1 | /* Dead code elimination pass for the GNU compiler. |
2 | Copyright (C) 2002-2023 Free Software Foundation, Inc. |
3 | Contributed by Ben Elliston <bje@redhat.com> |
4 | and Andrew MacLeod <amacleod@redhat.com> |
5 | Adapted to use control dependence by Steven Bosscher, SUSE Labs. |
6 | |
7 | This file is part of GCC. |
8 | |
9 | GCC is free software; you can redistribute it and/or modify it |
10 | under the terms of the GNU General Public License as published by the |
11 | Free Software Foundation; either version 3, or (at your option) any |
12 | later version. |
13 | |
14 | GCC is distributed in the hope that it will be useful, but WITHOUT |
15 | ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or |
16 | FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License |
17 | for more details. |
18 | |
19 | You should have received a copy of the GNU General Public License |
20 | along with GCC; see the file COPYING3. If not see |
21 | <http://www.gnu.org/licenses/>. */ |
22 | |
23 | /* Dead code elimination. |
24 | |
25 | References: |
26 | |
27 | Building an Optimizing Compiler, |
28 | Robert Morgan, Butterworth-Heinemann, 1998, Section 8.9. |
29 | |
30 | Advanced Compiler Design and Implementation, |
31 | Steven Muchnick, Morgan Kaufmann, 1997, Section 18.10. |
32 | |
33 | Dead-code elimination is the removal of statements which have no |
34 | impact on the program's output. "Dead statements" have no impact |
35 | on the program's output, while "necessary statements" may have |
36 | impact on the output. |
37 | |
38 | The algorithm consists of three phases: |
39 | 1. Marking as necessary all statements known to be necessary, |
40 | e.g. most function calls, writing a value to memory, etc; |
41 | 2. Propagating necessary statements, e.g., the statements |
42 | giving values to operands in necessary statements; and |
43 | 3. Removing dead statements. */ |
44 | |
45 | #include "config.h" |
46 | #include "system.h" |
47 | #include "coretypes.h" |
48 | #include "backend.h" |
49 | #include "rtl.h" |
50 | #include "tree.h" |
51 | #include "gimple.h" |
52 | #include "cfghooks.h" |
53 | #include "tree-pass.h" |
54 | #include "ssa.h" |
55 | #include "gimple-pretty-print.h" |
56 | #include "fold-const.h" |
57 | #include "calls.h" |
58 | #include "cfganal.h" |
59 | #include "tree-eh.h" |
60 | #include "gimplify.h" |
61 | #include "gimple-iterator.h" |
62 | #include "tree-cfg.h" |
63 | #include "tree-ssa-loop-niter.h" |
64 | #include "tree-into-ssa.h" |
65 | #include "tree-dfa.h" |
66 | #include "cfgloop.h" |
67 | #include "tree-scalar-evolution.h" |
68 | #include "tree-ssa-propagate.h" |
69 | #include "gimple-fold.h" |
70 | #include "tree-ssa.h" |
71 | |
72 | static struct stmt_stats |
73 | { |
74 | int total; |
75 | int total_phis; |
76 | int removed; |
77 | int removed_phis; |
78 | } stats; |
79 | |
80 | #define STMT_NECESSARY GF_PLF_1 |
81 | |
82 | static vec<gimple *> worklist; |
83 | |
84 | /* Vector indicating an SSA name has already been processed and marked |
85 | as necessary. */ |
86 | static sbitmap processed; |
87 | |
88 | /* Vector indicating that the last statement of a basic block has already |
89 | been marked as necessary. */ |
90 | static sbitmap last_stmt_necessary; |
91 | |
92 | /* Vector indicating that BB contains statements that are live. */ |
93 | static sbitmap bb_contains_live_stmts; |
94 | |
95 | /* Before we can determine whether a control branch is dead, we need to |
96 | compute which blocks are control dependent on which edges. |
97 | |
98 | We expect each block to be control dependent on very few edges so we |
99 | use a bitmap for each block recording its edges. An array holds the |
100 | bitmap. The Ith bit in the bitmap is set if that block is dependent |
101 | on the Ith edge. */ |
102 | static control_dependences *cd; |
103 | |
104 | /* Vector indicating that a basic block has already had all the edges |
105 | processed that it is control dependent on. */ |
106 | static sbitmap visited_control_parents; |
107 | |
108 | /* TRUE if this pass alters the CFG (by removing control statements). |
109 | FALSE otherwise. |
110 | |
111 | If this pass alters the CFG, then it will arrange for the dominators |
112 | to be recomputed. */ |
113 | static bool cfg_altered; |
114 | |
115 | /* When non-NULL holds map from basic block index into the postorder. */ |
116 | static int *bb_postorder; |
117 | |
118 | |
119 | /* True if we should treat any stmt with a vdef as necessary. */ |
120 | |
121 | static inline bool |
122 | keep_all_vdefs_p () |
123 | { |
124 | return optimize_debug; |
125 | } |
126 | |
127 | /* If STMT is not already marked necessary, mark it, and add it to the |
128 | worklist if ADD_TO_WORKLIST is true. */ |
129 | |
130 | static inline void |
131 | mark_stmt_necessary (gimple *stmt, bool add_to_worklist) |
132 | { |
133 | gcc_assert (stmt); |
134 | |
135 | if (gimple_plf (stmt, STMT_NECESSARY)) |
136 | return; |
137 | |
138 | if (dump_file && (dump_flags & TDF_DETAILS)) |
139 | { |
140 | fprintf (stream: dump_file, format: "Marking useful stmt: " ); |
141 | print_gimple_stmt (dump_file, stmt, 0, TDF_SLIM); |
142 | fprintf (stream: dump_file, format: "\n" ); |
143 | } |
144 | |
145 | gimple_set_plf (stmt, STMT_NECESSARY, val_p: true); |
146 | if (add_to_worklist) |
147 | worklist.safe_push (obj: stmt); |
148 | if (add_to_worklist && bb_contains_live_stmts && !is_gimple_debug (gs: stmt)) |
149 | bitmap_set_bit (map: bb_contains_live_stmts, bitno: gimple_bb (g: stmt)->index); |
150 | } |
151 | |
152 | |
153 | /* Mark the statement defining operand OP as necessary. */ |
154 | |
155 | static inline void |
156 | mark_operand_necessary (tree op) |
157 | { |
158 | gimple *stmt; |
159 | int ver; |
160 | |
161 | gcc_assert (op); |
162 | |
163 | ver = SSA_NAME_VERSION (op); |
164 | if (bitmap_bit_p (map: processed, bitno: ver)) |
165 | { |
166 | stmt = SSA_NAME_DEF_STMT (op); |
167 | gcc_assert (gimple_nop_p (stmt) |
168 | || gimple_plf (stmt, STMT_NECESSARY)); |
169 | return; |
170 | } |
171 | bitmap_set_bit (map: processed, bitno: ver); |
172 | |
173 | stmt = SSA_NAME_DEF_STMT (op); |
174 | gcc_assert (stmt); |
175 | |
176 | if (gimple_plf (stmt, STMT_NECESSARY) || gimple_nop_p (g: stmt)) |
177 | return; |
178 | |
179 | if (dump_file && (dump_flags & TDF_DETAILS)) |
180 | { |
181 | fprintf (stream: dump_file, format: "marking necessary through " ); |
182 | print_generic_expr (dump_file, op); |
183 | fprintf (stream: dump_file, format: " stmt " ); |
184 | print_gimple_stmt (dump_file, stmt, 0); |
185 | } |
186 | |
187 | gimple_set_plf (stmt, STMT_NECESSARY, val_p: true); |
188 | if (bb_contains_live_stmts) |
189 | bitmap_set_bit (map: bb_contains_live_stmts, bitno: gimple_bb (g: stmt)->index); |
190 | worklist.safe_push (obj: stmt); |
191 | } |
192 | |
193 | |
194 | /* Mark STMT as necessary if it obviously is. Add it to the worklist if |
195 | it can make other statements necessary. |
196 | |
197 | If AGGRESSIVE is false, control statements are conservatively marked as |
198 | necessary. */ |
199 | |
200 | static void |
201 | mark_stmt_if_obviously_necessary (gimple *stmt, bool aggressive) |
202 | { |
203 | /* Statements that are implicitly live. Most function calls, asm |
204 | and return statements are required. Labels and GIMPLE_BIND nodes |
205 | are kept because they are control flow, and we have no way of |
206 | knowing whether they can be removed. DCE can eliminate all the |
207 | other statements in a block, and CFG can then remove the block |
208 | and labels. */ |
209 | switch (gimple_code (g: stmt)) |
210 | { |
211 | case GIMPLE_PREDICT: |
212 | case GIMPLE_LABEL: |
213 | mark_stmt_necessary (stmt, add_to_worklist: false); |
214 | return; |
215 | |
216 | case GIMPLE_ASM: |
217 | case GIMPLE_RESX: |
218 | case GIMPLE_RETURN: |
219 | mark_stmt_necessary (stmt, add_to_worklist: true); |
220 | return; |
221 | |
222 | case GIMPLE_CALL: |
223 | { |
224 | /* Never elide a noreturn call we pruned control-flow for. */ |
225 | if ((gimple_call_flags (stmt) & ECF_NORETURN) |
226 | && gimple_call_ctrl_altering_p (gs: stmt)) |
227 | { |
228 | mark_stmt_necessary (stmt, add_to_worklist: true); |
229 | return; |
230 | } |
231 | |
232 | tree callee = gimple_call_fndecl (gs: stmt); |
233 | if (callee != NULL_TREE |
234 | && fndecl_built_in_p (node: callee, klass: BUILT_IN_NORMAL)) |
235 | switch (DECL_FUNCTION_CODE (decl: callee)) |
236 | { |
237 | case BUILT_IN_MALLOC: |
238 | case BUILT_IN_ALIGNED_ALLOC: |
239 | case BUILT_IN_CALLOC: |
240 | CASE_BUILT_IN_ALLOCA: |
241 | case BUILT_IN_STRDUP: |
242 | case BUILT_IN_STRNDUP: |
243 | case BUILT_IN_GOMP_ALLOC: |
244 | return; |
245 | |
246 | default:; |
247 | } |
248 | |
249 | if (callee != NULL_TREE |
250 | && flag_allocation_dce |
251 | && DECL_IS_REPLACEABLE_OPERATOR_NEW_P (callee)) |
252 | return; |
253 | |
254 | /* IFN_GOACC_LOOP calls are necessary in that they are used to |
255 | represent parameter (i.e. step, bound) of a lowered OpenACC |
256 | partitioned loop. But this kind of partitioned loop might not |
257 | survive from aggressive loop removal for it has loop exit and |
258 | is assumed to be finite. Therefore, we need to explicitly mark |
259 | these calls. (An example is libgomp.oacc-c-c++-common/pr84955.c) */ |
260 | if (gimple_call_internal_p (gs: stmt, fn: IFN_GOACC_LOOP)) |
261 | { |
262 | mark_stmt_necessary (stmt, add_to_worklist: true); |
263 | return; |
264 | } |
265 | break; |
266 | } |
267 | |
268 | case GIMPLE_DEBUG: |
269 | /* Debug temps without a value are not useful. ??? If we could |
270 | easily locate the debug temp bind stmt for a use thereof, |
271 | would could refrain from marking all debug temps here, and |
272 | mark them only if they're used. */ |
273 | if (gimple_debug_nonbind_marker_p (s: stmt) |
274 | || !gimple_debug_bind_p (s: stmt) |
275 | || gimple_debug_bind_has_value_p (dbg: stmt) |
276 | || TREE_CODE (gimple_debug_bind_get_var (stmt)) != DEBUG_EXPR_DECL) |
277 | mark_stmt_necessary (stmt, add_to_worklist: false); |
278 | return; |
279 | |
280 | case GIMPLE_GOTO: |
281 | gcc_assert (!simple_goto_p (stmt)); |
