198 | if (loop_vinfo->scalar_cond_masked_set.contains (k: cond)) |
12199 | masks = &LOOP_VINFO_MASKS (loop_vinfo); |
12200 | else |
12201 | { |
12202 | bool honor_nans = HONOR_NANS (TREE_TYPE (cond.op0)); |
12203 | tree_code orig_code = cond.code; |
12204 | cond.code = invert_tree_comparison (cond.code, honor_nans); |
12205 | if (!masked && loop_vinfo->scalar_cond_masked_set.contains (k: cond)) |
12206 | { |
12207 | masks = &LOOP_VINFO_MASKS (loop_vinfo); |
12208 | cond_code = cond.code; |
12209 | swap_cond_operands = true; |
12210 | } |
12211 | else |
12212 | { |
12213 | /* Try the inverse of the current mask. We check if the |
12214 | inverse mask is live and if so we generate a negate of |
12215 | the current mask such that we still honor NaNs. */ |
12216 | cond.inverted_p = true; |
12217 | cond.code = orig_code; |
12218 | if (loop_vinfo->scalar_cond_masked_set.contains (k: cond)) |
12219 | { |
12220 | masks = &LOOP_VINFO_MASKS (loop_vinfo); |
12221 | cond_code = cond.code; |
12222 | swap_cond_operands = true; |
12223 | must_invert_cmp_result = true; |
12224 | } |
12225 | } |
12226 | } |
12227 | } |
12228 | } |
12229 | |
12230 | /* Handle cond expr. */ |
12231 | if (masked) |
12232 | vect_get_vec_defs (vinfo, stmt_info, slp_node, ncopies, |
12233 | op0: cond_expr, vec_oprnds0: &vec_oprnds0, vectype0: comp_vectype, |
12234 | op1: then_clause, vec_oprnds1: &vec_oprnds2, vectype1: vectype, |
12235 | op2: reduction_type != EXTRACT_LAST_REDUCTION |
12236 | ? else_clause : NULL, vec_oprnds2: &vec_oprnds3, vectype2: vectype); |
12237 | else |
12238 | vect_get_vec_defs (vinfo, stmt_info, slp_node, ncopies, |
12239 | op0: cond_expr0, vec_oprnds0: &vec_oprnds0, vectype0: comp_vectype, |
12240 | op1: cond_expr1, vec_oprnds1: &vec_oprnds1, vectype1: comp_vectype, |
12241 | op2: then_clause, vec_oprnds2: &vec_oprnds2, vectype2: vectype, |
12242 | op3: reduction_type != EXTRACT_LAST_REDUCTION |
12243 | ? else_clause : NULL, vec_oprnds3: &vec_oprnds3, vectype3: vectype); |
12244 | |
12245 | /* Arguments are ready. Create the new vector stmt. */ |
12246 | FOR_EACH_VEC_ELT (vec_oprnds0, i, vec_cond_lhs) |
12247 | { |
12248 | vec_then_clause = vec_oprnds2[i]; |
12249 | if (reduction_type != EXTRACT_LAST_REDUCTION) |
12250 | vec_else_clause = vec_oprnds3[i]; |
12251 | |
12252 | if (swap_cond_operands) |
12253 | std::swap (a&: vec_then_clause, b&: vec_else_clause); |
12254 | |
12255 | if (masked) |
12256 | vec_compare = vec_cond_lhs; |
12257 | else |
12258 | { |
12259 | vec_cond_rhs = vec_oprnds1[i]; |
12260 | if (bitop1 == NOP_EXPR) |
12261 | { |
12262 | gimple_seq stmts = NULL; |
12263 | vec_compare = gimple_build (seq: &stmts, code: cond_code, type: vec_cmp_type, |
12264 | ops: vec_cond_lhs, ops: vec_cond_rhs); |
12265 | gsi_insert_before (gsi, stmts, GSI_SAME_STMT); |
12266 | } |
12267 | else |
12268 | { |
12269 | new_temp = make_ssa_name (var: vec_cmp_type); |
12270 | gassign *new_stmt; |
12271 | if (bitop1 == BIT_NOT_EXPR) |
12272 | new_stmt = gimple_build_assign (new_temp, bitop1, |
12273 | vec_cond_rhs); |
12274 | else |
12275 | new_stmt |
12276 | = gimple_build_assign (new_temp, bitop1, vec_cond_lhs, |
12277 | vec_cond_rhs); |
12278 | vect_finish_stmt_generation (vinfo, stmt_info, vec_stmt: new_stmt, gsi); |
12279 | if (bitop2 == NOP_EXPR) |
12280 | vec_compare = new_temp; |
12281 | else if (bitop2 == BIT_NOT_EXPR |
12282 | && reduction_type != EXTRACT_LAST_REDUCTION) |
12283 | { |
12284 | /* Instead of doing ~x ? y : z do x ? z : y. */ |
12285 | vec_compare = new_temp; |
12286 | std::swap (a&: vec_then_clause, b&: vec_else_clause); |
12287 | } |
12288 | else |
12289 | { |
12290 | vec_compare = make_ssa_name (var: vec_cmp_type); |
12291 | if (bitop2 == BIT_NOT_EXPR) |
12292 | new_stmt |
12293 | = gimple_build_assign (vec_compare, bitop2, new_temp); |
12294 | else |
12295 | new_stmt |
12296 | = gimple_build_assign (vec_compare, bitop2, |
12297 | vec_cond_lhs, new_temp); |
12298 | vect_finish_stmt_generation (vinfo, stmt_info, |
12299 | vec_stmt: new_stmt, gsi); |
12300 | } |
12301 | } |
12302 | } |
12303 | |
12304 | /* If we decided to apply a loop mask to the result of the vector |
12305 | comparison, AND the comparison with the mask now. Later passes |
12306 | should then be able to reuse the AND results between mulitple |
12307 | vector statements. |
12308 | |
12309 | For example: |
12310 | for (int i = 0; i < 100; ++i) |
12311 | x[i] = y[i] ? z[i] : 10; |
12312 | |
12313 | results in following optimized GIMPLE: |
12314 | |
12315 | mask__35.8_43 = vect__4.7_41 != { 0, ... }; |
12316 | vec_mask_and_46 = loop_mask_40 & mask__35.8_43; |
12317 | _19 = &MEM[base: z_12(D), index: ivtmp_56, step: 4, offset: 0B]; |
12318 | vect_iftmp.11_47 = .MASK_LOAD (_19, 4B, vec_mask_and_46); |
12319 | vect_iftmp.12_52 = VEC_COND_EXPR <vec_mask_and_46, |
12320 | vect_iftmp.11_47, { 10, ... }>; |
12321 | |
12322 | instead of using a masked and unmasked forms of |
12323 | vec != { 0, ... } (masked in the MASK_LOAD, |
12324 | unmasked in the VEC_COND_EXPR). */ |
12325 | |
12326 | /* Force vec_compare to be an SSA_NAME rather than a comparison, |
12327 | in cases where that's necessary. */ |
12328 | |
12329 | tree len = NULL_TREE, bias = NULL_TREE; |
12330 | if (masks || lens || reduction_type == EXTRACT_LAST_REDUCTION) |
12331 | { |
12332 | if (!is_gimple_val (vec_compare)) |
12333 | { |
12334 | tree vec_compare_name = make_ssa_name (var: vec_cmp_type); |
12335 | gassign *new_stmt = gimple_build_assign (vec_compare_name, |
12336 | vec_compare); |
12337 | vect_finish_stmt_generation (vinfo, stmt_info, vec_stmt: new_stmt, gsi); |
12338 | vec_compare = vec_compare_name; |
12339 | } |
12340 | |
12341 | if (must_invert_cmp_result) |
12342 | { |
12343 | tree vec_compare_name = make_ssa_name (var: vec_cmp_type); |
12344 | gassign *new_stmt = gimple_build_assign (vec_compare_name, |
12345 | BIT_NOT_EXPR, |
12346 | vec_compare); |
12347 | vect_finish_stmt_generation (vinfo, stmt_info, vec_stmt: new_stmt, gsi); |
12348 | vec_compare = vec_compare_name; |
12349 | } |
12350 | |
12351 | if (direct_internal_fn_supported_p (IFN_LEN_FOLD_EXTRACT_LAST, |
12352 | vectype, OPTIMIZE_FOR_SPEED)) |
12353 | { |
12354 | if (lens) |
12355 | { |
12356 | len = vect_get_loop_len (loop_vinfo, gsi, lens, |
12357 | vec_num * ncopies, vectype, i, 1); |
12358 | signed char biasval |
12359 | = LOOP_VINFO_PARTIAL_LOAD_STORE_BIAS (loop_vinfo); |
12360 | bias = build_int_cst (intQI_type_node, biasval); |
12361 | } |
12362 | else |
12363 | { |
12364 | len = size_int (TYPE_VECTOR_SUBPARTS (vectype)); |
12365 | bias = build_int_cst (intQI_type_node, 0); |
12366 | } |
12367 | } |
12368 | if (masks) |
12369 | { |
12370 | tree loop_mask |
12371 | = vect_get_loop_mask (loop_vinfo, gsi, masks, vec_num * ncopies, |
12372 | vectype, i); |
12373 | tree tmp2 = make_ssa_name (var: vec_cmp_type); |
12374 | gassign *g |
12375 | = gimple_build_assign (tmp2, BIT_AND_EXPR, vec_compare, |
12376 | loop_mask); |
12377 | vect_finish_stmt_generation (vinfo, stmt_info, vec_stmt: g, gsi); |
12378 | vec_compare = tmp2; |
12379 | } |
12380 | } |
12381 | |
12382 | gimple *new_stmt; |
12383 | if (reduction_type == EXTRACT_LAST_REDUCTION) |
12384 | { |
12385 | gimple *old_stmt = vect_orig_stmt (stmt_info)->stmt; |
12386 | tree lhs = gimple_get_lhs (old_stmt); |
12387 | if (len) |
12388 | new_stmt = gimple_build_call_internal |
12389 | (IFN_LEN_FOLD_EXTRACT_LAST, 5, else_clause, vec_compare, |
12390 | vec_then_clause, len, bias); |
12391 | else |
12392 | new_stmt = gimple_build_call_internal |
12393 | (IFN_FOLD_EXTRACT_LAST, 3, else_clause, vec_compare, |
12394 | vec_then_clause); |
12395 | gimple_call_set_lhs (gs: new_stmt, lhs); |
12396 | SSA_NAME_DEF_STMT (lhs) = new_stmt; |
12397 | if (old_stmt == gsi_stmt (i: *gsi)) |
12398 | vect_finish_replace_stmt (vinfo, stmt_info, vec_stmt: new_stmt); |
12399 | else |
12400 | { |
12401 | /* In this case we're moving the definition to later in the |
12402 | block. That doesn't matter because the only uses of the |
12403 | lhs are in phi statements. */ |
12404 | gimple_stmt_iterator old_gsi = gsi_for_stmt (old_stmt); |
12405 | gsi_remove (&old_gsi, true); |
12406 | vect_finish_stmt_generation (vinfo, stmt_info, vec_stmt: new_stmt, gsi); |
12407 | } |
12408 | } |
12409 | else |
12410 | { |
12411 | new_temp = make_ssa_name (var: vec_dest); |
12412 | new_stmt = gimple_build_assign (new_temp, VEC_COND_EXPR, vec_compare, |
12413 | vec_then_clause, vec_else_clause); |
12414 | vect_finish_stmt_generation (vinfo, stmt_info, vec_stmt: new_stmt, gsi); |
12415 | } |
12416 | if (slp_node) |
12417 | slp_node->push_vec_def (def: new_stmt); |
12418 | else |
12419 | STMT_VINFO_VEC_STMTS (stmt_info).safe_push (obj: new_stmt); |
12420 | } |
12421 | |
12422 | if (!slp_node) |
12423 | *vec_stmt = STMT_VINFO_VEC_STMTS (stmt_info)[0]; |
12424 | |
12425 | vec_oprnds0.release (); |
12426 | vec_oprnds1.release (); |
12427 | vec_oprnds2.release (); |
12428 | vec_oprnds3.release (); |
12429 | |
12430 | return true; |
12431 | } |
12432 | |
12433 | /* Helper of vectorizable_comparison. |
12434 | |
12435 | Check if STMT_INFO is comparison expression CODE that can be vectorized. |
12436 | If VEC_STMT is also passed, vectorize STMT_INFO: create a vectorized |
12437 | comparison, put it in VEC_STMT, and insert it at GSI. |
12438 | |
12439 | Return true if STMT_INFO is vectorizable in this way. */ |
12440 | |
12441 | static bool |
12442 | vectorizable_comparison_1 (vec_info *vinfo, tree vectype, |
12443 | stmt_vec_info stmt_info, tree_code code, |
12444 | gimple_stmt_iterator *gsi, gimple **vec_stmt, |
12445 | slp_tree slp_node, stmt_vector_for_cost *cost_vec) |
12446 | { |
12447 | tree lhs, rhs1, rhs2; |
12448 | tree vectype1 = NULL_TREE, vectype2 = NULL_TREE; |
12449 | tree vec_rhs1 = NULL_TREE, vec_rhs2 = NULL_TREE; |
12450 | tree new_temp; |
12451 | loop_vec_info loop_vinfo = dyn_cast <loop_vec_info> (p: vinfo); |
12452 | enum vect_def_type dts[2] = {vect_unknown_def_type, vect_unknown_def_type}; |
12453 | int ndts = 2; |
12454 | poly_uint64 nunits; |
12455 | int ncopies; |
12456 | enum tree_code bitop1 = NOP_EXPR, bitop2 = NOP_EXPR; |
12457 | int i; |
12458 | bb_vec_info bb_vinfo = dyn_cast <bb_vec_info> (p: vinfo); |
12459 | vec<tree> vec_oprnds0 = vNULL; |
12460 | vec<tree> vec_oprnds1 = vNULL; |
12461 | tree mask_type; |
12462 | tree mask; |
12463 | |
12464 | if (!STMT_VINFO_RELEVANT_P (stmt_info) && !bb_vinfo) |
12465 | return false; |
12466 | |
12467 | if (!vectype || !VECTOR_BOOLEAN_TYPE_P (vectype)) |
12468 | return false; |
12469 | |
12470 | mask_type = vectype; |
12471 | nunits = TYPE_VECTOR_SUBPARTS (node: vectype); |
12472 | |
12473 | if (slp_node) |
12474 | ncopies = 1; |
12475 | else |
12476 | ncopies = vect_get_num_copies (loop_vinfo, vectype); |
12477 | |
12478 | gcc_assert (ncopies >= 1); |
12479 | |
12480 | if (TREE_CODE_CLASS (code) != tcc_comparison) |
