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 32089623 : fwprop_set_lattice_val (tree name, tree val)
222 : {
223 32089623 : if (TREE_CODE (name) == SSA_NAME)
224 : {
225 32089623 : if (SSA_NAME_VERSION (name) >= lattice.length ())
226 : {
227 32193 : lattice.reserve (num_ssa_names - lattice.length ());
228 21462 : lattice.quick_grow_cleared (num_ssa_names);
229 : }
230 32089623 : lattice[SSA_NAME_VERSION (name)] = val;
231 : /* As this now constitutes a copy duplicate points-to
232 : and range info appropriately. */
233 32089623 : if (TREE_CODE (val) == SSA_NAME)
234 31647279 : maybe_duplicate_ssa_info_at_copy (name, val);
235 : }
236 32089623 : }
237 :
238 : /* Invalidate the lattice entry for NAME, done when releasing SSA names. */
239 : static void
240 903681 : fwprop_invalidate_lattice (tree name)
241 : {
242 903681 : if (name
243 901337 : && TREE_CODE (name) == SSA_NAME
244 1804891 : && SSA_NAME_VERSION (name) < lattice.length ())
245 901181 : lattice[SSA_NAME_VERSION (name)] = NULL_TREE;
246 903681 : }
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 27713632 : get_prop_source_stmt (tree name, bool single_use_only, bool *single_use_p)
258 : {
259 27713632 : bool single_use = true;
260 :
261 27714616 : do {
262 27714124 : gimple *def_stmt = SSA_NAME_DEF_STMT (name);
263 :
264 27714124 : if (!has_single_use (name))
265 : {
266 15128337 : single_use = false;
267 15128337 : 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 27712713 : if (!is_gimple_assign (def_stmt))
273 : return NULL;
274 :
275 : /* If def_stmt is a simple copy, continue looking. */
276 19530931 : if (gimple_assign_rhs_code (def_stmt) == SSA_NAME)
277 492 : name = gimple_assign_rhs1 (def_stmt);
278 : else
279 : {
280 19530439 : if (!single_use_only && single_use_p)
281 19237818 : *single_use_p = single_use;
282 :
283 19530439 : return def_stmt;
284 : }
285 492 : } 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 27454814 : can_propagate_from (gimple *def_stmt)
293 : {
294 27454814 : gcc_assert (is_gimple_assign (def_stmt));
295 :
296 : /* If the rhs has side-effects we cannot propagate from it. */
297 27454814 : if (gimple_has_volatile_ops (def_stmt))
298 : return false;
299 :
300 : /* If the rhs is a load we cannot propagate from it. */
301 26863964 : if (TREE_CODE_CLASS (gimple_assign_rhs_code (def_stmt)) == tcc_reference
302 26863964 : || TREE_CODE_CLASS (gimple_assign_rhs_code (def_stmt)) == tcc_declaration)
303 : return false;
304 :
305 : /* Constants can be always propagated. */
306 13369300 : if (gimple_assign_single_p (def_stmt)
307 13369300 : && 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 13369300 : 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 13368615 : if (CONVERT_EXPR_CODE_P (gimple_assign_rhs_code (def_stmt)))
319 : {
320 3224591 : tree rhs = gimple_assign_rhs1 (def_stmt);
321 3224591 : 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 235363 : remove_prop_source_from_use (tree name)
336 : {
337 296105 : gimple_stmt_iterator gsi;
338 296105 : gimple *stmt;
339 :
340 296105 : do {
341 296105 : basic_block bb;
342 :
343 296105 : if (SSA_NAME_IN_FREE_LIST (name)
344 296062 : || SSA_NAME_IS_DEFAULT_DEF (name)
345 588559 : || !has_zero_uses (name))
346 : break;
347 :
348 61199 : stmt = SSA_NAME_DEF_STMT (name);
349 61199 : if (gimple_code (stmt) == GIMPLE_PHI
350 61199 : || gimple_has_side_effects (stmt))
351 : break;
352 :
353 61199 : bb = gimple_bb (stmt);
354 61199 : gsi = gsi_for_stmt (stmt);
355 61199 : unlink_stmt_vdef (stmt);
356 61199 : if (gsi_remove (&gsi, true))
357 6 : bitmap_set_bit (to_purge, bb->index);
358 61199 : fwprop_invalidate_lattice (gimple_get_lhs (stmt));
359 61199 : release_defs (stmt);
360 :
361 61199 : name = is_gimple_assign (stmt) ? gimple_assign_rhs1 (stmt) : NULL_TREE;
362 61199 : } while (name && TREE_CODE (name) == SSA_NAME);
363 :
364 235363 : }
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 7319329 : rhs_to_tree (tree type, gimple *stmt)
375 : {
376 7319329 : location_t loc = gimple_location (stmt);
377 7319329 : enum tree_code code = gimple_assign_rhs_code (stmt);
378 7319329 : switch (get_gimple_rhs_class (code))
379 : {
380 11980 : case GIMPLE_TERNARY_RHS:
381 11980 : return fold_build3_loc (loc, code, type, gimple_assign_rhs1 (stmt),
382 : gimple_assign_rhs2 (stmt),
383 11980 : gimple_assign_rhs3 (stmt));
384 5021015 : case GIMPLE_BINARY_RHS:
385 5021015 : return fold_build2_loc (loc, code, type, gimple_assign_rhs1 (stmt),
386 5021015 : gimple_assign_rhs2 (stmt));
387 2026741 : case GIMPLE_UNARY_RHS:
388 2026741 : return build1 (code, type, gimple_assign_rhs1 (stmt));
389 259593 : case GIMPLE_SINGLE_RHS:
390 259593 : 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 8239151 : combine_cond_expr_cond (gimple *stmt, enum tree_code code, tree type,
404 : tree op0, tree op1, bool invariant_only)
405 : {
406 8239151 : tree t;
407 :
408 8239151 : gcc_assert (TREE_CODE_CLASS (code) == tcc_comparison);
409 :
410 8239151 : fold_defer_overflow_warnings ();
411 8239151 : t = fold_binary_loc (gimple_location (stmt), code, type, op0, op1);
412 8239151 : if (!t)
413 : {
414 4680362 : fold_undefer_overflow_warnings (false, NULL, 0);
415 4680362 : return NULL_TREE;
416 : }
417 :
418 : /* Require that we got a boolean type out if we put one in. */
419 3558789 : gcc_assert (TREE_CODE (TREE_TYPE (t)) == TREE_CODE (type));
420 :
421 : /* Canonicalize the combined condition for use in a COND_EXPR. */
422 3558789 : t = canonicalize_cond_expr_cond (t);
423 :
424 : /* Bail out if we required an invariant but didn't get one. */
425 3558789 : if (!t || (invariant_only && !is_gimple_min_invariant (t)))
426 : {
427 3325526 : fold_undefer_overflow_warnings (false, NULL, 0);
428 3325526 : return NULL_TREE;
429 : }
430 :
431 233263 : bool nowarn = warning_suppressed_p (stmt, OPT_Wstrict_overflow);
432 233263 : fold_undefer_overflow_warnings (!nowarn, stmt, 0);
433 :
434 233263 : return t;
435 : }
436 :
437 : /* Combine the comparison OP0 CODE OP1 at LOC with the defining statements
438 : of its operand. Return a new comparison tree or NULL_TREE if there
439 : were no simplifying combines. */
440 :
441 : static tree
442 21773655 : forward_propagate_into_comparison_1 (gimple *stmt,
443 : enum tree_code code, tree type,
444 : tree op0, tree op1)
445 : {
446 21773655 : tree tmp = NULL_TREE;
447 21773655 : tree rhs0 = NULL_TREE, rhs1 = NULL_TREE;
448 21773655 : bool single_use0_p = false, single_use1_p = false;
449 :
450 : /* For comparisons use the first operand, that is likely to
451 : simplify comparisons against constants. */
452 21773655 : if (TREE_CODE (op0) == SSA_NAME)
453 : {
454 21734408 : gimple *def_stmt = get_prop_source_stmt (op0, false, &single_use0_p);
455 21734408 : if (def_stmt && can_propagate_from (def_stmt))
456 : {
457 5553976 : enum tree_code def_code = gimple_assign_rhs_code (def_stmt);
458 5553976 : bool invariant_only_p = !single_use0_p;
459 :
460 5553976 : rhs0 = rhs_to_tree (TREE_TYPE (op1), def_stmt);
461 :
462 : /* Always combine comparisons or conversions from booleans. */
463 5553976 : if (TREE_CODE (op1) == INTEGER_CST
464 5553976 : && ((CONVERT_EXPR_CODE_P (def_code)
465 881500 : && TREE_CODE (TREE_TYPE (TREE_OPERAND (rhs0, 0)))
466 : == BOOLEAN_TYPE)
467 3576283 : || TREE_CODE_CLASS (def_code) == tcc_comparison))
468 : invariant_only_p = false;
469 :
470 5553976 : tmp = combine_cond_expr_cond (stmt, code, type,
471 : rhs0, op1, invariant_only_p);
472 5553976 : if (tmp)
473 : return tmp;
474 : }
475 : }
476 :
477 : /* If that wasn't successful, try the second operand. */
478 21548566 : if (TREE_CODE (op1) == SSA_NAME)
479 : {
480 5434936 : gimple *def_stmt = get_prop_source_stmt (op1, false, &single_use1_p);
481 5434936 : if (def_stmt && can_propagate_from (def_stmt))
482 : {
483 1765353 : rhs1 = rhs_to_tree (TREE_TYPE (op0), def_stmt);
484 3530706 : tmp = combine_cond_expr_cond (stmt, code, type,
485 1765353 : op0, rhs1, !single_use1_p);
486 1765353 : if (tmp)
487 : return tmp;
488 : }
489 : }
490 :
491 : /* If that wasn't successful either, try both operands. */
492 21542332 : if (rhs0 != NULL_TREE
493 21542332 : && rhs1 != NULL_TREE)
494 919822 : tmp = combine_cond_expr_cond (stmt, code, type,
495 : rhs0, rhs1,
496 919822 : !(single_use0_p && single_use1_p));
497 :
498 : return tmp;
499 : }
500 :
501 : /* Propagate from the ssa name definition statements of the assignment
502 : from a comparison at *GSI into the conditional if that simplifies it.
503 : Returns true if the stmt was modified. */
504 :
505 : static bool
506 2515964 : forward_propagate_into_comparison (gimple_stmt_iterator *gsi)
507 : {
508 2515964 : gimple *stmt = gsi_stmt (*gsi);
509 2515964 : tree tmp;
510 2515964 : tree type = TREE_TYPE (gimple_assign_lhs (stmt));
511 2515964 : tree rhs1 = gimple_assign_rhs1 (stmt);
512 2515964 : tree rhs2 = gimple_assign_rhs2 (stmt);
513 :
514 : /* Combine the comparison with defining statements. */
515 2515964 : tmp = forward_propagate_into_comparison_1 (stmt,
516 : gimple_assign_rhs_code (stmt),
517 : type, rhs1, rhs2);
518 2515964 : if (tmp && useless_type_conversion_p (type, TREE_TYPE (tmp)))
519 : {
520 7043 : if (dump_file)
521 : {
522 0 : fprintf (dump_file, " Replaced '");
523 0 : print_gimple_expr (dump_file, stmt, 0);
524 0 : fprintf (dump_file, "' with '");
525 0 : print_generic_expr (dump_file, tmp);
526 0 : fprintf (dump_file, "'\n");
527 : }
528 7043 : gimple_assign_set_rhs_from_tree (gsi, tmp);
529 7043 : fold_stmt (gsi);
530 7043 : update_stmt (gsi_stmt (*gsi));
531 :
532 7043 : if (TREE_CODE (rhs1) == SSA_NAME)
533 7043 : remove_prop_source_from_use (rhs1);
534 7043 : if (TREE_CODE (rhs2) == SSA_NAME)
535 2882 : remove_prop_source_from_use (rhs2);
536 7043 : return true;
537 : }
538 :
539 : return false;
540 : }
541 :
542 : /* Propagate from the ssa name definition statements of COND_EXPR
543 : in GIMPLE_COND statement STMT into the conditional if that simplifies it.
544 : Returns zero if no statement was changed, one if there were
545 : changes and two if cfg_cleanup needs to run. */
546 :
547 : static int
548 19257691 : forward_propagate_into_gimple_cond (gcond *stmt)
549 : {
550 19257691 : tree tmp;
551 19257691 : enum tree_code code = gimple_cond_code (stmt);
552 19257691 : tree rhs1 = gimple_cond_lhs (stmt);
553 19257691 : tree rhs2 = gimple_cond_rhs (stmt);
554 :
555 : /* GIMPLE_COND will always be a comparison. */
556 19257691 : gcc_assert (TREE_CODE_CLASS (gimple_cond_code (stmt)) == tcc_comparison);
557 :
558 19257691 : tmp = forward_propagate_into_comparison_1 (stmt, code,
559 : boolean_type_node,
560 : rhs1, rhs2);
561 19257691 : if (tmp
562 19257691 : && is_gimple_condexpr_for_cond (tmp))
563 : {
564 219885 : if (dump_file)
565 : {
566 9 : fprintf (dump_file, " Replaced '");
567 9 : print_gimple_expr (dump_file, stmt, 0);
568 9 : fprintf (dump_file, "' with '");
569 9 : print_generic_expr (dump_file, tmp);
570 9 : fprintf (dump_file, "'\n");
571 : }
572 :
573 219885 : gimple_cond_set_condition_from_tree (stmt, unshare_expr (tmp));
574 219885 : update_stmt (stmt);
575 :
576 219885 : if (TREE_CODE (rhs1) == SSA_NAME)
577 219885 : remove_prop_source_from_use (rhs1);
578 219885 : if (TREE_CODE (rhs2) == SSA_NAME)
579 5552 : remove_prop_source_from_use (rhs2);
580 219885 : return is_gimple_min_invariant (tmp) ? 2 : 1;
581 : }
582 :
583 19037806 : if (canonicalize_bool_cond (stmt, gimple_bb (stmt)))
584 : return 1;
585 :
586 : return 0;
587 : }
588 :
589 : /* We've just substituted an ADDR_EXPR into stmt. Update all the
590 : relevant data structures to match. */
591 :
592 : static void
593 1925713 : tidy_after_forward_propagate_addr (gimple *stmt)
594 : {
595 : /* We may have turned a trapping insn into a non-trapping insn. */
596 1925713 : if (maybe_clean_or_replace_eh_stmt (stmt, stmt))
597 131 : bitmap_set_bit (to_purge, gimple_bb (stmt)->index);
598 :
599 1925713 : if (TREE_CODE (gimple_assign_rhs1 (stmt)) == ADDR_EXPR)
600 247702 : recompute_tree_invariant_for_addr_expr (gimple_assign_rhs1 (stmt));
601 1925713 : }
602 :
603 : /* NAME is a SSA_NAME representing DEF_RHS which is of the form
604 : ADDR_EXPR <whatever>.
605 :
606 : Try to forward propagate the ADDR_EXPR into the use USE_STMT.
607 : Often this will allow for removal of an ADDR_EXPR and INDIRECT_REF
608 : node or for recovery of array indexing from pointer arithmetic.
609 :
610 : Return true if the propagation was successful (the propagation can
611 : be not totally successful, yet things may have been changed). */
612 :
613 : static bool
614 2746468 : forward_propagate_addr_expr_1 (tree name, tree def_rhs,
615 : gimple_stmt_iterator *use_stmt_gsi,
616 : bool single_use_p)
617 : {
618 2746468 : tree lhs, rhs, rhs2, array_ref;
619 2746468 : gimple *use_stmt = gsi_stmt (*use_stmt_gsi);
620 2746468 : enum tree_code rhs_code;
621 2746468 : bool res = true;
622 :
623 2746468 : gcc_assert (TREE_CODE (def_rhs) == ADDR_EXPR);
624 :
625 2746468 : lhs = gimple_assign_lhs (use_stmt);
626 2746468 : rhs_code = gimple_assign_rhs_code (use_stmt);
627 2746468 : rhs = gimple_assign_rhs1 (use_stmt);
628 :
629 : /* Do not perform copy-propagation but recurse through copy chains. */
630 2746468 : if (TREE_CODE (lhs) == SSA_NAME
631 1358613 : && rhs_code == SSA_NAME)
632 6888 : return forward_propagate_addr_expr (lhs, def_rhs, single_use_p);
633 :
634 : /* The use statement could be a conversion. Recurse to the uses of the
635 : lhs as copyprop does not copy through pointer to integer to pointer
636 : conversions and FRE does not catch all cases either.
637 : Treat the case of a single-use name and
638 : a conversion to def_rhs type separate, though. */
639 2739580 : if (TREE_CODE (lhs) == SSA_NAME
640 1351725 : && CONVERT_EXPR_CODE_P (rhs_code))
641 : {
642 : /* If there is a point in a conversion chain where the types match
643 : so we can remove a conversion re-materialize the address here
644 : and stop. */
645 23978 : if (single_use_p
646 23978 : && useless_type_conversion_p (TREE_TYPE (lhs), TREE_TYPE (def_rhs)))
647 : {
648 1 : gimple_assign_set_rhs1 (use_stmt, unshare_expr (def_rhs));
649 1 : gimple_assign_set_rhs_code (use_stmt, TREE_CODE (def_rhs));
650 1 : return true;
651 : }
652 :
653 : /* Else recurse if the conversion preserves the address value. */
654 47954 : if ((INTEGRAL_TYPE_P (TREE_TYPE (lhs))
655 2 : || POINTER_TYPE_P (TREE_TYPE (lhs)))
656 47954 : && (TYPE_PRECISION (TREE_TYPE (lhs))
657 23977 : >= TYPE_PRECISION (TREE_TYPE (def_rhs))))
658 23910 : return forward_propagate_addr_expr (lhs, def_rhs, single_use_p);
659 :
660 : return false;
661 : }
662 :
663 : /* If this isn't a conversion chain from this on we only can propagate
664 : into compatible pointer contexts. */
665 2715602 : if (!types_compatible_p (TREE_TYPE (name), TREE_TYPE (def_rhs)))
666 : return false;
667 :
668 : /* Propagate through constant pointer adjustments. */
669 2694944 : if (TREE_CODE (lhs) == SSA_NAME
670 1308262 : && rhs_code == POINTER_PLUS_EXPR
671 1308262 : && rhs == name
672 2858436 : && TREE_CODE (gimple_assign_rhs2 (use_stmt)) == INTEGER_CST)
673 : {
674 118042 : tree new_def_rhs;
675 : /* As we come here with non-invariant addresses in def_rhs we need
676 : to make sure we can build a valid constant offsetted address
677 : for further propagation. Simply rely on fold building that
678 : and check after the fact. */
679 118042 : new_def_rhs = fold_build2 (MEM_REF, TREE_TYPE (TREE_TYPE (rhs)),
680 : def_rhs,
681 : fold_convert (ptr_type_node,
682 : gimple_assign_rhs2 (use_stmt)));
683 118042 : if (TREE_CODE (new_def_rhs) == MEM_REF
684 118042 : && !is_gimple_mem_ref_addr (TREE_OPERAND (new_def_rhs, 0)))
685 : return false;
686 114094 : new_def_rhs = build1 (ADDR_EXPR, TREE_TYPE (rhs), new_def_rhs);
687 :
688 : /* Recurse. If we could propagate into all uses of lhs do not
689 : bother to replace into the current use but just pretend we did. */
690 114094 : if (forward_propagate_addr_expr (lhs, new_def_rhs, single_use_p))
691 : return true;
692 :
693 38084 : if (useless_type_conversion_p (TREE_TYPE (lhs),
694 38084 : TREE_TYPE (new_def_rhs)))
695 38084 : gimple_assign_set_rhs_with_ops (use_stmt_gsi, TREE_CODE (new_def_rhs),
696 : new_def_rhs);
697 0 : else if (is_gimple_min_invariant (new_def_rhs))
698 0 : gimple_assign_set_rhs_with_ops (use_stmt_gsi, NOP_EXPR, new_def_rhs);
699 : else
700 : return false;
701 38084 : gcc_assert (gsi_stmt (*use_stmt_gsi) == use_stmt);
702 38084 : update_stmt (use_stmt);
703 38084 : return true;
704 : }
705 :
706 : /* Now strip away any outer COMPONENT_REF/ARRAY_REF nodes from the LHS.
707 : ADDR_EXPR will not appear on the LHS. */
708 2576902 : tree *lhsp = gimple_assign_lhs_ptr (use_stmt);
709 3893500 : while (handled_component_p (*lhsp))
710 1316598 : lhsp = &TREE_OPERAND (*lhsp, 0);
711 2576902 : lhs = *lhsp;
712 :
713 : /* Now see if the LHS node is a MEM_REF using NAME. If so,
714 : propagate the ADDR_EXPR into the use of NAME and fold the result. */
715 2576902 : if (TREE_CODE (lhs) == MEM_REF
716 2576902 : && TREE_OPERAND (lhs, 0) == name)
717 : {
718 870808 : tree def_rhs_base;
719 870808 : poly_int64 def_rhs_offset;
720 : /* If the address is invariant we can always fold it. */
721 870808 : if ((def_rhs_base = get_addr_base_and_unit_offset (TREE_OPERAND (def_rhs, 0),
722 : &def_rhs_offset)))
723 : {
724 827868 : poly_offset_int off = mem_ref_offset (lhs);
725 827868 : tree new_ptr;
726 827868 : off += def_rhs_offset;
727 827868 : if (TREE_CODE (def_rhs_base) == MEM_REF)
728 : {
729 808493 : off += mem_ref_offset (def_rhs_base);
730 808493 : new_ptr = TREE_OPERAND (def_rhs_base, 0);
731 : }
732 : else
733 19375 : new_ptr = build_fold_addr_expr (def_rhs_base);
734 827868 : TREE_OPERAND (lhs, 0) = new_ptr;
735 827868 : TREE_OPERAND (lhs, 1)
736 827868 : = wide_int_to_tree (TREE_TYPE (TREE_OPERAND (lhs, 1)), off);
737 827868 : tidy_after_forward_propagate_addr (use_stmt);
738 : /* Continue propagating into the RHS if this was not the only use. */
739 827868 : if (single_use_p)
740 223145 : return true;
741 : }
742 : /* If the LHS is a plain dereference and the value type is the same as
743 : that of the pointed-to type of the address we can put the
744 : dereferenced address on the LHS preserving the original alias-type. */
745 42940 : else if (integer_zerop (TREE_OPERAND (lhs, 1))
746 17321 : && ((gimple_assign_lhs (use_stmt) == lhs
747 13967 : && useless_type_conversion_p
748 13967 : (TREE_TYPE (TREE_OPERAND (def_rhs, 0)),
749 13967 : TREE_TYPE (gimple_assign_rhs1 (use_stmt))))
750 13097 : || types_compatible_p (TREE_TYPE (lhs),
751 13097 : TREE_TYPE (TREE_OPERAND (def_rhs, 0))))
752 : /* Don't forward anything into clobber stmts if it would result
753 : in the lhs no longer being a MEM_REF. */
754 50171 : && (!gimple_clobber_p (use_stmt)
755 161 : || TREE_CODE (TREE_OPERAND (def_rhs, 0)) == MEM_REF))
756 : {
757 7070 : tree *def_rhs_basep = &TREE_OPERAND (def_rhs, 0);
758 7070 : tree new_offset, new_base, saved, new_lhs;
759 25480 : while (handled_component_p (*def_rhs_basep))
760 11340 : def_rhs_basep = &TREE_OPERAND (*def_rhs_basep, 0);
761 7070 : saved = *def_rhs_basep;
762 7070 : if (TREE_CODE (*def_rhs_basep) == MEM_REF)
763 : {
764 3712 : new_base = TREE_OPERAND (*def_rhs_basep, 0);
765 3712 : new_offset = fold_convert (TREE_TYPE (TREE_OPERAND (lhs, 1)),
766 : TREE_OPERAND (*def_rhs_basep, 1));
767 : }
768 : else
769 : {
770 3358 : new_base = build_fold_addr_expr (*def_rhs_basep);
771 3358 : new_offset = TREE_OPERAND (lhs, 1);
772 : }
773 7070 : *def_rhs_basep = build2 (MEM_REF, TREE_TYPE (*def_rhs_basep),
774 : new_base, new_offset);
775 7070 : TREE_THIS_VOLATILE (*def_rhs_basep) = TREE_THIS_VOLATILE (lhs);
776 7070 : TREE_SIDE_EFFECTS (*def_rhs_basep) = TREE_SIDE_EFFECTS (lhs);
777 7070 : TREE_THIS_NOTRAP (*def_rhs_basep) = TREE_THIS_NOTRAP (lhs);
778 7070 : new_lhs = unshare_expr (TREE_OPERAND (def_rhs, 0));
779 7070 : *lhsp = new_lhs;
780 7070 : TREE_THIS_VOLATILE (new_lhs) = TREE_THIS_VOLATILE (lhs);
781 7070 : TREE_SIDE_EFFECTS (new_lhs) = TREE_SIDE_EFFECTS (lhs);
782 7070 : *def_rhs_basep = saved;
783 7070 : tidy_after_forward_propagate_addr (use_stmt);
784 : /* Continue propagating into the RHS if this was not the
785 : only use. */
786 7070 : if (single_use_p)
787 : return true;
788 : }
789 : else
790 : /* We can have a struct assignment dereferencing our name twice.
791 : Note that we didn't propagate into the lhs to not falsely
792 : claim we did when propagating into the rhs. */
793 : res = false;
794 : }
795 :
796 : /* Strip away any outer COMPONENT_REF, ARRAY_REF or ADDR_EXPR
797 : nodes from the RHS. */
798 2350446 : tree *rhsp = gimple_assign_rhs1_ptr (use_stmt);
799 2350446 : if (TREE_CODE (*rhsp) == ADDR_EXPR)
800 236164 : rhsp = &TREE_OPERAND (*rhsp, 0);
801 3325284 : while (handled_component_p (*rhsp))
802 974838 : rhsp = &TREE_OPERAND (*rhsp, 0);
803 2350446 : rhs = *rhsp;
804 :
805 : /* Now see if the RHS node is a MEM_REF using NAME. If so,
806 : propagate the ADDR_EXPR into the use of NAME and fold the result. */
807 2350446 : if (TREE_CODE (rhs) == MEM_REF
808 2350446 : && TREE_OPERAND (rhs, 0) == name)
809 : {
810 1109609 : tree def_rhs_base;
811 1109609 : poly_int64 def_rhs_offset;
812 1109609 : if ((def_rhs_base = get_addr_base_and_unit_offset (TREE_OPERAND (def_rhs, 0),
813 : &def_rhs_offset)))
814 : {
815 1076314 : poly_offset_int off = mem_ref_offset (rhs);
816 1076314 : tree new_ptr;
817 1076314 : off += def_rhs_offset;
818 1076314 : if (TREE_CODE (def_rhs_base) == MEM_REF)
819 : {
820 1050208 : off += mem_ref_offset (def_rhs_base);
821 1050208 : new_ptr = TREE_OPERAND (def_rhs_base, 0);
822 : }
823 : else
824 26106 : new_ptr = build_fold_addr_expr (def_rhs_base);
825 1076314 : TREE_OPERAND (rhs, 0) = new_ptr;
826 1076314 : TREE_OPERAND (rhs, 1)
827 1076314 : = wide_int_to_tree (TREE_TYPE (TREE_OPERAND (rhs, 1)), off);
828 1076314 : fold_stmt_inplace (use_stmt_gsi);
829 1076314 : tidy_after_forward_propagate_addr (use_stmt);
830 1076314 : return res;
831 : }
832 : /* If the RHS is a plain dereference and the value type is the same as
833 : that of the pointed-to type of the address we can put the
834 : dereferenced address on the RHS preserving the original alias-type. */
835 33295 : else if (integer_zerop (TREE_OPERAND (rhs, 1))
836 33295 : && ((gimple_assign_rhs1 (use_stmt) == rhs
837 19321 : && useless_type_conversion_p
838 19321 : (TREE_TYPE (gimple_assign_lhs (use_stmt)),
839 19321 : TREE_TYPE (TREE_OPERAND (def_rhs, 0))))
840 22065 : || types_compatible_p (TREE_TYPE (rhs),
841 22065 : TREE_TYPE (TREE_OPERAND (def_rhs, 0)))))
842 : {
843 14461 : tree *def_rhs_basep = &TREE_OPERAND (def_rhs, 0);
844 14461 : tree new_offset, new_base, saved, new_rhs;
845 51172 : while (handled_component_p (*def_rhs_basep))
846 22250 : def_rhs_basep = &TREE_OPERAND (*def_rhs_basep, 0);
847 14461 : saved = *def_rhs_basep;
848 14461 : if (TREE_CODE (*def_rhs_basep) == MEM_REF)
849 : {
850 7025 : new_base = TREE_OPERAND (*def_rhs_basep, 0);
851 7025 : new_offset = fold_convert (TREE_TYPE (TREE_OPERAND (rhs, 1)),
852 : TREE_OPERAND (*def_rhs_basep, 1));
853 : }
854 : else
855 : {
856 7436 : new_base = build_fold_addr_expr (*def_rhs_basep);
857 7436 : new_offset = TREE_OPERAND (rhs, 1);
858 : }
859 14461 : *def_rhs_basep = build2 (MEM_REF, TREE_TYPE (*def_rhs_basep),
860 : new_base, new_offset);
861 14461 : TREE_THIS_VOLATILE (*def_rhs_basep) = TREE_THIS_VOLATILE (rhs);
862 14461 : TREE_SIDE_EFFECTS (*def_rhs_basep) = TREE_SIDE_EFFECTS (rhs);
863 14461 : TREE_THIS_NOTRAP (*def_rhs_basep) = TREE_THIS_NOTRAP (rhs);
864 14461 : new_rhs = unshare_expr (TREE_OPERAND (def_rhs, 0));
865 14461 : *rhsp = new_rhs;
866 14461 : TREE_THIS_VOLATILE (new_rhs) = TREE_THIS_VOLATILE (rhs);
867 14461 : TREE_SIDE_EFFECTS (new_rhs) = TREE_SIDE_EFFECTS (rhs);
868 14461 : *def_rhs_basep = saved;
869 14461 : fold_stmt_inplace (use_stmt_gsi);
870 14461 : tidy_after_forward_propagate_addr (use_stmt);
871 14461 : return res;
872 : }
873 : }
874 :
875 : /* If the use of the ADDR_EXPR is not a POINTER_PLUS_EXPR, there
876 : is nothing to do. */
877 1259671 : if (gimple_assign_rhs_code (use_stmt) != POINTER_PLUS_EXPR
878 1259671 : || gimple_assign_rhs1 (use_stmt) != name)
879 : return false;
880 :
881 : /* The remaining cases are all for turning pointer arithmetic into
882 : array indexing. They only apply when we have the address of
883 : element zero in an array. If that is not the case then there
884 : is nothing to do. */
885 45450 : array_ref = TREE_OPERAND (def_rhs, 0);
886 45450 : if ((TREE_CODE (array_ref) != ARRAY_REF
887 4543 : || TREE_CODE (TREE_TYPE (TREE_OPERAND (array_ref, 0))) != ARRAY_TYPE
888 4543 : || TREE_CODE (TREE_OPERAND (array_ref, 1)) != INTEGER_CST)
889 46928 : && TREE_CODE (TREE_TYPE (array_ref)) != ARRAY_TYPE)
890 : return false;
891 :
892 23006 : rhs2 = gimple_assign_rhs2 (use_stmt);
893 : /* Optimize &x[C1] p+ C2 to &x p+ C3 with C3 = C1 * element_size + C2. */
894 23006 : if (TREE_CODE (rhs2) == INTEGER_CST)
895 : {
896 0 : tree new_rhs = build1_loc (gimple_location (use_stmt),
897 0 : ADDR_EXPR, TREE_TYPE (def_rhs),
898 0 : fold_build2 (MEM_REF,
899 : TREE_TYPE (TREE_TYPE (def_rhs)),
900 : unshare_expr (def_rhs),
901 : fold_convert (ptr_type_node,
902 : rhs2)));
903 0 : gimple_assign_set_rhs_from_tree (use_stmt_gsi, new_rhs);
904 0 : use_stmt = gsi_stmt (*use_stmt_gsi);
905 0 : update_stmt (use_stmt);
906 0 : tidy_after_forward_propagate_addr (use_stmt);
907 0 : return true;
908 : }
909 :
910 : return false;
911 : }
912 :
913 : /* STMT is a statement of the form SSA_NAME = ADDR_EXPR <whatever>.
914 :
915 : Try to forward propagate the ADDR_EXPR into all uses of the SSA_NAME.
916 : Often this will allow for removal of an ADDR_EXPR and INDIRECT_REF
917 : node or for recovery of array indexing from pointer arithmetic.
918 :
919 : PARENT_SINGLE_USE_P tells if, when in a recursive invocation, NAME was
920 : the single use in the previous invocation. Pass true when calling
921 : this as toplevel.
922 :
923 : Returns true, if all uses have been propagated into. */
924 :
925 : static bool
926 3217223 : forward_propagate_addr_expr (tree name, tree rhs, bool parent_single_use_p)
927 : {
928 3217223 : bool all = true;
929 3217223 : bool single_use_p = parent_single_use_p && has_single_use (name);
930 :
931 16919958 : for (gimple *use_stmt : gather_imm_use_stmts (name))
932 : {
933 7268289 : bool result;
934 7268289 : tree use_rhs;
935 :
936 : /* If the use is not in a simple assignment statement, then
937 : there is nothing we can do. */
938 7268289 : if (!is_gimple_assign (use_stmt))
939 : {
940 4521821 : if (!is_gimple_debug (use_stmt))
941 1889326 : all = false;
942 4521821 : continue;
943 : }
944 :
945 2746468 : gimple_stmt_iterator gsi = gsi_for_stmt (use_stmt);
946 2746468 : result = forward_propagate_addr_expr_1 (name, rhs, &gsi,
947 : single_use_p);
948 : /* If the use has moved to a different statement adjust
949 : the update machinery for the old statement too. */
950 2746468 : if (use_stmt != gsi_stmt (gsi))
951 : {
952 0 : update_stmt (use_stmt);
953 0 : use_stmt = gsi_stmt (gsi);
954 : }
955 2746468 : update_stmt (use_stmt);
956 2746468 : all &= result;
957 :
958 : /* Remove intermediate now unused copy and conversion chains. */
959 2746468 : use_rhs = gimple_assign_rhs1 (use_stmt);
960 2746468 : if (result
961 1434941 : && TREE_CODE (gimple_assign_lhs (use_stmt)) == SSA_NAME
962 1196476 : && TREE_CODE (use_rhs) == SSA_NAME
963 2826093 : && has_zero_uses (gimple_assign_lhs (use_stmt)))
964 : {
965 79625 : gimple_stmt_iterator gsi = gsi_for_stmt (use_stmt);
966 79625 : fwprop_invalidate_lattice (gimple_get_lhs (use_stmt));
967 79625 : release_defs (use_stmt);
968 79625 : gsi_remove (&gsi, true);
969 : }
970 3217223 : }
971 :
972 3217223 : return all && has_zero_uses (name);
973 : }
974 :
975 :
976 : /* Helper function for simplify_gimple_switch. Remove case labels that
977 : have values outside the range of the new type. */
978 :
979 : static void
980 11784 : simplify_gimple_switch_label_vec (gswitch *stmt, tree index_type,
981 : vec<std::pair<int, int> > &edges_to_remove)
982 : {
983 11784 : unsigned int branch_num = gimple_switch_num_labels (stmt);
984 11784 : auto_vec<tree> labels (branch_num);
985 11784 : unsigned int i, len;
986 :
987 : /* Collect the existing case labels in a VEC, and preprocess it as if
988 : we are gimplifying a GENERIC SWITCH_EXPR. */
989 74031 : for (i = 1; i < branch_num; i++)
990 50463 : labels.quick_push (gimple_switch_label (stmt, i));
991 11784 : preprocess_case_label_vec_for_gimple (labels, index_type, NULL);
992 :
993 : /* If any labels were removed, replace the existing case labels
994 : in the GIMPLE_SWITCH statement with the correct ones.
