re PR tree-optimization/59221 (wrong code at -O2 and -O3 on x86_64-linux-gnu)
[gcc.git] / gcc / tree-ssa-threadedge.c
1 /* SSA Jump Threading
2 Copyright (C) 2005-2013 Free Software Foundation, Inc.
3 Contributed by Jeff Law <law@redhat.com>
4
5 This file is part of GCC.
6
7 GCC is free software; you can redistribute it and/or modify
8 it under the terms of the GNU General Public License as published by
9 the Free Software Foundation; either version 3, or (at your option)
10 any later version.
11
12 GCC is distributed in the hope that it will be useful,
13 but WITHOUT ANY WARRANTY; without even the implied warranty of
14 MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
15 GNU General Public License for more details.
16
17 You should have received a copy of the GNU General Public License
18 along with GCC; see the file COPYING3. If not see
19 <http://www.gnu.org/licenses/>. */
20
21 #include "config.h"
22 #include "system.h"
23 #include "coretypes.h"
24 #include "tm.h"
25 #include "tree.h"
26 #include "flags.h"
27 #include "tm_p.h"
28 #include "basic-block.h"
29 #include "cfgloop.h"
30 #include "function.h"
31 #include "timevar.h"
32 #include "dumpfile.h"
33 #include "gimple.h"
34 #include "gimple-iterator.h"
35 #include "gimple-ssa.h"
36 #include "tree-cfg.h"
37 #include "tree-phinodes.h"
38 #include "ssa-iterators.h"
39 #include "stringpool.h"
40 #include "tree-ssanames.h"
41 #include "tree-ssa-propagate.h"
42 #include "tree-ssa-threadupdate.h"
43 #include "langhooks.h"
44 #include "params.h"
45 #include "tree-ssa-threadedge.h"
46
47 /* To avoid code explosion due to jump threading, we limit the
48 number of statements we are going to copy. This variable
49 holds the number of statements currently seen that we'll have
50 to copy as part of the jump threading process. */
51 static int stmt_count;
52
53 /* Array to record value-handles per SSA_NAME. */
54 vec<tree> ssa_name_values;
55
56 /* Set the value for the SSA name NAME to VALUE. */
57
58 void
59 set_ssa_name_value (tree name, tree value)
60 {
61 if (SSA_NAME_VERSION (name) >= ssa_name_values.length ())
62 ssa_name_values.safe_grow_cleared (SSA_NAME_VERSION (name) + 1);
63 if (value && TREE_OVERFLOW_P (value))
64 value = drop_tree_overflow (value);
65 ssa_name_values[SSA_NAME_VERSION (name)] = value;
66 }
67
68 /* Initialize the per SSA_NAME value-handles array. Returns it. */
69 void
70 threadedge_initialize_values (void)
71 {
72 gcc_assert (!ssa_name_values.exists ());
73 ssa_name_values.create (num_ssa_names);
74 }
75
76 /* Free the per SSA_NAME value-handle array. */
77 void
78 threadedge_finalize_values (void)
79 {
80 ssa_name_values.release ();
81 }
82
83 /* Return TRUE if we may be able to thread an incoming edge into
84 BB to an outgoing edge from BB. Return FALSE otherwise. */
85
86 bool
87 potentially_threadable_block (basic_block bb)
88 {
89 gimple_stmt_iterator gsi;
90
91 /* If BB has a single successor or a single predecessor, then
92 there is no threading opportunity. */
93 if (single_succ_p (bb) || single_pred_p (bb))
94 return false;
95
96 /* If BB does not end with a conditional, switch or computed goto,
97 then there is no threading opportunity. */
98 gsi = gsi_last_bb (bb);
99 if (gsi_end_p (gsi)
100 || ! gsi_stmt (gsi)
101 || (gimple_code (gsi_stmt (gsi)) != GIMPLE_COND
102 && gimple_code (gsi_stmt (gsi)) != GIMPLE_GOTO
103 && gimple_code (gsi_stmt (gsi)) != GIMPLE_SWITCH))
104 return false;
105
106 return true;
107 }
108
109 /* Return the LHS of any ASSERT_EXPR where OP appears as the first
110 argument to the ASSERT_EXPR and in which the ASSERT_EXPR dominates
111 BB. If no such ASSERT_EXPR is found, return OP. */
112
113 static tree
114 lhs_of_dominating_assert (tree op, basic_block bb, gimple stmt)
115 {
116 imm_use_iterator imm_iter;
117 gimple use_stmt;
118 use_operand_p use_p;
119
120 FOR_EACH_IMM_USE_FAST (use_p, imm_iter, op)
121 {
122 use_stmt = USE_STMT (use_p);
123 if (use_stmt != stmt
124 && gimple_assign_single_p (use_stmt)
125 && TREE_CODE (gimple_assign_rhs1 (use_stmt)) == ASSERT_EXPR
126 && TREE_OPERAND (gimple_assign_rhs1 (use_stmt), 0) == op
127 && dominated_by_p (CDI_DOMINATORS, bb, gimple_bb (use_stmt)))
128 {
129 return gimple_assign_lhs (use_stmt);
130 }
131 }
132 return op;
133 }
134
135 /* We record temporary equivalences created by PHI nodes or
136 statements within the target block. Doing so allows us to
137 identify more jump threading opportunities, even in blocks
138 with side effects.