12481 | return false; |
12482 | |
12483 | slp_tree slp_rhs1, slp_rhs2; |
12484 | if (!vect_is_simple_use (vinfo, stmt_info, slp_node, |
12485 | 0, &rhs1, &slp_rhs1, &dts[0], &vectype1)) |
12486 | return false; |
12487 | |
12488 | if (!vect_is_simple_use (vinfo, stmt_info, slp_node, |
12489 | 1, &rhs2, &slp_rhs2, &dts[1], &vectype2)) |
12490 | return false; |
12491 | |
12492 | if (vectype1 && vectype2 |
12493 | && maybe_ne (a: TYPE_VECTOR_SUBPARTS (node: vectype1), |
12494 | b: TYPE_VECTOR_SUBPARTS (node: vectype2))) |
12495 | return false; |
12496 | |
12497 | vectype = vectype1 ? vectype1 : vectype2; |
12498 | |
12499 | /* Invariant comparison. */ |
12500 | if (!vectype) |
12501 | { |
12502 | if (VECT_SCALAR_BOOLEAN_TYPE_P (TREE_TYPE (rhs1))) |
12503 | vectype = mask_type; |
12504 | else |
12505 | vectype = get_vectype_for_scalar_type (vinfo, TREE_TYPE (rhs1), |
12506 | slp_node); |
12507 | if (!vectype || maybe_ne (a: TYPE_VECTOR_SUBPARTS (node: vectype), b: nunits)) |
12508 | return false; |
12509 | } |
12510 | else if (maybe_ne (a: nunits, b: TYPE_VECTOR_SUBPARTS (node: vectype))) |
12511 | return false; |
12512 | |
12513 | /* Can't compare mask and non-mask types. */ |
12514 | if (vectype1 && vectype2 |
12515 | && (VECTOR_BOOLEAN_TYPE_P (vectype1) ^ VECTOR_BOOLEAN_TYPE_P (vectype2))) |
12516 | return false; |
12517 | |
12518 | /* Boolean values may have another representation in vectors |
12519 | and therefore we prefer bit operations over comparison for |
12520 | them (which also works for scalar masks). We store opcodes |
12521 | to use in bitop1 and bitop2. Statement is vectorized as |
12522 | BITOP2 (rhs1 BITOP1 rhs2) or |
12523 | rhs1 BITOP2 (BITOP1 rhs2) |
12524 | depending on bitop1 and bitop2 arity. */ |
12525 | bool swap_p = false; |
12526 | if (VECTOR_BOOLEAN_TYPE_P (vectype)) |
12527 | { |
12528 | if (code == GT_EXPR) |
12529 | { |
12530 | bitop1 = BIT_NOT_EXPR; |
12531 | bitop2 = BIT_AND_EXPR; |
12532 | } |
12533 | else if (code == GE_EXPR) |
12534 | { |
12535 | bitop1 = BIT_NOT_EXPR; |
12536 | bitop2 = BIT_IOR_EXPR; |
12537 | } |
12538 | else if (code == LT_EXPR) |
12539 | { |
12540 | bitop1 = BIT_NOT_EXPR; |
12541 | bitop2 = BIT_AND_EXPR; |
12542 | swap_p = true; |
12543 | } |
12544 | else if (code == LE_EXPR) |
12545 | { |
12546 | bitop1 = BIT_NOT_EXPR; |
12547 | bitop2 = BIT_IOR_EXPR; |
12548 | swap_p = true; |
12549 | } |
12550 | else |
12551 | { |
12552 | bitop1 = BIT_XOR_EXPR; |
12553 | if (code == EQ_EXPR) |
12554 | bitop2 = BIT_NOT_EXPR; |
12555 | } |
12556 | } |
12557 | |
12558 | if (!vec_stmt) |
12559 | { |
12560 | if (bitop1 == NOP_EXPR) |
12561 | { |
12562 | if (!expand_vec_cmp_expr_p (vectype, mask_type, code)) |
12563 | return false; |
12564 | } |
12565 | else |
12566 | { |
12567 | machine_mode mode = TYPE_MODE (vectype); |
12568 | optab optab; |
12569 | |
12570 | optab = optab_for_tree_code (bitop1, vectype, optab_default); |
12571 | if (!optab || optab_handler (op: optab, mode) == CODE_FOR_nothing) |
12572 | return false; |
12573 | |
12574 | if (bitop2 != NOP_EXPR) |
12575 | { |
12576 | optab = optab_for_tree_code (bitop2, vectype, optab_default); |
12577 | if (!optab || optab_handler (op: optab, mode) == CODE_FOR_nothing) |
12578 | return false; |
12579 | } |
12580 | } |
12581 | |
12582 | /* Put types on constant and invariant SLP children. */ |
12583 | if (slp_node |
12584 | && (!vect_maybe_update_slp_op_vectype (slp_rhs1, vectype) |
12585 | || !vect_maybe_update_slp_op_vectype (slp_rhs2, vectype))) |
12586 | { |
12587 | if (dump_enabled_p ()) |
12588 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
12589 | "incompatible vector types for invariants\n" ); |
12590 | return false; |
12591 | } |
12592 | |
12593 | vect_model_simple_cost (vinfo, stmt_info, |
12594 | ncopies: ncopies * (1 + (bitop2 != NOP_EXPR)), |
12595 | dt: dts, ndts, node: slp_node, cost_vec); |
12596 | return true; |
12597 | } |
12598 | |
12599 | /* Transform. */ |
12600 | |
12601 | /* Handle def. */ |
12602 | lhs = gimple_assign_lhs (STMT_VINFO_STMT (stmt_info)); |
12603 | mask = vect_create_destination_var (lhs, mask_type); |
12604 | |
12605 | vect_get_vec_defs (vinfo, stmt_info, slp_node, ncopies, |
12606 | op0: rhs1, vec_oprnds0: &vec_oprnds0, vectype0: vectype, |
12607 | op1: rhs2, vec_oprnds1: &vec_oprnds1, vectype1: vectype); |
12608 | if (swap_p) |
12609 | std::swap (a&: vec_oprnds0, b&: vec_oprnds1); |
12610 | |
12611 | /* Arguments are ready. Create the new vector stmt. */ |
12612 | FOR_EACH_VEC_ELT (vec_oprnds0, i, vec_rhs1) |
12613 | { |
12614 | gimple *new_stmt; |
12615 | vec_rhs2 = vec_oprnds1[i]; |
12616 | |
12617 | new_temp = make_ssa_name (var: mask); |
12618 | if (bitop1 == NOP_EXPR) |
12619 | { |
12620 | new_stmt = gimple_build_assign (new_temp, code, |
12621 | vec_rhs1, vec_rhs2); |
12622 | vect_finish_stmt_generation (vinfo, stmt_info, vec_stmt: new_stmt, gsi); |
12623 | } |
12624 | else |
12625 | { |
12626 | if (bitop1 == BIT_NOT_EXPR) |
12627 | new_stmt = gimple_build_assign (new_temp, bitop1, vec_rhs2); |
12628 | else |
12629 | new_stmt = gimple_build_assign (new_temp, bitop1, vec_rhs1, |
12630 | vec_rhs2); |
12631 | vect_finish_stmt_generation (vinfo, stmt_info, vec_stmt: new_stmt, gsi); |
12632 | if (bitop2 != NOP_EXPR) |
12633 | { |
12634 | tree res = make_ssa_name (var: mask); |
12635 | if (bitop2 == BIT_NOT_EXPR) |
12636 | new_stmt = gimple_build_assign (res, bitop2, new_temp); |
12637 | else |
12638 | new_stmt = gimple_build_assign (res, bitop2, vec_rhs1, |
12639 | new_temp); |
12640 | vect_finish_stmt_generation (vinfo, stmt_info, vec_stmt: new_stmt, gsi); |
12641 | } |
12642 | } |
12643 | if (slp_node) |
12644 | slp_node->push_vec_def (def: new_stmt); |
12645 | else |
12646 | STMT_VINFO_VEC_STMTS (stmt_info).safe_push (obj: new_stmt); |
12647 | } |
12648 | |
12649 | if (!slp_node) |
12650 | *vec_stmt = STMT_VINFO_VEC_STMTS (stmt_info)[0]; |
12651 | |
12652 | vec_oprnds0.release (); |
12653 | vec_oprnds1.release (); |
12654 | |
12655 | return true; |
12656 | } |
12657 | |
12658 | /* vectorizable_comparison. |
12659 | |
12660 | Check if STMT_INFO is comparison expression that can be vectorized. |
12661 | If VEC_STMT is also passed, vectorize STMT_INFO: create a vectorized |
12662 | comparison, put it in VEC_STMT, and insert it at GSI. |
12663 | |
12664 | Return true if STMT_INFO is vectorizable in this way. */ |
12665 | |
12666 | static bool |
12667 | vectorizable_comparison (vec_info *vinfo, |
12668 | stmt_vec_info stmt_info, gimple_stmt_iterator *gsi, |
12669 | gimple **vec_stmt, |
12670 | slp_tree slp_node, stmt_vector_for_cost *cost_vec) |
12671 | { |
12672 | bb_vec_info bb_vinfo = dyn_cast <bb_vec_info> (p: vinfo); |
12673 | |
12674 | if (!STMT_VINFO_RELEVANT_P (stmt_info) && !bb_vinfo) |
12675 | return false; |
12676 | |
12677 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def) |
12678 | return false; |
12679 | |
12680 | gassign *stmt = dyn_cast <gassign *> (p: stmt_info->stmt); |
12681 | if (!stmt) |
12682 | return false; |
12683 | |
12684 | enum tree_code code = gimple_assign_rhs_code (gs: stmt); |
12685 | tree vectype = STMT_VINFO_VECTYPE (stmt_info); |
12686 | if (!vectorizable_comparison_1 (vinfo, vectype, stmt_info, code, gsi, |
12687 | vec_stmt, slp_node, cost_vec)) |
12688 | return false; |
12689 | |
12690 | if (!vec_stmt) |
12691 | STMT_VINFO_TYPE (stmt_info) = comparison_vec_info_type; |
12692 | |
12693 | return true; |
12694 | } |
12695 | |
12696 | /* If SLP_NODE is nonnull, return true if vectorizable_live_operation |
12697 | can handle all live statements in the node. Otherwise return true |
12698 | if STMT_INFO is not live or if vectorizable_live_operation can handle it. |
12699 | VEC_STMT_P is as for vectorizable_live_operation. */ |
12700 | |
12701 | static bool |
12702 | can_vectorize_live_stmts (vec_info *vinfo, stmt_vec_info stmt_info, |
12703 | slp_tree slp_node, slp_instance slp_node_instance, |
12704 | bool vec_stmt_p, |
12705 | stmt_vector_for_cost *cost_vec) |
12706 | { |
12707 | if (slp_node) |
12708 | { |
12709 | stmt_vec_info slp_stmt_info; |
12710 | unsigned int i; |
12711 | FOR_EACH_VEC_ELT (SLP_TREE_SCALAR_STMTS (slp_node), i, slp_stmt_info) |
12712 | { |
12713 | if (STMT_VINFO_LIVE_P (slp_stmt_info) |
12714 | && !vectorizable_live_operation (vinfo, slp_stmt_info, slp_node, |
12715 | slp_node_instance, i, |
12716 | vec_stmt_p, cost_vec)) |
12717 | return false; |
12718 | } |
12719 | } |
12720 | else if (STMT_VINFO_LIVE_P (stmt_info) |
12721 | && !vectorizable_live_operation (vinfo, stmt_info, |
12722 | slp_node, slp_node_instance, -1, |
12723 | vec_stmt_p, cost_vec)) |
12724 | return false; |
12725 | |
12726 | return true; |
12727 | } |
12728 | |
12729 | /* Make sure the statement is vectorizable. */ |
12730 | |
12731 | opt_result |
12732 | vect_analyze_stmt (vec_info *vinfo, |
12733 | stmt_vec_info stmt_info, bool *need_to_vectorize, |
12734 | slp_tree node, slp_instance node_instance, |
12735 | stmt_vector_for_cost *cost_vec) |
12736 | { |
12737 | bb_vec_info bb_vinfo = dyn_cast <bb_vec_info> (p: vinfo); |
12738 | enum vect_relevant relevance = STMT_VINFO_RELEVANT (stmt_info); |
12739 | bool ok; |
12740 | gimple_seq pattern_def_seq; |
12741 | |
12742 | if (dump_enabled_p ()) |
12743 | dump_printf_loc (MSG_NOTE, vect_location, "==> examining statement: %G" , |
12744 | stmt_info->stmt); |
12745 | |
12746 | if (gimple_has_volatile_ops (stmt: stmt_info->stmt)) |
12747 | return opt_result::failure_at (loc: stmt_info->stmt, |
12748 | fmt: "not vectorized:" |
12749 | " stmt has volatile operands: %G\n" , |
12750 | stmt_info->stmt); |
12751 | |
12752 | if (STMT_VINFO_IN_PATTERN_P (stmt_info) |
12753 | && node == NULL |
12754 | && (pattern_def_seq = STMT_VINFO_PATTERN_DEF_SEQ (stmt_info))) |
12755 | { |
12756 | gimple_stmt_iterator si; |
12757 | |
12758 | for (si = gsi_start (seq&: pattern_def_seq); !gsi_end_p (i: si); gsi_next (i: &si)) |
12759 | { |
12760 | stmt_vec_info pattern_def_stmt_info |
12761 | = vinfo->lookup_stmt (gsi_stmt (i: si)); |
12762 | if (STMT_VINFO_RELEVANT_P (pattern_def_stmt_info) |
12763 | || STMT_VINFO_LIVE_P (pattern_def_stmt_info)) |
12764 | { |
12765 | /* Analyze def stmt of STMT if it's a pattern stmt. */ |
12766 | if (dump_enabled_p ()) |
12767 | dump_printf_loc (MSG_NOTE, vect_location, |
12768 | "==> examining pattern def statement: %G" , |
12769 | pattern_def_stmt_info->stmt); |
12770 | |
12771 | opt_result res |
12772 | = vect_analyze_stmt (vinfo, stmt_info: pattern_def_stmt_info, |
12773 | need_to_vectorize, node, node_instance, |
12774 | cost_vec); |
12775 | if (!res) |
12776 | return res; |
12777 | } |
12778 | } |
12779 | } |
12780 | |
12781 | /* Skip stmts that do not need to be vectorized. In loops this is expected |
12782 | to include: |
12783 | - the COND_EXPR which is the loop exit condition |
12784 | - any LABEL_EXPRs in the loop |
12785 | - computations that are used only for array indexing or loop control. |