995 : Note that the type updates were done in-place on the case labels,
996 : so we only have to replace the case labels in the GIMPLE_SWITCH
997 : if the number of labels changed. */
998 11784 : len = labels.length ();
999 11784 : if (len < branch_num - 1)
1000 : {
1001 0 : bitmap target_blocks;
1002 0 : edge_iterator ei;
1003 0 : edge e;
1004 :
1005 : /* Corner case: *all* case labels have been removed as being
1006 : out-of-range for INDEX_TYPE. Push one label and let the
1007 : CFG cleanups deal with this further. */
1008 0 : if (len == 0)
1009 : {
1010 0 : tree label, elt;
1011 :
1012 0 : label = CASE_LABEL (gimple_switch_default_label (stmt));
1013 0 : elt = build_case_label (build_int_cst (index_type, 0), NULL, label);
1014 0 : labels.quick_push (elt);
1015 0 : len = 1;
1016 : }
1017 :
1018 0 : for (i = 0; i < labels.length (); i++)
1019 0 : gimple_switch_set_label (stmt, i + 1, labels[i]);
1020 0 : for (i++ ; i < branch_num; i++)
1021 0 : gimple_switch_set_label (stmt, i, NULL_TREE);
1022 0 : gimple_switch_set_num_labels (stmt, len + 1);
1023 :
1024 : /* Cleanup any edges that are now dead. */
1025 0 : target_blocks = BITMAP_ALLOC (NULL);
1026 0 : for (i = 0; i < gimple_switch_num_labels (stmt); i++)
1027 : {
1028 0 : tree elt = gimple_switch_label (stmt, i);
1029 0 : basic_block target = label_to_block (cfun, CASE_LABEL (elt));
1030 0 : bitmap_set_bit (target_blocks, target->index);
1031 : }
1032 0 : for (ei = ei_start (gimple_bb (stmt)->succs); (e = ei_safe_edge (ei)); )
1033 : {
1034 0 : if (! bitmap_bit_p (target_blocks, e->dest->index))
1035 0 : edges_to_remove.safe_push (std::make_pair (e->src->index,
1036 0 : e->dest->index));
1037 : else
1038 0 : ei_next (&ei);
1039 : }
1040 0 : BITMAP_FREE (target_blocks);
1041 : }
1042 11784 : }
1043 :
1044 : /* STMT is a SWITCH_EXPR for which we attempt to find equivalent forms of
1045 : the condition which we may be able to optimize better. */
1046 :
1047 : static bool
1048 110086 : simplify_gimple_switch (gswitch *stmt,
1049 : vec<std::pair<int, int> > &edges_to_remove,
1050 : bitmap simple_dce_worklist)
1051 : {
1052 : /* The optimization that we really care about is removing unnecessary
1053 : casts. That will let us do much better in propagating the inferred
1054 : constant at the switch target. */
1055 110086 : tree cond = gimple_switch_index (stmt);
1056 110086 : if (TREE_CODE (cond) == SSA_NAME)
1057 : {
1058 110085 : gimple *def_stmt = SSA_NAME_DEF_STMT (cond);
1059 110085 : if (gimple_assign_cast_p (def_stmt))
1060 : {
1061 12266 : tree def = gimple_assign_rhs1 (def_stmt);
1062 12266 : if (TREE_CODE (def) != SSA_NAME)
1063 : return false;
1064 :
1065 : /* If we have an extension or sign-change that preserves the
1066 : values we check against then we can copy the source value into
1067 : the switch. */
1068 12266 : tree ti = TREE_TYPE (def);
1069 12266 : if (INTEGRAL_TYPE_P (ti)
1070 12266 : && TYPE_PRECISION (ti) <= TYPE_PRECISION (TREE_TYPE (cond)))
1071 : {
1072 12021 : size_t n = gimple_switch_num_labels (stmt);
1073 12021 : tree min = NULL_TREE, max = NULL_TREE;
1074 12021 : if (n > 1)
1075 : {
1076 12021 : min = CASE_LOW (gimple_switch_label (stmt, 1));
1077 12021 : if (CASE_HIGH (gimple_switch_label (stmt, n - 1)))
1078 154 : max = CASE_HIGH (gimple_switch_label (stmt, n - 1));
1079 : else
1080 11867 : max = CASE_LOW (gimple_switch_label (stmt, n - 1));
1081 : }
1082 12021 : if ((!min || int_fits_type_p (min, ti))
1083 12017 : && (!max || int_fits_type_p (max, ti)))
1084 : {
1085 11784 : bitmap_set_bit (simple_dce_worklist,
1086 11784 : SSA_NAME_VERSION (cond));
1087 11784 : gimple_switch_set_index (stmt, def);
1088 11784 : simplify_gimple_switch_label_vec (stmt, ti,
1089 : edges_to_remove);
1090 11784 : update_stmt (stmt);
1091 11784 : return true;
1092 : }
1093 : }
1094 : }
1095 : }
1096 :
1097 : return false;
1098 : }
1099 :
1100 : /* For pointers p2 and p1 return p2 - p1 if the
1101 : difference is known and constant, otherwise return NULL. */
1102 :
1103 : static tree
1104 5234 : constant_pointer_difference (tree p1, tree p2)
1105 : {
1106 5234 : int i, j;
1107 : #define CPD_ITERATIONS 5
1108 5234 : tree exps[2][CPD_ITERATIONS];
1109 5234 : tree offs[2][CPD_ITERATIONS];
1110 5234 : int cnt[2];
1111 :
1112 15702 : for (i = 0; i < 2; i++)
1113 : {
1114 10468 : tree p = i ? p1 : p2;
1115 10468 : tree off = size_zero_node;
1116 10468 : gimple *stmt;
1117 10468 : enum tree_code code;
1118 :
1119 : /* For each of p1 and p2 we need to iterate at least
1120 : twice, to handle ADDR_EXPR directly in p1/p2,
1121 : SSA_NAME with ADDR_EXPR or POINTER_PLUS_EXPR etc.
1122 : on definition's stmt RHS. Iterate a few extra times. */
1123 10468 : j = 0;
1124 12412 : do
1125 : {
1126 12412 : if (!POINTER_TYPE_P (TREE_TYPE (p)))
1127 : break;
1128 12406 : if (TREE_CODE (p) == ADDR_EXPR)
1129 : {
1130 8988 : tree q = TREE_OPERAND (p, 0);
1131 8988 : poly_int64 offset;
1132 8988 : tree base = get_addr_base_and_unit_offset (q, &offset);
1133 8988 : if (base)
1134 : {
1135 8168 : q = base;
1136 8168 : if (maybe_ne (offset, 0))
1137 3407 : off = size_binop (PLUS_EXPR, off, size_int (offset));
1138 : }
1139 8988 : if (TREE_CODE (q) == MEM_REF
1140 8988 : && TREE_CODE (TREE_OPERAND (q, 0)) == SSA_NAME)
1141 : {
1142 213 : p = TREE_OPERAND (q, 0);
1143 213 : off = size_binop (PLUS_EXPR, off,
1144 : wide_int_to_tree (sizetype,
1145 : mem_ref_offset (q)));
1146 : }
1147 : else
1148 : {
1149 8775 : exps[i][j] = q;
1150 8775 : offs[i][j++] = off;
1151 8775 : break;
1152 : }
1153 : }
1154 3631 : if (TREE_CODE (p) != SSA_NAME)
1155 : break;
1156 3631 : exps[i][j] = p;
1157 3631 : offs[i][j++] = off;
1158 3631 : if (j == CPD_ITERATIONS)
1159 : break;
1160 3631 : stmt = SSA_NAME_DEF_STMT (p);
1161 3631 : if (!is_gimple_assign (stmt) || gimple_assign_lhs (stmt) != p)
1162 : break;
1163 2852 : code = gimple_assign_rhs_code (stmt);
1164 2852 : if (code == POINTER_PLUS_EXPR)
1165 : {
1166 1486 : if (TREE_CODE (gimple_assign_rhs2 (stmt)) != INTEGER_CST)
1167 : break;
1168 955 : off = size_binop (PLUS_EXPR, off, gimple_assign_rhs2 (stmt));
1169 955 : p = gimple_assign_rhs1 (stmt);
1170 : }
1171 1366 : else if (code == ADDR_EXPR || CONVERT_EXPR_CODE_P (code))
1172 989 : p = gimple_assign_rhs1 (stmt);
1173 : else
1174 : break;
1175 : }
1176 : while (1);
1177 10468 : cnt[i] = j;
1178 : }
1179 :
1180 7316 : for (i = 0; i < cnt[0]; i++)
1181 9755 : for (j = 0; j < cnt[1]; j++)
1182 7673 : if (exps[0][i] == exps[1][j])
1183 4344 : return size_binop (MINUS_EXPR, offs[0][i], offs[1][j]);
1184 :
1185 : return NULL_TREE;
1186 : }
1187 :
1188 : /* Helper function for optimize_aggr_zeroprop.
1189 : Props the zeroing (memset, VAL) that was done in DEST+OFFSET:LEN
1190 : (DEFSTMT) into the STMT. Returns true if the STMT was updated. */
1191 : static void
1192 20440779 : optimize_aggr_zeroprop_1 (gimple *defstmt, gimple *stmt,
1193 : tree dest, poly_int64 offset, tree val,
1194 : poly_offset_int len)
1195 : {
1196 20440779 : tree src2;
1197 20440779 : tree len2 = NULL_TREE;
1198 20440779 : poly_int64 offset2;
1199 :
1200 20440779 : if (gimple_call_builtin_p (stmt, BUILT_IN_MEMCPY)
1201 18085 : && TREE_CODE (gimple_call_arg (stmt, 1)) == ADDR_EXPR
1202 20454122 : && poly_int_tree_p (gimple_call_arg (stmt, 2)))
1203 : {
1204 12323 : src2 = TREE_OPERAND (gimple_call_arg (stmt, 1), 0);
1205 12323 : len2 = gimple_call_arg (stmt, 2);
1206 : }
1207 20428456 : else if (gimple_assign_load_p (stmt) && gimple_store_p (stmt))
1208 : {
1209 1846567 : src2 = gimple_assign_rhs1 (stmt);
1210 1846567 : len2 = (TREE_CODE (src2) == COMPONENT_REF
1211 1846567 : ? DECL_SIZE_UNIT (TREE_OPERAND (src2, 1))
1212 1679326 : : TYPE_SIZE_UNIT (TREE_TYPE (src2)));
1213 : /* Can only handle zero memsets. */
1214 1846567 : if (!integer_zerop (val))
1215 20419414 : return;
1216 : }
1217 : else
1218 18581889 : return;
1219 :
1220 1857928 : if (len2 == NULL_TREE
1221 1857928 : || !poly_int_tree_p (len2))
1222 : return;
1223 :
1224 1857928 : src2 = get_addr_base_and_unit_offset (src2, &offset2);
1225 1857928 : if (src2 == NULL_TREE
1226 1857928 : || maybe_lt (offset2, offset))
1227 : return;
1228 :
1229 857655 : if (!operand_equal_p (dest, src2, 0))
1230 : return;
1231 :
1232 : /* [ dest + offset, dest + offset + len - 1 ] is set to val.
1233 : Make sure that
1234 : [ dest + offset2, dest + offset2 + len2 - 1 ] is a subset of that. */
1235 130606 : if (maybe_gt (wi::to_poly_offset (len2) + (offset2 - offset),
1236 : len))
1237 : return;
1238 :
1239 21365 : if (dump_file && (dump_flags & TDF_DETAILS))
1240 : {
1241 32 : fprintf (dump_file, "Simplified\n ");
1242 32 : print_gimple_stmt (dump_file, stmt, 0, dump_flags);
1243 32 : fprintf (dump_file, "after previous\n ");
1244 32 : print_gimple_stmt (dump_file, defstmt, 0, dump_flags);
1245 : }
1246 21365 : gimple *orig_stmt = stmt;
1247 : /* For simplicity, don't change the kind of the stmt,
1248 : turn dest = src; into dest = {}; and memcpy (&dest, &src, len);
1249 : into memset (&dest, val, len);
1250 : In theory we could change dest = src into memset if dest
1251 : is addressable (maybe beneficial if val is not 0), or
1252 : memcpy (&dest, &src, len) into dest = {} if len is the size
1253 : of dest, dest isn't volatile. */
1254 21365 : if (is_gimple_assign (stmt))
1255 : {
1256 21360 : tree ctor_type = TREE_TYPE (gimple_assign_lhs (stmt));
1257 21360 : tree ctor = build_constructor (ctor_type, NULL);
1258 21360 : gimple_stmt_iterator gsi = gsi_for_stmt (stmt);
1259 21360 : gimple_assign_set_rhs_from_tree (&gsi, ctor);
1260 21360 : update_stmt (stmt);
1261 21360 : statistics_counter_event (cfun, "copy zeroing propagation of aggregate", 1);
1262 : }
1263 : else /* If stmt is memcpy, transform it into memset. */
1264 : {
1265 5 : gcall *call = as_a <gcall *> (stmt);
1266 5 : tree fndecl = builtin_decl_implicit (BUILT_IN_MEMSET);
1267 5 : gimple_call_set_fndecl (call, fndecl);
1268 5 : gimple_call_set_fntype (call, TREE_TYPE (fndecl));
1269 5 : gimple_call_set_arg (call, 1, val);
1270 5 : update_stmt (stmt);
1271 5 : statistics_counter_event (cfun, "memcpy to memset changed", 1);
1272 : }
1273 :
1274 21365 : if (dump_file && (dump_flags & TDF_DETAILS))
1275 : {
1276 32 : fprintf (dump_file, "into\n ");
1277 32 : print_gimple_stmt (dump_file, stmt, 0, dump_flags);
1278 : }
1279 :
1280 : /* Mark the bb for eh cleanup if needed. */
1281 21365 : if (maybe_clean_or_replace_eh_stmt (orig_stmt, stmt))
1282 6 : bitmap_set_bit (to_purge, gimple_bb (stmt)->index);
1283 : }
1284 :
1285 : /* Optimize
1286 : a = {}; // DEST = value ;; LEN(nullptr)
1287 : b = a;
1288 : into
1289 : a = {};
1290 : b = {};
1291 : Similarly for memset (&a, ..., sizeof (a)); instead of a = {};
1292 : and/or memcpy (&b, &a, sizeof (a)); instead of b = a; */
1293 :
1294 : static void
1295 30337969 : optimize_aggr_zeroprop (gimple *stmt, bool full_walk)
1296 : {
1297 30337969 : ao_ref read;
1298 60675938 : if (gimple_has_volatile_ops (stmt))
1299 26374970 : return;
1300 :
1301 29413053 : tree dest = NULL_TREE;
1302 29413053 : tree val = integer_zero_node;
1303 29413053 : tree len = NULL_TREE;
1304 29413053 : bool can_use_tbba = true;
1305 :
1306 29413053 : if (gimple_call_builtin_p (stmt, BUILT_IN_MEMSET)
1307 110687 : && TREE_CODE (gimple_call_arg (stmt, 0)) == ADDR_EXPR
1308 54846 : && TREE_CODE (gimple_call_arg (stmt, 1)) == INTEGER_CST
1309 29465620 : && poly_int_tree_p (gimple_call_arg (stmt, 2)))
1310 : {
1311 49791 : dest = TREE_OPERAND (gimple_call_arg (stmt, 0), 0);
1312 49791 : len = gimple_call_arg (stmt, 2);
1313 49791 : val = gimple_call_arg (stmt, 1);
1314 49791 : ao_ref_init_from_ptr_and_size (&read, gimple_call_arg (stmt, 0), len);
1315 49791 : can_use_tbba = false;
1316 : }
1317 29363262 : else if (gimple_store_p (stmt)
1318 29302206 : && gimple_assign_single_p (stmt)
1319 58665468 : && TREE_CODE (gimple_assign_rhs1 (stmt)) == STRING_CST)
1320 : {
1321 36400 : tree str = gimple_assign_rhs1 (stmt);
1322 36400 : dest = gimple_assign_lhs (stmt);
1323 36400 : ao_ref_init (&read, dest);
1324 : /* The string must contain all null char's for now. */
1325 41834 : for (int i = 0; i < TREE_STRING_LENGTH (str); i++)
1326 : {
1327 39151 : if (TREE_STRING_POINTER (str)[i] != 0)
1328 : {
1329 : dest = NULL_TREE;
1330 : break;
1331 : }
1332 : }
1333 : }
1334 : /* A store of integer (scalar, vector or complex) zeros is
1335 : a zero store. */
1336 29326862 : else if (gimple_store_p (stmt)
1337 29265806 : && gimple_assign_single_p (stmt)
1338 58592668 : && integer_zerop (gimple_assign_rhs1 (stmt)))
1339 : {
1340 3508116 : tree rhs = gimple_assign_rhs1 (stmt);
1341 3508116 : tree type = TREE_TYPE (rhs);
1342 3508116 : dest = gimple_assign_lhs (stmt);
1343 3508116 : ao_ref_init (&read, dest);
1344 : /* For integral types, the type precision needs to be a multiply of BITS_PER_UNIT. */
1345 3508116 : if (INTEGRAL_TYPE_P (type)
1346 3508116 : && (TYPE_PRECISION (type) % BITS_PER_UNIT) != 0)
1347 : dest = NULL_TREE;
1348 : }
1349 25818746 : else if (gimple_store_p (stmt)
1350 25757690 : && gimple_assign_single_p (stmt)
1351 25757690 : && TREE_CODE (gimple_assign_rhs1 (stmt)) == CONSTRUCTOR
1352 26515445 : && !gimple_clobber_p (stmt))
1353 : {
1354 696699 : dest = gimple_assign_lhs (stmt);
1355 696699 : ao_ref_init (&read, dest);
1356 : }
1357 :
1358 4082107 : if (dest == NULL_TREE)
1359 25364663 : return;
1360 :
1361 4048390 : if (len == NULL_TREE)
1362 3998599 : len = (TREE_CODE (dest) == COMPONENT_REF
1363 3998599 : ? DECL_SIZE_UNIT (TREE_OPERAND (dest, 1))
1364 1743150 : : TYPE_SIZE_UNIT (TREE_TYPE (dest)));
1365 3998599 : if (len == NULL_TREE
1366 4048390 : || !poly_int_tree_p (len))
1367 : return;
1368 :
1369 : /* This store needs to be on the byte boundary and pointing to an object. */
1370 4048390 : poly_int64 offset;
1371 4048390 : tree dest_base = get_addr_base_and_unit_offset (dest, &offset);
1372 4048390 : if (dest_base == NULL_TREE)
1373 : return;
1374 :
1375 : /* Setup the worklist. */
1376 3962999 : auto_vec<std::pair<tree, unsigned>> worklist;
1377 3962999 : unsigned limit = full_walk ? param_sccvn_max_alias_queries_per_access : 0;
1378 7925998 : worklist.safe_push (std::make_pair (gimple_vdef (stmt), limit));
1379 :
1380 26008848 : while (!worklist.is_empty ())
1381 : {
1382 18082850 : std::pair<tree, unsigned> top = worklist.pop ();
1383 18082850 : tree vdef = top.first;
1384 18082850 : limit = top.second;
1385 18082850 : gimple *use_stmt;
1386 18082850 : imm_use_iterator iter;
1387 58532664 : FOR_EACH_IMM_USE_STMT (use_stmt, iter, vdef)
1388 : {
1389 : /* Handling PHI nodes might not be worth it so don't. */
1390 22366964 : if (is_a <gphi*> (use_stmt))
1391 1926185 : continue;
1392 :
1393 : /* If this statement does not clobber add the vdef stmt to the
1394 : worklist.
1395 : After hitting the limit, allow clobbers to able to pass through. */
1396 2001838 : if ((limit != 0 || gimple_clobber_p (use_stmt))
1397 18474618 : && gimple_vdef (use_stmt)
1398 35968674 : && !stmt_may_clobber_ref_p_1 (use_stmt, &read,
1399 : /* tbaa_p = */ can_use_tbba))
1400 : {
1401 14119851 : unsigned new_limit = limit == 0 ? 0 : limit - 1;
1402 28239702 : worklist.safe_push (std::make_pair (gimple_vdef (use_stmt),
1403 : new_limit));
1404 : }
1405 :
1406 20440779 : optimize_aggr_zeroprop_1 (stmt, use_stmt, dest_base, offset,
1407 20440779 : val, wi::to_poly_offset (len));
1408 18082850 : }
1409 : }
1410 :
1411 3962999 : }
1412 :
1413 : /* Returns the pointer to the base of the object of the
1414 : reference EXPR and extracts the information about
1415 : the offset of the access, storing it to PBYTESIZE,
1416 : PBYTEPOS and PREVERSEP.
1417 : If the access is not a byte sized or position is not
1418 : on the byte, return NULL. */
1419 : static tree
1420 5178988 : split_core_and_offset_size (tree expr,
1421 : poly_int64 *pbytesize, poly_int64 *pbytepos,
1422 : tree *poffset, int *preversep)
1423 : {
1424 5178988 : tree core;
1425 5178988 : machine_mode mode;
1426 5178988 : int unsignedp, volatilep;
1427 5178988 : poly_int64 bitsize;
1428 5178988 : poly_int64 bitpos;
1429 5178988 : location_t loc = EXPR_LOCATION (expr);
1430 :
1431 5178988 : core = get_inner_reference (expr, &bitsize, &bitpos,
1432 : poffset, &mode, &unsignedp, preversep,
1433 : &volatilep);
1434 10357976 : if (!multiple_p (bitsize, BITS_PER_UNIT, pbytesize))
1435 : return NULL_TREE;
1436 5178988 : if (!multiple_p (bitpos, BITS_PER_UNIT, pbytepos))
1437 : return NULL_TREE;
1438 : /* If we are left with MEM[a + CST] strip that and add it to the
1439 : pbytepos and return a. */
1440 5178988 : if (TREE_CODE (core) == MEM_REF)
1441 : {
1442 1213911 : poly_offset_int tem;
1443 1213911 : tem = wi::to_poly_offset (TREE_OPERAND (core, 1));
1444 1213911 : tem += *pbytepos;
1445 1213911 : if (tem.to_shwi (pbytepos))
1446 1211988 : return TREE_OPERAND (core, 0);
1447 : }
1448 3967000 : core = build_fold_addr_expr_loc (loc, core);
1449 3967000 : STRIP_NOPS (core);
1450 3967000 : return core;
1451 : }
1452 :
1453 : /* Returns a new src based on the
1454 : copy `DEST = SRC` and for the old SRC2.
1455 : Returns null if SRC2 is not related to DEST. */
1456 :
1457 : static tree
1458 1198196 : new_src_based_on_copy (tree src2, tree dest, tree src)
1459 : {
1460 : /* If the second src is not exactly the same as dest,
1461 : try to handle it seperately; see it is address/size equivalent.
1462 : Handles `a` and `a.b` and `MEM<char[N]>(&a)` which all have
1463 : the same size and offsets as address/size equivalent.
1464 : This allows copying over a memcpy and also one for copying
1465 : where one field is the same size as the whole struct. */
1466 1198196 : if (operand_equal_p (dest, src2))
1467 : return src;
1468 : /* if both dest and src2 are decls, then we know these 2
1469 : accesses can't be the same. */
1470 698104 : if (DECL_P (dest) && DECL_P (src2))
1471 : return NULL_TREE;
1472 : /* A VCE can't be used with imag/real or BFR so reject them early. */
1473 365657 : if (TREE_CODE (src) == IMAGPART_EXPR
1474 365657 : || TREE_CODE (src) == REALPART_EXPR
1475 365657 : || TREE_CODE (src) == BIT_FIELD_REF)
1476 : return NULL_TREE;
1477 365657 : tree core1, core2;
1478 365657 : poly_int64 bytepos1, bytepos2;
1479 365657 : poly_int64 bytesize1, bytesize2;
1480 365657 : tree toffset1, toffset2;
1481 365657 : int reversep1 = 0;
1482 365657 : int reversep2 = 0;
1483 365657 : poly_int64 diff = 0;
1484 365657 : core1 = split_core_and_offset_size (dest, &bytesize1, &bytepos1,
1485 : &toffset1, &reversep1);
1486 365657 : core2 = split_core_and_offset_size (src2, &bytesize2, &bytepos2,
1487 : &toffset2, &reversep2);
1488 365657 : if (!core1 || !core2)
1489 : return NULL_TREE;
1490 365657 : if (reversep1 != reversep2)
1491 : return NULL_TREE;
1492 : /* The sizes of the 2 accesses need to be the same. */
1493 365657 : if (!known_eq (bytesize1, bytesize2))
1494 : return NULL_TREE;
1495 158203 : if (!operand_equal_p (core1, core2, 0))
1496 : return NULL_TREE;
1497 :
1498 22157 : if (toffset1 && toffset2)
1499 : {
1500 2 : tree type = TREE_TYPE (toffset1);
1501 2 : if (type != TREE_TYPE (toffset2))
1502 0 : toffset2 = fold_convert (type, toffset2);
1503 :
1504 2 : tree tdiff = fold_build2 (MINUS_EXPR, type, toffset1, toffset2);
1505 2 : if (!cst_and_fits_in_hwi (tdiff))
1506 : return NULL_TREE;
1507 :
1508 0 : diff = int_cst_value (tdiff);
1509 0 : }
1510 22155 : else if (toffset1 || toffset2)
1511 : {
1512 : /* If only one of the offsets is non-constant, the difference cannot
1513 : be a constant. */
1514 : return NULL_TREE;
1515 : }
1516 22123 : diff += bytepos1 - bytepos2;
1517 : /* The offset between the 2 need to be 0. */
1518 22123 : if (!known_eq (diff, 0))
1519 : return NULL_TREE;
1520 21402 : return fold_build1 (VIEW_CONVERT_EXPR,TREE_TYPE (src2), src);
1521 : }
1522 :
1523 : /* Returns true if SRC and DEST are the same address such that
1524 : `SRC == DEST;` is considered a nop. This is more than an
1525 : operand_equal_p check as it needs to be similar to
1526 : new_src_based_on_copy. */
1527 :
1528 : static bool
1529 4269459 : same_for_assignment (tree src, tree dest)
1530 : {
1531 4269459 : if (operand_equal_p (dest, src, 0))
1532 : return true;
1533 : /* if both dest and src2 are decls, then we know these 2
1534 : accesses can't be the same. */
1535 4266550 : if (DECL_P (dest) && DECL_P (src))
1536 : return false;
1537 :
1538 2223837 : tree core1, core2;
1539 2223837 : poly_int64 bytepos1, bytepos2;
1540 2223837 : poly_int64 bytesize1, bytesize2;
1541 2223837 : tree toffset1, toffset2;
1542 2223837 : int reversep1 = 0;
1543 2223837 : int reversep2 = 0;
1544 2223837 : poly_int64 diff = 0;
1545 2223837 : core1 = split_core_and_offset_size (dest, &bytesize1, &bytepos1,
1546 : &toffset1, &reversep1);
1547 2223837 : core2 = split_core_and_offset_size (src, &bytesize2, &bytepos2,
1548 : &toffset2, &reversep2);
1549 2223837 : if (!core1 || !core2)
1550 : return false;
1551 2223837 : if (reversep1 != reversep2)
1552 : return false;
1553 : /* The sizes of the 2 accesses need to be the same. */
1554 2223837 : if (!known_eq (bytesize1, bytesize2))
1555 : return false;
1556 2222923 : if (!operand_equal_p (core1, core2, 0))
1557 : return false;
1558 6091 : if (toffset1 && toffset2)
1559 : {
1560 343 : tree type = TREE_TYPE (toffset1);
1561 343 : if (type != TREE_TYPE (toffset2))
1562 0 : toffset2 = fold_convert (type, toffset2);
1563 :
1564 343 : tree tdiff = fold_build2 (MINUS_EXPR, type, toffset1, toffset2);
1565 343 : if (!cst_and_fits_in_hwi (tdiff))
1566 : return false;
1567 :
1568 0 : diff = int_cst_value (tdiff);
1569 0 : }
1570 5748 : else if (toffset1 || toffset2)
1571 : {
1572 : /* If only one of the offsets is non-constant, the difference cannot
1573 : be a constant. */
1574 : return false;
1575 : }
1576 5748 : diff += bytepos1 - bytepos2;
1577 : /* The offset between the 2 need to be 0. */
1578 5748 : if (!known_eq (diff, 0))
1579 : return false;
1580 : return true;
1581 : }
1582 :
1583 : /* Helper function for optimize_agr_copyprop.
1584 : For aggregate copies in USE_STMT, see if DEST
1585 : is on the lhs of USE_STMT and replace it with SRC. */
1586 : static void
1587 966785 : optimize_agr_copyprop_1 (gimple *stmt, gimple *use_stmt,
1588 : tree dest, tree src)
1589 : {
1590 966785 : gcc_assert (gimple_assign_load_p (use_stmt)
1591 : && gimple_store_p (use_stmt));
1592 1933570 : if (gimple_has_volatile_ops (use_stmt))
1593 589270 : return;
1594 966784 : tree dest2 = gimple_assign_lhs (use_stmt);
1595 966784 : tree src2 = gimple_assign_rhs1 (use_stmt);
1596 : /* If the new store is `src2 = src2;` skip over it. */
1597 966784 : if (same_for_assignment (src2, dest2))
1598 : return;
1599 966221 : src = new_src_based_on_copy (src2, dest, src);
1600 966221 : if (!src)
1601 : return;
1602 : /* For 2 memory refences and using a temporary to do the copy,
1603 : don't remove the temporary as the 2 memory references might overlap.
1604 : Note t does not need to be decl as it could be field.
1605 : See PR 22237 for full details.
1606 : E.g.
1607 : t = *a; #DEST = SRC;
1608 : *b = t; #DEST2 = SRC2;
1609 : Cannot be convert into
1610 : t = *a;
1611 : *b = *a;
1612 : Though the following is allowed to be done:
1613 : t = *a;
1614 : *a = t;
1615 : And convert it into:
1616 : t = *a;
1617 : *a = *a;
1618 : */
1619 402272 : if (!operand_equal_p (dest2, src, 0)
1620 402272 : && !DECL_P (dest2) && !DECL_P (src))
1621 : {
1622 : /* If *a and *b have the same base see if
1623 : the offset between the two is greater than
1624 : or equal to the size of the type. */
1625 27692 : poly_int64 offset1, offset2;
1626 27692 : tree len = TYPE_SIZE_UNIT (TREE_TYPE (src));
1627 27692 : if (len == NULL_TREE
1628 27692 : || !tree_fits_poly_int64_p (len))
1629 24757 : return;
1630 27692 : tree base1 = get_addr_base_and_unit_offset (dest2, &offset1);
1631 27692 : tree base2 = get_addr_base_and_unit_offset (src, &offset2);
1632 27692 : poly_int64 size = tree_to_poly_int64 (len);
1633 : /* If the bases are 2 different decls,
1634 : then there can be no overlapping. */
1635 27692 : if (base1 && base2
1636 27642 : && DECL_P (base1) && DECL_P (base2)
1637 1669 : && base1 != base2)
1638 : ;
1639 : /* If we can't figure out the base or the bases are
1640 : not equal then fall back to an alignment check. */
1641 26237 : else if (!base1
1642 26237 : || !base2
1643 26237 : || !operand_equal_p (base1, base2))
1644 : {
1645 25880 : unsigned int align1 = get_object_alignment (src);
1646 25880 : unsigned int align2 = get_object_alignment (dest2);
1647 25880 : align1 /= BITS_PER_UNIT;
1648 25880 : align2 /= BITS_PER_UNIT;
1649 : /* If the alignment of either object is less
1650 : than the size then there is a possibility
1651 : of overlapping. */
1652 25880 : if (maybe_lt (align1, size)
1653 25880 : || maybe_lt (align2, size))
1654 24757 : return;
1655 : }
1656 : /* Make sure [offset1, offset1 + len - 1] does
1657 : not overlap with [offset2, offset2 + len - 1],
1658 : it is ok if they are at the same location though. */
1659 357 : else if (ranges_maybe_overlap_p (offset1, size, offset2, size)
1660 357 : && !known_eq (offset2, offset1))
1661 : return;
1662 : }
1663 :
1664 377515 : if (dump_file && (dump_flags & TDF_DETAILS))
1665 : {
1666 11 : fprintf (dump_file, "Simplified\n ");
1667 11 : print_gimple_stmt (dump_file, use_stmt, 0, dump_flags);
1668 11 : fprintf (dump_file, "after previous\n ");
1669 11 : print_gimple_stmt (dump_file, stmt, 0, dump_flags);
1670 : }
1671 377515 : gimple *orig_stmt = use_stmt;
1672 377515 : gimple_stmt_iterator gsi = gsi_for_stmt (use_stmt);
1673 377515 : gimple_assign_set_rhs_from_tree (&gsi, unshare_expr (src));
1674 377515 : update_stmt (use_stmt);
1675 :
1676 377515 : if (dump_file && (dump_flags & TDF_DETAILS))
1677 : {
1678 11 : fprintf (dump_file, "into\n ");
1679 11 : print_gimple_stmt (dump_file, use_stmt, 0, dump_flags);
1680 : }
1681 377515 : if (maybe_clean_or_replace_eh_stmt (orig_stmt, use_stmt))
1682 0 : bitmap_set_bit (to_purge, gimple_bb (stmt)->index);
1683 377515 : statistics_counter_event (cfun, "copy prop for aggregate", 1);
1684 : }
1685 :
1686 : /* Helper function for optimize_agr_copyprop_1, propagate aggregates
1687 : into the arguments of USE_STMT if the argument matches with DEST;
1688 : replacing it with SRC. */
1689 : static void
1690 685516 : optimize_agr_copyprop_arg (gimple *defstmt, gcall *call,
1691 : tree dest, tree src)
1692 : {
1693 685516 : bool changed = false;
1694 2271149 : for (unsigned arg = 0; arg < gimple_call_num_args (call); arg++)
1695 : {
1696 1585633 : tree *argptr = gimple_call_arg_ptr (call, arg);
1697 2984446 : if (TREE_CODE (*argptr) == SSA_NAME
1698 908815 : || is_gimple_min_invariant (*argptr)
1699 1772453 : || TYPE_VOLATILE (TREE_TYPE (*argptr)))
1700 1398813 : continue;
1701 186820 : tree newsrc = new_src_based_on_copy (*argptr, dest, src);
1702 186820 : if (!newsrc)
1703 112540 : continue;
1704 :
1705 74280 : if (dump_file && (dump_flags & TDF_DETAILS))
1706 : {
1707 9 : fprintf (dump_file, "Simplified\n ");
1708 9 : print_gimple_stmt (dump_file, call, 0, dump_flags);
1709 9 : fprintf (dump_file, "after previous\n ");
1710 9 : print_gimple_stmt (dump_file, defstmt, 0, dump_flags);
1711 : }
1712 74280 : *argptr = unshare_expr (newsrc);
1713 74280 : changed = true;
1714 74280 : if (dump_file && (dump_flags & TDF_DETAILS))
1715 : {
1716 9 : fprintf (dump_file, "into\n ");
1717 9 : print_gimple_stmt (dump_file, call, 0, dump_flags);
1718 : }
1719 : }
1720 685516 : if (changed)
1721 74103 : update_stmt (call);
1722 685516 : }
1723 :
1724 : /* Helper function for optimize_agr_copyprop, propagate aggregates
1725 : into the return stmt USE if the operand of the return matches DEST;
1726 : replacing it with SRC. */
1727 : static void
1728 116888 : optimize_agr_copyprop_return (gimple *defstmt, greturn *use,
1729 : tree dest, tree src)
1730 : {
1731 116888 : tree rvalue = gimple_return_retval (use);
1732 116888 : if (!rvalue
1733 74674 : || TREE_CODE (rvalue) == SSA_NAME
1734 66440 : || is_gimple_min_invariant (rvalue)
1735 182933 : || TYPE_VOLATILE (TREE_TYPE (rvalue)))
1736 50844 : return;
1737 :
1738 : /* `return <retval>;` is already the best it could be.
1739 : Likewise `return *<retval>_N(D)`. */
1740 66044 : if (TREE_CODE (rvalue) == RESULT_DECL
1741 66044 : || (TREE_CODE (rvalue) == MEM_REF
1742 0 : && TREE_CODE (TREE_OPERAND (rvalue, 0)) == SSA_NAME
1743 0 : && TREE_CODE (SSA_NAME_VAR (TREE_OPERAND (rvalue, 0)))
1744 : == RESULT_DECL))
1745 : return;
1746 45155 : tree newsrc = new_src_based_on_copy (rvalue, dest, src);
1747 45155 : if (!newsrc)
1748 : return;
1749 : /* Currently only support non-global vars.
1750 : See PR 124099 on enumtls not supporting expanding for GIMPLE_RETURN.