282 | mark_stmt_necessary (stmt, add_to_worklist: true); |
283 | return; |
284 | |
285 | case GIMPLE_COND: |
286 | gcc_assert (EDGE_COUNT (gimple_bb (stmt)->succs) == 2); |
287 | /* Fall through. */ |
288 | |
289 | case GIMPLE_SWITCH: |
290 | if (! aggressive) |
291 | mark_stmt_necessary (stmt, add_to_worklist: true); |
292 | break; |
293 | |
294 | case GIMPLE_ASSIGN: |
295 | /* Mark indirect CLOBBERs to be lazily removed if their SSA operands |
296 | do not prevail. That also makes control flow leading to them |
297 | not necessary in aggressive mode. */ |
298 | if (gimple_clobber_p (s: stmt) && !zero_ssa_operands (stmt, SSA_OP_USE)) |
299 | return; |
300 | break; |
301 | |
302 | default: |
303 | break; |
304 | } |
305 | |
306 | /* If the statement has volatile operands, it needs to be preserved. |
307 | Same for statements that can alter control flow in unpredictable |
308 | ways. */ |
309 | if (gimple_has_side_effects (stmt) || is_ctrl_altering_stmt (stmt)) |
310 | { |
311 | mark_stmt_necessary (stmt, add_to_worklist: true); |
312 | return; |
313 | } |
314 | |
315 | /* If a statement could throw, it can be deemed necessary unless we |
316 | are allowed to remove dead EH. Test this after checking for |
317 | new/delete operators since we always elide their EH. */ |
318 | if (!cfun->can_delete_dead_exceptions |
319 | && stmt_could_throw_p (cfun, stmt)) |
320 | { |
321 | mark_stmt_necessary (stmt, add_to_worklist: true); |
322 | return; |
323 | } |
324 | |
325 | if ((gimple_vdef (g: stmt) && keep_all_vdefs_p ()) |
326 | || stmt_may_clobber_global_p (stmt, false)) |
327 | { |
328 | mark_stmt_necessary (stmt, add_to_worklist: true); |
329 | return; |
330 | } |
331 | |
332 | return; |
333 | } |
334 | |
335 | |
336 | /* Mark the last statement of BB as necessary. */ |
337 | |
338 | static bool |
339 | mark_last_stmt_necessary (basic_block bb) |
340 | { |
341 | if (!bitmap_set_bit (map: last_stmt_necessary, bitno: bb->index)) |
342 | return true; |
343 | |
344 | bitmap_set_bit (map: bb_contains_live_stmts, bitno: bb->index); |
345 | |
346 | /* We actually mark the statement only if it is a control statement. */ |
347 | gimple *stmt = *gsi_last_bb (bb); |
348 | if (stmt && is_ctrl_stmt (stmt)) |
349 | { |
350 | mark_stmt_necessary (stmt, add_to_worklist: true); |
351 | return true; |
352 | } |
353 | return false; |
354 | } |
355 | |
356 | |
357 | /* Mark control dependent edges of BB as necessary. We have to do this only |
358 | once for each basic block so we set the appropriate bit after we're done. |
359 | |
360 | When IGNORE_SELF is true, ignore BB in the list of control dependences. */ |
361 | |
362 | static void |
363 | mark_control_dependent_edges_necessary (basic_block bb, bool ignore_self) |
364 | { |
365 | bitmap_iterator bi; |
366 | unsigned edge_number; |
367 | bool skipped = false; |
368 | |
369 | gcc_assert (bb != EXIT_BLOCK_PTR_FOR_FN (cfun)); |
370 | |
371 | if (bb == ENTRY_BLOCK_PTR_FOR_FN (cfun)) |
372 | return; |
373 | |
374 | EXECUTE_IF_SET_IN_BITMAP (cd->get_edges_dependent_on (bb->index), |
375 | 0, edge_number, bi) |
376 | { |
377 | basic_block cd_bb = cd->get_edge_src (edge_number); |
378 | |
379 | if (ignore_self && cd_bb == bb) |
380 | { |
381 | skipped = true; |
382 | continue; |
383 | } |
384 | |
385 | if (!mark_last_stmt_necessary (bb: cd_bb)) |
386 | mark_control_dependent_edges_necessary (bb: cd_bb, ignore_self: false); |
387 | } |
388 | |
389 | if (!skipped) |
390 | bitmap_set_bit (map: visited_control_parents, bitno: bb->index); |
391 | } |
392 | |
393 | |
394 | /* Find obviously necessary statements. These are things like most function |
395 | calls, and stores to file level variables. |
396 | |
397 | If EL is NULL, control statements are conservatively marked as |
398 | necessary. Otherwise it contains the list of edges used by control |
399 | dependence analysis. */ |
400 | |
401 | static void |
402 | find_obviously_necessary_stmts (bool aggressive) |
403 | { |
404 | basic_block bb; |
405 | gimple_stmt_iterator gsi; |
406 | edge e; |
407 | gimple *phi, *stmt; |
408 | int flags; |
409 | |
410 | FOR_EACH_BB_FN (bb, cfun) |
411 | { |
412 | /* PHI nodes are never inherently necessary. */ |
413 | for (gsi = gsi_start_phis (bb); !gsi_end_p (i: gsi); gsi_next (i: &gsi)) |
414 | { |
415 | phi = gsi_stmt (i: gsi); |
416 | gimple_set_plf (stmt: phi, STMT_NECESSARY, val_p: false); |
417 | } |
418 | |
419 | /* Check all statements in the block. */ |
420 | for (gsi = gsi_start_bb (bb); !gsi_end_p (i: gsi); gsi_next (i: &gsi)) |
421 | { |
422 | stmt = gsi_stmt (i: gsi); |
423 | gimple_set_plf (stmt, STMT_NECESSARY, val_p: false); |
424 | mark_stmt_if_obviously_necessary (stmt, aggressive); |
425 | } |
426 | } |
427 | |
428 | /* Pure and const functions are finite and thus have no infinite loops in |
429 | them. */ |
430 | flags = flags_from_decl_or_type (current_function_decl); |
431 | if ((flags & (ECF_CONST|ECF_PURE)) && !(flags & ECF_LOOPING_CONST_OR_PURE)) |
432 | return; |
433 | |
434 | /* Prevent the empty possibly infinite loops from being removed. This is |
435 | needed to make the logic in remove_dead_stmt work to identify the |
436 | correct edge to keep when removing a controlling condition. */ |
437 | if (aggressive) |
438 | { |
439 | if (mark_irreducible_loops ()) |
440 | FOR_EACH_BB_FN (bb, cfun) |
441 | { |
442 | edge_iterator ei; |
443 | FOR_EACH_EDGE (e, ei, bb->succs) |
444 | if ((e->flags & EDGE_DFS_BACK) |
445 | && (e->flags & EDGE_IRREDUCIBLE_LOOP)) |
446 | { |
447 | if (dump_file) |
448 | fprintf (stream: dump_file, format: "Marking back edge of irreducible " |
449 | "loop %i->%i\n" , e->src->index, e->dest->index); |
450 | mark_control_dependent_edges_necessary (bb: e->dest, ignore_self: false); |
451 | } |
452 | } |
453 | |
454 | for (auto loop : loops_list (cfun, 0)) |
455 | /* For loops without an exit do not mark any condition. */ |
456 | if (loop->exits->next->e && !finite_loop_p (loop)) |
457 | { |
458 | if (dump_file) |
459 | fprintf (stream: dump_file, format: "cannot prove finiteness of loop %i\n" , |
460 | loop->num); |
461 | mark_control_dependent_edges_necessary (bb: loop->latch, ignore_self: false); |
462 | } |
463 | } |
464 | } |
465 | |
466 | |
467 | /* Return true if REF is based on an aliased base, otherwise false. */ |
468 | |
469 | static bool |
470 | ref_may_be_aliased (tree ref) |
471 | { |
472 | gcc_assert (TREE_CODE (ref) != WITH_SIZE_EXPR); |
473 | while (handled_component_p (t: ref)) |
474 | ref = TREE_OPERAND (ref, 0); |
475 | if ((TREE_CODE (ref) == MEM_REF || TREE_CODE (ref) == TARGET_MEM_REF) |
476 | && TREE_CODE (TREE_OPERAND (ref, 0)) == ADDR_EXPR) |
477 | ref = TREE_OPERAND (TREE_OPERAND (ref, 0), 0); |
478 | return !(DECL_P (ref) |
479 | && !may_be_aliased (var: ref)); |
480 | } |
481 | |
482 | static bitmap visited = NULL; |
483 | static unsigned int longest_chain = 0; |
484 | static unsigned int total_chain = 0; |
485 | static unsigned int nr_walks = 0; |
486 | static bool chain_ovfl = false; |
487 | |
488 | /* Worker for the walker that marks reaching definitions of REF, |
489 | which is based on a non-aliased decl, necessary. It returns |
490 | true whenever the defining statement of the current VDEF is |
491 | a kill for REF, as no dominating may-defs are necessary for REF |
492 | anymore. DATA points to the basic-block that contains the |
493 | stmt that refers to REF. */ |
494 | |
495 | static bool |
496 | mark_aliased_reaching_defs_necessary_1 (ao_ref *ref, tree vdef, void *data) |
497 | { |
498 | gimple *def_stmt = SSA_NAME_DEF_STMT (vdef); |
499 | |
500 | /* All stmts we visit are necessary. */ |
501 | if (! gimple_clobber_p (s: def_stmt)) |
502 | mark_operand_necessary (op: vdef); |
503 | |
504 | /* If the stmt lhs kills ref, then we can stop walking. */ |
505 | if (gimple_has_lhs (stmt: def_stmt) |
506 | && TREE_CODE (gimple_get_lhs (def_stmt)) != SSA_NAME |
507 | /* The assignment is not necessarily carried out if it can throw |
508 | and we can catch it in the current function where we could inspect |
509 | the previous value. |
510 | ??? We only need to care about the RHS throwing. For aggregate |
511 | assignments or similar calls and non-call exceptions the LHS |
512 | might throw as well. */ |
513 | && !stmt_can_throw_internal (cfun, def_stmt)) |
514 | { |
515 | tree base, lhs = gimple_get_lhs (def_stmt); |
516 | poly_int64 size, offset, max_size; |
517 | bool reverse; |
518 | ao_ref_base (ref); |
519 | base |
520 | = get_ref_base_and_extent (lhs, &offset, &size, &max_size, &reverse); |
521 | /* We can get MEM[symbol: sZ, index: D.8862_1] here, |
522 | so base == refd->base does not always hold. */ |
523 | if (base == ref->base) |
524 | { |
525 | /* For a must-alias check we need to be able to constrain |
526 | the accesses properly. */ |
527 | if (known_eq (size, max_size) |
528 | && known_subrange_p (pos1: ref->offset, size1: ref->max_size, pos2: offset, size2: size)) |
529 | return true; |
530 | /* Or they need to be exactly the same. */ |
531 | else if (ref->ref |
532 | /* Make sure there is no induction variable involved |
533 | in the references (gcc.c-torture/execute/pr42142.c). |
534 | The simplest way is to check if the kill dominates |
535 | the use. */ |
536 | /* But when both are in the same block we cannot |
537 | easily tell whether we came from a backedge |
538 | unless we decide to compute stmt UIDs |
539 | (see PR58246). */ |
540 | && (basic_block) data != gimple_bb (g: def_stmt) |
541 | && dominated_by_p (CDI_DOMINATORS, (basic_block) data, |
542 | gimple_bb (g: def_stmt)) |
543 | && operand_equal_p (ref->ref, lhs, flags: 0)) |
544 | return true; |
545 | } |
546 | } |
547 | |
548 | /* Otherwise keep walking. */ |
549 | return false; |
550 | } |
551 | |
552 | static void |
553 | mark_aliased_reaching_defs_necessary (gimple *stmt, tree ref) |
554 | { |
555 | /* Should have been caught before calling this function. */ |
556 | gcc_checking_assert (!keep_all_vdefs_p ()); |
557 | |
558 | unsigned int chain; |
559 | ao_ref refd; |
560 | gcc_assert (!chain_ovfl); |
561 | ao_ref_init (&refd, ref); |