12786 | In basic blocks we only analyze statements that are a part of some SLP |
12787 | instance, therefore, all the statements are relevant. |
12788 | |
12789 | Pattern statement needs to be analyzed instead of the original statement |
12790 | if the original statement is not relevant. Otherwise, we analyze both |
12791 | statements. In basic blocks we are called from some SLP instance |
12792 | traversal, don't analyze pattern stmts instead, the pattern stmts |
12793 | already will be part of SLP instance. */ |
12794 | |
12795 | stmt_vec_info pattern_stmt_info = STMT_VINFO_RELATED_STMT (stmt_info); |
12796 | if (!STMT_VINFO_RELEVANT_P (stmt_info) |
12797 | && !STMT_VINFO_LIVE_P (stmt_info)) |
12798 | { |
12799 | if (STMT_VINFO_IN_PATTERN_P (stmt_info) |
12800 | && pattern_stmt_info |
12801 | && (STMT_VINFO_RELEVANT_P (pattern_stmt_info) |
12802 | || STMT_VINFO_LIVE_P (pattern_stmt_info))) |
12803 | { |
12804 | /* Analyze PATTERN_STMT instead of the original stmt. */ |
12805 | stmt_info = pattern_stmt_info; |
12806 | if (dump_enabled_p ()) |
12807 | dump_printf_loc (MSG_NOTE, vect_location, |
12808 | "==> examining pattern statement: %G" , |
12809 | stmt_info->stmt); |
12810 | } |
12811 | else |
12812 | { |
12813 | if (dump_enabled_p ()) |
12814 | dump_printf_loc (MSG_NOTE, vect_location, "irrelevant.\n" ); |
12815 | |
12816 | return opt_result::success (); |
12817 | } |
12818 | } |
12819 | else if (STMT_VINFO_IN_PATTERN_P (stmt_info) |
12820 | && node == NULL |
12821 | && pattern_stmt_info |
12822 | && (STMT_VINFO_RELEVANT_P (pattern_stmt_info) |
12823 | || STMT_VINFO_LIVE_P (pattern_stmt_info))) |
12824 | { |
12825 | /* Analyze PATTERN_STMT too. */ |
12826 | if (dump_enabled_p ()) |
12827 | dump_printf_loc (MSG_NOTE, vect_location, |
12828 | "==> examining pattern statement: %G" , |
12829 | pattern_stmt_info->stmt); |
12830 | |
12831 | opt_result res |
12832 | = vect_analyze_stmt (vinfo, stmt_info: pattern_stmt_info, need_to_vectorize, node, |
12833 | node_instance, cost_vec); |
12834 | if (!res) |
12835 | return res; |
12836 | } |
12837 | |
12838 | switch (STMT_VINFO_DEF_TYPE (stmt_info)) |
12839 | { |
12840 | case vect_internal_def: |
12841 | break; |
12842 | |
12843 | case vect_reduction_def: |
12844 | case vect_nested_cycle: |
12845 | gcc_assert (!bb_vinfo |
12846 | && (relevance == vect_used_in_outer |
12847 | || relevance == vect_used_in_outer_by_reduction |
12848 | || relevance == vect_used_by_reduction |
12849 | || relevance == vect_unused_in_scope |
12850 | || relevance == vect_used_only_live)); |
12851 | break; |
12852 | |
12853 | case vect_induction_def: |
12854 | case vect_first_order_recurrence: |
12855 | gcc_assert (!bb_vinfo); |
12856 | break; |
12857 | |
12858 | case vect_constant_def: |
12859 | case vect_external_def: |
12860 | case vect_unknown_def_type: |
12861 | default: |
12862 | gcc_unreachable (); |
12863 | } |
12864 | |
12865 | tree saved_vectype = STMT_VINFO_VECTYPE (stmt_info); |
12866 | if (node) |
12867 | STMT_VINFO_VECTYPE (stmt_info) = SLP_TREE_VECTYPE (node); |
12868 | |
12869 | if (STMT_VINFO_RELEVANT_P (stmt_info)) |
12870 | { |
12871 | gcall *call = dyn_cast <gcall *> (p: stmt_info->stmt); |
12872 | gcc_assert (STMT_VINFO_VECTYPE (stmt_info) |
12873 | || (call && gimple_call_lhs (call) == NULL_TREE)); |
12874 | *need_to_vectorize = true; |
12875 | } |
12876 | |
12877 | if (PURE_SLP_STMT (stmt_info) && !node) |
12878 | { |
12879 | if (dump_enabled_p ()) |
12880 | dump_printf_loc (MSG_NOTE, vect_location, |
12881 | "handled only by SLP analysis\n" ); |
12882 | return opt_result::success (); |
12883 | } |
12884 | |
12885 | ok = true; |
12886 | if (!bb_vinfo |
12887 | && (STMT_VINFO_RELEVANT_P (stmt_info) |
12888 | || STMT_VINFO_DEF_TYPE (stmt_info) == vect_reduction_def)) |
12889 | /* Prefer vectorizable_call over vectorizable_simd_clone_call so |
12890 | -mveclibabi= takes preference over library functions with |
12891 | the simd attribute. */ |
12892 | ok = (vectorizable_call (vinfo, stmt_info, NULL, NULL, slp_node: node, cost_vec) |
12893 | || vectorizable_simd_clone_call (vinfo, stmt_info, NULL, NULL, slp_node: node, |
12894 | cost_vec) |
12895 | || vectorizable_conversion (vinfo, stmt_info, |
12896 | NULL, NULL, slp_node: node, cost_vec) |
12897 | || vectorizable_operation (vinfo, stmt_info, |
12898 | NULL, NULL, slp_node: node, cost_vec) |
12899 | || vectorizable_assignment (vinfo, stmt_info, |
12900 | NULL, NULL, slp_node: node, cost_vec) |
12901 | || vectorizable_load (vinfo, stmt_info, NULL, NULL, slp_node: node, cost_vec) |
12902 | || vectorizable_store (vinfo, stmt_info, NULL, NULL, slp_node: node, cost_vec) |
12903 | || vectorizable_reduction (as_a <loop_vec_info> (p: vinfo), stmt_info, |
12904 | node, node_instance, cost_vec) |
12905 | || vectorizable_induction (as_a <loop_vec_info> (p: vinfo), stmt_info, |
12906 | NULL, node, cost_vec) |
12907 | || vectorizable_shift (vinfo, stmt_info, NULL, NULL, slp_node: node, cost_vec) |
12908 | || vectorizable_condition (vinfo, stmt_info, |
12909 | NULL, NULL, slp_node: node, cost_vec) |
12910 | || vectorizable_comparison (vinfo, stmt_info, NULL, NULL, slp_node: node, |
12911 | cost_vec) |
12912 | || vectorizable_lc_phi (as_a <loop_vec_info> (p: vinfo), |
12913 | stmt_info, NULL, node) |
12914 | || vectorizable_recurr (as_a <loop_vec_info> (p: vinfo), |
12915 | stmt_info, NULL, node, cost_vec)); |
12916 | else |
12917 | { |
12918 | if (bb_vinfo) |
12919 | ok = (vectorizable_call (vinfo, stmt_info, NULL, NULL, slp_node: node, cost_vec) |
12920 | || vectorizable_simd_clone_call (vinfo, stmt_info, |
12921 | NULL, NULL, slp_node: node, cost_vec) |
12922 | || vectorizable_conversion (vinfo, stmt_info, NULL, NULL, slp_node: node, |
12923 | cost_vec) |
12924 | || vectorizable_shift (vinfo, stmt_info, |
12925 | NULL, NULL, slp_node: node, cost_vec) |
12926 | || vectorizable_operation (vinfo, stmt_info, |
12927 | NULL, NULL, slp_node: node, cost_vec) |
12928 | || vectorizable_assignment (vinfo, stmt_info, NULL, NULL, slp_node: node, |
12929 | cost_vec) |
12930 | || vectorizable_load (vinfo, stmt_info, |
12931 | NULL, NULL, slp_node: node, cost_vec) |
12932 | || vectorizable_store (vinfo, stmt_info, |
12933 | NULL, NULL, slp_node: node, cost_vec) |
12934 | || vectorizable_condition (vinfo, stmt_info, |
12935 | NULL, NULL, slp_node: node, cost_vec) |
12936 | || vectorizable_comparison (vinfo, stmt_info, NULL, NULL, slp_node: node, |
12937 | cost_vec) |
12938 | || vectorizable_phi (vinfo, stmt_info, NULL, node, cost_vec)); |
12939 | } |
12940 | |
12941 | if (node) |
12942 | STMT_VINFO_VECTYPE (stmt_info) = saved_vectype; |
12943 | |
12944 | if (!ok) |
12945 | return opt_result::failure_at (loc: stmt_info->stmt, |
12946 | fmt: "not vectorized:" |
12947 | " relevant stmt not supported: %G" , |
12948 | stmt_info->stmt); |
12949 | |
12950 | /* Stmts that are (also) "live" (i.e. - that are used out of the loop) |
12951 | need extra handling, except for vectorizable reductions. */ |
12952 | if (!bb_vinfo |
12953 | && STMT_VINFO_TYPE (stmt_info) != reduc_vec_info_type |
12954 | && STMT_VINFO_TYPE (stmt_info) != lc_phi_info_type |
12955 | && !can_vectorize_live_stmts (vinfo: as_a <loop_vec_info> (p: vinfo), |
12956 | stmt_info, slp_node: node, slp_node_instance: node_instance, |
12957 | vec_stmt_p: false, cost_vec)) |
12958 | return opt_result::failure_at (loc: stmt_info->stmt, |
12959 | fmt: "not vectorized:" |
12960 | " live stmt not supported: %G" , |
12961 | stmt_info->stmt); |
12962 | |
12963 | return opt_result::success (); |
12964 | } |
12965 | |
12966 | |
12967 | /* Function vect_transform_stmt. |
12968 | |
12969 | Create a vectorized stmt to replace STMT_INFO, and insert it at GSI. */ |
12970 | |
12971 | bool |
12972 | vect_transform_stmt (vec_info *vinfo, |
12973 | stmt_vec_info stmt_info, gimple_stmt_iterator *gsi, |
12974 | slp_tree slp_node, slp_instance slp_node_instance) |
12975 | { |
12976 | bool is_store = false; |
12977 | gimple *vec_stmt = NULL; |
12978 | bool done; |
12979 | |
12980 | gcc_assert (slp_node || !PURE_SLP_STMT (stmt_info)); |
12981 | |
12982 | tree saved_vectype = STMT_VINFO_VECTYPE (stmt_info); |
12983 | if (slp_node) |
12984 | STMT_VINFO_VECTYPE (stmt_info) = SLP_TREE_VECTYPE (slp_node); |
12985 | |
12986 | switch (STMT_VINFO_TYPE (stmt_info)) |
12987 | { |
12988 | case type_demotion_vec_info_type: |
12989 | case type_promotion_vec_info_type: |
12990 | case type_conversion_vec_info_type: |
12991 | done = vectorizable_conversion (vinfo, stmt_info, |
12992 | gsi, vec_stmt: &vec_stmt, slp_node, NULL); |
12993 | gcc_assert (done); |
12994 | break; |
12995 | |
12996 | case induc_vec_info_type: |
12997 | done = vectorizable_induction (as_a <loop_vec_info> (p: vinfo), |
12998 | stmt_info, &vec_stmt, slp_node, |
12999 | NULL); |
13000 | gcc_assert (done); |
13001 | break; |
13002 | |
13003 | case shift_vec_info_type: |
13004 | done = vectorizable_shift (vinfo, stmt_info, |
13005 | gsi, vec_stmt: &vec_stmt, slp_node, NULL); |
13006 | gcc_assert (done); |
13007 | break; |
13008 | |
13009 | case op_vec_info_type: |
13010 | done = vectorizable_operation (vinfo, stmt_info, gsi, vec_stmt: &vec_stmt, slp_node, |
13011 | NULL); |
13012 | gcc_assert (done); |
13013 | break; |
13014 | |
13015 | case assignment_vec_info_type: |
13016 | done = vectorizable_assignment (vinfo, stmt_info, |
13017 | gsi, vec_stmt: &vec_stmt, slp_node, NULL); |
13018 | gcc_assert (done); |
13019 | break; |
13020 | |
13021 | case load_vec_info_type: |
13022 | done = vectorizable_load (vinfo, stmt_info, gsi, vec_stmt: &vec_stmt, slp_node, |
13023 | NULL); |
13024 | gcc_assert (done); |
13025 | break; |
13026 | |
13027 | case store_vec_info_type: |
13028 | if (STMT_VINFO_GROUPED_ACCESS (stmt_info) |
13029 | && !slp_node |
13030 | && (++DR_GROUP_STORE_COUNT (DR_GROUP_FIRST_ELEMENT (stmt_info)) |
13031 | < DR_GROUP_SIZE (DR_GROUP_FIRST_ELEMENT (stmt_info)))) |
13032 | /* In case of interleaving, the whole chain is vectorized when the |
13033 | last store in the chain is reached. Store stmts before the last |
13034 | one are skipped, and there vec_stmt_info shouldn't be freed |
13035 | meanwhile. */ |
13036 | ; |
13037 | else |
13038 | { |
13039 | done = vectorizable_store (vinfo, stmt_info, |
13040 | gsi, vec_stmt: &vec_stmt, slp_node, NULL); |
13041 | gcc_assert (done); |
13042 | is_store = true; |
13043 | } |
13044 | break; |
13045 | |
13046 | case condition_vec_info_type: |
13047 | done = vectorizable_condition (vinfo, stmt_info, |
13048 | gsi, vec_stmt: &vec_stmt, slp_node, NULL); |
13049 | gcc_assert (done); |
13050 | break; |
13051 | |
13052 | case comparison_vec_info_type: |
13053 | done = vectorizable_comparison (vinfo, stmt_info, gsi, vec_stmt: &vec_stmt, |
13054 | slp_node, NULL); |
13055 | gcc_assert (done); |
13056 | break; |
13057 | |
13058 | case call_vec_info_type: |
13059 | done = vectorizable_call (vinfo, stmt_info, |