1751 : FIXME: could support VCEs too? */
1752 44942 : if (!VAR_P (newsrc) || is_global_var (newsrc))
1753 : return;
1754 20308 : if (dump_file && (dump_flags & TDF_DETAILS))
1755 : {
1756 1 : fprintf (dump_file, "Simplified\n ");
1757 1 : print_gimple_stmt (dump_file, use, 0, dump_flags);
1758 1 : fprintf (dump_file, "after previous\n ");
1759 1 : print_gimple_stmt (dump_file, defstmt, 0, dump_flags);
1760 : }
1761 20308 : gimple_return_set_retval (use, newsrc);
1762 20308 : if (dump_file && (dump_flags & TDF_DETAILS))
1763 : {
1764 1 : fprintf (dump_file, "into\n ");
1765 1 : print_gimple_stmt (dump_file, use, 0, dump_flags);
1766 : }
1767 20308 : update_stmt (use);
1768 : }
1769 :
1770 : /* Optimizes
1771 : DEST = SRC;
1772 : DEST2 = DEST; # DEST2 = SRC2;
1773 : into
1774 : DEST = SRC;
1775 : DEST2 = SRC;
1776 : STMT is the first statement and SRC is the common
1777 : between the statements.
1778 :
1779 : Also optimizes:
1780 : DEST = SRC;
1781 : call_func(..., DEST, ...);
1782 : into:
1783 : DEST = SRC;
1784 : call_func(..., SRC, ...);
1785 :
1786 : */
1787 : static void
1788 3711510 : optimize_agr_copyprop (gimple *stmt)
1789 : {
1790 7423020 : if (gimple_has_volatile_ops (stmt))
1791 411424 : return;
1792 :
1793 : /* Can't prop if the statement could throw. */
1794 3710371 : if (stmt_could_throw_p (cfun, stmt))
1795 : return;
1796 :
1797 3302675 : tree dest = gimple_assign_lhs (stmt);
1798 3302675 : tree src = gimple_assign_rhs1 (stmt);
1799 : /* If the statement is `src = src;` then ignore it. */
1800 3302675 : if (same_for_assignment (dest, src))
1801 : return;
1802 :
1803 3300086 : tree vdef = gimple_vdef (stmt);
1804 3300086 : imm_use_iterator iter;
1805 3300086 : gimple *use_stmt;
1806 12870295 : FOR_EACH_IMM_USE_STMT (use_stmt, iter, vdef)
1807 : {
1808 6270123 : if (gimple_assign_load_p (use_stmt)
1809 6270123 : && gimple_store_p (use_stmt))
1810 966785 : optimize_agr_copyprop_1 (stmt, use_stmt, dest, src);
1811 5303338 : else if (is_gimple_call (use_stmt))
1812 685516 : optimize_agr_copyprop_arg (stmt, as_a<gcall*>(use_stmt), dest, src);
1813 4617822 : else if (is_a<greturn*> (use_stmt))
1814 116888 : optimize_agr_copyprop_return (stmt, as_a<greturn*>(use_stmt), dest, src);
1815 3300086 : }
1816 : }
1817 :
1818 : /* Simple DSE of the lhs from a clobber STMT.
1819 : This is used mostly to clean up from optimize_agr_copyprop and
1820 : to remove (exactly one) extra copy that might later on confuse SRA.
1821 : An example is:
1822 : ;; write to a and such.
1823 : b = a; // This statement is to be removed
1824 : b = {CLOBBER};
1825 : SRA will totally scalarize b (which means also a) here for the extra copy
1826 : which is not something welcomed. So removing the copy will
1827 : allow SRA to move the scalarization of a further down or not at all.
1828 : */
1829 : static void
1830 6767334 : do_simple_agr_dse (gassign *stmt, bool full_walk)
1831 : {
1832 : /* Don't do this while in -Og as we want to keep around the copy
1833 : for debuggability. */
1834 6767334 : if (optimize_debug)
1835 4703042 : return;
1836 6763913 : ao_ref read;
1837 6763913 : basic_block bb = gimple_bb (stmt);
1838 6763913 : tree lhs = gimple_assign_lhs (stmt);
1839 : /* Only handle clobbers of a full decl. */
1840 6763913 : if (!DECL_P (lhs))
1841 : return;
1842 6062310 : ao_ref_init (&read, lhs);
1843 6062310 : tree vuse = gimple_vuse (stmt);
1844 6062310 : unsigned limit = full_walk ? param_sccvn_max_alias_queries_per_access : 4;
1845 15451746 : while (limit)
1846 : {
1847 15441289 : gimple *ostmt = SSA_NAME_DEF_STMT (vuse);
1848 : /* Don't handle phis, just declare to be done. */
1849 15441289 : if (is_a<gphi*>(ostmt) || gimple_nop_p (ostmt))
1850 : break;
1851 13387454 : basic_block obb = gimple_bb (ostmt);
1852 : /* If the clobber is not fully dominating the statement define,
1853 : then it is not "simple" to detect if the define is fully clobbered. */
1854 13387454 : if (obb != bb && !dominated_by_p (CDI_DOMINATORS, bb, obb))
1855 3998018 : return;
1856 13387454 : gimple *use_stmt;
1857 13387454 : imm_use_iterator iter;
1858 54137867 : FOR_EACH_IMM_USE_STMT (use_stmt, iter, gimple_vdef (ostmt))
1859 : {
1860 15749638 : basic_block ubb = gimple_bb (use_stmt);
1861 15749638 : if (stmt == use_stmt)
1862 4607193 : continue;
1863 : /* If the use is a clobber for lhs,
1864 : then it can be safely skipped; this happens with eh
1865 : and sometimes jump threading. */
1866 11142445 : if (gimple_clobber_p (use_stmt)
1867 11142445 : && lhs == gimple_assign_lhs (use_stmt))
1868 158540 : continue;
1869 : /* If the use is a phi and it is single use then check if that single use
1870 : is a clobber and lhs is the same. */
1871 10983905 : if (gphi *use_phi = dyn_cast<gphi*>(use_stmt))
1872 : {
1873 321133 : use_operand_p ou;
1874 321133 : gimple *ostmt;
1875 321133 : if (single_imm_use (gimple_phi_result (use_phi), &ou, &ostmt)
1876 273256 : && gimple_clobber_p (ostmt)
1877 536359 : && lhs == gimple_assign_lhs (ostmt))
1878 64940 : continue;
1879 : /* A phi node will never be dominating the clobber. */
1880 256193 : return;
1881 : }
1882 : /* The use needs to be dominating the clobber. */
1883 1331476 : if ((ubb != bb && !dominated_by_p (CDI_DOMINATORS, bb, ubb))
1884 11333618 : || ref_maybe_used_by_stmt_p (use_stmt, &read, false))
1885 1082304 : return;
1886 : /* Count the above alias lookup towards the limit. */
1887 9580468 : limit--;
1888 9580468 : if (limit == 0)
1889 : return;
1890 1774133 : }
1891 11613321 : vuse = gimple_vuse (ostmt);
1892 : /* This is a call with an assignment to the clobber decl,
1893 : remove the lhs or the whole stmt if it was pure/const. */
1894 11613321 : if (is_a <gcall*>(ostmt)
1895 11613321 : && lhs == gimple_call_lhs (ostmt))
1896 : {
1897 : /* Don't remove stores/statements that are needed for non-call
1898 : eh to work. */
1899 3191 : if (stmt_unremovable_because_of_non_call_eh_p (cfun, ostmt))
1900 : return;
1901 : /* If we delete a stmt that could throw, mark the block
1902 : in to_purge to cleanup afterwards. */
1903 3185 : if (stmt_could_throw_p (cfun, ostmt))
1904 907 : bitmap_set_bit (to_purge, obb->index);
1905 3185 : int flags = gimple_call_flags (ostmt);
1906 3185 : if ((flags & (ECF_PURE|ECF_CONST|ECF_NOVOPS))
1907 225 : && !(flags & (ECF_LOOPING_CONST_OR_PURE)))
1908 : {
1909 141 : gimple_stmt_iterator gsi = gsi_for_stmt (ostmt);
1910 141 : if (dump_file && (dump_flags & TDF_DETAILS))
1911 : {
1912 14 : fprintf (dump_file, "Removing dead call store stmt ");
1913 14 : print_gimple_stmt (dump_file, ostmt, 0);
1914 14 : fprintf (dump_file, "\n");
1915 : }
1916 141 : unlink_stmt_vdef (ostmt);
1917 141 : release_defs (ostmt);
1918 141 : gsi_remove (&gsi, true);
1919 141 : statistics_counter_event (cfun, "delete call dead store", 1);
1920 : /* Only remove the first store previous statement. */
1921 141 : return;
1922 : }
1923 : /* Make sure we do not remove a return slot we cannot reconstruct
1924 : later. */
1925 3044 : if (gimple_call_return_slot_opt_p (as_a <gcall *>(ostmt))
1926 3044 : && (TREE_ADDRESSABLE (TREE_TYPE (gimple_call_fntype (ostmt)))
1927 440 : || !poly_int_tree_p
1928 440 : (TYPE_SIZE (TREE_TYPE (gimple_call_fntype (ostmt))))))
1929 : return;
1930 545 : if (dump_file && (dump_flags & TDF_DETAILS))
1931 : {
1932 6 : fprintf (dump_file, "Removing lhs of call stmt ");
1933 6 : print_gimple_stmt (dump_file, ostmt, 0);
1934 6 : fprintf (dump_file, "\n");
1935 : }
1936 545 : gimple_call_set_lhs (ostmt, NULL_TREE);
1937 545 : update_stmt (ostmt);
1938 545 : statistics_counter_event (cfun, "removed lhs call", 1);
1939 545 : return;
1940 : }
1941 : /* This an assignment store to the clobbered decl,
1942 : then maybe remove it. */
1943 11610130 : if (is_a <gassign*>(ostmt)
1944 9788962 : && gimple_store_p (ostmt)
1945 9788962 : && !gimple_clobber_p (ostmt)
1946 14497999 : && lhs == gimple_assign_lhs (ostmt))
1947 : {
1948 : /* Don't remove stores/statements that are needed for non-call
1949 : eh to work. */
1950 151609 : if (stmt_unremovable_because_of_non_call_eh_p (cfun, ostmt))
1951 : return;
1952 : /* If we delete a stmt that could throw, mark the block
1953 : in to_purge to cleanup afterwards. */
1954 146525 : if (stmt_could_throw_p (cfun, ostmt))
1955 0 : bitmap_set_bit (to_purge, obb->index);
1956 146525 : gimple_stmt_iterator gsi = gsi_for_stmt (ostmt);
1957 146525 : if (dump_file && (dump_flags & TDF_DETAILS))
1958 : {
1959 12 : fprintf (dump_file, "Removing dead store stmt ");
1960 12 : print_gimple_stmt (dump_file, ostmt, 0);
1961 12 : fprintf (dump_file, "\n");
1962 : }
1963 146525 : unlink_stmt_vdef (ostmt);
1964 146525 : release_defs (ostmt);
1965 146525 : gsi_remove (&gsi, true);
1966 146525 : statistics_counter_event (cfun, "delete dead store", 1);
1967 : /* Only remove the first store previous statement. */
1968 146525 : return;
1969 : }
1970 : /* If the statement uses or maybe writes to the decl,
1971 : then nothing is to be removed. Don't know if the write
1972 : to the decl is partial write or a full one so the need
1973 : to stop.
1974 : e.g.
1975 : b.c = a;
1976 : Easier to stop here rather than do a full partial
1977 : dse of this statement.
1978 : b = {CLOBBER}; */
1979 11458521 : if (stmt_may_clobber_ref_p_1 (ostmt, &read, false)
1980 11458521 : || ref_maybe_used_by_stmt_p (ostmt, &read, false))
1981 2069085 : return;
1982 9389436 : limit--;
1983 : }
1984 : }
1985 :
1986 : /* Optimizes builtin memcmps for small constant sizes.
1987 : GSI_P is the GSI for the call. STMT is the call itself.
1988 : */
1989 :
1990 : static bool
1991 462827 : simplify_builtin_memcmp (gimple_stmt_iterator *gsi_p, gcall *stmt)
1992 : {
1993 : /* Make sure memcmp arguments are the correct type. */
1994 462827 : if (gimple_call_num_args (stmt) != 3)
1995 : return false;
1996 462827 : tree arg1 = gimple_call_arg (stmt, 0);
1997 462827 : tree arg2 = gimple_call_arg (stmt, 1);
1998 462827 : tree len = gimple_call_arg (stmt, 2);
1999 :
2000 462827 : if (!POINTER_TYPE_P (TREE_TYPE (arg1)))
2001 : return false;
2002 462827 : if (!POINTER_TYPE_P (TREE_TYPE (arg2)))
2003 : return false;
2004 462827 : if (!INTEGRAL_TYPE_P (TREE_TYPE (len)))
2005 : return false;
2006 :
2007 : /* The return value of the memcmp has to be used
2008 : equality comparison to zero. */
2009 462827 : tree res = gimple_call_lhs (stmt);
2010 :
2011 462827 : if (!res || !use_in_zero_equality (res))
2012 13932 : return false;
2013 :
2014 448895 : unsigned HOST_WIDE_INT leni;
2015 :
2016 448895 : if (tree_fits_uhwi_p (len)
2017 625396 : && (leni = tree_to_uhwi (len)) <= GET_MODE_SIZE (word_mode)
2018 529079 : && pow2p_hwi (leni))
2019 : {
2020 19045 : leni *= CHAR_TYPE_SIZE;
2021 19045 : unsigned align1 = get_pointer_alignment (arg1);
2022 19045 : unsigned align2 = get_pointer_alignment (arg2);
2023 19045 : unsigned align = MIN (align1, align2);
2024 19045 : scalar_int_mode mode;
2025 19045 : if (int_mode_for_size (leni, 1).exists (&mode)
2026 19045 : && (align >= leni || !targetm.slow_unaligned_access (mode, align)))
2027 : {
2028 19045 : location_t loc = gimple_location (stmt);
2029 19045 : tree type, off;
2030 19045 : type = build_nonstandard_integer_type (leni, 1);
2031 38090 : gcc_assert (known_eq (GET_MODE_BITSIZE (TYPE_MODE (type)), leni));
2032 19045 : tree ptrtype = build_pointer_type_for_mode (char_type_node,
2033 : ptr_mode, true);
2034 19045 : off = build_int_cst (ptrtype, 0);
2035 :
2036 : /* Create unaligned types if needed. */
2037 19045 : tree type1 = type, type2 = type;
2038 19045 : if (TYPE_ALIGN (type1) > align1)
2039 7721 : type1 = build_aligned_type (type1, align1);
2040 19045 : if (TYPE_ALIGN (type2) > align2)
2041 8221 : type2 = build_aligned_type (type2, align2);
2042 :
2043 19045 : arg1 = build2_loc (loc, MEM_REF, type1, arg1, off);
2044 19045 : arg2 = build2_loc (loc, MEM_REF, type2, arg2, off);
2045 19045 : tree tem1 = fold_const_aggregate_ref (arg1);
2046 19045 : if (tem1)
2047 219 : arg1 = tem1;
2048 19045 : tree tem2 = fold_const_aggregate_ref (arg2);
2049 19045 : if (tem2)
2050 7412 : arg2 = tem2;
2051 19045 : res = fold_convert_loc (loc, TREE_TYPE (res),
2052 : fold_build2_loc (loc, NE_EXPR,
2053 : boolean_type_node,
2054 : arg1, arg2));
2055 19045 : gimplify_and_update_call_from_tree (gsi_p, res);
2056 19045 : return true;
2057 : }
2058 : }
2059 :
2060 : /* Replace memcmp with memcmp_eq if the above fails. */
2061 429850 : if (DECL_FUNCTION_CODE (gimple_call_fndecl (stmt)) == BUILT_IN_MEMCMP_EQ)
2062 : return false;
2063 340554 : if (!fold_before_rtl_expansion_p ())
2064 : return false;
2065 89296 : gimple_call_set_fndecl (stmt, builtin_decl_explicit (BUILT_IN_MEMCMP_EQ));
2066 89296 : update_stmt (stmt);
2067 89296 : return true;
2068 : }
2069 :
2070 : /* Optimizes builtin memchrs for small constant sizes with a const string.
2071 : GSI_P is the GSI for the call. STMT is the call itself.
2072 : */
2073 :
2074 : static bool
2075 15253 : simplify_builtin_memchr (gimple_stmt_iterator *gsi_p, gcall *stmt)
2076 : {
2077 15253 : if (CHAR_BIT != 8 || BITS_PER_UNIT != 8)
2078 : return false;
2079 :
2080 15253 : if (gimple_call_num_args (stmt) != 3)
2081 : return false;
2082 :
2083 15253 : tree res = gimple_call_lhs (stmt);
2084 15253 : if (!res || !use_in_zero_equality (res))
2085 13828 : return false;
2086 :
2087 1425 : tree ptr = gimple_call_arg (stmt, 0);
2088 1425 : if (TREE_CODE (ptr) != ADDR_EXPR
2089 1425 : || TREE_CODE (TREE_OPERAND (ptr, 0)) != STRING_CST)
2090 : return false;
2091 :
2092 422 : unsigned HOST_WIDE_INT slen
2093 422 : = TREE_STRING_LENGTH (TREE_OPERAND (ptr, 0));
2094 : /* It must be a non-empty string constant. */
2095 422 : if (slen < 2)
2096 : return false;
2097 :
2098 : /* For -Os, only simplify strings with a single character. */
2099 418 : if (!optimize_bb_for_speed_p (gimple_bb (stmt))
2100 418 : && slen > 2)
2101 : return false;
2102 :
2103 402 : tree size = gimple_call_arg (stmt, 2);
2104 : /* Size must be a constant which is <= UNITS_PER_WORD and
2105 : <= the string length. */
2106 402 : if (!tree_fits_uhwi_p (size))
2107 : return false;
2108 :
2109 402 : unsigned HOST_WIDE_INT sz = tree_to_uhwi (size);
2110 403 : if (sz == 0 || sz > UNITS_PER_WORD || sz >= slen)
2111 : return false;
2112 :
2113 350 : tree ch = gimple_call_arg (stmt, 1);
2114 350 : location_t loc = gimple_location (stmt);
2115 350 : if (!useless_type_conversion_p (char_type_node,
2116 350 : TREE_TYPE (ch)))
2117 350 : ch = fold_convert_loc (loc, char_type_node, ch);
2118 350 : const char *p = TREE_STRING_POINTER (TREE_OPERAND (ptr, 0));
2119 350 : unsigned int isize = sz;
2120 350 : tree *op = XALLOCAVEC (tree, isize);
2121 1261 : for (unsigned int i = 0; i < isize; i++)
2122 : {
2123 911 : op[i] = build_int_cst (char_type_node, p[i]);
2124 911 : op[i] = fold_build2_loc (loc, EQ_EXPR, boolean_type_node,
2125 : op[i], ch);
2126 : }
2127 911 : for (unsigned int i = isize - 1; i >= 1; i--)
2128 561 : op[i - 1] = fold_convert_loc (loc, boolean_type_node,
2129 : fold_build2_loc (loc,
2130 : BIT_IOR_EXPR,
2131 : boolean_type_node,
2132 561 : op[i - 1],
2133 561 : op[i]));
2134 350 : res = fold_convert_loc (loc, TREE_TYPE (res), op[0]);
2135 350 : gimplify_and_update_call_from_tree (gsi_p, res);
2136 350 : return true;
2137 : }
2138 :
2139 : /* *GSI_P is a GIMPLE_CALL to a builtin function.
2140 : Optimize
2141 : memcpy (p, "abcd", 4); // STMT1
2142 : memset (p + 4, ' ', 3); // STMT2
2143 : into
2144 : memcpy (p, "abcd ", 7);
2145 : call if the latter can be stored by pieces during expansion.
2146 : */
2147 :
2148 : static bool
2149 110847 : simplify_builtin_memcpy_memset (gimple_stmt_iterator *gsi_p, gcall *stmt2)
2150 : {
2151 110847 : if (gimple_call_num_args (stmt2) != 3
2152 110847 : || gimple_call_lhs (stmt2)
2153 : || CHAR_BIT != 8
2154 110847 : || BITS_PER_UNIT != 8)
2155 : return false;
2156 :
2157 211873 : tree vuse = gimple_vuse (stmt2);
2158 103370 : if (vuse == NULL)
2159 : return false;
2160 103370 : gimple *stmt1 = SSA_NAME_DEF_STMT (vuse);
2161 :
2162 103370 : tree callee1;
2163 103370 : tree ptr1, src1, str1, off1, len1, lhs1;
2164 103370 : tree ptr2 = gimple_call_arg (stmt2, 0);
2165 103370 : tree val2 = gimple_call_arg (stmt2, 1);
2166 103370 : tree len2 = gimple_call_arg (stmt2, 2);
2167 103370 : tree diff, vdef, new_str_cst;
2168 103370 : gimple *use_stmt;
2169 103370 : unsigned int ptr1_align;
2170 103370 : unsigned HOST_WIDE_INT src_len;
2171 103370 : char *src_buf;
2172 103370 : use_operand_p use_p;
2173 :
2174 103370 : if (!tree_fits_shwi_p (val2)
2175 99147 : || !tree_fits_uhwi_p (len2)
2176 165363 : || compare_tree_int (len2, 1024) == 1)
2177 46460 : return false;
2178 :
2179 56910 : if (is_gimple_call (stmt1))
2180 : {
2181 : /* If first stmt is a call, it needs to be memcpy
2182 : or mempcpy, with string literal as second argument and
2183 : constant length. */
2184 29451 : callee1 = gimple_call_fndecl (stmt1);
2185 29451 : if (callee1 == NULL_TREE
2186 29335 : || !fndecl_built_in_p (callee1, BUILT_IN_NORMAL)
2187 55345 : || gimple_call_num_args (stmt1) != 3)
2188 : return false;
2189 24608 : if (DECL_FUNCTION_CODE (callee1) != BUILT_IN_MEMCPY
2190 24608 : && DECL_FUNCTION_CODE (callee1) != BUILT_IN_MEMPCPY)
2191 : return false;
2192 10867 : ptr1 = gimple_call_arg (stmt1, 0);
2193 10867 : src1 = gimple_call_arg (stmt1, 1);
2194 10867 : len1 = gimple_call_arg (stmt1, 2);
2195 10867 : lhs1 = gimple_call_lhs (stmt1);
2196 10867 : if (!tree_fits_uhwi_p (len1))
2197 : return false;
2198 10780 : str1 = string_constant (src1, &off1, NULL, NULL);
2199 10780 : if (str1 == NULL_TREE)
2200 : return false;
2201 4875 : if (!tree_fits_uhwi_p (off1)
2202 4875 : || compare_tree_int (off1, TREE_STRING_LENGTH (str1) - 1) > 0
2203 4875 : || compare_tree_int (len1, TREE_STRING_LENGTH (str1)
2204 4875 : - tree_to_uhwi (off1)) > 0
2205 4875 : || TREE_CODE (TREE_TYPE (str1)) != ARRAY_TYPE
2206 14625 : || TYPE_MODE (TREE_TYPE (TREE_TYPE (str1)))
2207 4875 : != TYPE_MODE (char_type_node))
2208 0 : return false;
2209 : }
2210 27459 : else if (gimple_assign_single_p (stmt1))
2211 : {
2212 : /* Otherwise look for length 1 memcpy optimized into
2213 : assignment. */
2214 16753 : ptr1 = gimple_assign_lhs (stmt1);
2215 16753 : src1 = gimple_assign_rhs1 (stmt1);
2216 16753 : if (TREE_CODE (ptr1) != MEM_REF
2217 3512 : || TYPE_MODE (TREE_TYPE (ptr1)) != TYPE_MODE (char_type_node)
2218 17787 : || !tree_fits_shwi_p (src1))
2219 16401 : return false;
2220 352 : ptr1 = build_fold_addr_expr (ptr1);
2221 352 : STRIP_USELESS_TYPE_CONVERSION (ptr1);
2222 352 : callee1 = NULL_TREE;
2223 352 : len1 = size_one_node;
2224 352 : lhs1 = NULL_TREE;
2225 352 : off1 = size_zero_node;
2226 352 : str1 = NULL_TREE;
2227 : }
2228 : else
2229 : return false;
2230 :
2231 5227 : diff = constant_pointer_difference (ptr1, ptr2);
2232 5227 : if (diff == NULL && lhs1 != NULL)
2233 : {
2234 7 : diff = constant_pointer_difference (lhs1, ptr2);
2235 7 : if (DECL_FUNCTION_CODE (callee1) == BUILT_IN_MEMPCPY
2236 7 : && diff != NULL)
2237 7 : diff = size_binop (PLUS_EXPR, diff,
2238 : fold_convert (sizetype, len1));
2239 : }
2240 : /* If the difference between the second and first destination pointer
2241 : is not constant, or is bigger than memcpy length, bail out. */
2242 5227 : if (diff == NULL
2243 4344 : || !tree_fits_uhwi_p (diff)
2244 4344 : || tree_int_cst_lt (len1, diff)
2245 9327 : || compare_tree_int (diff, 1024) == 1)
2246 1127 : return false;
2247 :
2248 : /* Use maximum of difference plus memset length and memcpy length
2249 : as the new memcpy length, if it is too big, bail out. */
2250 4100 : src_len = tree_to_uhwi (diff);
2251 4100 : src_len += tree_to_uhwi (len2);
2252 4100 : if (src_len < tree_to_uhwi (len1))
2253 : src_len = tree_to_uhwi (len1);
2254 4100 : if (src_len > 1024)
2255 : return false;
2256 :
2257 : /* If mempcpy value is used elsewhere, bail out, as mempcpy
2258 : with bigger length will return different result. */
2259 4100 : if (lhs1 != NULL_TREE
2260 193 : && DECL_FUNCTION_CODE (callee1) == BUILT_IN_MEMPCPY
2261 4107 : && (TREE_CODE (lhs1) != SSA_NAME
2262 7 : || !single_imm_use (lhs1, &use_p, &use_stmt)
2263 7 : || use_stmt != stmt2))
2264 0 : return false;
2265 :
2266 : /* If anything reads memory in between memcpy and memset
2267 : call, the modified memcpy call might change it. */
2268 4100 : vdef = gimple_vdef (stmt1);
2269 4100 : if (vdef != NULL
2270 4100 : && (!single_imm_use (vdef, &use_p, &use_stmt)
2271 3345 : || use_stmt != stmt2))
2272 : return false;
2273 :
2274 3345 : ptr1_align = get_pointer_alignment (ptr1);
2275 : /* Construct the new source string literal. */
2276 3345 : src_buf = XALLOCAVEC (char, src_len + 1);
2277 3345 : if (callee1)
2278 3184 : memcpy (src_buf,
2279 3184 : TREE_STRING_POINTER (str1) + tree_to_uhwi (off1),
2280 : tree_to_uhwi (len1));
2281 : else
2282 161 : src_buf[0] = tree_to_shwi (src1);
2283 3345 : memset (src_buf + tree_to_uhwi (diff),
2284 3345 : tree_to_shwi (val2), tree_to_uhwi (len2));
2285 3345 : src_buf[src_len] = '\0';
2286 : /* Neither builtin_strncpy_read_str nor builtin_memcpy_read_str
2287 : handle embedded '\0's. */
2288 3345 : if (strlen (src_buf) != src_len)
2289 : return false;
2290 3255 : rtl_profile_for_bb (gimple_bb (stmt2));
2291 : /* If the new memcpy wouldn't be emitted by storing the literal
2292 : by pieces, this optimization might enlarge .rodata too much,
2293 : as commonly used string literals couldn't be shared any
2294 : longer. */
2295 3255 : if (!can_store_by_pieces (src_len,
2296 : builtin_strncpy_read_str,
2297 : src_buf, ptr1_align, false))
2298 : return false;
2299 :
2300 2471 : new_str_cst = build_string_literal (src_len, src_buf);
2301 2471 : if (callee1)
2302 : {
2303 : /* If STMT1 is a mem{,p}cpy call, adjust it and remove
2304 : memset call. */
2305 2344 : if (lhs1 && DECL_FUNCTION_CODE (callee1) == BUILT_IN_MEMPCPY)
2306 7 : gimple_call_set_lhs (stmt1, NULL_TREE);
2307 2344 : gimple_call_set_arg (stmt1, 1, new_str_cst);
2308 2344 : gimple_call_set_arg (stmt1, 2,
2309 2344 : build_int_cst (TREE_TYPE (len1), src_len));
2310 2344 : update_stmt (stmt1);
2311 2344 : unlink_stmt_vdef (stmt2);
2312 2344 : gsi_replace (gsi_p, gimple_build_nop (), false);
2313 2344 : fwprop_invalidate_lattice (gimple_get_lhs (stmt2));
2314 2344 : release_defs (stmt2);
2315 2344 : if (lhs1 && DECL_FUNCTION_CODE (callee1) == BUILT_IN_MEMPCPY)
2316 : {
2317 7 : fwprop_invalidate_lattice (lhs1);
2318 7 : release_ssa_name (lhs1);
2319 : }
2320 2344 : return true;
2321 : }
2322 : else
2323 : {
2324 : /* Otherwise, if STMT1 is length 1 memcpy optimized into
2325 : assignment, remove STMT1 and change memset call into
2326 : memcpy call. */
2327 127 : gimple_stmt_iterator gsi = gsi_for_stmt (stmt1);
2328 :
2329 127 : if (!is_gimple_val (ptr1))
2330 12 : ptr1 = force_gimple_operand_gsi (gsi_p, ptr1, true, NULL_TREE,
2331 : true, GSI_SAME_STMT);
2332 127 : tree fndecl = builtin_decl_explicit (BUILT_IN_MEMCPY);
2333 127 : gimple_call_set_fndecl (stmt2, fndecl);
2334 127 : gimple_call_set_fntype (stmt2,
2335 127 : TREE_TYPE (fndecl));
2336 127 : gimple_call_set_arg (stmt2, 0, ptr1);
2337 127 : gimple_call_set_arg (stmt2, 1, new_str_cst);
2338 127 : gimple_call_set_arg (stmt2, 2,
2339 127 : build_int_cst (TREE_TYPE (len2), src_len));
2340 127 : unlink_stmt_vdef (stmt1);
2341 127 : gsi_remove (&gsi, true);
2342 127 : fwprop_invalidate_lattice (gimple_get_lhs (stmt1));
2343 127 : release_defs (stmt1);
2344 127 : update_stmt (stmt2);
2345 127 : return false;
2346 : }
2347 : }
2348 :
2349 :
2350 : /* Try to optimize out __builtin_stack_restore. Optimize it out
2351 : if there is another __builtin_stack_restore in the same basic
2352 : block and no calls or ASM_EXPRs are in between, or if this block's
2353 : only outgoing edge is to EXIT_BLOCK and there are no calls or
2354 : ASM_EXPRs after this __builtin_stack_restore.
2355 : Note restore right before a noreturn function is not needed.
2356 : And skip some cheap calls that will most likely become an instruction.
2357 : Restoring the stack before a call is important to be able to keep
2358 : stack usage down so that call does not run out of stack. */
2359 :
2360 :
2361 : static bool
2362 10378 : optimize_stack_restore (gimple_stmt_iterator *gsi, gimple *call)
2363 : {
2364 10378 : if (!fold_before_rtl_expansion_p ())
2365 : return false;
2366 2530 : tree callee;
2367 2530 : gimple *stmt;
2368 :
2369 2530 : basic_block bb = gsi_bb (*gsi);
2370 :
2371 2530 : if (gimple_call_num_args (call) != 1
2372 2530 : || TREE_CODE (gimple_call_arg (call, 0)) != SSA_NAME
2373 5060 : || !POINTER_TYPE_P (TREE_TYPE (gimple_call_arg (call, 0))))
2374 : return false;
2375 :
2376 2530 : gimple_stmt_iterator i = *gsi;
2377 6341 : for (gsi_next (&i); !gsi_end_p (i); gsi_next (&i))
2378 : {
2379 4272 : stmt = gsi_stmt (i);
2380 4272 : if (is_a<gasm*> (stmt))
2381 : return false;
2382 4271 : gcall *call = dyn_cast<gcall*>(stmt);
2383 4271 : if (!call)
2384 3602 : continue;
2385 :
2386 : /* We can remove the restore in front of noreturn
2387 : calls. Since the restore will happen either
2388 : via an unwind/longjmp or not at all. */
2389 669 : if (gimple_call_noreturn_p (call))
2390 : break;
2391 :
2392 : /* Internal calls are ok, to bypass
2393 : check first since fndecl will be null. */
2394 653 : if (gimple_call_internal_p (call))
2395 1 : continue;
2396 :
2397 652 : callee = gimple_call_fndecl (call);
2398 : /* Non-builtin calls are not ok. */
2399 652 : if (!callee
2400 652 : || !fndecl_built_in_p (callee))
2401 : return false;
2402 :
2403 : /* Do not remove stack updates before strub leave. */
2404 576 : if (fndecl_built_in_p (callee, BUILT_IN___STRUB_LEAVE)
2405 : /* Alloca calls are not ok either. */
2406 576 : || fndecl_builtin_alloc_p (callee))
2407 : return false;
2408 :
2409 364 : if (fndecl_built_in_p (callee, BUILT_IN_STACK_RESTORE))
2410 52 : goto second_stack_restore;
2411 :
2412 : /* If not a simple or inexpensive builtin, then it is not ok either. */
2413 312 : if (!is_simple_builtin (callee)
2414 312 : && !is_inexpensive_builtin (callee))
2415 : return false;
2416 : }
2417 :
2418 : /* Allow one successor of the exit block, or zero successors. */
2419 2085 : switch (EDGE_COUNT (bb->succs))
2420 : {
2421 : case 0:
2422 : break;
2423 1998 : case 1:
2424 1998 : if (single_succ_edge (bb)->dest != EXIT_BLOCK_PTR_FOR_FN (cfun))
2425 : return false;
2426 : break;
2427 : default:
2428 : return false;
2429 : }
2430 1728 : second_stack_restore:
2431 :
2432 : /* If there's exactly one use, then zap the call to __builtin_stack_save.
2433 : If there are multiple uses, then the last one should remove the call.
2434 : In any case, whether the call to __builtin_stack_save can be removed
2435 : or not is irrelevant to removing the call to __builtin_stack_restore. */
2436 1728 : if (has_single_use (gimple_call_arg (call, 0)))
2437 : {
2438 1551 : gimple *stack_save = SSA_NAME_DEF_STMT (gimple_call_arg (call, 0));
2439 1551 : if (is_gimple_call (stack_save))
2440 : {
2441 1549 : callee = gimple_call_fndecl (stack_save);
2442 1549 : if (callee && fndecl_built_in_p (callee, BUILT_IN_STACK_SAVE))
2443 : {
2444 1549 : gimple_stmt_iterator stack_save_gsi;
2445 1549 : tree rhs;
2446 :
2447 1549 : stack_save_gsi = gsi_for_stmt (stack_save);
2448 1549 : rhs = build_int_cst (TREE_TYPE (gimple_call_arg (call, 0)), 0);
2449 1549 : replace_call_with_value (&stack_save_gsi, rhs);
2450 : }
2451 : }
2452 : }
2453 :
2454 : /* No effect, so the statement will be deleted. */
2455 1728 : replace_call_with_value (gsi, NULL_TREE);
2456 1728 : return true;
2457 : }
2458 :
2459 : /* *GSI_P is a GIMPLE_CALL to a builtin function.
2460 : Optimize
2461 : memcpy (p, "abcd", 4);
2462 : memset (p + 4, ' ', 3);
2463 : into
2464 : memcpy (p, "abcd ", 7);
2465 : call if the latter can be stored by pieces during expansion.