562 | chain = walk_aliased_vdefs (&refd, gimple_vuse (g: stmt), |
563 | mark_aliased_reaching_defs_necessary_1, |
564 | gimple_bb (g: stmt), NULL); |
565 | if (chain > longest_chain) |
566 | longest_chain = chain; |
567 | total_chain += chain; |
568 | nr_walks++; |
569 | } |
570 | |
571 | /* Worker for the walker that marks reaching definitions of REF, which |
572 | is not based on a non-aliased decl. For simplicity we need to end |
573 | up marking all may-defs necessary that are not based on a non-aliased |
574 | decl. The only job of this walker is to skip may-defs based on |
575 | a non-aliased decl. */ |
576 | |
577 | static bool |
578 | mark_all_reaching_defs_necessary_1 (ao_ref *ref ATTRIBUTE_UNUSED, |
579 | tree vdef, void *data ATTRIBUTE_UNUSED) |
580 | { |
581 | gimple *def_stmt = SSA_NAME_DEF_STMT (vdef); |
582 | |
583 | /* We have to skip already visited (and thus necessary) statements |
584 | to make the chaining work after we dropped back to simple mode. */ |
585 | if (chain_ovfl |
586 | && bitmap_bit_p (map: processed, SSA_NAME_VERSION (vdef))) |
587 | { |
588 | gcc_assert (gimple_nop_p (def_stmt) |
589 | || gimple_plf (def_stmt, STMT_NECESSARY)); |
590 | return false; |
591 | } |
592 | |
593 | /* We want to skip stores to non-aliased variables. */ |
594 | if (!chain_ovfl |
595 | && gimple_assign_single_p (gs: def_stmt)) |
596 | { |
597 | tree lhs = gimple_assign_lhs (gs: def_stmt); |
598 | if (!ref_may_be_aliased (ref: lhs)) |
599 | return false; |
600 | } |
601 | |
602 | /* We want to skip statments that do not constitute stores but have |
603 | a virtual definition. */ |
604 | if (gcall *call = dyn_cast <gcall *> (p: def_stmt)) |
605 | { |
606 | tree callee = gimple_call_fndecl (gs: call); |
607 | if (callee != NULL_TREE |
608 | && fndecl_built_in_p (node: callee, klass: BUILT_IN_NORMAL)) |
609 | switch (DECL_FUNCTION_CODE (decl: callee)) |
610 | { |
611 | case BUILT_IN_MALLOC: |
612 | case BUILT_IN_ALIGNED_ALLOC: |
613 | case BUILT_IN_CALLOC: |
614 | CASE_BUILT_IN_ALLOCA: |
615 | case BUILT_IN_FREE: |
616 | case BUILT_IN_GOMP_ALLOC: |
617 | case BUILT_IN_GOMP_FREE: |
618 | return false; |
619 | |
620 | default:; |
621 | } |
622 | |
623 | if (callee != NULL_TREE |
624 | && (DECL_IS_REPLACEABLE_OPERATOR_NEW_P (callee) |
625 | || DECL_IS_OPERATOR_DELETE_P (callee)) |
626 | && gimple_call_from_new_or_delete (s: call)) |
627 | return false; |
628 | } |
629 | |
630 | if (! gimple_clobber_p (s: def_stmt)) |
631 | mark_operand_necessary (op: vdef); |
632 | |
633 | return false; |
634 | } |
635 | |
636 | static void |
637 | mark_all_reaching_defs_necessary (gimple *stmt) |
638 | { |
639 | /* Should have been caught before calling this function. */ |
640 | gcc_checking_assert (!keep_all_vdefs_p ()); |
641 | walk_aliased_vdefs (NULL, gimple_vuse (g: stmt), |
642 | mark_all_reaching_defs_necessary_1, NULL, &visited); |
643 | } |
644 | |
645 | /* Return true for PHI nodes with one or identical arguments |
646 | can be removed. */ |
647 | static bool |
648 | degenerate_phi_p (gimple *phi) |
649 | { |
650 | unsigned int i; |
651 | tree op = gimple_phi_arg_def (gs: phi, index: 0); |
652 | for (i = 1; i < gimple_phi_num_args (gs: phi); i++) |
653 | if (gimple_phi_arg_def (gs: phi, index: i) != op) |
654 | return false; |
655 | return true; |
656 | } |
657 | |
658 | /* Return that NEW_CALL and DELETE_CALL are a valid pair of new |
659 | and delete operators. */ |
660 | |
661 | static bool |
662 | valid_new_delete_pair_p (gimple *new_call, gimple *delete_call) |
663 | { |
664 | tree new_asm = DECL_ASSEMBLER_NAME (gimple_call_fndecl (new_call)); |
665 | tree delete_asm = DECL_ASSEMBLER_NAME (gimple_call_fndecl (delete_call)); |
666 | return valid_new_delete_pair_p (new_asm, delete_asm); |
667 | } |
668 | |
669 | /* Propagate necessity using the operands of necessary statements. |
670 | Process the uses on each statement in the worklist, and add all |
671 | feeding statements which contribute to the calculation of this |
672 | value to the worklist. |
673 | |
674 | In conservative mode, EL is NULL. */ |
675 | |
676 | static void |
677 | propagate_necessity (bool aggressive) |
678 | { |
679 | gimple *stmt; |
680 | |
681 | if (dump_file && (dump_flags & TDF_DETAILS)) |
682 | fprintf (stream: dump_file, format: "\nProcessing worklist:\n" ); |
683 | |
684 | while (worklist.length () > 0) |
685 | { |
686 | /* Take STMT from worklist. */ |
687 | stmt = worklist.pop (); |
688 | |
689 | if (dump_file && (dump_flags & TDF_DETAILS)) |
690 | { |
691 | fprintf (stream: dump_file, format: "processing: " ); |
692 | print_gimple_stmt (dump_file, stmt, 0, TDF_SLIM); |
693 | fprintf (stream: dump_file, format: "\n" ); |
694 | } |
695 | |
696 | if (aggressive) |
697 | { |
698 | /* Mark the last statement of the basic blocks on which the block |
699 | containing STMT is control dependent, but only if we haven't |
700 | already done so. */ |
701 | basic_block bb = gimple_bb (g: stmt); |
702 | if (bb != ENTRY_BLOCK_PTR_FOR_FN (cfun) |
703 | && !bitmap_bit_p (map: visited_control_parents, bitno: bb->index)) |
704 | mark_control_dependent_edges_necessary (bb, ignore_self: false); |
705 | } |
706 | |
707 | if (gimple_code (g: stmt) == GIMPLE_PHI |
708 | /* We do not process virtual PHI nodes nor do we track their |
709 | necessity. */ |
710 | && !virtual_operand_p (op: gimple_phi_result (gs: stmt))) |
711 | { |
712 | /* PHI nodes are somewhat special in that each PHI alternative has |
713 | data and control dependencies. All the statements feeding the |
714 | PHI node's arguments are always necessary. In aggressive mode, |
715 | we also consider the control dependent edges leading to the |
716 | predecessor block associated with each PHI alternative as |
717 | necessary. */ |
718 | gphi *phi = as_a <gphi *> (p: stmt); |
719 | size_t k; |
720 | |
721 | for (k = 0; k < gimple_phi_num_args (gs: stmt); k++) |
722 | { |
723 | tree arg = PHI_ARG_DEF (stmt, k); |
724 | if (TREE_CODE (arg) == SSA_NAME) |
725 | mark_operand_necessary (op: arg); |
726 | } |
727 | |
728 | /* For PHI operands it matters from where the control flow arrives |
729 | to the BB. Consider the following example: |
730 | |
731 | a=exp1; |
732 | b=exp2; |
733 | if (test) |
734 | ; |
735 | else |
736 | ; |
737 | c=PHI(a,b) |
738 | |
739 | We need to mark control dependence of the empty basic blocks, since they |
740 | contains computation of PHI operands. |
741 | |
742 | Doing so is too restrictive in the case the predecestor block is in |
743 | the loop. Consider: |
744 | |
745 | if (b) |
746 | { |
747 | int i; |
748 | for (i = 0; i<1000; ++i) |
749 | ; |
750 | j = 0; |
751 | } |
752 | return j; |
753 | |
754 | There is PHI for J in the BB containing return statement. |
755 | In this case the control dependence of predecestor block (that is |
756 | within the empty loop) also contains the block determining number |
757 | of iterations of the block that would prevent removing of empty |
758 | loop in this case. |
759 | |
760 | This scenario can be avoided by splitting critical edges. |
761 | To save the critical edge splitting pass we identify how the control |
762 | dependence would look like if the edge was split. |
763 | |
764 | Consider the modified CFG created from current CFG by splitting |
765 | edge B->C. In the postdominance tree of modified CFG, C' is |
766 | always child of C. There are two cases how chlids of C' can look |
767 | like: |
768 | |
769 | 1) C' is leaf |
770 | |
771 | In this case the only basic block C' is control dependent on is B. |
772 | |
773 | 2) C' has single child that is B |
774 | |
775 | In this case control dependence of C' is same as control |
776 | dependence of B in original CFG except for block B itself. |
777 | (since C' postdominate B in modified CFG) |
778 | |
779 | Now how to decide what case happens? There are two basic options: |
780 | |
781 | a) C postdominate B. Then C immediately postdominate B and |
782 | case 2 happens iff there is no other way from B to C except |
783 | the edge B->C. |
784 | |
785 | There is other way from B to C iff there is succesor of B that |
786 | is not postdominated by B. Testing this condition is somewhat |
787 | expensive, because we need to iterate all succesors of B. |
788 | We are safe to assume that this does not happen: we will mark B |
789 | as needed when processing the other path from B to C that is |
790 | conrol dependent on B and marking control dependencies of B |
791 | itself is harmless because they will be processed anyway after |
792 | processing control statement in B. |
793 | |
794 | b) C does not postdominate B. Always case 1 happens since there is |
795 | path from C to exit that does not go through B and thus also C'. */ |
796 | |
797 | if (aggressive && !degenerate_phi_p (phi: stmt)) |
798 | { |
799 | for (k = 0; k < gimple_phi_num_args (gs: stmt); k++) |
800 | { |
801 | basic_block arg_bb = gimple_phi_arg_edge (phi, i: k)->src; |
802 | |
803 | if (gimple_bb (g: stmt) |
804 | != get_immediate_dominator (CDI_POST_DOMINATORS, arg_bb)) |
805 | { |
806 | if (!mark_last_stmt_necessary (bb: arg_bb)) |
807 | mark_control_dependent_edges_necessary (bb: arg_bb, ignore_self: false); |
808 | } |
809 | else if (arg_bb != ENTRY_BLOCK_PTR_FOR_FN (cfun) |
810 | && !bitmap_bit_p (map: visited_control_parents, |
811 | bitno: arg_bb->index)) |
812 | mark_control_dependent_edges_necessary (bb: arg_bb, ignore_self: true); |
813 | } |
814 | } |
815 | } |
816 | else |
817 | { |
818 | /* Propagate through the operands. Examine all the USE, VUSE and |
819 | VDEF operands in this statement. Mark all the statements |
820 | which feed this statement's uses as necessary. */ |
821 | ssa_op_iter iter; |
822 | tree use; |
823 | |
824 | /* If this is a call to free which is directly fed by an |
825 | allocation function do not mark that necessary through |
826 | processing the argument. */ |
827 | bool is_delete_operator |
828 | = (is_gimple_call (gs: stmt) |
829 | && gimple_call_from_new_or_delete (s: as_a <gcall *> (p: stmt)) |
830 | && gimple_call_operator_delete_p (as_a <gcall *> (p: stmt))); |
