Line data Source code
1 : /* Forward propagation of expressions for single use variables.
2 : Copyright (C) 2004-2026 Free Software Foundation, Inc.
3 :
4 : This file is part of GCC.
5 :
6 : GCC is free software; you can redistribute it and/or modify
7 : it under the terms of the GNU General Public License as published by
8 : the Free Software Foundation; either version 3, or (at your option)
9 : any later version.
10 :
11 : GCC is distributed in the hope that it will be useful,
12 : but WITHOUT ANY WARRANTY; without even the implied warranty of
13 : MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
14 : GNU General Public License for more details.
15 :
16 : You should have received a copy of the GNU General Public License
17 : along with GCC; see the file COPYING3. If not see
18 : <http://www.gnu.org/licenses/>. */
19 :
20 : #include "config.h"
21 : #include "system.h"
22 : #include "coretypes.h"
23 : #include "backend.h"
24 : #include "rtl.h"
25 : #include "tree.h"
26 : #include "gimple.h"
27 : #include "cfghooks.h"
28 : #include "tree-pass.h"
29 : #include "ssa.h"
30 : #include "expmed.h"
31 : #include "optabs-query.h"
32 : #include "gimple-pretty-print.h"
33 : #include "fold-const.h"
34 : #include "stor-layout.h"
35 : #include "gimple-iterator.h"
36 : #include "gimple-fold.h"
37 : #include "tree-eh.h"
38 : #include "gimplify.h"
39 : #include "gimplify-me.h"
40 : #include "tree-cfg.h"
41 : #include "expr.h"
42 : #include "tree-dfa.h"
43 : #include "tree-ssa-propagate.h"
44 : #include "tree-ssa-dom.h"
45 : #include "tree-ssa-strlen.h"
46 : #include "builtins.h"
47 : #include "tree-cfgcleanup.h"
48 : #include "cfganal.h"
49 : #include "optabs-tree.h"
50 : #include "insn-config.h"
51 : #include "recog.h"
52 : #include "cfgloop.h"
53 : #include "tree-vectorizer.h"
54 : #include "tree-vector-builder.h"
55 : #include "vec-perm-indices.h"
56 : #include "internal-fn.h"
57 : #include "cgraph.h"
58 : #include "tree-ssa.h"
59 : #include "gimple-range.h"
60 : #include "tree-ssa-dce.h"
61 :
62 : /* This pass propagates the RHS of assignment statements into use
63 : sites of the LHS of the assignment. It's basically a specialized
64 : form of tree combination. It is hoped all of this can disappear
65 : when we have a generalized tree combiner.
66 :
67 : One class of common cases we handle is forward propagating a single use
68 : variable into a COND_EXPR.
69 :
70 : bb0:
71 : x = a COND b;
72 : if (x) goto ... else goto ...
73 :
74 : Will be transformed into:
75 :
76 : bb0:
77 : if (a COND b) goto ... else goto ...
78 :
79 : Similarly for the tests (x == 0), (x != 0), (x == 1) and (x != 1).
80 :
81 : Or (assuming c1 and c2 are constants):
82 :
83 : bb0:
84 : x = a + c1;
85 : if (x EQ/NEQ c2) goto ... else goto ...
86 :
87 : Will be transformed into:
88 :
89 : bb0:
90 : if (a EQ/NEQ (c2 - c1)) goto ... else goto ...
91 :
92 : Similarly for x = a - c1.
93 :
94 : Or
95 :
96 : bb0:
97 : x = !a
98 : if (x) goto ... else goto ...
99 :
100 : Will be transformed into:
101 :
102 : bb0:
103 : if (a == 0) goto ... else goto ...
104 :
105 : Similarly for the tests (x == 0), (x != 0), (x == 1) and (x != 1).
106 : For these cases, we propagate A into all, possibly more than one,
107 : COND_EXPRs that use X.
108 :
109 : Or
110 :
111 : bb0:
112 : x = (typecast) a
113 : if (x) goto ... else goto ...
114 :
115 : Will be transformed into:
116 :
117 : bb0:
118 : if (a != 0) goto ... else goto ...
119 :
120 : (Assuming a is an integral type and x is a boolean or x is an
121 : integral and a is a boolean.)
122 :
123 : Similarly for the tests (x == 0), (x != 0), (x == 1) and (x != 1).
124 : For these cases, we propagate A into all, possibly more than one,
125 : COND_EXPRs that use X.
126 :
127 : In addition to eliminating the variable and the statement which assigns
128 : a value to the variable, we may be able to later thread the jump without
129 : adding insane complexity in the dominator optimizer.
130 :
131 : Also note these transformations can cascade. We handle this by having
132 : a worklist of COND_EXPR statements to examine. As we make a change to
133 : a statement, we put it back on the worklist to examine on the next
134 : iteration of the main loop.
135 :
136 : A second class of propagation opportunities arises for ADDR_EXPR
137 : nodes.
138 :
139 : ptr = &x->y->z;
140 : res = *ptr;
141 :
142 : Will get turned into
143 :
144 : res = x->y->z;
145 :
146 : Or
147 : ptr = (type1*)&type2var;
148 : res = *ptr
149 :
150 : Will get turned into (if type1 and type2 are the same size
151 : and neither have volatile on them):
152 : res = VIEW_CONVERT_EXPR<type1>(type2var)
153 :
154 : Or
155 :
156 : ptr = &x[0];
157 : ptr2 = ptr + <constant>;
158 :
159 : Will get turned into
160 :
161 : ptr2 = &x[constant/elementsize];
162 :
163 : Or
164 :
165 : ptr = &x[0];
166 : offset = index * element_size;
167 : offset_p = (pointer) offset;
168 : ptr2 = ptr + offset_p
169 :
170 : Will get turned into:
171 :
172 : ptr2 = &x[index];
173 :
174 : Or
175 : ssa = (int) decl
176 : res = ssa & 1
177 :
178 : Provided that decl has known alignment >= 2, will get turned into
179 :
180 : res = 0
181 :
182 : We also propagate casts into SWITCH_EXPR and COND_EXPR conditions to
183 : allow us to remove the cast and {NOT_EXPR,NEG_EXPR} into a subsequent
184 : {NOT_EXPR,NEG_EXPR}.
185 :
186 : This will (of course) be extended as other needs arise. */
187 :
188 : /* Data structure that contains simplifiable vectorized permute sequences.
189 : See recognise_vec_perm_simplify_seq () for a description of the sequence. */
190 :
191 : struct _vec_perm_simplify_seq
192 : {
193 : /* Defining stmts of vectors in the sequence. */
194 : gassign *v_1_stmt;
195 : gassign *v_2_stmt;
196 : gassign *v_x_stmt;
197 : gassign *v_y_stmt;
198 : /* Final permute statement. */
199 : gassign *stmt;
200 : /* New selector indices for stmt. */
201 : tree new_sel;
202 : /* Elements of each vector and selector. */
203 : unsigned int nelts;
204 : };
205 : typedef struct _vec_perm_simplify_seq *vec_perm_simplify_seq;
206 :
207 : static bool forward_propagate_addr_expr (tree, tree, bool);
208 :
209 : /* Set to true if we delete dead edges during the optimization. */
210 : static bool cfg_changed;
211 :
212 : static tree rhs_to_tree (tree type, gimple *stmt);
213 :
214 : static bitmap to_purge;
215 :
216 : /* Const-and-copy lattice. */
217 : static vec<tree> lattice;
218 :
219 : /* Set the lattice entry for NAME to VAL. */
220 : static void
221 33132171 : fwprop_set_lattice_val (tree name, tree val)
222 : {
223 33132171 : if (TREE_CODE (name) == SSA_NAME)
224 : {
225 33132171 : if (SSA_NAME_VERSION (name) >= lattice.length ())
226 : {
227 32205 : lattice.reserve (num_ssa_names - lattice.length ());
228 21470 : lattice.quick_grow_cleared (num_ssa_names);
229 : }
230 33132171 : lattice[SSA_NAME_VERSION (name)] = val;
231 : /* As this now constitutes a copy duplicate points-to
232 : and range info appropriately. */
233 33132171 : if (TREE_CODE (val) == SSA_NAME)
234 32680742 : maybe_duplicate_ssa_info_at_copy (name, val);
235 : }
236 33132171 : }
237 :
238 : /* Invalidate the lattice entry for NAME, done when releasing SSA names. */
239 : static void
240 937570 : fwprop_invalidate_lattice (tree name)
241 : {
242 937570 : if (name
243 935171 : && TREE_CODE (name) == SSA_NAME
244 1872614 : && SSA_NAME_VERSION (name) < lattice.length ())
245 935011 : lattice[SSA_NAME_VERSION (name)] = NULL_TREE;
246 937570 : }
247 :
248 : /* Get the statement we can propagate from into NAME skipping
249 : trivial copies. Returns the statement which defines the
250 : propagation source or NULL_TREE if there is no such one.
251 : If SINGLE_USE_ONLY is set considers only sources which have
252 : a single use chain up to NAME. If SINGLE_USE_P is non-null,
253 : it is set to whether the chain to NAME is a single use chain
254 : or not. SINGLE_USE_P is not written to if SINGLE_USE_ONLY is set. */
255 :
256 : static gimple *
257 28391684 : get_prop_source_stmt (tree name, bool single_use_only, bool *single_use_p)
258 : {
259 28391684 : bool single_use = true;
260 :
261 28392672 : do {
262 28392178 : gimple *def_stmt = SSA_NAME_DEF_STMT (name);
263 :
264 28392178 : if (!has_single_use (name))
265 : {
266 15499260 : single_use = false;
267 15499260 : if (single_use_only)
268 : return NULL;
269 : }
270 :
271 : /* If name is defined by a PHI node or is the default def, bail out. */
272 28390695 : if (!is_gimple_assign (def_stmt))
273 : return NULL;
274 :
275 : /* If def_stmt is a simple copy, continue looking. */
276 20022804 : if (gimple_assign_rhs_code (def_stmt) == SSA_NAME)
277 494 : name = gimple_assign_rhs1 (def_stmt);
278 : else
279 : {
280 20022310 : if (!single_use_only && single_use_p)
281 19699633 : *single_use_p = single_use;
282 :
283 20022310 : return def_stmt;
284 : }
285 494 : } while (1);
286 : }
287 :
288 : /* Checks if the destination ssa name in DEF_STMT can be used as
289 : propagation source. Returns true if so, otherwise false. */
290 :
291 : static bool
292 28134727 : can_propagate_from (gimple *def_stmt)
293 : {
294 28134727 : gcc_assert (is_gimple_assign (def_stmt));
295 :
296 : /* If the rhs has side-effects we cannot propagate from it. */
297 28134727 : if (gimple_has_volatile_ops (def_stmt))
298 : return false;
299 :
300 : /* If the rhs is a load we cannot propagate from it. */
301 27542203 : if (TREE_CODE_CLASS (gimple_assign_rhs_code (def_stmt)) == tcc_reference
302 27542203 : || TREE_CODE_CLASS (gimple_assign_rhs_code (def_stmt)) == tcc_declaration)
303 : return false;
304 :
305 : /* Constants can be always propagated. */
306 13575029 : if (gimple_assign_single_p (def_stmt)
307 13575029 : && is_gimple_min_invariant (gimple_assign_rhs1 (def_stmt)))
308 : return true;
309 :
310 : /* We cannot propagate ssa names that occur in abnormal phi nodes. */
311 13575029 : if (stmt_references_abnormal_ssa_name (def_stmt))
312 : return false;
313 :
314 : /* If the definition is a conversion of a pointer to a function type,
315 : then we cannot apply optimizations as some targets require
316 : function pointers to be canonicalized and in this case this
317 : optimization could eliminate a necessary canonicalization. */
318 13574338 : if (CONVERT_EXPR_CODE_P (gimple_assign_rhs_code (def_stmt)))
319 : {
320 3261203 : tree rhs = gimple_assign_rhs1 (def_stmt);
321 3261203 : if (FUNCTION_POINTER_TYPE_P (TREE_TYPE (rhs)))
322 : return false;
323 : }
324 :
325 : return true;
326 : }
327 :
328 : /* Remove a chain of dead statements starting at the definition of
329 : NAME. The chain is linked via the first operand of the defining statements.
330 : If NAME was replaced in its only use then this function can be used
331 : to clean up dead stmts. The function handles already released SSA
332 : names gracefully. */
333 :
334 : static void
335 241229 : remove_prop_source_from_use (tree name)
336 : {
337 302035 : gimple_stmt_iterator gsi;
338 302035 : gimple *stmt;
339 :
340 302035 : do {
341 302035 : basic_block bb;
342 :
343 302035 : if (SSA_NAME_IN_FREE_LIST (name)
344 301992 : || SSA_NAME_IS_DEFAULT_DEF (name)
345 602093 : || !has_zero_uses (name))
346 : break;
347 :
348 61263 : stmt = SSA_NAME_DEF_STMT (name);
349 61263 : if (gimple_code (stmt) == GIMPLE_PHI
350 61263 : || gimple_has_side_effects (stmt))
351 : break;
352 :
353 61263 : bb = gimple_bb (stmt);
354 61263 : gsi = gsi_for_stmt (stmt);
355 61263 : unlink_stmt_vdef (stmt);
356 61263 : if (gsi_remove (&gsi, true))
357 6 : bitmap_set_bit (to_purge, bb->index);
358 61263 : fwprop_invalidate_lattice (gimple_get_lhs (stmt));
359 61263 : release_defs (stmt);
360 :
361 61263 : name = is_gimple_assign (stmt) ? gimple_assign_rhs1 (stmt) : NULL_TREE;
362 61263 : } while (name && TREE_CODE (name) == SSA_NAME);
363 :
364 241229 : }
365 :
366 : /* Return the rhs of a gassign *STMT in a form of a single tree,
367 : converted to type TYPE.
368 :
369 : This should disappear, but is needed so we can combine expressions and use
370 : the fold() interfaces. Long term, we need to develop folding and combine
371 : routines that deal with gimple exclusively . */
372 :
373 : static tree
374 7404996 : rhs_to_tree (tree type, gimple *stmt)
375 : {
376 7404996 : location_t loc = gimple_location (stmt);
377 7404996 : enum tree_code code = gimple_assign_rhs_code (stmt);
378 7404996 : switch (get_gimple_rhs_class (code))
379 : {
380 13315 : case GIMPLE_TERNARY_RHS:
381 13315 : return fold_build3_loc (loc, code, type, gimple_assign_rhs1 (stmt),
382 : gimple_assign_rhs2 (stmt),
383 13315 : gimple_assign_rhs3 (stmt));
384 5054395 : case GIMPLE_BINARY_RHS:
385 5054395 : return fold_build2_loc (loc, code, type, gimple_assign_rhs1 (stmt),
386 5054395 : gimple_assign_rhs2 (stmt));
387 2054397 : case GIMPLE_UNARY_RHS:
388 2054397 : return build1 (code, type, gimple_assign_rhs1 (stmt));
389 282889 : case GIMPLE_SINGLE_RHS:
390 282889 : return gimple_assign_rhs1 (stmt);
391 0 : default:
392 0 : gcc_unreachable ();
393 : }
394 : }
395 :
396 : /* Combine OP0 CODE OP1 in the context of a COND_EXPR. Returns
397 : the folded result in a form suitable for COND_EXPR_COND or
398 : NULL_TREE, if there is no suitable simplified form. If
399 : INVARIANT_ONLY is true only gimple_min_invariant results are
400 : considered simplified. */
401 :
402 : static tree
403 8329544 : combine_cond_expr_cond (gimple *stmt, enum tree_code code, tree type,
404 : tree op0, tree op1, bool invariant_only)
405 : {
406 8329544 : tree t;
407 :
408 8329544 : gcc_assert (TREE_CODE_CLASS (code) == tcc_comparison);
409 :
410 8329544 : t = fold_binary_loc (gimple_location (stmt), code, type, op0, op1);
411 8329544 : if (!t)
412 : return NULL_TREE;
413 :
414 : /* Require that we got a boolean type out if we put one in. */
415 3616584 : gcc_assert (TREE_CODE (TREE_TYPE (t)) == TREE_CODE (type));
416 :
417 : /* Canonicalize the combined condition for use in a COND_EXPR. */
418 3616584 : t = canonicalize_cond_expr_cond (t);
419 :
420 : /* Bail out if we required an invariant but didn't get one. */
421 3616584 : if (!t || (invariant_only && !is_gimple_min_invariant (t)))
422 3377981 : return NULL_TREE;
423 :
424 : return t;
425 : }
426 :
427 : /* Combine the comparison OP0 CODE OP1 at LOC with the defining statements
428 : of its operand. Return a new comparison tree or NULL_TREE if there
429 : were no simplifying combines. */
430 :
431 : static tree
432 22333486 : forward_propagate_into_comparison_1 (gimple *stmt,
433 : enum tree_code code, tree type,
434 : tree op0, tree op1)
435 : {
436 22333486 : tree tmp = NULL_TREE;
437 22333486 : tree rhs0 = NULL_TREE, rhs1 = NULL_TREE;
438 22333486 : bool single_use0_p = false, single_use1_p = false;
439 :
440 : /* For comparisons use the first operand, that is likely to
441 : simplify comparisons against constants. */
442 22333486 : if (TREE_CODE (op0) == SSA_NAME)
443 : {
444 22297091 : gimple *def_stmt = get_prop_source_stmt (op0, false, &single_use0_p);
445 22297091 : if (def_stmt && can_propagate_from (def_stmt))
446 : {
447 5620365 : enum tree_code def_code = gimple_assign_rhs_code (def_stmt);
448 5620365 : bool invariant_only_p = !single_use0_p;
449 :
450 5620365 : rhs0 = rhs_to_tree (TREE_TYPE (op1), def_stmt);
451 :
452 : /* Always combine comparisons or conversions from booleans. */
453 5620365 : if (TREE_CODE (op1) == INTEGER_CST
454 5620365 : && ((CONVERT_EXPR_CODE_P (def_code)
455 894755 : && TREE_CODE (TREE_TYPE (TREE_OPERAND (rhs0, 0)))
456 : == BOOLEAN_TYPE)
457 3604916 : || TREE_CODE_CLASS (def_code) == tcc_comparison))
458 : invariant_only_p = false;
459 :
460 5620365 : tmp = combine_cond_expr_cond (stmt, code, type,
461 : rhs0, op1, invariant_only_p);
462 5620365 : if (tmp)
463 : return tmp;
464 : }
465 : }
466 :
467 : /* If that wasn't successful, try the second operand. */
468 22103549 : if (TREE_CODE (op1) == SSA_NAME)
469 : {
470 5505662 : gimple *def_stmt = get_prop_source_stmt (op1, false, &single_use1_p);
471 5505662 : if (def_stmt && can_propagate_from (def_stmt))
472 : {
473 1784631 : rhs1 = rhs_to_tree (TREE_TYPE (op0), def_stmt);
474 3569262 : tmp = combine_cond_expr_cond (stmt, code, type,
475 1784631 : op0, rhs1, !single_use1_p);
476 1784631 : if (tmp)
477 : return tmp;
478 : }
479 : }
480 :
481 : /* If that wasn't successful either, try both operands. */
482 22096796 : if (rhs0 != NULL_TREE
483 22096796 : && rhs1 != NULL_TREE)
484 924548 : tmp = combine_cond_expr_cond (stmt, code, type,
485 : rhs0, rhs1,
486 924548 : !(single_use0_p && single_use1_p));
487 :
488 : return tmp;
489 : }
490 :
491 : /* Propagate from the ssa name definition statements of the assignment
492 : from a comparison at *GSI into the conditional if that simplifies it.
493 : Returns true if the stmt was modified. */
494 :
495 : static bool
496 2662715 : forward_propagate_into_comparison (gimple_stmt_iterator *gsi)
497 : {
498 2662715 : gimple *stmt = gsi_stmt (*gsi);
499 2662715 : tree tmp;
500 2662715 : tree type = TREE_TYPE (gimple_assign_lhs (stmt));
501 2662715 : tree rhs1 = gimple_assign_rhs1 (stmt);
502 2662715 : tree rhs2 = gimple_assign_rhs2 (stmt);
503 :
504 : /* Combine the comparison with defining statements. */
505 2662715 : tmp = forward_propagate_into_comparison_1 (stmt,
506 : gimple_assign_rhs_code (stmt),
507 : type, rhs1, rhs2);
508 2662715 : if (tmp && useless_type_conversion_p (type, TREE_TYPE (tmp)))
509 : {
510 7211 : if (dump_file)
511 : {
512 0 : fprintf (dump_file, " Replaced '");
513 0 : print_gimple_expr (dump_file, stmt, 0);
514 0 : fprintf (dump_file, "' with '");
515 0 : print_generic_expr (dump_file, tmp);
516 0 : fprintf (dump_file, "'\n");
517 : }
518 7211 : gimple_assign_set_rhs_from_tree (gsi, tmp);
519 7211 : fold_stmt (gsi);
520 7211 : update_stmt (gsi_stmt (*gsi));
521 :
522 7211 : if (TREE_CODE (rhs1) == SSA_NAME)
523 7211 : remove_prop_source_from_use (rhs1);
524 7211 : if (TREE_CODE (rhs2) == SSA_NAME)
525 3024 : remove_prop_source_from_use (rhs2);
526 7211 : return true;
527 : }
528 :
529 : return false;
530 : }
531 :
532 : /* Propagate from the ssa name definition statements of COND_EXPR
533 : in GIMPLE_COND statement STMT into the conditional if that simplifies it.
534 : Returns zero if no statement was changed, one if there were
535 : changes and two if cfg_cleanup needs to run. */
536 :
537 : static int
538 19670771 : forward_propagate_into_gimple_cond (gcond *stmt)
539 : {
540 19670771 : tree tmp;
541 19670771 : enum tree_code code = gimple_cond_code (stmt);
542 19670771 : tree rhs1 = gimple_cond_lhs (stmt);
543 19670771 : tree rhs2 = gimple_cond_rhs (stmt);
544 :
545 : /* GIMPLE_COND will always be a comparison. */
546 19670771 : gcc_assert (TREE_CODE_CLASS (gimple_cond_code (stmt)) == tcc_comparison);
547 :
548 19670771 : tmp = forward_propagate_into_comparison_1 (stmt, code,
549 : boolean_type_node,
550 : rhs1, rhs2);
551 19670771 : if (tmp
552 19670771 : && is_gimple_condexpr_for_cond (tmp))
553 : {
554 225067 : if (dump_file)
555 : {
556 9 : fprintf (dump_file, " Replaced '");
557 9 : print_gimple_expr (dump_file, stmt, 0);
558 9 : fprintf (dump_file, "' with '");
559 9 : print_generic_expr (dump_file, tmp);
560 9 : fprintf (dump_file, "'\n");
561 : }
562 :
563 225067 : gimple_cond_set_condition_from_tree (stmt, unshare_expr (tmp));
564 225067 : update_stmt (stmt);
565 :
566 225067 : if (TREE_CODE (rhs1) == SSA_NAME)
567 225067 : remove_prop_source_from_use (rhs1);
568 225067 : if (TREE_CODE (rhs2) == SSA_NAME)
569 5926 : remove_prop_source_from_use (rhs2);
570 225067 : return is_gimple_min_invariant (tmp) ? 2 : 1;
571 : }
572 :
573 19445704 : if (canonicalize_bool_cond (stmt, gimple_bb (stmt)))
574 : return 1;
575 :
576 : return 0;
577 : }
578 :
579 : /* We've just substituted an ADDR_EXPR into stmt. Update all the
580 : relevant data structures to match. */
581 :
582 : static void
583 2007544 : tidy_after_forward_propagate_addr (gimple *stmt)
584 : {
585 : /* We may have turned a trapping insn into a non-trapping insn. */
586 2007544 : if (maybe_clean_or_replace_eh_stmt (stmt, stmt))
587 131 : bitmap_set_bit (to_purge, gimple_bb (stmt)->index);
588 :
589 2007544 : if (TREE_CODE (gimple_assign_rhs1 (stmt)) == ADDR_EXPR)
590 253466 : recompute_tree_invariant_for_addr_expr (gimple_assign_rhs1 (stmt));
591 2007544 : }
592 :
593 : /* NAME is a SSA_NAME representing DEF_RHS which is of the form
594 : ADDR_EXPR <whatever>.
595 :
596 : Try to forward propagate the ADDR_EXPR into the use USE_STMT.
597 : Often this will allow for removal of an ADDR_EXPR and INDIRECT_REF
598 : node or for recovery of array indexing from pointer arithmetic.
599 :
600 : Return true if the propagation was successful (the propagation can
601 : be not totally successful, yet things may have been changed). */
602 :
603 : static bool
604 2848461 : forward_propagate_addr_expr_1 (tree name, tree def_rhs,
605 : gimple_stmt_iterator *use_stmt_gsi,
606 : bool single_use_p)
607 : {
608 2848461 : tree lhs, rhs, rhs2, array_ref;
609 2848461 : gimple *use_stmt = gsi_stmt (*use_stmt_gsi);
610 2848461 : enum tree_code rhs_code;
611 2848461 : bool res = true;
612 :
613 2848461 : gcc_assert (TREE_CODE (def_rhs) == ADDR_EXPR);
614 :
615 2848461 : lhs = gimple_assign_lhs (use_stmt);
616 2848461 : rhs_code = gimple_assign_rhs_code (use_stmt);
617 2848461 : rhs = gimple_assign_rhs1 (use_stmt);
618 :
619 : /* Do not perform copy-propagation but recurse through copy chains. */
620 2848461 : if (TREE_CODE (lhs) == SSA_NAME
621 1419608 : && rhs_code == SSA_NAME)
622 8616 : return forward_propagate_addr_expr (lhs, def_rhs, single_use_p);
623 :
624 : /* The use statement could be a conversion. Recurse to the uses of the
625 : lhs as copyprop does not copy through pointer to integer to pointer
626 : conversions and FRE does not catch all cases either.
627 : Treat the case of a single-use name and
628 : a conversion to def_rhs type separate, though. */
629 2839845 : if (TREE_CODE (lhs) == SSA_NAME
630 1410992 : && CONVERT_EXPR_CODE_P (rhs_code))
631 : {
632 : /* If there is a point in a conversion chain where the types match
633 : so we can remove a conversion re-materialize the address here
634 : and stop. */
635 23939 : if (single_use_p
636 23939 : && useless_type_conversion_p (TREE_TYPE (lhs), TREE_TYPE (def_rhs)))
637 : {
638 1 : gimple_assign_set_rhs1 (use_stmt, unshare_expr (def_rhs));
639 1 : gimple_assign_set_rhs_code (use_stmt, TREE_CODE (def_rhs));
640 1 : return true;
641 : }
642 :
643 : /* Else recurse if the conversion preserves the address value. */
644 47876 : if ((INTEGRAL_TYPE_P (TREE_TYPE (lhs))
645 2 : || POINTER_TYPE_P (TREE_TYPE (lhs)))
646 47876 : && (TYPE_PRECISION (TREE_TYPE (lhs))
647 23938 : >= TYPE_PRECISION (TREE_TYPE (def_rhs))))
648 23871 : return forward_propagate_addr_expr (lhs, def_rhs, single_use_p);
649 :
650 : return false;
651 : }
652 :
653 : /* If this isn't a conversion chain from this on we only can propagate
654 : into compatible pointer contexts. */
655 2815906 : if (!types_compatible_p (TREE_TYPE (name), TREE_TYPE (def_rhs)))
656 : return false;
657 :
658 : /* Propagate through constant pointer adjustments. */
659 2795308 : if (TREE_CODE (lhs) == SSA_NAME
660 1367628 : && rhs_code == POINTER_PLUS_EXPR
661 1367628 : && rhs == name
662 2962167 : && TREE_CODE (gimple_assign_rhs2 (use_stmt)) == INTEGER_CST)
663 : {
664 118428 : tree new_def_rhs;
665 : /* As we come here with non-invariant addresses in def_rhs we need
666 : to make sure we can build a valid constant offsetted address
667 : for further propagation. Simply rely on fold building that
668 : and check after the fact. */
669 118428 : new_def_rhs = fold_build2 (MEM_REF, TREE_TYPE (TREE_TYPE (rhs)),
670 : def_rhs,
671 : fold_convert (ptr_type_node,
672 : gimple_assign_rhs2 (use_stmt)));
673 118428 : if (TREE_CODE (new_def_rhs) == MEM_REF
674 118428 : && !is_gimple_mem_ref_addr (TREE_OPERAND (new_def_rhs, 0)))
675 : return false;
676 114467 : new_def_rhs = build1 (ADDR_EXPR, TREE_TYPE (rhs), new_def_rhs);
677 :
678 : /* Recurse. If we could propagate into all uses of lhs do not
679 : bother to replace into the current use but just pretend we did. */
680 114467 : if (forward_propagate_addr_expr (lhs, new_def_rhs, single_use_p))
681 : return true;
682 :
683 38371 : if (useless_type_conversion_p (TREE_TYPE (lhs),
684 38371 : TREE_TYPE (new_def_rhs)))
685 38371 : gimple_assign_set_rhs_with_ops (use_stmt_gsi, TREE_CODE (new_def_rhs),
686 : new_def_rhs);
687 0 : else if (is_gimple_min_invariant (new_def_rhs))
688 0 : gimple_assign_set_rhs_with_ops (use_stmt_gsi, NOP_EXPR, new_def_rhs);
689 : else
690 : return false;
691 38371 : gcc_assert (gsi_stmt (*use_stmt_gsi) == use_stmt);
692 38371 : update_stmt (use_stmt);
693 38371 : return true;
694 : }
695 :
696 : /* Now strip away any outer COMPONENT_REF/ARRAY_REF nodes from the LHS.
697 : ADDR_EXPR will not appear on the LHS. */
698 2676880 : tree *lhsp = gimple_assign_lhs_ptr (use_stmt);
699 4041822 : while (handled_component_p (*lhsp))
700 1364942 : lhsp = &TREE_OPERAND (*lhsp, 0);
701 2676880 : lhs = *lhsp;
702 :
703 : /* Now see if the LHS node is a MEM_REF using NAME. If so,
704 : propagate the ADDR_EXPR into the use of NAME and fold the result. */
705 2676880 : if (TREE_CODE (lhs) == MEM_REF
706 2676880 : && TREE_OPERAND (lhs, 0) == name)
707 : {
708 903734 : tree def_rhs_base;
709 903734 : poly_int64 def_rhs_offset;
710 : /* If the address is invariant we can always fold it. */
711 903734 : if ((def_rhs_base = get_addr_base_and_unit_offset (TREE_OPERAND (def_rhs, 0),
712 : &def_rhs_offset)))
713 : {
714 856628 : poly_offset_int off = mem_ref_offset (lhs);
715 856628 : tree new_ptr;
716 856628 : off += def_rhs_offset;
717 856628 : if (TREE_CODE (def_rhs_base) == MEM_REF)
718 : {
719 835250 : off += mem_ref_offset (def_rhs_base);
720 835250 : new_ptr = TREE_OPERAND (def_rhs_base, 0);
721 : }
722 : else
723 21378 : new_ptr = build_fold_addr_expr (def_rhs_base);
724 856628 : TREE_OPERAND (lhs, 0) = new_ptr;
725 856628 : TREE_OPERAND (lhs, 1)
726 856628 : = wide_int_to_tree (TREE_TYPE (TREE_OPERAND (lhs, 1)), off);
727 856628 : tidy_after_forward_propagate_addr (use_stmt);
728 : /* Continue propagating into the RHS if this was not the only use. */
729 856628 : if (single_use_p)
730 231194 : return true;
731 : }
732 : /* If the LHS is a plain dereference and the value type is the same as
733 : that of the pointed-to type of the address we can put the
734 : dereferenced address on the LHS preserving the original alias-type. */
735 47106 : else if (integer_zerop (TREE_OPERAND (lhs, 1))
736 18520 : && ((gimple_assign_lhs (use_stmt) == lhs
737 14893 : && useless_type_conversion_p
738 14893 : (TREE_TYPE (TREE_OPERAND (def_rhs, 0)),
739 14893 : TREE_TYPE (gimple_assign_rhs1 (use_stmt))))
740 13598 : || types_compatible_p (TREE_TYPE (lhs),
741 13598 : TREE_TYPE (TREE_OPERAND (def_rhs, 0))))
742 : /* Don't forward anything into clobber stmts if it would result
743 : in the lhs no longer being a MEM_REF. */
744 55284 : && (!gimple_clobber_p (use_stmt)
745 164 : || TREE_CODE (TREE_OPERAND (def_rhs, 0)) == MEM_REF))
746 : {
747 8014 : tree *def_rhs_basep = &TREE_OPERAND (def_rhs, 0);
748 8014 : tree new_offset, new_base, saved, new_lhs;
749 29081 : while (handled_component_p (*def_rhs_basep))
750 13053 : def_rhs_basep = &TREE_OPERAND (*def_rhs_basep, 0);
751 8014 : saved = *def_rhs_basep;
752 8014 : if (TREE_CODE (*def_rhs_basep) == MEM_REF)
753 : {
754 3941 : new_base = TREE_OPERAND (*def_rhs_basep, 0);
755 3941 : new_offset = fold_convert (TREE_TYPE (TREE_OPERAND (lhs, 1)),
756 : TREE_OPERAND (*def_rhs_basep, 1));
757 : }
758 : else
759 : {
760 4073 : new_base = build_fold_addr_expr (*def_rhs_basep);
761 4073 : new_offset = TREE_OPERAND (lhs, 1);
762 : }
763 8014 : tree atype = TREE_TYPE (*def_rhs_basep);
764 8014 : if (TYPE_ALIGN (TREE_TYPE (lhs)) < TYPE_ALIGN (atype))
765 312 : atype = build_aligned_type (atype, TYPE_ALIGN (TREE_TYPE (lhs)));
766 8014 : *def_rhs_basep = build2 (MEM_REF, atype, new_base, new_offset);
767 8014 : TREE_THIS_VOLATILE (*def_rhs_basep) = TREE_THIS_VOLATILE (lhs);
768 8014 : TREE_SIDE_EFFECTS (*def_rhs_basep) = TREE_SIDE_EFFECTS (lhs);
769 8014 : TREE_THIS_NOTRAP (*def_rhs_basep) = TREE_THIS_NOTRAP (lhs);
770 8014 : new_lhs = unshare_expr (TREE_OPERAND (def_rhs, 0));
771 8014 : *lhsp = new_lhs;
772 8014 : TREE_THIS_VOLATILE (new_lhs) = TREE_THIS_VOLATILE (lhs);
773 8014 : TREE_SIDE_EFFECTS (new_lhs) = TREE_SIDE_EFFECTS (lhs);
774 8014 : *def_rhs_basep = saved;
775 8014 : tidy_after_forward_propagate_addr (use_stmt);
776 : /* Continue propagating into the RHS if this was not the
777 : only use. */
778 8014 : if (single_use_p)
779 : return true;
780 : }
781 : else
782 : /* We can have a struct assignment dereferencing our name twice.