13060 | gsi, vec_stmt: &vec_stmt, slp_node, NULL); |
13061 | break; |
13062 | |
13063 | case call_simd_clone_vec_info_type: |
13064 | done = vectorizable_simd_clone_call (vinfo, stmt_info, gsi, vec_stmt: &vec_stmt, |
13065 | slp_node, NULL); |
13066 | break; |
13067 | |
13068 | case reduc_vec_info_type: |
13069 | done = vect_transform_reduction (as_a <loop_vec_info> (p: vinfo), stmt_info, |
13070 | gsi, &vec_stmt, slp_node); |
13071 | gcc_assert (done); |
13072 | break; |
13073 | |
13074 | case cycle_phi_info_type: |
13075 | done = vect_transform_cycle_phi (as_a <loop_vec_info> (p: vinfo), stmt_info, |
13076 | &vec_stmt, slp_node, slp_node_instance); |
13077 | gcc_assert (done); |
13078 | break; |
13079 | |
13080 | case lc_phi_info_type: |
13081 | done = vectorizable_lc_phi (as_a <loop_vec_info> (p: vinfo), |
13082 | stmt_info, &vec_stmt, slp_node); |
13083 | gcc_assert (done); |
13084 | break; |
13085 | |
13086 | case recurr_info_type: |
13087 | done = vectorizable_recurr (as_a <loop_vec_info> (p: vinfo), |
13088 | stmt_info, &vec_stmt, slp_node, NULL); |
13089 | gcc_assert (done); |
13090 | break; |
13091 | |
13092 | case phi_info_type: |
13093 | done = vectorizable_phi (vinfo, stmt_info, &vec_stmt, slp_node, NULL); |
13094 | gcc_assert (done); |
13095 | break; |
13096 | |
13097 | default: |
13098 | if (!STMT_VINFO_LIVE_P (stmt_info)) |
13099 | { |
13100 | if (dump_enabled_p ()) |
13101 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
13102 | "stmt not supported.\n" ); |
13103 | gcc_unreachable (); |
13104 | } |
13105 | done = true; |
13106 | } |
13107 | |
13108 | if (!slp_node && vec_stmt) |
13109 | gcc_assert (STMT_VINFO_VEC_STMTS (stmt_info).exists ()); |
13110 | |
13111 | if (STMT_VINFO_TYPE (stmt_info) != store_vec_info_type) |
13112 | { |
13113 | /* Handle stmts whose DEF is used outside the loop-nest that is |
13114 | being vectorized. */ |
13115 | done = can_vectorize_live_stmts (vinfo, stmt_info, slp_node, |
13116 | slp_node_instance, vec_stmt_p: true, NULL); |
13117 | gcc_assert (done); |
13118 | } |
13119 | |
13120 | if (slp_node) |
13121 | STMT_VINFO_VECTYPE (stmt_info) = saved_vectype; |
13122 | |
13123 | return is_store; |
13124 | } |
13125 | |
13126 | |
13127 | /* Remove a group of stores (for SLP or interleaving), free their |
13128 | stmt_vec_info. */ |
13129 | |
13130 | void |
13131 | vect_remove_stores (vec_info *vinfo, stmt_vec_info first_stmt_info) |
13132 | { |
13133 | stmt_vec_info next_stmt_info = first_stmt_info; |
13134 | |
13135 | while (next_stmt_info) |
13136 | { |
13137 | stmt_vec_info tmp = DR_GROUP_NEXT_ELEMENT (next_stmt_info); |
13138 | next_stmt_info = vect_orig_stmt (stmt_info: next_stmt_info); |
13139 | /* Free the attached stmt_vec_info and remove the stmt. */ |
13140 | vinfo->remove_stmt (next_stmt_info); |
13141 | next_stmt_info = tmp; |
13142 | } |
13143 | } |
13144 | |
13145 | /* If NUNITS is nonzero, return a vector type that contains NUNITS |
13146 | elements of type SCALAR_TYPE, or null if the target doesn't support |
13147 | such a type. |
13148 | |
13149 | If NUNITS is zero, return a vector type that contains elements of |
13150 | type SCALAR_TYPE, choosing whichever vector size the target prefers. |
13151 | |
13152 | If PREVAILING_MODE is VOIDmode, we have not yet chosen a vector mode |
13153 | for this vectorization region and want to "autodetect" the best choice. |
13154 | Otherwise, PREVAILING_MODE is a previously-chosen vector TYPE_MODE |
13155 | and we want the new type to be interoperable with it. PREVAILING_MODE |
13156 | in this case can be a scalar integer mode or a vector mode; when it |
13157 | is a vector mode, the function acts like a tree-level version of |
13158 | related_vector_mode. */ |
13159 | |
13160 | tree |
13161 | get_related_vectype_for_scalar_type (machine_mode prevailing_mode, |
13162 | tree scalar_type, poly_uint64 nunits) |
13163 | { |
13164 | tree orig_scalar_type = scalar_type; |
13165 | scalar_mode inner_mode; |
13166 | machine_mode simd_mode; |
13167 | tree vectype; |
13168 | |
13169 | if ((!INTEGRAL_TYPE_P (scalar_type) |
13170 | && !POINTER_TYPE_P (scalar_type) |
13171 | && !SCALAR_FLOAT_TYPE_P (scalar_type)) |
13172 | || (!is_int_mode (TYPE_MODE (scalar_type), int_mode: &inner_mode) |
13173 | && !is_float_mode (TYPE_MODE (scalar_type), float_mode: &inner_mode))) |
13174 | return NULL_TREE; |
13175 | |
13176 | unsigned int nbytes = GET_MODE_SIZE (mode: inner_mode); |
13177 | |
13178 | /* Interoperability between modes requires one to be a constant multiple |
13179 | of the other, so that the number of vectors required for each operation |
13180 | is a compile-time constant. */ |
13181 | if (prevailing_mode != VOIDmode |
13182 | && !constant_multiple_p (a: nunits * nbytes, |
13183 | b: GET_MODE_SIZE (mode: prevailing_mode)) |
13184 | && !constant_multiple_p (a: GET_MODE_SIZE (mode: prevailing_mode), |
13185 | b: nunits * nbytes)) |
13186 | return NULL_TREE; |
13187 | |
13188 | /* For vector types of elements whose mode precision doesn't |
13189 | match their types precision we use a element type of mode |
13190 | precision. The vectorization routines will have to make sure |
13191 | they support the proper result truncation/extension. |
13192 | We also make sure to build vector types with INTEGER_TYPE |
13193 | component type only. */ |
13194 | if (INTEGRAL_TYPE_P (scalar_type) |
13195 | && (GET_MODE_BITSIZE (mode: inner_mode) != TYPE_PRECISION (scalar_type) |
13196 | || TREE_CODE (scalar_type) != INTEGER_TYPE)) |
13197 | scalar_type = build_nonstandard_integer_type (GET_MODE_BITSIZE (mode: inner_mode), |
13198 | TYPE_UNSIGNED (scalar_type)); |
13199 | |
13200 | /* We shouldn't end up building VECTOR_TYPEs of non-scalar components. |
13201 | When the component mode passes the above test simply use a type |
13202 | corresponding to that mode. The theory is that any use that |
13203 | would cause problems with this will disable vectorization anyway. */ |
13204 | else if (!SCALAR_FLOAT_TYPE_P (scalar_type) |
13205 | && !INTEGRAL_TYPE_P (scalar_type)) |
13206 | scalar_type = lang_hooks.types.type_for_mode (inner_mode, 1); |
13207 | |
13208 | /* We can't build a vector type of elements with alignment bigger than |
13209 | their size. */ |
13210 | else if (nbytes < TYPE_ALIGN_UNIT (scalar_type)) |
13211 | scalar_type = lang_hooks.types.type_for_mode (inner_mode, |
13212 | TYPE_UNSIGNED (scalar_type)); |
13213 | |
13214 | /* If we felt back to using the mode fail if there was |
13215 | no scalar type for it. */ |
13216 | if (scalar_type == NULL_TREE) |
13217 | return NULL_TREE; |
13218 | |
13219 | /* If no prevailing mode was supplied, use the mode the target prefers. |
13220 | Otherwise lookup a vector mode based on the prevailing mode. */ |
13221 | if (prevailing_mode == VOIDmode) |
13222 | { |
13223 | gcc_assert (known_eq (nunits, 0U)); |
13224 | simd_mode = targetm.vectorize.preferred_simd_mode (inner_mode); |
13225 | if (SCALAR_INT_MODE_P (simd_mode)) |
13226 | { |
13227 | /* Traditional behavior is not to take the integer mode |
13228 | literally, but simply to use it as a way of determining |
13229 | the vector size. It is up to mode_for_vector to decide |
13230 | what the TYPE_MODE should be. |
13231 | |
13232 | Note that nunits == 1 is allowed in order to support single |
13233 | element vector types. */ |
13234 | if (!multiple_p (a: GET_MODE_SIZE (mode: simd_mode), b: nbytes, multiple: &nunits) |
13235 | || !mode_for_vector (inner_mode, nunits).exists (mode: &simd_mode)) |
13236 | return NULL_TREE; |
13237 | } |
13238 | } |
13239 | else if (SCALAR_INT_MODE_P (prevailing_mode) |
13240 | || !related_vector_mode (prevailing_mode, |
13241 | inner_mode, nunits).exists (mode: &simd_mode)) |
13242 | { |
13243 | /* Fall back to using mode_for_vector, mostly in the hope of being |
13244 | able to use an integer mode. */ |
13245 | if (known_eq (nunits, 0U) |
13246 | && !multiple_p (a: GET_MODE_SIZE (mode: prevailing_mode), b: nbytes, multiple: &nunits)) |
13247 | return NULL_TREE; |
13248 | |
13249 | if (!mode_for_vector (inner_mode, nunits).exists (mode: &simd_mode)) |
13250 | return NULL_TREE; |
13251 | } |
13252 | |
13253 | vectype = build_vector_type_for_mode (scalar_type, simd_mode); |
13254 | |
13255 | /* In cases where the mode was chosen by mode_for_vector, check that |
13256 | the target actually supports the chosen mode, or that it at least |
13257 | allows the vector mode to be replaced by a like-sized integer. */ |
13258 | if (!VECTOR_MODE_P (TYPE_MODE (vectype)) |
13259 | && !INTEGRAL_MODE_P (TYPE_MODE (vectype))) |
13260 | return NULL_TREE; |
13261 | |
13262 | /* Re-attach the address-space qualifier if we canonicalized the scalar |
13263 | type. */ |
13264 | if (TYPE_ADDR_SPACE (orig_scalar_type) != TYPE_ADDR_SPACE (vectype)) |
13265 | return build_qualified_type |
13266 | (vectype, KEEP_QUAL_ADDR_SPACE (TYPE_QUALS (orig_scalar_type))); |
13267 | |
13268 | return vectype; |
13269 | } |
13270 | |
13271 | /* Function get_vectype_for_scalar_type. |
13272 | |
13273 | Returns the vector type corresponding to SCALAR_TYPE as supported |
13274 | by the target. If GROUP_SIZE is nonzero and we're performing BB |
13275 | vectorization, make sure that the number of elements in the vector |
13276 | is no bigger than GROUP_SIZE. */ |
13277 | |
13278 | tree |
13279 | get_vectype_for_scalar_type (vec_info *vinfo, tree scalar_type, |
13280 | unsigned int group_size) |
13281 | { |
13282 | /* For BB vectorization, we should always have a group size once we've |
13283 | constructed the SLP tree; the only valid uses of zero GROUP_SIZEs |
13284 | are tentative requests during things like early data reference |
13285 | analysis and pattern recognition. */ |
13286 | if (is_a <bb_vec_info> (p: vinfo)) |
13287 | gcc_assert (vinfo->slp_instances.is_empty () || group_size != 0); |
13288 | else |
13289 | group_size = 0; |
13290 | |
13291 | tree vectype = get_related_vectype_for_scalar_type (prevailing_mode: vinfo->vector_mode, |
13292 | scalar_type); |
13293 | if (vectype && vinfo->vector_mode == VOIDmode) |
13294 | vinfo->vector_mode = TYPE_MODE (vectype); |
13295 | |
13296 | /* Register the natural choice of vector type, before the group size |
13297 | has been applied. */ |
13298 | if (vectype) |
13299 | vinfo->used_vector_modes.add (TYPE_MODE (vectype)); |
13300 | |
13301 | /* If the natural choice of vector type doesn't satisfy GROUP_SIZE, |
13302 | try again with an explicit number of elements. */ |
13303 | if (vectype |
13304 | && group_size |
13305 | && maybe_ge (TYPE_VECTOR_SUBPARTS (vectype), group_size)) |
13306 | { |
13307 | /* Start with the biggest number of units that fits within |
13308 | GROUP_SIZE and halve it until we find a valid vector type. |
13309 | Usually either the first attempt will succeed or all will |
13310 | fail (in the latter case because GROUP_SIZE is too small |