831 | if (is_delete_operator |
832 | || gimple_call_builtin_p (stmt, BUILT_IN_FREE) |
833 | || gimple_call_builtin_p (stmt, BUILT_IN_GOMP_FREE)) |
834 | { |
835 | tree ptr = gimple_call_arg (gs: stmt, index: 0); |
836 | gcall *def_stmt; |
837 | tree def_callee; |
838 | /* If the pointer we free is defined by an allocation |
839 | function do not add the call to the worklist. */ |
840 | if (TREE_CODE (ptr) == SSA_NAME |
841 | && (def_stmt = dyn_cast <gcall *> (SSA_NAME_DEF_STMT (ptr))) |
842 | && (def_callee = gimple_call_fndecl (gs: def_stmt)) |
843 | && ((DECL_BUILT_IN_CLASS (def_callee) == BUILT_IN_NORMAL |
844 | && (DECL_FUNCTION_CODE (decl: def_callee) == BUILT_IN_ALIGNED_ALLOC |
845 | || DECL_FUNCTION_CODE (decl: def_callee) == BUILT_IN_MALLOC |
846 | || DECL_FUNCTION_CODE (decl: def_callee) == BUILT_IN_CALLOC |
847 | || DECL_FUNCTION_CODE (decl: def_callee) == BUILT_IN_GOMP_ALLOC)) |
848 | || (DECL_IS_REPLACEABLE_OPERATOR_NEW_P (def_callee) |
849 | && gimple_call_from_new_or_delete (s: def_stmt)))) |
850 | { |
851 | if (is_delete_operator |
852 | && !valid_new_delete_pair_p (new_call: def_stmt, delete_call: stmt)) |
853 | mark_operand_necessary (op: gimple_call_arg (gs: stmt, index: 0)); |
854 | |
855 | /* Delete operators can have alignment and (or) size |
856 | as next arguments. When being a SSA_NAME, they |
857 | must be marked as necessary. Similarly GOMP_free. */ |
858 | if (gimple_call_num_args (gs: stmt) >= 2) |
859 | for (unsigned i = 1; i < gimple_call_num_args (gs: stmt); |
860 | i++) |
861 | { |
862 | tree arg = gimple_call_arg (gs: stmt, index: i); |
863 | if (TREE_CODE (arg) == SSA_NAME) |
864 | mark_operand_necessary (op: arg); |
865 | } |
866 | |
867 | continue; |
868 | } |
869 | } |
870 | |
871 | FOR_EACH_SSA_TREE_OPERAND (use, stmt, iter, SSA_OP_USE) |
872 | mark_operand_necessary (op: use); |
873 | |
874 | use = gimple_vuse (g: stmt); |
875 | if (!use) |
876 | continue; |
877 | |
878 | /* No need to search for vdefs if we intrinsicly keep them all. */ |
879 | if (keep_all_vdefs_p ()) |
880 | continue; |
881 | |
882 | /* If we dropped to simple mode make all immediately |
883 | reachable definitions necessary. */ |
884 | if (chain_ovfl) |
885 | { |
886 | mark_all_reaching_defs_necessary (stmt); |
887 | continue; |
888 | } |
889 | |
890 | /* For statements that may load from memory (have a VUSE) we |
891 | have to mark all reaching (may-)definitions as necessary. |
892 | We partition this task into two cases: |
893 | 1) explicit loads based on decls that are not aliased |
894 | 2) implicit loads (like calls) and explicit loads not |
895 | based on decls that are not aliased (like indirect |
896 | references or loads from globals) |
897 | For 1) we mark all reaching may-defs as necessary, stopping |
898 | at dominating kills. For 2) we want to mark all dominating |
899 | references necessary, but non-aliased ones which we handle |
900 | in 1). By keeping a global visited bitmap for references |
901 | we walk for 2) we avoid quadratic behavior for those. */ |
902 | |
903 | if (gcall *call = dyn_cast <gcall *> (p: stmt)) |
904 | { |
905 | tree callee = gimple_call_fndecl (gs: call); |
906 | unsigned i; |
907 | |
908 | /* Calls to functions that are merely acting as barriers |
909 | or that only store to memory do not make any previous |
910 | stores necessary. */ |
911 | if (callee != NULL_TREE |
912 | && DECL_BUILT_IN_CLASS (callee) == BUILT_IN_NORMAL |
913 | && (DECL_FUNCTION_CODE (decl: callee) == BUILT_IN_MEMSET |
914 | || DECL_FUNCTION_CODE (decl: callee) == BUILT_IN_MEMSET_CHK |
915 | || DECL_FUNCTION_CODE (decl: callee) == BUILT_IN_MALLOC |
916 | || DECL_FUNCTION_CODE (decl: callee) == BUILT_IN_ALIGNED_ALLOC |
917 | || DECL_FUNCTION_CODE (decl: callee) == BUILT_IN_CALLOC |
918 | || DECL_FUNCTION_CODE (decl: callee) == BUILT_IN_FREE |
919 | || DECL_FUNCTION_CODE (decl: callee) == BUILT_IN_VA_END |
920 | || ALLOCA_FUNCTION_CODE_P (DECL_FUNCTION_CODE (callee)) |
921 | || DECL_FUNCTION_CODE (decl: callee) == BUILT_IN_STACK_SAVE |
922 | || DECL_FUNCTION_CODE (decl: callee) == BUILT_IN_STACK_RESTORE |
923 | || DECL_FUNCTION_CODE (decl: callee) == BUILT_IN_ASSUME_ALIGNED)) |
924 | continue; |
925 | |
926 | if (callee != NULL_TREE |
927 | && (DECL_IS_REPLACEABLE_OPERATOR_NEW_P (callee) |
928 | || DECL_IS_OPERATOR_DELETE_P (callee)) |
929 | && gimple_call_from_new_or_delete (s: call)) |
930 | continue; |
931 | |
932 | /* Calls implicitly load from memory, their arguments |
933 | in addition may explicitly perform memory loads. */ |
934 | mark_all_reaching_defs_necessary (stmt: call); |
935 | for (i = 0; i < gimple_call_num_args (gs: call); ++i) |
936 | { |
937 | tree arg = gimple_call_arg (gs: call, index: i); |
938 | if (TREE_CODE (arg) == SSA_NAME |
939 | || is_gimple_min_invariant (arg)) |
940 | continue; |
941 | if (TREE_CODE (arg) == WITH_SIZE_EXPR) |
942 | arg = TREE_OPERAND (arg, 0); |
943 | if (!ref_may_be_aliased (ref: arg)) |
944 | mark_aliased_reaching_defs_necessary (stmt: call, ref: arg); |
945 | } |
946 | } |
947 | else if (gimple_assign_single_p (gs: stmt)) |
948 | { |
949 | tree rhs; |
950 | /* If this is a load mark things necessary. */ |
951 | rhs = gimple_assign_rhs1 (gs: stmt); |
952 | if (TREE_CODE (rhs) != SSA_NAME |
953 | && !is_gimple_min_invariant (rhs) |
954 | && TREE_CODE (rhs) != CONSTRUCTOR) |
955 | { |
956 | if (!ref_may_be_aliased (ref: rhs)) |
957 | mark_aliased_reaching_defs_necessary (stmt, ref: rhs); |
958 | else |
959 | mark_all_reaching_defs_necessary (stmt); |
960 | } |
961 | } |
962 | else if (greturn *return_stmt = dyn_cast <greturn *> (p: stmt)) |
963 | { |
964 | tree rhs = gimple_return_retval (gs: return_stmt); |
965 | /* A return statement may perform a load. */ |
966 | if (rhs |
967 | && TREE_CODE (rhs) != SSA_NAME |
968 | && !is_gimple_min_invariant (rhs) |
969 | && TREE_CODE (rhs) != CONSTRUCTOR) |
970 | { |
971 | if (!ref_may_be_aliased (ref: rhs)) |
972 | mark_aliased_reaching_defs_necessary (stmt, ref: rhs); |
973 | else |
974 | mark_all_reaching_defs_necessary (stmt); |
975 | } |
976 | } |
977 | else if (gasm *asm_stmt = dyn_cast <gasm *> (p: stmt)) |
978 | { |
979 | unsigned i; |
980 | mark_all_reaching_defs_necessary (stmt); |
981 | /* Inputs may perform loads. */ |
982 | for (i = 0; i < gimple_asm_ninputs (asm_stmt); ++i) |
983 | { |
984 | tree op = TREE_VALUE (gimple_asm_input_op (asm_stmt, i)); |
985 | if (TREE_CODE (op) != SSA_NAME |
986 | && !is_gimple_min_invariant (op) |
987 | && TREE_CODE (op) != CONSTRUCTOR |
988 | && !ref_may_be_aliased (ref: op)) |
989 | mark_aliased_reaching_defs_necessary (stmt, ref: op); |
990 | } |
991 | } |
992 | else if (gimple_code (g: stmt) == GIMPLE_TRANSACTION) |
993 | { |
994 | /* The beginning of a transaction is a memory barrier. */ |
995 | /* ??? If we were really cool, we'd only be a barrier |
996 | for the memories touched within the transaction. */ |
997 | mark_all_reaching_defs_necessary (stmt); |
998 | } |
999 | else |
1000 | gcc_unreachable (); |
1001 | |
1002 | /* If we over-used our alias oracle budget drop to simple |
1003 | mode. The cost metric allows quadratic behavior |
1004 | (number of uses times number of may-defs queries) up to |
1005 | a constant maximal number of queries and after that falls back to |
1006 | super-linear complexity. */ |
1007 | if (/* Constant but quadratic for small functions. */ |
1008 | total_chain > 128 * 128 |
1009 | /* Linear in the number of may-defs. */ |
1010 | && total_chain > 32 * longest_chain |
1011 | /* Linear in the number of uses. */ |
1012 | && total_chain > nr_walks * 32) |
1013 | { |
1014 | chain_ovfl = true; |
1015 | if (visited) |
1016 | bitmap_clear (visited); |
1017 | } |
1018 | } |
1019 | } |
1020 | } |
1021 | |
1022 | /* Remove dead PHI nodes from block BB. */ |
1023 | |
1024 | static bool |
1025 | remove_dead_phis (basic_block bb) |
1026 | { |
1027 | bool something_changed = false; |
1028 | gphi *phi; |
1029 | gphi_iterator gsi; |
1030 | |
1031 | for (gsi = gsi_start_phis (bb); !gsi_end_p (i: gsi);) |
1032 | { |
1033 | stats.total_phis++; |
1034 | phi = gsi.phi (); |
1035 | |
1036 | /* We do not track necessity of virtual PHI nodes. Instead do |
1037 | very simple dead PHI removal here. */ |
1038 | if (virtual_operand_p (op: gimple_phi_result (gs: phi))) |
1039 | { |
1040 | /* Virtual PHI nodes with one or identical arguments |
1041 | can be removed. */ |
1042 | if (!loops_state_satisfies_p (flags: LOOP_CLOSED_SSA) |
1043 | && degenerate_phi_p (phi)) |
1044 | { |
1045 | tree vdef = gimple_phi_result (gs: phi); |
1046 | tree vuse = gimple_phi_arg_def (gs: phi, index: 0); |
1047 | |
1048 | use_operand_p use_p; |
1049 | imm_use_iterator iter; |
1050 | gimple *use_stmt; |
1051 | FOR_EACH_IMM_USE_STMT (use_stmt, iter, vdef) |
1052 | FOR_EACH_IMM_USE_ON_STMT (use_p, iter) |
1053 | SET_USE (use_p, vuse); |
1054 | if (SSA_NAME_OCCURS_IN_ABNORMAL_PHI (vdef) |
1055 | && TREE_CODE (vuse) == SSA_NAME) |
1056 | SSA_NAME_OCCURS_IN_ABNORMAL_PHI (vuse) = 1; |
1057 | } |
1058 | else |
1059 | gimple_set_plf (stmt: phi, STMT_NECESSARY, val_p: true); |
1060 | } |
1061 | |
1062 | if (!gimple_plf (stmt: phi, STMT_NECESSARY)) |
1063 | { |
1064 | something_changed = true; |
1065 | if (dump_file && (dump_flags & TDF_DETAILS)) |
1066 | { |
1067 | fprintf (stream: dump_file, format: "Deleting : " ); |
1068 | print_gimple_stmt (dump_file, phi, 0, TDF_SLIM); |
1069 | fprintf (stream: dump_file, format: "\n" ); |
1070 | } |
1071 | |
1072 | remove_phi_node (&gsi, true); |
1073 | stats.removed_phis++; |
1074 | continue; |
1075 | } |
1076 | |
1077 | gsi_next (i: &gsi); |
1078 | } |
1079 | return something_changed; |
1080 | } |
1081 | |
1082 | |
1083 | /* Remove dead statement pointed to by iterator I. Receives the basic block BB |
1084 | containing I so that we don't have to look it up. */ |
1085 | |
1086 | static void |
1087 | remove_dead_stmt (gimple_stmt_iterator *i, basic_block bb, |
1088 | vec<edge> &to_remove_edges) |
1089 | { |
1090 | gimple *stmt = gsi_stmt (i: *i); |
1091 | |
1092 | if (dump_file && (dump_flags & TDF_DETAILS)) |
1093 | { |