139
140 We keep track of those temporary equivalences in a stack
141 structure so that we can unwind them when we're done processing
142 a particular edge. This routine handles unwinding the data
143 structures. */
144
145 static void
146 remove_temporary_equivalences (vec<tree> *stack)
147 {
148 while (stack->length () > 0)
149 {
150 tree prev_value, dest;
151
152 dest = stack->pop ();
153
154 /* A NULL value indicates we should stop unwinding, otherwise
155 pop off the next entry as they're recorded in pairs. */
156 if (dest == NULL)
157 break;
158
159 prev_value = stack->pop ();
160 set_ssa_name_value (dest, prev_value);
161 }
162 }
163
164 /* Record a temporary equivalence, saving enough information so that
165 we can restore the state of recorded equivalences when we're
166 done processing the current edge. */
167
168 static void
169 record_temporary_equivalence (tree x, tree y, vec<tree> *stack)
170 {
171 tree prev_x = SSA_NAME_VALUE (x);
172
173 if (TREE_CODE (y) == SSA_NAME)
174 {
175 tree tmp = SSA_NAME_VALUE (y);
176 y = tmp ? tmp : y;
177 }
178
179 set_ssa_name_value (x, y);
180 stack->reserve (2);
181 stack->quick_push (prev_x);
182 stack->quick_push (x);
183 }
184
185 /* Record temporary equivalences created by PHIs at the target of the
186 edge E. Record unwind information for the equivalences onto STACK.
187
188 If a PHI which prevents threading is encountered, then return FALSE
189 indicating we should not thread this edge, else return TRUE. */
190
191 static bool
192 record_temporary_equivalences_from_phis (edge e, vec<tree> *stack)
193 {
194 gimple_stmt_iterator gsi;
195
196 /* Each PHI creates a temporary equivalence, record them.
197 These are context sensitive equivalences and will be removed
198 later. */
199 for (gsi = gsi_start_phis (e->dest); !gsi_end_p (gsi); gsi_next (&gsi))
200 {
201 gimple phi = gsi_stmt (gsi);
202 tree src = PHI_ARG_DEF_FROM_EDGE (phi, e);
203 tree dst = gimple_phi_result (phi);
204
205 /* If the desired argument is not the same as this PHI's result
206 and it is set by a PHI in E->dest, then we can not thread
207 through E->dest. */
208 if (src != dst
209 && TREE_CODE (src) == SSA_NAME
210 && gimple_code (SSA_NAME_DEF_STMT (src)) == GIMPLE_PHI
211 && gimple_bb (SSA_NAME_DEF_STMT (src)) == e->dest)
212 return false;
213
214 /* We consider any non-virtual PHI as a statement since it
215 count result in a constant assignment or copy operation. */
216 if (!virtual_operand_p (dst))
217 stmt_count++;
218
219 record_temporary_equivalence (dst, src, stack);
220 }
221 return true;
222 }
223
224 /* Fold the RHS of an assignment statement and return it as a tree.