783 : Note that we didn't propagate into the lhs to not falsely
784 : claim we did when propagating into the rhs. */
785 : res = false;
786 : }
787 :
788 : /* Strip away any outer COMPONENT_REF, ARRAY_REF or ADDR_EXPR
789 : nodes from the RHS. */
790 2441479 : tree *rhsp = gimple_assign_rhs1_ptr (use_stmt);
791 2441479 : if (TREE_CODE (*rhsp) == ADDR_EXPR)
792 241434 : rhsp = &TREE_OPERAND (*rhsp, 0);
793 3461844 : while (handled_component_p (*rhsp))
794 1020365 : rhsp = &TREE_OPERAND (*rhsp, 0);
795 2441479 : rhs = *rhsp;
796 :
797 : /* Now see if the RHS node is a MEM_REF using NAME. If so,
798 : propagate the ADDR_EXPR into the use of NAME and fold the result. */
799 2441479 : if (TREE_CODE (rhs) == MEM_REF
800 2441479 : && TREE_OPERAND (rhs, 0) == name)
801 : {
802 1164654 : tree def_rhs_base;
803 1164654 : poly_int64 def_rhs_offset;
804 1164654 : if ((def_rhs_base = get_addr_base_and_unit_offset (TREE_OPERAND (def_rhs, 0),
805 : &def_rhs_offset)))
806 : {
807 1127633 : poly_offset_int off = mem_ref_offset (rhs);
808 1127633 : tree new_ptr;
809 1127633 : off += def_rhs_offset;
810 1127633 : if (TREE_CODE (def_rhs_base) == MEM_REF)
811 : {
812 1101147 : off += mem_ref_offset (def_rhs_base);
813 1101147 : new_ptr = TREE_OPERAND (def_rhs_base, 0);
814 : }
815 : else
816 26486 : new_ptr = build_fold_addr_expr (def_rhs_base);
817 1127633 : TREE_OPERAND (rhs, 0) = new_ptr;
818 1127633 : TREE_OPERAND (rhs, 1)
819 1127633 : = wide_int_to_tree (TREE_TYPE (TREE_OPERAND (rhs, 1)), off);
820 1127633 : fold_stmt_inplace (use_stmt_gsi);
821 1127633 : tidy_after_forward_propagate_addr (use_stmt);
822 1127633 : return res;
823 : }
824 : /* If the RHS is a plain dereference and the value type is the same as
825 : that of the pointed-to type of the address we can put the
826 : dereferenced address on the RHS preserving the original alias-type. */
827 37021 : else if (integer_zerop (TREE_OPERAND (rhs, 1))
828 37021 : && ((gimple_assign_rhs1 (use_stmt) == rhs
829 20263 : && useless_type_conversion_p
830 20263 : (TREE_TYPE (gimple_assign_lhs (use_stmt)),
831 20263 : TREE_TYPE (TREE_OPERAND (def_rhs, 0))))
832 22533 : || types_compatible_p (TREE_TYPE (rhs),
833 22533 : TREE_TYPE (TREE_OPERAND (def_rhs, 0)))))
834 : {
835 15269 : tree *def_rhs_basep = &TREE_OPERAND (def_rhs, 0);
836 15269 : tree new_offset, new_base, saved, new_rhs;
837 54333 : while (handled_component_p (*def_rhs_basep))
838 23795 : def_rhs_basep = &TREE_OPERAND (*def_rhs_basep, 0);
839 15269 : saved = *def_rhs_basep;
840 15269 : if (TREE_CODE (*def_rhs_basep) == MEM_REF)
841 : {
842 7359 : new_base = TREE_OPERAND (*def_rhs_basep, 0);
843 7359 : new_offset = fold_convert (TREE_TYPE (TREE_OPERAND (rhs, 1)),
844 : TREE_OPERAND (*def_rhs_basep, 1));
845 : }
846 : else
847 : {
848 7910 : new_base = build_fold_addr_expr (*def_rhs_basep);
849 7910 : new_offset = TREE_OPERAND (rhs, 1);
850 : }
851 15269 : tree atype = TREE_TYPE (*def_rhs_basep);
852 15269 : if (TYPE_ALIGN (TREE_TYPE (rhs)) < TYPE_ALIGN (atype))
853 526 : atype = build_aligned_type (atype, TYPE_ALIGN (TREE_TYPE (rhs)));
854 15269 : *def_rhs_basep = build2 (MEM_REF, atype, new_base, new_offset);
855 15269 : TREE_THIS_VOLATILE (*def_rhs_basep) = TREE_THIS_VOLATILE (rhs);
856 15269 : TREE_SIDE_EFFECTS (*def_rhs_basep) = TREE_SIDE_EFFECTS (rhs);
857 15269 : TREE_THIS_NOTRAP (*def_rhs_basep) = TREE_THIS_NOTRAP (rhs);
858 15269 : new_rhs = unshare_expr (TREE_OPERAND (def_rhs, 0));
859 15269 : *rhsp = new_rhs;
860 15269 : TREE_THIS_VOLATILE (new_rhs) = TREE_THIS_VOLATILE (rhs);
861 15269 : TREE_SIDE_EFFECTS (new_rhs) = TREE_SIDE_EFFECTS (rhs);
862 15269 : *def_rhs_basep = saved;
863 15269 : fold_stmt_inplace (use_stmt_gsi);
864 15269 : tidy_after_forward_propagate_addr (use_stmt);
865 15269 : return res;
866 : }
867 : }
868 :
869 : /* If the use of the ADDR_EXPR is not a POINTER_PLUS_EXPR, there
870 : is nothing to do. */
871 1298577 : if (gimple_assign_rhs_code (use_stmt) != POINTER_PLUS_EXPR
872 1298577 : || gimple_assign_rhs1 (use_stmt) != name)
873 : return false;
874 :
875 : /* The remaining cases are all for turning pointer arithmetic into
876 : array indexing. They only apply when we have the address of
877 : element zero in an array. If that is not the case then there
878 : is nothing to do. */
879 48431 : array_ref = TREE_OPERAND (def_rhs, 0);
880 48431 : if ((TREE_CODE (array_ref) != ARRAY_REF
881 4570 : || TREE_CODE (TREE_TYPE (TREE_OPERAND (array_ref, 0))) != ARRAY_TYPE
882 4570 : || TREE_CODE (TREE_OPERAND (array_ref, 1)) != INTEGER_CST)
883 49930 : && TREE_CODE (TREE_TYPE (array_ref)) != ARRAY_TYPE)
884 : return false;
885 :
886 24807 : rhs2 = gimple_assign_rhs2 (use_stmt);
887 : /* Optimize &x[C1] p+ C2 to &x p+ C3 with C3 = C1 * element_size + C2. */
888 24807 : if (TREE_CODE (rhs2) == INTEGER_CST)
889 : {
890 0 : tree new_rhs = build1_loc (gimple_location (use_stmt),
891 0 : ADDR_EXPR, TREE_TYPE (def_rhs),
892 0 : fold_build2 (MEM_REF,
893 : TREE_TYPE (TREE_TYPE (def_rhs)),
894 : unshare_expr (def_rhs),
895 : fold_convert (ptr_type_node,
896 : rhs2)));
897 0 : gimple_assign_set_rhs_from_tree (use_stmt_gsi, new_rhs);
898 0 : use_stmt = gsi_stmt (*use_stmt_gsi);
899 0 : update_stmt (use_stmt);
900 0 : tidy_after_forward_propagate_addr (use_stmt);
901 0 : return true;
902 : }
903 :
904 : return false;
905 : }
906 :
907 : /* STMT is a statement of the form SSA_NAME = ADDR_EXPR <whatever>.
908 :
909 : Try to forward propagate the ADDR_EXPR into all uses of the SSA_NAME.
910 : Often this will allow for removal of an ADDR_EXPR and INDIRECT_REF
911 : node or for recovery of array indexing from pointer arithmetic.
912 :
913 : PARENT_SINGLE_USE_P tells if, when in a recursive invocation, NAME was
914 : the single use in the previous invocation. Pass true when calling
915 : this as toplevel.
916 :
917 : Returns true, if all uses have been propagated into. */
918 :
919 : static bool
920 3301377 : forward_propagate_addr_expr (tree name, tree rhs, bool parent_single_use_p)
921 : {
922 3301377 : bool all = true;
923 3301377 : bool single_use_p = parent_single_use_p && has_single_use (name);
924 :
925 17458870 : for (gimple *use_stmt : gather_imm_use_stmts (name))
926 : {
927 7554739 : bool result;
928 7554739 : tree use_rhs;
929 :
930 : /* If the use is not in a simple assignment statement, then
931 : there is nothing we can do. */
932 7554739 : if (!is_gimple_assign (use_stmt))
933 : {
934 4706278 : if (!is_gimple_debug (use_stmt))
935 1929180 : all = false;
936 4706278 : continue;
937 : }
938 :
939 2848461 : gimple_stmt_iterator gsi = gsi_for_stmt (use_stmt);
940 2848461 : result = forward_propagate_addr_expr_1 (name, rhs, &gsi,
941 : single_use_p);
942 : /* If the use has moved to a different statement adjust
943 : the update machinery for the old statement too. */
944 2848461 : if (use_stmt != gsi_stmt (gsi))
945 : {
946 0 : update_stmt (use_stmt);
947 0 : use_stmt = gsi_stmt (gsi);
948 : }
949 2848461 : update_stmt (use_stmt);
950 2848461 : all &= result;
951 :
952 : /* Remove intermediate now unused copy and conversion chains. */
953 2848461 : use_rhs = gimple_assign_rhs1 (use_stmt);
954 2848461 : if (result
955 1497981 : && TREE_CODE (gimple_assign_lhs (use_stmt)) == SSA_NAME
956 1249937 : && TREE_CODE (use_rhs) == SSA_NAME
957 2929767 : && has_zero_uses (gimple_assign_lhs (use_stmt)))
958 : {
959 81306 : gimple_stmt_iterator gsi = gsi_for_stmt (use_stmt);
960 81306 : fwprop_invalidate_lattice (gimple_get_lhs (use_stmt));
961 81306 : release_defs (use_stmt);
962 81306 : gsi_remove (&gsi, true);
963 : }
964 3301377 : }
965 :
966 3301377 : return all && has_zero_uses (name);
967 : }
968 :
969 :
970 : /* Helper function for simplify_gimple_switch. Remove case labels that
971 : have values outside the range of the new type. */
972 :
973 : static void
974 11706 : simplify_gimple_switch_label_vec (gswitch *stmt, tree index_type,
975 : vec<std::pair<int, int> > &edges_to_remove)
976 : {
977 11706 : unsigned int branch_num = gimple_switch_num_labels (stmt);
978 11706 : auto_vec<tree> labels (branch_num);
979 11706 : unsigned int i, len;
980 :
981 : /* Collect the existing case labels in a VEC, and preprocess it as if
982 : we are gimplifying a GENERIC SWITCH_EXPR. */
983 73642 : for (i = 1; i < branch_num; i++)
984 50230 : labels.quick_push (gimple_switch_label (stmt, i));
985 11706 : preprocess_case_label_vec_for_gimple (labels, index_type, NULL);
986 :
987 : /* If any labels were removed, replace the existing case labels
988 : in the GIMPLE_SWITCH statement with the correct ones.
989 : Note that the type updates were done in-place on the case labels,
990 : so we only have to replace the case labels in the GIMPLE_SWITCH
991 : if the number of labels changed. */
992 11706 : len = labels.length ();
993 11706 : if (len < branch_num - 1)
994 : {
995 0 : bitmap target_blocks;
996 0 : edge_iterator ei;
997 0 : edge e;
998 :
999 : /* Corner case: *all* case labels have been removed as being
1000 : out-of-range for INDEX_TYPE. Push one label and let the
1001 : CFG cleanups deal with this further. */
1002 0 : if (len == 0)
1003 : {
1004 0 : tree label, elt;
1005 :
1006 0 : label = CASE_LABEL (gimple_switch_default_label (stmt));
1007 0 : elt = build_case_label (build_int_cst (index_type, 0), NULL, label);
1008 0 : labels.quick_push (elt);
1009 0 : len = 1;
1010 : }
1011 :
1012 0 : for (i = 0; i < labels.length (); i++)
1013 0 : gimple_switch_set_label (stmt, i + 1, labels[i]);
1014 0 : for (i++ ; i < branch_num; i++)
1015 0 : gimple_switch_set_label (stmt, i, NULL_TREE);
1016 0 : gimple_switch_set_num_labels (stmt, len + 1);
1017 :
1018 : /* Cleanup any edges that are now dead. */
1019 0 : target_blocks = BITMAP_ALLOC (NULL);
1020 0 : for (i = 0; i < gimple_switch_num_labels (stmt); i++)
1021 : {
1022 0 : tree elt = gimple_switch_label (stmt, i);
1023 0 : basic_block target = label_to_block (cfun, CASE_LABEL (elt));
1024 0 : bitmap_set_bit (target_blocks, target->index);
1025 : }
1026 0 : for (ei = ei_start (gimple_bb (stmt)->succs); (e = ei_safe_edge (ei)); )
1027 : {
1028 0 : if (! bitmap_bit_p (target_blocks, e->dest->index))
1029 0 : edges_to_remove.safe_push (std::make_pair (e->src->index,
1030 0 : e->dest->index));
1031 : else
1032 0 : ei_next (&ei);
1033 : }
1034 0 : BITMAP_FREE (target_blocks);
1035 : }
1036 11706 : }
1037 :
1038 : /* STMT is a SWITCH_EXPR for which we attempt to find equivalent forms of
1039 : the condition which we may be able to optimize better. */
1040 :
1041 : static bool
1042 107050 : simplify_gimple_switch (gswitch *stmt,
1043 : vec<std::pair<int, int> > &edges_to_remove,
1044 : bitmap simple_dce_worklist)
1045 : {
1046 : /* The optimization that we really care about is removing unnecessary
1047 : casts. That will let us do much better in propagating the inferred
1048 : constant at the switch target. */
1049 107050 : tree cond = gimple_switch_index (stmt);
1050 107050 : if (TREE_CODE (cond) == SSA_NAME)
1051 : {
1052 107049 : gimple *def_stmt = SSA_NAME_DEF_STMT (cond);
1053 107049 : if (gimple_assign_cast_p (def_stmt))
1054 : {
1055 12630 : tree def = gimple_assign_rhs1 (def_stmt);
1056 12630 : if (TREE_CODE (def) != SSA_NAME)
1057 : return false;
1058 :
1059 : /* If we have an extension or sign-change that preserves the
1060 : values we check against then we can copy the source value into
1061 : the switch. */
1062 12630 : tree ti = TREE_TYPE (def);
1063 12630 : if (INTEGRAL_TYPE_P (ti)
1064 12630 : && TYPE_PRECISION (ti) <= TYPE_PRECISION (TREE_TYPE (cond)))
1065 : {
1066 12385 : size_t n = gimple_switch_num_labels (stmt);
1067 12385 : tree min = NULL_TREE, max = NULL_TREE;
1068 12385 : if (n > 1)
1069 : {
1070 12385 : min = CASE_LOW (gimple_switch_label (stmt, 1));
1071 12385 : if (CASE_HIGH (gimple_switch_label (stmt, n - 1)))
1072 159 : max = CASE_HIGH (gimple_switch_label (stmt, n - 1));
1073 : else
1074 12226 : max = CASE_LOW (gimple_switch_label (stmt, n - 1));
1075 : }
1076 12385 : if ((!min || int_fits_type_p (min, ti))
1077 12381 : && (!max || int_fits_type_p (max, ti)))
1078 : {
1079 11706 : bitmap_set_bit (simple_dce_worklist,
1080 11706 : SSA_NAME_VERSION (cond));
1081 11706 : gimple_switch_set_index (stmt, def);
1082 11706 : simplify_gimple_switch_label_vec (stmt, ti,
1083 : edges_to_remove);
1084 11706 : update_stmt (stmt);
1085 11706 : return true;
1086 : }
1087 : }
1088 : }
1089 : }
1090 :
1091 : return false;
1092 : }
1093 :
1094 : /* For pointers p2 and p1 return p2 - p1 if the
1095 : difference is known and constant, otherwise return NULL. */
1096 :
1097 : static tree
1098 5360 : constant_pointer_difference (tree p1, tree p2)
1099 : {
1100 5360 : int i, j;
1101 : #define CPD_ITERATIONS 5
1102 5360 : tree exps[2][CPD_ITERATIONS];
1103 5360 : tree offs[2][CPD_ITERATIONS];
1104 5360 : int cnt[2];
1105 :
1106 16080 : for (i = 0; i < 2; i++)
1107 : {
1108 10720 : tree p = i ? p1 : p2;
1109 10720 : tree off = size_zero_node;
1110 10720 : gimple *stmt;
1111 10720 : enum tree_code code;
1112 :
1113 : /* For each of p1 and p2 we need to iterate at least
1114 : twice, to handle ADDR_EXPR directly in p1/p2,
1115 : SSA_NAME with ADDR_EXPR or POINTER_PLUS_EXPR etc.
1116 : on definition's stmt RHS. Iterate a few extra times. */
1117 10720 : j = 0;
1118 12659 : do
1119 : {
1120 12659 : if (!POINTER_TYPE_P (TREE_TYPE (p)))
1121 : break;
1122 12653 : if (TREE_CODE (p) == ADDR_EXPR)
1123 : {
1124 9241 : tree q = TREE_OPERAND (p, 0);
1125 9241 : poly_int64 offset;
1126 9241 : tree base = get_addr_base_and_unit_offset (q, &offset);
1127 9241 : if (base)
1128 : {
1129 8425 : q = base;
1130 8425 : if (maybe_ne (offset, 0))
1131 3537 : off = size_binop (PLUS_EXPR, off, size_int (offset));
1132 : }
1133 9241 : if (TREE_CODE (q) == MEM_REF
1134 9241 : && TREE_CODE (TREE_OPERAND (q, 0)) == SSA_NAME)
1135 : {
1136 213 : p = TREE_OPERAND (q, 0);
1137 213 : off = size_binop (PLUS_EXPR, off,
1138 : wide_int_to_tree (sizetype,
1139 : mem_ref_offset (q)));
1140 : }
1141 : else
1142 : {
1143 9028 : exps[i][j] = q;
1144 9028 : offs[i][j++] = off;
1145 9028 : break;
1146 : }
1147 : }
1148 3625 : if (TREE_CODE (p) != SSA_NAME)
1149 : break;
1150 3625 : exps[i][j] = p;
1151 3625 : offs[i][j++] = off;
1152 3625 : if (j == CPD_ITERATIONS)
1153 : break;
1154 3625 : stmt = SSA_NAME_DEF_STMT (p);
1155 3625 : if (!is_gimple_assign (stmt) || gimple_assign_lhs (stmt) != p)
1156 : break;
1157 2843 : code = gimple_assign_rhs_code (stmt);
1158 2843 : if (code == POINTER_PLUS_EXPR)
1159 : {
1160 1485 : if (TREE_CODE (gimple_assign_rhs2 (stmt)) != INTEGER_CST)
1161 : break;
1162 954 : off = size_binop (PLUS_EXPR, off, gimple_assign_rhs2 (stmt));
1163 954 : p = gimple_assign_rhs1 (stmt);
1164 : }
1165 1358 : else if (code == ADDR_EXPR || CONVERT_EXPR_CODE_P (code))
1166 985 : p = gimple_assign_rhs1 (stmt);
1167 : else
1168 : break;
1169 : }
1170 : while (1);
1171 10720 : cnt[i] = j;
1172 : }
1173 :
1174 7441 : for (i = 0; i < cnt[0]; i++)
1175 9877 : for (j = 0; j < cnt[1]; j++)
1176 7796 : if (exps[0][i] == exps[1][j])
1177 4467 : return size_binop (MINUS_EXPR, offs[0][i], offs[1][j]);
1178 :
1179 : return NULL_TREE;
1180 : }
1181 :
1182 : /* Helper function for optimize_aggr_zeroprop.
1183 : Props the zeroing (memset, VAL) that was done in DEST+OFFSET:LEN
1184 : (DEFSTMT) into the STMT. Returns true if the STMT was updated. */
1185 : static void
1186 22510976 : optimize_aggr_zeroprop_1 (gimple *defstmt, gimple *stmt,
1187 : tree dest, poly_int64 offset, tree val,
1188 : poly_offset_int len)
1189 : {
1190 22510976 : tree src2;
1191 22510976 : tree len2 = NULL_TREE;
1192 22510976 : poly_int64 offset2;
1193 :
1194 22510976 : if (gimple_call_builtin_p (stmt, BUILT_IN_MEMCPY)
1195 22136 : && TREE_CODE (gimple_call_arg (stmt, 1)) == ADDR_EXPR
1196 22524987 : && poly_int_tree_p (gimple_call_arg (stmt, 2)))
1197 : {
1198 12977 : src2 = TREE_OPERAND (gimple_call_arg (stmt, 1), 0);
1199 12977 : len2 = gimple_call_arg (stmt, 2);
1200 : }
1201 22497999 : else if (gimple_assign_load_p (stmt) && gimple_store_p (stmt))
1202 : {
1203 1908170 : src2 = gimple_assign_rhs1 (stmt);
1204 1908170 : len2 = (TREE_CODE (src2) == COMPONENT_REF
1205 1908170 : ? DECL_SIZE_UNIT (TREE_OPERAND (src2, 1))
1206 1740551 : : TYPE_SIZE_UNIT (TREE_TYPE (src2)));
1207 : /* Can only handle zero memsets. */
1208 1908170 : if (!integer_zerop (val))
1209 22488353 : return;
1210 : }
1211 : else
1212 20589829 : return;
1213 :
1214 1920141 : if (len2 == NULL_TREE
1215 1920141 : || !poly_int_tree_p (len2))
1216 : return;
1217 :
1218 1920141 : src2 = get_addr_base_and_unit_offset (src2, &offset2);
1219 1920141 : if (src2 == NULL_TREE
1220 1920141 : || maybe_lt (offset2, offset))
1221 : return;
1222 :
1223 875886 : if (!operand_equal_p (dest, src2, 0))
1224 : return;
1225 :
1226 : /* [ dest + offset, dest + offset + len - 1 ] is set to val.
1227 : Make sure that
1228 : [ dest + offset2, dest + offset2 + len2 - 1 ] is a subset of that. */
1229 132526 : if (maybe_gt (wi::to_poly_offset (len2) + (offset2 - offset),
1230 : len))
1231 : return;
1232 :
1233 22623 : if (dump_file && (dump_flags & TDF_DETAILS))
1234 : {
1235 32 : fprintf (dump_file, "Simplified\n ");
1236 32 : print_gimple_stmt (dump_file, stmt, 0, dump_flags);
1237 32 : fprintf (dump_file, "after previous\n ");
1238 32 : print_gimple_stmt (dump_file, defstmt, 0, dump_flags);
1239 : }
1240 22623 : gimple *orig_stmt = stmt;
1241 : /* For simplicity, don't change the kind of the stmt,
1242 : turn dest = src; into dest = {}; and memcpy (&dest, &src, len);
1243 : into memset (&dest, val, len);
1244 : In theory we could change dest = src into memset if dest
1245 : is addressable (maybe beneficial if val is not 0), or
1246 : memcpy (&dest, &src, len) into dest = {} if len is the size
1247 : of dest, dest isn't volatile. */
1248 22623 : if (is_gimple_assign (stmt))
1249 : {
1250 22618 : tree ctor_type = TREE_TYPE (gimple_assign_lhs (stmt));
1251 22618 : tree ctor = build_constructor (ctor_type, NULL);
1252 22618 : gimple_stmt_iterator gsi = gsi_for_stmt (stmt);
1253 22618 : gimple_assign_set_rhs_from_tree (&gsi, ctor);
1254 22618 : update_stmt (stmt);
1255 22618 : statistics_counter_event (cfun, "copy zeroing propagation of aggregate", 1);
1256 : }
1257 : else /* If stmt is memcpy, transform it into memset. */
1258 : {
1259 5 : gcall *call = as_a <gcall *> (stmt);
1260 5 : tree fndecl = builtin_decl_implicit (BUILT_IN_MEMSET);
1261 5 : gimple_call_set_fndecl (call, fndecl);
1262 5 : gimple_call_set_fntype (call, TREE_TYPE (fndecl));
1263 5 : gimple_call_set_arg (call, 1, val);
1264 5 : update_stmt (stmt);
1265 5 : statistics_counter_event (cfun, "memcpy to memset changed", 1);
1266 : }
1267 :
1268 22623 : if (dump_file && (dump_flags & TDF_DETAILS))
1269 : {
1270 32 : fprintf (dump_file, "into\n ");
1271 32 : print_gimple_stmt (dump_file, stmt, 0, dump_flags);
1272 : }
1273 :
1274 : /* Mark the bb for eh cleanup if needed. */
1275 22623 : if (maybe_clean_or_replace_eh_stmt (orig_stmt, stmt))
1276 6 : bitmap_set_bit (to_purge, gimple_bb (stmt)->index);
1277 : }
1278 :
1279 : /* Optimize
1280 : a = {}; // DEST = value ;; LEN(nullptr)
1281 : b = a;
1282 : into
1283 : a = {};
1284 : b = {};
1285 : Similarly for memset (&a, ..., sizeof (a)); instead of a = {};
1286 : and/or memcpy (&b, &a, sizeof (a)); instead of b = a; */
1287 :
1288 : static void
1289 31389573 : optimize_aggr_zeroprop (gimple *stmt, bool full_walk)
1290 : {
1291 31389573 : ao_ref read;
1292 62779146 : if (gimple_has_volatile_ops (stmt))
1293 27333649 : return;
1294 :
1295 30463407 : tree dest = NULL_TREE;
1296 30463407 : tree val = integer_zero_node;
1297 30463407 : tree len = NULL_TREE;
1298 30463407 : bool can_use_tbba = true;
1299 :
1300 30463407 : if (gimple_call_builtin_p (stmt, BUILT_IN_MEMSET)
1301 111851 : && TREE_CODE (gimple_call_arg (stmt, 0)) == ADDR_EXPR
1302 56975 : && TREE_CODE (gimple_call_arg (stmt, 1)) == INTEGER_CST
1303 30518066 : && poly_int_tree_p (gimple_call_arg (stmt, 2)))
1304 : {
1305 51894 : dest = TREE_OPERAND (gimple_call_arg (stmt, 0), 0);
1306 51894 : len = gimple_call_arg (stmt, 2);
1307 51894 : val = gimple_call_arg (stmt, 1);
1308 51894 : ao_ref_init_from_ptr_and_size (&read, gimple_call_arg (stmt, 0), len);
1309 51894 : can_use_tbba = false;
1310 : }
1311 30411513 : else if (gimple_store_p (stmt)
1312 30351396 : && gimple_assign_single_p (stmt)
1313 60762909 : && TREE_CODE (gimple_assign_rhs1 (stmt)) == STRING_CST)
1314 : {
1315 26377 : tree str = gimple_assign_rhs1 (stmt);
1316 26377 : dest = gimple_assign_lhs (stmt);
1317 26377 : ao_ref_init (&read, dest);
1318 : /* The string must contain all null char's for now. */
1319 31719 : for (int i = 0; i < TREE_STRING_LENGTH (str); i++)
1320 : {
1321 29131 : if (TREE_STRING_POINTER (str)[i] != 0)
1322 : {
1323 : dest = NULL_TREE;
1324 : break;
1325 : }
1326 : }
1327 : }
1328 : /* A store of integer (scalar, vector or complex) zeros is
1329 : a zero store. */
1330 30385136 : else if (gimple_store_p (stmt)
1331 30325019 : && gimple_assign_single_p (stmt)
1332 60710155 : && integer_zerop (gimple_assign_rhs1 (stmt)))
1333 : {
1334 3586595 : tree rhs = gimple_assign_rhs1 (stmt);
1335 3586595 : tree type = TREE_TYPE (rhs);
1336 3586595 : dest = gimple_assign_lhs (stmt);
1337 3586595 : ao_ref_init (&read, dest);
1338 : /* For integral types, the type precision needs to be a multiply of BITS_PER_UNIT. */
1339 3586595 : if (INTEGRAL_TYPE_P (type)
1340 3586595 : && (TYPE_PRECISION (type) % BITS_PER_UNIT) != 0)
1341 : dest = NULL_TREE;
1342 : }
1343 26798541 : else if (gimple_store_p (stmt)
1344 26738424 : && gimple_assign_single_p (stmt)
1345 26738424 : && TREE_CODE (gimple_assign_rhs1 (stmt)) == CONSTRUCTOR
1346 27510861 : && !gimple_clobber_p (stmt))
1347 : {
1348 712320 : dest = gimple_assign_lhs (stmt);
1349 712320 : ao_ref_init (&read, dest);
1350 : }
1351 :
1352 4165307 : if (dest == NULL_TREE)
1353 26321889 : return;
1354 :
1355 4141518 : if (len == NULL_TREE)
1356 4089624 : len = (TREE_CODE (dest) == COMPONENT_REF
1357 4089624 : ? DECL_SIZE_UNIT (TREE_OPERAND (dest, 1))
1358 1775910 : : TYPE_SIZE_UNIT (TREE_TYPE (dest)));
1359 4089624 : if (len == NULL_TREE
1360 4141518 : || !poly_int_tree_p (len))
1361 : return;
1362 :
1363 : /* Sometimes memset can have no vdef due to invalid declaration of memset (const, etc.). */
1364 35616661 : if (!gimple_vdef (stmt))
1365 : return;
1366 :
1367 : /* This store needs to be on the byte boundary and pointing to an object. */
1368 4141494 : poly_int64 offset;
1369 4141494 : tree dest_base = get_addr_base_and_unit_offset (dest, &offset);
1370 4141494 : if (dest_base == NULL_TREE)
1371 : return;
1372 :
1373 : /* Setup the worklist. */
1374 4055924 : auto_vec<std::pair<tree, unsigned>> worklist;
1375 4055924 : unsigned limit = full_walk ? param_sccvn_max_alias_queries_per_access : 0;
1376 8111848 : worklist.safe_push (std::make_pair (gimple_vdef (stmt), limit));
1377 :
1378 28109463 : while (!worklist.is_empty ())
1379 : {
1380 19997615 : std::pair<tree, unsigned> top = worklist.pop ();
1381 19997615 : tree vdef = top.first;
1382 19997615 : limit = top.second;
1383 19997615 : gimple *use_stmt;
1384 19997615 : imm_use_iterator iter;
1385 64527863 : FOR_EACH_IMM_USE_STMT (use_stmt, iter, vdef)
1386 : {
1387 : /* Handling PHI nodes might not be worth it so don't. */
1388 24532633 : if (is_a <gphi*> (use_stmt))
1389 2021657 : continue;
1390 :
1391 : /* If this statement does not clobber add the vdef stmt to the
1392 : worklist.
1393 : After hitting the limit, allow clobbers to able to pass through. */
1394 2048969 : if ((limit != 0 || gimple_clobber_p (use_stmt))
1395 20497963 : && gimple_vdef (use_stmt)
1396 39896685 : && !stmt_may_clobber_ref_p_1 (use_stmt, &read,
1397 : /* tbaa_p = */ can_use_tbba))
1398 : {
1399 15941691 : unsigned new_limit = limit == 0 ? 0 : limit - 1;
1400 31883382 : worklist.safe_push (std::make_pair (gimple_vdef (use_stmt),
1401 : new_limit));
1402 : }
1403 :
1404 22510976 : optimize_aggr_zeroprop_1 (stmt, use_stmt, dest_base, offset,
1405 22510976 : val, wi::to_poly_offset (len));
1406 19997615 : }
1407 : }
1408 :
1409 4055924 : }
1410 :
1411 : /* Returns the pointer to the base of the object of the
1412 : reference EXPR and extracts the information about
1413 : the offset of the access, storing it to PBYTESIZE,
1414 : PBYTEPOS and PREVERSEP.
1415 : If the access is not a byte sized or position is not
1416 : on the byte, return NULL. */
1417 : static tree
1418 5314778 : split_core_and_offset_size (tree expr,
1419 : poly_int64 *pbytesize, poly_int64 *pbytepos,
1420 : tree *poffset, int *preversep)
1421 : {
1422 5314778 : tree core;
1423 5314778 : machine_mode mode;
1424 5314778 : int unsignedp, volatilep;
1425 5314778 : poly_int64 bitsize;
1426 5314778 : poly_int64 bitpos;
1427 5314778 : location_t loc = EXPR_LOCATION (expr);
1428 :
1429 5314778 : core = get_inner_reference (expr, &bitsize, &bitpos,
1430 : poffset, &mode, &unsignedp, preversep,
1431 : &volatilep);
1432 10629556 : if (!multiple_p (bitsize, BITS_PER_UNIT, pbytesize))
1433 : return NULL_TREE;
1434 5314778 : if (!multiple_p (bitpos, BITS_PER_UNIT, pbytepos))
1435 : return NULL_TREE;
1436 : /* If we are left with MEM[a + CST] strip that and add it to the
1437 : pbytepos and return a. */
1438 5314778 : if (TREE_CODE (core) == MEM_REF)
1439 : {
1440 1240593 : poly_offset_int tem;
1441 1240593 : tem = wi::to_poly_offset (TREE_OPERAND (core, 1));
1442 1240593 : tem += *pbytepos;
1443 1240593 : if (tem.to_shwi (pbytepos))
1444 1238707 : return TREE_OPERAND (core, 0);
1445 : }
1446 4076071 : core = build_fold_addr_expr_loc (loc, core);
1447 4076071 : STRIP_NOPS (core);
1448 4076071 : return core;
1449 : }
1450 :
1451 : /* Returns a new src based on the
1452 : copy `DEST = SRC` and for the old SRC2.
1453 : Returns null if SRC2 is not related to DEST. */
1454 :
1455 : static tree
1456 1263223 : new_src_based_on_copy (tree src2, tree dest, tree src)
1457 : {
1458 : /* If the second src is not exactly the same as dest,
1459 : try to handle it separately; see it is address/size equivalent.
1460 : Handles `a` and `a.b` and `MEM<char[N]>(&a)` which all have
1461 : the same size and offsets as address/size equivalent.
1462 : This allows copying over a memcpy and also one for copying
1463 : where one field is the same size as the whole struct. */
1464 1263223 : if (operand_equal_p (dest, src2))
1465 : return src;
1466 : /* if both dest and src2 are decls, then we know these 2
1467 : accesses can't be the same. */
1468 720216 : if (DECL_P (dest) && DECL_P (src2))
1469 : return NULL_TREE;
1470 : /* A VCE can't be used with imag/real or BFR so reject them early. */
1471 379514 : if (TREE_CODE (src) == IMAGPART_EXPR
1472 379514 : || TREE_CODE (src) == REALPART_EXPR
1473 379514 : || TREE_CODE (src) == BIT_FIELD_REF)
1474 : return NULL_TREE;
1475 379514 : tree core1, core2;
1476 379514 : poly_int64 bytepos1, bytepos2;
1477 379514 : poly_int64 bytesize1, bytesize2;
1478 379514 : tree toffset1, toffset2;
1479 379514 : int reversep1 = 0;
1480 379514 : int reversep2 = 0;
1481 379514 : poly_int64 diff = 0;
1482 379514 : core1 = split_core_and_offset_size (dest, &bytesize1, &bytepos1,
1483 : &toffset1, &reversep1);
1484 379514 : core2 = split_core_and_offset_size (src2, &bytesize2, &bytepos2,
1485 : &toffset2, &reversep2);
1486 379514 : if (!core1 || !core2)
1487 : return NULL_TREE;
1488 379514 : if (reversep1 != reversep2)
1489 : return NULL_TREE;
1490 : /* The sizes of the 2 accesses need to be the same. */
1491 379514 : if (!known_eq (bytesize1, bytesize2))
1492 : return NULL_TREE;
1493 170340 : if (!operand_equal_p (core1, core2, 0))
1494 : return NULL_TREE;
1495 :
1496 22757 : if (toffset1 && toffset2)
1497 : {
1498 2 : tree type = TREE_TYPE (toffset1);
1499 2 : if (type != TREE_TYPE (toffset2))
1500 0 : toffset2 = fold_convert (type, toffset2);
1501 :
1502 2 : tree tdiff = fold_build2 (MINUS_EXPR, type, toffset1, toffset2);
1503 2 : if (!cst_and_fits_in_hwi (tdiff))
1504 : return NULL_TREE;
1505 :
1506 0 : diff = int_cst_value (tdiff);
1507 0 : }
1508 22755 : else if (toffset1 || toffset2)
1509 : {
1510 : /* If only one of the offsets is non-constant, the difference cannot
1511 : be a constant. */
1512 : return NULL_TREE;
1513 : }
1514 22723 : diff += bytepos1 - bytepos2;
1515 : /* The offset between the 2 need to be 0. */
1516 22723 : if (!known_eq (diff, 0))
1517 : return NULL_TREE;
1518 21888 : return fold_build1 (VIEW_CONVERT_EXPR,TREE_TYPE (src2), src);
1519 : }
1520 :
1521 : /* Returns true if SRC and DEST are the same address such that
1522 : `SRC == DEST;` is considered a nop. This is more than an
1523 : operand_equal_p check as it needs to be similar to
1524 : new_src_based_on_copy. */
1525 :
1526 : static bool
1527 4500207 : same_for_assignment (tree src, tree dest)
1528 : {
1529 4500207 : if (operand_equal_p (dest, src, 0))
1530 : return true;
1531 : /* if both dest and src2 are decls, then we know these 2
1532 : accesses can't be the same. */
1533 4497300 : if (DECL_P (dest) && DECL_P (src))
1534 : return false;
1535 :
1536 2277875 : tree core1, core2;
1537 2277875 : poly_int64 bytepos1, bytepos2;
1538 2277875 : poly_int64 bytesize1, bytesize2;
1539 2277875 : tree toffset1, toffset2;
1540 2277875 : int reversep1 = 0;
1541 2277875 : int reversep2 = 0;
1542 2277875 : poly_int64 diff = 0;
1543 2277875 : core1 = split_core_and_offset_size (dest, &bytesize1, &bytepos1,
1544 : &toffset1, &reversep1);
1545 2277875 : core2 = split_core_and_offset_size (src, &bytesize2, &bytepos2,
1546 : &toffset2, &reversep2);
1547 2277875 : if (!core1 || !core2)
1548 : return false;
1549 2277875 : if (reversep1 != reversep2)
1550 : return false;
1551 : /* The sizes of the 2 accesses need to be the same. */
1552 2277875 : if (!known_eq (bytesize1, bytesize2))
1553 : return false;
1554 2276951 : if (!operand_equal_p (core1, core2, 0))
1555 : return false;
1556 6084 : if (toffset1 && toffset2)
1557 : {
1558 313 : tree type = TREE_TYPE (toffset1);
1559 313 : if (type != TREE_TYPE (toffset2))
1560 0 : toffset2 = fold_convert (type, toffset2);
1561 :
1562 313 : tree tdiff = fold_build2 (MINUS_EXPR, type, toffset1, toffset2);
1563 313 : if (!cst_and_fits_in_hwi (tdiff))
1564 : return false;
1565 :
1566 0 : diff = int_cst_value (tdiff);
1567 0 : }
1568 5771 : else if (toffset1 || toffset2)
1569 : {
1570 : /* If only one of the offsets is non-constant, the difference cannot
1571 : be a constant. */
1572 : return false;
1573 : }
1574 5771 : diff += bytepos1 - bytepos2;
1575 : /* The offset between the 2 need to be 0. */
1576 5771 : if (!known_eq (diff, 0))
1577 : return false;
1578 : return true;
1579 : }
1580 :
1581 : /* Helper function for optimize_agr_copyprop.
1582 : For aggregate copies in USE_STMT, see if DEST
1583 : is on the lhs of USE_STMT and replace it with SRC. */
1584 : static void
1585 1022581 : optimize_agr_copyprop_1 (gimple *stmt, gimple *use_stmt,
1586 : tree dest, tree src)
1587 : {
1588 1022581 : gcc_assert (gimple_assign_load_p (use_stmt)
1589 : && gimple_store_p (use_stmt));
1590 2045162 : if (gimple_has_volatile_ops (use_stmt))
1591 612517 : return;
1592 1022580 : tree dest2 = gimple_assign_lhs (use_stmt);
1593 1022580 : tree src2 = gimple_assign_rhs1 (use_stmt);
1594 : /* If the new store is `src2 = src2;` skip over it. */
1595 1022580 : if (same_for_assignment (src2, dest2))
1596 : return;
1597 1022017 : src = new_src_based_on_copy (src2, dest, src);
1598 1022017 : if (!src)
1599 : return;
1600 : /* For 2 memory references and using a temporary to do the copy,
1601 : don't remove the temporary as the 2 memory references might overlap.
1602 : Note t does not need to be decl as it could be field.
1603 : See PR 22237 for full details.
1604 : E.g.