13311 | for the target), but it's possible that a target could have |
13312 | a hole between supported vector types. |
13313 | |
13314 | If GROUP_SIZE is not a power of 2, this has the effect of |
13315 | trying the largest power of 2 that fits within the group, |
13316 | even though the group is not a multiple of that vector size. |
13317 | The BB vectorizer will then try to carve up the group into |
13318 | smaller pieces. */ |
13319 | unsigned int nunits = 1 << floor_log2 (x: group_size); |
13320 | do |
13321 | { |
13322 | vectype = get_related_vectype_for_scalar_type (prevailing_mode: vinfo->vector_mode, |
13323 | scalar_type, nunits); |
13324 | nunits /= 2; |
13325 | } |
13326 | while (nunits > 1 && !vectype); |
13327 | } |
13328 | |
13329 | return vectype; |
13330 | } |
13331 | |
13332 | /* Return the vector type corresponding to SCALAR_TYPE as supported |
13333 | by the target. NODE, if nonnull, is the SLP tree node that will |
13334 | use the returned vector type. */ |
13335 | |
13336 | tree |
13337 | get_vectype_for_scalar_type (vec_info *vinfo, tree scalar_type, slp_tree node) |
13338 | { |
13339 | unsigned int group_size = 0; |
13340 | if (node) |
13341 | group_size = SLP_TREE_LANES (node); |
13342 | return get_vectype_for_scalar_type (vinfo, scalar_type, group_size); |
13343 | } |
13344 | |
13345 | /* Function get_mask_type_for_scalar_type. |
13346 | |
13347 | Returns the mask type corresponding to a result of comparison |
13348 | of vectors of specified SCALAR_TYPE as supported by target. |
13349 | If GROUP_SIZE is nonzero and we're performing BB vectorization, |
13350 | make sure that the number of elements in the vector is no bigger |
13351 | than GROUP_SIZE. */ |
13352 | |
13353 | tree |
13354 | get_mask_type_for_scalar_type (vec_info *vinfo, tree scalar_type, |
13355 | unsigned int group_size) |
13356 | { |
13357 | tree vectype = get_vectype_for_scalar_type (vinfo, scalar_type, group_size); |
13358 | |
13359 | if (!vectype) |
13360 | return NULL; |
13361 | |
13362 | return truth_type_for (vectype); |
13363 | } |
13364 | |
13365 | /* Function get_mask_type_for_scalar_type. |
13366 | |
13367 | Returns the mask type corresponding to a result of comparison |
13368 | of vectors of specified SCALAR_TYPE as supported by target. |
13369 | NODE, if nonnull, is the SLP tree node that will use the returned |
13370 | vector type. */ |
13371 | |
13372 | tree |
13373 | get_mask_type_for_scalar_type (vec_info *vinfo, tree scalar_type, |
13374 | slp_tree node) |
13375 | { |
13376 | tree vectype = get_vectype_for_scalar_type (vinfo, scalar_type, node); |
13377 | |
13378 | if (!vectype) |
13379 | return NULL; |
13380 | |
13381 | return truth_type_for (vectype); |
13382 | } |
13383 | |
13384 | /* Function get_same_sized_vectype |
13385 | |
13386 | Returns a vector type corresponding to SCALAR_TYPE of size |
13387 | VECTOR_TYPE if supported by the target. */ |
13388 | |
13389 | tree |
13390 | get_same_sized_vectype (tree scalar_type, tree vector_type) |
13391 | { |
13392 | if (VECT_SCALAR_BOOLEAN_TYPE_P (scalar_type)) |
13393 | return truth_type_for (vector_type); |
13394 | |
13395 | poly_uint64 nunits; |
13396 | if (!multiple_p (a: GET_MODE_SIZE (TYPE_MODE (vector_type)), |
13397 | b: GET_MODE_SIZE (TYPE_MODE (scalar_type)), multiple: &nunits)) |
13398 | return NULL_TREE; |
13399 | |
13400 | return get_related_vectype_for_scalar_type (TYPE_MODE (vector_type), |
13401 | scalar_type, nunits); |
13402 | } |
13403 | |
13404 | /* Return true if replacing LOOP_VINFO->vector_mode with VECTOR_MODE |
13405 | would not change the chosen vector modes. */ |
13406 | |
13407 | bool |
13408 | vect_chooses_same_modes_p (vec_info *vinfo, machine_mode vector_mode) |
13409 | { |
13410 | for (vec_info::mode_set::iterator i = vinfo->used_vector_modes.begin (); |
13411 | i != vinfo->used_vector_modes.end (); ++i) |
13412 | if (!VECTOR_MODE_P (*i) |
13413 | || related_vector_mode (vector_mode, GET_MODE_INNER (*i), 0) != *i) |
13414 | return false; |
13415 | return true; |
13416 | } |
13417 | |
13418 | /* Function vect_is_simple_use. |
13419 | |
13420 | Input: |
13421 | VINFO - the vect info of the loop or basic block that is being vectorized. |
13422 | OPERAND - operand in the loop or bb. |
13423 | Output: |
13424 | DEF_STMT_INFO_OUT (optional) - information about the defining stmt in |
13425 | case OPERAND is an SSA_NAME that is defined in the vectorizable region |
13426 | DEF_STMT_OUT (optional) - the defining stmt in case OPERAND is an SSA_NAME; |
13427 | the definition could be anywhere in the function |
13428 | DT - the type of definition |
13429 | |
13430 | Returns whether a stmt with OPERAND can be vectorized. |
13431 | For loops, supportable operands are constants, loop invariants, and operands |
13432 | that are defined by the current iteration of the loop. Unsupportable |
13433 | operands are those that are defined by a previous iteration of the loop (as |
13434 | is the case in reduction/induction computations). |
13435 | For basic blocks, supportable operands are constants and bb invariants. |
13436 | For now, operands defined outside the basic block are not supported. */ |
13437 | |
13438 | bool |
13439 | vect_is_simple_use (tree operand, vec_info *vinfo, enum vect_def_type *dt, |
13440 | stmt_vec_info *def_stmt_info_out, gimple **def_stmt_out) |
13441 | { |
13442 | if (def_stmt_info_out) |
13443 | *def_stmt_info_out = NULL; |
13444 | if (def_stmt_out) |
13445 | *def_stmt_out = NULL; |
13446 | *dt = vect_unknown_def_type; |
13447 | |
13448 | if (dump_enabled_p ()) |
13449 | { |
13450 | dump_printf_loc (MSG_NOTE, vect_location, |
13451 | "vect_is_simple_use: operand " ); |
13452 | if (TREE_CODE (operand) == SSA_NAME |
13453 | && !SSA_NAME_IS_DEFAULT_DEF (operand)) |
13454 | dump_gimple_expr (MSG_NOTE, TDF_SLIM, SSA_NAME_DEF_STMT (operand), 0); |
13455 | else |
13456 | dump_generic_expr (MSG_NOTE, TDF_SLIM, operand); |
13457 | } |
13458 | |
13459 | if (CONSTANT_CLASS_P (operand)) |
13460 | *dt = vect_constant_def; |
13461 | else if (is_gimple_min_invariant (operand)) |
13462 | *dt = vect_external_def; |
13463 | else if (TREE_CODE (operand) != SSA_NAME) |
13464 | *dt = vect_unknown_def_type; |
13465 | else if (SSA_NAME_IS_DEFAULT_DEF (operand)) |
13466 | *dt = vect_external_def; |
13467 | else |
13468 | { |
13469 | gimple *def_stmt = SSA_NAME_DEF_STMT (operand); |
13470 | stmt_vec_info stmt_vinfo = vinfo->lookup_def (operand); |
13471 | if (!stmt_vinfo) |
13472 | *dt = vect_external_def; |
13473 | else |
13474 | { |
13475 | stmt_vinfo = vect_stmt_to_vectorize (stmt_info: stmt_vinfo); |
13476 | def_stmt = stmt_vinfo->stmt; |
13477 | *dt = STMT_VINFO_DEF_TYPE (stmt_vinfo); |
13478 | if (def_stmt_info_out) |
13479 | *def_stmt_info_out = stmt_vinfo; |
13480 | } |
13481 | if (def_stmt_out) |
13482 | *def_stmt_out = def_stmt; |
13483 | } |
13484 | |
13485 | if (dump_enabled_p ()) |
13486 | { |
13487 | dump_printf (MSG_NOTE, ", type of def: " ); |
13488 | switch (*dt) |
13489 | { |
13490 | case vect_uninitialized_def: |
13491 | dump_printf (MSG_NOTE, "uninitialized\n" ); |
13492 | break; |
13493 | case vect_constant_def: |
13494 | dump_printf (MSG_NOTE, "constant\n" ); |
13495 | break; |
13496 | case vect_external_def: |
13497 | dump_printf (MSG_NOTE, "external\n" ); |
13498 | break; |
13499 | case vect_internal_def: |
13500 | dump_printf (MSG_NOTE, "internal\n" ); |
13501 | break; |
13502 | case vect_induction_def: |
13503 | dump_printf (MSG_NOTE, "induction\n" ); |
13504 | break; |
13505 | case vect_reduction_def: |
13506 | dump_printf (MSG_NOTE, "reduction\n" ); |
13507 | break; |
13508 | case vect_double_reduction_def: |
13509 | dump_printf (MSG_NOTE, "double reduction\n" ); |
13510 | break; |
13511 | case vect_nested_cycle: |
13512 | dump_printf (MSG_NOTE, "nested cycle\n" ); |
13513 | break; |
13514 | case vect_first_order_recurrence: |
13515 | dump_printf (MSG_NOTE, "first order recurrence\n" ); |
13516 | break; |
13517 | case vect_unknown_def_type: |
13518 | dump_printf (MSG_NOTE, "unknown\n" ); |
13519 | break; |
13520 | } |
13521 | } |
13522 | |
13523 | if (*dt == vect_unknown_def_type) |
13524 | { |
13525 | if (dump_enabled_p ()) |
13526 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
13527 | "Unsupported pattern.\n" ); |
13528 | return false; |
13529 | } |
13530 | |
13531 | return true; |
13532 | } |
13533 | |
13534 | /* Function vect_is_simple_use. |
13535 | |
13536 | Same as vect_is_simple_use but also determines the vector operand |
13537 | type of OPERAND and stores it to *VECTYPE. If the definition of |
13538 | OPERAND is vect_uninitialized_def, vect_constant_def or |
13539 | vect_external_def *VECTYPE will be set to NULL_TREE and the caller |
13540 | is responsible to compute the best suited vector type for the |
13541 | scalar operand. */ |
13542 | |
13543 | bool |
13544 | vect_is_simple_use (tree operand, vec_info *vinfo, enum vect_def_type *dt, |
13545 | tree *vectype, stmt_vec_info *def_stmt_info_out, |
13546 | gimple **def_stmt_out) |
13547 | { |
13548 | stmt_vec_info def_stmt_info; |
13549 | gimple *def_stmt; |
13550 | if (!vect_is_simple_use (operand, vinfo, dt, def_stmt_info_out: &def_stmt_info, def_stmt_out: &def_stmt)) |
13551 | return false; |
13552 | |
13553 | if (def_stmt_out) |
13554 | *def_stmt_out = def_stmt; |
13555 | if (def_stmt_info_out) |
13556 | *def_stmt_info_out = def_stmt_info; |
13557 | |
13558 | /* Now get a vector type if the def is internal, otherwise supply |
13559 | NULL_TREE and leave it up to the caller to figure out a proper |
13560 | type for the use stmt. */ |
13561 | if (*dt == vect_internal_def |
13562 | || *dt == vect_induction_def |
13563 | || *dt == vect_reduction_def |
13564 | || *dt == vect_double_reduction_def |
13565 | || *dt == vect_nested_cycle |
13566 | || *dt == vect_first_order_recurrence) |
13567 | { |
13568 | *vectype = STMT_VINFO_VECTYPE (def_stmt_info); |
13569 | gcc_assert (*vectype != NULL_TREE); |
13570 | if (dump_enabled_p ()) |
13571 | dump_printf_loc (MSG_NOTE, vect_location, |
13572 | "vect_is_simple_use: vectype %T\n" , *vectype); |
13573 | } |
13574 | else if (*dt == vect_uninitialized_def |
13575 | || *dt == vect_constant_def |
13576 | || *dt == vect_external_def) |
13577 | *vectype = NULL_TREE; |
13578 | else |
13579 | gcc_unreachable (); |
13580 | |
13581 | return true; |
13582 | } |
13583 | |
13584 | /* Function vect_is_simple_use. |
13585 | |
13586 | Same as vect_is_simple_use but determines the operand by operand |
13587 | position OPERAND from either STMT or SLP_NODE, filling in *OP |
13588 | and *SLP_DEF (when SLP_NODE is not NULL). */ |
13589 | |
13590 | bool |
13591 | vect_is_simple_use (vec_info *vinfo, stmt_vec_info stmt, slp_tree slp_node, |
13592 | unsigned operand, tree *op, slp_tree *slp_def, |
13593 | enum vect_def_type *dt, |
13594 | tree *vectype, stmt_vec_info *def_stmt_info_out) |
13595 | { |
13596 | if (slp_node) |
13597 | { |