225 May return NULL_TREE if no simplification is possible. */
226
227 static tree
228 fold_assignment_stmt (gimple stmt)
229 {
230 enum tree_code subcode = gimple_assign_rhs_code (stmt);
231
232 switch (get_gimple_rhs_class (subcode))
233 {
234 case GIMPLE_SINGLE_RHS:
235 return fold (gimple_assign_rhs1 (stmt));
236
237 case GIMPLE_UNARY_RHS:
238 {
239 tree lhs = gimple_assign_lhs (stmt);
240 tree op0 = gimple_assign_rhs1 (stmt);
241 return fold_unary (subcode, TREE_TYPE (lhs), op0);
242 }
243
244 case GIMPLE_BINARY_RHS:
245 {
246 tree lhs = gimple_assign_lhs (stmt);
247 tree op0 = gimple_assign_rhs1 (stmt);
248 tree op1 = gimple_assign_rhs2 (stmt);
249 return fold_binary (subcode, TREE_TYPE (lhs), op0, op1);
250 }
251
252 case GIMPLE_TERNARY_RHS:
253 {
254 tree lhs = gimple_assign_lhs (stmt);
255 tree op0 = gimple_assign_rhs1 (stmt);
256 tree op1 = gimple_assign_rhs2 (stmt);
257 tree op2 = gimple_assign_rhs3 (stmt);
258
259 /* Sadly, we have to handle conditional assignments specially
260 here, because fold expects all the operands of an expression
261 to be folded before the expression itself is folded, but we
262 can't just substitute the folded condition here. */
263 if (gimple_assign_rhs_code (stmt) == COND_EXPR)
264 op0 = fold (op0);
265
266 return fold_ternary (subcode, TREE_TYPE (lhs), op0, op1, op2);
267 }
268
269 default:
270 gcc_unreachable ();
271 }
272 }
273
274 /* Try to simplify each statement in E->dest, ultimately leading to
275 a simplification of the COND_EXPR at the end of E->dest.
276
277 Record unwind information for temporary equivalences onto STACK.
278
279 Use SIMPLIFY (a pointer to a callback function) to further simplify
280 statements using pass specific information.
281
282 We might consider marking just those statements which ultimately
283 feed the COND_EXPR. It's not clear if the overhead of bookkeeping
284 would be recovered by trying to simplify fewer statements.
285
286 If we are able to simplify a statement into the form
287 SSA_NAME = (SSA_NAME | gimple invariant), then we can record
288 a context sensitive equivalence which may help us simplify
289 later statements in E->dest. */
290
291 static gimple
292 record_temporary_equivalences_from_stmts_at_dest (edge e,
293 vec<tree> *stack,
294 tree (*simplify) (gimple,
295 gimple))
296 {
297 gimple stmt = NULL;
298 gimple_stmt_iterator gsi;
299 int max_stmt_count;
300
301 max_stmt_count = PARAM_VALUE (PARAM_MAX_JUMP_THREAD_DUPLICATION_STMTS);
302
303 /* Walk through each statement in the block recording equivalences
304 we discover. Note any equivalences we discover are context
305 sensitive (ie, are dependent on traversing E) and must be unwound
306 when we're finished processing E. */
307 for (gsi = gsi_start_bb (e->dest); !gsi_end_p (gsi); gsi_next (&gsi))
308 {
309 tree cached_lhs = NULL;
310
311 stmt = gsi_stmt (gsi);
312
313 /* Ignore empty statements and labels. */
314 if (gimple_code (stmt) == GIMPLE_NOP
315 || gimple_code (stmt) == GIMPLE_LABEL
316 || is_gimple_debug (stmt))
317 continue;
318
319 /* If the statement has volatile operands, then we assume we
320 can not thread through this block. This is overly
321 conservative in some ways. */
322 if (gimple_code (stmt) == GIMPLE_ASM && gimple_asm_volatile_p (stmt))
323 return NULL;
324
325 /* If duplicating this block is going to cause too much code
326 expansion, then do not thread through this block. */
327 stmt_count++;
328 if (stmt_count > max_stmt_count)
329 return NULL;
330
331 /* If this is not a statement that sets an SSA_NAME to a new
332 value, then do not try to simplify this statement as it will
333 not simplify in any way that is helpful for jump threading. */
334 if ((gimple_code (stmt) != GIMPLE_ASSIGN
335 || TREE_CODE (gimple_assign_lhs (stmt)) != SSA_NAME)
336 && (gimple_code (stmt) != GIMPLE_CALL
337 || gimple_call_lhs (stmt) == NULL_TREE
338 || TREE_CODE (gimple_call_lhs (stmt)) != SSA_NAME))
339 continue;
340
341 /* The result of __builtin_object_size depends on all the arguments
342 of a phi node. Temporarily using only one edge produces invalid
343 results. For example
344
345 if (x < 6)
346 goto l;
347 else
348 goto l;
349
350 l:
351 r = PHI <&w[2].a[1](2), &a.a[6](3)>
352 __builtin_object_size (r, 0)
353
354 The result of __builtin_object_size is defined to be the maximum of
355 remaining bytes. If we use only one edge on the phi, the result will
356 change to be the remaining bytes for the corresponding phi argument.