1605 : t = *a; #DEST = SRC;
1606 : *b = t; #DEST2 = SRC2;
1607 : Cannot be convert into
1608 : t = *a;
1609 : *b = *a;
1610 : Though the following is allowed to be done:
1611 : t = *a;
1612 : *a = t;
1613 : And convert it into:
1614 : t = *a;
1615 : *a = *a;
1616 : */
1617 438430 : if (!operand_equal_p (dest2, src, 0)
1618 438430 : && !DECL_P (dest2) && !DECL_P (src))
1619 : {
1620 : /* If *a and *b have the same base see if
1621 : the offset between the two is greater than
1622 : or equal to the size of the type. */
1623 31698 : poly_int64 offset1, offset2;
1624 31698 : tree len = TYPE_SIZE_UNIT (TREE_TYPE (src));
1625 31698 : if (len == NULL_TREE
1626 31698 : || !tree_fits_poly_int64_p (len))
1627 28366 : return;
1628 31698 : tree base1 = get_addr_base_and_unit_offset (dest2, &offset1);
1629 31698 : tree base2 = get_addr_base_and_unit_offset (src, &offset2);
1630 31698 : poly_int64 size = tree_to_poly_int64 (len);
1631 : /* If the bases are 2 different decls,
1632 : then there can be no overlapping. */
1633 31698 : if (base1 && base2
1634 30831 : && DECL_P (base1) && DECL_P (base2)
1635 1874 : && base1 != base2)
1636 : ;
1637 : /* If we can't figure out the base or the bases are
1638 : not equal then fall back to an alignment check. */
1639 30050 : else if (!base1
1640 30050 : || !base2
1641 30050 : || !operand_equal_p (base1, base2))
1642 : {
1643 29681 : unsigned int align1 = get_object_alignment (src);
1644 29681 : unsigned int align2 = get_object_alignment (dest2);
1645 29681 : align1 /= BITS_PER_UNIT;
1646 29681 : align2 /= BITS_PER_UNIT;
1647 : /* If the alignment of either object is less
1648 : than the size then there is a possibility
1649 : of overlapping. */
1650 29681 : if (maybe_lt (align1, size)
1651 29681 : || maybe_lt (align2, size))
1652 28366 : return;
1653 : }
1654 : /* Make sure [offset1, offset1 + len - 1] does
1655 : not overlap with [offset2, offset2 + len - 1],
1656 : it is ok if they are at the same location though. */
1657 369 : else if (ranges_maybe_overlap_p (offset1, size, offset2, size)
1658 369 : && !known_eq (offset2, offset1))
1659 : return;
1660 : }
1661 :
1662 410064 : if (dump_file && (dump_flags & TDF_DETAILS))
1663 : {
1664 11 : fprintf (dump_file, "Simplified\n ");
1665 11 : print_gimple_stmt (dump_file, use_stmt, 0, dump_flags);
1666 11 : fprintf (dump_file, "after previous\n ");
1667 11 : print_gimple_stmt (dump_file, stmt, 0, dump_flags);
1668 : }
1669 410064 : gimple *orig_stmt = use_stmt;
1670 410064 : gimple_stmt_iterator gsi = gsi_for_stmt (use_stmt);
1671 410064 : gimple_assign_set_rhs_from_tree (&gsi, unshare_expr (src));
1672 410064 : update_stmt (use_stmt);
1673 :
1674 410064 : if (dump_file && (dump_flags & TDF_DETAILS))
1675 : {
1676 11 : fprintf (dump_file, "into\n ");
1677 11 : print_gimple_stmt (dump_file, use_stmt, 0, dump_flags);
1678 : }
1679 410064 : if (maybe_clean_or_replace_eh_stmt (orig_stmt, use_stmt))
1680 0 : bitmap_set_bit (to_purge, gimple_bb (stmt)->index);
1681 410064 : statistics_counter_event (cfun, "copy prop for aggregate", 1);
1682 : }
1683 :
1684 : /* Helper function for optimize_agr_copyprop_1, propagate aggregates
1685 : into the arguments of USE_STMT if the argument matches with DEST;
1686 : replacing it with SRC. */
1687 : static void
1688 710397 : optimize_agr_copyprop_arg (gimple *defstmt, gcall *call,
1689 : tree dest, tree src)
1690 : {
1691 710397 : bool changed = false;
1692 2376921 : for (unsigned arg = 0; arg < gimple_call_num_args (call); arg++)
1693 : {
1694 1666524 : tree *argptr = gimple_call_arg_ptr (call, arg);
1695 3144061 : if (TREE_CODE (*argptr) == SSA_NAME
1696 947501 : || is_gimple_min_invariant (*argptr)
1697 1855511 : || TYPE_VOLATILE (TREE_TYPE (*argptr)))
1698 1477537 : continue;
1699 188987 : tree newsrc = new_src_based_on_copy (*argptr, dest, src);
1700 188987 : if (!newsrc)
1701 114647 : continue;
1702 :
1703 74340 : if (dump_file && (dump_flags & TDF_DETAILS))
1704 : {
1705 9 : fprintf (dump_file, "Simplified\n ");
1706 9 : print_gimple_stmt (dump_file, call, 0, dump_flags);
1707 9 : fprintf (dump_file, "after previous\n ");
1708 9 : print_gimple_stmt (dump_file, defstmt, 0, dump_flags);
1709 : }
1710 74340 : *argptr = unshare_expr (newsrc);
1711 74340 : changed = true;
1712 74340 : if (dump_file && (dump_flags & TDF_DETAILS))
1713 : {
1714 9 : fprintf (dump_file, "into\n ");
1715 9 : print_gimple_stmt (dump_file, call, 0, dump_flags);
1716 : }
1717 : }
1718 710397 : if (changed)
1719 74164 : update_stmt (call);
1720 710397 : }
1721 :
1722 : /* Helper function for optimize_agr_copyprop, propagate aggregates
1723 : into the return stmt USE if the operand of the return matches DEST;
1724 : replacing it with SRC. */
1725 : static void
1726 126957 : optimize_agr_copyprop_return (gimple *defstmt, greturn *use,
1727 : tree dest, tree src)
1728 : {
1729 126957 : tree rvalue = gimple_return_retval (use);
1730 126957 : if (!rvalue
1731 82892 : || TREE_CODE (rvalue) == SSA_NAME
1732 74115 : || is_gimple_min_invariant (rvalue)
1733 200673 : || TYPE_VOLATILE (TREE_TYPE (rvalue)))
1734 53242 : return;
1735 :
1736 : /* `return <retval>;` is already the best it could be.
1737 : Likewise `return *<retval>_N(D)`. */
1738 73715 : if (TREE_CODE (rvalue) == RESULT_DECL
1739 73715 : || (TREE_CODE (rvalue) == MEM_REF
1740 0 : && TREE_CODE (TREE_OPERAND (rvalue, 0)) == SSA_NAME
1741 0 : && TREE_CODE (SSA_NAME_VAR (TREE_OPERAND (rvalue, 0)))
1742 : == RESULT_DECL))
1743 : return;
1744 52219 : tree newsrc = new_src_based_on_copy (rvalue, dest, src);
1745 52219 : if (!newsrc)
1746 : return;
1747 : /* Currently only support non-global vars.
1748 : See PR 124099 on enumtls not supporting expanding for GIMPLE_RETURN.
1749 : FIXME: could support VCEs too? */
1750 52125 : if (!VAR_P (newsrc) || is_global_var (newsrc))
1751 : return;
1752 25989 : if (dump_file && (dump_flags & TDF_DETAILS))
1753 : {
1754 1 : fprintf (dump_file, "Simplified\n ");
1755 1 : print_gimple_stmt (dump_file, use, 0, dump_flags);
1756 1 : fprintf (dump_file, "after previous\n ");
1757 1 : print_gimple_stmt (dump_file, defstmt, 0, dump_flags);
1758 : }
1759 25989 : gimple_return_set_retval (use, newsrc);
1760 25989 : if (dump_file && (dump_flags & TDF_DETAILS))
1761 : {
1762 1 : fprintf (dump_file, "into\n ");
1763 1 : print_gimple_stmt (dump_file, use, 0, dump_flags);
1764 : }
1765 25989 : update_stmt (use);
1766 : }
1767 :
1768 : /* Optimizes
1769 : DEST = SRC;
1770 : DEST2 = DEST; # DEST2 = SRC2;
1771 : into
1772 : DEST = SRC;
1773 : DEST2 = SRC;
1774 : STMT is the first statement and SRC is the common
1775 : between the statements.
1776 :
1777 : Also optimizes:
1778 : DEST = SRC;
1779 : call_func(..., DEST, ...);
1780 : into:
1781 : DEST = SRC;
1782 : call_func(..., SRC, ...);
1783 :
1784 : */
1785 : static void
1786 3886465 : optimize_agr_copyprop (gimple *stmt)
1787 : {
1788 7772930 : if (gimple_has_volatile_ops (stmt))
1789 411428 : return;
1790 :
1791 : /* Can't prop if the statement could throw. */
1792 3885254 : if (stmt_could_throw_p (cfun, stmt))
1793 : return;
1794 :
1795 3477627 : tree dest = gimple_assign_lhs (stmt);
1796 3477627 : tree src = gimple_assign_rhs1 (stmt);
1797 : /* If the statement is `src = src;` then ignore it. */
1798 3477627 : if (same_for_assignment (dest, src))
1799 : return;
1800 :
1801 3475037 : tree vdef = gimple_vdef (stmt);
1802 3475037 : imm_use_iterator iter;
1803 3475037 : gimple *use_stmt;
1804 13503724 : FOR_EACH_IMM_USE_STMT (use_stmt, iter, vdef)
1805 : {
1806 6553650 : if (gimple_assign_load_p (use_stmt)
1807 6553650 : && gimple_store_p (use_stmt))
1808 1022581 : optimize_agr_copyprop_1 (stmt, use_stmt, dest, src);
1809 5531069 : else if (is_gimple_call (use_stmt))
1810 710397 : optimize_agr_copyprop_arg (stmt, as_a<gcall*>(use_stmt), dest, src);
1811 4820672 : else if (is_a<greturn*> (use_stmt))
1812 126957 : optimize_agr_copyprop_return (stmt, as_a<greturn*>(use_stmt), dest, src);
1813 3475037 : }
1814 : }
1815 :
1816 : /* Simple DSE of the lhs from a clobber STMT.
1817 : This is used mostly to clean up from optimize_agr_copyprop and
1818 : to remove (exactly one) extra copy that might later on confuse SRA.
1819 : An example is:
1820 : ;; write to a and such.
1821 : b = a; // This statement is to be removed
1822 : b = {CLOBBER};
1823 : SRA will totally scalarize b (which means also a) here for the extra copy
1824 : which is not something welcomed. So removing the copy will
1825 : allow SRA to move the scalarization of a further down or not at all.
1826 : */
1827 : static void
1828 7400238 : do_simple_agr_dse (gassign *stmt, bool full_walk)
1829 : {
1830 : /* Don't do this while in -Og as we want to keep around the copy
1831 : for debuggability. */
1832 7400238 : if (optimize_debug)
1833 5113167 : return;
1834 7396845 : ao_ref read;
1835 7396845 : basic_block bb = gimple_bb (stmt);
1836 7396845 : tree lhs = gimple_assign_lhs (stmt);
1837 : /* Only handle clobbers of a full decl. */
1838 7396845 : if (!DECL_P (lhs))
1839 : return;
1840 6673121 : ao_ref_init (&read, lhs);
1841 6673121 : tree vuse = gimple_vuse (stmt);
1842 6673121 : unsigned limit = full_walk ? param_sccvn_max_alias_queries_per_access : 4;
1843 16966896 : while (limit)
1844 : {
1845 16953729 : gimple *ostmt = SSA_NAME_DEF_STMT (vuse);
1846 : /* Don't handle phis, just declare to be done. */
1847 16953729 : if (is_a<gphi*>(ostmt) || gimple_nop_p (ostmt))
1848 : break;
1849 14679825 : basic_block obb = gimple_bb (ostmt);
1850 : /* If the clobber is not fully dominating the statement define,
1851 : then it is not "simple" to detect if the define is fully clobbered. */
1852 14679825 : if (obb != bb && !dominated_by_p (CDI_DOMINATORS, bb, obb))
1853 4386050 : return;
1854 14679825 : gimple *use_stmt;
1855 14679825 : imm_use_iterator iter;
1856 59387200 : FOR_EACH_IMM_USE_STMT (use_stmt, iter, gimple_vdef (ostmt))
1857 : {
1858 17310832 : basic_block ubb = gimple_bb (use_stmt);
1859 17310832 : if (stmt == use_stmt)
1860 5117919 : continue;
1861 : /* If the use is a clobber for lhs,
1862 : then it can be safely skipped; this happens with eh
1863 : and sometimes jump threading. */
1864 12192913 : if (gimple_clobber_p (use_stmt)
1865 12192913 : && lhs == gimple_assign_lhs (use_stmt))
1866 172961 : continue;
1867 : /* If the use is a phi and it is single use then check if that single use
1868 : is a clobber and lhs is the same. */
1869 12019952 : if (gphi *use_phi = dyn_cast<gphi*>(use_stmt))
1870 : {
1871 360135 : use_operand_p ou;
1872 360135 : gimple *ostmt;
1873 360135 : if (single_imm_use (gimple_phi_result (use_phi), &ou, &ostmt)
1874 300968 : && gimple_clobber_p (ostmt)
1875 601651 : && lhs == gimple_assign_lhs (ostmt))
1876 67865 : continue;
1877 : /* A phi node will never be dominating the clobber. */
1878 292270 : return;
1879 : }
1880 : /* The use needs to be dominating the clobber. */
1881 1472726 : if ((ubb != bb && !dominated_by_p (CDI_DOMINATORS, bb, ubb))
1882 12402620 : || ref_maybe_used_by_stmt_p (use_stmt, &read, false))
1883 1215064 : return;
1884 : /* Count the above alias lookup towards the limit. */
1885 10444753 : limit--;
1886 10444753 : if (limit == 0)
1887 : return;
1888 1963107 : }
1889 12716718 : vuse = gimple_vuse (ostmt);
1890 : /* This is a call with an assignment to the clobber decl,
1891 : remove the lhs or the whole stmt if it was pure/const. */
1892 12716718 : if (is_a <gcall*>(ostmt)
1893 12716718 : && lhs == gimple_call_lhs (ostmt))
1894 : {
1895 : /* Don't remove stores/statements that are needed for non-call
1896 : eh to work. */
1897 4029 : if (stmt_unremovable_because_of_non_call_eh_p (cfun, ostmt))
1898 : return;
1899 : /* If we delete a stmt that could throw, mark the block
1900 : in to_purge to cleanup afterwards. */
1901 4023 : if (stmt_could_throw_p (cfun, ostmt))
1902 1006 : bitmap_set_bit (to_purge, obb->index);
1903 4023 : int flags = gimple_call_flags (ostmt);
1904 4023 : if ((flags & (ECF_PURE|ECF_CONST|ECF_NOVOPS))
1905 203 : && !(flags & (ECF_LOOPING_CONST_OR_PURE)))
1906 : {
1907 119 : gimple_stmt_iterator gsi = gsi_for_stmt (ostmt);
1908 119 : if (dump_file && (dump_flags & TDF_DETAILS))
1909 : {
1910 14 : fprintf (dump_file, "Removing dead call store stmt ");
1911 14 : print_gimple_stmt (dump_file, ostmt, 0);
1912 14 : fprintf (dump_file, "\n");
1913 : }
1914 119 : unlink_stmt_vdef (ostmt);
1915 119 : release_defs (ostmt);
1916 119 : gsi_remove (&gsi, true);
1917 119 : statistics_counter_event (cfun, "delete call dead store", 1);
1918 : /* Only remove the first store previous statement. */
1919 119 : return;
1920 : }
1921 : /* Make sure we do not remove a return slot we cannot reconstruct
1922 : later. */
1923 3904 : if (gimple_call_return_slot_opt_p (as_a <gcall *>(ostmt))
1924 3904 : && (TREE_ADDRESSABLE (TREE_TYPE (gimple_call_fntype (ostmt)))
1925 541 : || !poly_int_tree_p
1926 541 : (TYPE_SIZE (TREE_TYPE (gimple_call_fntype (ostmt))))))
1927 : return;
1928 679 : if (dump_file && (dump_flags & TDF_DETAILS))
1929 : {
1930 6 : fprintf (dump_file, "Removing lhs of call stmt ");
1931 6 : print_gimple_stmt (dump_file, ostmt, 0);
1932 6 : fprintf (dump_file, "\n");
1933 : }
1934 679 : gimple_call_set_lhs (ostmt, NULL_TREE);
1935 679 : update_stmt (ostmt);
1936 679 : statistics_counter_event (cfun, "removed lhs call", 1);
1937 679 : return;
1938 : }
1939 : /* This an assignment store to the clobbered decl,
1940 : then maybe remove it. */
1941 12712689 : if (is_a <gassign*>(ostmt)
1942 10733456 : && gimple_store_p (ostmt)
1943 10733456 : && !gimple_clobber_p (ostmt)
1944 15903827 : && lhs == gimple_assign_lhs (ostmt))
1945 : {
1946 : /* Don't remove stores/statements that are needed for non-call
1947 : eh to work. */
1948 167004 : if (stmt_unremovable_because_of_non_call_eh_p (cfun, ostmt))
1949 : return;
1950 : /* If we delete a stmt that could throw, mark the block
1951 : in to_purge to cleanup afterwards. */
1952 161920 : if (stmt_could_throw_p (cfun, ostmt))
1953 0 : bitmap_set_bit (to_purge, obb->index);
1954 161920 : gimple_stmt_iterator gsi = gsi_for_stmt (ostmt);
1955 161920 : if (dump_file && (dump_flags & TDF_DETAILS))
1956 : {
1957 12 : fprintf (dump_file, "Removing dead store stmt ");
1958 12 : print_gimple_stmt (dump_file, ostmt, 0);
1959 12 : fprintf (dump_file, "\n");
1960 : }
1961 161920 : unlink_stmt_vdef (ostmt);
1962 161920 : release_defs (ostmt);
1963 161920 : gsi_remove (&gsi, true);
1964 161920 : statistics_counter_event (cfun, "delete dead store", 1);
1965 : /* Only remove the first store previous statement. */
1966 161920 : return;
1967 : }
1968 : /* If the statement uses or maybe writes to the decl,
1969 : then nothing is to be removed. Don't know if the write
1970 : to the decl is partial write or a full one so the need
1971 : to stop.
1972 : e.g.
1973 : b.c = a;
1974 : Easier to stop here rather than do a full partial
1975 : dse of this statement.
1976 : b = {CLOBBER}; */
1977 12545685 : if (stmt_may_clobber_ref_p_1 (ostmt, &read, false)
1978 12545685 : || ref_maybe_used_by_stmt_p (ostmt, &read, false))
1979 2251910 : return;
1980 10293775 : limit--;
1981 : }
1982 : }
1983 :
1984 : /* Optimizes builtin memcmps for small constant sizes.
1985 : GSI_P is the GSI for the call. STMT is the call itself.
1986 : */
1987 :
1988 : static bool
1989 470836 : simplify_builtin_memcmp (gimple_stmt_iterator *gsi_p, gcall *stmt)
1990 : {
1991 : /* Make sure memcmp arguments are the correct type. */
1992 470836 : if (gimple_call_num_args (stmt) != 3)
1993 : return false;
1994 470836 : tree arg1 = gimple_call_arg (stmt, 0);
1995 470836 : tree arg2 = gimple_call_arg (stmt, 1);
1996 470836 : tree len = gimple_call_arg (stmt, 2);
1997 :
1998 470836 : if (!POINTER_TYPE_P (TREE_TYPE (arg1)))
1999 : return false;
2000 470836 : if (!POINTER_TYPE_P (TREE_TYPE (arg2)))
2001 : return false;
2002 470836 : if (!INTEGRAL_TYPE_P (TREE_TYPE (len)))
2003 : return false;
2004 :
2005 : /* The return value of the memcmp has to be used
2006 : equality comparison to zero. */
2007 470836 : tree res = gimple_call_lhs (stmt);
2008 :
2009 470836 : if (!res || !use_in_zero_equality (res))
2010 15059 : return false;
2011 :
2012 455777 : unsigned HOST_WIDE_INT leni;
2013 :
2014 455777 : if (tree_fits_uhwi_p (len)
2015 634244 : && (leni = tree_to_uhwi (len)) <= GET_MODE_SIZE (word_mode)
2016 536912 : && pow2p_hwi (leni))
2017 : {
2018 19183 : leni *= CHAR_TYPE_SIZE;
2019 19183 : unsigned align1 = get_pointer_alignment (arg1);
2020 19183 : unsigned align2 = get_pointer_alignment (arg2);
2021 19183 : unsigned align = MIN (align1, align2);
2022 19183 : scalar_int_mode mode;
2023 19183 : if (int_mode_for_size (leni, 1).exists (&mode)
2024 19183 : && (align >= leni || !targetm.slow_unaligned_access (mode, align)))
2025 : {
2026 19183 : location_t loc = gimple_location (stmt);
2027 19183 : tree type, off;
2028 19183 : type = build_nonstandard_integer_type (leni, 1);
2029 38366 : gcc_assert (known_eq (GET_MODE_BITSIZE (TYPE_MODE (type)), leni));
2030 19183 : tree ptrtype = build_pointer_type_for_mode (char_type_node,
2031 : ptr_mode, true);
2032 19183 : off = build_int_cst (ptrtype, 0);
2033 :
2034 : /* Create unaligned types if needed. */
2035 19183 : tree type1 = type, type2 = type;
2036 19183 : if (TYPE_ALIGN (type1) > align1)
2037 7861 : type1 = build_aligned_type (type1, align1);
2038 19183 : if (TYPE_ALIGN (type2) > align2)
2039 8356 : type2 = build_aligned_type (type2, align2);
2040 :
2041 19183 : arg1 = build2_loc (loc, MEM_REF, type1, arg1, off);
2042 19183 : arg2 = build2_loc (loc, MEM_REF, type2, arg2, off);
2043 19183 : tree tem1 = fold_const_aggregate_ref (arg1);
2044 19183 : if (tem1)
2045 222 : arg1 = tem1;
2046 19183 : tree tem2 = fold_const_aggregate_ref (arg2);
2047 19183 : if (tem2)
2048 7529 : arg2 = tem2;
2049 19183 : res = fold_convert_loc (loc, TREE_TYPE (res),
2050 : fold_build2_loc (loc, NE_EXPR,
2051 : boolean_type_node,
2052 : arg1, arg2));
2053 19183 : gimplify_and_update_call_from_tree (gsi_p, res);
2054 19183 : return true;
2055 : }
2056 : }
2057 :
2058 : /* Replace memcmp with memcmp_eq if the above fails. */
2059 436594 : if (DECL_FUNCTION_CODE (gimple_call_fndecl (stmt)) == BUILT_IN_MEMCMP_EQ)
2060 : return false;
2061 345679 : if (!fold_before_rtl_expansion_p ())
2062 : return false;
2063 90915 : gimple_call_set_fndecl (stmt, builtin_decl_explicit (BUILT_IN_MEMCMP_EQ));
2064 90915 : update_stmt (stmt);
2065 90915 : return true;
2066 : }
2067 :
2068 : /* Optimizes builtin memchrs for small constant sizes with a const string.
2069 : GSI_P is the GSI for the call. STMT is the call itself.
2070 : */
2071 :
2072 : static bool
2073 14509 : simplify_builtin_memchr (gimple_stmt_iterator *gsi_p, gcall *stmt)
2074 : {
2075 14509 : if (CHAR_BIT != 8 || BITS_PER_UNIT != 8)
2076 : return false;
2077 :
2078 14509 : if (gimple_call_num_args (stmt) != 3)
2079 : return false;
2080 :
2081 14509 : tree res = gimple_call_lhs (stmt);
2082 14509 : if (!res || !use_in_zero_equality (res))
2083 13050 : return false;
2084 :
2085 1459 : tree ptr = gimple_call_arg (stmt, 0);
2086 1459 : if (TREE_CODE (ptr) != ADDR_EXPR
2087 1459 : || TREE_CODE (TREE_OPERAND (ptr, 0)) != STRING_CST)
2088 : return false;
2089 :
2090 452 : unsigned HOST_WIDE_INT slen
2091 452 : = TREE_STRING_LENGTH (TREE_OPERAND (ptr, 0));
2092 : /* It must be a non-empty string constant. */
2093 452 : if (slen < 2)
2094 : return false;
2095 :
2096 : /* For -Os, only simplify strings with a single character. */
2097 448 : if (!optimize_bb_for_speed_p (gimple_bb (stmt))
2098 448 : && slen > 2)
2099 : return false;
2100 :
2101 432 : tree size = gimple_call_arg (stmt, 2);
2102 : /* Size must be a constant which is <= UNITS_PER_WORD and
2103 : <= the string length. */
2104 432 : if (!tree_fits_uhwi_p (size))
2105 : return false;
2106 :
2107 432 : unsigned HOST_WIDE_INT sz = tree_to_uhwi (size);
2108 433 : if (sz == 0 || sz > UNITS_PER_WORD || sz >= slen)
2109 : return false;
2110 :
2111 380 : tree ch = gimple_call_arg (stmt, 1);
2112 380 : location_t loc = gimple_location (stmt);
2113 380 : if (!useless_type_conversion_p (char_type_node,
2114 380 : TREE_TYPE (ch)))
2115 380 : ch = fold_convert_loc (loc, char_type_node, ch);
2116 380 : const char *p = TREE_STRING_POINTER (TREE_OPERAND (ptr, 0));
2117 380 : unsigned int isize = sz;
2118 380 : tree *op = XALLOCAVEC (tree, isize);
2119 1369 : for (unsigned int i = 0; i < isize; i++)
2120 : {
2121 989 : op[i] = build_int_cst (char_type_node, p[i]);
2122 989 : op[i] = fold_build2_loc (loc, EQ_EXPR, boolean_type_node,
2123 : op[i], ch);
2124 : }
2125 989 : for (unsigned int i = isize - 1; i >= 1; i--)
2126 609 : op[i - 1] = fold_convert_loc (loc, boolean_type_node,
2127 : fold_build2_loc (loc,
2128 : BIT_IOR_EXPR,
2129 : boolean_type_node,
2130 609 : op[i - 1],
2131 609 : op[i]));
2132 380 : res = fold_convert_loc (loc, TREE_TYPE (res), op[0]);
2133 380 : gimplify_and_update_call_from_tree (gsi_p, res);
2134 380 : return true;
2135 : }
2136 :
2137 : /* *GSI_P is a GIMPLE_CALL to a builtin function.
2138 : Optimize
2139 : memcpy (p, "abcd", 4); // STMT1
2140 : memset (p + 4, ' ', 3); // STMT2
2141 : into
2142 : memcpy (p, "abcd ", 7);
2143 : call if the latter can be stored by pieces during expansion.
2144 : */
2145 :
2146 : static bool
2147 112011 : simplify_builtin_memcpy_memset (gimple_stmt_iterator *gsi_p, gcall *stmt2)
2148 : {
2149 112011 : if (gimple_call_num_args (stmt2) != 3
2150 112011 : || gimple_call_lhs (stmt2)
2151 : || CHAR_BIT != 8
2152 112011 : || BITS_PER_UNIT != 8)
2153 : return false;
2154 :
2155 215151 : tree vuse = gimple_vuse (stmt2);
2156 105539 : if (vuse == NULL)
2157 : return false;
2158 105523 : gimple *stmt1 = SSA_NAME_DEF_STMT (vuse);
2159 :
2160 105523 : tree callee1;
2161 105523 : tree ptr1, src1, str1, off1, len1, lhs1;
2162 105523 : tree ptr2 = gimple_call_arg (stmt2, 0);
2163 105523 : tree val2 = gimple_call_arg (stmt2, 1);
2164 105523 : tree len2 = gimple_call_arg (stmt2, 2);
2165 105523 : tree diff, vdef, new_str_cst;
2166 105523 : gimple *use_stmt;
2167 105523 : unsigned int ptr1_align;
2168 105523 : unsigned HOST_WIDE_INT src_len;
2169 105523 : char *src_buf;
2170 105523 : use_operand_p use_p;
2171 :
2172 105523 : if (!tree_fits_shwi_p (val2)
2173 101054 : || !tree_fits_uhwi_p (len2)
2174 169974 : || compare_tree_int (len2, 1024) == 1)
2175 46629 : return false;
2176 :
2177 58894 : if (is_gimple_call (stmt1))
2178 : {
2179 : /* If first stmt is a call, it needs to be memcpy
2180 : or mempcpy, with string literal as second argument and
2181 : constant length. */
2182 30491 : callee1 = gimple_call_fndecl (stmt1);
2183 30491 : if (callee1 == NULL_TREE
2184 30375 : || !fndecl_built_in_p (callee1, BUILT_IN_NORMAL)
2185 57196 : || gimple_call_num_args (stmt1) != 3)
2186 : return false;
2187 25413 : if (DECL_FUNCTION_CODE (callee1) != BUILT_IN_MEMCPY
2188 25413 : && DECL_FUNCTION_CODE (callee1) != BUILT_IN_MEMPCPY)
2189 : return false;
2190 10985 : ptr1 = gimple_call_arg (stmt1, 0);
2191 10985 : src1 = gimple_call_arg (stmt1, 1);
2192 10985 : len1 = gimple_call_arg (stmt1, 2);
2193 10985 : lhs1 = gimple_call_lhs (stmt1);
2194 10985 : if (!tree_fits_uhwi_p (len1))
2195 : return false;
2196 10898 : str1 = string_constant (src1, &off1, NULL, NULL);
2197 10898 : if (str1 == NULL_TREE)
2198 : return false;
2199 5001 : if (!tree_fits_uhwi_p (off1)
2200 5001 : || compare_tree_int (off1, TREE_STRING_LENGTH (str1) - 1) > 0
2201 5001 : || compare_tree_int (len1, TREE_STRING_LENGTH (str1)
2202 5001 : - tree_to_uhwi (off1)) > 0
2203 5001 : || TREE_CODE (TREE_TYPE (str1)) != ARRAY_TYPE
2204 15003 : || TYPE_MODE (TREE_TYPE (TREE_TYPE (str1)))
2205 5001 : != TYPE_MODE (char_type_node))
2206 0 : return false;
2207 : }
2208 28403 : else if (gimple_assign_single_p (stmt1))
2209 : {
2210 : /* Otherwise look for length 1 memcpy optimized into
2211 : assignment. */
2212 17382 : ptr1 = gimple_assign_lhs (stmt1);
2213 17382 : src1 = gimple_assign_rhs1 (stmt1);
2214 17382 : if (TREE_CODE (ptr1) != MEM_REF
2215 3550 : || TYPE_MODE (TREE_TYPE (ptr1)) != TYPE_MODE (char_type_node)
2216 18416 : || !tree_fits_shwi_p (src1))
2217 17030 : return false;
2218 352 : ptr1 = build_fold_addr_expr (ptr1);
2219 352 : STRIP_USELESS_TYPE_CONVERSION (ptr1);
2220 352 : callee1 = NULL_TREE;
2221 352 : len1 = size_one_node;
2222 352 : lhs1 = NULL_TREE;
2223 352 : off1 = size_zero_node;
2224 352 : str1 = NULL_TREE;
2225 : }
2226 : else
2227 : return false;
2228 :
2229 5353 : diff = constant_pointer_difference (ptr1, ptr2);
2230 5353 : if (diff == NULL && lhs1 != NULL)
2231 : {
2232 7 : diff = constant_pointer_difference (lhs1, ptr2);
2233 7 : if (DECL_FUNCTION_CODE (callee1) == BUILT_IN_MEMPCPY
2234 7 : && diff != NULL)
2235 7 : diff = size_binop (PLUS_EXPR, diff,
2236 : fold_convert (sizetype, len1));
2237 : }
2238 : /* If the difference between the second and first destination pointer
2239 : is not constant, or is bigger than memcpy length, bail out. */
2240 5353 : if (diff == NULL
2241 4467 : || !tree_fits_uhwi_p (diff)
2242 4467 : || tree_int_cst_lt (len1, diff)
2243 9564 : || compare_tree_int (diff, 1024) == 1)
2244 1142 : return false;
2245 :
2246 : /* Use maximum of difference plus memset length and memcpy length
2247 : as the new memcpy length, if it is too big, bail out. */
2248 4211 : src_len = tree_to_uhwi (diff);
2249 4211 : src_len += tree_to_uhwi (len2);
2250 4211 : if (src_len < tree_to_uhwi (len1))
2251 : src_len = tree_to_uhwi (len1);
2252 4211 : if (src_len > 1024)
2253 : return false;
2254 :
2255 : /* If mempcpy value is used elsewhere, bail out, as mempcpy
2256 : with bigger length will return different result. */
2257 4211 : if (lhs1 != NULL_TREE
2258 64 : && DECL_FUNCTION_CODE (callee1) == BUILT_IN_MEMPCPY
2259 4218 : && (TREE_CODE (lhs1) != SSA_NAME
2260 7 : || !single_imm_use (lhs1, &use_p, &use_stmt)
2261 7 : || use_stmt != stmt2))
2262 0 : return false;
2263 :
2264 : /* If anything reads memory in between memcpy and memset
2265 : call, the modified memcpy call might change it. */
2266 4211 : vdef = gimple_vdef (stmt1);
2267 4211 : if (vdef != NULL
2268 4211 : && (!single_imm_use (vdef, &use_p, &use_stmt)
2269 3456 : || use_stmt != stmt2))
2270 : return false;
2271 :
2272 3456 : ptr1_align = get_pointer_alignment (ptr1);
2273 : /* Construct the new source string literal. */
2274 3456 : src_buf = XALLOCAVEC (char, src_len + 1);
2275 3456 : if (callee1)
2276 3295 : memcpy (src_buf,
2277 3295 : TREE_STRING_POINTER (str1) + tree_to_uhwi (off1),
2278 : tree_to_uhwi (len1));
2279 : else
2280 161 : src_buf[0] = tree_to_shwi (src1);
2281 3456 : memset (src_buf + tree_to_uhwi (diff),
2282 3456 : tree_to_shwi (val2), tree_to_uhwi (len2));
2283 3456 : src_buf[src_len] = '\0';
2284 : /* Neither builtin_strncpy_read_str nor builtin_memcpy_read_str
2285 : handle embedded '\0's. */
2286 3456 : if (strlen (src_buf) != src_len)
2287 : return false;
2288 3366 : rtl_profile_for_bb (gimple_bb (stmt2));
2289 : /* If the new memcpy wouldn't be emitted by storing the literal
2290 : by pieces, this optimization might enlarge .rodata too much,
2291 : as commonly used string literals couldn't be shared any
2292 : longer. */
2293 3366 : if (!can_store_by_pieces (src_len,
2294 : builtin_strncpy_read_str,
2295 : src_buf, ptr1_align, false))
2296 : return false;
2297 :
2298 2526 : new_str_cst = build_string_literal (src_len, src_buf);
2299 2526 : if (callee1)
2300 : {
2301 : /* If STMT1 is a mem{,p}cpy call, adjust it and remove
2302 : memset call. */
2303 2399 : if (lhs1 && DECL_FUNCTION_CODE (callee1) == BUILT_IN_MEMPCPY)
2304 7 : gimple_call_set_lhs (stmt1, NULL_TREE);
2305 2399 : gimple_call_set_arg (stmt1, 1, new_str_cst);
2306 2399 : gimple_call_set_arg (stmt1, 2,
2307 2399 : build_int_cst (TREE_TYPE (len1), src_len));
2308 2399 : update_stmt (stmt1);
2309 2399 : unlink_stmt_vdef (stmt2);
2310 2399 : gsi_replace (gsi_p, gimple_build_nop (), false);
2311 2399 : fwprop_invalidate_lattice (gimple_get_lhs (stmt2));
2312 2399 : release_defs (stmt2);
2313 2399 : if (lhs1 && DECL_FUNCTION_CODE (callee1) == BUILT_IN_MEMPCPY)
2314 : {
2315 7 : fwprop_invalidate_lattice (lhs1);
2316 7 : release_ssa_name (lhs1);
2317 : }
2318 2399 : return true;
2319 : }
2320 : else
2321 : {
2322 : /* Otherwise, if STMT1 is length 1 memcpy optimized into
2323 : assignment, remove STMT1 and change memset call into
2324 : memcpy call. */
2325 127 : gimple_stmt_iterator gsi = gsi_for_stmt (stmt1);
2326 :
2327 127 : if (!is_gimple_val (ptr1))
2328 12 : ptr1 = force_gimple_operand_gsi (gsi_p, ptr1, true, NULL_TREE,
2329 : true, GSI_SAME_STMT);
2330 127 : tree fndecl = builtin_decl_explicit (BUILT_IN_MEMCPY);
2331 127 : gimple_call_set_fndecl (stmt2, fndecl);
2332 127 : gimple_call_set_fntype (stmt2,
2333 127 : TREE_TYPE (fndecl));
2334 127 : gimple_call_set_arg (stmt2, 0, ptr1);
2335 127 : gimple_call_set_arg (stmt2, 1, new_str_cst);
2336 127 : gimple_call_set_arg (stmt2, 2,
2337 127 : build_int_cst (TREE_TYPE (len2), src_len));
2338 127 : unlink_stmt_vdef (stmt1);
2339 127 : gsi_remove (&gsi, true);
2340 127 : fwprop_invalidate_lattice (gimple_get_lhs (stmt1));
2341 127 : release_defs (stmt1);
2342 127 : update_stmt (stmt2);
2343 127 : return false;
2344 : }
2345 : }
2346 :
2347 :
2348 : /* Try to optimize out __builtin_stack_restore. Optimize it out
2349 : if there is another __builtin_stack_restore in the same basic
2350 : block and no calls or ASM_EXPRs are in between, or if this block's
2351 : only outgoing edge is to EXIT_BLOCK and there are no calls or
2352 : ASM_EXPRs after this __builtin_stack_restore.
2353 : Note restore right before a noreturn function is not needed.
2354 : And skip some cheap calls that will most likely become an instruction.