1094 | fprintf (stream: dump_file, format: "Deleting : " ); |
1095 | print_gimple_stmt (dump_file, stmt, 0, TDF_SLIM); |
1096 | fprintf (stream: dump_file, format: "\n" ); |
1097 | } |
1098 | |
1099 | stats.removed++; |
1100 | |
1101 | /* If we have determined that a conditional branch statement contributes |
1102 | nothing to the program, then we not only remove it, but we need to update |
1103 | the CFG. We can chose any of edges out of BB as long as we are sure to not |
1104 | close infinite loops. This is done by always choosing the edge closer to |
1105 | exit in inverted_rev_post_order_compute order. */ |
1106 | if (is_ctrl_stmt (stmt)) |
1107 | { |
1108 | edge_iterator ei; |
1109 | edge e = NULL, e2; |
1110 | |
1111 | /* See if there is only one non-abnormal edge. */ |
1112 | if (single_succ_p (bb)) |
1113 | e = single_succ_edge (bb); |
1114 | /* Otherwise chose one that is closer to bb with live statement in it. |
1115 | To be able to chose one, we compute inverted post order starting from |
1116 | all BBs with live statements. */ |
1117 | if (!e) |
1118 | { |
1119 | if (!bb_postorder) |
1120 | { |
1121 | int *rpo = XNEWVEC (int, n_basic_blocks_for_fn (cfun)); |
1122 | int n = inverted_rev_post_order_compute (cfun, rpo, |
1123 | start_points: &bb_contains_live_stmts); |
1124 | bb_postorder = XNEWVEC (int, last_basic_block_for_fn (cfun)); |
1125 | for (int i = 0; i < n; ++i) |
1126 | bb_postorder[rpo[i]] = i; |
1127 | free (ptr: rpo); |
1128 | } |
1129 | FOR_EACH_EDGE (e2, ei, bb->succs) |
1130 | if (!e || e2->dest == EXIT_BLOCK_PTR_FOR_FN (cfun) |
1131 | || bb_postorder [e->dest->index] |
1132 | >= bb_postorder [e2->dest->index]) |
1133 | e = e2; |
1134 | } |
1135 | gcc_assert (e); |
1136 | e->probability = profile_probability::always (); |
1137 | |
1138 | /* The edge is no longer associated with a conditional, so it does |
1139 | not have TRUE/FALSE flags. |
1140 | We are also safe to drop EH/ABNORMAL flags and turn them into |
1141 | normal control flow, because we know that all the destinations (including |
1142 | those odd edges) are equivalent for program execution. */ |
1143 | e->flags &= ~(EDGE_TRUE_VALUE | EDGE_FALSE_VALUE | EDGE_EH | EDGE_ABNORMAL); |
1144 | |
1145 | /* The lone outgoing edge from BB will be a fallthru edge. */ |
1146 | e->flags |= EDGE_FALLTHRU; |
1147 | |
1148 | /* Remove the remaining outgoing edges. */ |
1149 | FOR_EACH_EDGE (e2, ei, bb->succs) |
1150 | if (e != e2) |
1151 | { |
1152 | /* If we made a BB unconditionally exit a loop or removed |
1153 | an entry into an irreducible region, then this transform |
1154 | alters the set of BBs in the loop. Schedule a fixup. */ |
1155 | if (loop_exit_edge_p (bb->loop_father, e) |
1156 | || (e2->dest->flags & BB_IRREDUCIBLE_LOOP)) |
1157 | loops_state_set (flags: LOOPS_NEED_FIXUP); |
1158 | to_remove_edges.safe_push (obj: e2); |
1159 | } |
1160 | } |
1161 | |
1162 | /* If this is a store into a variable that is being optimized away, |
1163 | add a debug bind stmt if possible. */ |
1164 | if (MAY_HAVE_DEBUG_BIND_STMTS |
1165 | && gimple_assign_single_p (gs: stmt) |
1166 | && is_gimple_val (gimple_assign_rhs1 (gs: stmt))) |
1167 | { |
1168 | tree lhs = gimple_assign_lhs (gs: stmt); |
1169 | if ((VAR_P (lhs) || TREE_CODE (lhs) == PARM_DECL) |
1170 | && !DECL_IGNORED_P (lhs) |
1171 | && is_gimple_reg_type (TREE_TYPE (lhs)) |
1172 | && !is_global_var (t: lhs) |
1173 | && !DECL_HAS_VALUE_EXPR_P (lhs)) |
1174 | { |
1175 | tree rhs = gimple_assign_rhs1 (gs: stmt); |
1176 | gdebug *note |
1177 | = gimple_build_debug_bind (lhs, unshare_expr (rhs), stmt); |
1178 | gsi_insert_after (i, note, GSI_SAME_STMT); |
1179 | } |
1180 | } |
1181 | |
1182 | unlink_stmt_vdef (stmt); |
1183 | gsi_remove (i, true); |
1184 | release_defs (stmt); |
1185 | } |
1186 | |
1187 | /* Helper for maybe_optimize_arith_overflow. Find in *TP if there are any |
1188 | uses of data (SSA_NAME) other than REALPART_EXPR referencing it. */ |
1189 | |
1190 | static tree |
1191 | find_non_realpart_uses (tree *tp, int *walk_subtrees, void *data) |
1192 | { |
1193 | if (TYPE_P (*tp) || TREE_CODE (*tp) == REALPART_EXPR) |
1194 | *walk_subtrees = 0; |
1195 | if (*tp == (tree) data) |
1196 | return *tp; |
1197 | return NULL_TREE; |
1198 | } |
1199 | |
1200 | /* If the IMAGPART_EXPR of the {ADD,SUB,MUL}_OVERFLOW result is never used, |
1201 | but REALPART_EXPR is, optimize the {ADD,SUB,MUL}_OVERFLOW internal calls |
1202 | into plain unsigned {PLUS,MINUS,MULT}_EXPR, and if needed reset debug |
1203 | uses. */ |
1204 | |
1205 | static void |
1206 | maybe_optimize_arith_overflow (gimple_stmt_iterator *gsi, |
1207 | enum tree_code subcode) |
1208 | { |
1209 | gimple *stmt = gsi_stmt (i: *gsi); |
1210 | tree lhs = gimple_call_lhs (gs: stmt); |
1211 | |
1212 | if (lhs == NULL || TREE_CODE (lhs) != SSA_NAME) |
1213 | return; |
1214 | |
1215 | imm_use_iterator imm_iter; |
1216 | use_operand_p use_p; |
1217 | bool has_debug_uses = false; |
1218 | bool has_realpart_uses = false; |
1219 | bool has_other_uses = false; |
1220 | FOR_EACH_IMM_USE_FAST (use_p, imm_iter, lhs) |
1221 | { |
1222 | gimple *use_stmt = USE_STMT (use_p); |
1223 | if (is_gimple_debug (gs: use_stmt)) |
1224 | has_debug_uses = true; |
1225 | else if (is_gimple_assign (gs: use_stmt) |
1226 | && gimple_assign_rhs_code (gs: use_stmt) == REALPART_EXPR |
1227 | && TREE_OPERAND (gimple_assign_rhs1 (use_stmt), 0) == lhs) |
1228 | has_realpart_uses = true; |
1229 | else |
1230 | { |
1231 | has_other_uses = true; |
1232 | break; |
1233 | } |
1234 | } |
1235 | |
1236 | if (!has_realpart_uses || has_other_uses) |
1237 | return; |
1238 | |
1239 | tree arg0 = gimple_call_arg (gs: stmt, index: 0); |
1240 | tree arg1 = gimple_call_arg (gs: stmt, index: 1); |
1241 | location_t loc = gimple_location (g: stmt); |
1242 | tree type = TREE_TYPE (TREE_TYPE (lhs)); |
1243 | tree utype = type; |
1244 | if (!TYPE_UNSIGNED (type)) |
1245 | utype = build_nonstandard_integer_type (TYPE_PRECISION (type), 1); |
1246 | tree result = fold_build2_loc (loc, subcode, utype, |
1247 | fold_convert_loc (loc, utype, arg0), |
1248 | fold_convert_loc (loc, utype, arg1)); |
1249 | result = fold_convert_loc (loc, type, result); |
1250 | |
1251 | if (has_debug_uses) |
1252 | { |
1253 | gimple *use_stmt; |
1254 | FOR_EACH_IMM_USE_STMT (use_stmt, imm_iter, lhs) |
1255 | { |
1256 | if (!gimple_debug_bind_p (s: use_stmt)) |
1257 | continue; |
1258 | tree v = gimple_debug_bind_get_value (dbg: use_stmt); |
1259 | if (walk_tree (&v, find_non_realpart_uses, lhs, NULL)) |
1260 | { |
1261 | gimple_debug_bind_reset_value (dbg: use_stmt); |
1262 | update_stmt (s: use_stmt); |
1263 | } |
1264 | } |
1265 | } |
1266 | |
1267 | if (TREE_CODE (result) == INTEGER_CST && TREE_OVERFLOW (result)) |
1268 | result = drop_tree_overflow (result); |
1269 | tree overflow = build_zero_cst (type); |
1270 | tree ctype = build_complex_type (type); |
1271 | if (TREE_CODE (result) == INTEGER_CST) |
1272 | result = build_complex (ctype, result, overflow); |
1273 | else |
1274 | result = build2_loc (loc: gimple_location (g: stmt), code: COMPLEX_EXPR, |
1275 | type: ctype, arg0: result, arg1: overflow); |
1276 | |
1277 | if (dump_file && (dump_flags & TDF_DETAILS)) |
1278 | { |
1279 | fprintf (stream: dump_file, format: "Transforming call: " ); |
1280 | print_gimple_stmt (dump_file, stmt, 0, TDF_SLIM); |
1281 | fprintf (stream: dump_file, format: "because the overflow result is never used into: " ); |
1282 | print_generic_stmt (dump_file, result, TDF_SLIM); |
1283 | fprintf (stream: dump_file, format: "\n" ); |
1284 | } |
1285 | |
1286 | gimplify_and_update_call_from_tree (gsi, result); |
1287 | } |
1288 | |
1289 | /* Returns whether the control parents of BB are preserved. */ |
1290 | |
1291 | static bool |
1292 | control_parents_preserved_p (basic_block bb) |
1293 | { |
1294 | /* If we marked the control parents from BB they are preserved. */ |
1295 | if (bitmap_bit_p (map: visited_control_parents, bitno: bb->index)) |
1296 | return true; |
1297 | |
1298 | /* But they can also end up being marked from elsewhere. */ |
1299 | bitmap_iterator bi; |
1300 | unsigned edge_number; |
1301 | EXECUTE_IF_SET_IN_BITMAP (cd->get_edges_dependent_on (bb->index), |
1302 | 0, edge_number, bi) |
1303 | { |
1304 | basic_block cd_bb = cd->get_edge_src (edge_number); |
1305 | if (cd_bb != bb |
1306 | && !bitmap_bit_p (map: last_stmt_necessary, bitno: cd_bb->index)) |
1307 | return false; |
1308 | } |
1309 | /* And cache the result. */ |
1310 | bitmap_set_bit (map: visited_control_parents, bitno: bb->index); |
1311 | return true; |
1312 | } |
1313 | |
1314 | /* Eliminate unnecessary statements. Any instruction not marked as necessary |
1315 | contributes nothing to the program, and can be deleted. */ |
1316 | |
1317 | static bool |
1318 | eliminate_unnecessary_stmts (bool aggressive) |
1319 | { |
1320 | bool something_changed = false; |
1321 | basic_block bb; |
1322 | gimple_stmt_iterator gsi, psi; |
1323 | gimple *stmt; |
1324 | tree call; |
1325 | auto_vec<edge> to_remove_edges; |
1326 | |
1327 | if (dump_file && (dump_flags & TDF_DETAILS)) |
1328 | fprintf (stream: dump_file, format: "\nEliminating unnecessary statements:\n" ); |
1329 | |
1330 | bool had_setjmp = cfun->calls_setjmp; |
1331 | clear_special_calls (); |
1332 | |
1333 | /* Walking basic blocks and statements in reverse order avoids |
1334 | releasing SSA names before any other DEFs that refer to them are |
1335 | released. This helps avoid loss of debug information, as we get |
1336 | a chance to propagate all RHSs of removed SSAs into debug uses, |
1337 | rather than only the latest ones. E.g., consider: |
1338 | |
1339 | x_3 = y_1 + z_2; |
1340 | a_5 = x_3 - b_4; |
1341 | # DEBUG a => a_5 |
1342 | |
1343 | If we were to release x_3 before a_5, when we reached a_5 and |
1344 | tried to substitute it into the debug stmt, we'd see x_3 there, |
1345 | but x_3's DEF, type, etc would have already been disconnected. |
1346 | By going backwards, the debug stmt first changes to: |
1347 | |
1348 | # DEBUG a => x_3 - b_4 |
1349 | |
1350 | and then to: |
1351 | |