13598 | slp_tree child = SLP_TREE_CHILDREN (slp_node)[operand]; |
13599 | *slp_def = child; |
13600 | *vectype = SLP_TREE_VECTYPE (child); |
13601 | if (SLP_TREE_DEF_TYPE (child) == vect_internal_def) |
13602 | { |
13603 | *op = gimple_get_lhs (SLP_TREE_REPRESENTATIVE (child)->stmt); |
13604 | return vect_is_simple_use (operand: *op, vinfo, dt, def_stmt_info_out); |
13605 | } |
13606 | else |
13607 | { |
13608 | if (def_stmt_info_out) |
13609 | *def_stmt_info_out = NULL; |
13610 | *op = SLP_TREE_SCALAR_OPS (child)[0]; |
13611 | *dt = SLP_TREE_DEF_TYPE (child); |
13612 | return true; |
13613 | } |
13614 | } |
13615 | else |
13616 | { |
13617 | *slp_def = NULL; |
13618 | if (gassign *ass = dyn_cast <gassign *> (p: stmt->stmt)) |
13619 | { |
13620 | if (gimple_assign_rhs_code (gs: ass) == COND_EXPR |
13621 | && COMPARISON_CLASS_P (gimple_assign_rhs1 (ass))) |
13622 | { |
13623 | if (operand < 2) |
13624 | *op = TREE_OPERAND (gimple_assign_rhs1 (ass), operand); |
13625 | else |
13626 | *op = gimple_op (gs: ass, i: operand); |
13627 | } |
13628 | else if (gimple_assign_rhs_code (gs: ass) == VIEW_CONVERT_EXPR) |
13629 | *op = TREE_OPERAND (gimple_assign_rhs1 (ass), 0); |
13630 | else |
13631 | *op = gimple_op (gs: ass, i: operand + 1); |
13632 | } |
13633 | else if (gcall *call = dyn_cast <gcall *> (p: stmt->stmt)) |
13634 | *op = gimple_call_arg (gs: call, index: operand); |
13635 | else |
13636 | gcc_unreachable (); |
13637 | return vect_is_simple_use (operand: *op, vinfo, dt, vectype, def_stmt_info_out); |
13638 | } |
13639 | } |
13640 | |
13641 | /* If OP is not NULL and is external or constant update its vector |
13642 | type with VECTYPE. Returns true if successful or false if not, |
13643 | for example when conflicting vector types are present. */ |
13644 | |
13645 | bool |
13646 | vect_maybe_update_slp_op_vectype (slp_tree op, tree vectype) |
13647 | { |
13648 | if (!op || SLP_TREE_DEF_TYPE (op) == vect_internal_def) |
13649 | return true; |
13650 | if (SLP_TREE_VECTYPE (op)) |
13651 | return types_compatible_p (SLP_TREE_VECTYPE (op), type2: vectype); |
13652 | /* For external defs refuse to produce VECTOR_BOOLEAN_TYPE_P, those |
13653 | should be handled by patters. Allow vect_constant_def for now. */ |
13654 | if (VECTOR_BOOLEAN_TYPE_P (vectype) |
13655 | && SLP_TREE_DEF_TYPE (op) == vect_external_def) |
13656 | return false; |
13657 | SLP_TREE_VECTYPE (op) = vectype; |
13658 | return true; |
13659 | } |
13660 | |
13661 | /* Function supportable_widening_operation |
13662 | |
13663 | Check whether an operation represented by the code CODE is a |
13664 | widening operation that is supported by the target platform in |
13665 | vector form (i.e., when operating on arguments of type VECTYPE_IN |
13666 | producing a result of type VECTYPE_OUT). |
13667 | |
13668 | Widening operations we currently support are NOP (CONVERT), FLOAT, |
13669 | FIX_TRUNC and WIDEN_MULT. This function checks if these operations |
13670 | are supported by the target platform either directly (via vector |
13671 | tree-codes), or via target builtins. |
13672 | |
13673 | Output: |
13674 | - CODE1 and CODE2 are codes of vector operations to be used when |
13675 | vectorizing the operation, if available. |
13676 | - MULTI_STEP_CVT determines the number of required intermediate steps in |
13677 | case of multi-step conversion (like char->short->int - in that case |
13678 | MULTI_STEP_CVT will be 1). |
13679 | - INTERM_TYPES contains the intermediate type required to perform the |
13680 | widening operation (short in the above example). */ |
13681 | |
13682 | bool |
13683 | supportable_widening_operation (vec_info *vinfo, |
13684 | code_helper code, |
13685 | stmt_vec_info stmt_info, |
13686 | tree vectype_out, tree vectype_in, |
13687 | code_helper *code1, |
13688 | code_helper *code2, |
13689 | int *multi_step_cvt, |
13690 | vec<tree> *interm_types) |
13691 | { |
13692 | loop_vec_info loop_info = dyn_cast <loop_vec_info> (p: vinfo); |
13693 | class loop *vect_loop = NULL; |
13694 | machine_mode vec_mode; |
13695 | enum insn_code icode1, icode2; |
13696 | optab optab1 = unknown_optab, optab2 = unknown_optab; |
13697 | tree vectype = vectype_in; |
13698 | tree wide_vectype = vectype_out; |
13699 | tree_code c1 = MAX_TREE_CODES, c2 = MAX_TREE_CODES; |
13700 | int i; |
13701 | tree prev_type, intermediate_type; |
13702 | machine_mode intermediate_mode, prev_mode; |
13703 | optab optab3, optab4; |
13704 | |
13705 | *multi_step_cvt = 0; |
13706 | if (loop_info) |
13707 | vect_loop = LOOP_VINFO_LOOP (loop_info); |
13708 | |
13709 | switch (code.safe_as_tree_code ()) |
13710 | { |
13711 | case MAX_TREE_CODES: |
13712 | /* Don't set c1 and c2 if code is not a tree_code. */ |
13713 | break; |
13714 | |
13715 | case WIDEN_MULT_EXPR: |
13716 | /* The result of a vectorized widening operation usually requires |
13717 | two vectors (because the widened results do not fit into one vector). |
13718 | The generated vector results would normally be expected to be |
13719 | generated in the same order as in the original scalar computation, |
13720 | i.e. if 8 results are generated in each vector iteration, they are |
13721 | to be organized as follows: |
13722 | vect1: [res1,res2,res3,res4], |
13723 | vect2: [res5,res6,res7,res8]. |
13724 | |
13725 | However, in the special case that the result of the widening |
13726 | operation is used in a reduction computation only, the order doesn't |
13727 | matter (because when vectorizing a reduction we change the order of |
13728 | the computation). Some targets can take advantage of this and |
13729 | generate more efficient code. For example, targets like Altivec, |
13730 | that support widen_mult using a sequence of {mult_even,mult_odd} |
13731 | generate the following vectors: |
13732 | vect1: [res1,res3,res5,res7], |
13733 | vect2: [res2,res4,res6,res8]. |
13734 | |
13735 | When vectorizing outer-loops, we execute the inner-loop sequentially |
13736 | (each vectorized inner-loop iteration contributes to VF outer-loop |
13737 | iterations in parallel). We therefore don't allow to change the |
13738 | order of the computation in the inner-loop during outer-loop |
13739 | vectorization. */ |
13740 | /* TODO: Another case in which order doesn't *really* matter is when we |
13741 | widen and then contract again, e.g. (short)((int)x * y >> 8). |
13742 | Normally, pack_trunc performs an even/odd permute, whereas the |
13743 | repack from an even/odd expansion would be an interleave, which |
13744 | would be significantly simpler for e.g. AVX2. */ |
13745 | /* In any case, in order to avoid duplicating the code below, recurse |
13746 | on VEC_WIDEN_MULT_EVEN_EXPR. If it succeeds, all the return values |
13747 | are properly set up for the caller. If we fail, we'll continue with |
13748 | a VEC_WIDEN_MULT_LO/HI_EXPR check. */ |
13749 | if (vect_loop |
13750 | && STMT_VINFO_RELEVANT (stmt_info) == vect_used_by_reduction |
13751 | && !nested_in_vect_loop_p (loop: vect_loop, stmt_info) |
13752 | && supportable_widening_operation (vinfo, code: VEC_WIDEN_MULT_EVEN_EXPR, |
13753 | stmt_info, vectype_out, |
13754 | vectype_in, code1, |
13755 | code2, multi_step_cvt, |
13756 | interm_types)) |
13757 | { |
13758 | /* Elements in a vector with vect_used_by_reduction property cannot |
13759 | be reordered if the use chain with this property does not have the |
13760 | same operation. One such an example is s += a * b, where elements |
13761 | in a and b cannot be reordered. Here we check if the vector defined |
13762 | by STMT is only directly used in the reduction statement. */ |
13763 | tree lhs = gimple_assign_lhs (gs: stmt_info->stmt); |
13764 | stmt_vec_info use_stmt_info = loop_info->lookup_single_use (lhs); |
13765 | if (use_stmt_info |
13766 | && STMT_VINFO_DEF_TYPE (use_stmt_info) == vect_reduction_def) |
13767 | return true; |
13768 | } |
13769 | c1 = VEC_WIDEN_MULT_LO_EXPR; |
13770 | c2 = VEC_WIDEN_MULT_HI_EXPR; |
13771 | break; |
13772 | |
13773 | case DOT_PROD_EXPR: |
13774 | c1 = DOT_PROD_EXPR; |
13775 | c2 = DOT_PROD_EXPR; |
13776 | break; |
13777 | |
13778 | case SAD_EXPR: |
13779 | c1 = SAD_EXPR; |
13780 | c2 = SAD_EXPR; |
13781 | break; |
13782 | |
13783 | case VEC_WIDEN_MULT_EVEN_EXPR: |
13784 | /* Support the recursion induced just above. */ |
13785 | c1 = VEC_WIDEN_MULT_EVEN_EXPR; |
13786 | c2 = VEC_WIDEN_MULT_ODD_EXPR; |
13787 | break; |
13788 | |
13789 | case WIDEN_LSHIFT_EXPR: |
13790 | c1 = VEC_WIDEN_LSHIFT_LO_EXPR; |
13791 | c2 = VEC_WIDEN_LSHIFT_HI_EXPR; |
13792 | break; |
13793 | |
13794 | CASE_CONVERT: |
13795 | c1 = VEC_UNPACK_LO_EXPR; |
13796 | c2 = VEC_UNPACK_HI_EXPR; |
13797 | break; |
13798 | |
13799 | case FLOAT_EXPR: |
13800 | c1 = VEC_UNPACK_FLOAT_LO_EXPR; |
13801 | c2 = VEC_UNPACK_FLOAT_HI_EXPR; |
13802 | break; |
13803 | |
13804 | case FIX_TRUNC_EXPR: |
13805 | c1 = VEC_UNPACK_FIX_TRUNC_LO_EXPR; |
13806 | c2 = VEC_UNPACK_FIX_TRUNC_HI_EXPR; |
13807 | break; |
13808 | |
13809 | default: |
13810 | gcc_unreachable (); |
13811 | } |
13812 | |
13813 | if (BYTES_BIG_ENDIAN && c1 != VEC_WIDEN_MULT_EVEN_EXPR) |
13814 | std::swap (a&: c1, b&: c2); |
13815 | |
13816 | if (code == FIX_TRUNC_EXPR) |
13817 | { |
13818 | /* The signedness is determined from output operand. */ |
13819 | optab1 = optab_for_tree_code (c1, vectype_out, optab_default); |
13820 | optab2 = optab_for_tree_code (c2, vectype_out, optab_default); |
13821 | } |
13822 | else if (CONVERT_EXPR_CODE_P (code.safe_as_tree_code ()) |
13823 | && VECTOR_BOOLEAN_TYPE_P (wide_vectype) |
13824 | && VECTOR_BOOLEAN_TYPE_P (vectype) |
13825 | && TYPE_MODE (wide_vectype) == TYPE_MODE (vectype) |
13826 | && SCALAR_INT_MODE_P (TYPE_MODE (vectype))) |
13827 | { |
13828 | /* If the input and result modes are the same, a different optab |
13829 | is needed where we pass in the number of units in vectype. */ |
13830 | optab1 = vec_unpacks_sbool_lo_optab; |
13831 | optab2 = vec_unpacks_sbool_hi_optab; |
13832 | } |
13833 | |
13834 | vec_mode = TYPE_MODE (vectype); |
13835 | if (widening_fn_p (code)) |
13836 | { |
13837 | /* If this is an internal fn then we must check whether the target |
13838 | supports either a low-high split or an even-odd split. */ |
13839 | internal_fn ifn = as_internal_fn (code: (combined_fn) code); |
13840 | |
13841 | internal_fn lo, hi, even, odd; |
13842 | lookup_hilo_internal_fn (ifn, &lo, &hi); |
13843 | *code1 = as_combined_fn (fn: lo); |
13844 | *code2 = as_combined_fn (fn: hi); |
13845 | optab1 = direct_internal_fn_optab (lo, {vectype, vectype}); |
13846 | optab2 = direct_internal_fn_optab (hi, {vectype, vectype}); |
13847 | |
13848 | /* If we don't support low-high, then check for even-odd. */ |
13849 | if (!optab1 |
13850 | || (icode1 = optab_handler (op: optab1, mode: vec_mode)) == CODE_FOR_nothing |
13851 | || !optab2 |
13852 | || (icode2 = optab_handler (op: optab2, mode: vec_mode)) == CODE_FOR_nothing) |
13853 | { |