2355 : Restoring the stack before a call is important to be able to keep
2356 : stack usage down so that call does not run out of stack. */
2357 :
2358 :
2359 : static bool
2360 10369 : optimize_stack_restore (gimple_stmt_iterator *gsi, gimple *call)
2361 : {
2362 10369 : if (!fold_before_rtl_expansion_p ())
2363 : return false;
2364 2525 : tree callee;
2365 2525 : gimple *stmt;
2366 :
2367 2525 : basic_block bb = gsi_bb (*gsi);
2368 :
2369 2525 : if (gimple_call_num_args (call) != 1
2370 2525 : || TREE_CODE (gimple_call_arg (call, 0)) != SSA_NAME
2371 5050 : || !POINTER_TYPE_P (TREE_TYPE (gimple_call_arg (call, 0))))
2372 : return false;
2373 :
2374 2525 : gimple_stmt_iterator i = *gsi;
2375 6288 : for (gsi_next (&i); !gsi_end_p (i); gsi_next (&i))
2376 : {
2377 4218 : stmt = gsi_stmt (i);
2378 4218 : if (is_a<gasm*> (stmt))
2379 : return false;
2380 4217 : gcall *call = dyn_cast<gcall*>(stmt);
2381 4217 : if (!call)
2382 3554 : continue;
2383 :
2384 : /* We can remove the restore in front of noreturn
2385 : calls. Since the restore will happen either
2386 : via an unwind/longjmp or not at all. */
2387 663 : if (gimple_call_noreturn_p (call))
2388 : break;
2389 :
2390 : /* Internal calls are ok, to bypass
2391 : check first since fndecl will be null. */
2392 647 : if (gimple_call_internal_p (call))
2393 1 : continue;
2394 :
2395 646 : callee = gimple_call_fndecl (call);
2396 : /* Non-builtin calls are not ok. */
2397 646 : if (!callee
2398 646 : || !fndecl_built_in_p (callee))
2399 : return false;
2400 :
2401 : /* Do not remove stack updates before strub leave. */
2402 570 : if (fndecl_built_in_p (callee, BUILT_IN___STRUB_LEAVE)
2403 : /* Alloca calls are not ok either. */
2404 570 : || fndecl_builtin_alloc_p (callee))
2405 : return false;
2406 :
2407 355 : if (fndecl_built_in_p (callee, BUILT_IN_STACK_RESTORE))
2408 52 : goto second_stack_restore;
2409 :
2410 : /* If not a simple or inexpensive builtin, then it is not ok either. */
2411 303 : if (!is_simple_builtin (callee)
2412 303 : && !is_inexpensive_builtin (callee))
2413 : return false;
2414 : }
2415 :
2416 : /* Allow one successor of the exit block, or zero successors. */
2417 2086 : switch (EDGE_COUNT (bb->succs))
2418 : {
2419 : case 0:
2420 : break;
2421 1999 : case 1:
2422 1999 : if (single_succ_edge (bb)->dest != EXIT_BLOCK_PTR_FOR_FN (cfun))
2423 : return false;
2424 : break;
2425 : default:
2426 : return false;
2427 : }
2428 1729 : second_stack_restore:
2429 :
2430 : /* If there's exactly one use, then zap the call to __builtin_stack_save.
2431 : If there are multiple uses, then the last one should remove the call.
2432 : In any case, whether the call to __builtin_stack_save can be removed
2433 : or not is irrelevant to removing the call to __builtin_stack_restore. */
2434 1729 : if (has_single_use (gimple_call_arg (call, 0)))
2435 : {
2436 1559 : gimple *stack_save = SSA_NAME_DEF_STMT (gimple_call_arg (call, 0));
2437 1559 : if (is_gimple_call (stack_save))
2438 : {
2439 1557 : callee = gimple_call_fndecl (stack_save);
2440 1557 : if (callee && fndecl_built_in_p (callee, BUILT_IN_STACK_SAVE))
2441 : {
2442 1557 : gimple_stmt_iterator stack_save_gsi;
2443 1557 : tree rhs;
2444 :
2445 1557 : stack_save_gsi = gsi_for_stmt (stack_save);
2446 1557 : rhs = build_int_cst (TREE_TYPE (gimple_call_arg (call, 0)), 0);
2447 1557 : replace_call_with_value (&stack_save_gsi, rhs);
2448 : }
2449 : }
2450 : }
2451 :
2452 : /* No effect, so the statement will be deleted. */
2453 1729 : replace_call_with_value (gsi, NULL_TREE);
2454 1729 : return true;
2455 : }
2456 :
2457 : /* *GSI_P is a GIMPLE_CALL to a builtin function.
2458 : Optimize
2459 : memcpy (p, "abcd", 4);
2460 : memset (p + 4, ' ', 3);
2461 : into
2462 : memcpy (p, "abcd ", 7);
2463 : call if the latter can be stored by pieces during expansion.
2464 :
2465 : Optimize
2466 : memchr ("abcd", a, 4) == 0;
2467 : or
2468 : memchr ("abcd", a, 4) != 0;
2469 : to
2470 : (a == 'a' || a == 'b' || a == 'c' || a == 'd') == 0
2471 : or
2472 : (a == 'a' || a == 'b' || a == 'c' || a == 'd') != 0
2473 :
2474 : Also canonicalize __atomic_fetch_op (p, x, y) op x
2475 : to __atomic_op_fetch (p, x, y) or
2476 : __atomic_op_fetch (p, x, y) iop x
2477 : to __atomic_fetch_op (p, x, y) when possible (also __sync). */
2478 :
2479 : static bool
2480 6293190 : simplify_builtin_call (gimple_stmt_iterator *gsi_p, tree callee2, bool full_walk)
2481 : {
2482 6293190 : gimple *stmt2 = gsi_stmt (*gsi_p);
2483 6293190 : enum built_in_function other_atomic = END_BUILTINS;
2484 6293190 : enum tree_code atomic_op = ERROR_MARK;
2485 :
2486 6293190 : switch (DECL_FUNCTION_CODE (callee2))
2487 : {
2488 10369 : case BUILT_IN_STACK_RESTORE:
2489 10369 : return optimize_stack_restore (gsi_p, as_a<gcall*>(stmt2));
2490 470836 : case BUILT_IN_MEMCMP:
2491 470836 : case BUILT_IN_MEMCMP_EQ:
2492 470836 : return simplify_builtin_memcmp (gsi_p, as_a<gcall*>(stmt2));
2493 14509 : case BUILT_IN_MEMCHR:
2494 14509 : return simplify_builtin_memchr (gsi_p, as_a<gcall*>(stmt2));
2495 :
2496 112011 : case BUILT_IN_MEMSET:
2497 112011 : if (gimple_call_num_args (stmt2) == 3)
2498 : {
2499 : /* Try to prop the zeroing/value of the memset to memcpy
2500 : if the dest is an address and the value is a constant. */
2501 112011 : optimize_aggr_zeroprop (stmt2, full_walk);
2502 : }
2503 112011 : return simplify_builtin_memcpy_memset (gsi_p, as_a<gcall*>(stmt2));
2504 :
2505 : #define CASE_ATOMIC(NAME, OTHER, OP) \
2506 : case BUILT_IN_##NAME##_1: \
2507 : case BUILT_IN_##NAME##_2: \
2508 : case BUILT_IN_##NAME##_4: \
2509 : case BUILT_IN_##NAME##_8: \
2510 : case BUILT_IN_##NAME##_16: \
2511 : atomic_op = OP; \
2512 : other_atomic \
2513 : = (enum built_in_function) (BUILT_IN_##OTHER##_1 \
2514 : + (DECL_FUNCTION_CODE (callee2) \
2515 : - BUILT_IN_##NAME##_1)); \
2516 : goto handle_atomic_fetch_op;
2517 :
2518 48548 : CASE_ATOMIC (ATOMIC_FETCH_ADD, ATOMIC_ADD_FETCH, PLUS_EXPR)
2519 7125 : CASE_ATOMIC (ATOMIC_FETCH_SUB, ATOMIC_SUB_FETCH, MINUS_EXPR)
2520 2876 : CASE_ATOMIC (ATOMIC_FETCH_AND, ATOMIC_AND_FETCH, BIT_AND_EXPR)
2521 2895 : CASE_ATOMIC (ATOMIC_FETCH_XOR, ATOMIC_XOR_FETCH, BIT_XOR_EXPR)
2522 3819 : CASE_ATOMIC (ATOMIC_FETCH_OR, ATOMIC_OR_FETCH, BIT_IOR_EXPR)
2523 :
2524 2365 : CASE_ATOMIC (SYNC_FETCH_AND_ADD, SYNC_ADD_AND_FETCH, PLUS_EXPR)
2525 2004 : CASE_ATOMIC (SYNC_FETCH_AND_SUB, SYNC_SUB_AND_FETCH, MINUS_EXPR)
2526 1876 : CASE_ATOMIC (SYNC_FETCH_AND_AND, SYNC_AND_AND_FETCH, BIT_AND_EXPR)
2527 2144 : CASE_ATOMIC (SYNC_FETCH_AND_XOR, SYNC_XOR_AND_FETCH, BIT_XOR_EXPR)
2528 1987 : CASE_ATOMIC (SYNC_FETCH_AND_OR, SYNC_OR_AND_FETCH, BIT_IOR_EXPR)
2529 :
2530 14400 : CASE_ATOMIC (ATOMIC_ADD_FETCH, ATOMIC_FETCH_ADD, MINUS_EXPR)
2531 8552 : CASE_ATOMIC (ATOMIC_SUB_FETCH, ATOMIC_FETCH_SUB, PLUS_EXPR)
2532 2380 : CASE_ATOMIC (ATOMIC_XOR_FETCH, ATOMIC_FETCH_XOR, BIT_XOR_EXPR)
2533 :
2534 846 : CASE_ATOMIC (SYNC_ADD_AND_FETCH, SYNC_FETCH_AND_ADD, MINUS_EXPR)
2535 732 : CASE_ATOMIC (SYNC_SUB_AND_FETCH, SYNC_FETCH_AND_SUB, PLUS_EXPR)
2536 800 : CASE_ATOMIC (SYNC_XOR_AND_FETCH, SYNC_FETCH_AND_XOR, BIT_XOR_EXPR)
2537 :
2538 : #undef CASE_ATOMIC
2539 :
2540 103349 : handle_atomic_fetch_op:
2541 103349 : if (gimple_call_num_args (stmt2) >= 2 && gimple_call_lhs (stmt2))
2542 : {
2543 60088 : tree lhs2 = gimple_call_lhs (stmt2), lhsc = lhs2;
2544 60088 : tree arg = gimple_call_arg (stmt2, 1);
2545 60088 : gimple *use_stmt, *cast_stmt = NULL;
2546 60088 : use_operand_p use_p;
2547 60088 : tree ndecl = builtin_decl_explicit (other_atomic);
2548 :
2549 60088 : if (ndecl == NULL_TREE || !single_imm_use (lhs2, &use_p, &use_stmt))
2550 : break;
2551 :
2552 58959 : if (gimple_assign_cast_p (use_stmt))
2553 : {
2554 31403 : cast_stmt = use_stmt;
2555 31403 : lhsc = gimple_assign_lhs (cast_stmt);
2556 31403 : if (lhsc == NULL_TREE
2557 31403 : || !INTEGRAL_TYPE_P (TREE_TYPE (lhsc))
2558 30852 : || (TYPE_PRECISION (TREE_TYPE (lhsc))
2559 30852 : != TYPE_PRECISION (TREE_TYPE (lhs2)))
2560 60729 : || !single_imm_use (lhsc, &use_p, &use_stmt))
2561 : {
2562 2605 : use_stmt = cast_stmt;
2563 2605 : cast_stmt = NULL;
2564 2605 : lhsc = lhs2;
2565 : }
2566 : }
2567 :
2568 58959 : bool ok = false;
2569 58959 : tree oarg = NULL_TREE;
2570 58959 : enum tree_code ccode = ERROR_MARK;
2571 58959 : tree crhs1 = NULL_TREE, crhs2 = NULL_TREE;
2572 58959 : if (is_gimple_assign (use_stmt)
2573 58959 : && gimple_assign_rhs_code (use_stmt) == atomic_op)
2574 : {
2575 1416 : if (gimple_assign_rhs1 (use_stmt) == lhsc)
2576 1016 : oarg = gimple_assign_rhs2 (use_stmt);
2577 400 : else if (atomic_op != MINUS_EXPR)
2578 : oarg = gimple_assign_rhs1 (use_stmt);
2579 : }
2580 57543 : else if (atomic_op == MINUS_EXPR
2581 13246 : && is_gimple_assign (use_stmt)
2582 3606 : && gimple_assign_rhs_code (use_stmt) == PLUS_EXPR
2583 199 : && TREE_CODE (arg) == INTEGER_CST
2584 57742 : && (TREE_CODE (gimple_assign_rhs2 (use_stmt))
2585 : == INTEGER_CST))
2586 : {
2587 183 : tree a = fold_convert (TREE_TYPE (lhs2), arg);
2588 183 : tree o = fold_convert (TREE_TYPE (lhs2),
2589 : gimple_assign_rhs2 (use_stmt));
2590 183 : if (wi::to_wide (a) == wi::neg (wi::to_wide (o)))
2591 : ok = true;
2592 : }
2593 57360 : else if (atomic_op == BIT_AND_EXPR || atomic_op == BIT_IOR_EXPR)
2594 : ;
2595 52118 : else if (gimple_code (use_stmt) == GIMPLE_COND)
2596 : {
2597 19580 : ccode = gimple_cond_code (use_stmt);
2598 19580 : crhs1 = gimple_cond_lhs (use_stmt);
2599 19580 : crhs2 = gimple_cond_rhs (use_stmt);
2600 : }
2601 32538 : else if (is_gimple_assign (use_stmt))
2602 : {
2603 9519 : if (gimple_assign_rhs_class (use_stmt) == GIMPLE_BINARY_RHS)
2604 : {
2605 3935 : ccode = gimple_assign_rhs_code (use_stmt);
2606 3935 : crhs1 = gimple_assign_rhs1 (use_stmt);
2607 3935 : crhs2 = gimple_assign_rhs2 (use_stmt);
2608 : }
2609 5584 : else if (gimple_assign_rhs_code (use_stmt) == COND_EXPR)
2610 : {
2611 0 : tree cond = gimple_assign_rhs1 (use_stmt);
2612 0 : if (COMPARISON_CLASS_P (cond))
2613 : {
2614 0 : ccode = TREE_CODE (cond);
2615 0 : crhs1 = TREE_OPERAND (cond, 0);
2616 0 : crhs2 = TREE_OPERAND (cond, 1);
2617 : }
2618 : }
2619 : }
2620 24531 : if (ccode == EQ_EXPR || ccode == NE_EXPR)
2621 : {
2622 : /* Deal with x - y == 0 or x ^ y == 0
2623 : being optimized into x == y and x + cst == 0
2624 : into x == -cst. */
2625 22331 : tree o = NULL_TREE;
2626 22331 : if (crhs1 == lhsc)
2627 : o = crhs2;
2628 133 : else if (crhs2 == lhsc)
2629 133 : o = crhs1;
2630 22331 : if (o && atomic_op != PLUS_EXPR)
2631 : oarg = o;
2632 10116 : else if (o
2633 10116 : && TREE_CODE (o) == INTEGER_CST
2634 10116 : && TREE_CODE (arg) == INTEGER_CST)
2635 : {
2636 9406 : tree a = fold_convert (TREE_TYPE (lhs2), arg);
2637 9406 : o = fold_convert (TREE_TYPE (lhs2), o);
2638 9406 : if (wi::to_wide (a) == wi::neg (wi::to_wide (o)))
2639 58959 : ok = true;
2640 : }
2641 : }
2642 58959 : if (oarg && !ok)
2643 : {
2644 13631 : if (operand_equal_p (arg, oarg, 0))
2645 : ok = true;
2646 12302 : else if (TREE_CODE (arg) == SSA_NAME
2647 2203 : && TREE_CODE (oarg) == SSA_NAME)
2648 : {
2649 745 : tree oarg2 = oarg;
2650 745 : if (gimple_assign_cast_p (SSA_NAME_DEF_STMT (oarg)))
2651 : {
2652 104 : gimple *g = SSA_NAME_DEF_STMT (oarg);
2653 104 : oarg2 = gimple_assign_rhs1 (g);
2654 104 : if (TREE_CODE (oarg2) != SSA_NAME
2655 104 : || !INTEGRAL_TYPE_P (TREE_TYPE (oarg2))
2656 208 : || (TYPE_PRECISION (TREE_TYPE (oarg2))
2657 104 : != TYPE_PRECISION (TREE_TYPE (oarg))))
2658 : oarg2 = oarg;
2659 : }
2660 745 : if (gimple_assign_cast_p (SSA_NAME_DEF_STMT (arg)))
2661 : {
2662 544 : gimple *g = SSA_NAME_DEF_STMT (arg);
2663 544 : tree rhs1 = gimple_assign_rhs1 (g);
2664 : /* Handle e.g.
2665 : x.0_1 = (long unsigned int) x_4(D);
2666 : _2 = __atomic_fetch_add_8 (&vlong, x.0_1, 0);
2667 : _3 = (long int) _2;
2668 : _7 = x_4(D) + _3; */
2669 544 : if (rhs1 == oarg || rhs1 == oarg2)
2670 : ok = true;
2671 : /* Handle e.g.
2672 : x.18_1 = (short unsigned int) x_5(D);
2673 : _2 = (int) x.18_1;
2674 : _3 = __atomic_fetch_xor_2 (&vshort, _2, 0);
2675 : _4 = (short int) _3;
2676 : _8 = x_5(D) ^ _4;
2677 : This happens only for char/short. */
2678 160 : else if (TREE_CODE (rhs1) == SSA_NAME
2679 160 : && INTEGRAL_TYPE_P (TREE_TYPE (rhs1))
2680 320 : && (TYPE_PRECISION (TREE_TYPE (rhs1))
2681 160 : == TYPE_PRECISION (TREE_TYPE (lhs2))))
2682 : {
2683 160 : g = SSA_NAME_DEF_STMT (rhs1);
2684 160 : if (gimple_assign_cast_p (g)
2685 160 : && (gimple_assign_rhs1 (g) == oarg
2686 0 : || gimple_assign_rhs1 (g) == oarg2))
2687 : ok = true;
2688 : }
2689 : }
2690 745 : if (!ok && arg == oarg2)
2691 : /* Handle e.g.
2692 : _1 = __sync_fetch_and_add_4 (&v, x_5(D));
2693 : _2 = (int) _1;
2694 : x.0_3 = (int) x_5(D);
2695 : _7 = _2 + x.0_3; */
2696 : ok = true;
2697 : }
2698 : }
2699 :
2700 57630 : if (ok)
2701 : {
2702 2605 : tree new_lhs = make_ssa_name (TREE_TYPE (lhs2));
2703 2605 : gimple_call_set_lhs (stmt2, new_lhs);
2704 2605 : gimple_call_set_fndecl (stmt2, ndecl);
2705 2605 : gimple_stmt_iterator gsi = gsi_for_stmt (use_stmt);
2706 2605 : if (ccode == ERROR_MARK)
2707 2000 : gimple_assign_set_rhs_with_ops (&gsi, cast_stmt
2708 : ? NOP_EXPR : SSA_NAME,
2709 : new_lhs);
2710 : else
2711 : {
2712 1382 : crhs1 = new_lhs;
2713 1382 : crhs2 = build_zero_cst (TREE_TYPE (lhs2));
2714 1382 : if (gimple_code (use_stmt) == GIMPLE_COND)
2715 : {
2716 1043 : gcond *cond_stmt = as_a <gcond *> (use_stmt);
2717 1043 : gimple_cond_set_lhs (cond_stmt, crhs1);
2718 1043 : gimple_cond_set_rhs (cond_stmt, crhs2);
2719 : }
2720 339 : else if (gimple_assign_rhs_class (use_stmt)
2721 : == GIMPLE_BINARY_RHS)
2722 : {
2723 339 : gimple_assign_set_rhs1 (use_stmt, crhs1);
2724 339 : gimple_assign_set_rhs2 (use_stmt, crhs2);
2725 : }
2726 : else
2727 : {
2728 0 : gcc_checking_assert (gimple_assign_rhs_code (use_stmt)
2729 : == COND_EXPR);
2730 0 : tree cond = build2 (ccode, boolean_type_node,
2731 : crhs1, crhs2);
2732 0 : gimple_assign_set_rhs1 (use_stmt, cond);
2733 : }
2734 : }
2735 2605 : update_stmt (use_stmt);
2736 2605 : if (atomic_op != BIT_AND_EXPR
2737 2605 : && atomic_op != BIT_IOR_EXPR
2738 2605 : && !stmt_ends_bb_p (stmt2))
2739 : {
2740 : /* For the benefit of debug stmts, emit stmt(s) to set
2741 : lhs2 to the value it had from the new builtin.
2742 : E.g. if it was previously:
2743 : lhs2 = __atomic_fetch_add_8 (ptr, arg, 0);
2744 : emit:
2745 : new_lhs = __atomic_add_fetch_8 (ptr, arg, 0);
2746 : lhs2 = new_lhs - arg;
2747 : We also keep cast_stmt if any in the IL for
2748 : the same reasons.
2749 : These stmts will be DCEd later and proper debug info
2750 : will be emitted.
2751 : This is only possible for reversible operations
2752 : (+/-/^) and without -fnon-call-exceptions. */
2753 2264 : gsi = gsi_for_stmt (stmt2);
2754 2264 : tree type = TREE_TYPE (lhs2);
2755 2264 : if (TREE_CODE (arg) == INTEGER_CST)
2756 1682 : arg = fold_convert (type, arg);
2757 582 : else if (!useless_type_conversion_p (type, TREE_TYPE (arg)))
2758 : {
2759 0 : tree narg = make_ssa_name (type);
2760 0 : gimple *g = gimple_build_assign (narg, NOP_EXPR, arg);
2761 0 : gsi_insert_after (&gsi, g, GSI_NEW_STMT);
2762 0 : arg = narg;
2763 : }
2764 2264 : enum tree_code rcode;
2765 2264 : switch (atomic_op)
2766 : {
2767 : case PLUS_EXPR: rcode = MINUS_EXPR; break;
2768 727 : case MINUS_EXPR: rcode = PLUS_EXPR; break;
2769 492 : case BIT_XOR_EXPR: rcode = atomic_op; break;
2770 0 : default: gcc_unreachable ();
2771 : }
2772 2264 : gimple *g = gimple_build_assign (lhs2, rcode, new_lhs, arg);
2773 2264 : gsi_insert_after (&gsi, g, GSI_NEW_STMT);
2774 2264 : update_stmt (stmt2);
2775 : }
2776 : else
2777 : {
2778 : /* For e.g.
2779 : lhs2 = __atomic_fetch_or_8 (ptr, arg, 0);
2780 : after we change it to
2781 : new_lhs = __atomic_or_fetch_8 (ptr, arg, 0);
2782 : there is no way to find out the lhs2 value (i.e.
2783 : what the atomic memory contained before the operation),
2784 : values of some bits are lost. We have checked earlier
2785 : that we don't have any non-debug users except for what
2786 : we are already changing, so we need to reset the
2787 : debug stmts and remove the cast_stmt if any. */
2788 341 : imm_use_iterator iter;
2789 676 : FOR_EACH_IMM_USE_STMT (use_stmt, iter, lhs2)
2790 335 : if (use_stmt != cast_stmt)
2791 : {
2792 168 : gcc_assert (is_gimple_debug (use_stmt));
2793 168 : gimple_debug_bind_reset_value (use_stmt);
2794 168 : update_stmt (use_stmt);
2795 341 : }
2796 341 : if (cast_stmt)
2797 : {
2798 167 : gsi = gsi_for_stmt (cast_stmt);
2799 167 : gsi_remove (&gsi, true);
2800 : }
2801 341 : update_stmt (stmt2);
2802 341 : release_ssa_name (lhs2);
2803 : }
2804 : }
2805 : }
2806 : break;
2807 :
2808 : default:
2809 : break;
2810 : }
2811 : return false;
2812 : }
2813 :
2814 : /* Given a ssa_name in NAME see if it was defined by an assignment and
2815 : set CODE to be the code and ARG1 to the first operand on the rhs and ARG2
2816 : to the second operand on the rhs. */
2817 :
2818 : static inline void
2819 17566123 : defcodefor_name (tree name, enum tree_code *code, tree *arg1, tree *arg2)
2820 : {
2821 17566123 : gimple *def;
2822 17566123 : enum tree_code code1;
2823 17566123 : tree arg11;
2824 17566123 : tree arg21;
2825 17566123 : tree arg31;
2826 17566123 : enum gimple_rhs_class grhs_class;
2827 :
2828 17566123 : code1 = TREE_CODE (name);
2829 17566123 : arg11 = name;
2830 17566123 : arg21 = NULL_TREE;
2831 17566123 : arg31 = NULL_TREE;
2832 17566123 : grhs_class = get_gimple_rhs_class (code1);
2833 :
2834 17566123 : if (code1 == SSA_NAME)
2835 : {
2836 11703723 : def = SSA_NAME_DEF_STMT (name);
2837 :
2838 11703723 : if (def && is_gimple_assign (def)
2839 19012879 : && can_propagate_from (def))
2840 : {
2841 5017144 : code1 = gimple_assign_rhs_code (def);
2842 5017144 : arg11 = gimple_assign_rhs1 (def);
2843 5017144 : arg21 = gimple_assign_rhs2 (def);
2844 5017144 : arg31 = gimple_assign_rhs3 (def);
2845 : }
2846 : }
2847 5862400 : else if (grhs_class != GIMPLE_SINGLE_RHS)
2848 0 : code1 = ERROR_MARK;
2849 :
2850 17566123 : *code = code1;
2851 17566123 : *arg1 = arg11;
2852 17566123 : if (arg2)
2853 17548863 : *arg2 = arg21;
2854 17566123 : if (arg31)
2855 2343 : *code = ERROR_MARK;
2856 17566123 : }
2857 :
2858 :
2859 : /* Recognize rotation patterns. Return true if a transformation
2860 : applied, otherwise return false.
2861 :
2862 : We are looking for X with unsigned type T with bitsize B, OP being
2863 : +, | or ^, some type T2 wider than T. For:
2864 : (X << CNT1) OP (X >> CNT2) iff CNT1 + CNT2 == B
2865 : ((T) ((T2) X << CNT1)) OP ((T) ((T2) X >> CNT2)) iff CNT1 + CNT2 == B
2866 :
2867 : transform these into:
2868 : X r<< CNT1
2869 :
2870 : Or for:
2871 : (X << Y) OP (X >> (B - Y))
2872 : (X << (int) Y) OP (X >> (int) (B - Y))
2873 : ((T) ((T2) X << Y)) OP ((T) ((T2) X >> (B - Y)))
2874 : ((T) ((T2) X << (int) Y)) OP ((T) ((T2) X >> (int) (B - Y)))
2875 : (X << Y) | (X >> ((-Y) & (B - 1)))
2876 : (X << (int) Y) | (X >> (int) ((-Y) & (B - 1)))
2877 : ((T) ((T2) X << Y)) | ((T) ((T2) X >> ((-Y) & (B - 1))))
2878 : ((T) ((T2) X << (int) Y)) | ((T) ((T2) X >> (int) ((-Y) & (B - 1))))
2879 :
2880 : transform these into (last 2 only if ranger can prove Y < B
2881 : or Y = N * B):
2882 : X r<< Y
2883 : or
2884 : X r<< (& & (B - 1))
2885 : The latter for the forms with T2 wider than T if ranger can't prove Y < B.
2886 :
2887 : Or for:
2888 : (X << (Y & (B - 1))) | (X >> ((-Y) & (B - 1)))
2889 : (X << (int) (Y & (B - 1))) | (X >> (int) ((-Y) & (B - 1)))
2890 : ((T) ((T2) X << (Y & (B - 1)))) | ((T) ((T2) X >> ((-Y) & (B - 1))))
2891 : ((T) ((T2) X << (int) (Y & (B - 1)))) \
2892 : | ((T) ((T2) X >> (int) ((-Y) & (B - 1))))
2893 :
2894 : transform these into:
2895 : X r<< (Y & (B - 1))
2896 :
2897 : Note, in the patterns with T2 type, the type of OP operands
2898 : might be even a signed type, but should have precision B.
2899 : Expressions with & (B - 1) should be recognized only if B is
2900 : a power of 2. */
2901 :
2902 : static bool
2903 10312220 : simplify_rotate (gimple_stmt_iterator *gsi)
2904 : {
2905 10312220 : gimple *stmt = gsi_stmt (*gsi);
2906 10312220 : tree arg[2], rtype, rotcnt = NULL_TREE;
2907 10312220 : tree def_arg1[2], def_arg2[2];
2908 10312220 : enum tree_code def_code[2];
2909 10312220 : tree lhs;
2910 10312220 : int i;
2911 10312220 : bool swapped_p = false;
2912 10312220 : gimple *g;
2913 10312220 : gimple *def_arg_stmt[2] = { NULL, NULL };
2914 10312220 : int wider_prec = 0;
2915 10312220 : bool add_masking = false;
2916 :
2917 10312220 : arg[0] = gimple_assign_rhs1 (stmt);
2918 10312220 : arg[1] = gimple_assign_rhs2 (stmt);
2919 10312220 : rtype = TREE_TYPE (arg[0]);
2920 :
2921 : /* Only create rotates in complete modes. Other cases are not
2922 : expanded properly. */
2923 10312220 : if (!INTEGRAL_TYPE_P (rtype)
2924 10312220 : || !type_has_mode_precision_p (rtype))
2925 1579065 : return false;
2926 :
2927 26199465 : for (i = 0; i < 2; i++)
2928 : {
2929 17466310 : defcodefor_name (arg[i], &def_code[i], &def_arg1[i], &def_arg2[i]);
2930 17466310 : if (TREE_CODE (arg[i]) == SSA_NAME)
2931 11603910 : def_arg_stmt[i] = SSA_NAME_DEF_STMT (arg[i]);
2932 : }
2933 :
2934 : /* Look through narrowing (or same precision) conversions. */
2935 7766183 : if (CONVERT_EXPR_CODE_P (def_code[0])
2936 966972 : && CONVERT_EXPR_CODE_P (def_code[1])
2937 140889 : && INTEGRAL_TYPE_P (TREE_TYPE (def_arg1[0]))
2938 117207 : && INTEGRAL_TYPE_P (TREE_TYPE (def_arg1[1]))
2939 109663 : && TYPE_PRECISION (TREE_TYPE (def_arg1[0]))
2940 109663 : == TYPE_PRECISION (TREE_TYPE (def_arg1[1]))
2941 63714 : && TYPE_PRECISION (TREE_TYPE (def_arg1[0])) >= TYPE_PRECISION (rtype)
2942 44690 : && has_single_use (arg[0])
2943 8766261 : && has_single_use (arg[1]))
2944 : {
2945 28621 : wider_prec = TYPE_PRECISION (TREE_TYPE (def_arg1[0]));
2946 85863 : for (i = 0; i < 2; i++)
2947 : {
2948 57242 : arg[i] = def_arg1[i];
2949 57242 : defcodefor_name (arg[i], &def_code[i], &def_arg1[i], &def_arg2[i]);
2950 57242 : if (TREE_CODE (arg[i]) == SSA_NAME)
2951 57242 : def_arg_stmt[i] = SSA_NAME_DEF_STMT (arg[i]);
2952 : }
2953 : }
2954 : else
2955 : {
2956 : /* Handle signed rotate; the RSHIFT_EXPR has to be done
2957 : in unsigned type but LSHIFT_EXPR could be signed. */
2958 8704534 : i = (def_code[0] == LSHIFT_EXPR || def_code[0] == RSHIFT_EXPR);
2959 7748643 : if (CONVERT_EXPR_CODE_P (def_code[i])
2960 955891 : && (def_code[1 - i] == LSHIFT_EXPR || def_code[1 - i] == RSHIFT_EXPR)
2961 30654 : && INTEGRAL_TYPE_P (TREE_TYPE (def_arg1[i]))
2962 29391 : && TYPE_PRECISION (rtype) == TYPE_PRECISION (TREE_TYPE (def_arg1[i]))
2963 8709436 : && has_single_use (arg[i]))
2964 : {
2965 2051 : arg[i] = def_arg1[i];
2966 2051 : defcodefor_name (arg[i], &def_code[i], &def_arg1[i], &def_arg2[i]);
2967 2051 : if (TREE_CODE (arg[i]) == SSA_NAME)
2968 2051 : def_arg_stmt[i] = SSA_NAME_DEF_STMT (arg[i]);
2969 : }
2970 : }
2971 :
2972 : /* One operand has to be LSHIFT_EXPR and one RSHIFT_EXPR. */
2973 8931969 : for (i = 0; i < 2; i++)
2974 8907233 : if (def_code[i] != LSHIFT_EXPR && def_code[i] != RSHIFT_EXPR)
2975 : return false;
2976 239439 : else if (!has_single_use (arg[i]))
2977 : return false;
2978 24736 : if (def_code[0] == def_code[1])
2979 : return false;
2980 :
2981 : /* If we've looked through narrowing conversions before, look through
2982 : widening conversions from unsigned type with the same precision
2983 : as rtype here. */
2984 20366 : if (TYPE_PRECISION (TREE_TYPE (def_arg1[0])) != TYPE_PRECISION (rtype))
2985 19348 : for (i = 0; i < 2; i++)
2986 : {
2987 12900 : tree tem;
2988 12900 : enum tree_code code;
2989 12900 : defcodefor_name (def_arg1[i], &code, &tem, NULL);
2990 4 : if (!CONVERT_EXPR_CODE_P (code)
2991 12896 : || !INTEGRAL_TYPE_P (TREE_TYPE (tem))
2992 25796 : || TYPE_PRECISION (TREE_TYPE (tem)) != TYPE_PRECISION (rtype))
2993 4 : return false;
2994 12896 : def_arg1[i] = tem;
2995 : }
2996 : /* Both shifts have to use the same first operand. */
2997 20362 : if (!operand_equal_for_phi_arg_p (def_arg1[0], def_arg1[1])
2998 32299 : || !types_compatible_p (TREE_TYPE (def_arg1[0]),
2999 11937 : TREE_TYPE (def_arg1[1])))
3000 : {
3001 8425 : if ((TYPE_PRECISION (TREE_TYPE (def_arg1[0]))
3002 8425 : != TYPE_PRECISION (TREE_TYPE (def_arg1[1])))
3003 8425 : || (TYPE_UNSIGNED (TREE_TYPE (def_arg1[0]))
3004 8425 : == TYPE_UNSIGNED (TREE_TYPE (def_arg1[1]))))
3005 8401 : return false;
3006 :
3007 : /* Handle signed rotate; the RSHIFT_EXPR has to be done
3008 : in unsigned type but LSHIFT_EXPR could be signed. */
3009 546 : i = def_code[0] != RSHIFT_EXPR;
3010 546 : if (!TYPE_UNSIGNED (TREE_TYPE (def_arg1[i])))
3011 : return false;
3012 :
3013 507 : tree tem;
3014 507 : enum tree_code code;
3015 507 : defcodefor_name (def_arg1[i], &code, &tem, NULL);
3016 304 : if (!CONVERT_EXPR_CODE_P (code)
3017 203 : || !INTEGRAL_TYPE_P (TREE_TYPE (tem))
3018 710 : || TYPE_PRECISION (TREE_TYPE (tem)) != TYPE_PRECISION (rtype))
3019 : return false;
3020 194 : def_arg1[i] = tem;
3021 194 : if (!operand_equal_for_phi_arg_p (def_arg1[0], def_arg1[1])
3022 218 : || !types_compatible_p (TREE_TYPE (def_arg1[0]),
3023 24 : TREE_TYPE (def_arg1[1])))
3024 170 : return false;
3025 : }
3026 11937 : else if (!TYPE_UNSIGNED (TREE_TYPE (def_arg1[0])))
3027 : return false;
3028 :
3029 : /* CNT1 + CNT2 == B case above. */
3030 10706 : if (tree_fits_uhwi_p (def_arg2[0])
3031 1210 : && tree_fits_uhwi_p (def_arg2[1])
3032 10706 : && tree_to_uhwi (def_arg2[0])
3033 1210 : + tree_to_uhwi (def_arg2[1]) == TYPE_PRECISION (rtype))
3034 : rotcnt = def_arg2[0];
3035 9776 : else if (TREE_CODE (def_arg2[0]) != SSA_NAME
3036 9496 : || TREE_CODE (def_arg2[1]) != SSA_NAME)
3037 : return false;
3038 : else
3039 : {
3040 9496 : tree cdef_arg1[2], cdef_arg2[2], def_arg2_alt[2];
3041 9496 : enum tree_code cdef_code[2];
3042 9496 : gimple *def_arg_alt_stmt[2] = { NULL, NULL };
3043 9496 : int check_range = 0;
3044 9496 : gimple *check_range_stmt = NULL;
3045 : /* Look through conversion of the shift count argument.
3046 : The C/C++ FE cast any shift count argument to integer_type_node.
3047 : The only problem might be if the shift count type maximum value
3048 : is equal or smaller than number of bits in rtype. */
3049 28488 : for (i = 0; i < 2; i++)
3050 : {
3051 18992 : def_arg2_alt[i] = def_arg2[i];
3052 18992 : defcodefor_name (def_arg2[i], &cdef_code[i],
3053 : &cdef_arg1[i], &cdef_arg2[i]);
3054 14724 : if (CONVERT_EXPR_CODE_P (cdef_code[i])
3055 4268 : && INTEGRAL_TYPE_P (TREE_TYPE (cdef_arg1[i]))
3056 4268 : && TYPE_PRECISION (TREE_TYPE (cdef_arg1[i]))
3057 8536 : > floor_log2 (TYPE_PRECISION (rtype))
3058 23260 : && type_has_mode_precision_p (TREE_TYPE (cdef_arg1[i])))
3059 : {
3060 4268 : def_arg2_alt[i] = cdef_arg1[i];
3061 4268 : if (TREE_CODE (def_arg2[i]) == SSA_NAME)
3062 4268 : def_arg_alt_stmt[i] = SSA_NAME_DEF_STMT (def_arg2[i]);
3063 4268 : defcodefor_name (def_arg2_alt[i], &cdef_code[i],
3064 : &cdef_arg1[i], &cdef_arg2[i]);
3065 : }
3066 : else
3067 14724 : def_arg_alt_stmt[i] = def_arg_stmt[i];
3068 : }
3069 25812 : for (i = 0; i < 2; i++)
3070 : /* Check for one shift count being Y and the other B - Y,
3071 : with optional casts. */
3072 18641 : if (cdef_code[i] == MINUS_EXPR
3073 862 : && tree_fits_shwi_p (cdef_arg1[i])
3074 862 : && tree_to_shwi (cdef_arg1[i]) == TYPE_PRECISION (rtype)
3075 19463 : && TREE_CODE (cdef_arg2[i]) == SSA_NAME)
3076 : {
3077 822 : tree tem;
3078 822 : enum tree_code code;
3079 :
3080 822 : if (cdef_arg2[i] == def_arg2[1 - i]
3081 472 : || cdef_arg2[i] == def_arg2_alt[1 - i])
3082 : {
3083 350 : rotcnt = cdef_arg2[i];
3084 350 : check_range = -1;
3085 350 : if (cdef_arg2[i] == def_arg2[1 - i])
3086 350 : check_range_stmt = def_arg_stmt[1 - i];
3087 : else
3088 0 : check_range_stmt = def_arg_alt_stmt[1 - i];
3089 806 : break;
3090 : }
3091 472 : defcodefor_name (cdef_arg2[i], &code, &tem, NULL);
3092 16 : if (CONVERT_EXPR_CODE_P (code)
3093 456 : && INTEGRAL_TYPE_P (TREE_TYPE (tem))
3094 456 : && TYPE_PRECISION (TREE_TYPE (tem))
3095 912 : > floor_log2 (TYPE_PRECISION (rtype))
3096 456 : && type_has_mode_precision_p (TREE_TYPE (tem))
3097 928 : && (tem == def_arg2[1 - i]
3098 288 : || tem == def_arg2_alt[1 - i]))
3099 : {
3100 456 : rotcnt = tem;
3101 456 : check_range = -1;
3102 456 : if (tem == def_arg2[1 - i])
3103 168 : check_range_stmt = def_arg_stmt[1 - i];
3104 : else
3105 288 : check_range_stmt = def_arg_alt_stmt[1 - i];
3106 : break;
3107 : }
3108 : }
3109 : /* The above sequence isn't safe for Y being 0,
3110 : because then one of the shifts triggers undefined behavior.