1352 | # DEBUG a => y_1 + z_2 - b_4 |
1353 | |
1354 | as desired. */ |
1355 | gcc_assert (dom_info_available_p (CDI_DOMINATORS)); |
1356 | auto_vec<basic_block> h; |
1357 | h = get_all_dominated_blocks (CDI_DOMINATORS, |
1358 | single_succ (ENTRY_BLOCK_PTR_FOR_FN (cfun))); |
1359 | |
1360 | while (h.length ()) |
1361 | { |
1362 | bb = h.pop (); |
1363 | |
1364 | /* Remove dead statements. */ |
1365 | auto_bitmap debug_seen; |
1366 | for (gsi = gsi_last_bb (bb); !gsi_end_p (i: gsi); gsi = psi) |
1367 | { |
1368 | stmt = gsi_stmt (i: gsi); |
1369 | |
1370 | psi = gsi; |
1371 | gsi_prev (i: &psi); |
1372 | |
1373 | stats.total++; |
1374 | |
1375 | /* We can mark a call to free as not necessary if the |
1376 | defining statement of its argument is not necessary |
1377 | (and thus is getting removed). */ |
1378 | if (gimple_plf (stmt, STMT_NECESSARY) |
1379 | && (gimple_call_builtin_p (stmt, BUILT_IN_FREE) |
1380 | || (is_gimple_call (gs: stmt) |
1381 | && gimple_call_from_new_or_delete (s: as_a <gcall *> (p: stmt)) |
1382 | && gimple_call_operator_delete_p (as_a <gcall *> (p: stmt))))) |
1383 | { |
1384 | tree ptr = gimple_call_arg (gs: stmt, index: 0); |
1385 | if (TREE_CODE (ptr) == SSA_NAME) |
1386 | { |
1387 | gimple *def_stmt = SSA_NAME_DEF_STMT (ptr); |
1388 | if (!gimple_nop_p (g: def_stmt) |
1389 | && !gimple_plf (stmt: def_stmt, STMT_NECESSARY)) |
1390 | gimple_set_plf (stmt, STMT_NECESSARY, val_p: false); |
1391 | } |
1392 | } |
1393 | |
1394 | /* If GSI is not necessary then remove it. */ |
1395 | if (!gimple_plf (stmt, STMT_NECESSARY)) |
1396 | { |
1397 | /* Keep clobbers that we can keep live live. */ |
1398 | if (gimple_clobber_p (s: stmt)) |
1399 | { |
1400 | ssa_op_iter iter; |
1401 | use_operand_p use_p; |
1402 | bool dead = false; |
1403 | FOR_EACH_SSA_USE_OPERAND (use_p, stmt, iter, SSA_OP_USE) |
1404 | { |
1405 | tree name = USE_FROM_PTR (use_p); |
1406 | if (!SSA_NAME_IS_DEFAULT_DEF (name) |
1407 | && !bitmap_bit_p (map: processed, SSA_NAME_VERSION (name))) |
1408 | { |
1409 | dead = true; |
1410 | break; |
1411 | } |
1412 | } |
1413 | if (!dead |
1414 | /* When doing CD-DCE we have to ensure all controls |
1415 | of the stmt are still live. */ |
1416 | && (!aggressive || control_parents_preserved_p (bb))) |
1417 | { |
1418 | bitmap_clear (debug_seen); |
1419 | continue; |
1420 | } |
1421 | } |
1422 | if (!is_gimple_debug (gs: stmt)) |
1423 | something_changed = true; |
1424 | remove_dead_stmt (i: &gsi, bb, to_remove_edges); |
1425 | continue; |
1426 | } |
1427 | else if (is_gimple_call (gs: stmt)) |
1428 | { |
1429 | tree name = gimple_call_lhs (gs: stmt); |
1430 | |
1431 | notice_special_calls (as_a <gcall *> (p: stmt)); |
1432 | |
1433 | /* When LHS of var = call (); is dead, simplify it into |
1434 | call (); saving one operand. */ |
1435 | if (name |
1436 | && TREE_CODE (name) == SSA_NAME |
1437 | && !bitmap_bit_p (map: processed, SSA_NAME_VERSION (name)) |
1438 | /* Avoid doing so for allocation calls which we |
1439 | did not mark as necessary, it will confuse the |
1440 | special logic we apply to malloc/free pair removal. */ |
1441 | && (!(call = gimple_call_fndecl (gs: stmt)) |
1442 | || ((DECL_BUILT_IN_CLASS (call) != BUILT_IN_NORMAL |
1443 | || (DECL_FUNCTION_CODE (decl: call) != BUILT_IN_ALIGNED_ALLOC |
1444 | && DECL_FUNCTION_CODE (decl: call) != BUILT_IN_MALLOC |
1445 | && DECL_FUNCTION_CODE (decl: call) != BUILT_IN_CALLOC |
1446 | && !ALLOCA_FUNCTION_CODE_P |
1447 | (DECL_FUNCTION_CODE (call)))) |
1448 | && !DECL_IS_REPLACEABLE_OPERATOR_NEW_P (call)))) |
1449 | { |
1450 | something_changed = true; |
1451 | if (dump_file && (dump_flags & TDF_DETAILS)) |
1452 | { |
1453 | fprintf (stream: dump_file, format: "Deleting LHS of call: " ); |
1454 | print_gimple_stmt (dump_file, stmt, 0, TDF_SLIM); |
1455 | fprintf (stream: dump_file, format: "\n" ); |
1456 | } |
1457 | |
1458 | gimple_call_set_lhs (gs: stmt, NULL_TREE); |
1459 | maybe_clean_or_replace_eh_stmt (stmt, stmt); |
1460 | update_stmt (s: stmt); |
1461 | release_ssa_name (name); |
1462 | |
1463 | /* GOMP_SIMD_LANE (unless three argument) or ASAN_POISON |
1464 | without lhs is not needed. */ |
1465 | if (gimple_call_internal_p (gs: stmt)) |
1466 | switch (gimple_call_internal_fn (gs: stmt)) |
1467 | { |
1468 | case IFN_GOMP_SIMD_LANE: |
1469 | if (gimple_call_num_args (gs: stmt) >= 3 |
1470 | && !integer_nonzerop (gimple_call_arg (gs: stmt, index: 2))) |
1471 | break; |
1472 | /* FALLTHRU */ |
1473 | case IFN_ASAN_POISON: |
1474 | remove_dead_stmt (i: &gsi, bb, to_remove_edges); |
1475 | break; |
1476 | default: |
1477 | break; |
1478 | } |
1479 | } |
1480 | else if (gimple_call_internal_p (gs: stmt)) |
1481 | switch (gimple_call_internal_fn (gs: stmt)) |
1482 | { |
1483 | case IFN_ADD_OVERFLOW: |
1484 | maybe_optimize_arith_overflow (gsi: &gsi, subcode: PLUS_EXPR); |
1485 | break; |
1486 | case IFN_SUB_OVERFLOW: |
1487 | maybe_optimize_arith_overflow (gsi: &gsi, subcode: MINUS_EXPR); |
1488 | break; |
1489 | case IFN_MUL_OVERFLOW: |
1490 | maybe_optimize_arith_overflow (gsi: &gsi, subcode: MULT_EXPR); |
1491 | break; |
1492 | case IFN_UADDC: |
1493 | if (integer_zerop (gimple_call_arg (gs: stmt, index: 2))) |
1494 | maybe_optimize_arith_overflow (gsi: &gsi, subcode: PLUS_EXPR); |
1495 | break; |
1496 | case IFN_USUBC: |
1497 | if (integer_zerop (gimple_call_arg (gs: stmt, index: 2))) |
1498 | maybe_optimize_arith_overflow (gsi: &gsi, subcode: MINUS_EXPR); |
1499 | break; |
1500 | default: |
1501 | break; |
1502 | } |
1503 | } |
1504 | else if (gimple_debug_bind_p (s: stmt)) |
1505 | { |
1506 | /* We are only keeping the last debug-bind of a |
1507 | non-DEBUG_EXPR_DECL variable in a series of |
1508 | debug-bind stmts. */ |
1509 | tree var = gimple_debug_bind_get_var (dbg: stmt); |
1510 | if (TREE_CODE (var) != DEBUG_EXPR_DECL |
1511 | && !bitmap_set_bit (debug_seen, DECL_UID (var))) |
1512 | remove_dead_stmt (i: &gsi, bb, to_remove_edges); |
1513 | continue; |
1514 | } |
1515 | bitmap_clear (debug_seen); |
1516 | } |
1517 | |
1518 | /* Remove dead PHI nodes. */ |
1519 | something_changed |= remove_dead_phis (bb); |
1520 | } |
1521 | |
1522 | /* First remove queued edges. */ |
1523 | if (!to_remove_edges.is_empty ()) |
1524 | { |
1525 | /* Remove edges. We've delayed this to not get bogus debug stmts |
1526 | during PHI node removal. */ |
1527 | for (unsigned i = 0; i < to_remove_edges.length (); ++i) |
1528 | remove_edge (to_remove_edges[i]); |
1529 | cfg_altered = true; |
1530 | } |
1531 | /* When we cleared calls_setjmp we can purge all abnormal edges. Do so. |
1532 | ??? We'd like to assert that setjmp calls do not pop out of nothing |
1533 | but we currently lack a per-stmt way of noting whether a call was |
1534 | recognized as returns-twice (or rather receives-control). */ |
1535 | if (!cfun->calls_setjmp && had_setjmp) |
1536 | { |
1537 | /* Make sure we only remove the edges, not dominated blocks. Using |
1538 | gimple_purge_dead_abnormal_call_edges would do that and we |
1539 | cannot free dominators yet. */ |
1540 | FOR_EACH_BB_FN (bb, cfun) |
1541 | if (gcall *stmt = safe_dyn_cast <gcall *> (p: *gsi_last_bb (bb))) |
1542 | if (!stmt_can_make_abnormal_goto (stmt)) |
1543 | { |
1544 | edge_iterator ei; |
1545 | edge e; |
1546 | for (ei = ei_start (bb->succs); (e = ei_safe_edge (i: ei)); ) |
1547 | { |
1548 | if (e->flags & EDGE_ABNORMAL) |
1549 | { |
1550 | if (e->flags & EDGE_FALLTHRU) |
1551 | e->flags &= ~EDGE_ABNORMAL; |
1552 | else |
1553 | remove_edge (e); |
1554 | cfg_altered = true; |
1555 | } |
1556 | else |
1557 | ei_next (i: &ei); |
1558 | } |
1559 | } |
1560 | } |
1561 | |
1562 | /* Now remove the unreachable blocks. */ |
1563 | if (cfg_altered) |
1564 | { |
1565 | basic_block prev_bb; |
1566 | |
1567 | find_unreachable_blocks (); |
1568 | |
1569 | /* Delete all unreachable basic blocks in reverse dominator order. */ |
1570 | for (bb = EXIT_BLOCK_PTR_FOR_FN (cfun)->prev_bb; |
1571 | bb != ENTRY_BLOCK_PTR_FOR_FN (cfun); bb = prev_bb) |
1572 | { |
1573 | prev_bb = bb->prev_bb; |
1574 | |
1575 | if ((bb_contains_live_stmts |
1576 | && !bitmap_bit_p (map: bb_contains_live_stmts, bitno: bb->index)) |
1577 | || !(bb->flags & BB_REACHABLE)) |
1578 | { |
1579 | /* Since we don't track liveness of virtual PHI nodes, it is |
1580 | possible that we rendered some PHI nodes unreachable while |
1581 | they are still in use. Mark them for renaming. */ |
1582 | for (gphi_iterator gsi = gsi_start_phis (bb); !gsi_end_p (i: gsi); |
1583 | gsi_next (i: &gsi)) |
1584 | if (virtual_operand_p (op: gimple_phi_result (gs: gsi.phi ()))) |
1585 | { |
1586 | bool found = false; |
1587 | imm_use_iterator iter; |
1588 | |
1589 | FOR_EACH_IMM_USE_STMT (stmt, iter, |
1590 | gimple_phi_result (gsi.phi ())) |
1591 | { |
1592 | if (!(gimple_bb (g: stmt)->flags & BB_REACHABLE)) |
1593 | continue; |
1594 | if (gimple_code (g: stmt) == GIMPLE_PHI |
1595 | || gimple_plf (stmt, STMT_NECESSARY)) |
1596 | { |
1597 | found = true; |
1598 | break; |
1599 | } |
1600 | } |
1601 | if (found) |
1602 | mark_virtual_phi_result_for_renaming (gsi.phi ()); |
1603 | } |
1604 | |
1605 | if (!(bb->flags & BB_REACHABLE)) |
1606 | { |
1607 | /* Speed up the removal of blocks that don't |
1608 | dominate others. Walking backwards, this should |
1609 | be the common case. ??? Do we need to recompute |
1610 | dominators because of cfg_altered? */ |
1611 | if (!first_dom_son (CDI_DOMINATORS, bb)) |
1612 | delete_basic_block (bb); |
1613 | else |
1614 | { |
1615 | h = get_all_dominated_blocks (CDI_DOMINATORS, bb); |
1616 | |
1617 | while (h.length ()) |
1618 | { |
1619 | bb = h.pop (); |
1620 | prev_bb = bb->prev_bb; |
1621 | /* Rearrangements to the CFG may have failed |
1622 | to update the dominators tree, so that |
1623 | formerly-dominated blocks are now |
1624 | otherwise reachable. */ |
1625 | if (!!(bb->flags & BB_REACHABLE)) |
1626 | continue; |
1627 | delete_basic_block (bb); |
1628 | } |
1629 | |
1630 | h.release (); |
1631 | } |
1632 | } |
1633 | } |
1634 | } |
1635 | } |
1636 | |
1637 | if (bb_postorder) |
1638 | free (ptr: bb_postorder); |