357
358 Similarly for __builtin_constant_p:
359
360 r = PHI <1(2), 2(3)>
361 __builtin_constant_p (r)
362
363 Both PHI arguments are constant, but x ? 1 : 2 is still not
364 constant. */
365
366 if (is_gimple_call (stmt))
367 {
368 tree fndecl = gimple_call_fndecl (stmt);
369 if (fndecl
370 && (DECL_FUNCTION_CODE (fndecl) == BUILT_IN_OBJECT_SIZE
371 || DECL_FUNCTION_CODE (fndecl) == BUILT_IN_CONSTANT_P))
372 continue;
373 }
374
375 /* At this point we have a statement which assigns an RHS to an
376 SSA_VAR on the LHS. We want to try and simplify this statement
377 to expose more context sensitive equivalences which in turn may
378 allow us to simplify the condition at the end of the loop.
379
380 Handle simple copy operations as well as implied copies from
381 ASSERT_EXPRs. */
382 if (gimple_assign_single_p (stmt)
383 && TREE_CODE (gimple_assign_rhs1 (stmt)) == SSA_NAME)
384 cached_lhs = gimple_assign_rhs1 (stmt);
385 else if (gimple_assign_single_p (stmt)
386 && TREE_CODE (gimple_assign_rhs1 (stmt)) == ASSERT_EXPR)
387 cached_lhs = TREE_OPERAND (gimple_assign_rhs1 (stmt), 0);
388 else
389 {
390 /* A statement that is not a trivial copy or ASSERT_EXPR.
391 We're going to temporarily copy propagate the operands
392 and see if that allows us to simplify this statement. */
393 tree *copy;
394 ssa_op_iter iter;
395 use_operand_p use_p;
396 unsigned int num, i = 0;
397
398 num = NUM_SSA_OPERANDS (stmt, (SSA_OP_USE | SSA_OP_VUSE));
399 copy = XCNEWVEC (tree, num);
400
401 /* Make a copy of the uses & vuses into USES_COPY, then cprop into
402 the operands. */
403 FOR_EACH_SSA_USE_OPERAND (use_p, stmt, iter, SSA_OP_USE | SSA_OP_VUSE)
404 {
405 tree tmp = NULL;
406 tree use = USE_FROM_PTR (use_p);
407
408 copy[i++] = use;
409 if (TREE_CODE (use) == SSA_NAME)
410 tmp = SSA_NAME_VALUE (use);
411 if (tmp)
412 SET_USE (use_p, tmp);
413 }
414
415 /* Try to fold/lookup the new expression. Inserting the
416 expression into the hash table is unlikely to help. */
417 if (is_gimple_call (stmt))
418 cached_lhs = fold_call_stmt (stmt, false);
419 else
420 cached_lhs = fold_assignment_stmt (stmt);
421
422 if (!cached_lhs
423 || (TREE_CODE (cached_lhs) != SSA_NAME
424 && !is_gimple_min_invariant (cached_lhs)))
425 cached_lhs = (*simplify) (stmt, stmt);
426
427 /* Restore the statement's original uses/defs. */
428 i = 0;
429 FOR_EACH_SSA_USE_OPERAND (use_p, stmt, iter, SSA_OP_USE | SSA_OP_VUSE)
430 SET_USE (use_p, copy[i++]);
431
432 free (copy);
433 }
434
435 /* Record the context sensitive equivalence if we were able
436 to simplify this statement. */
437 if (cached_lhs
438 && (TREE_CODE (cached_lhs) == SSA_NAME
439 || is_gimple_min_invariant (cached_lhs)))
440 record_temporary_equivalence (gimple_get_lhs (stmt), cached_lhs, stack);
441 }
442 return stmt;
443 }
444
445 /* Simplify the control statement at the end of the block E->dest.
446
447 To avoid allocating memory unnecessarily, a scratch GIMPLE_COND
448 is available to use/clobber in DUMMY_COND.
449
450 Use SIMPLIFY (a pointer to a callback function) to further simplify
451 a condition using pass specific information.