3111 : This alternative is safe even for rotation count of 0.
3112 : One shift count is Y and the other (-Y) & (B - 1).
3113 : Or one shift count is Y & (B - 1) and the other (-Y) & (B - 1). */
3114 17819 : else if (cdef_code[i] == BIT_AND_EXPR
3115 28748 : && pow2p_hwi (TYPE_PRECISION (rtype))
3116 12432 : && tree_fits_shwi_p (cdef_arg2[i])
3117 24864 : && tree_to_shwi (cdef_arg2[i])
3118 12432 : == TYPE_PRECISION (rtype) - 1
3119 12372 : && TREE_CODE (cdef_arg1[i]) == SSA_NAME
3120 30191 : && gimple_assign_rhs_code (stmt) == BIT_IOR_EXPR)
3121 : {
3122 2312 : tree tem;
3123 2312 : enum tree_code code;
3124 :
3125 2312 : defcodefor_name (cdef_arg1[i], &code, &tem, NULL);
3126 2115 : if (CONVERT_EXPR_CODE_P (code)
3127 197 : && INTEGRAL_TYPE_P (TREE_TYPE (tem))
3128 197 : && TYPE_PRECISION (TREE_TYPE (tem))
3129 394 : > floor_log2 (TYPE_PRECISION (rtype))
3130 2509 : && type_has_mode_precision_p (TREE_TYPE (tem)))
3131 197 : defcodefor_name (tem, &code, &tem, NULL);
3132 :
3133 2312 : if (code == NEGATE_EXPR)
3134 : {
3135 1533 : if (tem == def_arg2[1 - i] || tem == def_arg2_alt[1 - i])
3136 : {
3137 854 : rotcnt = tem;
3138 854 : check_range = 1;
3139 854 : if (tem == def_arg2[1 - i])
3140 846 : check_range_stmt = def_arg_stmt[1 - i];
3141 : else
3142 8 : check_range_stmt = def_arg_alt_stmt[1 - i];
3143 1519 : break;
3144 : }
3145 679 : tree tem2;
3146 679 : defcodefor_name (tem, &code, &tem2, NULL);
3147 237 : if (CONVERT_EXPR_CODE_P (code)
3148 442 : && INTEGRAL_TYPE_P (TREE_TYPE (tem2))
3149 442 : && TYPE_PRECISION (TREE_TYPE (tem2))
3150 884 : > floor_log2 (TYPE_PRECISION (rtype))
3151 1121 : && type_has_mode_precision_p (TREE_TYPE (tem2)))
3152 : {
3153 442 : if (tem2 == def_arg2[1 - i]
3154 442 : || tem2 == def_arg2_alt[1 - i])
3155 : {
3156 228 : rotcnt = tem2;
3157 228 : check_range = 1;
3158 228 : if (tem2 == def_arg2[1 - i])
3159 0 : check_range_stmt = def_arg_stmt[1 - i];
3160 : else
3161 228 : check_range_stmt = def_arg_alt_stmt[1 - i];
3162 : break;
3163 : }
3164 : }
3165 : else
3166 237 : tem2 = NULL_TREE;
3167 :
3168 451 : if (cdef_code[1 - i] == BIT_AND_EXPR
3169 438 : && tree_fits_shwi_p (cdef_arg2[1 - i])
3170 876 : && tree_to_shwi (cdef_arg2[1 - i])
3171 438 : == TYPE_PRECISION (rtype) - 1
3172 889 : && TREE_CODE (cdef_arg1[1 - i]) == SSA_NAME)
3173 : {
3174 438 : if (tem == cdef_arg1[1 - i]
3175 213 : || tem2 == cdef_arg1[1 - i])
3176 : {
3177 : rotcnt = def_arg2[1 - i];
3178 437 : break;
3179 : }
3180 193 : tree tem3;
3181 193 : defcodefor_name (cdef_arg1[1 - i], &code, &tem3, NULL);
3182 0 : if (CONVERT_EXPR_CODE_P (code)
3183 193 : && INTEGRAL_TYPE_P (TREE_TYPE (tem3))
3184 193 : && TYPE_PRECISION (TREE_TYPE (tem3))
3185 386 : > floor_log2 (TYPE_PRECISION (rtype))
3186 386 : && type_has_mode_precision_p (TREE_TYPE (tem3)))
3187 : {
3188 193 : if (tem == tem3 || tem2 == tem3)
3189 : {
3190 : rotcnt = def_arg2[1 - i];
3191 : break;
3192 : }
3193 : }
3194 : }
3195 : }
3196 : }
3197 2325 : if (check_range && wider_prec > TYPE_PRECISION (rtype))
3198 : {
3199 1533 : if (TREE_CODE (rotcnt) != SSA_NAME)
3200 573 : return false;
3201 1533 : int_range_max r;
3202 1533 : range_query *q = get_range_query (cfun);
3203 1533 : if (q == get_global_range_query ())
3204 1522 : q = enable_ranger (cfun);
3205 1533 : if (!q->range_of_expr (r, rotcnt, check_range_stmt))
3206 : {
3207 0 : if (check_range > 0)
3208 : return false;
3209 0 : r.set_varying (TREE_TYPE (rotcnt));
3210 : }
3211 1533 : int prec = TYPE_PRECISION (TREE_TYPE (rotcnt));
3212 1533 : signop sign = TYPE_SIGN (TREE_TYPE (rotcnt));
3213 1533 : wide_int min = wide_int::from (TYPE_PRECISION (rtype), prec, sign);
3214 1533 : wide_int max = wide_int::from (wider_prec - 1, prec, sign);
3215 1533 : if (check_range < 0)
3216 616 : max = min;
3217 1533 : int_range<1> r2 (TREE_TYPE (rotcnt), min, max);
3218 1533 : r.intersect (r2);
3219 1533 : if (!r.undefined_p ())
3220 : {
3221 1181 : if (check_range > 0)
3222 : {
3223 589 : int_range_max r3;
3224 1844 : for (int i = TYPE_PRECISION (rtype) + 1; i < wider_prec;
3225 1255 : i += TYPE_PRECISION (rtype))
3226 : {
3227 1255 : int j = i + TYPE_PRECISION (rtype) - 2;
3228 1255 : min = wide_int::from (i, prec, sign);
3229 1255 : max = wide_int::from (MIN (j, wider_prec - 1),
3230 1255 : prec, sign);
3231 1255 : int_range<1> r4 (TREE_TYPE (rotcnt), min, max);
3232 1255 : r3.union_ (r4);
3233 1255 : }
3234 589 : r.intersect (r3);
3235 589 : if (!r.undefined_p ())
3236 573 : return false;
3237 589 : }
3238 : add_masking = true;
3239 : }
3240 1533 : }
3241 8923 : if (rotcnt == NULL_TREE)
3242 : return false;
3243 1752 : swapped_p = i != 1;
3244 : }
3245 :
3246 2682 : if (!useless_type_conversion_p (TREE_TYPE (def_arg2[0]),
3247 2682 : TREE_TYPE (rotcnt)))
3248 : {
3249 496 : g = gimple_build_assign (make_ssa_name (TREE_TYPE (def_arg2[0])),
3250 : NOP_EXPR, rotcnt);
3251 496 : gsi_insert_before (gsi, g, GSI_SAME_STMT);
3252 496 : rotcnt = gimple_assign_lhs (g);
3253 : }
3254 2682 : if (add_masking)
3255 : {
3256 608 : g = gimple_build_assign (make_ssa_name (TREE_TYPE (rotcnt)),
3257 : BIT_AND_EXPR, rotcnt,
3258 608 : build_int_cst (TREE_TYPE (rotcnt),
3259 608 : TYPE_PRECISION (rtype) - 1));
3260 608 : gsi_insert_before (gsi, g, GSI_SAME_STMT);
3261 608 : rotcnt = gimple_assign_lhs (g);
3262 : }
3263 2682 : lhs = gimple_assign_lhs (stmt);
3264 2682 : if (!useless_type_conversion_p (rtype, TREE_TYPE (def_arg1[0])))
3265 1010 : lhs = make_ssa_name (TREE_TYPE (def_arg1[0]));
3266 2682 : g = gimple_build_assign (lhs,
3267 2682 : ((def_code[0] == LSHIFT_EXPR) ^ swapped_p)
3268 : ? LROTATE_EXPR : RROTATE_EXPR, def_arg1[0], rotcnt);
3269 2682 : if (!useless_type_conversion_p (rtype, TREE_TYPE (def_arg1[0])))
3270 : {
3271 1010 : gsi_insert_before (gsi, g, GSI_SAME_STMT);
3272 1010 : g = gimple_build_assign (gimple_assign_lhs (stmt), NOP_EXPR, lhs);
3273 : }
3274 2682 : gsi_replace (gsi, g, false);
3275 2682 : return true;
3276 : }
3277 :
3278 :
3279 : /* Check whether an array contains a valid table according to VALIDATE_FN. */
3280 : template<typename ValidateFn>
3281 : static bool
3282 17 : check_table_array (tree ctor, HOST_WIDE_INT &zero_val, unsigned bits,
3283 : ValidateFn validate_fn)
3284 : {
3285 : tree elt, idx;
3286 17 : unsigned HOST_WIDE_INT i, raw_idx = 0;
3287 17 : unsigned matched = 0;
3288 :
3289 17 : zero_val = 0;
3290 :
3291 734 : FOR_EACH_CONSTRUCTOR_ELT (CONSTRUCTOR_ELTS (ctor), i, idx, elt)
3292 : {
3293 734 : if (!tree_fits_shwi_p (idx))
3294 : return false;
3295 734 : if (!tree_fits_shwi_p (elt) && TREE_CODE (elt) != RAW_DATA_CST)
3296 : return false;
3297 :
3298 734 : unsigned HOST_WIDE_INT index = tree_to_shwi (idx);
3299 : HOST_WIDE_INT val;
3300 :
3301 734 : if (TREE_CODE (elt) == INTEGER_CST)
3302 670 : val = tree_to_shwi (elt);
3303 : else
3304 : {
3305 64 : if (raw_idx == (unsigned) RAW_DATA_LENGTH (elt))
3306 : {
3307 0 : raw_idx = 0;
3308 0 : continue;
3309 : }
3310 64 : if (TYPE_UNSIGNED (TREE_TYPE (elt)))
3311 0 : val = RAW_DATA_UCHAR_ELT (elt, raw_idx);
3312 : else
3313 64 : val = RAW_DATA_SCHAR_ELT (elt, raw_idx);
3314 64 : index += raw_idx;
3315 64 : raw_idx++;
3316 64 : i--;
3317 : }
3318 :
3319 734 : if (index > bits * 2)
3320 : return false;
3321 :
3322 734 : if (index == 0)
3323 : {
3324 17 : zero_val = val;
3325 17 : matched++;
3326 : }
3327 :
3328 734 : if (val >= 0 && val < bits && validate_fn (val, index))
3329 672 : matched++;
3330 :
3331 734 : if (matched > bits)
3332 : return true;
3333 : }
3334 :
3335 : return false;
3336 : }
3337 :
3338 : /* Check whether a string contains a valid table according to VALIDATE_FN. */
3339 : template<typename ValidateFn>
3340 : static bool
3341 4 : check_table_string (tree string, HOST_WIDE_INT &zero_val,unsigned bits,
3342 : ValidateFn validate_fn)
3343 : {
3344 4 : unsigned HOST_WIDE_INT len = TREE_STRING_LENGTH (string);
3345 4 : unsigned matched = 0;
3346 4 : const unsigned char *p = (const unsigned char *) TREE_STRING_POINTER (string);
3347 :
3348 4 : if (len < bits || len > bits * 2)
3349 : return false;
3350 :
3351 4 : zero_val = p[0];
3352 :
3353 164 : for (unsigned i = 0; i < len; i++)
3354 160 : if (p[i] < bits && validate_fn (p[i], i))
3355 160 : matched++;
3356 :
3357 4 : return matched == bits;
3358 : }
3359 :
3360 : /* Check whether CTOR contains a valid table according to VALIDATE_FN. */
3361 : template<typename ValidateFn>
3362 : static bool
3363 29 : check_table (tree ctor, tree type, HOST_WIDE_INT &zero_val, unsigned bits,
3364 : ValidateFn validate_fn)
3365 : {
3366 29 : if (TREE_CODE (ctor) == CONSTRUCTOR)
3367 17 : return check_table_array (ctor, zero_val, bits, validate_fn);
3368 : else if (TREE_CODE (ctor) == STRING_CST
3369 12 : && TYPE_PRECISION (type) == CHAR_TYPE_SIZE)
3370 4 : return check_table_string (ctor, zero_val, bits, validate_fn);
3371 : return false;
3372 : }
3373 :
3374 : /* Match.pd function to match the ctz expression. */
3375 : extern bool gimple_ctz_table_index (tree, tree *, tree (*)(tree));
3376 : extern bool gimple_clz_table_index (tree, tree *, tree (*)(tree));
3377 : extern bool gimple_clz_msb_iso_table_index (tree, tree *, tree (*)(tree));
3378 :
3379 : /* Recognize count leading and trailing zeroes idioms.
3380 : The canonical form is array[((x & -x) * C) >> SHIFT] where C is a magic
3381 : constant which when multiplied by a power of 2 creates a unique value
3382 : in the top 5 or 6 bits. This is then indexed into a table which maps it
3383 : to the number of trailing zeroes. Array[0] is returned so the caller can
3384 : emit an appropriate sequence depending on whether ctz (0) is defined on
3385 : the target. */
3386 :
3387 : static bool
3388 1996254 : simplify_count_zeroes (gimple_stmt_iterator *gsi)
3389 : {
3390 1996254 : gimple *stmt = gsi_stmt (*gsi);
3391 1996254 : tree array_ref = gimple_assign_rhs1 (stmt);
3392 1996254 : tree res_ops[3];
3393 :
3394 1996254 : gcc_checking_assert (TREE_CODE (array_ref) == ARRAY_REF);
3395 :
3396 1996254 : internal_fn fn = IFN_LAST;
3397 : /* When true, the matched idiom is a CLZ using DeBruijn CTZ on the
3398 : isolated MSB -- see clz_msb_iso_table_index in match.pd. The
3399 : table stores MSB positions and must satisfy the direct CTZ
3400 : DeBruijn property, so we validate it with the CTZ checkfn even
3401 : though we emit IFN_CLZ code. */
3402 1996254 : bool clz_via_ctz = false;
3403 : /* For CTZ we recognize ((x & -x) * C) >> SHIFT where the array data
3404 : represents the number of trailing zeros. */
3405 1996254 : if (gimple_ctz_table_index (TREE_OPERAND (array_ref, 1), &res_ops[0], NULL))
3406 : fn = IFN_CTZ;
3407 : /* For CLZ we recognize
3408 : x |= x >> 1;
3409 : x |= x >> 2;
3410 : x |= x >> 4;
3411 : x |= x >> 8;
3412 : x |= x >> 16;
3413 : (x * C) >> SHIFT
3414 : where 31 minus the array data represents the number of leading zeros. */
3415 1996231 : else if (gimple_clz_table_index (TREE_OPERAND (array_ref, 1), &res_ops[0],
3416 : NULL))
3417 : fn = IFN_CLZ;
3418 : /* Variant CLZ idiom: after the OR-cascade sets all bits from 0 to
3419 : the original MSB, (value - (value >> 1)) isolates the MSB as a
3420 : power of two (2^k), and the subsequent DeBruijn multiply-and-shift
3421 : is a CTZ-style lookup on 2^k. The table stores MSB positions
3422 : directly. */
3423 1996221 : else if (gimple_clz_msb_iso_table_index (TREE_OPERAND (array_ref, 1),
3424 : &res_ops[0], NULL))
3425 : {
3426 : fn = IFN_CLZ;
3427 : clz_via_ctz = true;
3428 : }
3429 : else
3430 : return false;
3431 :
3432 34 : HOST_WIDE_INT zero_val;
3433 34 : tree type = TREE_TYPE (array_ref);
3434 34 : tree array = TREE_OPERAND (array_ref, 0);
3435 34 : tree input_type = TREE_TYPE (res_ops[0]);
3436 34 : unsigned input_bits = tree_to_shwi (TYPE_SIZE (input_type));
3437 :
3438 : /* Check the array element type is integral and not wider than 64 bits,
3439 : and the input is an unsigned 32-bit or 64-bit type. The table values
3440 : are bit positions in [0, input_bits - 1], so any integer element type
3441 : with at least 6 bits of precision suffices; the cap is just to keep
3442 : the transformation simple. */
3443 34 : if (!INTEGRAL_TYPE_P (type) || TYPE_PRECISION (type) > 64
3444 68 : || !TYPE_UNSIGNED (input_type))
3445 : return false;
3446 30 : if (input_bits != 32 && input_bits != 64)
3447 : return false;
3448 :
3449 30 : if (!direct_internal_fn_supported_p (fn, input_type, OPTIMIZE_FOR_BOTH))
3450 : return false;
3451 :
3452 : /* Check the lower bound of the array is zero. */
3453 30 : tree low = array_ref_low_bound (array_ref);
3454 30 : if (!low || !integer_zerop (low))
3455 0 : return false;
3456 :
3457 : /* Check the shift extracts the top 5..7 bits. */
3458 30 : unsigned shiftval = tree_to_shwi (res_ops[2]);
3459 30 : if (shiftval < input_bits - 7 || shiftval > input_bits - 5)
3460 : return false;
3461 :
3462 29 : tree ctor = ctor_for_folding (array);
3463 29 : if (!ctor)
3464 : return false;
3465 29 : unsigned HOST_WIDE_INT mulval = tree_to_uhwi (res_ops[1]);
3466 : /* CTZ and the MSB-isolation CLZ variant both use the direct CTZ
3467 : DeBruijn check (table[(magic << data) >> shift] == data). */
3468 29 : if (fn == IFN_CTZ || clz_via_ctz)
3469 : {
3470 559 : auto checkfn = [&](unsigned data, unsigned i) -> bool
3471 : {
3472 540 : unsigned HOST_WIDE_INT mask
3473 540 : = ((HOST_WIDE_INT_1U << (input_bits - shiftval)) - 1) << shiftval;
3474 540 : return (((mulval << data) & mask) >> shiftval) == i;
3475 19 : };
3476 19 : if (!check_table (ctor, type, zero_val, input_bits, checkfn))
3477 8 : return false;
3478 : }
3479 10 : else if (fn == IFN_CLZ)
3480 : {
3481 362 : auto checkfn = [&](unsigned data, unsigned i) -> bool
3482 : {
3483 352 : unsigned HOST_WIDE_INT mask
3484 352 : = ((HOST_WIDE_INT_1U << (input_bits - shiftval)) - 1) << shiftval;
3485 : /* The OR-cascade produces a value with all bits from 0 to the
3486 : original MSB set. Compute (1 << (data + 1)) - 1 to simulate
3487 : that value. When data + 1 equals HOST_BITS_PER_WIDE_INT
3488 : (i.e. data is the MSB position of a 64-bit input) the shift
3489 : is undefined behavior, so handle that case explicitly using
3490 : all-ones. Without this, any well-formed 64-bit DeBruijn CLZ
3491 : table is rejected because its entry for the all-ones input
3492 : correctly maps to the MSB (e.g. table[...] == 63).
3493 : PR tree-optimization/122569. */
3494 703 : unsigned HOST_WIDE_INT all_bits_below
3495 : = (data + 1 == HOST_BITS_PER_WIDE_INT)
3496 352 : ? HOST_WIDE_INT_M1U
3497 351 : : ((HOST_WIDE_INT_1U << (data + 1)) - 1);
3498 352 : return (((all_bits_below * mulval) & mask) >> shiftval) == i;
3499 10 : };
3500 10 : if (!check_table (ctor, type, zero_val, input_bits, checkfn))
3501 0 : return false;
3502 : }
3503 :
3504 21 : HOST_WIDE_INT ctz_val = -1;
3505 21 : bool zero_ok;
3506 21 : if (fn == IFN_CTZ)
3507 : {
3508 10 : ctz_val = 0;
3509 20 : zero_ok = CTZ_DEFINED_VALUE_AT_ZERO (SCALAR_INT_TYPE_MODE (input_type),
3510 : ctz_val) == 2;
3511 : }
3512 11 : else if (fn == IFN_CLZ)
3513 : {
3514 11 : ctz_val = 32;
3515 11 : zero_ok = CLZ_DEFINED_VALUE_AT_ZERO (SCALAR_INT_TYPE_MODE (input_type),
3516 : ctz_val) == 2;
3517 11 : zero_val = input_bits - 1 - zero_val;
3518 : }
3519 21 : int nargs = 2;
3520 :
3521 : /* If the input value can't be zero, don't special case ctz (0). */
3522 21 : range_query *q = get_range_query (cfun);
3523 21 : if (q == get_global_range_query ())
3524 21 : q = enable_ranger (cfun);
3525 21 : int_range_max vr;
3526 21 : if (q->range_of_expr (vr, res_ops[0], stmt)
3527 21 : && !range_includes_zero_p (vr))
3528 : {
3529 4 : zero_ok = true;
3530 4 : zero_val = 0;
3531 4 : ctz_val = 0;
3532 4 : nargs = 1;
3533 : }
3534 :
3535 21 : gimple_seq seq = NULL;
3536 21 : gimple *g;
3537 21 : gcall *call = gimple_build_call_internal (fn, nargs, res_ops[0],
3538 : nargs == 1 ? NULL_TREE
3539 38 : : build_int_cst (integer_type_node,
3540 17 : ctz_val));
3541 21 : gimple_set_location (call, gimple_location (stmt));
3542 21 : gimple_set_lhs (call, make_ssa_name (integer_type_node));
3543 21 : gimple_seq_add_stmt (&seq, call);
3544 :
3545 21 : tree prev_lhs = gimple_call_lhs (call);
3546 :
3547 21 : if (zero_ok && zero_val == ctz_val)
3548 : ;
3549 : /* Emit ctz (x) & 31 if ctz (0) is 32 but we need to return 0. */
3550 6 : else if (zero_ok && zero_val == 0 && ctz_val == input_bits)
3551 : {
3552 5 : g = gimple_build_assign (make_ssa_name (integer_type_node),
3553 : BIT_AND_EXPR, prev_lhs,
3554 : build_int_cst (integer_type_node,
3555 5 : input_bits - 1));
3556 5 : gimple_set_location (g, gimple_location (stmt));
3557 5 : gimple_seq_add_stmt (&seq, g);
3558 5 : prev_lhs = gimple_assign_lhs (g);
3559 : }
3560 : /* As fallback emit a conditional move. */
3561 : else
3562 : {
3563 10 : g = gimple_build_assign (make_ssa_name (boolean_type_node), EQ_EXPR,
3564 : res_ops[0], build_zero_cst (input_type));
3565 10 : gimple_set_location (g, gimple_location (stmt));
3566 10 : gimple_seq_add_stmt (&seq, g);
3567 10 : tree cond = gimple_assign_lhs (g);
3568 10 : g = gimple_build_assign (make_ssa_name (integer_type_node),
3569 : COND_EXPR, cond,
3570 10 : build_int_cst (integer_type_node, zero_val),
3571 : prev_lhs);
3572 10 : gimple_set_location (g, gimple_location (stmt));
3573 10 : gimple_seq_add_stmt (&seq, g);
3574 10 : prev_lhs = gimple_assign_lhs (g);
3575 : }
3576 :
3577 21 : if (fn == IFN_CLZ)
3578 : {
3579 11 : g = gimple_build_assign (make_ssa_name (integer_type_node),
3580 : MINUS_EXPR,
3581 : build_int_cst (integer_type_node,
3582 11 : input_bits - 1),
3583 : prev_lhs);
3584 11 : gimple_set_location (g, gimple_location (stmt));
3585 11 : gimple_seq_add_stmt (&seq, g);
3586 11 : prev_lhs = gimple_assign_lhs (g);
3587 : }
3588 :
3589 21 : g = gimple_build_assign (gimple_assign_lhs (stmt), NOP_EXPR, prev_lhs);
3590 21 : gimple_seq_add_stmt (&seq, g);
3591 21 : gsi_replace_with_seq (gsi, seq, true);
3592 21 : return true;
3593 21 : }
3594 :
3595 :
3596 : /* Determine whether applying the 2 permutations (mask1 then mask2)
3597 : gives back one of the input. */
3598 :
3599 : static int
3600 42 : is_combined_permutation_identity (tree mask1, tree mask2)
3601 : {
3602 42 : tree mask;
3603 42 : unsigned HOST_WIDE_INT nelts, i, j;
3604 42 : bool maybe_identity1 = true;
3605 42 : bool maybe_identity2 = true;
3606 :
3607 42 : gcc_checking_assert (TREE_CODE (mask1) == VECTOR_CST
3608 : && TREE_CODE (mask2) == VECTOR_CST);
3609 :
3610 : /* For VLA masks, check for the following pattern:
3611 : v1 = VEC_PERM_EXPR (v0, ..., mask1)
3612 : v2 = VEC_PERM_EXPR (v1, ..., mask2)
3613 : -->
3614 : v2 = v0
3615 : if mask1 == mask2 == {nelts - 1, nelts - 2, ...}. */
3616 :
3617 42 : if (operand_equal_p (mask1, mask2, 0)
3618 42 : && !VECTOR_CST_NELTS (mask1).is_constant ())
3619 : {
3620 : vec_perm_builder builder;
3621 : if (tree_to_vec_perm_builder (&builder, mask1))
3622 : {
3623 : poly_uint64 nelts = TYPE_VECTOR_SUBPARTS (TREE_TYPE (mask1));
3624 : vec_perm_indices sel (builder, 1, nelts);
3625 : if (sel.series_p (0, 1, nelts - 1, -1))
3626 : return 1;
3627 : }
3628 : }
3629 :
3630 42 : mask = fold_ternary (VEC_PERM_EXPR, TREE_TYPE (mask1), mask1, mask1, mask2);
3631 42 : if (mask == NULL_TREE || TREE_CODE (mask) != VECTOR_CST)
3632 : return 0;
3633 :
3634 42 : if (!VECTOR_CST_NELTS (mask).is_constant (&nelts))
3635 : return 0;
3636 72 : for (i = 0; i < nelts; i++)
3637 : {
3638 72 : tree val = VECTOR_CST_ELT (mask, i);
3639 72 : gcc_assert (TREE_CODE (val) == INTEGER_CST);
3640 72 : j = TREE_INT_CST_LOW (val) & (2 * nelts - 1);
3641 72 : if (j == i)
3642 : maybe_identity2 = false;
3643 55 : else if (j == i + nelts)
3644 : maybe_identity1 = false;
3645 : else
3646 : return 0;
3647 : }
3648 0 : return maybe_identity1 ? 1 : maybe_identity2 ? 2 : 0;
3649 : }
3650 :
3651 : /* Combine a shuffle with its arguments. Returns true if there were any
3652 : changes made. */
3653 :
3654 : static bool
3655 190530 : simplify_permutation (gimple_stmt_iterator *gsi)
3656 : {
3657 190530 : gimple *stmt = gsi_stmt (*gsi);
3658 190530 : gimple *def_stmt = NULL;
3659 190530 : tree op0, op1, op2, op3, arg0, arg1;
3660 190530 : enum tree_code code, code2 = ERROR_MARK;
3661 190530 : bool single_use_op0 = false;
3662 :
3663 190530 : gcc_checking_assert (gimple_assign_rhs_code (stmt) == VEC_PERM_EXPR);
3664 :
3665 190530 : op0 = gimple_assign_rhs1 (stmt);
3666 190530 : op1 = gimple_assign_rhs2 (stmt);
3667 190530 : op2 = gimple_assign_rhs3 (stmt);
3668 :
3669 190530 : if (TREE_CODE (op2) != VECTOR_CST)
3670 : return false;
3671 :
3672 187761 : if (TREE_CODE (op0) == VECTOR_CST)
3673 : {
3674 : code = VECTOR_CST;
3675 : arg0 = op0;
3676 : }
3677 185895 : else if (TREE_CODE (op0) == SSA_NAME)
3678 : {
3679 185895 : def_stmt = get_prop_source_stmt (op0, false, &single_use_op0);
3680 185895 : if (!def_stmt)
3681 : return false;
3682 177555 : code = gimple_assign_rhs_code (def_stmt);
3683 177555 : if (code == VIEW_CONVERT_EXPR)
3684 : {
3685 1621 : tree rhs = gimple_assign_rhs1 (def_stmt);
3686 1621 : tree name = TREE_OPERAND (rhs, 0);
3687 1621 : if (TREE_CODE (name) != SSA_NAME)
3688 : return false;
3689 1621 : if (!has_single_use (name))
3690 250 : single_use_op0 = false;
3691 : /* Here we update the def_stmt through this VIEW_CONVERT_EXPR,
3692 : but still keep the code to indicate it comes from
3693 : VIEW_CONVERT_EXPR. */
3694 1621 : def_stmt = SSA_NAME_DEF_STMT (name);
3695 1621 : if (!def_stmt || !is_gimple_assign (def_stmt))
3696 : return false;
3697 834 : if (gimple_assign_rhs_code (def_stmt) != CONSTRUCTOR)
3698 : return false;
3699 : }
3700 176200 : if (!can_propagate_from (def_stmt))
3701 : return false;
3702 24440 : arg0 = gimple_assign_rhs1 (def_stmt);
3703 : }
3704 : else
3705 : return false;
3706 :
3707 : /* Two consecutive shuffles. */
3708 24440 : if (code == VEC_PERM_EXPR)
3709 : {
3710 6657 : tree orig;
3711 6657 : int ident;
3712 :
3713 6657 : if (op0 != op1)
3714 : return false;
3715 42 : op3 = gimple_assign_rhs3 (def_stmt);
3716 42 : if (TREE_CODE (op3) != VECTOR_CST)
3717 : return false;
3718 42 : ident = is_combined_permutation_identity (op3, op2);
3719 42 : if (!ident)
3720 : return false;
3721 0 : orig = (ident == 1) ? gimple_assign_rhs1 (def_stmt)
3722 0 : : gimple_assign_rhs2 (def_stmt);
3723 0 : gimple_assign_set_rhs1 (stmt, unshare_expr (orig));
3724 0 : gimple_assign_set_rhs_code (stmt, TREE_CODE (orig));
3725 0 : gimple_set_num_ops (stmt, 2);
3726 0 : update_stmt (stmt);
3727 0 : remove_prop_source_from_use (op0);
3728 0 : return true;
3729 : }
3730 19649 : else if (code == CONSTRUCTOR
3731 19649 : || code == VECTOR_CST
3732 : || code == VIEW_CONVERT_EXPR)
3733 : {
3734 4658 : if (op0 != op1)
3735 : {
3736 4474 : if (TREE_CODE (op0) == SSA_NAME && !single_use_op0)
3737 : return false;
3738 :
3739 3819 : if (TREE_CODE (op1) == VECTOR_CST)
3740 : arg1 = op1;
3741 3186 : else if (TREE_CODE (op1) == SSA_NAME)
3742 : {
3743 3186 : gimple *def_stmt2 = get_prop_source_stmt (op1, true, NULL);
3744 3186 : if (!def_stmt2)
3745 : return false;
3746 1678 : code2 = gimple_assign_rhs_code (def_stmt2);
3747 1678 : if (code2 == VIEW_CONVERT_EXPR)
3748 : {
3749 4 : tree rhs = gimple_assign_rhs1 (def_stmt2);
3750 4 : tree name = TREE_OPERAND (rhs, 0);
3751 4 : if (TREE_CODE (name) != SSA_NAME)
3752 : return false;
3753 4 : if (!has_single_use (name))
3754 : return false;
3755 3 : def_stmt2 = SSA_NAME_DEF_STMT (name);
3756 3 : if (!def_stmt2 || !is_gimple_assign (def_stmt2))
3757 : return false;
3758 0 : if (gimple_assign_rhs_code (def_stmt2) != CONSTRUCTOR)
3759 : return false;
3760 : }
3761 1674 : else if (code2 != CONSTRUCTOR && code2 != VECTOR_CST)
3762 : return false;
3763 1485 : if (!can_propagate_from (def_stmt2))
3764 : return false;
3765 1485 : arg1 = gimple_assign_rhs1 (def_stmt2);
3766 : }
3767 : else
3768 : return false;
3769 : }
3770 : else
3771 : {
3772 : /* Already used twice in this statement. */
3773 184 : if (TREE_CODE (op0) == SSA_NAME && num_imm_uses (op0) > 2)
3774 : return false;
3775 : arg1 = arg0;
3776 : }
3777 :
3778 : /* If there are any VIEW_CONVERT_EXPRs found when finding permutation
3779 : operands source, check whether it's valid to transform and prepare
3780 : the required new operands. */
3781 2234 : if (code == VIEW_CONVERT_EXPR || code2 == VIEW_CONVERT_EXPR)
3782 : {
3783 : /* Figure out the target vector type to which operands should be
3784 : converted. If both are CONSTRUCTOR, the types should be the
3785 : same, otherwise, use the one of CONSTRUCTOR. */
3786 24 : tree tgt_type = NULL_TREE;
3787 24 : if (code == VIEW_CONVERT_EXPR)
3788 : {
3789 24 : gcc_assert (gimple_assign_rhs_code (def_stmt) == CONSTRUCTOR);
3790 24 : code = CONSTRUCTOR;
3791 24 : tgt_type = TREE_TYPE (arg0);
3792 : }
3793 24 : if (code2 == VIEW_CONVERT_EXPR)
3794 : {
3795 0 : tree arg1_type = TREE_TYPE (arg1);
3796 0 : if (tgt_type == NULL_TREE)
3797 : tgt_type = arg1_type;
3798 0 : else if (tgt_type != arg1_type)
3799 23 : return false;
3800 : }
3801 :
3802 24 : if (!VECTOR_TYPE_P (tgt_type))
3803 : return false;
3804 24 : tree op2_type = TREE_TYPE (op2);
3805 :
3806 : /* Figure out the shrunk factor. */
3807 24 : poly_uint64 tgt_units = TYPE_VECTOR_SUBPARTS (tgt_type);
3808 24 : poly_uint64 op2_units = TYPE_VECTOR_SUBPARTS (op2_type);
3809 24 : if (maybe_gt (tgt_units, op2_units))
3810 : return false;
3811 24 : unsigned int factor;
3812 47 : if (!constant_multiple_p (op2_units, tgt_units, &factor))
3813 : return false;
3814 :
3815 : /* Build the new permutation control vector as target vector. */
3816 24 : vec_perm_builder builder;
3817 24 : if (!tree_to_vec_perm_builder (&builder, op2))
3818 : return false;
3819 24 : vec_perm_indices indices (builder, 2, op2_units);
3820 24 : vec_perm_indices new_indices;
3821 24 : if (new_indices.new_shrunk_vector (indices, factor))
3822 : {
3823 1 : tree mask_type = tgt_type;
3824 1 : if (!VECTOR_INTEGER_TYPE_P (mask_type))
3825 : {
3826 0 : tree elem_type = TREE_TYPE (mask_type);
3827 0 : unsigned elem_size = TREE_INT_CST_LOW (TYPE_SIZE (elem_type));
3828 0 : tree int_type = build_nonstandard_integer_type (elem_size, 0);
3829 0 : mask_type = build_vector_type (int_type, tgt_units);
3830 : }
3831 1 : op2 = vec_perm_indices_to_tree (mask_type, new_indices);
3832 : }
3833 : else
3834 23 : return false;
3835 :
3836 : /* Convert the VECTOR_CST to the appropriate vector type. */
3837 1 : if (tgt_type != TREE_TYPE (arg0))
3838 0 : arg0 = fold_build1 (VIEW_CONVERT_EXPR, tgt_type, arg0);
3839 1 : else if (tgt_type != TREE_TYPE (arg1))
3840 0 : arg1 = fold_build1 (VIEW_CONVERT_EXPR, tgt_type, arg1);
3841 47 : }
3842 :
3843 : /* VIEW_CONVERT_EXPR should be updated to CONSTRUCTOR before. */
3844 2211 : gcc_assert (code == CONSTRUCTOR || code == VECTOR_CST);
3845 :
3846 : /* Shuffle of a constructor. */
3847 2211 : tree res_type
3848 2211 : = build_vector_type (TREE_TYPE (TREE_TYPE (arg0)),
3849 2211 : TYPE_VECTOR_SUBPARTS (TREE_TYPE (op2)));
3850 2211 : tree opt = fold_ternary (VEC_PERM_EXPR, res_type, arg0, arg1, op2);
3851 2211 : if (!opt
3852 280 : || (TREE_CODE (opt) != CONSTRUCTOR && TREE_CODE (opt) != VECTOR_CST))
3853 : return false;
3854 : /* Found VIEW_CONVERT_EXPR before, need one explicit conversion. */
3855 280 : if (res_type != TREE_TYPE (op0))
3856 : {
3857 1 : tree name = make_ssa_name (TREE_TYPE (opt));
3858 1 : gimple *ass_stmt = gimple_build_assign (name, opt);
3859 1 : gsi_insert_before (gsi, ass_stmt, GSI_SAME_STMT);
3860 1 : opt = build1 (VIEW_CONVERT_EXPR, TREE_TYPE (op0), name);
3861 : }
3862 280 : gimple_assign_set_rhs_from_tree (gsi, opt);
3863 280 : update_stmt (gsi_stmt (*gsi));
3864 280 : if (TREE_CODE (op0) == SSA_NAME)
3865 1 : remove_prop_source_from_use (op0);
3866 280 : if (op0 != op1 && TREE_CODE (op1) == SSA_NAME)
3867 0 : remove_prop_source_from_use (op1);
3868 280 : return true;
3869 : }
3870 :
3871 : return false;
3872 : }
3873 :
3874 : /* Get the BIT_FIELD_REF definition of VAL, if any, looking through
3875 : conversions with code CONV_CODE or update it if still ERROR_MARK.