1639 | bb_postorder = NULL; |
1640 | |
1641 | return something_changed; |
1642 | } |
1643 | |
1644 | |
1645 | /* Print out removed statement statistics. */ |
1646 | |
1647 | static void |
1648 | print_stats (void) |
1649 | { |
1650 | float percg; |
1651 | |
1652 | percg = ((float) stats.removed / (float) stats.total) * 100; |
1653 | fprintf (stream: dump_file, format: "Removed %d of %d statements (%d%%)\n" , |
1654 | stats.removed, stats.total, (int) percg); |
1655 | |
1656 | if (stats.total_phis == 0) |
1657 | percg = 0; |
1658 | else |
1659 | percg = ((float) stats.removed_phis / (float) stats.total_phis) * 100; |
1660 | |
1661 | fprintf (stream: dump_file, format: "Removed %d of %d PHI nodes (%d%%)\n" , |
1662 | stats.removed_phis, stats.total_phis, (int) percg); |
1663 | } |
1664 | |
1665 | /* Initialization for this pass. Set up the used data structures. */ |
1666 | |
1667 | static void |
1668 | tree_dce_init (bool aggressive) |
1669 | { |
1670 | memset (s: (void *) &stats, c: 0, n: sizeof (stats)); |
1671 | |
1672 | if (aggressive) |
1673 | { |
1674 | last_stmt_necessary = sbitmap_alloc (last_basic_block_for_fn (cfun)); |
1675 | bitmap_clear (last_stmt_necessary); |
1676 | bb_contains_live_stmts = sbitmap_alloc (last_basic_block_for_fn (cfun)); |
1677 | bitmap_clear (bb_contains_live_stmts); |
1678 | } |
1679 | |
1680 | processed = sbitmap_alloc (num_ssa_names + 1); |
1681 | bitmap_clear (processed); |
1682 | |
1683 | worklist.create (nelems: 64); |
1684 | cfg_altered = false; |
1685 | } |
1686 | |
1687 | /* Cleanup after this pass. */ |
1688 | |
1689 | static void |
1690 | tree_dce_done (bool aggressive) |
1691 | { |
1692 | if (aggressive) |
1693 | { |
1694 | delete cd; |
1695 | sbitmap_free (map: visited_control_parents); |
1696 | sbitmap_free (map: last_stmt_necessary); |
1697 | sbitmap_free (map: bb_contains_live_stmts); |
1698 | bb_contains_live_stmts = NULL; |
1699 | } |
1700 | |
1701 | sbitmap_free (map: processed); |
1702 | |
1703 | worklist.release (); |
1704 | } |
1705 | |
1706 | /* Sort PHI argument values for make_forwarders_with_degenerate_phis. */ |
1707 | |
1708 | static int |
1709 | sort_phi_args (const void *a_, const void *b_) |
1710 | { |
1711 | auto *a = (const std::pair<edge, hashval_t> *) a_; |
1712 | auto *b = (const std::pair<edge, hashval_t> *) b_; |
1713 | hashval_t ha = a->second; |
1714 | hashval_t hb = b->second; |
1715 | if (ha < hb) |
1716 | return -1; |
1717 | else if (ha > hb) |
1718 | return 1; |
1719 | else if (a->first->dest_idx < b->first->dest_idx) |
1720 | return -1; |
1721 | else if (a->first->dest_idx > b->first->dest_idx) |
1722 | return 1; |
1723 | else |
1724 | return 0; |
1725 | } |
1726 | |
1727 | /* Look for a non-virtual PHIs and make a forwarder block when all PHIs |
1728 | have the same argument on a set of edges. This is to not consider |
1729 | control dependences of individual edges for same values but only for |
1730 | the common set. */ |
1731 | |
1732 | static unsigned |
1733 | make_forwarders_with_degenerate_phis (function *fn) |
1734 | { |
1735 | unsigned todo = 0; |
1736 | |
1737 | basic_block bb; |
1738 | FOR_EACH_BB_FN (bb, fn) |
1739 | { |
1740 | /* Only PHIs with three or more arguments have opportunities. */ |
1741 | if (EDGE_COUNT (bb->preds) < 3) |
1742 | continue; |
1743 | /* Do not touch loop headers or blocks with abnormal predecessors. |
1744 | ??? This is to avoid creating valid loops here, see PR103458. |
1745 | We might want to improve things to either explicitely add those |
1746 | loops or at least consider blocks with no backedges. */ |
1747 | if (bb->loop_father->header == bb |
1748 | || bb_has_abnormal_pred (bb)) |
1749 | continue; |
1750 | |
1751 | /* Take one PHI node as template to look for identical |
1752 | arguments. Build a vector of candidates forming sets |
1753 | of argument edges with equal values. Note optimality |
1754 | depends on the particular choice of the template PHI |
1755 | since equal arguments are unordered leaving other PHIs |
1756 | with more than one set of equal arguments within this |
1757 | argument range unsorted. We'd have to break ties by |
1758 | looking at other PHI nodes. */ |
1759 | gphi_iterator gsi = gsi_start_nonvirtual_phis (bb); |
1760 | if (gsi_end_p (i: gsi)) |
1761 | continue; |
1762 | gphi *phi = gsi.phi (); |
1763 | auto_vec<std::pair<edge, hashval_t>, 8> args; |
1764 | bool need_resort = false; |
1765 | for (unsigned i = 0; i < gimple_phi_num_args (gs: phi); ++i) |
1766 | { |
1767 | edge e = gimple_phi_arg_edge (phi, i); |
1768 | /* Skip abnormal edges since we cannot redirect them. */ |
1769 | if (e->flags & EDGE_ABNORMAL) |
1770 | continue; |
1771 | /* Skip loop exit edges when we are in loop-closed SSA form |
1772 | since the forwarder we'd create does not have a PHI node. */ |
1773 | if (loops_state_satisfies_p (flags: LOOP_CLOSED_SSA) |
1774 | && loop_exit_edge_p (e->src->loop_father, e)) |
1775 | continue; |
1776 | |
1777 | tree arg = gimple_phi_arg_def (gs: phi, index: i); |
1778 | if (!CONSTANT_CLASS_P (arg) && TREE_CODE (arg) != SSA_NAME) |
1779 | need_resort = true; |
1780 | args.safe_push (obj: std::make_pair (x&: e, y: iterative_hash_expr (tree: arg, seed: 0))); |
1781 | } |
1782 | if (args.length () < 2) |
1783 | continue; |
1784 | args.qsort (sort_phi_args); |
1785 | /* The above sorting can be different between -g and -g0, as e.g. decls |
1786 | can have different uids (-g could have bigger gaps in between them). |
1787 | So, only use that to determine which args are equal, then change |
1788 | second from hash value to smallest dest_idx of the edges which have |
1789 | equal argument and sort again. If all the phi arguments are |
1790 | constants or SSA_NAME, there is no need for the second sort, the hash |
1791 | values are stable in that case. */ |
1792 | hashval_t hash = args[0].second; |
1793 | args[0].second = args[0].first->dest_idx; |
1794 | bool any_equal = false; |
1795 | for (unsigned i = 1; i < args.length (); ++i) |
1796 | if (hash == args[i].second |
1797 | && operand_equal_p (PHI_ARG_DEF_FROM_EDGE (phi, args[i - 1].first), |
1798 | PHI_ARG_DEF_FROM_EDGE (phi, args[i].first))) |
1799 | { |
1800 | args[i].second = args[i - 1].second; |
1801 | any_equal = true; |
1802 | } |
1803 | else |
1804 | { |
1805 | hash = args[i].second; |
1806 | args[i].second = args[i].first->dest_idx; |
1807 | } |
1808 | if (!any_equal) |
1809 | continue; |
1810 | if (need_resort) |
1811 | args.qsort (sort_phi_args); |
1812 | |
1813 | /* From the candidates vector now verify true candidates for |
1814 | forwarders and create them. */ |
1815 | gphi *vphi = get_virtual_phi (bb); |
1816 | unsigned start = 0; |
1817 | while (start < args.length () - 1) |
1818 | { |
1819 | unsigned i; |
1820 | for (i = start + 1; i < args.length (); ++i) |
1821 | if (args[start].second != args[i].second) |
1822 | break; |
1823 | /* args[start]..args[i-1] are equal. */ |
1824 | if (start != i - 1) |
1825 | { |
1826 | /* Check all PHI nodes for argument equality. */ |
1827 | bool equal = true; |
1828 | gphi_iterator gsi2 = gsi; |
1829 | gsi_next (i: &gsi2); |
1830 | for (; !gsi_end_p (i: gsi2); gsi_next (i: &gsi2)) |
1831 | { |
1832 | gphi *phi2 = gsi2.phi (); |
1833 | if (virtual_operand_p (op: gimple_phi_result (gs: phi2))) |
1834 | continue; |
1835 | tree start_arg |
1836 | = PHI_ARG_DEF_FROM_EDGE (phi2, args[start].first); |
1837 | for (unsigned j = start + 1; j < i; ++j) |
1838 | { |
1839 | if (!operand_equal_p (start_arg, |
1840 | PHI_ARG_DEF_FROM_EDGE |
1841 | (phi2, args[j].first))) |
1842 | { |
1843 | /* Another PHI might have a shorter set of |
1844 | equivalent args. Go for that. */ |
1845 | i = j; |
1846 | if (j == start + 1) |
1847 | equal = false; |
1848 | break; |
1849 | } |
1850 | } |
1851 | if (!equal) |
1852 | break; |
1853 | } |
1854 | if (equal) |
1855 | { |
1856 | /* If we are asked to forward all edges the block |
1857 | has all degenerate PHIs. Do nothing in that case. */ |
1858 | if (start == 0 |
1859 | && i == args.length () |
1860 | && args.length () == gimple_phi_num_args (gs: phi)) |
1861 | break; |
1862 | /* Instead of using make_forwarder_block we are |
1863 | rolling our own variant knowing that the forwarder |
1864 | does not need PHI nodes apart from eventually |
1865 | a virtual one. */ |
1866 | auto_vec<tree, 8> vphi_args; |
1867 | if (vphi) |
1868 | { |
1869 | vphi_args.reserve_exact (nelems: i - start); |
1870 | for (unsigned j = start; j < i; ++j) |
1871 | vphi_args.quick_push |
1872 | (PHI_ARG_DEF_FROM_EDGE (vphi, args[j].first)); |
1873 | } |
1874 | free_dominance_info (fn, CDI_DOMINATORS); |
1875 | basic_block forwarder = split_edge (args[start].first); |
1876 | profile_count count = profile_count::zero (); |
1877 | for (unsigned j = start + 1; j < i; ++j) |
1878 | { |
1879 | edge e = args[j].first; |
1880 | redirect_edge_and_branch_force (e, forwarder); |
1881 | redirect_edge_var_map_clear (e); |
1882 | count += e->count (); |
1883 | } |
1884 | forwarder->count = count; |
1885 | if (vphi) |
1886 | { |
1887 | tree def = copy_ssa_name (var: vphi_args[0]); |
1888 | gphi *vphi_copy = create_phi_node (def, forwarder); |
1889 | for (unsigned j = start; j < i; ++j) |
1890 | add_phi_arg (vphi_copy, vphi_args[j - start], |
1891 | args[j].first, UNKNOWN_LOCATION); |
1892 | SET_PHI_ARG_DEF |
1893 | (vphi, single_succ_edge (forwarder)->dest_idx, def); |
1894 | } |
1895 | todo |= TODO_cleanup_cfg; |
1896 | } |
1897 | } |
1898 | /* Continue searching for more opportunities. */ |
1899 | start = i; |
1900 | } |
1901 | } |
1902 | return todo; |
1903 | } |
1904 | |
1905 | /* Main routine to eliminate dead code. |
1906 | |
1907 | AGGRESSIVE controls the aggressiveness of the algorithm. |
1908 | In conservative mode, we ignore control dependence and simply declare |
1909 | all but the most trivially dead branches necessary. This mode is fast. |
1910 | In aggressive mode, control dependences are taken into account, which |
1911 | results in more dead code elimination, but at the cost of some time. |
1912 | |
1913 | FIXME: Aggressive mode before PRE doesn't work currently because |