13854 | lookup_evenodd_internal_fn (ifn, &even, &odd); |
13855 | *code1 = as_combined_fn (fn: even); |
13856 | *code2 = as_combined_fn (fn: odd); |
13857 | optab1 = direct_internal_fn_optab (even, {vectype, vectype}); |
13858 | optab2 = direct_internal_fn_optab (odd, {vectype, vectype}); |
13859 | } |
13860 | } |
13861 | else if (code.is_tree_code ()) |
13862 | { |
13863 | if (code == FIX_TRUNC_EXPR) |
13864 | { |
13865 | /* The signedness is determined from output operand. */ |
13866 | optab1 = optab_for_tree_code (c1, vectype_out, optab_default); |
13867 | optab2 = optab_for_tree_code (c2, vectype_out, optab_default); |
13868 | } |
13869 | else if (CONVERT_EXPR_CODE_P ((tree_code) code.safe_as_tree_code ()) |
13870 | && VECTOR_BOOLEAN_TYPE_P (wide_vectype) |
13871 | && VECTOR_BOOLEAN_TYPE_P (vectype) |
13872 | && TYPE_MODE (wide_vectype) == TYPE_MODE (vectype) |
13873 | && SCALAR_INT_MODE_P (TYPE_MODE (vectype))) |
13874 | { |
13875 | /* If the input and result modes are the same, a different optab |
13876 | is needed where we pass in the number of units in vectype. */ |
13877 | optab1 = vec_unpacks_sbool_lo_optab; |
13878 | optab2 = vec_unpacks_sbool_hi_optab; |
13879 | } |
13880 | else |
13881 | { |
13882 | optab1 = optab_for_tree_code (c1, vectype, optab_default); |
13883 | optab2 = optab_for_tree_code (c2, vectype, optab_default); |
13884 | } |
13885 | *code1 = c1; |
13886 | *code2 = c2; |
13887 | } |
13888 | |
13889 | if (!optab1 || !optab2) |
13890 | return false; |
13891 | |
13892 | if ((icode1 = optab_handler (op: optab1, mode: vec_mode)) == CODE_FOR_nothing |
13893 | || (icode2 = optab_handler (op: optab2, mode: vec_mode)) == CODE_FOR_nothing) |
13894 | return false; |
13895 | |
13896 | |
13897 | if (insn_data[icode1].operand[0].mode == TYPE_MODE (wide_vectype) |
13898 | && insn_data[icode2].operand[0].mode == TYPE_MODE (wide_vectype)) |
13899 | { |
13900 | if (!VECTOR_BOOLEAN_TYPE_P (vectype)) |
13901 | return true; |
13902 | /* For scalar masks we may have different boolean |
13903 | vector types having the same QImode. Thus we |
13904 | add additional check for elements number. */ |
13905 | if (known_eq (TYPE_VECTOR_SUBPARTS (vectype), |
13906 | TYPE_VECTOR_SUBPARTS (wide_vectype) * 2)) |
13907 | return true; |
13908 | } |
13909 | |
13910 | /* Check if it's a multi-step conversion that can be done using intermediate |
13911 | types. */ |
13912 | |
13913 | prev_type = vectype; |
13914 | prev_mode = vec_mode; |
13915 | |
13916 | if (!CONVERT_EXPR_CODE_P (code.safe_as_tree_code ())) |
13917 | return false; |
13918 | |
13919 | /* We assume here that there will not be more than MAX_INTERM_CVT_STEPS |
13920 | intermediate steps in promotion sequence. We try |
13921 | MAX_INTERM_CVT_STEPS to get to NARROW_VECTYPE, and fail if we do |
13922 | not. */ |
13923 | interm_types->create (MAX_INTERM_CVT_STEPS); |
13924 | for (i = 0; i < MAX_INTERM_CVT_STEPS; i++) |
13925 | { |
13926 | intermediate_mode = insn_data[icode1].operand[0].mode; |
13927 | if (VECTOR_BOOLEAN_TYPE_P (prev_type)) |
13928 | intermediate_type |
13929 | = vect_halve_mask_nunits (prev_type, intermediate_mode); |
13930 | else if (VECTOR_MODE_P (intermediate_mode)) |
13931 | { |
13932 | tree intermediate_element_type |
13933 | = lang_hooks.types.type_for_mode (GET_MODE_INNER (intermediate_mode), |
13934 | TYPE_UNSIGNED (prev_type)); |
13935 | intermediate_type |
13936 | = build_vector_type_for_mode (intermediate_element_type, |
13937 | intermediate_mode); |
13938 | } |
13939 | else |
13940 | intermediate_type |
13941 | = lang_hooks.types.type_for_mode (intermediate_mode, |
13942 | TYPE_UNSIGNED (prev_type)); |
13943 | |
13944 | if (VECTOR_BOOLEAN_TYPE_P (intermediate_type) |
13945 | && VECTOR_BOOLEAN_TYPE_P (prev_type) |
13946 | && intermediate_mode == prev_mode |
13947 | && SCALAR_INT_MODE_P (prev_mode)) |
13948 | { |
13949 | /* If the input and result modes are the same, a different optab |
13950 | is needed where we pass in the number of units in vectype. */ |
13951 | optab3 = vec_unpacks_sbool_lo_optab; |
13952 | optab4 = vec_unpacks_sbool_hi_optab; |
13953 | } |
13954 | else |
13955 | { |
13956 | optab3 = optab_for_tree_code (c1, intermediate_type, optab_default); |
13957 | optab4 = optab_for_tree_code (c2, intermediate_type, optab_default); |
13958 | } |
13959 | |
13960 | if (!optab3 || !optab4 |
13961 | || (icode1 = optab_handler (op: optab1, mode: prev_mode)) == CODE_FOR_nothing |
13962 | || insn_data[icode1].operand[0].mode != intermediate_mode |
13963 | || (icode2 = optab_handler (op: optab2, mode: prev_mode)) == CODE_FOR_nothing |
13964 | || insn_data[icode2].operand[0].mode != intermediate_mode |
13965 | || ((icode1 = optab_handler (op: optab3, mode: intermediate_mode)) |
13966 | == CODE_FOR_nothing) |
13967 | || ((icode2 = optab_handler (op: optab4, mode: intermediate_mode)) |
13968 | == CODE_FOR_nothing)) |
13969 | break; |
13970 | |
13971 | interm_types->quick_push (obj: intermediate_type); |
13972 | (*multi_step_cvt)++; |
13973 | |
13974 | if (insn_data[icode1].operand[0].mode == TYPE_MODE (wide_vectype) |
13975 | && insn_data[icode2].operand[0].mode == TYPE_MODE (wide_vectype)) |
13976 | { |
13977 | if (!VECTOR_BOOLEAN_TYPE_P (vectype)) |
13978 | return true; |
13979 | if (known_eq (TYPE_VECTOR_SUBPARTS (intermediate_type), |
13980 | TYPE_VECTOR_SUBPARTS (wide_vectype) * 2)) |
13981 | return true; |
13982 | } |
13983 | |
13984 | prev_type = intermediate_type; |
13985 | prev_mode = intermediate_mode; |
13986 | } |
13987 | |
13988 | interm_types->release (); |
13989 | return false; |
13990 | } |
13991 | |
13992 | |
13993 | /* Function supportable_narrowing_operation |
13994 | |
13995 | Check whether an operation represented by the code CODE is a |
13996 | narrowing operation that is supported by the target platform in |
13997 | vector form (i.e., when operating on arguments of type VECTYPE_IN |
13998 | and producing a result of type VECTYPE_OUT). |
13999 | |
14000 | Narrowing operations we currently support are NOP (CONVERT), FIX_TRUNC |
14001 | and FLOAT. This function checks if these operations are supported by |
14002 | the target platform directly via vector tree-codes. |
14003 | |
14004 | Output: |
14005 | - CODE1 is the code of a vector operation to be used when |
14006 | vectorizing the operation, if available. |
14007 | - MULTI_STEP_CVT determines the number of required intermediate steps in |
14008 | case of multi-step conversion (like int->short->char - in that case |
14009 | MULTI_STEP_CVT will be 1). |
14010 | - INTERM_TYPES contains the intermediate type required to perform the |
14011 | narrowing operation (short in the above example). */ |
14012 | |
14013 | bool |
14014 | supportable_narrowing_operation (code_helper code, |
14015 | tree vectype_out, tree vectype_in, |
14016 | code_helper *code1, int *multi_step_cvt, |
14017 | vec<tree> *interm_types) |
14018 | { |
14019 | machine_mode vec_mode; |
14020 | enum insn_code icode1; |
14021 | optab optab1, interm_optab; |
14022 | tree vectype = vectype_in; |
14023 | tree narrow_vectype = vectype_out; |
14024 | enum tree_code c1; |
14025 | tree intermediate_type, prev_type; |
14026 | machine_mode intermediate_mode, prev_mode; |
14027 | int i; |
14028 | unsigned HOST_WIDE_INT n_elts; |
14029 | bool uns; |
14030 | |
14031 | if (!code.is_tree_code ()) |
14032 | return false; |
14033 | |
14034 | *multi_step_cvt = 0; |
14035 | switch ((tree_code) code) |
14036 | { |
14037 | CASE_CONVERT: |
14038 | c1 = VEC_PACK_TRUNC_EXPR; |
14039 | if (VECTOR_BOOLEAN_TYPE_P (narrow_vectype) |
14040 | && VECTOR_BOOLEAN_TYPE_P (vectype) |
14041 | && SCALAR_INT_MODE_P (TYPE_MODE (vectype)) |
14042 | && TYPE_VECTOR_SUBPARTS (node: vectype).is_constant (const_value: &n_elts) |
14043 | && n_elts < BITS_PER_UNIT) |
14044 | optab1 = vec_pack_sbool_trunc_optab; |
14045 | else |
14046 | optab1 = optab_for_tree_code (c1, vectype, optab_default); |
14047 | break; |
14048 | |
14049 | case FIX_TRUNC_EXPR: |
14050 | c1 = VEC_PACK_FIX_TRUNC_EXPR; |
14051 | /* The signedness is determined from output operand. */ |
14052 | optab1 = optab_for_tree_code (c1, vectype_out, optab_default); |
14053 | break; |
14054 | |
14055 | case FLOAT_EXPR: |
14056 | c1 = VEC_PACK_FLOAT_EXPR; |
14057 | optab1 = optab_for_tree_code (c1, vectype, optab_default); |
14058 | break; |
14059 | |
14060 | default: |
14061 | gcc_unreachable (); |
14062 | } |
14063 | |
14064 | if (!optab1) |
14065 | return false; |
14066 | |
14067 | vec_mode = TYPE_MODE (vectype); |
14068 | if ((icode1 = optab_handler (op: optab1, mode: vec_mode)) == CODE_FOR_nothing) |
14069 | return false; |
14070 | |
14071 | *code1 = c1; |
14072 | |
14073 | if (insn_data[icode1].operand[0].mode == TYPE_MODE (narrow_vectype)) |
14074 | { |
14075 | if (!VECTOR_BOOLEAN_TYPE_P (vectype)) |
14076 | return true; |
14077 | /* For scalar masks we may have different boolean |
14078 | vector types having the same QImode. Thus we |
14079 | add additional check for elements number. */ |
14080 | if (known_eq (TYPE_VECTOR_SUBPARTS (vectype) * 2, |
14081 | TYPE_VECTOR_SUBPARTS (narrow_vectype))) |
14082 | return true; |
14083 | } |
14084 | |
14085 | if (code == FLOAT_EXPR) |
14086 | return false; |
14087 | |
14088 | /* Check if it's a multi-step conversion that can be done using intermediate |
14089 | types. */ |
14090 | prev_mode = vec_mode; |
14091 | prev_type = vectype; |
14092 | if (code == FIX_TRUNC_EXPR) |
14093 | uns = TYPE_UNSIGNED (vectype_out); |
14094 | else |
14095 | uns = TYPE_UNSIGNED (vectype); |
14096 | |
14097 | /* For multi-step FIX_TRUNC_EXPR prefer signed floating to integer |
14098 | conversion over unsigned, as unsigned FIX_TRUNC_EXPR is often more |
14099 | costly than signed. */ |
14100 | if (code == FIX_TRUNC_EXPR && uns) |
14101 | { |
14102 | enum insn_code icode2; |
14103 | |
14104 | intermediate_type |
14105 | = lang_hooks.types.type_for_mode (TYPE_MODE (vectype_out), 0); |
14106 | interm_optab |
14107 | = optab_for_tree_code (c1, intermediate_type, optab_default); |
14108 | if (interm_optab != unknown_optab |
14109 | && (icode2 = optab_handler (op: optab1, mode: vec_mode)) != CODE_FOR_nothing |
14110 | && insn_data[icode1].operand[0].mode |
14111 | == insn_data[icode2].operand[0].mode) |
14112 | { |
14113 | uns = false; |
14114 | optab1 = interm_optab; |
14115 | icode1 = icode2; |
14116 | } |
14117 | } |
14118 | |
14119 | /* We assume here that there will not be more than MAX_INTERM_CVT_STEPS |
14120 | intermediate steps in promotion sequence. We try |
14121 | MAX_INTERM_CVT_STEPS to get to NARROW_VECTYPE, and fail if we do not. */ |
14122 | interm_types->create (MAX_INTERM_CVT_STEPS); |
14123 | for (i = 0; i < MAX_INTERM_CVT_STEPS; i++) |
14124 | { |
14125 | intermediate_mode = insn_data[icode1].operand[0].mode; |
14126 | if (VECTOR_BOOLEAN_TYPE_P (prev_type)) |
14127 | intermediate_type |
14128 | = vect_double_mask_nunits (prev_type, intermediate_mode); |