2466 :
2467 : Optimize
2468 : memchr ("abcd", a, 4) == 0;
2469 : or
2470 : memchr ("abcd", a, 4) != 0;
2471 : to
2472 : (a == 'a' || a == 'b' || a == 'c' || a == 'd') == 0
2473 : or
2474 : (a == 'a' || a == 'b' || a == 'c' || a == 'd') != 0
2475 :
2476 : Also canonicalize __atomic_fetch_op (p, x, y) op x
2477 : to __atomic_op_fetch (p, x, y) or
2478 : __atomic_op_fetch (p, x, y) iop x
2479 : to __atomic_fetch_op (p, x, y) when possible (also __sync). */
2480 :
2481 : static bool
2482 6128230 : simplify_builtin_call (gimple_stmt_iterator *gsi_p, tree callee2, bool full_walk)
2483 : {
2484 6128230 : gimple *stmt2 = gsi_stmt (*gsi_p);
2485 6128230 : enum built_in_function other_atomic = END_BUILTINS;
2486 6128230 : enum tree_code atomic_op = ERROR_MARK;
2487 :
2488 6128230 : switch (DECL_FUNCTION_CODE (callee2))
2489 : {
2490 10378 : case BUILT_IN_STACK_RESTORE:
2491 10378 : return optimize_stack_restore (gsi_p, as_a<gcall*>(stmt2));
2492 462827 : case BUILT_IN_MEMCMP:
2493 462827 : case BUILT_IN_MEMCMP_EQ:
2494 462827 : return simplify_builtin_memcmp (gsi_p, as_a<gcall*>(stmt2));
2495 15253 : case BUILT_IN_MEMCHR:
2496 15253 : return simplify_builtin_memchr (gsi_p, as_a<gcall*>(stmt2));
2497 :
2498 110847 : case BUILT_IN_MEMSET:
2499 110847 : if (gimple_call_num_args (stmt2) == 3)
2500 : {
2501 : /* Try to prop the zeroing/value of the memset to memcpy
2502 : if the dest is an address and the value is a constant. */
2503 110847 : optimize_aggr_zeroprop (stmt2, full_walk);
2504 : }
2505 110847 : return simplify_builtin_memcpy_memset (gsi_p, as_a<gcall*>(stmt2));
2506 :
2507 : #define CASE_ATOMIC(NAME, OTHER, OP) \
2508 : case BUILT_IN_##NAME##_1: \
2509 : case BUILT_IN_##NAME##_2: \
2510 : case BUILT_IN_##NAME##_4: \
2511 : case BUILT_IN_##NAME##_8: \
2512 : case BUILT_IN_##NAME##_16: \
2513 : atomic_op = OP; \
2514 : other_atomic \
2515 : = (enum built_in_function) (BUILT_IN_##OTHER##_1 \
2516 : + (DECL_FUNCTION_CODE (callee2) \
2517 : - BUILT_IN_##NAME##_1)); \
2518 : goto handle_atomic_fetch_op;
2519 :
2520 48981 : CASE_ATOMIC (ATOMIC_FETCH_ADD, ATOMIC_ADD_FETCH, PLUS_EXPR)
2521 7125 : CASE_ATOMIC (ATOMIC_FETCH_SUB, ATOMIC_SUB_FETCH, MINUS_EXPR)
2522 2876 : CASE_ATOMIC (ATOMIC_FETCH_AND, ATOMIC_AND_FETCH, BIT_AND_EXPR)
2523 2895 : CASE_ATOMIC (ATOMIC_FETCH_XOR, ATOMIC_XOR_FETCH, BIT_XOR_EXPR)
2524 3823 : CASE_ATOMIC (ATOMIC_FETCH_OR, ATOMIC_OR_FETCH, BIT_IOR_EXPR)
2525 :
2526 2363 : CASE_ATOMIC (SYNC_FETCH_AND_ADD, SYNC_ADD_AND_FETCH, PLUS_EXPR)
2527 2004 : CASE_ATOMIC (SYNC_FETCH_AND_SUB, SYNC_SUB_AND_FETCH, MINUS_EXPR)
2528 1876 : CASE_ATOMIC (SYNC_FETCH_AND_AND, SYNC_AND_AND_FETCH, BIT_AND_EXPR)
2529 2144 : CASE_ATOMIC (SYNC_FETCH_AND_XOR, SYNC_XOR_AND_FETCH, BIT_XOR_EXPR)
2530 1987 : CASE_ATOMIC (SYNC_FETCH_AND_OR, SYNC_OR_AND_FETCH, BIT_IOR_EXPR)
2531 :
2532 14324 : CASE_ATOMIC (ATOMIC_ADD_FETCH, ATOMIC_FETCH_ADD, MINUS_EXPR)
2533 8528 : CASE_ATOMIC (ATOMIC_SUB_FETCH, ATOMIC_FETCH_SUB, PLUS_EXPR)
2534 2380 : CASE_ATOMIC (ATOMIC_XOR_FETCH, ATOMIC_FETCH_XOR, BIT_XOR_EXPR)
2535 :
2536 821 : CASE_ATOMIC (SYNC_ADD_AND_FETCH, SYNC_FETCH_AND_ADD, MINUS_EXPR)
2537 732 : CASE_ATOMIC (SYNC_SUB_AND_FETCH, SYNC_FETCH_AND_SUB, PLUS_EXPR)
2538 800 : CASE_ATOMIC (SYNC_XOR_AND_FETCH, SYNC_FETCH_AND_XOR, BIT_XOR_EXPR)
2539 :
2540 : #undef CASE_ATOMIC
2541 :
2542 103659 : handle_atomic_fetch_op:
2543 103659 : if (gimple_call_num_args (stmt2) >= 2 && gimple_call_lhs (stmt2))
2544 : {
2545 60238 : tree lhs2 = gimple_call_lhs (stmt2), lhsc = lhs2;
2546 60238 : tree arg = gimple_call_arg (stmt2, 1);
2547 60238 : gimple *use_stmt, *cast_stmt = NULL;
2548 60238 : use_operand_p use_p;
2549 60238 : tree ndecl = builtin_decl_explicit (other_atomic);
2550 :
2551 60238 : if (ndecl == NULL_TREE || !single_imm_use (lhs2, &use_p, &use_stmt))
2552 : break;
2553 :
2554 59109 : if (gimple_assign_cast_p (use_stmt))
2555 : {
2556 31652 : cast_stmt = use_stmt;
2557 31652 : lhsc = gimple_assign_lhs (cast_stmt);
2558 31652 : if (lhsc == NULL_TREE
2559 31652 : || !INTEGRAL_TYPE_P (TREE_TYPE (lhsc))
2560 31101 : || (TYPE_PRECISION (TREE_TYPE (lhsc))
2561 31101 : != TYPE_PRECISION (TREE_TYPE (lhs2)))
2562 61227 : || !single_imm_use (lhsc, &use_p, &use_stmt))
2563 : {
2564 2605 : use_stmt = cast_stmt;
2565 2605 : cast_stmt = NULL;
2566 2605 : lhsc = lhs2;
2567 : }
2568 : }
2569 :
2570 59109 : bool ok = false;
2571 59109 : tree oarg = NULL_TREE;
2572 59109 : enum tree_code ccode = ERROR_MARK;
2573 59109 : tree crhs1 = NULL_TREE, crhs2 = NULL_TREE;
2574 59109 : if (is_gimple_assign (use_stmt)
2575 59109 : && gimple_assign_rhs_code (use_stmt) == atomic_op)
2576 : {
2577 1416 : if (gimple_assign_rhs1 (use_stmt) == lhsc)
2578 1016 : oarg = gimple_assign_rhs2 (use_stmt);
2579 400 : else if (atomic_op != MINUS_EXPR)
2580 : oarg = gimple_assign_rhs1 (use_stmt);
2581 : }
2582 57693 : else if (atomic_op == MINUS_EXPR
2583 13160 : && is_gimple_assign (use_stmt)
2584 3609 : && gimple_assign_rhs_code (use_stmt) == PLUS_EXPR
2585 199 : && TREE_CODE (arg) == INTEGER_CST
2586 57892 : && (TREE_CODE (gimple_assign_rhs2 (use_stmt))
2587 : == INTEGER_CST))
2588 : {
2589 183 : tree a = fold_convert (TREE_TYPE (lhs2), arg);
2590 183 : tree o = fold_convert (TREE_TYPE (lhs2),
2591 : gimple_assign_rhs2 (use_stmt));
2592 183 : if (wi::to_wide (a) == wi::neg (wi::to_wide (o)))
2593 : ok = true;
2594 : }
2595 57510 : else if (atomic_op == BIT_AND_EXPR || atomic_op == BIT_IOR_EXPR)
2596 : ;
2597 52268 : else if (gimple_code (use_stmt) == GIMPLE_COND)
2598 : {
2599 19429 : ccode = gimple_cond_code (use_stmt);
2600 19429 : crhs1 = gimple_cond_lhs (use_stmt);
2601 19429 : crhs2 = gimple_cond_rhs (use_stmt);
2602 : }
2603 32839 : else if (is_gimple_assign (use_stmt))
2604 : {
2605 9522 : if (gimple_assign_rhs_class (use_stmt) == GIMPLE_BINARY_RHS)
2606 : {
2607 3941 : ccode = gimple_assign_rhs_code (use_stmt);
2608 3941 : crhs1 = gimple_assign_rhs1 (use_stmt);
2609 3941 : crhs2 = gimple_assign_rhs2 (use_stmt);
2610 : }
2611 5581 : else if (gimple_assign_rhs_code (use_stmt) == COND_EXPR)
2612 : {
2613 0 : tree cond = gimple_assign_rhs1 (use_stmt);
2614 0 : if (COMPARISON_CLASS_P (cond))
2615 : {
2616 0 : ccode = TREE_CODE (cond);
2617 0 : crhs1 = TREE_OPERAND (cond, 0);
2618 0 : crhs2 = TREE_OPERAND (cond, 1);
2619 : }
2620 : }
2621 : }
2622 24386 : if (ccode == EQ_EXPR || ccode == NE_EXPR)
2623 : {
2624 : /* Deal with x - y == 0 or x ^ y == 0
2625 : being optimized into x == y and x + cst == 0
2626 : into x == -cst. */
2627 22186 : tree o = NULL_TREE;
2628 22186 : if (crhs1 == lhsc)
2629 : o = crhs2;
2630 133 : else if (crhs2 == lhsc)
2631 133 : o = crhs1;
2632 22186 : if (o && atomic_op != PLUS_EXPR)
2633 : oarg = o;
2634 10054 : else if (o
2635 10054 : && TREE_CODE (o) == INTEGER_CST
2636 10054 : && TREE_CODE (arg) == INTEGER_CST)
2637 : {
2638 9344 : tree a = fold_convert (TREE_TYPE (lhs2), arg);
2639 9344 : o = fold_convert (TREE_TYPE (lhs2), o);
2640 9344 : if (wi::to_wide (a) == wi::neg (wi::to_wide (o)))
2641 59109 : ok = true;
2642 : }
2643 : }
2644 59109 : if (oarg && !ok)
2645 : {
2646 13548 : if (operand_equal_p (arg, oarg, 0))
2647 : ok = true;
2648 12220 : else if (TREE_CODE (arg) == SSA_NAME
2649 2179 : && TREE_CODE (oarg) == SSA_NAME)
2650 : {
2651 745 : tree oarg2 = oarg;
2652 745 : if (gimple_assign_cast_p (SSA_NAME_DEF_STMT (oarg)))
2653 : {
2654 104 : gimple *g = SSA_NAME_DEF_STMT (oarg);
2655 104 : oarg2 = gimple_assign_rhs1 (g);
2656 104 : if (TREE_CODE (oarg2) != SSA_NAME
2657 104 : || !INTEGRAL_TYPE_P (TREE_TYPE (oarg2))
2658 208 : || (TYPE_PRECISION (TREE_TYPE (oarg2))
2659 104 : != TYPE_PRECISION (TREE_TYPE (oarg))))
2660 : oarg2 = oarg;
2661 : }
2662 745 : if (gimple_assign_cast_p (SSA_NAME_DEF_STMT (arg)))
2663 : {
2664 544 : gimple *g = SSA_NAME_DEF_STMT (arg);
2665 544 : tree rhs1 = gimple_assign_rhs1 (g);
2666 : /* Handle e.g.
2667 : x.0_1 = (long unsigned int) x_4(D);
2668 : _2 = __atomic_fetch_add_8 (&vlong, x.0_1, 0);
2669 : _3 = (long int) _2;
2670 : _7 = x_4(D) + _3; */
2671 544 : if (rhs1 == oarg || rhs1 == oarg2)
2672 : ok = true;
2673 : /* Handle e.g.
2674 : x.18_1 = (short unsigned int) x_5(D);
2675 : _2 = (int) x.18_1;
2676 : _3 = __atomic_fetch_xor_2 (&vshort, _2, 0);
2677 : _4 = (short int) _3;
2678 : _8 = x_5(D) ^ _4;
2679 : This happens only for char/short. */
2680 160 : else if (TREE_CODE (rhs1) == SSA_NAME
2681 160 : && INTEGRAL_TYPE_P (TREE_TYPE (rhs1))
2682 320 : && (TYPE_PRECISION (TREE_TYPE (rhs1))
2683 160 : == TYPE_PRECISION (TREE_TYPE (lhs2))))
2684 : {
2685 160 : g = SSA_NAME_DEF_STMT (rhs1);
2686 160 : if (gimple_assign_cast_p (g)
2687 160 : && (gimple_assign_rhs1 (g) == oarg
2688 0 : || gimple_assign_rhs1 (g) == oarg2))
2689 : ok = true;
2690 : }
2691 : }
2692 745 : if (!ok && arg == oarg2)
2693 : /* Handle e.g.
2694 : _1 = __sync_fetch_and_add_4 (&v, x_5(D));
2695 : _2 = (int) _1;
2696 : x.0_3 = (int) x_5(D);
2697 : _7 = _2 + x.0_3; */
2698 : ok = true;
2699 : }
2700 : }
2701 :
2702 57781 : if (ok)
2703 : {
2704 2544 : tree new_lhs = make_ssa_name (TREE_TYPE (lhs2));
2705 2544 : gimple_call_set_lhs (stmt2, new_lhs);
2706 2544 : gimple_call_set_fndecl (stmt2, ndecl);
2707 2544 : gimple_stmt_iterator gsi = gsi_for_stmt (use_stmt);
2708 2544 : if (ccode == ERROR_MARK)
2709 2000 : gimple_assign_set_rhs_with_ops (&gsi, cast_stmt
2710 : ? NOP_EXPR : SSA_NAME,
2711 : new_lhs);
2712 : else
2713 : {
2714 1321 : crhs1 = new_lhs;
2715 1321 : crhs2 = build_zero_cst (TREE_TYPE (lhs2));
2716 1321 : if (gimple_code (use_stmt) == GIMPLE_COND)
2717 : {
2718 982 : gcond *cond_stmt = as_a <gcond *> (use_stmt);
2719 982 : gimple_cond_set_lhs (cond_stmt, crhs1);
2720 982 : gimple_cond_set_rhs (cond_stmt, crhs2);
2721 : }
2722 339 : else if (gimple_assign_rhs_class (use_stmt)
2723 : == GIMPLE_BINARY_RHS)
2724 : {
2725 339 : gimple_assign_set_rhs1 (use_stmt, crhs1);
2726 339 : gimple_assign_set_rhs2 (use_stmt, crhs2);
2727 : }
2728 : else
2729 : {
2730 0 : gcc_checking_assert (gimple_assign_rhs_code (use_stmt)
2731 : == COND_EXPR);
2732 0 : tree cond = build2 (ccode, boolean_type_node,
2733 : crhs1, crhs2);
2734 0 : gimple_assign_set_rhs1 (use_stmt, cond);
2735 : }
2736 : }
2737 2544 : update_stmt (use_stmt);
2738 2544 : if (atomic_op != BIT_AND_EXPR
2739 2544 : && atomic_op != BIT_IOR_EXPR
2740 2544 : && !stmt_ends_bb_p (stmt2))
2741 : {
2742 : /* For the benefit of debug stmts, emit stmt(s) to set
2743 : lhs2 to the value it had from the new builtin.
2744 : E.g. if it was previously:
2745 : lhs2 = __atomic_fetch_add_8 (ptr, arg, 0);
2746 : emit:
2747 : new_lhs = __atomic_add_fetch_8 (ptr, arg, 0);
2748 : lhs2 = new_lhs - arg;
2749 : We also keep cast_stmt if any in the IL for
2750 : the same reasons.
2751 : These stmts will be DCEd later and proper debug info
2752 : will be emitted.
2753 : This is only possible for reversible operations
2754 : (+/-/^) and without -fnon-call-exceptions. */
2755 2203 : gsi = gsi_for_stmt (stmt2);
2756 2203 : tree type = TREE_TYPE (lhs2);
2757 2203 : if (TREE_CODE (arg) == INTEGER_CST)
2758 1621 : arg = fold_convert (type, arg);
2759 582 : else if (!useless_type_conversion_p (type, TREE_TYPE (arg)))
2760 : {
2761 0 : tree narg = make_ssa_name (type);
2762 0 : gimple *g = gimple_build_assign (narg, NOP_EXPR, arg);
2763 0 : gsi_insert_after (&gsi, g, GSI_NEW_STMT);
2764 0 : arg = narg;
2765 : }
2766 2203 : enum tree_code rcode;
2767 2203 : switch (atomic_op)
2768 : {
2769 : case PLUS_EXPR: rcode = MINUS_EXPR; break;
2770 726 : case MINUS_EXPR: rcode = PLUS_EXPR; break;
2771 492 : case BIT_XOR_EXPR: rcode = atomic_op; break;
2772 0 : default: gcc_unreachable ();
2773 : }
2774 2203 : gimple *g = gimple_build_assign (lhs2, rcode, new_lhs, arg);
2775 2203 : gsi_insert_after (&gsi, g, GSI_NEW_STMT);
2776 2203 : update_stmt (stmt2);
2777 : }
2778 : else
2779 : {
2780 : /* For e.g.
2781 : lhs2 = __atomic_fetch_or_8 (ptr, arg, 0);
2782 : after we change it to
2783 : new_lhs = __atomic_or_fetch_8 (ptr, arg, 0);
2784 : there is no way to find out the lhs2 value (i.e.
2785 : what the atomic memory contained before the operation),
2786 : values of some bits are lost. We have checked earlier
2787 : that we don't have any non-debug users except for what
2788 : we are already changing, so we need to reset the
2789 : debug stmts and remove the cast_stmt if any. */
2790 341 : imm_use_iterator iter;
2791 676 : FOR_EACH_IMM_USE_STMT (use_stmt, iter, lhs2)
2792 335 : if (use_stmt != cast_stmt)
2793 : {
2794 168 : gcc_assert (is_gimple_debug (use_stmt));
2795 168 : gimple_debug_bind_reset_value (use_stmt);
2796 168 : update_stmt (use_stmt);
2797 341 : }
2798 341 : if (cast_stmt)
2799 : {
2800 167 : gsi = gsi_for_stmt (cast_stmt);
2801 167 : gsi_remove (&gsi, true);
2802 : }
2803 341 : update_stmt (stmt2);
2804 341 : release_ssa_name (lhs2);
2805 : }
2806 : }
2807 : }
2808 : break;
2809 :
2810 : default:
2811 : break;
2812 : }
2813 : return false;
2814 : }
2815 :
2816 : /* Given a ssa_name in NAME see if it was defined by an assignment and
2817 : set CODE to be the code and ARG1 to the first operand on the rhs and ARG2
2818 : to the second operand on the rhs. */
2819 :
2820 : static inline void
2821 17234633 : defcodefor_name (tree name, enum tree_code *code, tree *arg1, tree *arg2)
2822 : {
2823 17234633 : gimple *def;
2824 17234633 : enum tree_code code1;
2825 17234633 : tree arg11;
2826 17234633 : tree arg21;
2827 17234633 : tree arg31;
2828 17234633 : enum gimple_rhs_class grhs_class;
2829 :
2830 17234633 : code1 = TREE_CODE (name);
2831 17234633 : arg11 = name;
2832 17234633 : arg21 = NULL_TREE;
2833 17234633 : arg31 = NULL_TREE;
2834 17234633 : grhs_class = get_gimple_rhs_class (code1);
2835 :
2836 17234633 : if (code1 == SSA_NAME)
2837 : {
2838 11484783 : def = SSA_NAME_DEF_STMT (name);
2839 :
2840 11484783 : if (def && is_gimple_assign (def)
2841 18583310 : && can_propagate_from (def))
2842 : {
2843 4909519 : code1 = gimple_assign_rhs_code (def);
2844 4909519 : arg11 = gimple_assign_rhs1 (def);
2845 4909519 : arg21 = gimple_assign_rhs2 (def);
2846 4909519 : arg31 = gimple_assign_rhs3 (def);
2847 : }
2848 : }
2849 5749850 : else if (grhs_class != GIMPLE_SINGLE_RHS)
2850 0 : code1 = ERROR_MARK;
2851 :
2852 17234633 : *code = code1;
2853 17234633 : *arg1 = arg11;
2854 17234633 : if (arg2)
2855 17217397 : *arg2 = arg21;
2856 17234633 : if (arg31)
2857 2206 : *code = ERROR_MARK;
2858 17234633 : }
2859 :
2860 :
2861 : /* Recognize rotation patterns. Return true if a transformation
2862 : applied, otherwise return false.
2863 :
2864 : We are looking for X with unsigned type T with bitsize B, OP being
2865 : +, | or ^, some type T2 wider than T. For:
2866 : (X << CNT1) OP (X >> CNT2) iff CNT1 + CNT2 == B
2867 : ((T) ((T2) X << CNT1)) OP ((T) ((T2) X >> CNT2)) iff CNT1 + CNT2 == B
2868 :
2869 : transform these into:
2870 : X r<< CNT1
2871 :
2872 : Or for:
2873 : (X << Y) OP (X >> (B - Y))
2874 : (X << (int) Y) OP (X >> (int) (B - Y))
2875 : ((T) ((T2) X << Y)) OP ((T) ((T2) X >> (B - Y)))
2876 : ((T) ((T2) X << (int) Y)) OP ((T) ((T2) X >> (int) (B - Y)))
2877 : (X << Y) | (X >> ((-Y) & (B - 1)))
2878 : (X << (int) Y) | (X >> (int) ((-Y) & (B - 1)))
2879 : ((T) ((T2) X << Y)) | ((T) ((T2) X >> ((-Y) & (B - 1))))
2880 : ((T) ((T2) X << (int) Y)) | ((T) ((T2) X >> (int) ((-Y) & (B - 1))))
2881 :
2882 : transform these into (last 2 only if ranger can prove Y < B
2883 : or Y = N * B):
2884 : X r<< Y
2885 : or
2886 : X r<< (& & (B - 1))
2887 : The latter for the forms with T2 wider than T if ranger can't prove Y < B.
2888 :
2889 : Or for:
2890 : (X << (Y & (B - 1))) | (X >> ((-Y) & (B - 1)))
2891 : (X << (int) (Y & (B - 1))) | (X >> (int) ((-Y) & (B - 1)))
2892 : ((T) ((T2) X << (Y & (B - 1)))) | ((T) ((T2) X >> ((-Y) & (B - 1))))
2893 : ((T) ((T2) X << (int) (Y & (B - 1)))) \
2894 : | ((T) ((T2) X >> (int) ((-Y) & (B - 1))))
2895 :
2896 : transform these into:
2897 : X r<< (Y & (B - 1))
2898 :
2899 : Note, in the patterns with T2 type, the type of OP operands
2900 : might be even a signed type, but should have precision B.
2901 : Expressions with & (B - 1) should be recognized only if B is
2902 : a power of 2. */
2903 :
2904 : static bool
2905 10121909 : simplify_rotate (gimple_stmt_iterator *gsi)
2906 : {
2907 10121909 : gimple *stmt = gsi_stmt (*gsi);
2908 10121909 : tree arg[2], rtype, rotcnt = NULL_TREE;
2909 10121909 : tree def_arg1[2], def_arg2[2];
2910 10121909 : enum tree_code def_code[2];
2911 10121909 : tree lhs;
2912 10121909 : int i;
2913 10121909 : bool swapped_p = false;
2914 10121909 : gimple *g;
2915 10121909 : gimple *def_arg_stmt[2] = { NULL, NULL };
2916 10121909 : int wider_prec = 0;
2917 10121909 : bool add_masking = false;
2918 :
2919 10121909 : arg[0] = gimple_assign_rhs1 (stmt);
2920 10121909 : arg[1] = gimple_assign_rhs2 (stmt);
2921 10121909 : rtype = TREE_TYPE (arg[0]);
2922 :
2923 : /* Only create rotates in complete modes. Other cases are not
2924 : expanded properly. */
2925 10121909 : if (!INTEGRAL_TYPE_P (rtype)
2926 10121909 : || !type_has_mode_precision_p (rtype))
2927 1554305 : return false;
2928 :
2929 25702812 : for (i = 0; i < 2; i++)
2930 : {
2931 17135208 : defcodefor_name (arg[i], &def_code[i], &def_arg1[i], &def_arg2[i]);
2932 17135208 : if (TREE_CODE (arg[i]) == SSA_NAME)
2933 11385358 : def_arg_stmt[i] = SSA_NAME_DEF_STMT (arg[i]);
2934 : }
2935 :
2936 : /* Look through narrowing (or same precision) conversions. */
2937 7606332 : if (CONVERT_EXPR_CODE_P (def_code[0])
2938 961272 : && CONVERT_EXPR_CODE_P (def_code[1])
2939 139870 : && INTEGRAL_TYPE_P (TREE_TYPE (def_arg1[0]))
2940 116861 : && INTEGRAL_TYPE_P (TREE_TYPE (def_arg1[1]))
2941 109262 : && TYPE_PRECISION (TREE_TYPE (def_arg1[0]))
2942 109262 : == TYPE_PRECISION (TREE_TYPE (def_arg1[1]))
2943 63577 : && TYPE_PRECISION (TREE_TYPE (def_arg1[0])) >= TYPE_PRECISION (rtype)
2944 44040 : && has_single_use (arg[0])
2945 8600406 : && has_single_use (arg[1]))
2946 : {
2947 28479 : wider_prec = TYPE_PRECISION (TREE_TYPE (def_arg1[0]));
2948 85437 : for (i = 0; i < 2; i++)
2949 : {
2950 56958 : arg[i] = def_arg1[i];
2951 56958 : defcodefor_name (arg[i], &def_code[i], &def_arg1[i], &def_arg2[i]);
2952 56958 : if (TREE_CODE (arg[i]) == SSA_NAME)
2953 56958 : def_arg_stmt[i] = SSA_NAME_DEF_STMT (arg[i]);
2954 : }
2955 : }
2956 : else
2957 : {
2958 : /* Handle signed rotate; the RSHIFT_EXPR has to be done
2959 : in unsigned type but LSHIFT_EXPR could be signed. */
2960 8539125 : i = (def_code[0] == LSHIFT_EXPR || def_code[0] == RSHIFT_EXPR);
2961 7589468 : if (CONVERT_EXPR_CODE_P (def_code[i])
2962 949657 : && (def_code[1 - i] == LSHIFT_EXPR || def_code[1 - i] == RSHIFT_EXPR)
2963 28168 : && INTEGRAL_TYPE_P (TREE_TYPE (def_arg1[i]))
2964 27084 : && TYPE_PRECISION (rtype) == TYPE_PRECISION (TREE_TYPE (def_arg1[i]))
2965 8542465 : && has_single_use (arg[i]))
2966 : {
2967 1995 : arg[i] = def_arg1[i];
2968 1995 : defcodefor_name (arg[i], &def_code[i], &def_arg1[i], &def_arg2[i]);
2969 1995 : if (TREE_CODE (arg[i]) == SSA_NAME)
2970 1995 : def_arg_stmt[i] = SSA_NAME_DEF_STMT (arg[i]);
2971 : }
2972 : }
2973 :
2974 : /* One operand has to be LSHIFT_EXPR and one RSHIFT_EXPR. */
2975 8765389 : for (i = 0; i < 2; i++)
2976 8740851 : if (def_code[i] != LSHIFT_EXPR && def_code[i] != RSHIFT_EXPR)
2977 : return false;
2978 239667 : else if (!has_single_use (arg[i]))
2979 : return false;
2980 24538 : if (def_code[0] == def_code[1])
2981 : return false;
2982 :
2983 : /* If we've looked through narrowing conversions before, look through
2984 : widening conversions from unsigned type with the same precision
2985 : as rtype here. */
2986 20241 : if (TYPE_PRECISION (TREE_TYPE (def_arg1[0])) != TYPE_PRECISION (rtype))
2987 19348 : for (i = 0; i < 2; i++)
2988 : {
2989 12900 : tree tem;
2990 12900 : enum tree_code code;
2991 12900 : defcodefor_name (def_arg1[i], &code, &tem, NULL);
2992 4 : if (!CONVERT_EXPR_CODE_P (code)
2993 12896 : || !INTEGRAL_TYPE_P (TREE_TYPE (tem))
2994 25796 : || TYPE_PRECISION (TREE_TYPE (tem)) != TYPE_PRECISION (rtype))
2995 4 : return false;
2996 12896 : def_arg1[i] = tem;
2997 : }
2998 : /* Both shifts have to use the same first operand. */
2999 20237 : if (!operand_equal_for_phi_arg_p (def_arg1[0], def_arg1[1])
3000 32166 : || !types_compatible_p (TREE_TYPE (def_arg1[0]),
3001 11929 : TREE_TYPE (def_arg1[1])))
3002 : {
3003 8308 : if ((TYPE_PRECISION (TREE_TYPE (def_arg1[0]))
3004 8308 : != TYPE_PRECISION (TREE_TYPE (def_arg1[1])))
3005 8308 : || (TYPE_UNSIGNED (TREE_TYPE (def_arg1[0]))
3006 8308 : == TYPE_UNSIGNED (TREE_TYPE (def_arg1[1]))))
3007 8284 : return false;
3008 :
3009 : /* Handle signed rotate; the RSHIFT_EXPR has to be done
3010 : in unsigned type but LSHIFT_EXPR could be signed. */
3011 540 : i = def_code[0] != RSHIFT_EXPR;
3012 540 : if (!TYPE_UNSIGNED (TREE_TYPE (def_arg1[i])))
3013 : return false;
3014 :
3015 507 : tree tem;
3016 507 : enum tree_code code;
3017 507 : defcodefor_name (def_arg1[i], &code, &tem, NULL);
3018 304 : if (!CONVERT_EXPR_CODE_P (code)
3019 203 : || !INTEGRAL_TYPE_P (TREE_TYPE (tem))
3020 710 : || TYPE_PRECISION (TREE_TYPE (tem)) != TYPE_PRECISION (rtype))
3021 : return false;
3022 194 : def_arg1[i] = tem;
3023 194 : if (!operand_equal_for_phi_arg_p (def_arg1[0], def_arg1[1])
3024 218 : || !types_compatible_p (TREE_TYPE (def_arg1[0]),
3025 24 : TREE_TYPE (def_arg1[1])))
3026 170 : return false;
3027 : }
3028 11929 : else if (!TYPE_UNSIGNED (TREE_TYPE (def_arg1[0])))
3029 : return false;
3030 :
3031 : /* CNT1 + CNT2 == B case above. */
3032 10698 : if (tree_fits_uhwi_p (def_arg2[0])
3033 1210 : && tree_fits_uhwi_p (def_arg2[1])
3034 10698 : && tree_to_uhwi (def_arg2[0])
3035 1210 : + tree_to_uhwi (def_arg2[1]) == TYPE_PRECISION (rtype))
3036 : rotcnt = def_arg2[0];
3037 9768 : else if (TREE_CODE (def_arg2[0]) != SSA_NAME
3038 9488 : || TREE_CODE (def_arg2[1]) != SSA_NAME)
3039 : return false;
3040 : else
3041 : {
3042 9488 : tree cdef_arg1[2], cdef_arg2[2], def_arg2_alt[2];
3043 9488 : enum tree_code cdef_code[2];
3044 9488 : gimple *def_arg_alt_stmt[2] = { NULL, NULL };
3045 9488 : int check_range = 0;
3046 9488 : gimple *check_range_stmt = NULL;
3047 : /* Look through conversion of the shift count argument.
3048 : The C/C++ FE cast any shift count argument to integer_type_node.
3049 : The only problem might be if the shift count type maximum value
3050 : is equal or smaller than number of bits in rtype. */
3051 28464 : for (i = 0; i < 2; i++)
3052 : {
3053 18976 : def_arg2_alt[i] = def_arg2[i];
3054 18976 : defcodefor_name (def_arg2[i], &cdef_code[i],
3055 : &cdef_arg1[i], &cdef_arg2[i]);
3056 14716 : if (CONVERT_EXPR_CODE_P (cdef_code[i])
3057 4260 : && INTEGRAL_TYPE_P (TREE_TYPE (cdef_arg1[i]))
3058 4260 : && TYPE_PRECISION (TREE_TYPE (cdef_arg1[i]))
3059 8520 : > floor_log2 (TYPE_PRECISION (rtype))
3060 23236 : && type_has_mode_precision_p (TREE_TYPE (cdef_arg1[i])))
3061 : {
3062 4260 : def_arg2_alt[i] = cdef_arg1[i];
3063 4260 : if (TREE_CODE (def_arg2[i]) == SSA_NAME)
3064 4260 : def_arg_alt_stmt[i] = SSA_NAME_DEF_STMT (def_arg2[i]);
3065 4260 : defcodefor_name (def_arg2_alt[i], &cdef_code[i],
3066 : &cdef_arg1[i], &cdef_arg2[i]);
3067 : }
3068 : else
3069 14716 : def_arg_alt_stmt[i] = def_arg_stmt[i];
3070 : }
3071 25796 : for (i = 0; i < 2; i++)
3072 : /* Check for one shift count being Y and the other B - Y,
3073 : with optional casts. */
3074 18625 : if (cdef_code[i] == MINUS_EXPR
3075 862 : && tree_fits_shwi_p (cdef_arg1[i])
3076 862 : && tree_to_shwi (cdef_arg1[i]) == TYPE_PRECISION (rtype)
3077 19447 : && TREE_CODE (cdef_arg2[i]) == SSA_NAME)
3078 : {
3079 822 : tree tem;
3080 822 : enum tree_code code;
3081 :
3082 822 : if (cdef_arg2[i] == def_arg2[1 - i]
3083 472 : || cdef_arg2[i] == def_arg2_alt[1 - i])
3084 : {
3085 350 : rotcnt = cdef_arg2[i];
3086 350 : check_range = -1;
3087 350 : if (cdef_arg2[i] == def_arg2[1 - i])
3088 350 : check_range_stmt = def_arg_stmt[1 - i];
3089 : else
3090 0 : check_range_stmt = def_arg_alt_stmt[1 - i];
3091 806 : break;
3092 : }
3093 472 : defcodefor_name (cdef_arg2[i], &code, &tem, NULL);
3094 16 : if (CONVERT_EXPR_CODE_P (code)
3095 456 : && INTEGRAL_TYPE_P (TREE_TYPE (tem))
3096 456 : && TYPE_PRECISION (TREE_TYPE (tem))
3097 912 : > floor_log2 (TYPE_PRECISION (rtype))
3098 456 : && type_has_mode_precision_p (TREE_TYPE (tem))
3099 928 : && (tem == def_arg2[1 - i]
3100 288 : || tem == def_arg2_alt[1 - i]))
3101 : {
3102 456 : rotcnt = tem;
3103 456 : check_range = -1;
3104 456 : if (tem == def_arg2[1 - i])
3105 168 : check_range_stmt = def_arg_stmt[1 - i];
3106 : else
3107 288 : check_range_stmt = def_arg_alt_stmt[1 - i];
3108 : break;
3109 : }
3110 : }
3111 : /* The above sequence isn't safe for Y being 0,
3112 : because then one of the shifts triggers undefined behavior.
3113 : This alternative is safe even for rotation count of 0.
3114 : One shift count is Y and the other (-Y) & (B - 1).