3876 : Return NULL_TREE if no such matching def was found. */
3877 :
3878 : static tree
3879 426781 : get_bit_field_ref_def (tree val, enum tree_code &conv_code)
3880 : {
3881 426781 : if (TREE_CODE (val) != SSA_NAME)
3882 : return NULL_TREE ;
3883 399850 : gimple *def_stmt = get_prop_source_stmt (val, false, NULL);
3884 399850 : if (!def_stmt)
3885 : return NULL_TREE;
3886 320999 : enum tree_code code = gimple_assign_rhs_code (def_stmt);
3887 320999 : if (code == FLOAT_EXPR
3888 320999 : || code == FIX_TRUNC_EXPR
3889 : || CONVERT_EXPR_CODE_P (code))
3890 : {
3891 186937 : tree op1 = gimple_assign_rhs1 (def_stmt);
3892 186937 : if (conv_code == ERROR_MARK)
3893 89509 : conv_code = code;
3894 97428 : else if (conv_code != code)
3895 : return NULL_TREE;
3896 186912 : if (TREE_CODE (op1) != SSA_NAME)
3897 : return NULL_TREE;
3898 78799 : def_stmt = SSA_NAME_DEF_STMT (op1);
3899 78799 : if (! is_gimple_assign (def_stmt))
3900 : return NULL_TREE;
3901 63056 : code = gimple_assign_rhs_code (def_stmt);
3902 : }
3903 197118 : if (code != BIT_FIELD_REF)
3904 : return NULL_TREE;
3905 24393 : return gimple_assign_rhs1 (def_stmt);
3906 : }
3907 :
3908 : /* Recognize a VEC_PERM_EXPR. Returns true if there were any changes. */
3909 :
3910 : static bool
3911 164380 : simplify_vector_constructor (gimple_stmt_iterator *gsi)
3912 : {
3913 164380 : gimple *stmt = gsi_stmt (*gsi);
3914 164380 : tree op, orig[2], type;
3915 164380 : unsigned i;
3916 164380 : unsigned HOST_WIDE_INT nelts;
3917 164380 : unsigned HOST_WIDE_INT refnelts;
3918 164380 : enum tree_code conv_code;
3919 164380 : constructor_elt *elt;
3920 :
3921 164380 : op = gimple_assign_rhs1 (stmt);
3922 164380 : type = TREE_TYPE (op);
3923 164380 : gcc_checking_assert (TREE_CODE (op) == CONSTRUCTOR
3924 : && TREE_CODE (type) == VECTOR_TYPE);
3925 :
3926 164380 : if (!TYPE_VECTOR_SUBPARTS (type).is_constant (&nelts))
3927 : return false;
3928 :
3929 164380 : orig[0] = NULL;
3930 164380 : orig[1] = NULL;
3931 164380 : tree orig_elem_type[2] = {};
3932 164380 : conv_code = ERROR_MARK;
3933 164380 : bool maybe_ident = true;
3934 164380 : bool maybe_blend[2] = { true, true };
3935 164380 : tree one_constant = NULL_TREE;
3936 164380 : tree one_nonconstant = NULL_TREE;
3937 164380 : tree subelt;
3938 164380 : auto_vec<tree> constants;
3939 164380 : constants.safe_grow_cleared (nelts, true);
3940 164380 : auto_vec<std::pair<unsigned, unsigned>, 64> elts;
3941 164380 : unsigned int tsubelts = 0;
3942 460197 : FOR_EACH_VEC_SAFE_ELT (CONSTRUCTOR_ELTS (op), i, elt)
3943 : {
3944 426781 : tree ref, op1;
3945 426781 : unsigned int elem, src_elem_size;
3946 426781 : unsigned HOST_WIDE_INT nsubelts = 1;
3947 :
3948 426781 : if (i >= nelts)
3949 164380 : return false;
3950 :
3951 : /* Look for elements extracted and possibly converted from
3952 : another vector. */
3953 426781 : op1 = get_bit_field_ref_def (elt->value, conv_code);
3954 426781 : if (op1
3955 24393 : && TREE_CODE ((ref = TREE_OPERAND (op1, 0))) == SSA_NAME
3956 6028 : && VECTOR_TYPE_P (TREE_TYPE (ref))
3957 6017 : && (tree_nop_conversion_p (TREE_TYPE (op1),
3958 6017 : TREE_TYPE (TREE_TYPE (ref)))
3959 893 : || (VECTOR_TYPE_P (TREE_TYPE (op1))
3960 133 : && tree_nop_conversion_p (TREE_TYPE (TREE_TYPE (op1)),
3961 133 : TREE_TYPE (TREE_TYPE (ref)))
3962 133 : && TYPE_VECTOR_SUBPARTS (TREE_TYPE (op1))
3963 133 : .is_constant (&nsubelts)))
3964 5257 : && constant_multiple_p (bit_field_size (op1), nsubelts,
3965 : &src_elem_size)
3966 432038 : && constant_multiple_p (bit_field_offset (op1), src_elem_size, &elem)
3967 432038 : && TYPE_VECTOR_SUBPARTS (TREE_TYPE (ref)).is_constant (&refnelts))
3968 : {
3969 : unsigned int j;
3970 5626 : for (j = 0; j < 2; ++j)
3971 : {
3972 5597 : if (!orig[j])
3973 : {
3974 2401 : if (j == 0
3975 2602 : || useless_type_conversion_p (TREE_TYPE (orig[0]),
3976 201 : TREE_TYPE (ref)))
3977 : break;
3978 : }
3979 3196 : else if (ref == orig[j])
3980 : break;
3981 : }
3982 : /* Found a suitable vector element. */
3983 5257 : if (j < 2)
3984 : {
3985 5228 : orig[j] = ref;
3986 : /* Track what element type was actually extracted (which may
3987 : differ in signedness from the vector's element type due to
3988 : tree_nop_conversion_p). */
3989 5228 : if (!orig_elem_type[j])
3990 2397 : orig_elem_type[j] = TREE_TYPE (op1);
3991 5228 : if (elem != i || j != 0)
3992 2154 : maybe_ident = false;
3993 5228 : if (elem != i)
3994 2094 : maybe_blend[j] = false;
3995 10715 : for (unsigned int k = 0; k < nsubelts; ++k)
3996 5487 : elts.safe_push (std::make_pair (j, elem + k));
3997 5228 : tsubelts += nsubelts;
3998 5228 : continue;
3999 5228 : }
4000 : /* Else fallthru. */
4001 : }
4002 : /* Handle elements not extracted from a vector.
4003 : 1. constants by permuting with constant vector
4004 : 2. a unique non-constant element by permuting with a splat vector */
4005 421553 : if (orig[1]
4006 258303 : && orig[1] != error_mark_node)
4007 : return false;
4008 421524 : orig[1] = error_mark_node;
4009 421524 : if (VECTOR_TYPE_P (TREE_TYPE (elt->value))
4010 421524 : && !TYPE_VECTOR_SUBPARTS (TREE_TYPE (elt->value))
4011 5713 : .is_constant (&nsubelts))
4012 : return false;
4013 421524 : if (CONSTANT_CLASS_P (elt->value))
4014 : {
4015 26927 : if (one_nonconstant)
4016 : return false;
4017 18350 : if (!one_constant)
4018 8942 : one_constant = TREE_CODE (elt->value) == VECTOR_CST
4019 8942 : ? VECTOR_CST_ELT (elt->value, 0)
4020 : : elt->value;
4021 18350 : if (TREE_CODE (elt->value) == VECTOR_CST)
4022 : {
4023 687 : for (unsigned int k = 0; k < nsubelts; k++)
4024 507 : constants[tsubelts + k] = VECTOR_CST_ELT (elt->value, k);
4025 : }
4026 : else
4027 18170 : constants[tsubelts] = elt->value;
4028 : }
4029 : else
4030 : {
4031 394597 : if (one_constant)
4032 : return false;
4033 386120 : subelt = VECTOR_TYPE_P (TREE_TYPE (elt->value))
4034 386120 : ? ssa_uniform_vector_p (elt->value)
4035 : : elt->value;
4036 386120 : if (!subelt)
4037 : return false;
4038 380829 : if (!one_nonconstant)
4039 : one_nonconstant = subelt;
4040 231812 : else if (!operand_equal_p (one_nonconstant, subelt, 0))
4041 : return false;
4042 : }
4043 581507 : for (unsigned int k = 0; k < nsubelts; ++k)
4044 290918 : elts.safe_push (std::make_pair (1, tsubelts + k));
4045 290589 : tsubelts += nsubelts;
4046 290589 : maybe_ident = false;
4047 : }
4048 :
4049 66832 : if (elts.length () < nelts)
4050 : return false;
4051 :
4052 32165 : if (! orig[0]
4053 32165 : || ! VECTOR_TYPE_P (TREE_TYPE (orig[0])))
4054 : return false;
4055 1577 : refnelts = TYPE_VECTOR_SUBPARTS (TREE_TYPE (orig[0])).to_constant ();
4056 : /* We currently do not handle larger destination vectors. */
4057 1577 : if (refnelts < nelts)
4058 : return false;
4059 :
4060 : /* Determine the element type for the conversion source.
4061 : As orig_elem_type keeps track of the original type, check
4062 : if we need to perform a sign swap after permuting.
4063 : We need to be able to construct a vector type from the element
4064 : type which is not possible for e.g. BitInt or pointers
4065 : so pun with an integer type if needed. */
4066 1337 : tree perm_eltype = TREE_TYPE (TREE_TYPE (orig[0]));
4067 1337 : bool sign_change_p = false;
4068 1337 : if (conv_code != ERROR_MARK
4069 349 : && orig_elem_type[0]
4070 1686 : && TYPE_SIGN (orig_elem_type[0]) != TYPE_SIGN (perm_eltype))
4071 : {
4072 38 : perm_eltype = signed_or_unsigned_type_for
4073 38 : (TYPE_UNSIGNED (orig_elem_type[0]), perm_eltype);
4074 38 : sign_change_p = true;
4075 : }
4076 1337 : tree conv_src_type = build_vector_type (perm_eltype, nelts);
4077 :
4078 1337 : if (maybe_ident)
4079 : {
4080 : /* When there is no conversion, use the target type directly. */
4081 495 : if (conv_code == ERROR_MARK && nelts != refnelts)
4082 495 : conv_src_type = type;
4083 495 : if (conv_code != ERROR_MARK
4084 495 : && !supportable_convert_operation (conv_code, type, conv_src_type,
4085 : &conv_code))
4086 : {
4087 : /* Only few targets implement direct conversion patterns so try
4088 : some simple special cases via VEC_[UN]PACK[_FLOAT]_LO_EXPR. */
4089 93 : optab optab;
4090 93 : insn_code icode;
4091 93 : tree halfvectype, dblvectype;
4092 93 : enum tree_code unpack_op;
4093 :
4094 93 : if (!BYTES_BIG_ENDIAN)
4095 169 : unpack_op = (FLOAT_TYPE_P (TREE_TYPE (type))
4096 93 : ? VEC_UNPACK_FLOAT_LO_EXPR
4097 : : VEC_UNPACK_LO_EXPR);
4098 : else
4099 : unpack_op = (FLOAT_TYPE_P (TREE_TYPE (type))
4100 : ? VEC_UNPACK_FLOAT_HI_EXPR
4101 : : VEC_UNPACK_HI_EXPR);
4102 :
4103 : /* Conversions between DFP and FP have no special tree code
4104 : but we cannot handle those since all relevant vector conversion
4105 : optabs only have a single mode. */
4106 15 : if (CONVERT_EXPR_CODE_P (conv_code)
4107 78 : && FLOAT_TYPE_P (TREE_TYPE (type))
4108 97 : && (DECIMAL_FLOAT_TYPE_P (TREE_TYPE (type))
4109 2 : != DECIMAL_FLOAT_TYPE_P (TREE_TYPE (conv_src_type))))
4110 : return false;
4111 :
4112 15 : if (CONVERT_EXPR_CODE_P (conv_code)
4113 77 : && (2 * TYPE_PRECISION (TREE_TYPE (TREE_TYPE (orig[0])))
4114 77 : == TYPE_PRECISION (TREE_TYPE (type)))
4115 0 : && orig_elem_type[0]
4116 0 : && useless_type_conversion_p (orig_elem_type[0],
4117 0 : TREE_TYPE (TREE_TYPE (orig[0])))
4118 0 : && mode_for_vector (as_a <scalar_mode>
4119 0 : (TYPE_MODE (TREE_TYPE (TREE_TYPE (orig[0])))),
4120 0 : nelts * 2).exists ()
4121 0 : && (dblvectype
4122 0 : = build_vector_type (TREE_TYPE (TREE_TYPE (orig[0])),
4123 0 : nelts * 2))
4124 : /* Only use it for vector modes or for vector booleans
4125 : represented as scalar bitmasks. See PR95528. */
4126 0 : && (VECTOR_MODE_P (TYPE_MODE (dblvectype))
4127 0 : || VECTOR_BOOLEAN_TYPE_P (dblvectype))
4128 0 : && (optab = optab_for_tree_code (unpack_op,
4129 : dblvectype,
4130 : optab_default))
4131 0 : && ((icode = optab_handler (optab, TYPE_MODE (dblvectype)))
4132 : != CODE_FOR_nothing)
4133 92 : && (insn_data[icode].operand[0].mode == TYPE_MODE (type)))
4134 : {
4135 0 : gimple_seq stmts = NULL;
4136 0 : tree dbl;
4137 0 : if (refnelts == nelts)
4138 : {
4139 : /* ??? Paradoxical subregs don't exist, so insert into
4140 : the lower half of a wider zero vector. */
4141 0 : dbl = gimple_build (&stmts, BIT_INSERT_EXPR, dblvectype,
4142 : build_zero_cst (dblvectype), orig[0],
4143 0 : bitsize_zero_node);
4144 : }
4145 0 : else if (refnelts == 2 * nelts)
4146 : dbl = orig[0];
4147 : else
4148 0 : dbl = gimple_build (&stmts, BIT_FIELD_REF, dblvectype,
4149 0 : orig[0], TYPE_SIZE (dblvectype),
4150 0 : bitsize_zero_node);
4151 0 : gsi_insert_seq_before (gsi, stmts, GSI_SAME_STMT);
4152 0 : gimple_assign_set_rhs_with_ops (gsi, unpack_op, dbl);
4153 : }
4154 15 : else if (CONVERT_EXPR_CODE_P (conv_code)
4155 77 : && (TYPE_PRECISION (TREE_TYPE (TREE_TYPE (orig[0])))
4156 77 : == 2 * TYPE_PRECISION (TREE_TYPE (type)))
4157 1 : && orig_elem_type[0]
4158 1 : && useless_type_conversion_p (orig_elem_type[0],
4159 1 : TREE_TYPE (TREE_TYPE (orig[0])))
4160 1 : && mode_for_vector (as_a <scalar_mode>
4161 1 : (TYPE_MODE
4162 : (TREE_TYPE (TREE_TYPE (orig[0])))),
4163 2 : nelts / 2).exists ()
4164 1 : && (halfvectype
4165 1 : = build_vector_type (TREE_TYPE (TREE_TYPE (orig[0])),
4166 1 : nelts / 2))
4167 : /* Only use it for vector modes or for vector booleans
4168 : represented as scalar bitmasks. See PR95528. */
4169 1 : && (VECTOR_MODE_P (TYPE_MODE (halfvectype))
4170 0 : || VECTOR_BOOLEAN_TYPE_P (halfvectype))
4171 1 : && (optab = optab_for_tree_code (VEC_PACK_TRUNC_EXPR,
4172 : halfvectype,
4173 : optab_default))
4174 1 : && ((icode = optab_handler (optab, TYPE_MODE (halfvectype)))
4175 : != CODE_FOR_nothing)
4176 93 : && (insn_data[icode].operand[0].mode == TYPE_MODE (type)))
4177 : {
4178 0 : gimple_seq stmts = NULL;
4179 0 : tree low = gimple_build (&stmts, BIT_FIELD_REF, halfvectype,
4180 0 : orig[0], TYPE_SIZE (halfvectype),
4181 0 : bitsize_zero_node);
4182 0 : tree hig = gimple_build (&stmts, BIT_FIELD_REF, halfvectype,
4183 0 : orig[0], TYPE_SIZE (halfvectype),
4184 0 : TYPE_SIZE (halfvectype));
4185 0 : gsi_insert_seq_before (gsi, stmts, GSI_SAME_STMT);
4186 0 : gimple_assign_set_rhs_with_ops (gsi, VEC_PACK_TRUNC_EXPR,
4187 : low, hig);
4188 : }
4189 : else
4190 92 : return false;
4191 0 : update_stmt (gsi_stmt (*gsi));
4192 0 : return true;
4193 : }
4194 402 : if (nelts != refnelts)
4195 : {
4196 14 : gassign *lowpart
4197 14 : = gimple_build_assign (make_ssa_name (conv_src_type),
4198 : build3 (BIT_FIELD_REF, conv_src_type,
4199 14 : orig[0], TYPE_SIZE (conv_src_type),
4200 : bitsize_zero_node));
4201 14 : gsi_insert_before (gsi, lowpart, GSI_SAME_STMT);
4202 14 : orig[0] = gimple_assign_lhs (lowpart);
4203 : }
4204 388 : else if (sign_change_p)
4205 : {
4206 0 : gassign *conv
4207 0 : = gimple_build_assign (make_ssa_name (conv_src_type),
4208 : build1 (VIEW_CONVERT_EXPR, conv_src_type,
4209 : orig[0]));
4210 0 : gsi_insert_before (gsi, conv, GSI_SAME_STMT);
4211 0 : orig[0] = gimple_assign_lhs (conv);
4212 : }
4213 402 : if (conv_code == ERROR_MARK)
4214 : {
4215 385 : tree src_type = TREE_TYPE (orig[0]);
4216 385 : if (!useless_type_conversion_p (type, src_type))
4217 : {
4218 0 : gcc_assert (known_eq (TYPE_VECTOR_SUBPARTS (type),
4219 : TYPE_VECTOR_SUBPARTS (src_type))
4220 : && tree_nop_conversion_p (TREE_TYPE (type),
4221 : TREE_TYPE (src_type)));
4222 0 : tree rhs = build1 (VIEW_CONVERT_EXPR, type, orig[0]);
4223 0 : orig[0] = make_ssa_name (type);
4224 0 : gassign *assign = gimple_build_assign (orig[0], rhs);
4225 0 : gsi_insert_before (gsi, assign, GSI_SAME_STMT);
4226 : }
4227 385 : gimple_assign_set_rhs_from_tree (gsi, orig[0]);
4228 : }
4229 : else
4230 17 : gimple_assign_set_rhs_with_ops (gsi, conv_code, orig[0],
4231 : NULL_TREE, NULL_TREE);
4232 : }
4233 : else
4234 : {
4235 : /* If we combine a vector with a non-vector avoid cases where
4236 : we'll obviously end up with more GIMPLE stmts which is when
4237 : we'll later not fold this to a single insert into the vector
4238 : and we had a single extract originally. See PR92819. */
4239 842 : if (nelts == 2
4240 482 : && refnelts > 2
4241 102 : && orig[1] == error_mark_node
4242 31 : && !maybe_blend[0])
4243 256 : return false;
4244 817 : tree mask_type, perm_type;
4245 817 : perm_type = TREE_TYPE (orig[0]);
4246 817 : if (conv_code != ERROR_MARK
4247 817 : && !supportable_convert_operation (conv_code, type, conv_src_type,
4248 : &conv_code))
4249 : return false;
4250 :
4251 : /* Now that we know the number of elements of the source build the
4252 : permute vector.
4253 : ??? When the second vector has constant values we can shuffle
4254 : it and its source indexes to make the permutation supported.
4255 : For now it mimics a blend. */
4256 643 : vec_perm_builder sel (refnelts, refnelts, 1);
4257 643 : bool all_same_p = true;
4258 7302 : for (i = 0; i < elts.length (); ++i)
4259 : {
4260 3008 : sel.quick_push (elts[i].second + elts[i].first * refnelts);
4261 3008 : all_same_p &= known_eq (sel[i], sel[0]);
4262 : }
4263 : /* And fill the tail with "something". It's really don't care,
4264 : and ideally we'd allow VEC_PERM to have a smaller destination
4265 : vector. As a heuristic:
4266 :
4267 : (a) if what we have so far duplicates a single element, make the
4268 : tail do the same
4269 :
4270 : (b) otherwise preserve a uniform orig[0]. This facilitates
4271 : later pattern-matching of VEC_PERM_EXPR to a BIT_INSERT_EXPR. */
4272 1151 : for (; i < refnelts; ++i)
4273 1016 : sel.quick_push (all_same_p
4274 1524 : ? sel[0]
4275 140 : : (elts[0].second == 0 && elts[0].first == 0
4276 700 : ? 0 : refnelts) + i);
4277 815 : vec_perm_indices indices (sel, orig[1] ? 2 : 1, refnelts);
4278 643 : machine_mode vmode = TYPE_MODE (perm_type);
4279 643 : if ((cfun->curr_properties & PROP_gimple_lvec)
4280 643 : && !can_vec_perm_const_p (vmode, vmode, indices))
4281 : return false;
4282 586 : mask_type = build_vector_type (ssizetype, refnelts);
4283 586 : tree op2 = vec_perm_indices_to_tree (mask_type, indices);
4284 586 : bool converted_orig1 = false;
4285 586 : gimple_seq stmts = NULL;
4286 586 : if (!orig[1])
4287 153 : orig[1] = orig[0];
4288 433 : else if (orig[1] == error_mark_node
4289 314 : && one_nonconstant)
4290 : {
4291 : /* ??? We can see if we can safely convert to the original
4292 : element type. */
4293 90 : converted_orig1 = conv_code != ERROR_MARK;
4294 90 : tree target_type = converted_orig1 ? type : perm_type;
4295 90 : tree nonconstant_for_splat = one_nonconstant;
4296 : /* If there's a nop conversion between the target element type and
4297 : the nonconstant's type, convert it. */
4298 90 : if (!useless_type_conversion_p (TREE_TYPE (target_type),
4299 90 : TREE_TYPE (one_nonconstant)))
4300 0 : nonconstant_for_splat
4301 0 : = gimple_build (&stmts, NOP_EXPR, TREE_TYPE (target_type),
4302 : one_nonconstant);
4303 90 : orig[1] = gimple_build_vector_from_val (&stmts, UNKNOWN_LOCATION,
4304 : target_type,
4305 : nonconstant_for_splat);
4306 90 : }
4307 343 : else if (orig[1] == error_mark_node)
4308 : {
4309 : /* ??? See if we can convert the vector to the original type. */
4310 224 : converted_orig1 = conv_code != ERROR_MARK;
4311 224 : unsigned n = converted_orig1 ? nelts : refnelts;
4312 207 : tree target_type = converted_orig1 ? type : perm_type;
4313 224 : tree_vector_builder vec (target_type, n, 1);
4314 1748 : for (unsigned i = 0; i < n; ++i)
4315 2924 : if (i < nelts && constants[i])
4316 : {
4317 778 : tree constant = constants[i];
4318 : /* If there's a nop conversion, convert the constant. */
4319 778 : if (!useless_type_conversion_p (TREE_TYPE (target_type),
4320 778 : TREE_TYPE (constant)))
4321 2 : constant = fold_convert (TREE_TYPE (target_type), constant);
4322 778 : vec.quick_push (constant);
4323 : }
4324 : else
4325 : {
4326 : /* ??? Push a don't-care value. */
4327 746 : tree constant = one_constant;
4328 746 : if (!useless_type_conversion_p (TREE_TYPE (target_type),
4329 746 : TREE_TYPE (constant)))
4330 2 : constant = fold_convert (TREE_TYPE (target_type), constant);
4331 746 : vec.quick_push (constant);
4332 : }
4333 224 : orig[1] = vec.build ();
4334 224 : }
4335 467 : tree blend_op2 = NULL_TREE;
4336 467 : if (converted_orig1)
4337 : {
4338 : /* Make sure we can do a blend in the target type. */
4339 19 : vec_perm_builder sel (nelts, nelts, 1);
4340 87 : for (i = 0; i < elts.length (); ++i)
4341 68 : sel.quick_push (elts[i].first
4342 68 : ? elts[i].second + nelts : i);
4343 19 : vec_perm_indices indices (sel, 2, nelts);
4344 19 : machine_mode vmode = TYPE_MODE (type);
4345 19 : if ((cfun->curr_properties & PROP_gimple_lvec)
4346 19 : && !can_vec_perm_const_p (vmode, vmode, indices))
4347 0 : return false;
4348 19 : mask_type = build_vector_type (ssizetype, nelts);
4349 19 : blend_op2 = vec_perm_indices_to_tree (mask_type, indices);
4350 19 : }
4351 :
4352 : /* For a real orig[1] (no splat, constant etc.) we might need to
4353 : nop-convert it. Do so here. */
4354 586 : if (orig[1] && orig[1] != error_mark_node
4355 586 : && !useless_type_conversion_p (perm_type, TREE_TYPE (orig[1]))
4356 605 : && tree_nop_conversion_p (TREE_TYPE (perm_type),
4357 19 : TREE_TYPE (TREE_TYPE (orig[1]))))
4358 0 : orig[1] = gimple_build (&stmts, VIEW_CONVERT_EXPR, perm_type,
4359 : orig[1]);
4360 :
4361 586 : tree orig1_for_perm
4362 586 : = converted_orig1 ? build_zero_cst (perm_type) : orig[1];
4363 586 : tree res = gimple_build (&stmts, VEC_PERM_EXPR, perm_type,
4364 : orig[0], orig1_for_perm, op2);
4365 : /* If we're building a smaller vector, extract the element
4366 : with the proper type. */
4367 586 : if (nelts != refnelts)
4368 208 : res = gimple_build (&stmts, BIT_FIELD_REF,
4369 : conv_code != ERROR_MARK ? conv_src_type : type,
4370 : res,
4371 104 : TYPE_SIZE (conv_code != ERROR_MARK ? conv_src_type
4372 : : type),
4373 104 : bitsize_zero_node);
4374 : /* Otherwise, we can still have an intermediate sign change.
4375 : ??? In that case we have two subsequent conversions.
4376 : We should be able to merge them. */
4377 482 : else if (sign_change_p)
4378 14 : res = gimple_build (&stmts, VIEW_CONVERT_EXPR, conv_src_type, res);
4379 : /* Finally, apply the conversion. */
4380 586 : if (conv_code != ERROR_MARK)
4381 52 : res = gimple_build (&stmts, conv_code, type, res);
4382 534 : else if (!useless_type_conversion_p (type, TREE_TYPE (res)))
4383 : {
4384 3 : gcc_assert (known_eq (TYPE_VECTOR_SUBPARTS (type),
4385 : TYPE_VECTOR_SUBPARTS (perm_type))
4386 : && tree_nop_conversion_p (TREE_TYPE (type),
4387 : TREE_TYPE (perm_type)));
4388 3 : res = gimple_build (&stmts, VIEW_CONVERT_EXPR, type, res);
4389 : }
4390 : /* Blend in the actual constant. */
4391 586 : if (converted_orig1)
4392 19 : res = gimple_build (&stmts, VEC_PERM_EXPR, type,
4393 19 : res, orig[1], blend_op2);
4394 586 : gsi_insert_seq_before (gsi, stmts, GSI_SAME_STMT);
4395 586 : gimple_assign_set_rhs_with_ops (gsi, SSA_NAME, res);
4396 643 : }
4397 988 : update_stmt (gsi_stmt (*gsi));
4398 988 : return true;
4399 164380 : }
4400 :
4401 : /* Prepare a TARGET_MEM_REF ref so that it can be subsetted as
4402 : lvalue. This splits out an address computation stmt before *GSI
4403 : and returns a MEM_REF wrapping the address. */
4404 :
4405 : static tree
4406 1104 : prepare_target_mem_ref_lvalue (tree ref, gimple_stmt_iterator *gsi)
4407 : {
4408 1104 : if (TREE_CODE (TREE_OPERAND (ref, 0)) == ADDR_EXPR)
4409 218 : mark_addressable (TREE_OPERAND (TREE_OPERAND (ref, 0), 0));
4410 1104 : tree ptrtype = build_pointer_type (TREE_TYPE (ref));
4411 1104 : tree tem = make_ssa_name (ptrtype);
4412 1104 : gimple *new_stmt
4413 1104 : = gimple_build_assign (tem, build1 (ADDR_EXPR, TREE_TYPE (tem),
4414 : unshare_expr (ref)));
4415 1104 : gsi_insert_before (gsi, new_stmt, GSI_SAME_STMT);
4416 2208 : ref = build2_loc (EXPR_LOCATION (ref),
4417 1104 : MEM_REF, TREE_TYPE (ref), tem,
4418 1104 : build_int_cst (TREE_TYPE (TREE_OPERAND (ref, 1)), 0));
4419 1104 : return ref;
4420 : }
4421 :
4422 : /* Rewrite the vector load at *GSI to component-wise loads if the load
4423 : is only used in BIT_FIELD_REF extractions with eventual intermediate
4424 : widening. */
4425 :
4426 : static void
4427 293226 : optimize_vector_load (gimple_stmt_iterator *gsi)
4428 : {
4429 293226 : gimple *stmt = gsi_stmt (*gsi);
4430 293226 : tree lhs = gimple_assign_lhs (stmt);
4431 293226 : tree rhs = gimple_assign_rhs1 (stmt);
4432 293226 : tree vuse = gimple_vuse (stmt);
4433 :
4434 : /* Gather BIT_FIELD_REFs to rewrite, looking through
4435 : VEC_UNPACK_{LO,HI}_EXPR. */
4436 293226 : use_operand_p use_p;
4437 293226 : imm_use_iterator iter;
4438 293226 : bool rewrite = true;
4439 293226 : bool scalar_use = false;
4440 293226 : bool unpack_use = false;
4441 293226 : auto_vec<gimple *, 8> bf_stmts;
4442 293226 : auto_vec<tree, 8> worklist;
4443 293226 : worklist.quick_push (lhs);
4444 295168 : do
4445 : {
4446 295168 : tree def = worklist.pop ();
4447 295168 : unsigned HOST_WIDE_INT def_eltsize
4448 295168 : = TREE_INT_CST_LOW (TYPE_SIZE (TREE_TYPE (TREE_TYPE (def))));
4449 667815 : FOR_EACH_IMM_USE_FAST (use_p, iter, def)
4450 : {
4451 352071 : gimple *use_stmt = USE_STMT (use_p);
4452 352071 : if (is_gimple_debug (use_stmt))
4453 77479 : continue;
4454 350683 : tree use_lhs;
4455 350683 : if (!is_gimple_assign (use_stmt)
4456 : /* For alias reasons we move the use to the place of the
4457 : load. Avoid this when abnormals are involved. */
4458 350683 : || ((TREE_CODE ((use_lhs = gimple_assign_lhs (use_stmt)))
4459 : == SSA_NAME)
4460 239381 : && SSA_NAME_OCCURS_IN_ABNORMAL_PHI (use_lhs)))
4461 : {
4462 : rewrite = false;
4463 274592 : break;
4464 : }
4465 314713 : enum tree_code use_code = gimple_assign_rhs_code (use_stmt);
4466 314713 : tree use_rhs = gimple_assign_rhs1 (use_stmt);
4467 386939 : if (use_code == BIT_FIELD_REF
4468 72227 : && TREE_OPERAND (use_rhs, 0) == def
4469 : /* If its on the VEC_UNPACK_{HI,LO}_EXPR
4470 : def need to verify it is element aligned. */
4471 386940 : && (def == lhs
4472 153 : || (known_eq (bit_field_size (use_rhs), def_eltsize)
4473 153 : && constant_multiple_p (bit_field_offset (use_rhs),
4474 : def_eltsize)
4475 : /* We can simulate the VEC_UNPACK_{HI,LO}_EXPR
4476 : via a NOP_EXPR only for integral types.
4477 : ??? Support VEC_UNPACK_FLOAT_{HI,LO}_EXPR. */
4478 153 : && INTEGRAL_TYPE_P (TREE_TYPE (use_rhs)))))
4479 : {
4480 72226 : if (!VECTOR_TYPE_P (TREE_TYPE (gimple_assign_lhs (use_stmt))))
4481 70067 : scalar_use = true;
4482 72226 : bf_stmts.safe_push (use_stmt);
4483 72226 : continue;
4484 : }
4485 : /* Walk through one level of VEC_UNPACK_{LO,HI}_EXPR. */
4486 242487 : if (def == lhs
4487 240626 : && (use_code == VEC_UNPACK_HI_EXPR
4488 240626 : || use_code == VEC_UNPACK_LO_EXPR)
4489 3865 : && use_rhs == lhs)
4490 : {
4491 3865 : unpack_use = true;
4492 3865 : worklist.safe_push (gimple_assign_lhs (use_stmt));
4493 3865 : continue;
4494 : }
4495 : rewrite = false;
4496 : break;
4497 295168 : }
4498 295168 : if (!rewrite)
4499 : break;
4500 : }
4501 41152 : while (!worklist.is_empty ());
4502 :
4503 293226 : rewrite = rewrite && (scalar_use
4504 18634 : || unpack_use
4505 616 : || !can_implement_p (mov_optab,
4506 616 : TYPE_MODE (TREE_TYPE (lhs))));
4507 293226 : if (!rewrite)
4508 : {
4509 274774 : gsi_next (gsi);
4510 274774 : return;
4511 : }
4512 : /* We now have all ultimate uses of the load to rewrite in bf_stmts. */
4513 :
4514 : /* Prepare the original ref to be wrapped in adjusted BIT_FIELD_REFs.