1914 | the dominance info is not invalidated after DCE1. This is |
1915 | not an issue right now because we only run aggressive DCE |
1916 | as the last tree SSA pass, but keep this in mind when you |
1917 | start experimenting with pass ordering. */ |
1918 | |
1919 | static unsigned int |
1920 | perform_tree_ssa_dce (bool aggressive) |
1921 | { |
1922 | bool something_changed = 0; |
1923 | unsigned todo = 0; |
1924 | |
1925 | /* Preheaders are needed for SCEV to work. |
1926 | Simple lateches and recorded exits improve chances that loop will |
1927 | proved to be finite in testcases such as in loop-15.c and loop-24.c */ |
1928 | bool in_loop_pipeline = scev_initialized_p (); |
1929 | if (aggressive && ! in_loop_pipeline) |
1930 | { |
1931 | loop_optimizer_init (LOOPS_NORMAL |
1932 | | LOOPS_HAVE_RECORDED_EXITS); |
1933 | scev_initialize (); |
1934 | } |
1935 | |
1936 | if (aggressive) |
1937 | todo |= make_forwarders_with_degenerate_phis (cfun); |
1938 | |
1939 | calculate_dominance_info (CDI_DOMINATORS); |
1940 | |
1941 | tree_dce_init (aggressive); |
1942 | |
1943 | if (aggressive) |
1944 | { |
1945 | /* Compute control dependence. */ |
1946 | calculate_dominance_info (CDI_POST_DOMINATORS); |
1947 | cd = new control_dependences (); |
1948 | |
1949 | visited_control_parents = |
1950 | sbitmap_alloc (last_basic_block_for_fn (cfun)); |
1951 | bitmap_clear (visited_control_parents); |
1952 | |
1953 | mark_dfs_back_edges (); |
1954 | } |
1955 | |
1956 | find_obviously_necessary_stmts (aggressive); |
1957 | |
1958 | if (aggressive && ! in_loop_pipeline) |
1959 | { |
1960 | scev_finalize (); |
1961 | loop_optimizer_finalize (); |
1962 | } |
1963 | |
1964 | longest_chain = 0; |
1965 | total_chain = 0; |
1966 | nr_walks = 0; |
1967 | chain_ovfl = false; |
1968 | visited = BITMAP_ALLOC (NULL); |
1969 | propagate_necessity (aggressive); |
1970 | BITMAP_FREE (visited); |
1971 | |
1972 | something_changed |= eliminate_unnecessary_stmts (aggressive); |
1973 | something_changed |= cfg_altered; |
1974 | |
1975 | /* We do not update postdominators, so free them unconditionally. */ |
1976 | free_dominance_info (CDI_POST_DOMINATORS); |
1977 | |
1978 | /* If we removed paths in the CFG, then we need to update |
1979 | dominators as well. I haven't investigated the possibility |
1980 | of incrementally updating dominators. */ |
1981 | if (cfg_altered) |
1982 | free_dominance_info (CDI_DOMINATORS); |
1983 | |
1984 | statistics_counter_event (cfun, "Statements deleted" , stats.removed); |
1985 | statistics_counter_event (cfun, "PHI nodes deleted" , stats.removed_phis); |
1986 | |
1987 | /* Debugging dumps. */ |
1988 | if (dump_file && (dump_flags & (TDF_STATS|TDF_DETAILS))) |
1989 | print_stats (); |
1990 | |
1991 | tree_dce_done (aggressive); |
1992 | |
1993 | if (something_changed) |
1994 | { |
1995 | free_numbers_of_iterations_estimates (cfun); |
1996 | if (in_loop_pipeline) |
1997 | scev_reset (); |
1998 | todo |= TODO_update_ssa | TODO_cleanup_cfg; |
1999 | } |
2000 | return todo; |
2001 | } |
2002 | |
2003 | /* Pass entry points. */ |
2004 | static unsigned int |
2005 | tree_ssa_dce (void) |
2006 | { |
2007 | return perform_tree_ssa_dce (/*aggressive=*/false); |
2008 | } |
2009 | |
2010 | static unsigned int |
2011 | tree_ssa_cd_dce (void) |
2012 | { |
2013 | return perform_tree_ssa_dce (/*aggressive=*/optimize >= 2); |
2014 | } |
2015 | |
2016 | namespace { |
2017 | |
2018 | const pass_data pass_data_dce = |
2019 | { |
2020 | .type: GIMPLE_PASS, /* type */ |
2021 | .name: "dce" , /* name */ |
2022 | .optinfo_flags: OPTGROUP_NONE, /* optinfo_flags */ |
2023 | .tv_id: TV_TREE_DCE, /* tv_id */ |
2024 | .properties_required: ( PROP_cfg | PROP_ssa ), /* properties_required */ |
2025 | .properties_provided: 0, /* properties_provided */ |
2026 | .properties_destroyed: 0, /* properties_destroyed */ |
2027 | .todo_flags_start: 0, /* todo_flags_start */ |
2028 | .todo_flags_finish: 0, /* todo_flags_finish */ |
2029 | }; |
2030 | |
2031 | class pass_dce : public gimple_opt_pass |
2032 | { |
2033 | public: |
2034 | pass_dce (gcc::context *ctxt) |
2035 | : gimple_opt_pass (pass_data_dce, ctxt), update_address_taken_p (false) |
2036 | {} |
2037 | |
2038 | /* opt_pass methods: */ |
2039 | opt_pass * clone () final override { return new pass_dce (m_ctxt); } |
2040 | void set_pass_param (unsigned n, bool param) final override |
2041 | { |
2042 | gcc_assert (n == 0); |
2043 | update_address_taken_p = param; |
2044 | } |
2045 | bool gate (function *) final override { return flag_tree_dce != 0; } |
2046 | unsigned int execute (function *) final override |
2047 | { |
2048 | return (tree_ssa_dce () |
2049 | | (update_address_taken_p ? TODO_update_address_taken : 0)); |
2050 | } |
2051 | |
2052 | private: |
2053 | bool update_address_taken_p; |
2054 | }; // class pass_dce |
2055 | |
2056 | } // anon namespace |
2057 | |
2058 | gimple_opt_pass * |
2059 | make_pass_dce (gcc::context *ctxt) |
2060 | { |
2061 | return new pass_dce (ctxt); |
2062 | } |
2063 | |
2064 | namespace { |
2065 | |
2066 | const pass_data pass_data_cd_dce = |
2067 | { |
2068 | .type: GIMPLE_PASS, /* type */ |
2069 | .name: "cddce" , /* name */ |
2070 | .optinfo_flags: OPTGROUP_NONE, /* optinfo_flags */ |
2071 | .tv_id: TV_TREE_CD_DCE, /* tv_id */ |
2072 | .properties_required: ( PROP_cfg | PROP_ssa ), /* properties_required */ |
2073 | .properties_provided: 0, /* properties_provided */ |
2074 | .properties_destroyed: 0, /* properties_destroyed */ |
2075 | .todo_flags_start: 0, /* todo_flags_start */ |
2076 | .todo_flags_finish: 0, /* todo_flags_finish */ |
2077 | }; |
2078 | |
2079 | class pass_cd_dce : public gimple_opt_pass |
2080 | { |
2081 | public: |
2082 | pass_cd_dce (gcc::context *ctxt) |
2083 | : gimple_opt_pass (pass_data_cd_dce, ctxt), update_address_taken_p (false) |
2084 | {} |
2085 | |
2086 | /* opt_pass methods: */ |
2087 | opt_pass * clone () final override { return new pass_cd_dce (m_ctxt); } |
2088 | void set_pass_param (unsigned n, bool param) final override |
2089 | { |
2090 | gcc_assert (n == 0); |
2091 | update_address_taken_p = param; |
2092 | } |
2093 | bool gate (function *) final override { return flag_tree_dce != 0; } |
2094 | unsigned int execute (function *) final override |
2095 | { |
2096 | return (tree_ssa_cd_dce () |
2097 | | (update_address_taken_p ? TODO_update_address_taken : 0)); |
2098 | } |
2099 | |
2100 | private: |
2101 | bool update_address_taken_p; |
2102 | }; // class pass_cd_dce |
2103 | |
2104 | } // anon namespace |
2105 | |
2106 | gimple_opt_pass * |
2107 | make_pass_cd_dce (gcc::context *ctxt) |
2108 | { |
2109 | return new pass_cd_dce (ctxt); |
2110 | } |
2111 | |
2112 | |
2113 | /* A cheap DCE interface. WORKLIST is a list of possibly dead stmts and |
2114 | is consumed by this function. The function has linear complexity in |
2115 | the number of dead stmts with a constant factor like the average SSA |
2116 | use operands number. */ |
2117 | |
2118 | void |
2119 | simple_dce_from_worklist (bitmap worklist, bitmap need_eh_cleanup) |
2120 | { |
2121 | int phiremoved = 0; |
2122 | int stmtremoved = 0; |
2123 | while (! bitmap_empty_p (map: worklist)) |
2124 | { |
2125 | /* Pop item. */ |
2126 | unsigned i = bitmap_clear_first_set_bit (worklist); |
2127 | |
2128 | tree def = ssa_name (i); |
2129 | /* Removed by somebody else or still in use. |
2130 | Note use in itself for a phi node is not counted as still in use. */ |
2131 | if (!def) |
2132 | continue; |
2133 | if (!has_zero_uses (var: def)) |
2134 | { |
2135 | gimple *def_stmt = SSA_NAME_DEF_STMT (def); |
2136 | |
2137 | if (gimple_code (g: def_stmt) != GIMPLE_PHI) |
2138 | continue; |
2139 | |
2140 | gimple *use_stmt; |
2141 | imm_use_iterator use_iter; |
2142 | bool canremove = true; |
2143 | |
2144 | FOR_EACH_IMM_USE_STMT (use_stmt, use_iter, def) |
2145 | { |
2146 | /* Ignore debug statements. */ |
2147 | if (is_gimple_debug (gs: use_stmt)) |
2148 | continue; |
2149 | if (use_stmt != def_stmt) |
2150 | { |
2151 | canremove = false; |
2152 | break; |
2153 | } |
2154 | } |
2155 | if (!canremove) |
2156 | continue; |
2157 | } |
2158 | |
2159 | gimple *t = SSA_NAME_DEF_STMT (def); |
2160 | if (gimple_has_side_effects (t)) |
2161 | continue; |
2162 | |
2163 | /* The defining statement needs to be defining only this name. |
2164 | ASM is the only statement that can define more than one |
2165 | name. */ |
2166 | if (is_a<gasm *>(p: t) |
2167 | && !single_ssa_def_operand (stmt: t, SSA_OP_ALL_DEFS)) |
2168 | continue; |
2169 | |
2170 | /* Don't remove statements that are needed for non-call |
2171 | eh to work. */ |
2172 | if (stmt_unremovable_because_of_non_call_eh_p (cfun, t)) |
2173 | continue; |
2174 | |
2175 | /* Tell the caller that we removed a statement that might |
2176 | throw so it could cleanup the cfg for that block. */ |
2177 | if (need_eh_cleanup && stmt_could_throw_p (cfun, t)) |
2178 | bitmap_set_bit (need_eh_cleanup, gimple_bb (g: t)->index); |
2179 | |
2180 | /* Add uses to the worklist. */ |
2181 | ssa_op_iter iter; |
2182 | use_operand_p use_p; |
2183 | FOR_EACH_PHI_OR_STMT_USE (use_p, t, iter, SSA_OP_USE) |
2184 | { |
2185 | tree use = USE_FROM_PTR (use_p); |
2186 | if (TREE_CODE (use) == SSA_NAME |
2187 | && ! SSA_NAME_IS_DEFAULT_DEF (use)) |
2188 | bitmap_set_bit (worklist, SSA_NAME_VERSION (use)); |
2189 | } |
2190 | |
2191 | /* Remove stmt. */ |
2192 | if (dump_file && (dump_flags & TDF_DETAILS)) |
2193 | { |
2194 | fprintf (stream: dump_file, format: "Removing dead stmt:" ); |
2195 | print_gimple_stmt (dump_file, t, 0); |
2196 | } |
2197 | gimple_stmt_iterator gsi = gsi_for_stmt (t); |
2198 | if (gimple_code (g: t) == GIMPLE_PHI) |
2199 | { |
2200 | remove_phi_node (&gsi, true); |
2201 | phiremoved++; |
2202 | } |
2203 | else |
2204 | { |
2205 | unlink_stmt_vdef (t); |
2206 | gsi_remove (&gsi, true); |
2207 | release_defs (t); |
2208 | stmtremoved++; |
2209 | } |
2210 | } |
2211 | statistics_counter_event (cfun, "PHIs removed" , |
2212 | phiremoved); |
2213 | statistics_counter_event (cfun, "Statements removed" , |
2214 | stmtremoved); |
2215 | } |
2216 | |