3115 : Or one shift count is Y & (B - 1) and the other (-Y) & (B - 1). */
3116 17803 : else if (cdef_code[i] == BIT_AND_EXPR
3117 28724 : && pow2p_hwi (TYPE_PRECISION (rtype))
3118 12416 : && tree_fits_shwi_p (cdef_arg2[i])
3119 24832 : && tree_to_shwi (cdef_arg2[i])
3120 12416 : == TYPE_PRECISION (rtype) - 1
3121 12356 : && TREE_CODE (cdef_arg1[i]) == SSA_NAME
3122 30159 : && gimple_assign_rhs_code (stmt) == BIT_IOR_EXPR)
3123 : {
3124 2296 : tree tem;
3125 2296 : enum tree_code code;
3126 :
3127 2296 : defcodefor_name (cdef_arg1[i], &code, &tem, NULL);
3128 2099 : if (CONVERT_EXPR_CODE_P (code)
3129 197 : && INTEGRAL_TYPE_P (TREE_TYPE (tem))
3130 197 : && TYPE_PRECISION (TREE_TYPE (tem))
3131 394 : > floor_log2 (TYPE_PRECISION (rtype))
3132 2493 : && type_has_mode_precision_p (TREE_TYPE (tem)))
3133 197 : defcodefor_name (tem, &code, &tem, NULL);
3134 :
3135 2296 : if (code == NEGATE_EXPR)
3136 : {
3137 1525 : if (tem == def_arg2[1 - i] || tem == def_arg2_alt[1 - i])
3138 : {
3139 854 : rotcnt = tem;
3140 854 : check_range = 1;
3141 854 : if (tem == def_arg2[1 - i])
3142 846 : check_range_stmt = def_arg_stmt[1 - i];
3143 : else
3144 8 : check_range_stmt = def_arg_alt_stmt[1 - i];
3145 1511 : break;
3146 : }
3147 671 : tree tem2;
3148 671 : defcodefor_name (tem, &code, &tem2, NULL);
3149 237 : if (CONVERT_EXPR_CODE_P (code)
3150 434 : && INTEGRAL_TYPE_P (TREE_TYPE (tem2))
3151 434 : && TYPE_PRECISION (TREE_TYPE (tem2))
3152 868 : > floor_log2 (TYPE_PRECISION (rtype))
3153 1105 : && type_has_mode_precision_p (TREE_TYPE (tem2)))
3154 : {
3155 434 : if (tem2 == def_arg2[1 - i]
3156 434 : || tem2 == def_arg2_alt[1 - i])
3157 : {
3158 228 : rotcnt = tem2;
3159 228 : check_range = 1;
3160 228 : if (tem2 == def_arg2[1 - i])
3161 0 : check_range_stmt = def_arg_stmt[1 - i];
3162 : else
3163 228 : check_range_stmt = def_arg_alt_stmt[1 - i];
3164 : break;
3165 : }
3166 : }
3167 : else
3168 237 : tem2 = NULL_TREE;
3169 :
3170 443 : if (cdef_code[1 - i] == BIT_AND_EXPR
3171 430 : && tree_fits_shwi_p (cdef_arg2[1 - i])
3172 860 : && tree_to_shwi (cdef_arg2[1 - i])
3173 430 : == TYPE_PRECISION (rtype) - 1
3174 873 : && TREE_CODE (cdef_arg1[1 - i]) == SSA_NAME)
3175 : {
3176 430 : if (tem == cdef_arg1[1 - i]
3177 205 : || tem2 == cdef_arg1[1 - i])
3178 : {
3179 : rotcnt = def_arg2[1 - i];
3180 429 : break;
3181 : }
3182 193 : tree tem3;
3183 193 : defcodefor_name (cdef_arg1[1 - i], &code, &tem3, NULL);
3184 0 : if (CONVERT_EXPR_CODE_P (code)
3185 193 : && INTEGRAL_TYPE_P (TREE_TYPE (tem3))
3186 193 : && TYPE_PRECISION (TREE_TYPE (tem3))
3187 386 : > floor_log2 (TYPE_PRECISION (rtype))
3188 386 : && type_has_mode_precision_p (TREE_TYPE (tem3)))
3189 : {
3190 193 : if (tem == tem3 || tem2 == tem3)
3191 : {
3192 : rotcnt = def_arg2[1 - i];
3193 : break;
3194 : }
3195 : }
3196 : }
3197 : }
3198 : }
3199 2317 : if (check_range && wider_prec > TYPE_PRECISION (rtype))
3200 : {
3201 1533 : if (TREE_CODE (rotcnt) != SSA_NAME)
3202 573 : return false;
3203 1533 : int_range_max r;
3204 1533 : range_query *q = get_range_query (cfun);
3205 1533 : if (q == get_global_range_query ())
3206 1522 : q = enable_ranger (cfun);
3207 1533 : if (!q->range_of_expr (r, rotcnt, check_range_stmt))
3208 : {
3209 0 : if (check_range > 0)
3210 : return false;
3211 0 : r.set_varying (TREE_TYPE (rotcnt));
3212 : }
3213 1533 : int prec = TYPE_PRECISION (TREE_TYPE (rotcnt));
3214 1533 : signop sign = TYPE_SIGN (TREE_TYPE (rotcnt));
3215 1533 : wide_int min = wide_int::from (TYPE_PRECISION (rtype), prec, sign);
3216 1533 : wide_int max = wide_int::from (wider_prec - 1, prec, sign);
3217 1533 : if (check_range < 0)
3218 616 : max = min;
3219 1533 : int_range<1> r2 (TREE_TYPE (rotcnt), min, max);
3220 1533 : r.intersect (r2);
3221 1533 : if (!r.undefined_p ())
3222 : {
3223 1181 : if (check_range > 0)
3224 : {
3225 589 : int_range_max r3;
3226 1844 : for (int i = TYPE_PRECISION (rtype) + 1; i < wider_prec;
3227 1255 : i += TYPE_PRECISION (rtype))
3228 : {
3229 1255 : int j = i + TYPE_PRECISION (rtype) - 2;
3230 1255 : min = wide_int::from (i, prec, sign);
3231 1255 : max = wide_int::from (MIN (j, wider_prec - 1),
3232 1255 : prec, sign);
3233 1255 : int_range<1> r4 (TREE_TYPE (rotcnt), min, max);
3234 1255 : r3.union_ (r4);
3235 1255 : }
3236 589 : r.intersect (r3);
3237 589 : if (!r.undefined_p ())
3238 573 : return false;
3239 589 : }
3240 : add_masking = true;
3241 : }
3242 1533 : }
3243 8915 : if (rotcnt == NULL_TREE)
3244 : return false;
3245 1744 : swapped_p = i != 1;
3246 : }
3247 :
3248 2674 : if (!useless_type_conversion_p (TREE_TYPE (def_arg2[0]),
3249 2674 : TREE_TYPE (rotcnt)))
3250 : {
3251 496 : g = gimple_build_assign (make_ssa_name (TREE_TYPE (def_arg2[0])),
3252 : NOP_EXPR, rotcnt);
3253 496 : gsi_insert_before (gsi, g, GSI_SAME_STMT);
3254 496 : rotcnt = gimple_assign_lhs (g);
3255 : }
3256 2674 : if (add_masking)
3257 : {
3258 608 : g = gimple_build_assign (make_ssa_name (TREE_TYPE (rotcnt)),
3259 : BIT_AND_EXPR, rotcnt,
3260 608 : build_int_cst (TREE_TYPE (rotcnt),
3261 608 : TYPE_PRECISION (rtype) - 1));
3262 608 : gsi_insert_before (gsi, g, GSI_SAME_STMT);
3263 608 : rotcnt = gimple_assign_lhs (g);
3264 : }
3265 2674 : lhs = gimple_assign_lhs (stmt);
3266 2674 : if (!useless_type_conversion_p (rtype, TREE_TYPE (def_arg1[0])))
3267 1010 : lhs = make_ssa_name (TREE_TYPE (def_arg1[0]));
3268 2674 : g = gimple_build_assign (lhs,
3269 2674 : ((def_code[0] == LSHIFT_EXPR) ^ swapped_p)
3270 : ? LROTATE_EXPR : RROTATE_EXPR, def_arg1[0], rotcnt);
3271 2674 : if (!useless_type_conversion_p (rtype, TREE_TYPE (def_arg1[0])))
3272 : {
3273 1010 : gsi_insert_before (gsi, g, GSI_SAME_STMT);
3274 1010 : g = gimple_build_assign (gimple_assign_lhs (stmt), NOP_EXPR, lhs);
3275 : }
3276 2674 : gsi_replace (gsi, g, false);
3277 2674 : return true;
3278 : }
3279 :
3280 :
3281 : /* Check whether an array contains a valid table according to VALIDATE_FN. */
3282 : template<typename ValidateFn>
3283 : static bool
3284 14 : check_table_array (tree ctor, HOST_WIDE_INT &zero_val, unsigned bits,
3285 : ValidateFn validate_fn)
3286 : {
3287 : tree elt, idx;
3288 14 : unsigned HOST_WIDE_INT i, raw_idx = 0;
3289 14 : unsigned matched = 0;
3290 :
3291 14 : zero_val = 0;
3292 :
3293 542 : FOR_EACH_CONSTRUCTOR_ELT (CONSTRUCTOR_ELTS (ctor), i, idx, elt)
3294 : {
3295 542 : if (!tree_fits_shwi_p (idx))
3296 : return false;
3297 542 : if (!tree_fits_shwi_p (elt) && TREE_CODE (elt) != RAW_DATA_CST)
3298 : return false;
3299 :
3300 542 : unsigned HOST_WIDE_INT index = tree_to_shwi (idx);
3301 : HOST_WIDE_INT val;
3302 :
3303 542 : if (TREE_CODE (elt) == INTEGER_CST)
3304 478 : val = tree_to_shwi (elt);
3305 : else
3306 : {
3307 64 : if (raw_idx == (unsigned) RAW_DATA_LENGTH (elt))
3308 : {
3309 0 : raw_idx = 0;
3310 0 : continue;
3311 : }
3312 64 : if (TYPE_UNSIGNED (TREE_TYPE (elt)))
3313 0 : val = RAW_DATA_UCHAR_ELT (elt, raw_idx);
3314 : else
3315 64 : val = RAW_DATA_SCHAR_ELT (elt, raw_idx);
3316 64 : index += raw_idx;
3317 64 : raw_idx++;
3318 64 : i--;
3319 : }
3320 :
3321 542 : if (index > bits * 2)
3322 : return false;
3323 :
3324 542 : if (index == 0)
3325 : {
3326 14 : zero_val = val;
3327 14 : matched++;
3328 : }
3329 :
3330 542 : if (val >= 0 && val < bits && validate_fn (val, index))
3331 480 : matched++;
3332 :
3333 542 : if (matched > bits)
3334 : return true;
3335 : }
3336 :
3337 : return false;
3338 : }
3339 :
3340 : /* Check whether a string contains a valid table according to VALIDATE_FN. */
3341 : template<typename ValidateFn>
3342 : static bool
3343 4 : check_table_string (tree string, HOST_WIDE_INT &zero_val,unsigned bits,
3344 : ValidateFn validate_fn)
3345 : {
3346 4 : unsigned HOST_WIDE_INT len = TREE_STRING_LENGTH (string);
3347 4 : unsigned matched = 0;
3348 4 : const unsigned char *p = (const unsigned char *) TREE_STRING_POINTER (string);
3349 :
3350 4 : if (len < bits || len > bits * 2)
3351 : return false;
3352 :
3353 4 : zero_val = p[0];
3354 :
3355 164 : for (unsigned i = 0; i < len; i++)
3356 160 : if (p[i] < bits && validate_fn (p[i], i))
3357 160 : matched++;
3358 :
3359 4 : return matched == bits;
3360 : }
3361 :
3362 : /* Check whether CTOR contains a valid table according to VALIDATE_FN. */
3363 : template<typename ValidateFn>
3364 : static bool
3365 26 : check_table (tree ctor, tree type, HOST_WIDE_INT &zero_val, unsigned bits,
3366 : ValidateFn validate_fn)
3367 : {
3368 26 : if (TREE_CODE (ctor) == CONSTRUCTOR)
3369 14 : return check_table_array (ctor, zero_val, bits, validate_fn);
3370 : else if (TREE_CODE (ctor) == STRING_CST
3371 12 : && TYPE_PRECISION (type) == CHAR_TYPE_SIZE)
3372 4 : return check_table_string (ctor, zero_val, bits, validate_fn);
3373 : return false;
3374 : }
3375 :
3376 : /* Match.pd function to match the ctz expression. */
3377 : extern bool gimple_ctz_table_index (tree, tree *, tree (*)(tree));
3378 : extern bool gimple_clz_table_index (tree, tree *, tree (*)(tree));
3379 :
3380 : /* Recognize count leading and trailing zeroes idioms.
3381 : The canonical form is array[((x & -x) * C) >> SHIFT] where C is a magic
3382 : constant which when multiplied by a power of 2 creates a unique value
3383 : in the top 5 or 6 bits. This is then indexed into a table which maps it
3384 : to the number of trailing zeroes. Array[0] is returned so the caller can
3385 : emit an appropriate sequence depending on whether ctz (0) is defined on
3386 : the target. */
3387 :
3388 : static bool
3389 1962912 : simplify_count_zeroes (gimple_stmt_iterator *gsi)
3390 : {
3391 1962912 : gimple *stmt = gsi_stmt (*gsi);
3392 1962912 : tree array_ref = gimple_assign_rhs1 (stmt);
3393 1962912 : tree res_ops[3];
3394 :
3395 1962912 : gcc_checking_assert (TREE_CODE (array_ref) == ARRAY_REF);
3396 :
3397 1962912 : internal_fn fn = IFN_LAST;
3398 : /* For CTZ we recognize ((x & -x) * C) >> SHIFT where the array data
3399 : represents the number of trailing zeros. */
3400 1962912 : if (gimple_ctz_table_index (TREE_OPERAND (array_ref, 1), &res_ops[0], NULL))
3401 : fn = IFN_CTZ;
3402 : /* For CLZ we recognize
3403 : x |= x >> 1;
3404 : x |= x >> 2;
3405 : x |= x >> 4;
3406 : x |= x >> 8;
3407 : x |= x >> 16;
3408 : (x * C) >> SHIFT
3409 : where 31 minus the array data represents the number of leading zeros. */
3410 1962890 : else if (gimple_clz_table_index (TREE_OPERAND (array_ref, 1), &res_ops[0],
3411 : NULL))
3412 : fn = IFN_CLZ;
3413 : else
3414 : return false;
3415 :
3416 31 : HOST_WIDE_INT zero_val;
3417 31 : tree type = TREE_TYPE (array_ref);
3418 31 : tree array = TREE_OPERAND (array_ref, 0);
3419 31 : tree input_type = TREE_TYPE (res_ops[0]);
3420 31 : unsigned input_bits = tree_to_shwi (TYPE_SIZE (input_type));
3421 :
3422 : /* Check the array element type is not wider than 32 bits and the input is
3423 : an unsigned 32-bit or 64-bit type. */
3424 31 : if (TYPE_PRECISION (type) > 32 || !TYPE_UNSIGNED (input_type))
3425 : return false;
3426 27 : if (input_bits != 32 && input_bits != 64)
3427 : return false;
3428 :
3429 27 : if (!direct_internal_fn_supported_p (fn, input_type, OPTIMIZE_FOR_BOTH))
3430 : return false;
3431 :
3432 : /* Check the lower bound of the array is zero. */
3433 27 : tree low = array_ref_low_bound (array_ref);
3434 27 : if (!low || !integer_zerop (low))
3435 0 : return false;
3436 :
3437 : /* Check the shift extracts the top 5..7 bits. */
3438 27 : unsigned shiftval = tree_to_shwi (res_ops[2]);
3439 27 : if (shiftval < input_bits - 7 || shiftval > input_bits - 5)
3440 : return false;
3441 :
3442 26 : tree ctor = ctor_for_folding (array);
3443 26 : if (!ctor)
3444 : return false;
3445 26 : unsigned HOST_WIDE_INT mulval = tree_to_uhwi (res_ops[1]);
3446 26 : if (fn == IFN_CTZ)
3447 : {
3448 429 : auto checkfn = [&](unsigned data, unsigned i) -> bool
3449 : {
3450 412 : unsigned HOST_WIDE_INT mask
3451 412 : = ((HOST_WIDE_INT_1U << (input_bits - shiftval)) - 1) << shiftval;
3452 412 : return (((mulval << data) & mask) >> shiftval) == i;
3453 17 : };
3454 17 : if (!check_table (ctor, type, zero_val, input_bits, checkfn))
3455 8 : return false;
3456 : }
3457 9 : else if (fn == IFN_CLZ)
3458 : {
3459 297 : auto checkfn = [&](unsigned data, unsigned i) -> bool
3460 : {
3461 288 : unsigned HOST_WIDE_INT mask
3462 288 : = ((HOST_WIDE_INT_1U << (input_bits - shiftval)) - 1) << shiftval;
3463 288 : return (((((HOST_WIDE_INT_1U << (data + 1)) - 1) * mulval) & mask)
3464 288 : >> shiftval) == i;
3465 9 : };
3466 9 : if (!check_table (ctor, type, zero_val, input_bits, checkfn))
3467 0 : return false;
3468 : }
3469 :
3470 18 : HOST_WIDE_INT ctz_val = -1;
3471 18 : bool zero_ok;
3472 18 : if (fn == IFN_CTZ)
3473 : {
3474 9 : ctz_val = 0;
3475 18 : zero_ok = CTZ_DEFINED_VALUE_AT_ZERO (SCALAR_INT_TYPE_MODE (input_type),
3476 : ctz_val) == 2;
3477 : }
3478 9 : else if (fn == IFN_CLZ)
3479 : {
3480 9 : ctz_val = 32;
3481 9 : zero_ok = CLZ_DEFINED_VALUE_AT_ZERO (SCALAR_INT_TYPE_MODE (input_type),
3482 : ctz_val) == 2;
3483 9 : zero_val = input_bits - 1 - zero_val;
3484 : }
3485 18 : int nargs = 2;
3486 :
3487 : /* If the input value can't be zero, don't special case ctz (0). */
3488 18 : range_query *q = get_range_query (cfun);
3489 18 : if (q == get_global_range_query ())
3490 18 : q = enable_ranger (cfun);
3491 18 : int_range_max vr;
3492 18 : if (q->range_of_expr (vr, res_ops[0], stmt)
3493 18 : && !range_includes_zero_p (vr))
3494 : {
3495 4 : zero_ok = true;
3496 4 : zero_val = 0;
3497 4 : ctz_val = 0;
3498 4 : nargs = 1;
3499 : }
3500 :
3501 18 : gimple_seq seq = NULL;
3502 18 : gimple *g;
3503 18 : gcall *call = gimple_build_call_internal (fn, nargs, res_ops[0],
3504 : nargs == 1 ? NULL_TREE
3505 32 : : build_int_cst (integer_type_node,
3506 14 : ctz_val));
3507 18 : gimple_set_location (call, gimple_location (stmt));
3508 18 : gimple_set_lhs (call, make_ssa_name (integer_type_node));
3509 18 : gimple_seq_add_stmt (&seq, call);
3510 :
3511 18 : tree prev_lhs = gimple_call_lhs (call);
3512 :
3513 18 : if (zero_ok && zero_val == ctz_val)
3514 : ;
3515 : /* Emit ctz (x) & 31 if ctz (0) is 32 but we need to return 0. */
3516 6 : else if (zero_ok && zero_val == 0 && ctz_val == input_bits)
3517 : {
3518 5 : g = gimple_build_assign (make_ssa_name (integer_type_node),
3519 : BIT_AND_EXPR, prev_lhs,
3520 : build_int_cst (integer_type_node,
3521 5 : input_bits - 1));
3522 5 : gimple_set_location (g, gimple_location (stmt));
3523 5 : gimple_seq_add_stmt (&seq, g);
3524 5 : prev_lhs = gimple_assign_lhs (g);
3525 : }
3526 : /* As fallback emit a conditional move. */
3527 : else
3528 : {
3529 7 : g = gimple_build_assign (make_ssa_name (boolean_type_node), EQ_EXPR,
3530 : res_ops[0], build_zero_cst (input_type));
3531 7 : gimple_set_location (g, gimple_location (stmt));
3532 7 : gimple_seq_add_stmt (&seq, g);
3533 7 : tree cond = gimple_assign_lhs (g);
3534 7 : g = gimple_build_assign (make_ssa_name (integer_type_node),
3535 : COND_EXPR, cond,
3536 7 : build_int_cst (integer_type_node, zero_val),
3537 : prev_lhs);
3538 7 : gimple_set_location (g, gimple_location (stmt));
3539 7 : gimple_seq_add_stmt (&seq, g);
3540 7 : prev_lhs = gimple_assign_lhs (g);
3541 : }
3542 :
3543 18 : if (fn == IFN_CLZ)
3544 : {
3545 9 : g = gimple_build_assign (make_ssa_name (integer_type_node),
3546 : MINUS_EXPR,
3547 : build_int_cst (integer_type_node,
3548 9 : input_bits - 1),
3549 : prev_lhs);
3550 9 : gimple_set_location (g, gimple_location (stmt));
3551 9 : gimple_seq_add_stmt (&seq, g);
3552 9 : prev_lhs = gimple_assign_lhs (g);
3553 : }
3554 :
3555 18 : g = gimple_build_assign (gimple_assign_lhs (stmt), NOP_EXPR, prev_lhs);
3556 18 : gimple_seq_add_stmt (&seq, g);
3557 18 : gsi_replace_with_seq (gsi, seq, true);
3558 18 : return true;
3559 18 : }
3560 :
3561 :
3562 : /* Determine whether applying the 2 permutations (mask1 then mask2)
3563 : gives back one of the input. */
3564 :
3565 : static int
3566 34 : is_combined_permutation_identity (tree mask1, tree mask2)
3567 : {
3568 34 : tree mask;
3569 34 : unsigned HOST_WIDE_INT nelts, i, j;
3570 34 : bool maybe_identity1 = true;
3571 34 : bool maybe_identity2 = true;
3572 :
3573 34 : gcc_checking_assert (TREE_CODE (mask1) == VECTOR_CST
3574 : && TREE_CODE (mask2) == VECTOR_CST);
3575 :
3576 : /* For VLA masks, check for the following pattern:
3577 : v1 = VEC_PERM_EXPR (v0, ..., mask1)
3578 : v2 = VEC_PERM_EXPR (v1, ..., mask2)
3579 : -->
3580 : v2 = v0
3581 : if mask1 == mask2 == {nelts - 1, nelts - 2, ...}. */
3582 :
3583 34 : if (operand_equal_p (mask1, mask2, 0)
3584 34 : && !VECTOR_CST_NELTS (mask1).is_constant ())
3585 : {
3586 : vec_perm_builder builder;
3587 : if (tree_to_vec_perm_builder (&builder, mask1))
3588 : {
3589 : poly_uint64 nelts = TYPE_VECTOR_SUBPARTS (TREE_TYPE (mask1));
3590 : vec_perm_indices sel (builder, 1, nelts);
3591 : if (sel.series_p (0, 1, nelts - 1, -1))
3592 : return 1;
3593 : }
3594 : }
3595 :
3596 34 : mask = fold_ternary (VEC_PERM_EXPR, TREE_TYPE (mask1), mask1, mask1, mask2);
3597 34 : if (mask == NULL_TREE || TREE_CODE (mask) != VECTOR_CST)
3598 : return 0;
3599 :
3600 34 : if (!VECTOR_CST_NELTS (mask).is_constant (&nelts))
3601 : return 0;
3602 60 : for (i = 0; i < nelts; i++)
3603 : {
3604 60 : tree val = VECTOR_CST_ELT (mask, i);
3605 60 : gcc_assert (TREE_CODE (val) == INTEGER_CST);
3606 60 : j = TREE_INT_CST_LOW (val) & (2 * nelts - 1);
3607 60 : if (j == i)
3608 : maybe_identity2 = false;
3609 47 : else if (j == i + nelts)
3610 : maybe_identity1 = false;
3611 : else
3612 : return 0;
3613 : }
3614 0 : return maybe_identity1 ? 1 : maybe_identity2 ? 2 : 0;
3615 : }
3616 :
3617 : /* Combine a shuffle with its arguments. Returns true if there were any
3618 : changes made. */
3619 :
3620 : static bool
3621 181766 : simplify_permutation (gimple_stmt_iterator *gsi)
3622 : {
3623 181766 : gimple *stmt = gsi_stmt (*gsi);
3624 181766 : gimple *def_stmt = NULL;
3625 181766 : tree op0, op1, op2, op3, arg0, arg1;
3626 181766 : enum tree_code code, code2 = ERROR_MARK;
3627 181766 : bool single_use_op0 = false;
3628 :
3629 181766 : gcc_checking_assert (gimple_assign_rhs_code (stmt) == VEC_PERM_EXPR);
3630 :
3631 181766 : op0 = gimple_assign_rhs1 (stmt);
3632 181766 : op1 = gimple_assign_rhs2 (stmt);
3633 181766 : op2 = gimple_assign_rhs3 (stmt);
3634 :
3635 181766 : if (TREE_CODE (op2) != VECTOR_CST)
3636 : return false;
3637 :
3638 179029 : if (TREE_CODE (op0) == VECTOR_CST)
3639 : {
3640 : code = VECTOR_CST;
3641 : arg0 = op0;
3642 : }
3643 177167 : else if (TREE_CODE (op0) == SSA_NAME)
3644 : {
3645 177167 : def_stmt = get_prop_source_stmt (op0, false, &single_use_op0);
3646 177167 : if (!def_stmt)
3647 : return false;
3648 169148 : code = gimple_assign_rhs_code (def_stmt);
3649 169148 : if (code == VIEW_CONVERT_EXPR)
3650 : {
3651 1385 : tree rhs = gimple_assign_rhs1 (def_stmt);
3652 1385 : tree name = TREE_OPERAND (rhs, 0);
3653 1385 : if (TREE_CODE (name) != SSA_NAME)
3654 : return false;
3655 1385 : if (!has_single_use (name))
3656 218 : single_use_op0 = false;
3657 : /* Here we update the def_stmt through this VIEW_CONVERT_EXPR,
3658 : but still keep the code to indicate it comes from
3659 : VIEW_CONVERT_EXPR. */
3660 1385 : def_stmt = SSA_NAME_DEF_STMT (name);
3661 1385 : if (!def_stmt || !is_gimple_assign (def_stmt))
3662 : return false;
3663 598 : if (gimple_assign_rhs_code (def_stmt) != CONSTRUCTOR)
3664 : return false;
3665 : }
3666 167862 : if (!can_propagate_from (def_stmt))
3667 : return false;
3668 19542 : arg0 = gimple_assign_rhs1 (def_stmt);
3669 : }
3670 : else
3671 : return false;
3672 :
3673 : /* Two consecutive shuffles. */
3674 19542 : if (code == VEC_PERM_EXPR)
3675 : {
3676 6287 : tree orig;
3677 6287 : int ident;
3678 :
3679 6287 : if (op0 != op1)
3680 : return false;
3681 34 : op3 = gimple_assign_rhs3 (def_stmt);
3682 34 : if (TREE_CODE (op3) != VECTOR_CST)
3683 : return false;
3684 34 : ident = is_combined_permutation_identity (op3, op2);
3685 34 : if (!ident)
3686 : return false;
3687 0 : orig = (ident == 1) ? gimple_assign_rhs1 (def_stmt)
3688 0 : : gimple_assign_rhs2 (def_stmt);
3689 0 : gimple_assign_set_rhs1 (stmt, unshare_expr (orig));
3690 0 : gimple_assign_set_rhs_code (stmt, TREE_CODE (orig));
3691 0 : gimple_set_num_ops (stmt, 2);
3692 0 : update_stmt (stmt);
3693 0 : remove_prop_source_from_use (op0);
3694 0 : return true;
3695 : }
3696 15117 : else if (code == CONSTRUCTOR
3697 15117 : || code == VECTOR_CST
3698 : || code == VIEW_CONVERT_EXPR)
3699 : {
3700 2622 : if (op0 != op1)
3701 : {
3702 2445 : if (TREE_CODE (op0) == SSA_NAME && !single_use_op0)
3703 : return false;
3704 :
3705 2090 : if (TREE_CODE (op1) == VECTOR_CST)
3706 : arg1 = op1;
3707 1601 : else if (TREE_CODE (op1) == SSA_NAME)
3708 : {
3709 1601 : gimple *def_stmt2 = get_prop_source_stmt (op1, true, NULL);
3710 1601 : if (!def_stmt2)
3711 : return false;
3712 165 : code2 = gimple_assign_rhs_code (def_stmt2);
3713 165 : if (code2 == VIEW_CONVERT_EXPR)
3714 : {
3715 4 : tree rhs = gimple_assign_rhs1 (def_stmt2);
3716 4 : tree name = TREE_OPERAND (rhs, 0);
3717 4 : if (TREE_CODE (name) != SSA_NAME)
3718 : return false;
3719 4 : if (!has_single_use (name))
3720 : return false;
3721 3 : def_stmt2 = SSA_NAME_DEF_STMT (name);
3722 3 : if (!def_stmt2 || !is_gimple_assign (def_stmt2))
3723 : return false;
3724 0 : if (gimple_assign_rhs_code (def_stmt2) != CONSTRUCTOR)
3725 : return false;
3726 : }
3727 161 : else if (code2 != CONSTRUCTOR && code2 != VECTOR_CST)
3728 : return false;
3729 47 : if (!can_propagate_from (def_stmt2))
3730 : return false;
3731 47 : arg1 = gimple_assign_rhs1 (def_stmt2);
3732 : }
3733 : else
3734 : return false;
3735 : }
3736 : else
3737 : {
3738 : /* Already used twice in this statement. */
3739 177 : if (TREE_CODE (op0) == SSA_NAME && num_imm_uses (op0) > 2)
3740 : return false;
3741 : arg1 = arg0;
3742 : }
3743 :
3744 : /* If there are any VIEW_CONVERT_EXPRs found when finding permutation
3745 : operands source, check whether it's valid to transform and prepare
3746 : the required new operands. */
3747 616 : if (code == VIEW_CONVERT_EXPR || code2 == VIEW_CONVERT_EXPR)
3748 : {
3749 : /* Figure out the target vector type to which operands should be
3750 : converted. If both are CONSTRUCTOR, the types should be the
3751 : same, otherwise, use the one of CONSTRUCTOR. */
3752 18 : tree tgt_type = NULL_TREE;
3753 18 : if (code == VIEW_CONVERT_EXPR)
3754 : {
3755 18 : gcc_assert (gimple_assign_rhs_code (def_stmt) == CONSTRUCTOR);
3756 18 : code = CONSTRUCTOR;
3757 18 : tgt_type = TREE_TYPE (arg0);
3758 : }
3759 18 : if (code2 == VIEW_CONVERT_EXPR)
3760 : {
3761 0 : tree arg1_type = TREE_TYPE (arg1);
3762 0 : if (tgt_type == NULL_TREE)
3763 : tgt_type = arg1_type;
3764 0 : else if (tgt_type != arg1_type)
3765 17 : return false;
3766 : }
3767 :
3768 18 : if (!VECTOR_TYPE_P (tgt_type))
3769 : return false;
3770 18 : tree op2_type = TREE_TYPE (op2);
3771 :
3772 : /* Figure out the shrunk factor. */
3773 18 : poly_uint64 tgt_units = TYPE_VECTOR_SUBPARTS (tgt_type);
3774 18 : poly_uint64 op2_units = TYPE_VECTOR_SUBPARTS (op2_type);
3775 18 : if (maybe_gt (tgt_units, op2_units))
3776 : return false;
3777 18 : unsigned int factor;
3778 35 : if (!constant_multiple_p (op2_units, tgt_units, &factor))
3779 : return false;
3780 :
3781 : /* Build the new permutation control vector as target vector. */
3782 18 : vec_perm_builder builder;
3783 18 : if (!tree_to_vec_perm_builder (&builder, op2))
3784 : return false;
3785 18 : vec_perm_indices indices (builder, 2, op2_units);
3786 18 : vec_perm_indices new_indices;
3787 18 : if (new_indices.new_shrunk_vector (indices, factor))
3788 : {
3789 1 : tree mask_type = tgt_type;
3790 1 : if (!VECTOR_INTEGER_TYPE_P (mask_type))
3791 : {
3792 0 : tree elem_type = TREE_TYPE (mask_type);
3793 0 : unsigned elem_size = TREE_INT_CST_LOW (TYPE_SIZE (elem_type));
3794 0 : tree int_type = build_nonstandard_integer_type (elem_size, 0);
3795 0 : mask_type = build_vector_type (int_type, tgt_units);
3796 : }
3797 1 : op2 = vec_perm_indices_to_tree (mask_type, new_indices);
3798 : }
3799 : else
3800 17 : return false;
3801 :
3802 : /* Convert the VECTOR_CST to the appropriate vector type. */
3803 1 : if (tgt_type != TREE_TYPE (arg0))
3804 0 : arg0 = fold_build1 (VIEW_CONVERT_EXPR, tgt_type, arg0);
3805 1 : else if (tgt_type != TREE_TYPE (arg1))
3806 0 : arg1 = fold_build1 (VIEW_CONVERT_EXPR, tgt_type, arg1);
3807 35 : }
3808 :
3809 : /* VIEW_CONVERT_EXPR should be updated to CONSTRUCTOR before. */
3810 599 : gcc_assert (code == CONSTRUCTOR || code == VECTOR_CST);
3811 :
3812 : /* Shuffle of a constructor. */
3813 599 : tree res_type
3814 599 : = build_vector_type (TREE_TYPE (TREE_TYPE (arg0)),
3815 599 : TYPE_VECTOR_SUBPARTS (TREE_TYPE (op2)));
3816 599 : tree opt = fold_ternary (VEC_PERM_EXPR, res_type, arg0, arg1, op2);
3817 599 : if (!opt
3818 280 : || (TREE_CODE (opt) != CONSTRUCTOR && TREE_CODE (opt) != VECTOR_CST))
3819 : return false;
3820 : /* Found VIEW_CONVERT_EXPR before, need one explicit conversion. */
3821 280 : if (res_type != TREE_TYPE (op0))
3822 : {
3823 1 : tree name = make_ssa_name (TREE_TYPE (opt));
3824 1 : gimple *ass_stmt = gimple_build_assign (name, opt);
3825 1 : gsi_insert_before (gsi, ass_stmt, GSI_SAME_STMT);
3826 1 : opt = build1 (VIEW_CONVERT_EXPR, TREE_TYPE (op0), name);
3827 : }
3828 280 : gimple_assign_set_rhs_from_tree (gsi, opt);
3829 280 : update_stmt (gsi_stmt (*gsi));
3830 280 : if (TREE_CODE (op0) == SSA_NAME)
3831 1 : remove_prop_source_from_use (op0);
3832 280 : if (op0 != op1 && TREE_CODE (op1) == SSA_NAME)
3833 0 : remove_prop_source_from_use (op1);
3834 280 : return true;
3835 : }
3836 :
3837 : return false;
3838 : }
3839 :
3840 : /* Get the BIT_FIELD_REF definition of VAL, if any, looking through
3841 : conversions with code CONV_CODE or update it if still ERROR_MARK.