452
453 Return the simplified condition or NULL if simplification could
454 not be performed. */
455
456 static tree
457 simplify_control_stmt_condition (edge e,
458 gimple stmt,
459 gimple dummy_cond,
460 tree (*simplify) (gimple, gimple),
461 bool handle_dominating_asserts)
462 {
463 tree cond, cached_lhs;
464 enum gimple_code code = gimple_code (stmt);
465
466 /* For comparisons, we have to update both operands, then try
467 to simplify the comparison. */
468 if (code == GIMPLE_COND)
469 {
470 tree op0, op1;
471 enum tree_code cond_code;
472
473 op0 = gimple_cond_lhs (stmt);
474 op1 = gimple_cond_rhs (stmt);
475 cond_code = gimple_cond_code (stmt);
476
477 /* Get the current value of both operands. */
478 if (TREE_CODE (op0) == SSA_NAME)
479 {
480 tree tmp = SSA_NAME_VALUE (op0);
481 if (tmp)
482 op0 = tmp;
483 }
484
485 if (TREE_CODE (op1) == SSA_NAME)
486 {
487 tree tmp = SSA_NAME_VALUE (op1);
488 if (tmp)
489 op1 = tmp;
490 }
491
492 if (handle_dominating_asserts)
493 {
494 /* Now see if the operand was consumed by an ASSERT_EXPR
495 which dominates E->src. If so, we want to replace the
496 operand with the LHS of the ASSERT_EXPR. */
497 if (TREE_CODE (op0) == SSA_NAME)
498 op0 = lhs_of_dominating_assert (op0, e->src, stmt);
499
500 if (TREE_CODE (op1) == SSA_NAME)
501 op1 = lhs_of_dominating_assert (op1, e->src, stmt);
502 }
503
504 /* We may need to canonicalize the comparison. For
505 example, op0 might be a constant while op1 is an
506 SSA_NAME. Failure to canonicalize will cause us to
507 miss threading opportunities. */
508 if (tree_swap_operands_p (op0, op1, false))
509 {
510 tree tmp;
511 cond_code = swap_tree_comparison (cond_code);
512 tmp = op0;
513 op0 = op1;
514 op1 = tmp;
515 }
516
517 /* Stuff the operator and operands into our dummy conditional
518 expression. */
519 gimple_cond_set_code (dummy_cond, cond_code);
520 gimple_cond_set_lhs (dummy_cond, op0);
521 gimple_cond_set_rhs (dummy_cond, op1);
522
523 /* We absolutely do not care about any type conversions
524 we only care about a zero/nonzero value. */
525 fold_defer_overflow_warnings ();
526
527 cached_lhs = fold_binary (cond_code, boolean_type_node, op0, op1);
528 if (cached_lhs)
529 while (CONVERT_EXPR_P (cached_lhs))
530 cached_lhs = TREE_OPERAND (cached_lhs, 0);
531
532 fold_undefer_overflow_warnings ((cached_lhs
533 && is_gimple_min_invariant (cached_lhs)),
534 stmt, WARN_STRICT_OVERFLOW_CONDITIONAL);
535
536 /* If we have not simplified the condition down to an invariant,
537 then use the pass specific callback to simplify the condition. */
538 if (!cached_lhs
539 || !is_gimple_min_invariant (cached_lhs))
540 cached_lhs = (*simplify) (dummy_cond, stmt);
541
542 return cached_lhs;
543 }
544
545 if (code == GIMPLE_SWITCH)
546 cond = gimple_switch_index (stmt);
547 else if (code == GIMPLE_GOTO)
548 cond = gimple_goto_dest (stmt);
549 else
550 gcc_unreachable ();
551
552 /* We can have conditionals which just test the state of a variable
553 rather than use a relational operator. These are simpler to handle. */
554 if (TREE_CODE (cond) == SSA_NAME)
555 {
556 cached_lhs = cond;
557
558 /* Get the variable's current value from the equivalence chains.