4515 : For TARGET_MEM_REFs we have to separate the LEA from the reference. */
4516 18452 : tree load_rhs = rhs;
4517 18452 : if (TREE_CODE (load_rhs) == TARGET_MEM_REF)
4518 1103 : load_rhs = prepare_target_mem_ref_lvalue (load_rhs, gsi);
4519 :
4520 : /* Rewrite the BIT_FIELD_REFs to be actual loads, re-emitting them at
4521 : the place of the original load. */
4522 121414 : for (gimple *use_stmt : bf_stmts)
4523 : {
4524 66058 : tree bfr = gimple_assign_rhs1 (use_stmt);
4525 66058 : tree new_rhs = unshare_expr (load_rhs);
4526 66058 : if (TREE_OPERAND (bfr, 0) != lhs)
4527 : {
4528 : /* When the BIT_FIELD_REF is on the promoted vector we have to
4529 : adjust it and emit a conversion afterwards. */
4530 152 : gimple *def_stmt
4531 152 : = SSA_NAME_DEF_STMT (TREE_OPERAND (bfr, 0));
4532 152 : enum tree_code def_code
4533 152 : = gimple_assign_rhs_code (def_stmt);
4534 :
4535 : /* The adjusted BIT_FIELD_REF is of the promotion source
4536 : vector size and at half of the offset... */
4537 152 : new_rhs = fold_build3 (BIT_FIELD_REF,
4538 : TREE_TYPE (TREE_TYPE (lhs)),
4539 : new_rhs,
4540 : TYPE_SIZE (TREE_TYPE (TREE_TYPE (lhs))),
4541 : size_binop (EXACT_DIV_EXPR,
4542 : TREE_OPERAND (bfr, 2),
4543 : bitsize_int (2)));
4544 : /* ... and offsetted by half of the vector if VEC_UNPACK_HI_EXPR. */
4545 152 : if (def_code == (!BYTES_BIG_ENDIAN
4546 : ? VEC_UNPACK_HI_EXPR : VEC_UNPACK_LO_EXPR))
4547 76 : TREE_OPERAND (new_rhs, 2)
4548 152 : = size_binop (PLUS_EXPR, TREE_OPERAND (new_rhs, 2),
4549 : size_binop (EXACT_DIV_EXPR,
4550 : TYPE_SIZE (TREE_TYPE (lhs)),
4551 : bitsize_int (2)));
4552 152 : tree tem = make_ssa_name (TREE_TYPE (TREE_TYPE (lhs)));
4553 152 : gimple *new_stmt = gimple_build_assign (tem, new_rhs);
4554 152 : location_t loc = gimple_location (use_stmt);
4555 152 : gimple_set_location (new_stmt, loc);
4556 152 : gimple_set_vuse (new_stmt, vuse);
4557 152 : gsi_insert_before (gsi, new_stmt, GSI_SAME_STMT);
4558 : /* Perform scalar promotion. */
4559 152 : new_stmt = gimple_build_assign (gimple_assign_lhs (use_stmt),
4560 : NOP_EXPR, tem);
4561 152 : gimple_set_location (new_stmt, loc);
4562 152 : gsi_insert_before (gsi, new_stmt, GSI_SAME_STMT);
4563 : }
4564 : else
4565 : {
4566 : /* When the BIT_FIELD_REF is on the original load result
4567 : we can just wrap that. */
4568 65906 : tree new_rhs = fold_build3 (BIT_FIELD_REF, TREE_TYPE (bfr),
4569 : unshare_expr (load_rhs),
4570 : TREE_OPERAND (bfr, 1),
4571 : TREE_OPERAND (bfr, 2));
4572 65906 : gimple *new_stmt = gimple_build_assign (gimple_assign_lhs (use_stmt),
4573 : new_rhs);
4574 65906 : location_t loc = gimple_location (use_stmt);
4575 65906 : gimple_set_location (new_stmt, loc);
4576 65906 : gimple_set_vuse (new_stmt, vuse);
4577 65906 : gsi_insert_before (gsi, new_stmt, GSI_SAME_STMT);
4578 : }
4579 66058 : gimple_stmt_iterator gsi2 = gsi_for_stmt (use_stmt);
4580 66058 : unlink_stmt_vdef (use_stmt);
4581 66058 : gsi_remove (&gsi2, true);
4582 : }
4583 :
4584 : /* Finally get rid of the intermediate stmts. */
4585 18452 : gimple *use_stmt;
4586 37508 : FOR_EACH_IMM_USE_STMT (use_stmt, iter, lhs)
4587 : {
4588 604 : if (is_gimple_debug (use_stmt))
4589 : {
4590 542 : if (gimple_debug_bind_p (use_stmt))
4591 : {
4592 542 : gimple_debug_bind_reset_value (use_stmt);
4593 542 : update_stmt (use_stmt);
4594 : }
4595 542 : continue;
4596 : }
4597 62 : gimple_stmt_iterator gsi2 = gsi_for_stmt (use_stmt);
4598 62 : unlink_stmt_vdef (use_stmt);
4599 62 : release_defs (use_stmt);
4600 62 : gsi_remove (&gsi2, true);
4601 18452 : }
4602 : /* And the original load. */
4603 18452 : release_defs (stmt);
4604 18452 : gsi_remove (gsi, true);
4605 293226 : }
4606 :
4607 :
4608 : /* Primitive "lattice" function for gimple_simplify. */
4609 :
4610 : static tree
4611 1659491741 : fwprop_ssa_val (tree name)
4612 : {
4613 : /* First valueize NAME. */
4614 1659491741 : if (TREE_CODE (name) == SSA_NAME
4615 1659491741 : && SSA_NAME_VERSION (name) < lattice.length ())
4616 : {
4617 1658646299 : tree val = lattice[SSA_NAME_VERSION (name)];
4618 1658646299 : if (val)
4619 1659491741 : name = val;
4620 : }
4621 : /* We continue matching along SSA use-def edges for SSA names
4622 : that are not single-use. Currently there are no patterns
4623 : that would cause any issues with that. */
4624 1659491741 : return name;
4625 : }
4626 :
4627 : /* Search for opportunities to free half of the lanes in the following pattern:
4628 :
4629 : v_in = {e0, e1, e2, e3}
4630 : v_1 = VEC_PERM <v_in, v_in, {0, 2, 0, 2}>
4631 : // v_1 = {e0, e2, e0, e2}
4632 : v_2 = VEC_PERM <v_in, v_in, {1, 3, 1, 3}>
4633 : // v_2 = {e1, e3, e1, e3}
4634 :
4635 : v_x = v_1 + v_2
4636 : // v_x = {e0+e1, e2+e3, e0+e1, e2+e3}
4637 : v_y = v_1 - v_2
4638 : // v_y = {e0-e1, e2-e3, e0-e1, e2-e3}
4639 :
4640 : v_out = VEC_PERM <v_x, v_y, {0, 1, 6, 7}>
4641 : // v_out = {e0+e1, e2+e3, e0-e1, e2-e3}
4642 :
4643 : The last statement could be simplified to:
4644 : v_out' = VEC_PERM <v_x, v_y, {0, 1, 4, 5}>
4645 : // v_out' = {e0+e1, e2+e3, e0-e1, e2-e3}
4646 :
4647 : Characteristic properties:
4648 : - v_1 and v_2 are created from the same input vector v_in and introduce the
4649 : lane duplication (in the selection operand) that we can eliminate.
4650 : - v_x and v_y are results from lane-preserving operations that use v_1 and
4651 : v_2 as inputs.
4652 : - v_out is created by selecting from duplicated lanes. */
4653 :
4654 : static bool
4655 188381 : recognise_vec_perm_simplify_seq (gassign *stmt, vec_perm_simplify_seq *seq)
4656 : {
4657 188381 : unsigned HOST_WIDE_INT nelts;
4658 :
4659 188381 : gcc_checking_assert (stmt);
4660 188381 : gcc_checking_assert (gimple_assign_rhs_code (stmt) == VEC_PERM_EXPR);
4661 188381 : basic_block bb = gimple_bb (stmt);
4662 :
4663 : /* Decompose the final vec permute statement. */
4664 188381 : tree v_x = gimple_assign_rhs1 (stmt);
4665 188381 : tree v_y = gimple_assign_rhs2 (stmt);
4666 188381 : tree sel = gimple_assign_rhs3 (stmt);
4667 :
4668 188381 : if (TREE_CODE (sel) != VECTOR_CST
4669 185612 : || !VECTOR_CST_NELTS (sel).is_constant (&nelts)
4670 185612 : || TREE_CODE (v_x) != SSA_NAME
4671 183756 : || TREE_CODE (v_y) != SSA_NAME
4672 178487 : || !has_single_use (v_x)
4673 299540 : || !has_single_use (v_y))
4674 79026 : return false;
4675 :
4676 : /* Don't analyse sequences with many lanes. */
4677 109355 : if (nelts > 4)
4678 : return false;
4679 :
4680 : /* Lookup the definition of v_x and v_y. */
4681 105743 : gassign *v_x_stmt = dyn_cast<gassign *> (SSA_NAME_DEF_STMT (v_x));
4682 105743 : gassign *v_y_stmt = dyn_cast<gassign *> (SSA_NAME_DEF_STMT (v_y));
4683 105366 : if (!v_x_stmt || gimple_bb (v_x_stmt) != bb
4684 211109 : || !v_y_stmt || gimple_bb (v_y_stmt) != bb)
4685 : return false;
4686 :
4687 : /* Check the operations that define v_x and v_y. */
4688 105359 : if (TREE_CODE_CLASS (gimple_assign_rhs_code (v_x_stmt)) != tcc_binary
4689 107437 : || TREE_CODE_CLASS (gimple_assign_rhs_code (v_y_stmt)) != tcc_binary)
4690 : return false;
4691 :
4692 2078 : tree v_x_1 = gimple_assign_rhs1 (v_x_stmt);
4693 2078 : tree v_x_2 = gimple_assign_rhs2 (v_x_stmt);
4694 2078 : tree v_y_1 = gimple_assign_rhs1 (v_y_stmt);
4695 2078 : tree v_y_2 = gimple_assign_rhs2 (v_y_stmt);
4696 :
4697 2078 : if (v_x_stmt == v_y_stmt
4698 2078 : || TREE_CODE (v_x_1) != SSA_NAME
4699 2075 : || TREE_CODE (v_x_2) != SSA_NAME
4700 2051 : || num_imm_uses (v_x_1) != 2
4701 3971 : || num_imm_uses (v_x_2) != 2)
4702 : return false;
4703 :
4704 1857 : if (v_x_1 != v_y_1 || v_x_2 != v_y_2)
4705 : {
4706 : /* Allow operands of commutative operators to swap. */
4707 651 : if (commutative_tree_code (gimple_assign_rhs_code (v_x_stmt)))
4708 : {
4709 : /* Keep v_x_1 the first operand for non-commutative operators. */
4710 244 : std::swap (v_x_1, v_x_2);
4711 244 : if (v_x_1 != v_y_1 || v_x_2 != v_y_2)
4712 : return false;
4713 : }
4714 407 : else if (commutative_tree_code (gimple_assign_rhs_code (v_y_stmt)))
4715 : {
4716 407 : if (v_x_1 != v_y_2 || v_x_2 != v_y_1)
4717 : return false;
4718 : }
4719 : else
4720 : return false;
4721 : }
4722 1857 : gassign *v_1_stmt = dyn_cast<gassign *> (SSA_NAME_DEF_STMT (v_x_1));
4723 1857 : gassign *v_2_stmt = dyn_cast<gassign *> (SSA_NAME_DEF_STMT (v_x_2));
4724 1793 : if (!v_1_stmt || gimple_bb (v_1_stmt) != bb
4725 3650 : || !v_2_stmt || gimple_bb (v_2_stmt) != bb)
4726 : return false;
4727 :
4728 1789 : if (gimple_assign_rhs_code (v_1_stmt) != VEC_PERM_EXPR
4729 1911 : || gimple_assign_rhs_code (v_2_stmt) != VEC_PERM_EXPR)
4730 : return false;
4731 :
4732 : /* Decompose initial VEC_PERM_EXPRs. */
4733 108 : tree v_in = gimple_assign_rhs1 (v_1_stmt);
4734 108 : tree v_1_sel = gimple_assign_rhs3 (v_1_stmt);
4735 108 : tree v_2_sel = gimple_assign_rhs3 (v_2_stmt);
4736 108 : if (v_in != gimple_assign_rhs2 (v_1_stmt)
4737 103 : || v_in != gimple_assign_rhs1 (v_2_stmt)
4738 209 : || v_in != gimple_assign_rhs2 (v_2_stmt))
4739 : return false;
4740 :
4741 101 : unsigned HOST_WIDE_INT v_1_nelts, v_2_nelts;
4742 101 : if (TREE_CODE (v_1_sel) != VECTOR_CST
4743 101 : || !VECTOR_CST_NELTS (v_1_sel).is_constant (&v_1_nelts)
4744 101 : || TREE_CODE (v_2_sel) != VECTOR_CST
4745 202 : || !VECTOR_CST_NELTS (v_2_sel).is_constant (&v_2_nelts))
4746 0 : return false;
4747 :
4748 101 : if (nelts != v_1_nelts || nelts != v_2_nelts)
4749 : return false;
4750 :
4751 : /* Create the new selector. */
4752 101 : vec_perm_builder new_sel_perm (nelts, nelts, 1);
4753 101 : auto_vec<bool> lanes (nelts);
4754 101 : lanes.quick_grow_cleared (nelts);
4755 505 : for (unsigned int i = 0; i < nelts; i++)
4756 : {
4757 : /* Extract the i-th value from the selector. */
4758 404 : unsigned int sel_cst = TREE_INT_CST_LOW (VECTOR_CST_ELT (sel, i));
4759 404 : unsigned int lane = sel_cst % nelts;
4760 404 : unsigned int offs = sel_cst / nelts;
4761 :
4762 : /* Check what's in the lane. */
4763 404 : unsigned int e_1 = TREE_INT_CST_LOW (VECTOR_CST_ELT (v_1_sel, lane));
4764 404 : unsigned int e_2 = TREE_INT_CST_LOW (VECTOR_CST_ELT (v_2_sel, lane));
4765 :
4766 : /* Reuse previous lane (if any). */
4767 404 : unsigned int l = 0;
4768 687 : for (; l < lane; l++)
4769 : {
4770 481 : if ((TREE_INT_CST_LOW (VECTOR_CST_ELT (v_1_sel, l)) == e_1)
4771 481 : && (TREE_INT_CST_LOW (VECTOR_CST_ELT (v_2_sel, l)) == e_2))
4772 : break;
4773 : }
4774 :
4775 : /* Add to narrowed selector. */
4776 404 : new_sel_perm.quick_push (l + offs * nelts);
4777 :
4778 : /* Mark lane as used. */
4779 404 : lanes[l] = true;
4780 : }
4781 :
4782 : /* Count how many lanes are need. */
4783 : unsigned int cnt = 0;
4784 505 : for (unsigned int i = 0; i < nelts; i++)
4785 404 : cnt += lanes[i];
4786 :
4787 : /* If more than (nelts/2) lanes are needed, skip the sequence. */
4788 101 : if (cnt > nelts / 2)
4789 : return false;
4790 :
4791 : /* Check if the resulting permutation is cheap. */
4792 101 : vec_perm_indices new_indices (new_sel_perm, 2, nelts);
4793 101 : tree vectype = TREE_TYPE (gimple_assign_lhs (stmt));
4794 101 : machine_mode vmode = TYPE_MODE (vectype);
4795 101 : if (!can_vec_perm_const_p (vmode, vmode, new_indices, false))
4796 : return false;
4797 :
4798 101 : *seq = XNEW (struct _vec_perm_simplify_seq);
4799 101 : (*seq)->stmt = stmt;
4800 101 : (*seq)->v_1_stmt = v_1_stmt;
4801 101 : (*seq)->v_2_stmt = v_2_stmt;
4802 101 : (*seq)->v_x_stmt = v_x_stmt;
4803 101 : (*seq)->v_y_stmt = v_y_stmt;
4804 101 : (*seq)->nelts = nelts;
4805 101 : (*seq)->new_sel = vect_gen_perm_mask_checked (vectype, new_indices);
4806 :
4807 101 : if (dump_file)
4808 : {
4809 28 : fprintf (dump_file, "Found vec perm simplify sequence ending with:\n\t");
4810 28 : print_gimple_stmt (dump_file, stmt, 0);
4811 :
4812 28 : if (dump_flags & TDF_DETAILS)
4813 : {
4814 28 : fprintf (dump_file, "\tNarrowed vec_perm selector: ");
4815 28 : print_generic_expr (dump_file, (*seq)->new_sel);
4816 28 : fprintf (dump_file, "\n");
4817 : }
4818 : }
4819 :
4820 : return true;
4821 202 : }
4822 :
4823 : /* Reduce the lane consumption of a simplifiable vec perm sequence. */
4824 :
4825 : static void
4826 74 : narrow_vec_perm_simplify_seq (const vec_perm_simplify_seq &seq)
4827 : {
4828 74 : gassign *stmt = seq->stmt;
4829 74 : if (dump_file && (dump_flags & TDF_DETAILS))
4830 : {
4831 22 : fprintf (dump_file, "Updating VEC_PERM statement:\n");
4832 22 : fprintf (dump_file, "Old stmt: ");
4833 22 : print_gimple_stmt (dump_file, stmt, 0);
4834 : }
4835 :
4836 : /* Update the last VEC_PERM statement. */
4837 74 : gimple_assign_set_rhs3 (stmt, seq->new_sel);
4838 74 : update_stmt (stmt);
4839 :
4840 74 : if (dump_file && (dump_flags & TDF_DETAILS))
4841 : {
4842 22 : fprintf (dump_file, "New stmt: ");
4843 22 : print_gimple_stmt (dump_file, stmt, 0);
4844 : }
4845 74 : }
4846 :
4847 : /* Test if we can blend two simplifiable vec permute sequences.
4848 : NEED_SWAP will be set, if sequences must be swapped for blending. */
4849 :
4850 : static bool
4851 47 : can_blend_vec_perm_simplify_seqs_p (vec_perm_simplify_seq seq1,
4852 : vec_perm_simplify_seq seq2,
4853 : bool *need_swap)
4854 : {
4855 47 : unsigned int nelts = seq1->nelts;
4856 47 : basic_block bb = gimple_bb (seq1->stmt);
4857 :
4858 47 : gcc_assert (gimple_bb (seq2->stmt) == bb);
4859 :
4860 : /* BBs and number of elements must be equal. */
4861 47 : if (gimple_bb (seq2->stmt) != bb || seq2->nelts != nelts)
4862 : return false;
4863 :
4864 : /* We need vectors of the same type. */
4865 47 : if (TREE_TYPE (gimple_assign_lhs (seq1->stmt))
4866 47 : != TREE_TYPE (gimple_assign_lhs (seq2->stmt)))
4867 : return false;
4868 :
4869 : /* We require isomorphic operators. */
4870 41 : if (((gimple_assign_rhs_code (seq1->v_x_stmt)
4871 41 : != gimple_assign_rhs_code (seq2->v_x_stmt))
4872 41 : || (gimple_assign_rhs_code (seq1->v_y_stmt)
4873 41 : != gimple_assign_rhs_code (seq2->v_y_stmt))))
4874 : return false;
4875 :
4876 : /* We cannot have any dependencies between the sequences.
4877 :
4878 : For merging, we will reuse seq1->v_1_stmt and seq1->v_2_stmt.
4879 : seq1's v_in is defined before these statements, but we need
4880 : to check if seq2's v_in is defined before them as well.
4881 :
4882 : Further, we will reuse seq2->stmt. We need to ensure that
4883 : seq1->v_x_stmt and seq1->v_y_stmt are before it.
4884 :
4885 : Note, that we don't need to check the BBs here, because all
4886 : statements of both sequences have to be in the same BB. */
4887 :
4888 41 : tree seq2_v_in = gimple_assign_rhs1 (seq2->v_1_stmt);
4889 41 : if (TREE_CODE (seq2_v_in) != SSA_NAME)
4890 : return false;
4891 :
4892 41 : gassign *seq2_v_in_stmt = dyn_cast<gassign *> (SSA_NAME_DEF_STMT (seq2_v_in));
4893 41 : if (!seq2_v_in_stmt || gimple_bb (seq2_v_in_stmt) != bb
4894 41 : || (gimple_uid (seq2_v_in_stmt) > gimple_uid (seq1->v_1_stmt))
4895 37 : || (gimple_uid (seq1->v_x_stmt) > gimple_uid (seq2->stmt))
4896 37 : || (gimple_uid (seq1->v_y_stmt) > gimple_uid (seq2->stmt)))
4897 : {
4898 4 : tree seq1_v_in = gimple_assign_rhs1 (seq1->v_1_stmt);
4899 4 : if (TREE_CODE (seq1_v_in) != SSA_NAME)
4900 : return false;
4901 :
4902 4 : gassign *seq1_v_in_stmt
4903 4 : = dyn_cast<gassign *> (SSA_NAME_DEF_STMT (seq1_v_in));
4904 : /* Let's try to see if we succeed when swapping the sequences. */
4905 4 : if (!seq1_v_in_stmt || gimple_bb (seq1_v_in_stmt)
4906 0 : || (gimple_uid (seq1_v_in_stmt) > gimple_uid (seq2->v_1_stmt))
4907 0 : || (gimple_uid (seq2->v_x_stmt) > gimple_uid (seq1->stmt))
4908 0 : || (gimple_uid (seq2->v_y_stmt) > gimple_uid (seq1->stmt)))
4909 : return false;
4910 0 : *need_swap = true;
4911 : }
4912 : else
4913 37 : *need_swap = false;
4914 :
4915 37 : if (dump_file && (dump_flags & TDF_DETAILS))
4916 11 : fprintf (dump_file, "Found vec perm simplify sequence pair.\n");
4917 :
4918 : return true;
4919 : }
4920 :
4921 : /* Calculate the permutations for blending the two given vec permute
4922 : sequences. This may fail if the resulting permutation is not
4923 : supported. */
4924 :
4925 : static bool
4926 37 : calc_perm_vec_perm_simplify_seqs (vec_perm_simplify_seq seq1,
4927 : vec_perm_simplify_seq seq2,
4928 : vec_perm_indices *seq2_stmt_indices,
4929 : vec_perm_indices *seq1_v_1_stmt_indices,
4930 : vec_perm_indices *seq1_v_2_stmt_indices)
4931 : {
4932 37 : unsigned int i;
4933 37 : unsigned int nelts = seq1->nelts;
4934 37 : auto_vec<unsigned int> lane_assignment;
4935 37 : lane_assignment.create (nelts);
4936 :
4937 : /* Mark all lanes as free. */
4938 37 : lane_assignment.quick_grow_cleared (nelts);
4939 :
4940 : /* Allocate lanes for seq1. */
4941 185 : for (i = 0; i < nelts; i++)
4942 : {
4943 148 : unsigned int l = TREE_INT_CST_LOW (VECTOR_CST_ELT (seq1->new_sel, i));
4944 148 : l %= nelts;
4945 148 : lane_assignment[l] = 1;
4946 : }
4947 :
4948 : /* Allocate lanes for seq2 and calculate selector for seq2->stmt. */
4949 37 : vec_perm_builder seq2_stmt_sel_perm (nelts, nelts, 1);
4950 185 : for (i = 0; i < nelts; i++)
4951 : {
4952 148 : unsigned int sel = TREE_INT_CST_LOW (VECTOR_CST_ELT (seq2->new_sel, i));
4953 148 : unsigned int lane = sel % nelts;
4954 148 : unsigned int offs = sel / nelts;
4955 148 : unsigned int new_sel;
4956 :
4957 : /* Check if we already allocated the lane for seq2. */
4958 148 : unsigned int j = 0;
4959 263 : for (; j < i; j++)
4960 : {
4961 189 : unsigned int sel_old;
4962 189 : sel_old = TREE_INT_CST_LOW (VECTOR_CST_ELT (seq2->new_sel, j));
4963 189 : unsigned int lane_old = sel_old % nelts;
4964 189 : if (lane == lane_old)
4965 : {
4966 74 : new_sel = seq2_stmt_sel_perm[j].to_constant ();
4967 74 : new_sel = (new_sel % nelts) + offs * nelts;
4968 74 : break;
4969 : }
4970 : }
4971 :
4972 : /* If the lane is not allocated, we need to do that now. */
4973 148 : if (j == i)
4974 : {
4975 : unsigned int l_orig = lane;
4976 182 : while (lane_assignment[lane] != 0)
4977 : {
4978 108 : lane = (lane + 1) % nelts;
4979 :
4980 : /* This should not happen if both sequences utilize no more than
4981 : half of the lanes. Test anyway to guarantee termination. */
4982 108 : if (lane == l_orig)
4983 0 : return false;
4984 : }
4985 :
4986 : /* Allocate lane. */
4987 74 : lane_assignment[lane] = 2 + l_orig;
4988 74 : new_sel = lane + offs * nelts;
4989 : }
4990 :
4991 148 : seq2_stmt_sel_perm.quick_push (new_sel);
4992 : }
4993 :
4994 : /* Check if the resulting permutation is cheap. */
4995 37 : seq2_stmt_indices->new_vector (seq2_stmt_sel_perm, 2, nelts);
4996 37 : tree vectype = TREE_TYPE (gimple_assign_lhs (seq2->stmt));
4997 37 : machine_mode vmode = TYPE_MODE (vectype);
4998 37 : if (!can_vec_perm_const_p (vmode, vmode, *seq2_stmt_indices, false))
4999 : return false;
5000 :
5001 : /* Calculate selectors for seq1->v_1_stmt and seq1->v_2_stmt. */
5002 37 : vec_perm_builder seq1_v_1_stmt_sel_perm (nelts, nelts, 1);
5003 37 : vec_perm_builder seq1_v_2_stmt_sel_perm (nelts, nelts, 1);
5004 185 : for (i = 0; i < nelts; i++)
5005 : {
5006 148 : bool use_seq1 = lane_assignment[i] < 2;
5007 148 : unsigned int l1, l2;
5008 :
5009 148 : if (use_seq1)
5010 : {
5011 : /* Just reuse the selector indices. */
5012 74 : tree s1 = gimple_assign_rhs3 (seq1->v_1_stmt);
5013 74 : tree s2 = gimple_assign_rhs3 (seq1->v_2_stmt);
5014 74 : l1 = TREE_INT_CST_LOW (VECTOR_CST_ELT (s1, i));
5015 74 : l2 = TREE_INT_CST_LOW (VECTOR_CST_ELT (s2, i));
5016 : }
5017 : else
5018 : {
5019 : /* We moved the lanes for seq2, so we need to adjust for that. */
5020 74 : tree s1 = gimple_assign_rhs3 (seq2->v_1_stmt);
5021 74 : tree s2 = gimple_assign_rhs3 (seq2->v_2_stmt);
5022 74 : l1 = TREE_INT_CST_LOW (VECTOR_CST_ELT (s1, lane_assignment[i] - 2));
5023 74 : l2 = TREE_INT_CST_LOW (VECTOR_CST_ELT (s2, lane_assignment[i] - 2));
5024 : }
5025 :
5026 148 : l1 %= nelts;
5027 148 : l2 %= nelts;
5028 222 : seq1_v_1_stmt_sel_perm.quick_push (l1 + (use_seq1 ? 0 : nelts));
5029 148 : seq1_v_2_stmt_sel_perm.quick_push (l2 + (use_seq1 ? 0 : nelts));
5030 : }
5031 :
5032 37 : seq1_v_1_stmt_indices->new_vector (seq1_v_1_stmt_sel_perm, 2, nelts);
5033 37 : vectype = TREE_TYPE (gimple_assign_lhs (seq1->v_1_stmt));
5034 37 : vmode = TYPE_MODE (vectype);
5035 37 : if (!can_vec_perm_const_p (vmode, vmode, *seq1_v_1_stmt_indices, false))
5036 : return false;
5037 :
5038 37 : seq1_v_2_stmt_indices->new_vector (seq1_v_2_stmt_sel_perm, 2, nelts);
5039 37 : vectype = TREE_TYPE (gimple_assign_lhs (seq1->v_2_stmt));
5040 37 : vmode = TYPE_MODE (vectype);
5041 37 : if (!can_vec_perm_const_p (vmode, vmode, *seq1_v_2_stmt_indices, false))
5042 : return false;
5043 :
5044 : return true;
5045 74 : }
5046 :
5047 : /* Blend the two given simplifiable vec permute sequences using the
5048 : given permutations. */
5049 :
5050 : static void
5051 37 : blend_vec_perm_simplify_seqs (vec_perm_simplify_seq seq1,
5052 : vec_perm_simplify_seq seq2,
5053 : const vec_perm_indices &seq2_stmt_indices,
5054 : const vec_perm_indices &seq1_v_1_stmt_indices,
5055 : const vec_perm_indices &seq1_v_2_stmt_indices)
5056 : {
5057 : /* We don't need to adjust seq1->stmt because its lanes consumption
5058 : was already narrowed before entering this function. */
5059 :
5060 : /* Adjust seq2->stmt: copy RHS1/RHS2 from seq1->stmt and set new sel. */
5061 37 : if (dump_file && (dump_flags & TDF_DETAILS))
5062 : {
5063 11 : fprintf (dump_file, "Updating VEC_PERM statement:\n");
5064 11 : fprintf (dump_file, "Old stmt: ");
5065 11 : print_gimple_stmt (dump_file, seq2->stmt, 0);
5066 : }
5067 :
5068 37 : gimple_assign_set_rhs1 (seq2->stmt, gimple_assign_rhs1 (seq1->stmt));
5069 74 : gimple_assign_set_rhs2 (seq2->stmt, gimple_assign_rhs2 (seq1->stmt));
5070 37 : tree vectype = TREE_TYPE (gimple_assign_lhs (seq2->stmt));
5071 37 : tree sel = vect_gen_perm_mask_checked (vectype, seq2_stmt_indices);
5072 37 : gimple_assign_set_rhs3 (seq2->stmt, sel);
5073 37 : update_stmt (seq2->stmt);
5074 :
5075 37 : if (dump_file && (dump_flags & TDF_DETAILS))
5076 : {
5077 11 : fprintf (dump_file, "New stmt: ");
5078 11 : print_gimple_stmt (dump_file, seq2->stmt, 0);
5079 : }
5080 :
5081 : /* Adjust seq1->v_1_stmt: copy RHS2 from seq2->v_1_stmt and set new sel. */
5082 37 : if (dump_file && (dump_flags & TDF_DETAILS))
5083 : {
5084 11 : fprintf (dump_file, "Updating VEC_PERM statement:\n");
5085 11 : fprintf (dump_file, "Old stmt: ");
5086 11 : print_gimple_stmt (dump_file, seq1->v_1_stmt, 0);
5087 : }
5088 :
5089 37 : gimple_assign_set_rhs2 (seq1->v_1_stmt, gimple_assign_rhs1 (seq2->v_1_stmt));
5090 37 : vectype = TREE_TYPE (gimple_assign_lhs (seq1->v_1_stmt));
5091 37 : sel = vect_gen_perm_mask_checked (vectype, seq1_v_1_stmt_indices);
5092 37 : gimple_assign_set_rhs3 (seq1->v_1_stmt, sel);
5093 37 : update_stmt (seq1->v_1_stmt);
5094 :
5095 37 : if (dump_file && (dump_flags & TDF_DETAILS))
5096 : {
5097 11 : fprintf (dump_file, "New stmt: ");
5098 11 : print_gimple_stmt (dump_file, seq1->v_1_stmt, 0);
5099 : }
5100 :
5101 : /* Adjust seq1->v_2_stmt: copy RHS2 from seq2->v_2_stmt and set new sel. */
5102 37 : if (dump_file && (dump_flags & TDF_DETAILS))
5103 : {
5104 11 : fprintf (dump_file, "Updating VEC_PERM statement:\n");
5105 11 : fprintf (dump_file, "Old stmt: ");
5106 11 : print_gimple_stmt (dump_file, seq1->v_2_stmt, 0);
5107 : }
5108 :
5109 37 : gimple_assign_set_rhs2 (seq1->v_2_stmt, gimple_assign_rhs1 (seq2->v_2_stmt));
5110 37 : vectype = TREE_TYPE (gimple_assign_lhs (seq1->v_2_stmt));
5111 37 : sel = vect_gen_perm_mask_checked (vectype, seq1_v_2_stmt_indices);
5112 37 : gimple_assign_set_rhs3 (seq1->v_2_stmt, sel);
5113 37 : update_stmt (seq1->v_2_stmt);
5114 :
5115 37 : if (dump_file && (dump_flags & TDF_DETAILS))
5116 : {
5117 11 : fprintf (dump_file, "New stmt: ");
5118 11 : print_gimple_stmt (dump_file, seq1->v_2_stmt, 0);
5119 : }
5120 :
5121 : /* At this point, we have four unmodified seq2 stmts, which will be
5122 : eliminated by DCE. */
5123 :
5124 37 : if (dump_file)
5125 11 : fprintf (dump_file, "Vec perm simplify sequences have been blended.\n\n");
5126 37 : }
5127 :
5128 : /* Try to blend narrowed vec_perm_simplify_seqs pairwise.
5129 : The provided list will be empty after this call. */
5130 :
5131 : static void
5132 330805910 : process_vec_perm_simplify_seq_list (vec<vec_perm_simplify_seq> *l)
5133 : {
5134 330805910 : unsigned int i, j;
5135 330805910 : vec_perm_simplify_seq seq1, seq2;
5136 :
5137 330805910 : if (l->is_empty ())
5138 330805865 : return;
5139 :
5140 45 : if (dump_file && (dump_flags & TDF_DETAILS))
5141 13 : fprintf (dump_file, "\nProcessing %u vec perm simplify sequences.\n",
5142 : l->length ());
5143 :
5144 109 : FOR_EACH_VEC_ELT (*l, i, seq1)
5145 : {
5146 64 : if (i + 1 < l->length ())
5147 : {
5148 51 : FOR_EACH_VEC_ELT_FROM (*l, j, seq2, i + 1)
5149 : {
5150 47 : bool swap = false;
5151 47 : if (can_blend_vec_perm_simplify_seqs_p (seq1, seq2, &swap))
5152 : {
5153 37 : vec_perm_indices seq2_stmt_indices;
5154 37 : vec_perm_indices seq1_v_1_stmt_indices;
5155 37 : vec_perm_indices seq1_v_2_stmt_indices;
5156 111 : if (calc_perm_vec_perm_simplify_seqs (swap ? seq2 : seq1,
5157 : swap ? seq1 : seq2,
5158 : &seq2_stmt_indices,
5159 : &seq1_v_1_stmt_indices,
5160 : &seq1_v_2_stmt_indices))
5161 : {
5162 : /* Narrow lane usage. */
5163 37 : narrow_vec_perm_simplify_seq (seq1);
5164 37 : narrow_vec_perm_simplify_seq (seq2);
5165 :
5166 : /* Blend sequences. */
5167 37 : blend_vec_perm_simplify_seqs (swap ? seq2 : seq1,
5168 : swap ? seq1 : seq2,
5169 : seq2_stmt_indices,
5170 : seq1_v_1_stmt_indices,
5171 : seq1_v_2_stmt_indices);
5172 :
5173 : /* We can use unordered_remove as we break the loop. */
5174 37 : l->unordered_remove (j);
5175 37 : XDELETE (seq2);
5176 37 : break;
5177 : }
5178 37 : }
5179 : }
5180 : }
5181 :
5182 : /* We don't need to call l->remove for seq1. */
5183 64 : XDELETE (seq1);
5184 : }
5185 :
5186 45 : l->truncate (0);
5187 : }
5188 :
5189 : static void
5190 101 : append_vec_perm_simplify_seq_list (vec<vec_perm_simplify_seq> *l,
5191 : const vec_perm_simplify_seq &seq)
5192 : {
5193 : /* If no space on list left, then process the list. */
5194 101 : if (!l->space (1))
5195 0 : process_vec_perm_simplify_seq_list (l);
5196 :
5197 101 : l->quick_push (seq);
5198 101 : }
5199 :
5200 : /* Main entry point for the forward propagation and statement combine
5201 : optimizer. */
5202 :
5203 : namespace {
5204 :
5205 : const pass_data pass_data_forwprop =
5206 : {
5207 : GIMPLE_PASS, /* type */
5208 : "forwprop", /* name */
5209 : OPTGROUP_NONE, /* optinfo_flags */
5210 : TV_TREE_FORWPROP, /* tv_id */
5211 : ( PROP_cfg | PROP_ssa ), /* properties_required */
5212 : 0, /* properties_provided */
5213 : 0, /* properties_destroyed */
5214 : 0, /* todo_flags_start */
5215 : 0, /* todo_flags_finish */
5216 : };
5217 :
5218 : class pass_forwprop : public gimple_opt_pass
5219 : {
5220 : public:
5221 1461855 : pass_forwprop (gcc::context *ctxt)
5222 2923710 : : gimple_opt_pass (pass_data_forwprop, ctxt), last_p (false)
5223 : {}
5224 :
5225 : /* opt_pass methods: */
5226 1169484 : opt_pass * clone () final override { return new pass_forwprop (m_ctxt); }
5227 1754226 : void set_pass_param (unsigned int n, bool param) final override
5228 : {
5229 1754226 : switch (n)
5230 : {
5231 1169484 : case 0:
5232 1169484 : m_full_walk = param;
5233 1169484 : break;
5234 584742 : case 1:
5235 584742 : last_p = param;
5236 584742 : break;
5237 0 : default:
5238 0 : gcc_unreachable();
5239 : }
5240 1754226 : }
5241 5695907 : bool gate (function *) final override { return flag_tree_forwprop; }
5242 : unsigned int execute (function *) final override;
5243 :
5244 : private:
5245 : /* Determines whether the pass instance should set PROP_last_full_fold. */
5246 : bool last_p;
5247 :
5248 : /* True if the aggregate props are doing a full walk or not. */
5249 : bool m_full_walk = false;
5250 : }; // class pass_forwprop
5251 :
5252 : /* Attempt to make the BB block of __builtin_unreachable unreachable by changing
5253 : the incoming jumps. Return true if at least one jump was changed. */
5254 :
5255 : static bool
5256 1131 : optimize_unreachable (basic_block bb)
5257 : {
5258 1131 : gimple_stmt_iterator gsi;
5259 1131 : gimple *stmt;
5260 1131 : edge_iterator ei;
5261 1131 : edge e;
5262 1131 : bool ret;
5263 :
5264 1131 : ret = false;
5265 2331 : FOR_EACH_EDGE (e, ei, bb->preds)
5266 : {
5267 1200 : gsi = gsi_last_bb (e->src);
5268 1200 : if (gsi_end_p (gsi))
5269 328 : continue;
5270 :
5271 872 : stmt = gsi_stmt (gsi);
5272 872 : if (gcond *cond_stmt = dyn_cast <gcond *> (stmt))
5273 : {
5274 : /* If the condition is already true/false
5275 : ignore it. This can happen during copy prop of forwprop. */
5276 686 : if (gimple_cond_true_p (cond_stmt)
5277 678 : || gimple_cond_false_p (cond_stmt))
5278 8 : continue;
5279 670 : else if (e->flags & EDGE_TRUE_VALUE)
5280 582 : gimple_cond_make_false (cond_stmt);
5281 88 : else if (e->flags & EDGE_FALSE_VALUE)
5282 88 : gimple_cond_make_true (cond_stmt);
5283 : else
5284 0 : gcc_unreachable ();
5285 670 : update_stmt (cond_stmt);
5286 : }
5287 : else
5288 : {
5289 : /* Todo: handle other cases. e.g. switch. */
5290 194 : continue;
5291 : }
5292 :
5293 670 : ret = true;
5294 : }
5295 :
5296 1131 : return ret;
5297 : }
5298 :
5299 : unsigned int
5300 5693299 : pass_forwprop::execute (function *fun)
5301 : {
5302 5693299 : unsigned int todoflags = 0;
5303 : /* Handle a full walk only when expensive optimizations are on. */
5304 5693299 : bool full_walk = m_full_walk && flag_expensive_optimizations;
5305 :
5306 5693299 : cfg_changed = false;
5307 5693299 : if (last_p)
5308 1056998 : fun->curr_properties |= PROP_last_full_fold;
5309 :
5310 5693299 : calculate_dominance_info (CDI_DOMINATORS);
5311 :
5312 : /* Combine stmts with the stmts defining their operands. Do that
5313 : in an order that guarantees visiting SSA defs before SSA uses. */
5314 11386598 : lattice.create (num_ssa_names);
5315 11386598 : lattice.quick_grow_cleared (num_ssa_names);
5316 5693299 : int *postorder = XNEWVEC (int, n_basic_blocks_for_fn (fun));
5317 5693299 : int postorder_num = pre_and_rev_post_order_compute_fn (fun, NULL,
5318 : postorder, false);
5319 5693299 : int *bb_to_rpo = XNEWVEC (int, last_basic_block_for_fn (fun));
5320 51311617 : for (int i = 0; i < postorder_num; ++i)
5321 : {
5322 45618318 : bb_to_rpo[postorder[i]] = i;
5323 45618318 : edge_iterator ei;
5324 45618318 : edge e;
5325 109755600 : FOR_EACH_EDGE (e, ei, BASIC_BLOCK_FOR_FN (fun, postorder[i])->succs)
5326 64137282 : e->flags &= ~EDGE_EXECUTABLE;
5327 : }
5328 5693299 : single_succ_edge (BASIC_BLOCK_FOR_FN (fun, ENTRY_BLOCK))->flags
5329 5693299 : |= EDGE_EXECUTABLE;
5330 5693299 : auto_vec<gimple *, 4> to_fixup;
5331 5693299 : auto_vec<gimple *, 32> to_remove;
5332 5693299 : auto_vec<unsigned, 32> to_remove_defs;
5333 5693299 : auto_vec<std::pair<int, int>, 10> edges_to_remove;
5334 5693299 : auto_bitmap simple_dce_worklist;
5335 5693299 : auto_bitmap need_ab_cleanup;
5336 5693299 : to_purge = BITMAP_ALLOC (NULL);
5337 5693299 : auto_vec<vec_perm_simplify_seq, 8> vec_perm_simplify_seq_list;
5338 51311617 : for (int i = 0; i < postorder_num; ++i)
5339 : {
5340 45618318 : gimple_stmt_iterator gsi;
5341 45618318 : basic_block bb = BASIC_BLOCK_FOR_FN (fun, postorder[i]);
5342 45618318 : edge_iterator ei;
5343 45618318 : edge e;
5344 :
5345 : /* Skip processing not executable blocks. We could improve
5346 : single_use tracking by at least unlinking uses from unreachable
5347 : blocks but since blocks with uses are not processed in a
5348 : meaningful order this is probably not worth it. */
5349 45618318 : bool any = false;
5350 46766058 : FOR_EACH_EDGE (e, ei, bb->preds)
5351 : {
5352 46753092 : if ((e->flags & EDGE_EXECUTABLE)
5353 : /* We can handle backedges in natural loops correctly but
5354 : for irreducible regions we have to take all backedges
5355 : conservatively when we did not visit the source yet. */
5356 46753092 : || (bb_to_rpo[e->src->index] > i
5357 674007 : && !dominated_by_p (CDI_DOMINATORS, e->src, e->dest)))
5358 : {
5359 : any = true;
5360 : break;
5361 : }
5362 : }
5363 45618318 : if (!any)
5364 13603 : continue;
5365 :
5366 : /* Remove conditions that go directly to unreachable when this is the last forwprop. */
5367 45605352 : if (last_p
5368 9934221 : && !(flag_sanitize & SANITIZE_UNREACHABLE))
5369 : {
5370 9929246 : gimple_stmt_iterator gsi;
5371 9929246 : gsi = gsi_start_nondebug_after_labels_bb (bb);
5372 9929883 : if (!gsi_end_p (gsi)
5373 9077333 : && gimple_call_builtin_p (*gsi, BUILT_IN_UNREACHABLE)
5374 9930377 : && optimize_unreachable (bb))
5375 : {
5376 637 : cfg_changed = true;
5377 637 : continue;
5378 : }
5379 : }
5380 :
5381 : /* Record degenerate PHIs in the lattice. */
5382 61620861 : for (gphi_iterator si = gsi_start_phis (bb); !gsi_end_p (si);
5383 16016146 : gsi_next (&si))
5384 : {
5385 16016146 : gphi *phi = si.phi ();
5386 16016146 : tree res = gimple_phi_result (phi);
5387 32032292 : if (virtual_operand_p (res))
5388 7359050 : continue;
5389 :
5390 8657096 : tree first = NULL_TREE;
5391 8657096 : bool all_same = true;
5392 8657096 : edge_iterator ei;
5393 8657096 : edge e;
5394 17812132 : FOR_EACH_EDGE (e, ei, bb->preds)
5395 : {
5396 : /* Ignore not executable forward edges. */
5397 17599711 : if (!(e->flags & EDGE_EXECUTABLE))
5398 : {
5399 4059499 : if (bb_to_rpo[e->src->index] < i)
5400 5589 : continue;
5401 : /* Avoid equivalences from backedges - while we might
5402 : be able to make irreducible regions reducible and
5403 : thus turning a back into a forward edge we do not
5404 : want to deal with the intermediate SSA issues that
5405 : exposes. */
5406 : all_same = false;
5407 : }
5408 17594122 : tree use = PHI_ARG_DEF_FROM_EDGE (phi, e);
5409 17594122 : if (use == res)
5410 : /* The PHI result can also appear on a backedge, if so
5411 : we can ignore this case for the purpose of determining
5412 : the singular value. */
5413 : ;
5414 17581375 : else if (! first)
5415 : first = use;
5416 8924279 : else if (! operand_equal_p (first, use, 0))
5417 : {
5418 : all_same = false;
5419 : break;
5420 : }
5421 : }
5422 8657096 : if (all_same)
5423 : {
5424 207819 : if (may_propagate_copy (res, first))
5425 207264 : to_remove_defs.safe_push (SSA_NAME_VERSION (res));
5426 207819 : fwprop_set_lattice_val (res, first);
5427 : }
5428 : }
5429 :
5430 : /* Apply forward propagation to all stmts in the basic-block.