3842 : Return NULL_TREE if no such matching def was found. */
3843 :
3844 : static tree
3845 391344 : get_bit_field_ref_def (tree val, enum tree_code &conv_code)
3846 : {
3847 391344 : if (TREE_CODE (val) != SSA_NAME)
3848 : return NULL_TREE ;
3849 365520 : gimple *def_stmt = get_prop_source_stmt (val, false, NULL);
3850 365520 : if (!def_stmt)
3851 : return NULL_TREE;
3852 292456 : enum tree_code code = gimple_assign_rhs_code (def_stmt);
3853 292456 : if (code == FLOAT_EXPR
3854 292456 : || code == FIX_TRUNC_EXPR
3855 : || CONVERT_EXPR_CODE_P (code))
3856 : {
3857 178852 : tree op1 = gimple_assign_rhs1 (def_stmt);
3858 178852 : if (conv_code == ERROR_MARK)
3859 86117 : conv_code = code;
3860 92735 : else if (conv_code != code)
3861 : return NULL_TREE;
3862 178827 : if (TREE_CODE (op1) != SSA_NAME)
3863 : return NULL_TREE;
3864 71971 : def_stmt = SSA_NAME_DEF_STMT (op1);
3865 71971 : if (! is_gimple_assign (def_stmt))
3866 : return NULL_TREE;
3867 57228 : code = gimple_assign_rhs_code (def_stmt);
3868 : }
3869 170832 : if (code != BIT_FIELD_REF)
3870 : return NULL_TREE;
3871 22859 : return gimple_assign_rhs1 (def_stmt);
3872 : }
3873 :
3874 : /* Recognize a VEC_PERM_EXPR. Returns true if there were any changes. */
3875 :
3876 : static bool
3877 147562 : simplify_vector_constructor (gimple_stmt_iterator *gsi)
3878 : {
3879 147562 : gimple *stmt = gsi_stmt (*gsi);
3880 147562 : tree op, orig[2], type;
3881 147562 : unsigned i;
3882 147562 : unsigned HOST_WIDE_INT nelts;
3883 147562 : unsigned HOST_WIDE_INT refnelts;
3884 147562 : enum tree_code conv_code;
3885 147562 : constructor_elt *elt;
3886 :
3887 147562 : op = gimple_assign_rhs1 (stmt);
3888 147562 : type = TREE_TYPE (op);
3889 147562 : gcc_checking_assert (TREE_CODE (op) == CONSTRUCTOR
3890 : && TREE_CODE (type) == VECTOR_TYPE);
3891 :
3892 147562 : if (!TYPE_VECTOR_SUBPARTS (type).is_constant (&nelts))
3893 : return false;
3894 :
3895 147562 : orig[0] = NULL;
3896 147562 : orig[1] = NULL;
3897 147562 : tree orig_elem_type[2] = {};
3898 147562 : conv_code = ERROR_MARK;
3899 147562 : bool maybe_ident = true;
3900 147562 : bool maybe_blend[2] = { true, true };
3901 147562 : tree one_constant = NULL_TREE;
3902 147562 : tree one_nonconstant = NULL_TREE;
3903 147562 : tree subelt;
3904 147562 : auto_vec<tree> constants;
3905 147562 : constants.safe_grow_cleared (nelts, true);
3906 147562 : auto_vec<std::pair<unsigned, unsigned>, 64> elts;
3907 147562 : unsigned int tsubelts = 0;
3908 422149 : FOR_EACH_VEC_SAFE_ELT (CONSTRUCTOR_ELTS (op), i, elt)
3909 : {
3910 391344 : tree ref, op1;
3911 391344 : unsigned int elem, src_elem_size;
3912 391344 : unsigned HOST_WIDE_INT nsubelts = 1;
3913 :
3914 391344 : if (i >= nelts)
3915 147562 : return false;
3916 :
3917 : /* Look for elements extracted and possibly converted from
3918 : another vector. */
3919 391344 : op1 = get_bit_field_ref_def (elt->value, conv_code);
3920 391344 : if (op1
3921 22859 : && TREE_CODE ((ref = TREE_OPERAND (op1, 0))) == SSA_NAME
3922 4944 : && VECTOR_TYPE_P (TREE_TYPE (ref))
3923 4941 : && (tree_nop_conversion_p (TREE_TYPE (op1),
3924 4941 : TREE_TYPE (TREE_TYPE (ref)))
3925 673 : || (VECTOR_TYPE_P (TREE_TYPE (op1))
3926 29 : && tree_nop_conversion_p (TREE_TYPE (TREE_TYPE (op1)),
3927 29 : TREE_TYPE (TREE_TYPE (ref)))
3928 29 : && TYPE_VECTOR_SUBPARTS (TREE_TYPE (op1))
3929 29 : .is_constant (&nsubelts)))
3930 4297 : && constant_multiple_p (bit_field_size (op1), nsubelts,
3931 : &src_elem_size)
3932 395641 : && constant_multiple_p (bit_field_offset (op1), src_elem_size, &elem)
3933 395641 : && TYPE_VECTOR_SUBPARTS (TREE_TYPE (ref)).is_constant (&refnelts))
3934 : {
3935 : unsigned int j;
3936 4558 : for (j = 0; j < 2; ++j)
3937 : {
3938 4541 : if (!orig[j])
3939 : {
3940 2112 : if (j == 0
3941 2302 : || useless_type_conversion_p (TREE_TYPE (orig[0]),
3942 190 : TREE_TYPE (ref)))
3943 : break;
3944 : }
3945 2429 : else if (ref == orig[j])
3946 : break;
3947 : }
3948 : /* Found a suitable vector element. */
3949 4297 : if (j < 2)
3950 : {
3951 4280 : orig[j] = ref;
3952 : /* Track what element type was actually extracted (which may
3953 : differ in signedness from the vector's element type due to
3954 : tree_nop_conversion_p). */
3955 4280 : if (!orig_elem_type[j])
3956 2106 : orig_elem_type[j] = TREE_TYPE (op1);
3957 4280 : if (elem != i || j != 0)
3958 2088 : maybe_ident = false;
3959 4280 : if (elem != i)
3960 2019 : maybe_blend[j] = false;
3961 8581 : for (unsigned int k = 0; k < nsubelts; ++k)
3962 4301 : elts.safe_push (std::make_pair (j, elem + k));
3963 4280 : tsubelts += nsubelts;
3964 4280 : continue;
3965 4280 : }
3966 : /* Else fallthru. */
3967 : }
3968 : /* Handle elements not extracted from a vector.
3969 : 1. constants by permuting with constant vector
3970 : 2. a unique non-constant element by permuting with a splat vector */
3971 387064 : if (orig[1]
3972 240551 : && orig[1] != error_mark_node)
3973 : return false;
3974 387048 : orig[1] = error_mark_node;
3975 387048 : if (VECTOR_TYPE_P (TREE_TYPE (elt->value))
3976 387048 : && !TYPE_VECTOR_SUBPARTS (TREE_TYPE (elt->value))
3977 1761 : .is_constant (&nsubelts))
3978 : return false;
3979 387048 : if (CONSTANT_CLASS_P (elt->value))
3980 : {
3981 25820 : if (one_nonconstant)
3982 : return false;
3983 17389 : if (!one_constant)
3984 8167 : one_constant = TREE_CODE (elt->value) == VECTOR_CST
3985 8167 : ? VECTOR_CST_ELT (elt->value, 0)
3986 : : elt->value;
3987 17389 : if (TREE_CODE (elt->value) == VECTOR_CST)
3988 : {
3989 37 : for (unsigned int k = 0; k < nsubelts; k++)
3990 23 : constants[tsubelts + k] = VECTOR_CST_ELT (elt->value, k);
3991 : }
3992 : else
3993 17375 : constants[tsubelts] = elt->value;
3994 : }
3995 : else
3996 : {
3997 361228 : if (one_constant)
3998 : return false;
3999 353340 : subelt = VECTOR_TYPE_P (TREE_TYPE (elt->value))
4000 353340 : ? ssa_uniform_vector_p (elt->value)
4001 : : elt->value;
4002 353340 : if (!subelt)
4003 : return false;
4004 351603 : if (!one_nonconstant)
4005 : one_nonconstant = subelt;
4006 214994 : else if (!operand_equal_p (one_nonconstant, subelt, 0))
4007 : return false;
4008 : }
4009 540625 : for (unsigned int k = 0; k < nsubelts; ++k)
4010 270318 : elts.safe_push (std::make_pair (1, tsubelts + k));
4011 270307 : tsubelts += nsubelts;
4012 270307 : maybe_ident = false;
4013 : }
4014 :
4015 61610 : if (elts.length () < nelts)
4016 : return false;
4017 :
4018 30020 : if (! orig[0]
4019 30020 : || ! VECTOR_TYPE_P (TREE_TYPE (orig[0])))
4020 : return false;
4021 1484 : refnelts = TYPE_VECTOR_SUBPARTS (TREE_TYPE (orig[0])).to_constant ();
4022 : /* We currently do not handle larger destination vectors. */
4023 1484 : if (refnelts < nelts)
4024 : return false;
4025 :
4026 : /* Determine the element type for the conversion source.
4027 : As orig_elem_type keeps track of the original type, check
4028 : if we need to perform a sign swap after permuting.
4029 : We need to be able to construct a vector type from the element
4030 : type which is not possible for e.g. BitInt or pointers
4031 : so pun with an integer type if needed. */
4032 1343 : tree perm_eltype = TREE_TYPE (TREE_TYPE (orig[0]));
4033 1343 : bool sign_change_p = false;
4034 1343 : if (conv_code != ERROR_MARK
4035 322 : && orig_elem_type[0]
4036 1665 : && TYPE_SIGN (orig_elem_type[0]) != TYPE_SIGN (perm_eltype))
4037 : {
4038 34 : perm_eltype = signed_or_unsigned_type_for
4039 34 : (TYPE_UNSIGNED (orig_elem_type[0]), perm_eltype);
4040 34 : sign_change_p = true;
4041 : }
4042 1343 : tree conv_src_type = build_vector_type (perm_eltype, nelts);
4043 :
4044 1343 : if (maybe_ident)
4045 : {
4046 : /* When there is no conversion, use the target type directly. */
4047 376 : if (conv_code == ERROR_MARK && nelts != refnelts)
4048 376 : conv_src_type = type;
4049 376 : if (conv_code != ERROR_MARK
4050 376 : && !supportable_convert_operation (conv_code, type, conv_src_type,
4051 : &conv_code))
4052 : {
4053 : /* Only few targets implement direct conversion patterns so try
4054 : some simple special cases via VEC_[UN]PACK[_FLOAT]_LO_EXPR. */
4055 5 : optab optab;
4056 5 : insn_code icode;
4057 5 : tree halfvectype, dblvectype;
4058 5 : enum tree_code unpack_op;
4059 :
4060 5 : if (!BYTES_BIG_ENDIAN)
4061 5 : unpack_op = (FLOAT_TYPE_P (TREE_TYPE (type))
4062 5 : ? VEC_UNPACK_FLOAT_LO_EXPR
4063 : : VEC_UNPACK_LO_EXPR);
4064 : else
4065 : unpack_op = (FLOAT_TYPE_P (TREE_TYPE (type))
4066 : ? VEC_UNPACK_FLOAT_HI_EXPR
4067 : : VEC_UNPACK_HI_EXPR);
4068 :
4069 : /* Conversions between DFP and FP have no special tree code
4070 : but we cannot handle those since all relevant vector conversion
4071 : optabs only have a single mode. */
4072 3 : if (CONVERT_EXPR_CODE_P (conv_code)
4073 2 : && FLOAT_TYPE_P (TREE_TYPE (type))
4074 9 : && (DECIMAL_FLOAT_TYPE_P (TREE_TYPE (type))
4075 2 : != DECIMAL_FLOAT_TYPE_P (TREE_TYPE (conv_src_type))))
4076 : return false;
4077 :
4078 3 : if (CONVERT_EXPR_CODE_P (conv_code)
4079 1 : && (2 * TYPE_PRECISION (TREE_TYPE (TREE_TYPE (orig[0])))
4080 1 : == TYPE_PRECISION (TREE_TYPE (type)))
4081 0 : && orig_elem_type[0]
4082 0 : && useless_type_conversion_p (orig_elem_type[0],
4083 0 : TREE_TYPE (TREE_TYPE (orig[0])))
4084 0 : && mode_for_vector (as_a <scalar_mode>
4085 0 : (TYPE_MODE (TREE_TYPE (TREE_TYPE (orig[0])))),
4086 0 : nelts * 2).exists ()
4087 0 : && (dblvectype
4088 0 : = build_vector_type (TREE_TYPE (TREE_TYPE (orig[0])),
4089 0 : nelts * 2))
4090 : /* Only use it for vector modes or for vector booleans
4091 : represented as scalar bitmasks. See PR95528. */
4092 0 : && (VECTOR_MODE_P (TYPE_MODE (dblvectype))
4093 0 : || VECTOR_BOOLEAN_TYPE_P (dblvectype))
4094 0 : && (optab = optab_for_tree_code (unpack_op,
4095 : dblvectype,
4096 : optab_default))
4097 0 : && ((icode = optab_handler (optab, TYPE_MODE (dblvectype)))
4098 : != CODE_FOR_nothing)
4099 4 : && (insn_data[icode].operand[0].mode == TYPE_MODE (type)))
4100 : {
4101 0 : gimple_seq stmts = NULL;
4102 0 : tree dbl;
4103 0 : if (refnelts == nelts)
4104 : {
4105 : /* ??? Paradoxical subregs don't exist, so insert into
4106 : the lower half of a wider zero vector. */
4107 0 : dbl = gimple_build (&stmts, BIT_INSERT_EXPR, dblvectype,
4108 : build_zero_cst (dblvectype), orig[0],
4109 0 : bitsize_zero_node);
4110 : }
4111 0 : else if (refnelts == 2 * nelts)
4112 : dbl = orig[0];
4113 : else
4114 0 : dbl = gimple_build (&stmts, BIT_FIELD_REF, dblvectype,
4115 0 : orig[0], TYPE_SIZE (dblvectype),
4116 0 : bitsize_zero_node);
4117 0 : gsi_insert_seq_before (gsi, stmts, GSI_SAME_STMT);
4118 0 : gimple_assign_set_rhs_with_ops (gsi, unpack_op, dbl);
4119 : }
4120 3 : else if (CONVERT_EXPR_CODE_P (conv_code)
4121 1 : && (TYPE_PRECISION (TREE_TYPE (TREE_TYPE (orig[0])))
4122 1 : == 2 * TYPE_PRECISION (TREE_TYPE (type)))
4123 1 : && orig_elem_type[0]
4124 1 : && useless_type_conversion_p (orig_elem_type[0],
4125 1 : TREE_TYPE (TREE_TYPE (orig[0])))
4126 1 : && mode_for_vector (as_a <scalar_mode>
4127 1 : (TYPE_MODE
4128 : (TREE_TYPE (TREE_TYPE (orig[0])))),
4129 2 : nelts / 2).exists ()
4130 1 : && (halfvectype
4131 1 : = build_vector_type (TREE_TYPE (TREE_TYPE (orig[0])),
4132 1 : nelts / 2))
4133 : /* Only use it for vector modes or for vector booleans
4134 : represented as scalar bitmasks. See PR95528. */
4135 1 : && (VECTOR_MODE_P (TYPE_MODE (halfvectype))
4136 0 : || VECTOR_BOOLEAN_TYPE_P (halfvectype))
4137 1 : && (optab = optab_for_tree_code (VEC_PACK_TRUNC_EXPR,
4138 : halfvectype,
4139 : optab_default))
4140 1 : && ((icode = optab_handler (optab, TYPE_MODE (halfvectype)))
4141 : != CODE_FOR_nothing)
4142 5 : && (insn_data[icode].operand[0].mode == TYPE_MODE (type)))
4143 : {
4144 0 : gimple_seq stmts = NULL;
4145 0 : tree low = gimple_build (&stmts, BIT_FIELD_REF, halfvectype,
4146 0 : orig[0], TYPE_SIZE (halfvectype),
4147 0 : bitsize_zero_node);
4148 0 : tree hig = gimple_build (&stmts, BIT_FIELD_REF, halfvectype,
4149 0 : orig[0], TYPE_SIZE (halfvectype),
4150 0 : TYPE_SIZE (halfvectype));
4151 0 : gsi_insert_seq_before (gsi, stmts, GSI_SAME_STMT);
4152 0 : gimple_assign_set_rhs_with_ops (gsi, VEC_PACK_TRUNC_EXPR,
4153 : low, hig);
4154 : }
4155 : else
4156 4 : return false;
4157 0 : update_stmt (gsi_stmt (*gsi));
4158 0 : return true;
4159 : }
4160 371 : if (nelts != refnelts)
4161 : {
4162 20 : gassign *lowpart
4163 20 : = gimple_build_assign (make_ssa_name (conv_src_type),
4164 : build3 (BIT_FIELD_REF, conv_src_type,
4165 20 : orig[0], TYPE_SIZE (conv_src_type),
4166 : bitsize_zero_node));
4167 20 : gsi_insert_before (gsi, lowpart, GSI_SAME_STMT);
4168 20 : orig[0] = gimple_assign_lhs (lowpart);
4169 : }
4170 351 : else if (sign_change_p)
4171 : {
4172 0 : gassign *conv
4173 0 : = gimple_build_assign (make_ssa_name (conv_src_type),
4174 : build1 (VIEW_CONVERT_EXPR, conv_src_type,
4175 : orig[0]));
4176 0 : gsi_insert_before (gsi, conv, GSI_SAME_STMT);
4177 0 : orig[0] = gimple_assign_lhs (conv);
4178 : }
4179 371 : if (conv_code == ERROR_MARK)
4180 : {
4181 354 : tree src_type = TREE_TYPE (orig[0]);
4182 354 : if (!useless_type_conversion_p (type, src_type))
4183 : {
4184 0 : gcc_assert (known_eq (TYPE_VECTOR_SUBPARTS (type),
4185 : TYPE_VECTOR_SUBPARTS (src_type))
4186 : && tree_nop_conversion_p (TREE_TYPE (type),
4187 : TREE_TYPE (src_type)));
4188 0 : tree rhs = build1 (VIEW_CONVERT_EXPR, type, orig[0]);
4189 0 : orig[0] = make_ssa_name (type);
4190 0 : gassign *assign = gimple_build_assign (orig[0], rhs);
4191 0 : gsi_insert_before (gsi, assign, GSI_SAME_STMT);
4192 : }
4193 354 : gimple_assign_set_rhs_from_tree (gsi, orig[0]);
4194 : }
4195 : else
4196 17 : gimple_assign_set_rhs_with_ops (gsi, conv_code, orig[0],
4197 : NULL_TREE, NULL_TREE);
4198 : }
4199 : else
4200 : {
4201 : /* If we combine a vector with a non-vector avoid cases where
4202 : we'll obviously end up with more GIMPLE stmts which is when
4203 : we'll later not fold this to a single insert into the vector
4204 : and we had a single extract originally. See PR92819. */
4205 967 : if (nelts == 2
4206 758 : && refnelts > 2
4207 168 : && orig[1] == error_mark_node
4208 33 : && !maybe_blend[0])
4209 379 : return false;
4210 940 : tree mask_type, perm_type;
4211 940 : perm_type = TREE_TYPE (orig[0]);
4212 940 : if (conv_code != ERROR_MARK
4213 940 : && !supportable_convert_operation (conv_code, type, conv_src_type,
4214 : &conv_code))
4215 : return false;
4216 :
4217 : /* Now that we know the number of elements of the source build the
4218 : permute vector.
4219 : ??? When the second vector has constant values we can shuffle
4220 : it and its source indexes to make the permutation supported.
4221 : For now it mimics a blend. */
4222 705 : vec_perm_builder sel (refnelts, refnelts, 1);
4223 705 : bool all_same_p = true;
4224 6050 : for (i = 0; i < elts.length (); ++i)
4225 : {
4226 2320 : sel.quick_push (elts[i].second + elts[i].first * refnelts);
4227 2320 : all_same_p &= known_eq (sel[i], sel[0]);
4228 : }
4229 : /* And fill the tail with "something". It's really don't care,
4230 : and ideally we'd allow VEC_PERM to have a smaller destination
4231 : vector. As a heuristic:
4232 :
4233 : (a) if what we have so far duplicates a single element, make the
4234 : tail do the same
4235 :
4236 : (b) otherwise preserve a uniform orig[0]. This facilitates
4237 : later pattern-matching of VEC_PERM_EXPR to a BIT_INSERT_EXPR. */
4238 1471 : for (; i < refnelts; ++i)
4239 1532 : sel.quick_push (all_same_p
4240 2298 : ? sel[0]
4241 48 : : (elts[0].second == 0 && elts[0].first == 0
4242 1012 : ? 0 : refnelts) + i);
4243 945 : vec_perm_indices indices (sel, orig[1] ? 2 : 1, refnelts);
4244 705 : machine_mode vmode = TYPE_MODE (perm_type);
4245 705 : if (!can_vec_perm_const_p (vmode, vmode, indices))
4246 : return false;
4247 588 : mask_type = build_vector_type (ssizetype, refnelts);
4248 588 : tree op2 = vec_perm_indices_to_tree (mask_type, indices);
4249 588 : bool converted_orig1 = false;
4250 588 : gimple_seq stmts = NULL;
4251 588 : if (!orig[1])
4252 197 : orig[1] = orig[0];
4253 391 : else if (orig[1] == error_mark_node
4254 242 : && one_nonconstant)
4255 : {
4256 : /* ??? We can see if we can safely convert to the original
4257 : element type. */
4258 155 : converted_orig1 = conv_code != ERROR_MARK;
4259 155 : tree target_type = converted_orig1 ? type : perm_type;
4260 155 : tree nonconstant_for_splat = one_nonconstant;
4261 : /* If there's a nop conversion between the target element type and
4262 : the nonconstant's type, convert it. */
4263 155 : if (!useless_type_conversion_p (TREE_TYPE (target_type),
4264 155 : TREE_TYPE (one_nonconstant)))
4265 0 : nonconstant_for_splat
4266 0 : = gimple_build (&stmts, NOP_EXPR, TREE_TYPE (target_type),
4267 : one_nonconstant);
4268 155 : orig[1] = gimple_build_vector_from_val (&stmts, UNKNOWN_LOCATION,
4269 : target_type,
4270 : nonconstant_for_splat);
4271 155 : }
4272 236 : else if (orig[1] == error_mark_node)
4273 : {
4274 : /* ??? See if we can convert the vector to the original type. */
4275 87 : converted_orig1 = conv_code != ERROR_MARK;
4276 87 : unsigned n = converted_orig1 ? nelts : refnelts;
4277 70 : tree target_type = converted_orig1 ? type : perm_type;
4278 87 : tree_vector_builder vec (target_type, n, 1);
4279 533 : for (unsigned i = 0; i < n; ++i)
4280 864 : if (i < nelts && constants[i])
4281 : {
4282 225 : tree constant = constants[i];
4283 : /* If there's a nop conversion, convert the constant. */
4284 225 : if (!useless_type_conversion_p (TREE_TYPE (target_type),
4285 225 : TREE_TYPE (constant)))
4286 0 : constant = fold_convert (TREE_TYPE (target_type), constant);
4287 225 : vec.quick_push (constant);
4288 : }
4289 : else
4290 : {
4291 : /* ??? Push a don't-care value. */
4292 221 : tree constant = one_constant;
4293 221 : if (!useless_type_conversion_p (TREE_TYPE (target_type),
4294 221 : TREE_TYPE (constant)))
4295 0 : constant = fold_convert (TREE_TYPE (target_type), constant);
4296 221 : vec.quick_push (constant);
4297 : }
4298 87 : orig[1] = vec.build ();
4299 87 : }
4300 439 : tree blend_op2 = NULL_TREE;
4301 439 : if (converted_orig1)
4302 : {
4303 : /* Make sure we can do a blend in the target type. */
4304 19 : vec_perm_builder sel (nelts, nelts, 1);
4305 87 : for (i = 0; i < elts.length (); ++i)
4306 68 : sel.quick_push (elts[i].first
4307 68 : ? elts[i].second + nelts : i);
4308 19 : vec_perm_indices indices (sel, 2, nelts);
4309 19 : machine_mode vmode = TYPE_MODE (type);
4310 19 : if (!can_vec_perm_const_p (vmode, vmode, indices))
4311 0 : return false;
4312 19 : mask_type = build_vector_type (ssizetype, nelts);
4313 19 : blend_op2 = vec_perm_indices_to_tree (mask_type, indices);
4314 19 : }
4315 :
4316 : /* For a real orig[1] (no splat, constant etc.) we might need to
4317 : nop-convert it. Do so here. */
4318 588 : if (orig[1] && orig[1] != error_mark_node
4319 588 : && !useless_type_conversion_p (perm_type, TREE_TYPE (orig[1]))
4320 607 : && tree_nop_conversion_p (TREE_TYPE (perm_type),
4321 19 : TREE_TYPE (TREE_TYPE (orig[1]))))
4322 0 : orig[1] = gimple_build (&stmts, VIEW_CONVERT_EXPR, perm_type,
4323 : orig[1]);
4324 :
4325 588 : tree orig1_for_perm
4326 588 : = converted_orig1 ? build_zero_cst (perm_type) : orig[1];
4327 588 : tree res = gimple_build (&stmts, VEC_PERM_EXPR, perm_type,
4328 : orig[0], orig1_for_perm, op2);
4329 : /* If we're building a smaller vector, extract the element
4330 : with the proper type. */
4331 588 : if (nelts != refnelts)
4332 294 : res = gimple_build (&stmts, BIT_FIELD_REF,
4333 : conv_code != ERROR_MARK ? conv_src_type : type,
4334 : res,
4335 147 : TYPE_SIZE (conv_code != ERROR_MARK ? conv_src_type
4336 : : type),
4337 147 : bitsize_zero_node);
4338 : /* Otherwise, we can still have an intermediate sign change.
4339 : ??? In that case we have two subsequent conversions.
4340 : We should be able to merge them. */
4341 441 : else if (sign_change_p)
4342 14 : res = gimple_build (&stmts, VIEW_CONVERT_EXPR, conv_src_type, res);
4343 : /* Finally, apply the conversion. */
4344 588 : if (conv_code != ERROR_MARK)
4345 52 : res = gimple_build (&stmts, conv_code, type, res);
4346 536 : else if (!useless_type_conversion_p (type, TREE_TYPE (res)))
4347 : {
4348 1 : gcc_assert (known_eq (TYPE_VECTOR_SUBPARTS (type),
4349 : TYPE_VECTOR_SUBPARTS (perm_type))
4350 : && tree_nop_conversion_p (TREE_TYPE (type),
4351 : TREE_TYPE (perm_type)));
4352 1 : res = gimple_build (&stmts, VIEW_CONVERT_EXPR, type, res);
4353 : }
4354 : /* Blend in the actual constant. */
4355 588 : if (converted_orig1)
4356 19 : res = gimple_build (&stmts, VEC_PERM_EXPR, type,
4357 19 : res, orig[1], blend_op2);
4358 588 : gsi_insert_seq_before (gsi, stmts, GSI_SAME_STMT);
4359 588 : gimple_assign_set_rhs_with_ops (gsi, SSA_NAME, res);
4360 705 : }
4361 959 : update_stmt (gsi_stmt (*gsi));
4362 959 : return true;
4363 147562 : }
4364 :
4365 : /* Prepare a TARGET_MEM_REF ref so that it can be subsetted as
4366 : lvalue. This splits out an address computation stmt before *GSI
4367 : and returns a MEM_REF wrapping the address. */
4368 :
4369 : static tree
4370 1096 : prepare_target_mem_ref_lvalue (tree ref, gimple_stmt_iterator *gsi)
4371 : {
4372 1096 : if (TREE_CODE (TREE_OPERAND (ref, 0)) == ADDR_EXPR)
4373 215 : mark_addressable (TREE_OPERAND (TREE_OPERAND (ref, 0), 0));
4374 1096 : tree ptrtype = build_pointer_type (TREE_TYPE (ref));
4375 1096 : tree tem = make_ssa_name (ptrtype);
4376 1096 : gimple *new_stmt
4377 1096 : = gimple_build_assign (tem, build1 (ADDR_EXPR, TREE_TYPE (tem),
4378 : unshare_expr (ref)));
4379 1096 : gsi_insert_before (gsi, new_stmt, GSI_SAME_STMT);
4380 2192 : ref = build2_loc (EXPR_LOCATION (ref),
4381 1096 : MEM_REF, TREE_TYPE (ref), tem,
4382 1096 : build_int_cst (TREE_TYPE (TREE_OPERAND (ref, 1)), 0));
4383 1096 : return ref;
4384 : }
4385 :
4386 : /* Rewrite the vector load at *GSI to component-wise loads if the load
4387 : is only used in BIT_FIELD_REF extractions with eventual intermediate
4388 : widening. */
4389 :
4390 : static void
4391 283431 : optimize_vector_load (gimple_stmt_iterator *gsi)
4392 : {
4393 283431 : gimple *stmt = gsi_stmt (*gsi);
4394 283431 : tree lhs = gimple_assign_lhs (stmt);
4395 283431 : tree rhs = gimple_assign_rhs1 (stmt);
4396 283431 : tree vuse = gimple_vuse (stmt);
4397 :
4398 : /* Gather BIT_FIELD_REFs to rewrite, looking through
4399 : VEC_UNPACK_{LO,HI}_EXPR. */
4400 283431 : use_operand_p use_p;
4401 283431 : imm_use_iterator iter;
4402 283431 : bool rewrite = true;
4403 283431 : bool scalar_use = false;
4404 283431 : bool unpack_use = false;
4405 283431 : auto_vec<gimple *, 8> bf_stmts;
4406 283431 : auto_vec<tree, 8> worklist;
4407 283431 : worklist.quick_push (lhs);
4408 285185 : do
4409 : {
4410 285185 : tree def = worklist.pop ();
4411 285185 : unsigned HOST_WIDE_INT def_eltsize
4412 285185 : = TREE_INT_CST_LOW (TYPE_SIZE (TREE_TYPE (TREE_TYPE (def))));
4413 644874 : FOR_EACH_IMM_USE_FAST (use_p, iter, def)
4414 : {
4415 339502 : gimple *use_stmt = USE_STMT (use_p);
4416 339502 : if (is_gimple_debug (use_stmt))
4417 74504 : continue;
4418 339439 : if (!is_gimple_assign (use_stmt))
4419 : {
4420 : rewrite = false;
4421 264998 : break;
4422 : }
4423 304177 : enum tree_code use_code = gimple_assign_rhs_code (use_stmt);
4424 304177 : tree use_rhs = gimple_assign_rhs1 (use_stmt);
4425 375099 : if (use_code == BIT_FIELD_REF
4426 70923 : && TREE_OPERAND (use_rhs, 0) == def
4427 : /* If its on the VEC_UNPACK_{HI,LO}_EXPR
4428 : def need to verify it is element aligned. */
4429 375100 : && (def == lhs
4430 85 : || (known_eq (bit_field_size (use_rhs), def_eltsize)
4431 85 : && constant_multiple_p (bit_field_offset (use_rhs),
4432 : def_eltsize)
4433 : /* We can simulate the VEC_UNPACK_{HI,LO}_EXPR
4434 : via a NOP_EXPR only for integral types.
4435 : ??? Support VEC_UNPACK_FLOAT_{HI,LO}_EXPR. */
4436 85 : && INTEGRAL_TYPE_P (TREE_TYPE (use_rhs)))))
4437 : {
4438 70922 : if (!VECTOR_TYPE_P (TREE_TYPE (gimple_assign_lhs (use_stmt))))
4439 68854 : scalar_use = true;
4440 70922 : bf_stmts.safe_push (use_stmt);
4441 70922 : continue;
4442 : }
4443 : /* Walk through one level of VEC_UNPACK_{LO,HI}_EXPR. */
4444 233255 : if (def == lhs
4445 231548 : && (use_code == VEC_UNPACK_HI_EXPR
4446 231548 : || use_code == VEC_UNPACK_LO_EXPR)
4447 3519 : && use_rhs == lhs)
4448 : {
4449 3519 : unpack_use = true;
4450 3519 : worklist.safe_push (gimple_assign_lhs (use_stmt));
4451 3519 : continue;
4452 : }
4453 : rewrite = false;
4454 : break;
4455 285185 : }
4456 285185 : if (!rewrite)
4457 : break;
4458 : }
4459 40374 : while (!worklist.is_empty ());
4460 :
4461 283431 : rewrite = rewrite && (scalar_use
4462 18433 : || unpack_use
4463 575 : || !can_implement_p (mov_optab,
4464 575 : TYPE_MODE (TREE_TYPE (lhs))));
4465 283431 : if (!rewrite)
4466 : {
4467 265173 : gsi_next (gsi);
4468 265173 : return;
4469 : }
4470 : /* We now have all ultimate uses of the load to rewrite in bf_stmts. */
4471 :
4472 : /* Prepare the original ref to be wrapped in adjusted BIT_FIELD_REFs.
4473 : For TARGET_MEM_REFs we have to separate the LEA from the reference. */
4474 18258 : tree load_rhs = rhs;
4475 18258 : if (TREE_CODE (load_rhs) == TARGET_MEM_REF)
4476 1095 : load_rhs = prepare_target_mem_ref_lvalue (load_rhs, gsi);
4477 :
4478 : /* Rewrite the BIT_FIELD_REFs to be actual loads, re-emitting them at
4479 : the place of the original load. */
4480 120133 : for (gimple *use_stmt : bf_stmts)
4481 : {
4482 65359 : tree bfr = gimple_assign_rhs1 (use_stmt);
4483 65359 : tree new_rhs = unshare_expr (load_rhs);
4484 65359 : if (TREE_OPERAND (bfr, 0) != lhs)
4485 : {
4486 : /* When the BIT_FIELD_REF is on the promoted vector we have to
4487 : adjust it and emit a conversion afterwards. */
4488 84 : gimple *def_stmt
4489 84 : = SSA_NAME_DEF_STMT (TREE_OPERAND (bfr, 0));
4490 84 : enum tree_code def_code
4491 84 : = gimple_assign_rhs_code (def_stmt);
4492 :
4493 : /* The adjusted BIT_FIELD_REF is of the promotion source
4494 : vector size and at half of the offset... */
4495 84 : new_rhs = fold_build3 (BIT_FIELD_REF,
4496 : TREE_TYPE (TREE_TYPE (lhs)),
4497 : new_rhs,
4498 : TYPE_SIZE (TREE_TYPE (TREE_TYPE (lhs))),
4499 : size_binop (EXACT_DIV_EXPR,
4500 : TREE_OPERAND (bfr, 2),
4501 : bitsize_int (2)));
4502 : /* ... and offsetted by half of the vector if VEC_UNPACK_HI_EXPR. */
4503 84 : if (def_code == (!BYTES_BIG_ENDIAN
4504 : ? VEC_UNPACK_HI_EXPR : VEC_UNPACK_LO_EXPR))
4505 42 : TREE_OPERAND (new_rhs, 2)
4506 84 : = size_binop (PLUS_EXPR, TREE_OPERAND (new_rhs, 2),
4507 : size_binop (EXACT_DIV_EXPR,
4508 : TYPE_SIZE (TREE_TYPE (lhs)),
4509 : bitsize_int (2)));
4510 84 : tree tem = make_ssa_name (TREE_TYPE (TREE_TYPE (lhs)));
4511 84 : gimple *new_stmt = gimple_build_assign (tem, new_rhs);
4512 84 : location_t loc = gimple_location (use_stmt);
4513 84 : gimple_set_location (new_stmt, loc);
4514 84 : gimple_set_vuse (new_stmt, vuse);
4515 84 : gsi_insert_before (gsi, new_stmt, GSI_SAME_STMT);
4516 : /* Perform scalar promotion. */
4517 84 : new_stmt = gimple_build_assign (gimple_assign_lhs (use_stmt),
4518 : NOP_EXPR, tem);
4519 84 : gimple_set_location (new_stmt, loc);
4520 84 : gsi_insert_before (gsi, new_stmt, GSI_SAME_STMT);
4521 : }
4522 : else
4523 : {
4524 : /* When the BIT_FIELD_REF is on the original load result
4525 : we can just wrap that. */
4526 65275 : tree new_rhs = fold_build3 (BIT_FIELD_REF, TREE_TYPE (bfr),
4527 : unshare_expr (load_rhs),
4528 : TREE_OPERAND (bfr, 1),
4529 : TREE_OPERAND (bfr, 2));
4530 65275 : gimple *new_stmt = gimple_build_assign (gimple_assign_lhs (use_stmt),
4531 : new_rhs);
4532 65275 : location_t loc = gimple_location (use_stmt);
4533 65275 : gimple_set_location (new_stmt, loc);
4534 65275 : gimple_set_vuse (new_stmt, vuse);
4535 65275 : gsi_insert_before (gsi, new_stmt, GSI_SAME_STMT);
4536 : }
4537 65359 : gimple_stmt_iterator gsi2 = gsi_for_stmt (use_stmt);
4538 65359 : unlink_stmt_vdef (use_stmt);
4539 65359 : gsi_remove (&gsi2, true);
4540 : }
4541 :
4542 : /* Finally get rid of the intermediate stmts. */
4543 18258 : gimple *use_stmt;
4544 36636 : FOR_EACH_IMM_USE_STMT (use_stmt, iter, lhs)
4545 : {
4546 120 : if (is_gimple_debug (use_stmt))
4547 : {
4548 92 : if (gimple_debug_bind_p (use_stmt))
4549 : {
4550 92 : gimple_debug_bind_reset_value (use_stmt);
4551 92 : update_stmt (use_stmt);
4552 : }
4553 92 : continue;
4554 : }
4555 28 : gimple_stmt_iterator gsi2 = gsi_for_stmt (use_stmt);
4556 28 : unlink_stmt_vdef (use_stmt);
4557 28 : release_defs (use_stmt);
4558 28 : gsi_remove (&gsi2, true);
4559 18258 : }
4560 : /* And the original load. */
4561 18258 : release_defs (stmt);
4562 18258 : gsi_remove (gsi, true);
4563 283431 : }
4564 :
4565 :
4566 : /* Primitive "lattice" function for gimple_simplify. */
4567 :
4568 : static tree
4569 1590646283 : fwprop_ssa_val (tree name)
4570 : {
4571 : /* First valueize NAME. */
4572 1590646283 : if (TREE_CODE (name) == SSA_NAME
4573 1590646283 : && SSA_NAME_VERSION (name) < lattice.length ())
4574 : {
4575 1589804864 : tree val = lattice[SSA_NAME_VERSION (name)];
4576 1589804864 : if (val)
4577 1590646283 : name = val;
4578 : }
4579 : /* We continue matching along SSA use-def edges for SSA names
4580 : that are not single-use. Currently there are no patterns
4581 : that would cause any issues with that. */
4582 1590646283 : return name;
4583 : }
4584 :
4585 : /* Search for opportunities to free half of the lanes in the following pattern:
4586 :
4587 : v_in = {e0, e1, e2, e3}
4588 : v_1 = VEC_PERM <v_in, v_in, {0, 2, 0, 2}>
4589 : // v_1 = {e0, e2, e0, e2}
4590 : v_2 = VEC_PERM <v_in, v_in, {1, 3, 1, 3}>
4591 : // v_2 = {e1, e3, e1, e3}
4592 :
4593 : v_x = v_1 + v_2
4594 : // v_x = {e0+e1, e2+e3, e0+e1, e2+e3}
4595 : v_y = v_1 - v_2
4596 : // v_y = {e0-e1, e2-e3, e0-e1, e2-e3}
4597 :
4598 : v_out = VEC_PERM <v_x, v_y, {0, 1, 6, 7}>
4599 : // v_out = {e0+e1, e2+e3, e0-e1, e2-e3}
4600 :
4601 : The last statement could be simplified to:
4602 : v_out' = VEC_PERM <v_x, v_y, {0, 1, 4, 5}>
4603 : // v_out' = {e0+e1, e2+e3, e0-e1, e2-e3}
4604 :
4605 : Characteristic properties:
4606 : - v_1 and v_2 are created from the same input vector v_in and introduce the
4607 : lane duplication (in the selection operand) that we can eliminate.