14129 | else |
14130 | intermediate_type |
14131 | = lang_hooks.types.type_for_mode (intermediate_mode, uns); |
14132 | if (VECTOR_BOOLEAN_TYPE_P (intermediate_type) |
14133 | && VECTOR_BOOLEAN_TYPE_P (prev_type) |
14134 | && SCALAR_INT_MODE_P (prev_mode) |
14135 | && TYPE_VECTOR_SUBPARTS (node: intermediate_type).is_constant (const_value: &n_elts) |
14136 | && n_elts < BITS_PER_UNIT) |
14137 | interm_optab = vec_pack_sbool_trunc_optab; |
14138 | else |
14139 | interm_optab |
14140 | = optab_for_tree_code (VEC_PACK_TRUNC_EXPR, intermediate_type, |
14141 | optab_default); |
14142 | if (!interm_optab |
14143 | || ((icode1 = optab_handler (op: optab1, mode: prev_mode)) == CODE_FOR_nothing) |
14144 | || insn_data[icode1].operand[0].mode != intermediate_mode |
14145 | || ((icode1 = optab_handler (op: interm_optab, mode: intermediate_mode)) |
14146 | == CODE_FOR_nothing)) |
14147 | break; |
14148 | |
14149 | interm_types->quick_push (obj: intermediate_type); |
14150 | (*multi_step_cvt)++; |
14151 | |
14152 | if (insn_data[icode1].operand[0].mode == TYPE_MODE (narrow_vectype)) |
14153 | { |
14154 | if (!VECTOR_BOOLEAN_TYPE_P (vectype)) |
14155 | return true; |
14156 | if (known_eq (TYPE_VECTOR_SUBPARTS (intermediate_type) * 2, |
14157 | TYPE_VECTOR_SUBPARTS (narrow_vectype))) |
14158 | return true; |
14159 | } |
14160 | |
14161 | prev_mode = intermediate_mode; |
14162 | prev_type = intermediate_type; |
14163 | optab1 = interm_optab; |
14164 | } |
14165 | |
14166 | interm_types->release (); |
14167 | return false; |
14168 | } |
14169 | |
14170 | /* Generate and return a vector mask of MASK_TYPE such that |
14171 | mask[I] is true iff J + START_INDEX < END_INDEX for all J <= I. |
14172 | Add the statements to SEQ. */ |
14173 | |
14174 | tree |
14175 | vect_gen_while (gimple_seq *seq, tree mask_type, tree start_index, |
14176 | tree end_index, const char *name) |
14177 | { |
14178 | tree cmp_type = TREE_TYPE (start_index); |
14179 | gcc_checking_assert (direct_internal_fn_supported_p (IFN_WHILE_ULT, |
14180 | cmp_type, mask_type, |
14181 | OPTIMIZE_FOR_SPEED)); |
14182 | gcall *call = gimple_build_call_internal (IFN_WHILE_ULT, 3, |
14183 | start_index, end_index, |
14184 | build_zero_cst (mask_type)); |
14185 | tree tmp; |
14186 | if (name) |
14187 | tmp = make_temp_ssa_name (type: mask_type, NULL, name); |
14188 | else |
14189 | tmp = make_ssa_name (var: mask_type); |
14190 | gimple_call_set_lhs (gs: call, lhs: tmp); |
14191 | gimple_seq_add_stmt (seq, call); |
14192 | return tmp; |
14193 | } |
14194 | |
14195 | /* Generate a vector mask of type MASK_TYPE for which index I is false iff |
14196 | J + START_INDEX < END_INDEX for all J <= I. Add the statements to SEQ. */ |
14197 | |
14198 | tree |
14199 | vect_gen_while_not (gimple_seq *seq, tree mask_type, tree start_index, |
14200 | tree end_index) |
14201 | { |
14202 | tree tmp = vect_gen_while (seq, mask_type, start_index, end_index); |
14203 | return gimple_build (seq, code: BIT_NOT_EXPR, type: mask_type, ops: tmp); |
14204 | } |
14205 | |
14206 | /* Try to compute the vector types required to vectorize STMT_INFO, |
14207 | returning true on success and false if vectorization isn't possible. |
14208 | If GROUP_SIZE is nonzero and we're performing BB vectorization, |
14209 | take sure that the number of elements in the vectors is no bigger |
14210 | than GROUP_SIZE. |
14211 | |
14212 | On success: |
14213 | |
14214 | - Set *STMT_VECTYPE_OUT to: |
14215 | - NULL_TREE if the statement doesn't need to be vectorized; |
14216 | - the equivalent of STMT_VINFO_VECTYPE otherwise. |
14217 | |
14218 | - Set *NUNITS_VECTYPE_OUT to the vector type that contains the maximum |
14219 | number of units needed to vectorize STMT_INFO, or NULL_TREE if the |
14220 | statement does not help to determine the overall number of units. */ |
14221 | |
14222 | opt_result |
14223 | vect_get_vector_types_for_stmt (vec_info *vinfo, stmt_vec_info stmt_info, |
14224 | tree *stmt_vectype_out, |
14225 | tree *nunits_vectype_out, |
14226 | unsigned int group_size) |
14227 | { |
14228 | gimple *stmt = stmt_info->stmt; |
14229 | |
14230 | /* For BB vectorization, we should always have a group size once we've |
14231 | constructed the SLP tree; the only valid uses of zero GROUP_SIZEs |
14232 | are tentative requests during things like early data reference |
14233 | analysis and pattern recognition. */ |
14234 | if (is_a <bb_vec_info> (p: vinfo)) |
14235 | gcc_assert (vinfo->slp_instances.is_empty () || group_size != 0); |
14236 | else |
14237 | group_size = 0; |
14238 | |
14239 | *stmt_vectype_out = NULL_TREE; |
14240 | *nunits_vectype_out = NULL_TREE; |
14241 | |
14242 | if (gimple_get_lhs (stmt) == NULL_TREE |
14243 | /* MASK_STORE has no lhs, but is ok. */ |
14244 | && !gimple_call_internal_p (gs: stmt, fn: IFN_MASK_STORE)) |
14245 | { |
14246 | if (is_a <gcall *> (p: stmt)) |
14247 | { |
14248 | /* Ignore calls with no lhs. These must be calls to |
14249 | #pragma omp simd functions, and what vectorization factor |
14250 | it really needs can't be determined until |
14251 | vectorizable_simd_clone_call. */ |
14252 | if (dump_enabled_p ()) |
14253 | dump_printf_loc (MSG_NOTE, vect_location, |
14254 | "defer to SIMD clone analysis.\n" ); |
14255 | return opt_result::success (); |
14256 | } |
14257 | |
14258 | return opt_result::failure_at (loc: stmt, |
14259 | fmt: "not vectorized: irregular stmt.%G" , stmt); |
14260 | } |
14261 | |
14262 | tree vectype; |
14263 | tree scalar_type = NULL_TREE; |
14264 | if (group_size == 0 && STMT_VINFO_VECTYPE (stmt_info)) |
14265 | { |
14266 | vectype = STMT_VINFO_VECTYPE (stmt_info); |
14267 | if (dump_enabled_p ()) |
14268 | dump_printf_loc (MSG_NOTE, vect_location, |
14269 | "precomputed vectype: %T\n" , vectype); |
14270 | } |
14271 | else if (vect_use_mask_type_p (stmt_info)) |
14272 | { |
14273 | unsigned int precision = stmt_info->mask_precision; |
14274 | scalar_type = build_nonstandard_integer_type (precision, 1); |
14275 | vectype = get_mask_type_for_scalar_type (vinfo, scalar_type, group_size); |
14276 | if (!vectype) |
14277 | return opt_result::failure_at (loc: stmt, fmt: "not vectorized: unsupported" |
14278 | " data-type %T\n" , scalar_type); |
14279 | if (dump_enabled_p ()) |
14280 | dump_printf_loc (MSG_NOTE, vect_location, "vectype: %T\n" , vectype); |
14281 | } |
14282 | else |
14283 | { |
14284 | if (data_reference *dr = STMT_VINFO_DATA_REF (stmt_info)) |
14285 | scalar_type = TREE_TYPE (DR_REF (dr)); |
14286 | else if (gimple_call_internal_p (gs: stmt, fn: IFN_MASK_STORE)) |
14287 | scalar_type = TREE_TYPE (gimple_call_arg (stmt, 3)); |
14288 | else |
14289 | scalar_type = TREE_TYPE (gimple_get_lhs (stmt)); |
14290 | |
14291 | if (dump_enabled_p ()) |
14292 | { |
14293 | if (group_size) |
14294 | dump_printf_loc (MSG_NOTE, vect_location, |
14295 | "get vectype for scalar type (group size %d):" |
14296 | " %T\n" , group_size, scalar_type); |
14297 | else |
14298 | dump_printf_loc (MSG_NOTE, vect_location, |
14299 | "get vectype for scalar type: %T\n" , scalar_type); |
14300 | } |
14301 | vectype = get_vectype_for_scalar_type (vinfo, scalar_type, group_size); |
14302 | if (!vectype) |
14303 | return opt_result::failure_at (loc: stmt, |
14304 | fmt: "not vectorized:" |
14305 | " unsupported data-type %T\n" , |
14306 | scalar_type); |
14307 | |
14308 | if (dump_enabled_p ()) |
14309 | dump_printf_loc (MSG_NOTE, vect_location, "vectype: %T\n" , vectype); |
14310 | } |
14311 | |
14312 | if (scalar_type && VECTOR_MODE_P (TYPE_MODE (scalar_type))) |
14313 | return opt_result::failure_at (loc: stmt, |
14314 | fmt: "not vectorized: vector stmt in loop:%G" , |
14315 | stmt); |
14316 | |
14317 | *stmt_vectype_out = vectype; |
14318 | |
14319 | /* Don't try to compute scalar types if the stmt produces a boolean |
14320 | vector; use the existing vector type instead. */ |
14321 | tree nunits_vectype = vectype; |
14322 | if (!VECTOR_BOOLEAN_TYPE_P (vectype)) |
14323 | { |
14324 | /* The number of units is set according to the smallest scalar |
14325 | type (or the largest vector size, but we only support one |
14326 | vector size per vectorization). */ |
14327 | scalar_type = vect_get_smallest_scalar_type (stmt_info, |
14328 | TREE_TYPE (vectype)); |
14329 | if (scalar_type != TREE_TYPE (vectype)) |
14330 | { |
14331 | if (dump_enabled_p ()) |
14332 | dump_printf_loc (MSG_NOTE, vect_location, |
14333 | "get vectype for smallest scalar type: %T\n" , |
14334 | scalar_type); |
14335 | nunits_vectype = get_vectype_for_scalar_type (vinfo, scalar_type, |
14336 | group_size); |
14337 | if (!nunits_vectype) |
14338 | return opt_result::failure_at |
14339 | (loc: stmt, fmt: "not vectorized: unsupported data-type %T\n" , |
14340 | scalar_type); |
14341 | if (dump_enabled_p ()) |
14342 | dump_printf_loc (MSG_NOTE, vect_location, "nunits vectype: %T\n" , |
14343 | nunits_vectype); |
14344 | } |
14345 | } |
14346 | |
14347 | if (!multiple_p (a: TYPE_VECTOR_SUBPARTS (node: nunits_vectype), |
14348 | b: TYPE_VECTOR_SUBPARTS (node: *stmt_vectype_out))) |
14349 | return opt_result::failure_at (loc: stmt, |
14350 | fmt: "Not vectorized: Incompatible number " |
14351 | "of vector subparts between %T and %T\n" , |
14352 | nunits_vectype, *stmt_vectype_out); |
14353 | |
14354 | if (dump_enabled_p ()) |
14355 | { |
14356 | dump_printf_loc (MSG_NOTE, vect_location, "nunits = " ); |
14357 | dump_dec (MSG_NOTE, TYPE_VECTOR_SUBPARTS (node: nunits_vectype)); |
14358 | dump_printf (MSG_NOTE, "\n" ); |
14359 | } |
14360 | |
14361 | *nunits_vectype_out = nunits_vectype; |
14362 | return opt_result::success (); |
14363 | } |
14364 | |
14365 | /* Generate and return statement sequence that sets vector length LEN that is: |
14366 | |
14367 | min_of_start_and_end = min (START_INDEX, END_INDEX); |
14368 | left_len = END_INDEX - min_of_start_and_end; |
14369 | rhs = min (left_len, LEN_LIMIT); |
14370 | LEN = rhs; |
14371 | |
14372 | Note: the cost of the code generated by this function is modeled |
14373 | by vect_estimate_min_profitable_iters, so changes here may need |
14374 | corresponding changes there. */ |
14375 | |
14376 | gimple_seq |
14377 | vect_gen_len (tree len, tree start_index, tree end_index, tree len_limit) |
14378 | { |
14379 | gimple_seq stmts = NULL; |
14380 | tree len_type = TREE_TYPE (len); |
14381 | gcc_assert (TREE_TYPE (start_index) == len_type); |
14382 | |
14383 | tree min = gimple_build (seq: &stmts, code: MIN_EXPR, type: len_type, ops: start_index, ops: end_index); |
14384 | tree left_len = gimple_build (seq: &stmts, code: MINUS_EXPR, type: len_type, ops: end_index, ops: min); |
14385 | tree rhs = gimple_build (seq: &stmts, code: MIN_EXPR, type: len_type, ops: left_len, ops: len_limit); |
14386 | gimple* stmt = gimple_build_assign (len, rhs); |
14387 | gimple_seq_add_stmt (&stmts, stmt); |
14388 | |
14389 | return stmts; |
14390 | } |
14391 | |
14392 | |