559
560 It is possible to get loops in the SSA_NAME_VALUE chains
561 (consider threading the backedge of a loop where we have
562 a loop invariant SSA_NAME used in the condition. */
563 if (cached_lhs
564 && TREE_CODE (cached_lhs) == SSA_NAME
565 && SSA_NAME_VALUE (cached_lhs))
566 cached_lhs = SSA_NAME_VALUE (cached_lhs);
567
568 /* If we're dominated by a suitable ASSERT_EXPR, then
569 update CACHED_LHS appropriately. */
570 if (handle_dominating_asserts && TREE_CODE (cached_lhs) == SSA_NAME)
571 cached_lhs = lhs_of_dominating_assert (cached_lhs, e->src, stmt);
572
573 /* If we haven't simplified to an invariant yet, then use the
574 pass specific callback to try and simplify it further. */
575 if (cached_lhs && ! is_gimple_min_invariant (cached_lhs))
576 cached_lhs = (*simplify) (stmt, stmt);
577 }
578 else
579 cached_lhs = NULL;
580
581 return cached_lhs;
582 }
583
584 /* Return TRUE if the statement at the end of e->dest depends on
585 the output of any statement in BB. Otherwise return FALSE.
586
587 This is used when we are threading a backedge and need to ensure
588 that temporary equivalences from BB do not affect the condition
589 in e->dest. */
590
591 static bool
592 cond_arg_set_in_bb (edge e, basic_block bb)
593 {
594 ssa_op_iter iter;
595 use_operand_p use_p;
596 gimple last = last_stmt (e->dest);
597
598 /* E->dest does not have to end with a control transferring
599 instruction. This can occur when we try to extend a jump
600 threading opportunity deeper into the CFG. In that case
601 it is safe for this check to return false. */
602 if (!last)
603 return false;
604
605 if (gimple_code (last) != GIMPLE_COND
606 && gimple_code (last) != GIMPLE_GOTO
607 && gimple_code (last) != GIMPLE_SWITCH)
608 return false;
609
610 FOR_EACH_SSA_USE_OPERAND (use_p, last, iter, SSA_OP_USE | SSA_OP_VUSE)
611 {
612 tree use = USE_FROM_PTR (use_p);
613
614 if (TREE_CODE (use) == SSA_NAME
615 && gimple_code (SSA_NAME_DEF_STMT (use)) != GIMPLE_PHI
616 && gimple_bb (SSA_NAME_DEF_STMT (use)) == bb)
617 return true;
618 }
619 return false;
620 }
621
622 /* Copy debug stmts from DEST's chain of single predecessors up to
623 SRC, so that we don't lose the bindings as PHI nodes are introduced
624 when DEST gains new predecessors. */
625 void
626 propagate_threaded_block_debug_into (basic_block dest, basic_block src)
627 {
628 if (!MAY_HAVE_DEBUG_STMTS)
629 return;
630
631 if (!single_pred_p (dest))
632 return;
633
634 gcc_checking_assert (dest != src);
635
636 gimple_stmt_iterator gsi = gsi_after_labels (dest);
637 int i = 0;
638 const int alloc_count = 16; // ?? Should this be a PARAM?
639
640 /* Estimate the number of debug vars overridden in the beginning of
641 DEST, to tell how many we're going to need to begin with. */
642 for (gimple_stmt_iterator si = gsi;
643 i * 4 <= alloc_count * 3 && !gsi_end_p (si); gsi_next (&si))
644 {
645 gimple stmt = gsi_stmt (si);
646 if (!is_gimple_debug (stmt))
647 break;
648 i++;
649 }
650
651 stack_vec<tree, alloc_count> fewvars;
652 pointer_set_t *vars = NULL;
653
654 /* If we're already starting with 3/4 of alloc_count, go for a
655 pointer_set, otherwise start with an unordered stack-allocated
656 VEC. */
657 if (i * 4 > alloc_count * 3)
658 vars = pointer_set_create ();
659
660 /* Now go through the initial debug stmts in DEST again, this time
661 actually inserting in VARS or FEWVARS. Don't bother checking for
662 duplicates in FEWVARS. */
663 for (gimple_stmt_iterator si = gsi; !gsi_end_p (si); gsi_next (&si))
664 {
665 gimple stmt = gsi_stmt (si);
666 if (!is_gimple_debug (stmt))
667 break;
668
669 tree var;
670
671 if (gimple_debug_bind_p (stmt))
672 var = gimple_debug_bind_get_var (stmt);
673 else if (gimple_debug_source_bind_p (stmt))
674 var = gimple_debug_source_bind_get_var (stmt);
675 else
676 gcc_unreachable ();
677
678 if (vars)
679 pointer_set_insert (vars, var);
680 else
681 fewvars.quick_push (var);
682 }
683