4608 : - v_x and v_y are results from lane-preserving operations that use v_1 and
4609 : v_2 as inputs.
4610 : - v_out is created by selecting from duplicated lanes. */
4611 :
4612 : static bool
4613 179602 : recognise_vec_perm_simplify_seq (gassign *stmt, vec_perm_simplify_seq *seq)
4614 : {
4615 179602 : unsigned HOST_WIDE_INT nelts;
4616 :
4617 179602 : gcc_checking_assert (stmt);
4618 179602 : gcc_checking_assert (gimple_assign_rhs_code (stmt) == VEC_PERM_EXPR);
4619 179602 : basic_block bb = gimple_bb (stmt);
4620 :
4621 : /* Decompose the final vec permute statement. */
4622 179602 : tree v_x = gimple_assign_rhs1 (stmt);
4623 179602 : tree v_y = gimple_assign_rhs2 (stmt);
4624 179602 : tree sel = gimple_assign_rhs3 (stmt);
4625 :
4626 179602 : if (TREE_CODE (sel) != VECTOR_CST
4627 176865 : || !VECTOR_CST_NELTS (sel).is_constant (&nelts)
4628 176865 : || TREE_CODE (v_x) != SSA_NAME
4629 175013 : || TREE_CODE (v_y) != SSA_NAME
4630 172470 : || !has_single_use (v_x)
4631 288189 : || !has_single_use (v_y))
4632 72721 : return false;
4633 :
4634 : /* Don't analyse sequences with many lanes. */
4635 106881 : if (nelts > 4)
4636 : return false;
4637 :
4638 : /* Lookup the definition of v_x and v_y. */
4639 104895 : gassign *v_x_stmt = dyn_cast<gassign *> (SSA_NAME_DEF_STMT (v_x));
4640 104895 : gassign *v_y_stmt = dyn_cast<gassign *> (SSA_NAME_DEF_STMT (v_y));
4641 104518 : if (!v_x_stmt || gimple_bb (v_x_stmt) != bb
4642 209413 : || !v_y_stmt || gimple_bb (v_y_stmt) != bb)
4643 : return false;
4644 :
4645 : /* Check the operations that define v_x and v_y. */
4646 104511 : if (TREE_CODE_CLASS (gimple_assign_rhs_code (v_x_stmt)) != tcc_binary
4647 106544 : || TREE_CODE_CLASS (gimple_assign_rhs_code (v_y_stmt)) != tcc_binary)
4648 : return false;
4649 :
4650 2033 : tree v_x_1 = gimple_assign_rhs1 (v_x_stmt);
4651 2033 : tree v_x_2 = gimple_assign_rhs2 (v_x_stmt);
4652 2033 : tree v_y_1 = gimple_assign_rhs1 (v_y_stmt);
4653 2033 : tree v_y_2 = gimple_assign_rhs2 (v_y_stmt);
4654 :
4655 2033 : if (v_x_stmt == v_y_stmt
4656 2033 : || TREE_CODE (v_x_1) != SSA_NAME
4657 2030 : || TREE_CODE (v_x_2) != SSA_NAME
4658 2028 : || num_imm_uses (v_x_1) != 2
4659 3903 : || num_imm_uses (v_x_2) != 2)
4660 : return false;
4661 :
4662 1834 : if (v_x_1 != v_y_1 || v_x_2 != v_y_2)
4663 : {
4664 : /* Allow operands of commutative operators to swap. */
4665 646 : if (commutative_tree_code (gimple_assign_rhs_code (v_x_stmt)))
4666 : {
4667 : /* Keep v_x_1 the first operand for non-commutative operators. */
4668 266 : std::swap (v_x_1, v_x_2);
4669 266 : if (v_x_1 != v_y_1 || v_x_2 != v_y_2)
4670 : return false;
4671 : }
4672 380 : else if (commutative_tree_code (gimple_assign_rhs_code (v_y_stmt)))
4673 : {
4674 380 : if (v_x_1 != v_y_2 || v_x_2 != v_y_1)
4675 : return false;
4676 : }
4677 : else
4678 : return false;
4679 : }
4680 1834 : gassign *v_1_stmt = dyn_cast<gassign *> (SSA_NAME_DEF_STMT (v_x_1));
4681 1834 : gassign *v_2_stmt = dyn_cast<gassign *> (SSA_NAME_DEF_STMT (v_x_2));
4682 1770 : if (!v_1_stmt || gimple_bb (v_1_stmt) != bb
4683 3604 : || !v_2_stmt || gimple_bb (v_2_stmt) != bb)
4684 : return false;
4685 :
4686 1766 : if (gimple_assign_rhs_code (v_1_stmt) != VEC_PERM_EXPR
4687 1888 : || gimple_assign_rhs_code (v_2_stmt) != VEC_PERM_EXPR)
4688 : return false;
4689 :
4690 : /* Decompose initial VEC_PERM_EXPRs. */
4691 108 : tree v_in = gimple_assign_rhs1 (v_1_stmt);
4692 108 : tree v_1_sel = gimple_assign_rhs3 (v_1_stmt);
4693 108 : tree v_2_sel = gimple_assign_rhs3 (v_2_stmt);
4694 108 : if (v_in != gimple_assign_rhs2 (v_1_stmt)
4695 103 : || v_in != gimple_assign_rhs1 (v_2_stmt)
4696 209 : || v_in != gimple_assign_rhs2 (v_2_stmt))
4697 : return false;
4698 :
4699 101 : unsigned HOST_WIDE_INT v_1_nelts, v_2_nelts;
4700 101 : if (TREE_CODE (v_1_sel) != VECTOR_CST
4701 101 : || !VECTOR_CST_NELTS (v_1_sel).is_constant (&v_1_nelts)
4702 101 : || TREE_CODE (v_2_sel) != VECTOR_CST
4703 202 : || !VECTOR_CST_NELTS (v_2_sel).is_constant (&v_2_nelts))
4704 0 : return false;
4705 :
4706 101 : if (nelts != v_1_nelts || nelts != v_2_nelts)
4707 : return false;
4708 :
4709 : /* Create the new selector. */
4710 101 : vec_perm_builder new_sel_perm (nelts, nelts, 1);
4711 101 : auto_vec<bool> lanes (nelts);
4712 101 : lanes.quick_grow_cleared (nelts);
4713 505 : for (unsigned int i = 0; i < nelts; i++)
4714 : {
4715 : /* Extract the i-th value from the selector. */
4716 404 : unsigned int sel_cst = TREE_INT_CST_LOW (VECTOR_CST_ELT (sel, i));
4717 404 : unsigned int lane = sel_cst % nelts;
4718 404 : unsigned int offs = sel_cst / nelts;
4719 :
4720 : /* Check what's in the lane. */
4721 404 : unsigned int e_1 = TREE_INT_CST_LOW (VECTOR_CST_ELT (v_1_sel, lane));
4722 404 : unsigned int e_2 = TREE_INT_CST_LOW (VECTOR_CST_ELT (v_2_sel, lane));
4723 :
4724 : /* Reuse previous lane (if any). */
4725 404 : unsigned int l = 0;
4726 687 : for (; l < lane; l++)
4727 : {
4728 481 : if ((TREE_INT_CST_LOW (VECTOR_CST_ELT (v_1_sel, l)) == e_1)
4729 481 : && (TREE_INT_CST_LOW (VECTOR_CST_ELT (v_2_sel, l)) == e_2))
4730 : break;
4731 : }
4732 :
4733 : /* Add to narrowed selector. */
4734 404 : new_sel_perm.quick_push (l + offs * nelts);
4735 :
4736 : /* Mark lane as used. */
4737 404 : lanes[l] = true;
4738 : }
4739 :
4740 : /* Count how many lanes are need. */
4741 : unsigned int cnt = 0;
4742 505 : for (unsigned int i = 0; i < nelts; i++)
4743 404 : cnt += lanes[i];
4744 :
4745 : /* If more than (nelts/2) lanes are needed, skip the sequence. */
4746 101 : if (cnt > nelts / 2)
4747 : return false;
4748 :
4749 : /* Check if the resulting permutation is cheap. */
4750 101 : vec_perm_indices new_indices (new_sel_perm, 2, nelts);
4751 101 : tree vectype = TREE_TYPE (gimple_assign_lhs (stmt));
4752 101 : machine_mode vmode = TYPE_MODE (vectype);
4753 101 : if (!can_vec_perm_const_p (vmode, vmode, new_indices, false))
4754 : return false;
4755 :
4756 101 : *seq = XNEW (struct _vec_perm_simplify_seq);
4757 101 : (*seq)->stmt = stmt;
4758 101 : (*seq)->v_1_stmt = v_1_stmt;
4759 101 : (*seq)->v_2_stmt = v_2_stmt;
4760 101 : (*seq)->v_x_stmt = v_x_stmt;
4761 101 : (*seq)->v_y_stmt = v_y_stmt;
4762 101 : (*seq)->nelts = nelts;
4763 101 : (*seq)->new_sel = vect_gen_perm_mask_checked (vectype, new_indices);
4764 :
4765 101 : if (dump_file)
4766 : {
4767 28 : fprintf (dump_file, "Found vec perm simplify sequence ending with:\n\t");
4768 28 : print_gimple_stmt (dump_file, stmt, 0);
4769 :
4770 28 : if (dump_flags & TDF_DETAILS)
4771 : {
4772 28 : fprintf (dump_file, "\tNarrowed vec_perm selector: ");
4773 28 : print_generic_expr (dump_file, (*seq)->new_sel);
4774 28 : fprintf (dump_file, "\n");
4775 : }
4776 : }
4777 :
4778 : return true;
4779 202 : }
4780 :
4781 : /* Reduce the lane consumption of a simplifiable vec perm sequence. */
4782 :
4783 : static void
4784 74 : narrow_vec_perm_simplify_seq (const vec_perm_simplify_seq &seq)
4785 : {
4786 74 : gassign *stmt = seq->stmt;
4787 74 : if (dump_file && (dump_flags & TDF_DETAILS))
4788 : {
4789 22 : fprintf (dump_file, "Updating VEC_PERM statement:\n");
4790 22 : fprintf (dump_file, "Old stmt: ");
4791 22 : print_gimple_stmt (dump_file, stmt, 0);
4792 : }
4793 :
4794 : /* Update the last VEC_PERM statement. */
4795 74 : gimple_assign_set_rhs3 (stmt, seq->new_sel);
4796 74 : update_stmt (stmt);
4797 :
4798 74 : if (dump_file && (dump_flags & TDF_DETAILS))
4799 : {
4800 22 : fprintf (dump_file, "New stmt: ");
4801 22 : print_gimple_stmt (dump_file, stmt, 0);
4802 : }
4803 74 : }
4804 :
4805 : /* Test if we can blend two simplifiable vec permute sequences.
4806 : NEED_SWAP will be set, if sequences must be swapped for blending. */
4807 :
4808 : static bool
4809 47 : can_blend_vec_perm_simplify_seqs_p (vec_perm_simplify_seq seq1,
4810 : vec_perm_simplify_seq seq2,
4811 : bool *need_swap)
4812 : {
4813 47 : unsigned int nelts = seq1->nelts;
4814 47 : basic_block bb = gimple_bb (seq1->stmt);
4815 :
4816 47 : gcc_assert (gimple_bb (seq2->stmt) == bb);
4817 :
4818 : /* BBs and number of elements must be equal. */
4819 47 : if (gimple_bb (seq2->stmt) != bb || seq2->nelts != nelts)
4820 : return false;
4821 :
4822 : /* We need vectors of the same type. */
4823 47 : if (TREE_TYPE (gimple_assign_lhs (seq1->stmt))
4824 47 : != TREE_TYPE (gimple_assign_lhs (seq2->stmt)))
4825 : return false;
4826 :
4827 : /* We require isomorphic operators. */
4828 41 : if (((gimple_assign_rhs_code (seq1->v_x_stmt)
4829 41 : != gimple_assign_rhs_code (seq2->v_x_stmt))
4830 41 : || (gimple_assign_rhs_code (seq1->v_y_stmt)
4831 41 : != gimple_assign_rhs_code (seq2->v_y_stmt))))
4832 : return false;
4833 :
4834 : /* We cannot have any dependencies between the sequences.
4835 :
4836 : For merging, we will reuse seq1->v_1_stmt and seq1->v_2_stmt.
4837 : seq1's v_in is defined before these statements, but we need
4838 : to check if seq2's v_in is defined before them as well.
4839 :
4840 : Further, we will reuse seq2->stmt. We need to ensure that
4841 : seq1->v_x_stmt and seq1->v_y_stmt are before it.
4842 :
4843 : Note, that we don't need to check the BBs here, because all
4844 : statements of both sequences have to be in the same BB. */
4845 :
4846 41 : tree seq2_v_in = gimple_assign_rhs1 (seq2->v_1_stmt);
4847 41 : if (TREE_CODE (seq2_v_in) != SSA_NAME)
4848 : return false;
4849 :
4850 41 : gassign *seq2_v_in_stmt = dyn_cast<gassign *> (SSA_NAME_DEF_STMT (seq2_v_in));
4851 41 : if (!seq2_v_in_stmt || gimple_bb (seq2_v_in_stmt) != bb
4852 41 : || (gimple_uid (seq2_v_in_stmt) > gimple_uid (seq1->v_1_stmt))
4853 37 : || (gimple_uid (seq1->v_x_stmt) > gimple_uid (seq2->stmt))
4854 37 : || (gimple_uid (seq1->v_y_stmt) > gimple_uid (seq2->stmt)))
4855 : {
4856 4 : tree seq1_v_in = gimple_assign_rhs1 (seq1->v_1_stmt);
4857 4 : if (TREE_CODE (seq1_v_in) != SSA_NAME)
4858 : return false;
4859 :
4860 4 : gassign *seq1_v_in_stmt
4861 4 : = dyn_cast<gassign *> (SSA_NAME_DEF_STMT (seq1_v_in));
4862 : /* Let's try to see if we succeed when swapping the sequences. */
4863 4 : if (!seq1_v_in_stmt || gimple_bb (seq1_v_in_stmt)
4864 0 : || (gimple_uid (seq1_v_in_stmt) > gimple_uid (seq2->v_1_stmt))
4865 0 : || (gimple_uid (seq2->v_x_stmt) > gimple_uid (seq1->stmt))
4866 0 : || (gimple_uid (seq2->v_y_stmt) > gimple_uid (seq1->stmt)))
4867 : return false;
4868 0 : *need_swap = true;
4869 : }
4870 : else
4871 37 : *need_swap = false;
4872 :
4873 37 : if (dump_file && (dump_flags & TDF_DETAILS))
4874 11 : fprintf (dump_file, "Found vec perm simplify sequence pair.\n");
4875 :
4876 : return true;
4877 : }
4878 :
4879 : /* Calculate the permutations for blending the two given vec permute
4880 : sequences. This may fail if the resulting permutation is not
4881 : supported. */
4882 :
4883 : static bool
4884 37 : calc_perm_vec_perm_simplify_seqs (vec_perm_simplify_seq seq1,
4885 : vec_perm_simplify_seq seq2,
4886 : vec_perm_indices *seq2_stmt_indices,
4887 : vec_perm_indices *seq1_v_1_stmt_indices,
4888 : vec_perm_indices *seq1_v_2_stmt_indices)
4889 : {
4890 37 : unsigned int i;
4891 37 : unsigned int nelts = seq1->nelts;
4892 37 : auto_vec<unsigned int> lane_assignment;
4893 37 : lane_assignment.create (nelts);
4894 :
4895 : /* Mark all lanes as free. */
4896 37 : lane_assignment.quick_grow_cleared (nelts);
4897 :
4898 : /* Allocate lanes for seq1. */
4899 185 : for (i = 0; i < nelts; i++)
4900 : {
4901 148 : unsigned int l = TREE_INT_CST_LOW (VECTOR_CST_ELT (seq1->new_sel, i));
4902 148 : l %= nelts;
4903 148 : lane_assignment[l] = 1;
4904 : }
4905 :
4906 : /* Allocate lanes for seq2 and calculate selector for seq2->stmt. */
4907 37 : vec_perm_builder seq2_stmt_sel_perm (nelts, nelts, 1);
4908 185 : for (i = 0; i < nelts; i++)
4909 : {
4910 148 : unsigned int sel = TREE_INT_CST_LOW (VECTOR_CST_ELT (seq2->new_sel, i));
4911 148 : unsigned int lane = sel % nelts;
4912 148 : unsigned int offs = sel / nelts;
4913 148 : unsigned int new_sel;
4914 :
4915 : /* Check if we already allocated the lane for seq2. */
4916 148 : unsigned int j = 0;
4917 263 : for (; j < i; j++)
4918 : {
4919 189 : unsigned int sel_old;
4920 189 : sel_old = TREE_INT_CST_LOW (VECTOR_CST_ELT (seq2->new_sel, j));
4921 189 : unsigned int lane_old = sel_old % nelts;
4922 189 : if (lane == lane_old)
4923 : {
4924 74 : new_sel = seq2_stmt_sel_perm[j].to_constant ();
4925 74 : new_sel = (new_sel % nelts) + offs * nelts;
4926 74 : break;
4927 : }
4928 : }
4929 :
4930 : /* If the lane is not allocated, we need to do that now. */
4931 148 : if (j == i)
4932 : {
4933 : unsigned int l_orig = lane;
4934 182 : while (lane_assignment[lane] != 0)
4935 : {
4936 108 : lane = (lane + 1) % nelts;
4937 :
4938 : /* This should not happen if both sequences utilize no more than
4939 : half of the lanes. Test anyway to guarantee termination. */
4940 108 : if (lane == l_orig)
4941 0 : return false;
4942 : }
4943 :
4944 : /* Allocate lane. */
4945 74 : lane_assignment[lane] = 2 + l_orig;
4946 74 : new_sel = lane + offs * nelts;
4947 : }
4948 :
4949 148 : seq2_stmt_sel_perm.quick_push (new_sel);
4950 : }
4951 :
4952 : /* Check if the resulting permutation is cheap. */
4953 37 : seq2_stmt_indices->new_vector (seq2_stmt_sel_perm, 2, nelts);
4954 37 : tree vectype = TREE_TYPE (gimple_assign_lhs (seq2->stmt));
4955 37 : machine_mode vmode = TYPE_MODE (vectype);
4956 37 : if (!can_vec_perm_const_p (vmode, vmode, *seq2_stmt_indices, false))
4957 : return false;
4958 :
4959 : /* Calculate selectors for seq1->v_1_stmt and seq1->v_2_stmt. */
4960 37 : vec_perm_builder seq1_v_1_stmt_sel_perm (nelts, nelts, 1);
4961 37 : vec_perm_builder seq1_v_2_stmt_sel_perm (nelts, nelts, 1);
4962 185 : for (i = 0; i < nelts; i++)
4963 : {
4964 148 : bool use_seq1 = lane_assignment[i] < 2;
4965 148 : unsigned int l1, l2;
4966 :
4967 148 : if (use_seq1)
4968 : {
4969 : /* Just reuse the selector indices. */
4970 74 : tree s1 = gimple_assign_rhs3 (seq1->v_1_stmt);
4971 74 : tree s2 = gimple_assign_rhs3 (seq1->v_2_stmt);
4972 74 : l1 = TREE_INT_CST_LOW (VECTOR_CST_ELT (s1, i));
4973 74 : l2 = TREE_INT_CST_LOW (VECTOR_CST_ELT (s2, i));
4974 : }
4975 : else
4976 : {
4977 : /* We moved the lanes for seq2, so we need to adjust for that. */
4978 74 : tree s1 = gimple_assign_rhs3 (seq2->v_1_stmt);
4979 74 : tree s2 = gimple_assign_rhs3 (seq2->v_2_stmt);
4980 74 : l1 = TREE_INT_CST_LOW (VECTOR_CST_ELT (s1, lane_assignment[i] - 2));
4981 74 : l2 = TREE_INT_CST_LOW (VECTOR_CST_ELT (s2, lane_assignment[i] - 2));
4982 : }
4983 :
4984 148 : l1 %= nelts;
4985 148 : l2 %= nelts;
4986 222 : seq1_v_1_stmt_sel_perm.quick_push (l1 + (use_seq1 ? 0 : nelts));
4987 148 : seq1_v_2_stmt_sel_perm.quick_push (l2 + (use_seq1 ? 0 : nelts));
4988 : }
4989 :
4990 37 : seq1_v_1_stmt_indices->new_vector (seq1_v_1_stmt_sel_perm, 2, nelts);
4991 37 : vectype = TREE_TYPE (gimple_assign_lhs (seq1->v_1_stmt));
4992 37 : vmode = TYPE_MODE (vectype);
4993 37 : if (!can_vec_perm_const_p (vmode, vmode, *seq1_v_1_stmt_indices, false))
4994 : return false;
4995 :
4996 37 : seq1_v_2_stmt_indices->new_vector (seq1_v_2_stmt_sel_perm, 2, nelts);
4997 37 : vectype = TREE_TYPE (gimple_assign_lhs (seq1->v_2_stmt));
4998 37 : vmode = TYPE_MODE (vectype);
4999 37 : if (!can_vec_perm_const_p (vmode, vmode, *seq1_v_2_stmt_indices, false))
5000 : return false;
5001 :
5002 : return true;
5003 74 : }
5004 :
5005 : /* Blend the two given simplifiable vec permute sequences using the
5006 : given permutations. */
5007 :
5008 : static void
5009 37 : blend_vec_perm_simplify_seqs (vec_perm_simplify_seq seq1,
5010 : vec_perm_simplify_seq seq2,
5011 : const vec_perm_indices &seq2_stmt_indices,
5012 : const vec_perm_indices &seq1_v_1_stmt_indices,
5013 : const vec_perm_indices &seq1_v_2_stmt_indices)
5014 : {
5015 : /* We don't need to adjust seq1->stmt because its lanes consumption
5016 : was already narrowed before entering this function. */
5017 :
5018 : /* Adjust seq2->stmt: copy RHS1/RHS2 from seq1->stmt and set new sel. */
5019 37 : if (dump_file && (dump_flags & TDF_DETAILS))
5020 : {
5021 11 : fprintf (dump_file, "Updating VEC_PERM statement:\n");
5022 11 : fprintf (dump_file, "Old stmt: ");
5023 11 : print_gimple_stmt (dump_file, seq2->stmt, 0);
5024 : }
5025 :
5026 37 : gimple_assign_set_rhs1 (seq2->stmt, gimple_assign_rhs1 (seq1->stmt));
5027 74 : gimple_assign_set_rhs2 (seq2->stmt, gimple_assign_rhs2 (seq1->stmt));
5028 37 : tree vectype = TREE_TYPE (gimple_assign_lhs (seq2->stmt));
5029 37 : tree sel = vect_gen_perm_mask_checked (vectype, seq2_stmt_indices);
5030 37 : gimple_assign_set_rhs3 (seq2->stmt, sel);
5031 37 : update_stmt (seq2->stmt);
5032 :
5033 37 : if (dump_file && (dump_flags & TDF_DETAILS))
5034 : {
5035 11 : fprintf (dump_file, "New stmt: ");
5036 11 : print_gimple_stmt (dump_file, seq2->stmt, 0);
5037 : }
5038 :
5039 : /* Adjust seq1->v_1_stmt: copy RHS2 from seq2->v_1_stmt and set new sel. */
5040 37 : if (dump_file && (dump_flags & TDF_DETAILS))
5041 : {
5042 11 : fprintf (dump_file, "Updating VEC_PERM statement:\n");
5043 11 : fprintf (dump_file, "Old stmt: ");
5044 11 : print_gimple_stmt (dump_file, seq1->v_1_stmt, 0);
5045 : }
5046 :
5047 37 : gimple_assign_set_rhs2 (seq1->v_1_stmt, gimple_assign_rhs1 (seq2->v_1_stmt));
5048 37 : vectype = TREE_TYPE (gimple_assign_lhs (seq1->v_1_stmt));
5049 37 : sel = vect_gen_perm_mask_checked (vectype, seq1_v_1_stmt_indices);
5050 37 : gimple_assign_set_rhs3 (seq1->v_1_stmt, sel);
5051 37 : update_stmt (seq1->v_1_stmt);
5052 :
5053 37 : if (dump_file && (dump_flags & TDF_DETAILS))
5054 : {
5055 11 : fprintf (dump_file, "New stmt: ");
5056 11 : print_gimple_stmt (dump_file, seq1->v_1_stmt, 0);
5057 : }
5058 :
5059 : /* Adjust seq1->v_2_stmt: copy RHS2 from seq2->v_2_stmt and set new sel. */
5060 37 : if (dump_file && (dump_flags & TDF_DETAILS))
5061 : {
5062 11 : fprintf (dump_file, "Updating VEC_PERM statement:\n");
5063 11 : fprintf (dump_file, "Old stmt: ");
5064 11 : print_gimple_stmt (dump_file, seq1->v_2_stmt, 0);
5065 : }
5066 :
5067 37 : gimple_assign_set_rhs2 (seq1->v_2_stmt, gimple_assign_rhs1 (seq2->v_2_stmt));
5068 37 : vectype = TREE_TYPE (gimple_assign_lhs (seq1->v_2_stmt));
5069 37 : sel = vect_gen_perm_mask_checked (vectype, seq1_v_2_stmt_indices);
5070 37 : gimple_assign_set_rhs3 (seq1->v_2_stmt, sel);
5071 37 : update_stmt (seq1->v_2_stmt);
5072 :
5073 37 : if (dump_file && (dump_flags & TDF_DETAILS))
5074 : {
5075 11 : fprintf (dump_file, "New stmt: ");
5076 11 : print_gimple_stmt (dump_file, seq1->v_2_stmt, 0);
5077 : }
5078 :
5079 : /* At this point, we have four unmodified seq2 stmts, which will be
5080 : eliminated by DCE. */
5081 :
5082 37 : if (dump_file)
5083 11 : fprintf (dump_file, "Vec perm simplify sequences have been blended.\n\n");
5084 37 : }
5085 :
5086 : /* Try to blend narrowed vec_perm_simplify_seqs pairwise.
5087 : The provided list will be empty after this call. */
5088 :
5089 : static void
5090 311349188 : process_vec_perm_simplify_seq_list (vec<vec_perm_simplify_seq> *l)
5091 : {
5092 311349188 : unsigned int i, j;
5093 311349188 : vec_perm_simplify_seq seq1, seq2;
5094 :
5095 311349188 : if (l->is_empty ())
5096 311349143 : return;
5097 :
5098 45 : if (dump_file && (dump_flags & TDF_DETAILS))
5099 13 : fprintf (dump_file, "\nProcessing %u vec perm simplify sequences.\n",
5100 : l->length ());
5101 :
5102 109 : FOR_EACH_VEC_ELT (*l, i, seq1)
5103 : {
5104 64 : if (i + 1 < l->length ())
5105 : {
5106 51 : FOR_EACH_VEC_ELT_FROM (*l, j, seq2, i + 1)
5107 : {
5108 47 : bool swap = false;
5109 47 : if (can_blend_vec_perm_simplify_seqs_p (seq1, seq2, &swap))
5110 : {
5111 37 : vec_perm_indices seq2_stmt_indices;
5112 37 : vec_perm_indices seq1_v_1_stmt_indices;
5113 37 : vec_perm_indices seq1_v_2_stmt_indices;
5114 111 : if (calc_perm_vec_perm_simplify_seqs (swap ? seq2 : seq1,
5115 : swap ? seq1 : seq2,
5116 : &seq2_stmt_indices,
5117 : &seq1_v_1_stmt_indices,
5118 : &seq1_v_2_stmt_indices))
5119 : {
5120 : /* Narrow lane usage. */
5121 37 : narrow_vec_perm_simplify_seq (seq1);
5122 37 : narrow_vec_perm_simplify_seq (seq2);
5123 :
5124 : /* Blend sequences. */
5125 37 : blend_vec_perm_simplify_seqs (swap ? seq2 : seq1,
5126 : swap ? seq1 : seq2,
5127 : seq2_stmt_indices,
5128 : seq1_v_1_stmt_indices,
5129 : seq1_v_2_stmt_indices);
5130 :
5131 : /* We can use unordered_remove as we break the loop. */
5132 37 : l->unordered_remove (j);
5133 37 : XDELETE (seq2);
5134 37 : break;
5135 : }
5136 37 : }
5137 : }
5138 : }
5139 :
5140 : /* We don't need to call l->remove for seq1. */
5141 64 : XDELETE (seq1);
5142 : }
5143 :
5144 45 : l->truncate (0);
5145 : }
5146 :
5147 : static void
5148 101 : append_vec_perm_simplify_seq_list (vec<vec_perm_simplify_seq> *l,
5149 : const vec_perm_simplify_seq &seq)
5150 : {
5151 : /* If no space on list left, then process the list. */
5152 101 : if (!l->space (1))
5153 0 : process_vec_perm_simplify_seq_list (l);
5154 :
5155 101 : l->quick_push (seq);
5156 101 : }
5157 :
5158 : /* Main entry point for the forward propagation and statement combine
5159 : optimizer. */
5160 :
5161 : namespace {
5162 :
5163 : const pass_data pass_data_forwprop =
5164 : {
5165 : GIMPLE_PASS, /* type */
5166 : "forwprop", /* name */
5167 : OPTGROUP_NONE, /* optinfo_flags */
5168 : TV_TREE_FORWPROP, /* tv_id */
5169 : ( PROP_cfg | PROP_ssa ), /* properties_required */
5170 : 0, /* properties_provided */
5171 : 0, /* properties_destroyed */
5172 : 0, /* todo_flags_start */
5173 : 0, /* todo_flags_finish */
5174 : };
5175 :
5176 : class pass_forwprop : public gimple_opt_pass
5177 : {
5178 : public:
5179 1428610 : pass_forwprop (gcc::context *ctxt)
5180 2857220 : : gimple_opt_pass (pass_data_forwprop, ctxt), last_p (false)
5181 : {}
5182 :
5183 : /* opt_pass methods: */
5184 1142888 : opt_pass * clone () final override { return new pass_forwprop (m_ctxt); }
5185 1714332 : void set_pass_param (unsigned int n, bool param) final override
5186 : {
5187 1714332 : switch (n)
5188 : {
5189 1142888 : case 0:
5190 1142888 : m_full_walk = param;
5191 1142888 : break;
5192 571444 : case 1:
5193 571444 : last_p = param;
5194 571444 : break;
5195 0 : default:
5196 0 : gcc_unreachable();
5197 : }
5198 1714332 : }
5199 5539535 : bool gate (function *) final override { return flag_tree_forwprop; }
5200 : unsigned int execute (function *) final override;
5201 :
5202 : private:
5203 : /* Determines whether the pass instance should set PROP_last_full_fold. */
5204 : bool last_p;
5205 :
5206 : /* True if the aggregate props are doing a full walk or not. */
5207 : bool m_full_walk = false;
5208 : }; // class pass_forwprop
5209 :
5210 : /* Attemp to make the BB block of __builtin_unreachable unreachable by changing
5211 : the incoming jumps. Return true if at least one jump was changed. */
5212 :
5213 : static bool
5214 3013 : optimize_unreachable (basic_block bb)
5215 : {
5216 3013 : gimple_stmt_iterator gsi;
5217 3013 : gimple *stmt;
5218 3013 : edge_iterator ei;
5219 3013 : edge e;
5220 3013 : bool ret;
5221 :
5222 3013 : ret = false;
5223 7412 : FOR_EACH_EDGE (e, ei, bb->preds)
5224 : {
5225 4399 : gsi = gsi_last_bb (e->src);
5226 4399 : if (gsi_end_p (gsi))
5227 330 : continue;
5228 :
5229 4069 : stmt = gsi_stmt (gsi);
5230 4069 : if (gcond *cond_stmt = dyn_cast <gcond *> (stmt))
5231 : {
5232 : /* If the condition is already true/false
5233 : ignore it. This can happen during copy prop of forwprop. */
5234 634 : if (gimple_cond_true_p (cond_stmt)
5235 626 : || gimple_cond_false_p (cond_stmt))
5236 8 : continue;
5237 618 : else if (e->flags & EDGE_TRUE_VALUE)
5238 537 : gimple_cond_make_false (cond_stmt);
5239 81 : else if (e->flags & EDGE_FALSE_VALUE)
5240 81 : gimple_cond_make_true (cond_stmt);
5241 : else
5242 0 : gcc_unreachable ();
5243 618 : update_stmt (cond_stmt);
5244 : }
5245 : else
5246 : {
5247 : /* Todo: handle other cases. Note that unreachable switch case
5248 : statements have already been removed. */
5249 3443 : continue;
5250 : }
5251 :
5252 618 : ret = true;
5253 : }
5254 :
5255 3013 : return ret;
5256 : }
5257 :
5258 : unsigned int
5259 5536978 : pass_forwprop::execute (function *fun)
5260 : {
5261 5536978 : unsigned int todoflags = 0;
5262 : /* Handle a full walk only when expensive optimizations are on. */
5263 5536978 : bool full_walk = m_full_walk && flag_expensive_optimizations;
5264 :
5265 5536978 : cfg_changed = false;
5266 5536978 : if (last_p)
5267 1043629 : fun->curr_properties |= PROP_last_full_fold;
5268 :
5269 5536978 : calculate_dominance_info (CDI_DOMINATORS);
5270 :
5271 : /* Combine stmts with the stmts defining their operands. Do that
5272 : in an order that guarantees visiting SSA defs before SSA uses. */
5273 11073956 : lattice.create (num_ssa_names);
5274 11073956 : lattice.quick_grow_cleared (num_ssa_names);
5275 5536978 : int *postorder = XNEWVEC (int, n_basic_blocks_for_fn (fun));
5276 5536978 : int postorder_num = pre_and_rev_post_order_compute_fn (fun, NULL,
5277 : postorder, false);
5278 5536978 : int *bb_to_rpo = XNEWVEC (int, last_basic_block_for_fn (fun));
5279 50057573 : for (int i = 0; i < postorder_num; ++i)
5280 : {
5281 44520595 : bb_to_rpo[postorder[i]] = i;
5282 44520595 : edge_iterator ei;
5283 44520595 : edge e;
5284 107212567 : FOR_EACH_EDGE (e, ei, BASIC_BLOCK_FOR_FN (fun, postorder[i])->succs)
5285 62691972 : e->flags &= ~EDGE_EXECUTABLE;
5286 : }
5287 5536978 : single_succ_edge (BASIC_BLOCK_FOR_FN (fun, ENTRY_BLOCK))->flags
5288 5536978 : |= EDGE_EXECUTABLE;
5289 5536978 : auto_vec<gimple *, 4> to_fixup;
5290 5536978 : auto_vec<gimple *, 32> to_remove;
5291 5536978 : auto_vec<unsigned, 32> to_remove_defs;
5292 5536978 : auto_vec<std::pair<int, int>, 10> edges_to_remove;
5293 5536978 : auto_bitmap simple_dce_worklist;
5294 5536978 : auto_bitmap need_ab_cleanup;
5295 5536978 : to_purge = BITMAP_ALLOC (NULL);
5296 5536978 : auto_vec<vec_perm_simplify_seq, 8> vec_perm_simplify_seq_list;
5297 50057573 : for (int i = 0; i < postorder_num; ++i)
5298 : {
5299 44520595 : gimple_stmt_iterator gsi;
5300 44520595 : basic_block bb = BASIC_BLOCK_FOR_FN (fun, postorder[i]);
5301 44520595 : edge_iterator ei;
5302 44520595 : edge e;
5303 :
5304 : /* Skip processing not executable blocks. We could improve
5305 : single_use tracking by at least unlinking uses from unreachable
5306 : blocks but since blocks with uses are not processed in a
5307 : meaningful order this is probably not worth it. */
5308 44520595 : bool any = false;
5309 45658696 : FOR_EACH_EDGE (e, ei, bb->preds)
5310 : {
5311 45644347 : if ((e->flags & EDGE_EXECUTABLE)
5312 : /* We can handle backedges in natural loops correctly but
5313 : for irreducible regions we have to take all backedges
5314 : conservatively when we did not visit the source yet. */
5315 45644347 : || (bb_to_rpo[e->src->index] > i
5316 664043 : && !dominated_by_p (CDI_DOMINATORS, e->src, e->dest)))
5317 : {
5318 : any = true;
5319 : break;
5320 : }
5321 : }
5322 44520595 : if (!any)
5323 14939 : continue;
5324 :
5325 : /* Remove conditions that go directly to unreachable when this is the last forwprop. */
5326 44506246 : if (last_p
5327 9778706 : && !(flag_sanitize & SANITIZE_UNREACHABLE))
5328 : {
5329 9773736 : gimple_stmt_iterator gsi;
5330 9773736 : gsi = gsi_start_nondebug_after_labels_bb (bb);
5331 9774326 : if (!gsi_end_p (gsi)
5332 8951056 : && gimple_call_builtin_p (*gsi, BUILT_IN_UNREACHABLE)
5333 9776749 : && optimize_unreachable (bb))
5334 : {
5335 590 : cfg_changed = true;
5336 590 : continue;
5337 : }
5338 : }
5339 :
5340 : /* Record degenerate PHIs in the lattice. */
5341 60432801 : for (gphi_iterator si = gsi_start_phis (bb); !gsi_end_p (si);
5342 15927145 : gsi_next (&si))
5343 : {
5344 15927145 : gphi *phi = si.phi ();
5345 15927145 : tree res = gimple_phi_result (phi);
5346 31854290 : if (virtual_operand_p (res))
5347 7273477 : continue;
5348 :
5349 8653668 : tree first = NULL_TREE;
5350 8653668 : bool all_same = true;
5351 8653668 : edge_iterator ei;
5352 8653668 : edge e;
5353 17797881 : FOR_EACH_EDGE (e, ei, bb->preds)
5354 : {
5355 : /* Ignore not executable forward edges. */
5356 17577009 : if (!(e->flags & EDGE_EXECUTABLE))
5357 : {
5358 4036298 : if (bb_to_rpo[e->src->index] < i)
5359 5506 : continue;
5360 : /* Avoid equivalences from backedges - while we might
5361 : be able to make irreducible regions reducible and
5362 : thus turning a back into a forward edge we do not
5363 : want to deal with the intermediate SSA issues that
5364 : exposes. */
5365 : all_same = false;
5366 : }
5367 17571503 : tree use = PHI_ARG_DEF_FROM_EDGE (phi, e);
5368 17571503 : if (use == res)
5369 : /* The PHI result can also appear on a backedge, if so
5370 : we can ignore this case for the purpose of determining
5371 : the singular value. */
5372 : ;
5373 17558886 : else if (! first)
5374 : first = use;
5375 8905218 : else if (! operand_equal_p (first, use, 0))
5376 : {
5377 : all_same = false;
5378 : break;
5379 : }
5380 : }
5381 8653668 : if (all_same)
5382 : {
5383 216148 : if (may_propagate_copy (res, first))
5384 215659 : to_remove_defs.safe_push (SSA_NAME_VERSION (res));
5385 216148 : fwprop_set_lattice_val (res, first);
5386 : }
5387 : }
5388 :
5389 : /* Apply forward propagation to all stmts in the basic-block.