5431 : Note we update GSI within the loop as necessary. */
5432 45604715 : unsigned int uid = 1;
5433 444224632 : for (gsi = gsi_start_bb (bb); !gsi_end_p (gsi); )
5434 : {
5435 353015202 : gimple *stmt = gsi_stmt (gsi);
5436 353015202 : tree lhs, rhs;
5437 353015202 : enum tree_code code;
5438 :
5439 353015202 : gimple_set_uid (stmt, uid++);
5440 :
5441 353015202 : if (!is_gimple_assign (stmt))
5442 : {
5443 245900527 : process_vec_perm_simplify_seq_list (&vec_perm_simplify_seq_list);
5444 245900527 : gsi_next (&gsi);
5445 245900527 : continue;
5446 : }
5447 :
5448 107114675 : lhs = gimple_assign_lhs (stmt);
5449 107114675 : rhs = gimple_assign_rhs1 (stmt);
5450 107114675 : code = gimple_assign_rhs_code (stmt);
5451 :
5452 146415343 : if (TREE_CODE (lhs) != SSA_NAME
5453 107114675 : || has_zero_uses (lhs))
5454 : {
5455 39300668 : process_vec_perm_simplify_seq_list (&vec_perm_simplify_seq_list);
5456 39300668 : gsi_next (&gsi);
5457 39300668 : continue;
5458 : }
5459 :
5460 : /* If this statement sets an SSA_NAME to an address,
5461 : try to propagate the address into the uses of the SSA_NAME. */
5462 67814007 : if ((code == ADDR_EXPR
5463 : /* Handle pointer conversions on invariant addresses
5464 : as well, as this is valid gimple. */
5465 65489243 : || (CONVERT_EXPR_CODE_P (code)
5466 9010003 : && TREE_CODE (rhs) == ADDR_EXPR
5467 357130 : && POINTER_TYPE_P (TREE_TYPE (lhs))))
5468 67814231 : && TREE_CODE (TREE_OPERAND (rhs, 0)) != TARGET_MEM_REF)
5469 : {
5470 2324389 : tree base = get_base_address (TREE_OPERAND (rhs, 0));
5471 2324389 : if ((!base
5472 2324389 : || !DECL_P (base)
5473 134124 : || decl_address_invariant_p (base))
5474 2324389 : && !stmt_references_abnormal_ssa_name (stmt)
5475 4648762 : && forward_propagate_addr_expr (lhs, rhs, true))
5476 : {
5477 475635 : fwprop_invalidate_lattice (gimple_get_lhs (stmt));
5478 475635 : release_defs (stmt);
5479 475635 : gsi_remove (&gsi, true);
5480 : }
5481 : else
5482 1848754 : gsi_next (&gsi);
5483 : }
5484 65489618 : else if (code == POINTER_PLUS_EXPR)
5485 : {
5486 3660843 : tree off = gimple_assign_rhs2 (stmt);
5487 3660843 : if (TREE_CODE (off) == INTEGER_CST
5488 1125808 : && can_propagate_from (stmt)
5489 1125455 : && !simple_iv_increment_p (stmt)
5490 : /* ??? Better adjust the interface to that function
5491 : instead of building new trees here. */
5492 4490893 : && forward_propagate_addr_expr
5493 2490150 : (lhs,
5494 : build1_loc (gimple_location (stmt),
5495 830050 : ADDR_EXPR, TREE_TYPE (rhs),
5496 830050 : fold_build2 (MEM_REF,
5497 : TREE_TYPE (TREE_TYPE (rhs)),
5498 : rhs,
5499 : fold_convert (ptr_type_node,
5500 : off))), true))
5501 : {
5502 316833 : fwprop_invalidate_lattice (gimple_get_lhs (stmt));
5503 316833 : release_defs (stmt);
5504 316833 : gsi_remove (&gsi, true);
5505 : }
5506 3344010 : else if (is_gimple_min_invariant (rhs))
5507 : {
5508 : /* Make sure to fold &a[0] + off_1 here. */
5509 416717 : fold_stmt_inplace (&gsi);
5510 416717 : update_stmt (stmt);
5511 416717 : if (gimple_assign_rhs_code (stmt) == POINTER_PLUS_EXPR)
5512 416699 : gsi_next (&gsi);
5513 : }
5514 : else
5515 2927293 : gsi_next (&gsi);
5516 : }
5517 61828775 : else if (TREE_CODE (TREE_TYPE (lhs)) == COMPLEX_TYPE
5518 214034 : && gimple_assign_load_p (stmt)
5519 134964 : && !gimple_has_volatile_ops (stmt)
5520 40889 : && TREE_CODE (rhs) != TARGET_MEM_REF
5521 40860 : && TREE_CODE (rhs) != BIT_FIELD_REF
5522 61869631 : && !stmt_can_throw_internal (fun, stmt))
5523 : {
5524 : /* Rewrite loads used only in real/imagpart extractions to
5525 : component-wise loads. */
5526 40731 : use_operand_p use_p;
5527 40731 : imm_use_iterator iter;
5528 40731 : tree vuse = gimple_vuse (stmt);
5529 40731 : bool rewrite = true;
5530 86645 : FOR_EACH_IMM_USE_FAST (use_p, iter, lhs)
5531 : {
5532 43788 : gimple *use_stmt = USE_STMT (use_p);
5533 43788 : if (is_gimple_debug (use_stmt))
5534 1011 : continue;
5535 42777 : if (!is_gimple_assign (use_stmt)
5536 28092 : || (gimple_assign_rhs_code (use_stmt) != REALPART_EXPR
5537 25991 : && gimple_assign_rhs_code (use_stmt) != IMAGPART_EXPR)
5538 46949 : || TREE_OPERAND (gimple_assign_rhs1 (use_stmt), 0) != lhs)
5539 : {
5540 : rewrite = false;
5541 : break;
5542 : }
5543 40731 : }
5544 40731 : if (rewrite)
5545 : {
5546 2126 : gimple *use_stmt;
5547 8889 : FOR_EACH_IMM_USE_STMT (use_stmt, iter, lhs)
5548 : {
5549 4637 : if (is_gimple_debug (use_stmt))
5550 : {
5551 498 : if (gimple_debug_bind_p (use_stmt))
5552 : {
5553 498 : gimple_debug_bind_reset_value (use_stmt);
5554 498 : update_stmt (use_stmt);
5555 : }
5556 498 : continue;
5557 : }
5558 :
5559 8278 : tree new_rhs = build1 (gimple_assign_rhs_code (use_stmt),
5560 4139 : TREE_TYPE (TREE_TYPE (rhs)),
5561 : unshare_expr (rhs));
5562 4139 : gimple *new_stmt
5563 4139 : = gimple_build_assign (gimple_assign_lhs (use_stmt),
5564 : new_rhs);
5565 :
5566 4139 : location_t loc = gimple_location (use_stmt);
5567 4139 : gimple_set_location (new_stmt, loc);
5568 4139 : gimple_set_vuse (new_stmt, vuse);
5569 4139 : gimple_stmt_iterator gsi2 = gsi_for_stmt (use_stmt);
5570 4139 : unlink_stmt_vdef (use_stmt);
5571 4139 : gsi_remove (&gsi2, true);
5572 :
5573 4139 : gsi_insert_before (&gsi, new_stmt, GSI_SAME_STMT);
5574 2126 : }
5575 :
5576 2126 : release_defs (stmt);
5577 2126 : gsi_remove (&gsi, true);
5578 : }
5579 : else
5580 38605 : gsi_next (&gsi);
5581 : }
5582 61788044 : else if (TREE_CODE (TREE_TYPE (lhs)) == VECTOR_TYPE
5583 1804584 : && (TYPE_MODE (TREE_TYPE (lhs)) == BLKmode
5584 : /* After vector lowering rewrite all loads, but
5585 : initially do not since this conflicts with
5586 : vector CONSTRUCTOR to shuffle optimization. */
5587 1780532 : || (fun->curr_properties & PROP_gimple_lvec))
5588 925086 : && gimple_assign_load_p (stmt)
5589 307597 : && !gimple_has_volatile_ops (stmt)
5590 293728 : && !stmt_can_throw_internal (fun, stmt)
5591 62081772 : && (!VAR_P (rhs) || !DECL_HARD_REGISTER (rhs)))
5592 293226 : optimize_vector_load (&gsi);
5593 :
5594 61494818 : else if (code == COMPLEX_EXPR)
5595 : {
5596 : /* Rewrite stores of a single-use complex build expression
5597 : to component-wise stores. */
5598 37861 : use_operand_p use_p;
5599 37861 : gimple *use_stmt, *def1, *def2;
5600 37861 : tree rhs2;
5601 37861 : if (single_imm_use (lhs, &use_p, &use_stmt)
5602 35678 : && gimple_store_p (use_stmt)
5603 42088 : && !gimple_has_volatile_ops (use_stmt)
5604 3128 : && is_gimple_assign (use_stmt)
5605 3124 : && (TREE_CODE (TREE_TYPE (gimple_assign_lhs (use_stmt)))
5606 : == COMPLEX_TYPE)
5607 40980 : && (TREE_CODE (gimple_assign_lhs (use_stmt))
5608 : != TARGET_MEM_REF))
5609 : {
5610 3115 : tree use_lhs = gimple_assign_lhs (use_stmt);
5611 3115 : if (auto_var_p (use_lhs))
5612 601 : DECL_NOT_GIMPLE_REG_P (use_lhs) = 1;
5613 6230 : tree new_lhs = build1 (REALPART_EXPR,
5614 3115 : TREE_TYPE (TREE_TYPE (use_lhs)),
5615 : unshare_expr (use_lhs));
5616 3115 : gimple *new_stmt = gimple_build_assign (new_lhs, rhs);
5617 3115 : location_t loc = gimple_location (use_stmt);
5618 3115 : gimple_set_location (new_stmt, loc);
5619 6230 : gimple_set_vuse (new_stmt, gimple_vuse (use_stmt));
5620 3115 : gimple_set_vdef (new_stmt, make_ssa_name (gimple_vop (fun)));
5621 6230 : SSA_NAME_DEF_STMT (gimple_vdef (new_stmt)) = new_stmt;
5622 6230 : gimple_set_vuse (use_stmt, gimple_vdef (new_stmt));
5623 3115 : gimple_stmt_iterator gsi2 = gsi_for_stmt (use_stmt);
5624 3115 : gsi_insert_before (&gsi2, new_stmt, GSI_SAME_STMT);
5625 :
5626 6230 : new_lhs = build1 (IMAGPART_EXPR,
5627 3115 : TREE_TYPE (TREE_TYPE (use_lhs)),
5628 : unshare_expr (use_lhs));
5629 3115 : gimple_assign_set_lhs (use_stmt, new_lhs);
5630 3115 : gimple_assign_set_rhs1 (use_stmt, gimple_assign_rhs2 (stmt));
5631 3115 : update_stmt (use_stmt);
5632 :
5633 3115 : release_defs (stmt);
5634 3115 : gsi_remove (&gsi, true);
5635 : }
5636 : /* Rewrite a component-wise load of a complex to a complex
5637 : load if the components are not used separately. */
5638 34746 : else if (TREE_CODE (rhs) == SSA_NAME
5639 34305 : && has_single_use (rhs)
5640 30769 : && ((rhs2 = gimple_assign_rhs2 (stmt)), true)
5641 30769 : && TREE_CODE (rhs2) == SSA_NAME
5642 28958 : && has_single_use (rhs2)
5643 28538 : && (def1 = SSA_NAME_DEF_STMT (rhs),
5644 28538 : gimple_assign_load_p (def1))
5645 1088 : && (def2 = SSA_NAME_DEF_STMT (rhs2),
5646 1088 : gimple_assign_load_p (def2))
5647 1588 : && (gimple_vuse (def1) == gimple_vuse (def2))
5648 791 : && !gimple_has_volatile_ops (def1)
5649 791 : && !gimple_has_volatile_ops (def2)
5650 791 : && !stmt_can_throw_internal (fun, def1)
5651 791 : && !stmt_can_throw_internal (fun, def2)
5652 791 : && gimple_assign_rhs_code (def1) == REALPART_EXPR
5653 545 : && gimple_assign_rhs_code (def2) == IMAGPART_EXPR
5654 35291 : && operand_equal_p (TREE_OPERAND (gimple_assign_rhs1
5655 : (def1), 0),
5656 545 : TREE_OPERAND (gimple_assign_rhs1
5657 : (def2), 0)))
5658 : {
5659 545 : tree cl = TREE_OPERAND (gimple_assign_rhs1 (def1), 0);
5660 545 : gimple_assign_set_rhs_from_tree (&gsi, unshare_expr (cl));
5661 545 : gcc_assert (gsi_stmt (gsi) == stmt);
5662 1090 : gimple_set_vuse (stmt, gimple_vuse (def1));
5663 545 : gimple_set_modified (stmt, true);
5664 545 : gimple_stmt_iterator gsi2 = gsi_for_stmt (def1);
5665 545 : gsi_remove (&gsi, false);
5666 545 : gsi_insert_after (&gsi2, stmt, GSI_SAME_STMT);
5667 : }
5668 : else
5669 34201 : gsi_next (&gsi);
5670 : }
5671 61456957 : else if (code == CONSTRUCTOR
5672 166586 : && VECTOR_TYPE_P (TREE_TYPE (rhs))
5673 166586 : && TYPE_MODE (TREE_TYPE (rhs)) == BLKmode
5674 4492 : && CONSTRUCTOR_NELTS (rhs) > 0
5675 61461449 : && (!VECTOR_TYPE_P (TREE_TYPE (CONSTRUCTOR_ELT (rhs, 0)->value))
5676 2049 : || (TYPE_MODE (TREE_TYPE (CONSTRUCTOR_ELT (rhs, 0)->value))
5677 : != BLKmode)))
5678 : {
5679 : /* Rewrite stores of a single-use vector constructors
5680 : to component-wise stores if the mode isn't supported. */
5681 4135 : use_operand_p use_p;
5682 4135 : gimple *use_stmt;
5683 4135 : if (single_imm_use (lhs, &use_p, &use_stmt)
5684 3692 : && gimple_store_p (use_stmt)
5685 3024 : && !gimple_has_volatile_ops (use_stmt)
5686 1506 : && !stmt_can_throw_internal (fun, use_stmt)
5687 5634 : && is_gimple_assign (use_stmt))
5688 : {
5689 1499 : tree elt_t = TREE_TYPE (CONSTRUCTOR_ELT (rhs, 0)->value);
5690 1499 : unsigned HOST_WIDE_INT elt_w
5691 1499 : = tree_to_uhwi (TYPE_SIZE (elt_t));
5692 1499 : unsigned HOST_WIDE_INT n
5693 1499 : = tree_to_uhwi (TYPE_SIZE (TREE_TYPE (rhs)));
5694 1499 : tree use_lhs = gimple_assign_lhs (use_stmt);
5695 1499 : if (auto_var_p (use_lhs))
5696 567 : DECL_NOT_GIMPLE_REG_P (use_lhs) = 1;
5697 932 : else if (TREE_CODE (use_lhs) == TARGET_MEM_REF)
5698 : {
5699 1 : gimple_stmt_iterator gsi2 = gsi_for_stmt (use_stmt);
5700 1 : use_lhs = prepare_target_mem_ref_lvalue (use_lhs, &gsi2);
5701 : }
5702 33388 : for (unsigned HOST_WIDE_INT bi = 0; bi < n; bi += elt_w)
5703 : {
5704 31889 : unsigned HOST_WIDE_INT ci = bi / elt_w;
5705 31889 : tree new_rhs;
5706 31889 : if (ci < CONSTRUCTOR_NELTS (rhs))
5707 31271 : new_rhs = CONSTRUCTOR_ELT (rhs, ci)->value;
5708 : else
5709 618 : new_rhs = build_zero_cst (elt_t);
5710 31889 : tree new_lhs = build3 (BIT_FIELD_REF,
5711 : elt_t,
5712 : unshare_expr (use_lhs),
5713 31889 : bitsize_int (elt_w),
5714 31889 : bitsize_int (bi));
5715 31889 : gimple *new_stmt = gimple_build_assign (new_lhs, new_rhs);
5716 31889 : location_t loc = gimple_location (use_stmt);
5717 31889 : gimple_set_location (new_stmt, loc);
5718 63778 : gimple_set_vuse (new_stmt, gimple_vuse (use_stmt));
5719 31889 : gimple_set_vdef (new_stmt,
5720 : make_ssa_name (gimple_vop (fun)));
5721 63778 : SSA_NAME_DEF_STMT (gimple_vdef (new_stmt)) = new_stmt;
5722 63778 : gimple_set_vuse (use_stmt, gimple_vdef (new_stmt));
5723 31889 : gimple_stmt_iterator gsi2 = gsi_for_stmt (use_stmt);
5724 31889 : gsi_insert_before (&gsi2, new_stmt, GSI_SAME_STMT);
5725 : }
5726 1499 : gimple_stmt_iterator gsi2 = gsi_for_stmt (use_stmt);
5727 1499 : unlink_stmt_vdef (use_stmt);
5728 1499 : release_defs (use_stmt);
5729 1499 : gsi_remove (&gsi2, true);
5730 1499 : release_defs (stmt);
5731 1499 : gsi_remove (&gsi, true);
5732 : }
5733 : else
5734 2636 : gsi_next (&gsi);
5735 : }
5736 61452822 : else if (code == VEC_PERM_EXPR)
5737 : {
5738 : /* Find vectorized sequences, where we can reduce the lane
5739 : utilization. The narrowing will be donw later and only
5740 : if we find a pair of sequences that can be blended. */
5741 188381 : gassign *assign = dyn_cast <gassign *> (stmt);
5742 188381 : vec_perm_simplify_seq seq;
5743 188381 : if (recognise_vec_perm_simplify_seq (assign, &seq))
5744 101 : append_vec_perm_simplify_seq_list (&vec_perm_simplify_seq_list,
5745 : seq);
5746 :
5747 188381 : gsi_next (&gsi);
5748 : }
5749 : else
5750 61264441 : gsi_next (&gsi);
5751 : }
5752 :
5753 45604715 : process_vec_perm_simplify_seq_list (&vec_perm_simplify_seq_list);
5754 :
5755 : /* Combine stmts with the stmts defining their operands.
5756 : Note we update GSI within the loop as necessary. */
5757 443864940 : for (gsi = gsi_start_bb (bb); !gsi_end_p (gsi); gsi_next (&gsi))
5758 : {
5759 352655510 : gimple *stmt = gsi_stmt (gsi);
5760 :
5761 : /* Mark stmt as potentially needing revisiting. */
5762 352655510 : gimple_set_plf (stmt, GF_PLF_1, false);
5763 :
5764 352655510 : bool can_make_abnormal_goto = (is_gimple_call (stmt)
5765 352655510 : && stmt_can_make_abnormal_goto (stmt));
5766 :
5767 : /* Substitute from our lattice. We need to do so only once. */
5768 352655510 : bool substituted_p = false;
5769 352655510 : use_operand_p usep;
5770 352655510 : ssa_op_iter iter;
5771 517911638 : FOR_EACH_SSA_USE_OPERAND (usep, stmt, iter, SSA_OP_USE)
5772 : {
5773 165256128 : tree use = USE_FROM_PTR (usep);
5774 165256128 : tree val = fwprop_ssa_val (use);
5775 165256128 : if (val && val != use)
5776 : {
5777 1849228 : if (!is_gimple_debug (stmt))
5778 1542182 : bitmap_set_bit (simple_dce_worklist, SSA_NAME_VERSION (use));
5779 1849228 : if (may_propagate_copy (use, val))
5780 : {
5781 1846000 : propagate_value (usep, val);
5782 1846000 : substituted_p = true;
5783 : }
5784 : }
5785 : }
5786 352655510 : if (substituted_p)
5787 1791734 : update_stmt (stmt);
5788 1791734 : if (substituted_p
5789 1791734 : && is_gimple_assign (stmt)
5790 1076470 : && gimple_assign_rhs_code (stmt) == ADDR_EXPR)
5791 20012 : recompute_tree_invariant_for_addr_expr (gimple_assign_rhs1 (stmt));
5792 352655510 : if (substituted_p
5793 352655510 : && can_make_abnormal_goto
5794 352655510 : && !stmt_can_make_abnormal_goto (stmt))
5795 3 : bitmap_set_bit (need_ab_cleanup, bb->index);
5796 :
5797 355482704 : bool changed;
5798 710965408 : do
5799 : {
5800 355482704 : gimple *orig_stmt = stmt = gsi_stmt (gsi);
5801 355482704 : bool was_call = is_gimple_call (stmt);
5802 355482704 : bool was_noreturn = (was_call
5803 355482704 : && gimple_call_noreturn_p (stmt));
5804 355482704 : changed = false;
5805 :
5806 355482704 : auto_vec<tree, 8> uses;
5807 523752888 : FOR_EACH_SSA_USE_OPERAND (usep, stmt, iter, SSA_OP_USE)
5808 168270184 : if (uses.space (1))
5809 167879317 : uses.quick_push (USE_FROM_PTR (usep));
5810 :
5811 355482704 : if (fold_stmt (&gsi, fwprop_ssa_val, simple_dce_worklist))
5812 : {
5813 2506212 : changed = true;
5814 : /* There is no updating of the address
5815 : taken after the last forwprop so update
5816 : the addresses when a folding happened to a call.
5817 : The va_* builtins can remove taking of the address so
5818 : can the sincos->cexpi transformation. See PR 39643 and PR 20983. */
5819 2506212 : if (was_call && last_p)
5820 2506212 : todoflags |= TODO_update_address_taken;
5821 2506212 : stmt = gsi_stmt (gsi);
5822 : /* Cleanup the CFG if we simplified a condition to
5823 : true or false. */
5824 2506212 : if (gcond *cond = dyn_cast <gcond *> (stmt))
5825 1001117 : if (gimple_cond_true_p (cond)
5826 1001117 : || gimple_cond_false_p (cond))
5827 13956 : cfg_changed = true;
5828 : /* Queue old uses for simple DCE if not debug statement. */
5829 2506212 : if (!is_gimple_debug (stmt))
5830 10599242 : for (tree use : uses)
5831 3101984 : if (TREE_CODE (use) == SSA_NAME
5832 3101984 : && !SSA_NAME_IS_DEFAULT_DEF (use))
5833 2904271 : bitmap_set_bit (simple_dce_worklist,
5834 2904271 : SSA_NAME_VERSION (use));
5835 2506212 : update_stmt (stmt);
5836 : }
5837 :
5838 355482704 : switch (gimple_code (stmt))
5839 : {
5840 108094092 : case GIMPLE_ASSIGN:
5841 108094092 : {
5842 108094092 : tree rhs1 = gimple_assign_rhs1 (stmt);
5843 108094092 : enum tree_code code = gimple_assign_rhs_code (stmt);
5844 108094092 : if (gimple_clobber_p (stmt))
5845 7400238 : do_simple_agr_dse (as_a<gassign*>(stmt), full_walk);
5846 100693854 : else if (gimple_store_p (stmt))
5847 : {
5848 31277562 : optimize_aggr_zeroprop (stmt, full_walk);
5849 31277562 : if (gimple_assign_load_p (stmt))
5850 3886465 : optimize_agr_copyprop (stmt);
5851 : }
5852 69416292 : else if (TREE_CODE_CLASS (code) == tcc_comparison)
5853 2662715 : changed |= forward_propagate_into_comparison (&gsi);
5854 66753577 : else if ((code == PLUS_EXPR
5855 66753577 : || code == BIT_IOR_EXPR
5856 56579243 : || code == BIT_XOR_EXPR)
5857 66891463 : && simplify_rotate (&gsi))
5858 : changed = true;
5859 66750895 : else if (code == VEC_PERM_EXPR)
5860 190530 : changed |= simplify_permutation (&gsi);
5861 66560365 : else if (code == CONSTRUCTOR
5862 66560365 : && TREE_CODE (TREE_TYPE (rhs1)) == VECTOR_TYPE)
5863 164380 : changed |= simplify_vector_constructor (&gsi);
5864 66395985 : else if (code == ARRAY_REF)
5865 1996254 : changed |= simplify_count_zeroes (&gsi);
5866 : break;
5867 : }
5868 :
5869 107050 : case GIMPLE_SWITCH:
5870 107050 : changed |= simplify_gimple_switch (as_a <gswitch *> (stmt),
5871 : edges_to_remove,
5872 : simple_dce_worklist);
5873 107050 : break;
5874 :
5875 19670771 : case GIMPLE_COND:
5876 19670771 : {
5877 19670771 : int did_something = forward_propagate_into_gimple_cond
5878 19670771 : (as_a <gcond *> (stmt));
5879 19670771 : if (did_something == 2)
5880 1653 : cfg_changed = true;
5881 19670771 : changed |= did_something != 0;
5882 19670771 : break;
5883 : }
5884 :
5885 23765488 : case GIMPLE_CALL:
5886 23765488 : {
5887 23765488 : tree callee = gimple_call_fndecl (stmt);
5888 23765488 : if (callee != NULL_TREE
5889 23765488 : && fndecl_built_in_p (callee, BUILT_IN_NORMAL))
5890 6293190 : changed |= simplify_builtin_call (&gsi, callee, full_walk);
5891 : break;
5892 : }
5893 :
5894 355480022 : default:;
5895 : }
5896 :
5897 355480022 : if (changed || substituted_p)
5898 : {
5899 4097687 : substituted_p = false;
5900 4097687 : stmt = gsi_stmt (gsi);
5901 4097687 : if (maybe_clean_or_replace_eh_stmt (orig_stmt, stmt))
5902 70 : bitmap_set_bit (to_purge, bb->index);
5903 4097687 : if (!was_noreturn
5904 4097687 : && is_gimple_call (stmt) && gimple_call_noreturn_p (stmt))
5905 12 : to_fixup.safe_push (stmt);
5906 : }
5907 4097687 : if (changed)
5908 : {
5909 : /* If the stmt changed then re-visit it and the statements
5910 : inserted before it. */
5911 8845892 : for (; !gsi_end_p (gsi); gsi_prev (&gsi))
5912 5615289 : if (gimple_plf (gsi_stmt (gsi), GF_PLF_1))
5913 : break;
5914 2827194 : if (gsi_end_p (gsi))
5915 442508 : gsi = gsi_start_bb (bb);
5916 : else
5917 2605940 : gsi_next (&gsi);
5918 : }
5919 355482704 : }
5920 : while (changed);
5921 :
5922 : /* Stmt no longer needs to be revisited. */
5923 352655510 : stmt = gsi_stmt (gsi);
5924 352655510 : gcc_checking_assert (!gimple_plf (stmt, GF_PLF_1));
5925 352655510 : gimple_set_plf (stmt, GF_PLF_1, true);
5926 :
5927 : /* Fill up the lattice. */
5928 352655510 : if (gimple_assign_single_p (stmt))
5929 : {
5930 71590327 : tree lhs = gimple_assign_lhs (stmt);
5931 71590327 : tree rhs = gimple_assign_rhs1 (stmt);
5932 71590327 : if (TREE_CODE (lhs) == SSA_NAME)
5933 : {
5934 32924352 : tree val = lhs;
5935 32924352 : if (TREE_CODE (rhs) == SSA_NAME)
5936 787453 : val = fwprop_ssa_val (rhs);
5937 32136899 : else if (is_gimple_min_invariant (rhs))
5938 428087 : val = rhs;
5939 : /* If we can propagate the lattice-value mark the
5940 : stmt for removal. */
5941 32924352 : if (val != lhs
5942 32924352 : && may_propagate_copy (lhs, val))
5943 1212110 : to_remove_defs.safe_push (SSA_NAME_VERSION (lhs));
5944 32924352 : fwprop_set_lattice_val (lhs, val);
5945 : }
5946 : }
5947 281065183 : else if (gimple_nop_p (stmt))
5948 89584 : to_remove.safe_push (stmt);
5949 : }
5950 :
5951 : /* Substitute in destination PHI arguments. */
5952 109730910 : FOR_EACH_EDGE (e, ei, bb->succs)
5953 64126195 : for (gphi_iterator gsi = gsi_start_phis (e->dest);
5954 106467047 : !gsi_end_p (gsi); gsi_next (&gsi))
5955 : {
5956 42340852 : gphi *phi = gsi.phi ();
5957 42340852 : use_operand_p use_p = PHI_ARG_DEF_PTR_FROM_EDGE (phi, e);
5958 42340852 : tree arg = USE_FROM_PTR (use_p);
5959 69793774 : if (TREE_CODE (arg) != SSA_NAME
5960 42340852 : || virtual_operand_p (arg))
5961 27452922 : continue;
5962 14887930 : tree val = fwprop_ssa_val (arg);
5963 14887930 : if (val != arg
5964 14887930 : && may_propagate_copy (arg, val, !(e->flags & EDGE_ABNORMAL)))
5965 244029 : propagate_value (use_p, val);
5966 : }
5967 :
5968 : /* Mark outgoing executable edges. */
5969 45604715 : if (edge e = find_taken_edge (bb, NULL))
5970 : {
5971 19364628 : e->flags |= EDGE_EXECUTABLE;
5972 45624653 : if (EDGE_COUNT (bb->succs) > 1)
5973 19938 : cfg_changed = true;
5974 : }
5975 : else
5976 : {
5977 70981715 : FOR_EACH_EDGE (e, ei, bb->succs)
5978 44741628 : e->flags |= EDGE_EXECUTABLE;
5979 : }
5980 : }
5981 5693299 : free (postorder);
5982 5693299 : free (bb_to_rpo);
5983 5693299 : lattice.release ();
5984 :
5985 : /* First remove chains of stmts where we check no uses remain. */
5986 5693299 : simple_dce_from_worklist (simple_dce_worklist, to_purge);
5987 :
5988 6036201 : auto remove = [](gimple *stmt)
5989 : {
5990 342902 : if (dump_file && (dump_flags & TDF_DETAILS))
5991 : {
5992 1 : fprintf (dump_file, "Removing dead stmt ");
5993 1 : print_gimple_stmt (dump_file, stmt, 0);
5994 1 : fprintf (dump_file, "\n");
5995 : }
5996 342902 : gimple_stmt_iterator gsi = gsi_for_stmt (stmt);
5997 342902 : if (gimple_code (stmt) == GIMPLE_PHI)
5998 92699 : remove_phi_node (&gsi, true);
5999 : else
6000 : {
6001 250203 : unlink_stmt_vdef (stmt);
6002 250203 : gsi_remove (&gsi, true);
6003 250203 : release_defs (stmt);
6004 : }
6005 342902 : };
6006 :
6007 : /* Then remove stmts we know we can remove even though we did not
6008 : substitute in dead code regions, so uses can remain. Do so in reverse
6009 : order to make debug stmt creation possible. */
6010 12805972 : while (!to_remove_defs.is_empty())
6011 : {
6012 1419374 : tree def = ssa_name (to_remove_defs.pop ());
6013 : /* For example remove_prop_source_from_use can remove stmts queued
6014 : for removal. Deal with this gracefully. */
6015 1419374 : if (!def)
6016 1166056 : continue;
6017 253318 : gimple *stmt = SSA_NAME_DEF_STMT (def);
6018 253318 : remove (stmt);
6019 : }
6020 :
6021 : /* Wipe other queued stmts that do not have SSA defs. */
6022 5782883 : while (!to_remove.is_empty())
6023 : {
6024 89584 : gimple *stmt = to_remove.pop ();
6025 89584 : remove (stmt);
6026 : }
6027 :
6028 : /* Fixup stmts that became noreturn calls. This may require splitting
6029 : blocks and thus isn't possible during the walk. Do this
6030 : in reverse order so we don't inadvertently remove a stmt we want to
6031 : fixup by visiting a dominating now noreturn call first. */
6032 5693311 : while (!to_fixup.is_empty ())
6033 : {
6034 12 : gimple *stmt = to_fixup.pop ();
6035 12 : if (dump_file && dump_flags & TDF_DETAILS)
6036 : {
6037 0 : fprintf (dump_file, "Fixing up noreturn call ");
6038 0 : print_gimple_stmt (dump_file, stmt, 0);
6039 0 : fprintf (dump_file, "\n");
6040 : }
6041 12 : cfg_changed |= fixup_noreturn_call (stmt);
6042 : }
6043 :
6044 5693299 : cfg_changed |= gimple_purge_all_dead_eh_edges (to_purge);
6045 5693299 : cfg_changed |= gimple_purge_all_dead_abnormal_call_edges (need_ab_cleanup);
6046 5693299 : BITMAP_FREE (to_purge);
6047 :
6048 : /* Remove edges queued from switch stmt simplification. */
6049 17079897 : for (auto ep : edges_to_remove)
6050 : {
6051 0 : basic_block src = BASIC_BLOCK_FOR_FN (fun, ep.first);
6052 0 : basic_block dest = BASIC_BLOCK_FOR_FN (fun, ep.second);
6053 0 : edge e;
6054 0 : if (src && dest && (e = find_edge (src, dest)))
6055 : {
6056 0 : free_dominance_info (CDI_DOMINATORS);
6057 0 : remove_edge (e);
6058 0 : cfg_changed = true;
6059 : }
6060 : }
6061 :
6062 11385055 : if (get_range_query (fun) != get_global_range_query ())
6063 1543 : disable_ranger (fun);
6064 :
6065 5693299 : if (cfg_changed)
6066 9409 : todoflags |= TODO_cleanup_cfg;
6067 :
6068 5693299 : return todoflags;
6069 5693299 : }
6070 :
6071 : } // anon namespace
6072 :
6073 : gimple_opt_pass *
6074 292371 : make_pass_forwprop (gcc::context *ctxt)
6075 : {
6076 292371 : return new pass_forwprop (ctxt);
6077 : }
|