5390 : Note we update GSI within the loop as necessary. */
5391 44505656 : unsigned int uid = 1;
5392 421628347 : for (gsi = gsi_start_bb (bb); !gsi_end_p (gsi); )
5393 : {
5394 332617035 : gimple *stmt = gsi_stmt (gsi);
5395 332617035 : tree lhs, rhs;
5396 332617035 : enum tree_code code;
5397 :
5398 332617035 : gimple_set_uid (stmt, uid++);
5399 :
5400 332617035 : if (!is_gimple_assign (stmt))
5401 : {
5402 229231274 : process_vec_perm_simplify_seq_list (&vec_perm_simplify_seq_list);
5403 229231274 : gsi_next (&gsi);
5404 229231274 : continue;
5405 : }
5406 :
5407 103385761 : lhs = gimple_assign_lhs (stmt);
5408 103385761 : rhs = gimple_assign_rhs1 (stmt);
5409 103385761 : code = gimple_assign_rhs_code (stmt);
5410 :
5411 140998019 : if (TREE_CODE (lhs) != SSA_NAME
5412 103385761 : || has_zero_uses (lhs))
5413 : {
5414 37612258 : process_vec_perm_simplify_seq_list (&vec_perm_simplify_seq_list);
5415 37612258 : gsi_next (&gsi);
5416 37612258 : continue;
5417 : }
5418 :
5419 : /* If this statement sets an SSA_NAME to an address,
5420 : try to propagate the address into the uses of the SSA_NAME. */
5421 65773503 : if ((code == ADDR_EXPR
5422 : /* Handle pointer conversions on invariant addresses
5423 : as well, as this is valid gimple. */
5424 63528046 : || (CONVERT_EXPR_CODE_P (code)
5425 8783294 : && TREE_CODE (rhs) == ADDR_EXPR
5426 350078 : && POINTER_TYPE_P (TREE_TYPE (lhs))))
5427 65773727 : && TREE_CODE (TREE_OPERAND (rhs, 0)) != TARGET_MEM_REF)
5428 : {
5429 2245027 : tree base = get_base_address (TREE_OPERAND (rhs, 0));
5430 2245027 : if ((!base
5431 2245027 : || !DECL_P (base)
5432 129739 : || decl_address_invariant_p (base))
5433 2245027 : && !stmt_references_abnormal_ssa_name (stmt)
5434 4490038 : && forward_propagate_addr_expr (lhs, rhs, true))
5435 : {
5436 454901 : fwprop_invalidate_lattice (gimple_get_lhs (stmt));
5437 454901 : release_defs (stmt);
5438 454901 : gsi_remove (&gsi, true);
5439 : }
5440 : else
5441 1790126 : gsi_next (&gsi);
5442 : }
5443 63528476 : else if (code == POINTER_PLUS_EXPR)
5444 : {
5445 3614823 : tree off = gimple_assign_rhs2 (stmt);
5446 3614823 : if (TREE_CODE (off) == INTEGER_CST
5447 1119708 : && can_propagate_from (stmt)
5448 1119355 : && !simple_iv_increment_p (stmt)
5449 : /* ??? Better adjust the interface to that function
5450 : instead of building new trees here. */
5451 4442143 : && forward_propagate_addr_expr
5452 2481960 : (lhs,
5453 : build1_loc (gimple_location (stmt),
5454 827320 : ADDR_EXPR, TREE_TYPE (rhs),
5455 827320 : fold_build2 (MEM_REF,
5456 : TREE_TYPE (TREE_TYPE (rhs)),
5457 : rhs,
5458 : fold_convert (ptr_type_node,
5459 : off))), true))
5460 : {
5461 305478 : fwprop_invalidate_lattice (gimple_get_lhs (stmt));
5462 305478 : release_defs (stmt);
5463 305478 : gsi_remove (&gsi, true);
5464 : }
5465 3309345 : else if (is_gimple_min_invariant (rhs))
5466 : {
5467 : /* Make sure to fold &a[0] + off_1 here. */
5468 409139 : fold_stmt_inplace (&gsi);
5469 409139 : update_stmt (stmt);
5470 409139 : if (gimple_assign_rhs_code (stmt) == POINTER_PLUS_EXPR)
5471 409121 : gsi_next (&gsi);
5472 : }
5473 : else
5474 2900206 : gsi_next (&gsi);
5475 : }
5476 59913653 : else if (TREE_CODE (TREE_TYPE (lhs)) == COMPLEX_TYPE
5477 211847 : && gimple_assign_load_p (stmt)
5478 134462 : && !gimple_has_volatile_ops (stmt)
5479 40387 : && TREE_CODE (rhs) != TARGET_MEM_REF
5480 40362 : && TREE_CODE (rhs) != BIT_FIELD_REF
5481 59954011 : && !stmt_can_throw_internal (fun, stmt))
5482 : {
5483 : /* Rewrite loads used only in real/imagpart extractions to
5484 : component-wise loads. */
5485 40233 : use_operand_p use_p;
5486 40233 : imm_use_iterator iter;
5487 40233 : tree vuse = gimple_vuse (stmt);
5488 40233 : bool rewrite = true;
5489 85223 : FOR_EACH_IMM_USE_FAST (use_p, iter, lhs)
5490 : {
5491 42905 : gimple *use_stmt = USE_STMT (use_p);
5492 42905 : if (is_gimple_debug (use_stmt))
5493 691 : continue;
5494 42214 : if (!is_gimple_assign (use_stmt)
5495 27591 : || (gimple_assign_rhs_code (use_stmt) != REALPART_EXPR
5496 25543 : && gimple_assign_rhs_code (use_stmt) != IMAGPART_EXPR)
5497 46280 : || TREE_OPERAND (gimple_assign_rhs1 (use_stmt), 0) != lhs)
5498 : {
5499 : rewrite = false;
5500 : break;
5501 : }
5502 40233 : }
5503 40233 : if (rewrite)
5504 : {
5505 2085 : gimple *use_stmt;
5506 8661 : FOR_EACH_IMM_USE_STMT (use_stmt, iter, lhs)
5507 : {
5508 4491 : if (is_gimple_debug (use_stmt))
5509 : {
5510 454 : if (gimple_debug_bind_p (use_stmt))
5511 : {
5512 454 : gimple_debug_bind_reset_value (use_stmt);
5513 454 : update_stmt (use_stmt);
5514 : }
5515 454 : continue;
5516 : }
5517 :
5518 8074 : tree new_rhs = build1 (gimple_assign_rhs_code (use_stmt),
5519 4037 : TREE_TYPE (TREE_TYPE (rhs)),
5520 : unshare_expr (rhs));
5521 4037 : gimple *new_stmt
5522 4037 : = gimple_build_assign (gimple_assign_lhs (use_stmt),
5523 : new_rhs);
5524 :
5525 4037 : location_t loc = gimple_location (use_stmt);
5526 4037 : gimple_set_location (new_stmt, loc);
5527 4037 : gimple_set_vuse (new_stmt, vuse);
5528 4037 : gimple_stmt_iterator gsi2 = gsi_for_stmt (use_stmt);
5529 4037 : unlink_stmt_vdef (use_stmt);
5530 4037 : gsi_remove (&gsi2, true);
5531 :
5532 4037 : gsi_insert_before (&gsi, new_stmt, GSI_SAME_STMT);
5533 2085 : }
5534 :
5535 2085 : release_defs (stmt);
5536 2085 : gsi_remove (&gsi, true);
5537 : }
5538 : else
5539 38148 : gsi_next (&gsi);
5540 : }
5541 59873420 : else if (TREE_CODE (TREE_TYPE (lhs)) == VECTOR_TYPE
5542 1615007 : && (TYPE_MODE (TREE_TYPE (lhs)) == BLKmode
5543 : /* After vector lowering rewrite all loads, but
5544 : initially do not since this conflicts with
5545 : vector CONSTRUCTOR to shuffle optimization. */
5546 1594025 : || (fun->curr_properties & PROP_gimple_lvec))
5547 889665 : && gimple_assign_load_p (stmt)
5548 297802 : && !gimple_has_volatile_ops (stmt)
5549 283933 : && !stmt_can_throw_internal (fun, stmt)
5550 60157353 : && (!VAR_P (rhs) || !DECL_HARD_REGISTER (rhs)))
5551 283431 : optimize_vector_load (&gsi);
5552 :
5553 59589989 : else if (code == COMPLEX_EXPR)
5554 : {
5555 : /* Rewrite stores of a single-use complex build expression
5556 : to component-wise stores. */
5557 36579 : use_operand_p use_p;
5558 36579 : gimple *use_stmt, *def1, *def2;
5559 36579 : tree rhs2;
5560 36579 : if (single_imm_use (lhs, &use_p, &use_stmt)
5561 34419 : && gimple_store_p (use_stmt)
5562 41038 : && !gimple_has_volatile_ops (use_stmt)
5563 2603 : && is_gimple_assign (use_stmt)
5564 2599 : && (TREE_CODE (TREE_TYPE (gimple_assign_lhs (use_stmt)))
5565 : == COMPLEX_TYPE)
5566 39173 : && (TREE_CODE (gimple_assign_lhs (use_stmt))
5567 : != TARGET_MEM_REF))
5568 : {
5569 2590 : tree use_lhs = gimple_assign_lhs (use_stmt);
5570 2590 : if (auto_var_p (use_lhs))
5571 601 : DECL_NOT_GIMPLE_REG_P (use_lhs) = 1;
5572 5180 : tree new_lhs = build1 (REALPART_EXPR,
5573 2590 : TREE_TYPE (TREE_TYPE (use_lhs)),
5574 : unshare_expr (use_lhs));
5575 2590 : gimple *new_stmt = gimple_build_assign (new_lhs, rhs);
5576 2590 : location_t loc = gimple_location (use_stmt);
5577 2590 : gimple_set_location (new_stmt, loc);
5578 5180 : gimple_set_vuse (new_stmt, gimple_vuse (use_stmt));
5579 2590 : gimple_set_vdef (new_stmt, make_ssa_name (gimple_vop (fun)));
5580 5180 : SSA_NAME_DEF_STMT (gimple_vdef (new_stmt)) = new_stmt;
5581 5180 : gimple_set_vuse (use_stmt, gimple_vdef (new_stmt));
5582 2590 : gimple_stmt_iterator gsi2 = gsi_for_stmt (use_stmt);
5583 2590 : gsi_insert_before (&gsi2, new_stmt, GSI_SAME_STMT);
5584 :
5585 5180 : new_lhs = build1 (IMAGPART_EXPR,
5586 2590 : TREE_TYPE (TREE_TYPE (use_lhs)),
5587 : unshare_expr (use_lhs));
5588 2590 : gimple_assign_set_lhs (use_stmt, new_lhs);
5589 2590 : gimple_assign_set_rhs1 (use_stmt, gimple_assign_rhs2 (stmt));
5590 2590 : update_stmt (use_stmt);
5591 :
5592 2590 : release_defs (stmt);
5593 2590 : gsi_remove (&gsi, true);
5594 : }
5595 : /* Rewrite a component-wise load of a complex to a complex
5596 : load if the components are not used separately. */
5597 33989 : else if (TREE_CODE (rhs) == SSA_NAME
5598 33548 : && has_single_use (rhs)
5599 30060 : && ((rhs2 = gimple_assign_rhs2 (stmt)), true)
5600 30060 : && TREE_CODE (rhs2) == SSA_NAME
5601 28316 : && has_single_use (rhs2)
5602 27895 : && (def1 = SSA_NAME_DEF_STMT (rhs),
5603 27895 : gimple_assign_load_p (def1))
5604 1097 : && (def2 = SSA_NAME_DEF_STMT (rhs2),
5605 1097 : gimple_assign_load_p (def2))
5606 1606 : && (gimple_vuse (def1) == gimple_vuse (def2))
5607 800 : && !gimple_has_volatile_ops (def1)
5608 800 : && !gimple_has_volatile_ops (def2)
5609 800 : && !stmt_can_throw_internal (fun, def1)
5610 800 : && !stmt_can_throw_internal (fun, def2)
5611 800 : && gimple_assign_rhs_code (def1) == REALPART_EXPR
5612 542 : && gimple_assign_rhs_code (def2) == IMAGPART_EXPR
5613 34531 : && operand_equal_p (TREE_OPERAND (gimple_assign_rhs1
5614 : (def1), 0),
5615 542 : TREE_OPERAND (gimple_assign_rhs1
5616 : (def2), 0)))
5617 : {
5618 542 : tree cl = TREE_OPERAND (gimple_assign_rhs1 (def1), 0);
5619 542 : gimple_assign_set_rhs_from_tree (&gsi, unshare_expr (cl));
5620 542 : gcc_assert (gsi_stmt (gsi) == stmt);
5621 1084 : gimple_set_vuse (stmt, gimple_vuse (def1));
5622 542 : gimple_set_modified (stmt, true);
5623 542 : gimple_stmt_iterator gsi2 = gsi_for_stmt (def1);
5624 542 : gsi_remove (&gsi, false);
5625 542 : gsi_insert_after (&gsi2, stmt, GSI_SAME_STMT);
5626 : }
5627 : else
5628 33447 : gsi_next (&gsi);
5629 : }
5630 59553410 : else if (code == CONSTRUCTOR
5631 149639 : && VECTOR_TYPE_P (TREE_TYPE (rhs))
5632 149639 : && TYPE_MODE (TREE_TYPE (rhs)) == BLKmode
5633 2906 : && CONSTRUCTOR_NELTS (rhs) > 0
5634 59556316 : && (!VECTOR_TYPE_P (TREE_TYPE (CONSTRUCTOR_ELT (rhs, 0)->value))
5635 545 : || (TYPE_MODE (TREE_TYPE (CONSTRUCTOR_ELT (rhs, 0)->value))
5636 : != BLKmode)))
5637 : {
5638 : /* Rewrite stores of a single-use vector constructors
5639 : to component-wise stores if the mode isn't supported. */
5640 2905 : use_operand_p use_p;
5641 2905 : gimple *use_stmt;
5642 2905 : if (single_imm_use (lhs, &use_p, &use_stmt)
5643 2474 : && gimple_store_p (use_stmt)
5644 2918 : && !gimple_has_volatile_ops (use_stmt)
5645 1453 : && !stmt_can_throw_internal (fun, use_stmt)
5646 4348 : && is_gimple_assign (use_stmt))
5647 : {
5648 1443 : tree elt_t = TREE_TYPE (CONSTRUCTOR_ELT (rhs, 0)->value);
5649 1443 : unsigned HOST_WIDE_INT elt_w
5650 1443 : = tree_to_uhwi (TYPE_SIZE (elt_t));
5651 1443 : unsigned HOST_WIDE_INT n
5652 1443 : = tree_to_uhwi (TYPE_SIZE (TREE_TYPE (rhs)));
5653 1443 : tree use_lhs = gimple_assign_lhs (use_stmt);
5654 1443 : if (auto_var_p (use_lhs))
5655 540 : DECL_NOT_GIMPLE_REG_P (use_lhs) = 1;
5656 903 : else if (TREE_CODE (use_lhs) == TARGET_MEM_REF)
5657 : {
5658 1 : gimple_stmt_iterator gsi2 = gsi_for_stmt (use_stmt);
5659 1 : use_lhs = prepare_target_mem_ref_lvalue (use_lhs, &gsi2);
5660 : }
5661 32696 : for (unsigned HOST_WIDE_INT bi = 0; bi < n; bi += elt_w)
5662 : {
5663 31253 : unsigned HOST_WIDE_INT ci = bi / elt_w;
5664 31253 : tree new_rhs;
5665 31253 : if (ci < CONSTRUCTOR_NELTS (rhs))
5666 30635 : new_rhs = CONSTRUCTOR_ELT (rhs, ci)->value;
5667 : else
5668 618 : new_rhs = build_zero_cst (elt_t);
5669 31253 : tree new_lhs = build3 (BIT_FIELD_REF,
5670 : elt_t,
5671 : unshare_expr (use_lhs),
5672 31253 : bitsize_int (elt_w),
5673 31253 : bitsize_int (bi));
5674 31253 : gimple *new_stmt = gimple_build_assign (new_lhs, new_rhs);
5675 31253 : location_t loc = gimple_location (use_stmt);
5676 31253 : gimple_set_location (new_stmt, loc);
5677 62506 : gimple_set_vuse (new_stmt, gimple_vuse (use_stmt));
5678 31253 : gimple_set_vdef (new_stmt,
5679 : make_ssa_name (gimple_vop (fun)));
5680 62506 : SSA_NAME_DEF_STMT (gimple_vdef (new_stmt)) = new_stmt;
5681 62506 : gimple_set_vuse (use_stmt, gimple_vdef (new_stmt));
5682 31253 : gimple_stmt_iterator gsi2 = gsi_for_stmt (use_stmt);
5683 31253 : gsi_insert_before (&gsi2, new_stmt, GSI_SAME_STMT);
5684 : }
5685 1443 : gimple_stmt_iterator gsi2 = gsi_for_stmt (use_stmt);
5686 1443 : unlink_stmt_vdef (use_stmt);
5687 1443 : release_defs (use_stmt);
5688 1443 : gsi_remove (&gsi2, true);
5689 1443 : release_defs (stmt);
5690 1443 : gsi_remove (&gsi, true);
5691 : }
5692 : else
5693 1462 : gsi_next (&gsi);
5694 : }
5695 59550505 : else if (code == VEC_PERM_EXPR)
5696 : {
5697 : /* Find vectorized sequences, where we can reduce the lane
5698 : utilization. The narrowing will be donw later and only
5699 : if we find a pair of sequences that can be blended. */
5700 179602 : gassign *assign = dyn_cast <gassign *> (stmt);
5701 179602 : vec_perm_simplify_seq seq;
5702 179602 : if (recognise_vec_perm_simplify_seq (assign, &seq))
5703 101 : append_vec_perm_simplify_seq_list (&vec_perm_simplify_seq_list,
5704 : seq);
5705 :
5706 179602 : gsi_next (&gsi);
5707 : }
5708 : else
5709 59370903 : gsi_next (&gsi);
5710 : }
5711 :
5712 44505656 : process_vec_perm_simplify_seq_list (&vec_perm_simplify_seq_list);
5713 :
5714 : /* Combine stmts with the stmts defining their operands.
5715 : Note we update GSI within the loop as necessary. */
5716 421287733 : for (gsi = gsi_start_bb (bb); !gsi_end_p (gsi); gsi_next (&gsi))
5717 : {
5718 332276421 : gimple *stmt = gsi_stmt (gsi);
5719 :
5720 : /* Mark stmt as potentially needing revisiting. */
5721 332276421 : gimple_set_plf (stmt, GF_PLF_1, false);
5722 :
5723 332276421 : bool can_make_abnormal_goto = (is_gimple_call (stmt)
5724 332276421 : && stmt_can_make_abnormal_goto (stmt));
5725 :
5726 : /* Substitute from our lattice. We need to do so only once. */
5727 332276421 : bool substituted_p = false;
5728 332276421 : use_operand_p usep;
5729 332276421 : ssa_op_iter iter;
5730 492249980 : FOR_EACH_SSA_USE_OPERAND (usep, stmt, iter, SSA_OP_USE)
5731 : {
5732 159973559 : tree use = USE_FROM_PTR (usep);
5733 159973559 : tree val = fwprop_ssa_val (use);
5734 159973559 : if (val && val != use)
5735 : {
5736 1853289 : if (!is_gimple_debug (stmt))
5737 1540427 : bitmap_set_bit (simple_dce_worklist, SSA_NAME_VERSION (use));
5738 1853289 : if (may_propagate_copy (use, val))
5739 : {
5740 1850106 : propagate_value (usep, val);
5741 1850106 : substituted_p = true;
5742 : }
5743 : }
5744 : }
5745 332276421 : if (substituted_p)
5746 1797045 : update_stmt (stmt);
5747 1797045 : if (substituted_p
5748 1797045 : && is_gimple_assign (stmt)
5749 1083013 : && gimple_assign_rhs_code (stmt) == ADDR_EXPR)
5750 19858 : recompute_tree_invariant_for_addr_expr (gimple_assign_rhs1 (stmt));
5751 332276421 : if (substituted_p
5752 332276421 : && can_make_abnormal_goto
5753 332276421 : && !stmt_can_make_abnormal_goto (stmt))
5754 3 : bitmap_set_bit (need_ab_cleanup, bb->index);
5755 :
5756 335071446 : bool changed;
5757 670142892 : do
5758 : {
5759 335071446 : gimple *orig_stmt = stmt = gsi_stmt (gsi);
5760 335071446 : bool was_call = is_gimple_call (stmt);
5761 335071446 : bool was_noreturn = (was_call
5762 335071446 : && gimple_call_noreturn_p (stmt));
5763 335071446 : changed = false;
5764 :
5765 335071446 : auto_vec<tree, 8> uses;
5766 498038286 : FOR_EACH_SSA_USE_OPERAND (usep, stmt, iter, SSA_OP_USE)
5767 162966840 : if (uses.space (1))
5768 162581482 : uses.quick_push (USE_FROM_PTR (usep));
5769 :
5770 335071446 : if (fold_stmt (&gsi, fwprop_ssa_val, simple_dce_worklist))
5771 : {
5772 2480771 : changed = true;
5773 : /* There is no updating of the address
5774 : taken after the last forwprop so update
5775 : the addresses when a folding happened to a call.
5776 : The va_* builtins can remove taking of the address so
5777 : can the sincos->cexpi transformation. See PR 39643 and PR 20983. */
5778 2480771 : if (was_call && last_p)
5779 2480771 : todoflags |= TODO_update_address_taken;
5780 2480771 : stmt = gsi_stmt (gsi);
5781 : /* Cleanup the CFG if we simplified a condition to
5782 : true or false. */
5783 2480771 : if (gcond *cond = dyn_cast <gcond *> (stmt))
5784 976952 : if (gimple_cond_true_p (cond)
5785 976952 : || gimple_cond_false_p (cond))
5786 15436 : cfg_changed = true;
5787 : /* Queue old uses for simple DCE if not debug statement. */
5788 2480771 : if (!is_gimple_debug (stmt))
5789 10504199 : for (tree use : uses)
5790 3081800 : if (TREE_CODE (use) == SSA_NAME
5791 3081800 : && !SSA_NAME_IS_DEFAULT_DEF (use))
5792 2885582 : bitmap_set_bit (simple_dce_worklist,
5793 2885582 : SSA_NAME_VERSION (use));
5794 2480771 : update_stmt (stmt);
5795 : }
5796 :
5797 335071446 : switch (gimple_code (stmt))
5798 : {
5799 104395493 : case GIMPLE_ASSIGN:
5800 104395493 : {
5801 104395493 : tree rhs1 = gimple_assign_rhs1 (stmt);
5802 104395493 : enum tree_code code = gimple_assign_rhs_code (stmt);
5803 104395493 : if (gimple_clobber_p (stmt))
5804 6767334 : do_simple_agr_dse (as_a<gassign*>(stmt), full_walk);
5805 97628159 : else if (gimple_store_p (stmt))
5806 : {
5807 30227122 : optimize_aggr_zeroprop (stmt, full_walk);
5808 30227122 : if (gimple_assign_load_p (stmt))
5809 3711510 : optimize_agr_copyprop (stmt);
5810 : }
5811 67401037 : else if (TREE_CODE_CLASS (code) == tcc_comparison)
5812 2515964 : changed |= forward_propagate_into_comparison (&gsi);
5813 64885073 : else if ((code == PLUS_EXPR
5814 64885073 : || code == BIT_IOR_EXPR
5815 54891757 : || code == BIT_XOR_EXPR)
5816 65013666 : && simplify_rotate (&gsi))
5817 : changed = true;
5818 64882399 : else if (code == VEC_PERM_EXPR)
5819 181766 : changed |= simplify_permutation (&gsi);
5820 64700633 : else if (code == CONSTRUCTOR
5821 64700633 : && TREE_CODE (TREE_TYPE (rhs1)) == VECTOR_TYPE)
5822 147562 : changed |= simplify_vector_constructor (&gsi);
5823 64553071 : else if (code == ARRAY_REF)
5824 1962912 : changed |= simplify_count_zeroes (&gsi);
5825 : break;
5826 : }
5827 :
5828 110086 : case GIMPLE_SWITCH:
5829 110086 : changed |= simplify_gimple_switch (as_a <gswitch *> (stmt),
5830 : edges_to_remove,
5831 : simple_dce_worklist);
5832 110086 : break;
5833 :
5834 19257691 : case GIMPLE_COND:
5835 19257691 : {
5836 19257691 : int did_something = forward_propagate_into_gimple_cond
5837 19257691 : (as_a <gcond *> (stmt));
5838 19257691 : if (did_something == 2)
5839 1700 : cfg_changed = true;
5840 19257691 : changed |= did_something != 0;
5841 19257691 : break;
5842 : }
5843 :
5844 23060349 : case GIMPLE_CALL:
5845 23060349 : {
5846 23060349 : tree callee = gimple_call_fndecl (stmt);
5847 23060349 : if (callee != NULL_TREE
5848 23060349 : && fndecl_built_in_p (callee, BUILT_IN_NORMAL))
5849 6128230 : changed |= simplify_builtin_call (&gsi, callee, full_walk);
5850 : break;
5851 : }
5852 :
5853 335068772 : default:;
5854 : }
5855 :
5856 335068772 : if (changed || substituted_p)
5857 : {
5858 4065547 : substituted_p = false;
5859 4065547 : stmt = gsi_stmt (gsi);
5860 4065547 : if (maybe_clean_or_replace_eh_stmt (orig_stmt, stmt))
5861 70 : bitmap_set_bit (to_purge, bb->index);
5862 4065547 : if (!was_noreturn
5863 4065547 : && is_gimple_call (stmt) && gimple_call_noreturn_p (stmt))
5864 12 : to_fixup.safe_push (stmt);
5865 : }
5866 4065547 : if (changed)
5867 : {
5868 : /* If the stmt changed then re-visit it and the statements
5869 : inserted before it. */
5870 8748481 : for (; !gsi_end_p (gsi); gsi_prev (&gsi))
5871 5547820 : if (gimple_plf (gsi_stmt (gsi), GF_PLF_1))
5872 : break;
5873 2795025 : if (gsi_end_p (gsi))
5874 447866 : gsi = gsi_start_bb (bb);
5875 : else
5876 2571092 : gsi_next (&gsi);
5877 : }
5878 335071446 : }
5879 : while (changed);
5880 :
5881 : /* Stmt no longer needs to be revisited. */
5882 332276421 : stmt = gsi_stmt (gsi);
5883 332276421 : gcc_checking_assert (!gimple_plf (stmt, GF_PLF_1));
5884 332276421 : gimple_set_plf (stmt, GF_PLF_1, true);
5885 :
5886 : /* Fill up the lattice. */
5887 332276421 : if (gimple_assign_single_p (stmt))
5888 : {
5889 68853156 : tree lhs = gimple_assign_lhs (stmt);
5890 68853156 : tree rhs = gimple_assign_rhs1 (stmt);
5891 68853156 : if (TREE_CODE (lhs) == SSA_NAME)
5892 : {
5893 31873475 : tree val = lhs;
5894 31873475 : if (TREE_CODE (rhs) == SSA_NAME)
5895 783917 : val = fwprop_ssa_val (rhs);
5896 31089558 : else if (is_gimple_min_invariant (rhs))
5897 419705 : val = rhs;
5898 : /* If we can propagate the lattice-value mark the
5899 : stmt for removal. */
5900 31873475 : if (val != lhs
5901 31873475 : && may_propagate_copy (lhs, val))
5902 1200254 : to_remove_defs.safe_push (SSA_NAME_VERSION (lhs));
5903 31873475 : fwprop_set_lattice_val (lhs, val);
5904 : }
5905 : }
5906 263423265 : else if (gimple_nop_p (stmt))
5907 88049 : to_remove.safe_push (stmt);
5908 : }
5909 :
5910 : /* Substitute in destination PHI arguments. */
5911 107186852 : FOR_EACH_EDGE (e, ei, bb->succs)
5912 62681196 : for (gphi_iterator gsi = gsi_start_phis (e->dest);
5913 104709957 : !gsi_end_p (gsi); gsi_next (&gsi))
5914 : {
5915 42028761 : gphi *phi = gsi.phi ();
5916 42028761 : use_operand_p use_p = PHI_ARG_DEF_PTR_FROM_EDGE (phi, e);
5917 42028761 : tree arg = USE_FROM_PTR (use_p);
5918 69248367 : if (TREE_CODE (arg) != SSA_NAME
5919 42028761 : || virtual_operand_p (arg))
5920 27219606 : continue;
5921 14809155 : tree val = fwprop_ssa_val (arg);
5922 14809155 : if (val != arg
5923 14809155 : && may_propagate_copy (arg, val, !(e->flags & EDGE_ABNORMAL)))
5924 233591 : propagate_value (use_p, val);
5925 : }
5926 :
5927 : /* Mark outgoing exectuable edges. */
5928 44505656 : if (edge e = find_taken_edge (bb, NULL))
5929 : {
5930 18903290 : e->flags |= EDGE_EXECUTABLE;
5931 44527003 : if (EDGE_COUNT (bb->succs) > 1)
5932 21347 : cfg_changed = true;
5933 : }
5934 : else
5935 : {
5936 69358924 : FOR_EACH_EDGE (e, ei, bb->succs)
5937 43756558 : e->flags |= EDGE_EXECUTABLE;
5938 : }
5939 : }
5940 5536978 : free (postorder);
5941 5536978 : free (bb_to_rpo);
5942 5536978 : lattice.release ();
5943 :
5944 : /* First remove chains of stmts where we check no uses remain. */
5945 5536978 : simple_dce_from_worklist (simple_dce_worklist, to_purge);
5946 :
5947 5872715 : auto remove = [](gimple *stmt)
5948 : {
5949 335737 : if (dump_file && (dump_flags & TDF_DETAILS))
5950 : {
5951 1 : fprintf (dump_file, "Removing dead stmt ");
5952 1 : print_gimple_stmt (dump_file, stmt, 0);
5953 1 : fprintf (dump_file, "\n");
5954 : }
5955 335737 : gimple_stmt_iterator gsi = gsi_for_stmt (stmt);
5956 335737 : if (gimple_code (stmt) == GIMPLE_PHI)
5957 83200 : remove_phi_node (&gsi, true);
5958 : else
5959 : {
5960 252537 : unlink_stmt_vdef (stmt);
5961 252537 : gsi_remove (&gsi, true);
5962 252537 : release_defs (stmt);
5963 : }
5964 335737 : };
5965 :
5966 : /* Then remove stmts we know we can remove even though we did not
5967 : substitute in dead code regions, so uses can remain. Do so in reverse
5968 : order to make debug stmt creation possible. */
5969 12489869 : while (!to_remove_defs.is_empty())
5970 : {
5971 1415913 : tree def = ssa_name (to_remove_defs.pop ());
5972 : /* For example remove_prop_source_from_use can remove stmts queued
5973 : for removal. Deal with this gracefully. */
5974 1415913 : if (!def)
5975 1168225 : continue;
5976 247688 : gimple *stmt = SSA_NAME_DEF_STMT (def);
5977 247688 : remove (stmt);
5978 : }
5979 :
5980 : /* Wipe other queued stmts that do not have SSA defs. */
5981 5625027 : while (!to_remove.is_empty())
5982 : {
5983 88049 : gimple *stmt = to_remove.pop ();
5984 88049 : remove (stmt);
5985 : }
5986 :
5987 : /* Fixup stmts that became noreturn calls. This may require splitting
5988 : blocks and thus isn't possible during the walk. Do this
5989 : in reverse order so we don't inadvertedly remove a stmt we want to
5990 : fixup by visiting a dominating now noreturn call first. */
5991 5536990 : while (!to_fixup.is_empty ())
5992 : {
5993 12 : gimple *stmt = to_fixup.pop ();
5994 12 : if (dump_file && dump_flags & TDF_DETAILS)
5995 : {
5996 0 : fprintf (dump_file, "Fixing up noreturn call ");
5997 0 : print_gimple_stmt (dump_file, stmt, 0);
5998 0 : fprintf (dump_file, "\n");
5999 : }
6000 12 : cfg_changed |= fixup_noreturn_call (stmt);
6001 : }
6002 :
6003 5536978 : cfg_changed |= gimple_purge_all_dead_eh_edges (to_purge);
6004 5536978 : cfg_changed |= gimple_purge_all_dead_abnormal_call_edges (need_ab_cleanup);
6005 5536978 : BITMAP_FREE (to_purge);
6006 :
6007 : /* Remove edges queued from switch stmt simplification. */
6008 16610934 : for (auto ep : edges_to_remove)
6009 : {
6010 0 : basic_block src = BASIC_BLOCK_FOR_FN (fun, ep.first);
6011 0 : basic_block dest = BASIC_BLOCK_FOR_FN (fun, ep.second);
6012 0 : edge e;
6013 0 : if (src && dest && (e = find_edge (src, dest)))
6014 : {
6015 0 : free_dominance_info (CDI_DOMINATORS);
6016 0 : remove_edge (e);
6017 0 : cfg_changed = true;
6018 : }
6019 : }
6020 :
6021 11072416 : if (get_range_query (fun) != get_global_range_query ())
6022 1540 : disable_ranger (fun);
6023 :
6024 5536978 : if (cfg_changed)
6025 9897 : todoflags |= TODO_cleanup_cfg;
6026 :
6027 5536978 : return todoflags;
6028 5536978 : }
6029 :
6030 : } // anon namespace
6031 :
6032 : gimple_opt_pass *
6033 285722 : make_pass_forwprop (gcc::context *ctxt)
6034 : {
6035 285722 : return new pass_forwprop (ctxt);
6036 : }
|