684 basic_block bb = dest;
685
686 do
687 {
688 bb = single_pred (bb);
689 for (gimple_stmt_iterator si = gsi_last_bb (bb);
690 !gsi_end_p (si); gsi_prev (&si))
691 {
692 gimple stmt = gsi_stmt (si);
693 if (!is_gimple_debug (stmt))
694 continue;
695
696 tree var;
697
698 if (gimple_debug_bind_p (stmt))
699 var = gimple_debug_bind_get_var (stmt);
700 else if (gimple_debug_source_bind_p (stmt))
701 var = gimple_debug_source_bind_get_var (stmt);
702 else
703 gcc_unreachable ();
704
705 /* Discard debug bind overlaps. ??? Unlike stmts from src,
706 copied into a new block that will precede BB, debug bind
707 stmts in bypassed BBs may actually be discarded if
708 they're overwritten by subsequent debug bind stmts, which
709 might be a problem once we introduce stmt frontier notes
710 or somesuch. Adding `&& bb == src' to the condition
711 below will preserve all potentially relevant debug
712 notes. */
713 if (vars && pointer_set_insert (vars, var))
714 continue;
715 else if (!vars)
716 {
717 int i = fewvars.length ();
718 while (i--)
719 if (fewvars[i] == var)
720 break;
721 if (i >= 0)
722 continue;
723
724 if (fewvars.length () < (unsigned) alloc_count)
725 fewvars.quick_push (var);
726 else
727 {
728 vars = pointer_set_create ();
729 for (i = 0; i < alloc_count; i++)
730 pointer_set_insert (vars, fewvars[i]);
731 fewvars.release ();
732 pointer_set_insert (vars, var);
733 }
734 }
735
736 stmt = gimple_copy (stmt);
737 /* ??? Should we drop the location of the copy to denote
738 they're artificial bindings? */
739 gsi_insert_before (&gsi, stmt, GSI_NEW_STMT);
740 }
741 }
742 while (bb != src && single_pred_p (bb));
743
744 if (vars)
745 pointer_set_destroy (vars);
746 else if (fewvars.exists ())
747 fewvars.release ();
748 }
749
750 /* See if TAKEN_EDGE->dest is a threadable block with no side effecs (ie, it
751 need not be duplicated as part of the CFG/SSA updating process).
752
753 If it is threadable, add it to PATH and VISITED and recurse, ultimately
754 returning TRUE from the toplevel call. Otherwise do nothing and
755 return false.
756
757 DUMMY_COND, HANDLE_DOMINATING_ASSERTS and SIMPLIFY are used to
758 try and simplify the condition at the end of TAKEN_EDGE->dest. */
759 static bool
760 thread_around_empty_blocks (edge taken_edge,
761 gimple dummy_cond,
762 bool handle_dominating_asserts,
763 tree (*simplify) (gimple, gimple),
764 bitmap visited,
765 vec<jump_thread_edge *> *path,
766 bool *backedge_seen_p)
767 {
768 basic_block bb = taken_edge->dest;
769 gimple_stmt_iterator gsi;
770 gimple stmt;
771 tree cond;
772
773 /* The key property of these blocks is that they need not be duplicated
774 when threading. Thus they can not have visible side effects such
775 as PHI nodes. */
776 if (!gsi_end_p (gsi_start_phis (bb)))
777 return false;
778
779 /* Skip over DEBUG statements at the start of the block. */
780 gsi = gsi_start_nondebug_bb (bb);
781
782 /* If the block has no statements, but does have a single successor, then
783 it's just a forwarding block and we can thread through it trivially.
784
785 However, note that just threading through empty blocks with single
786 successors is not inherently profitable. For the jump thread to
787 be profitable, we must avoid a runtime conditional.
788
789 By taking the return value from the recursive call, we get the
790 desired effect of returning TRUE when we found a profitable jump
791 threading opportunity and FALSE otherwise.
792
793 This is particularly important when this routine is called after
794 processing a joiner block. Returning TRUE too aggressively in
795 that case results in pointless duplication of the joiner block. */
796 if (gsi_end_p (gsi))
797 {
798 if (single_succ_p (bb))
799 {
800 taken_edge = single_succ_edge (bb);
801 if (!bitmap_bit_p (visited, taken_edge->dest->index))
802 {
803 jump_thread_edge *x
804 = new jump_thread_edge (taken_edge, EDGE_NO_COPY_SRC_BLOCK);
805 path->safe_push (x);
806 bitmap_set_bit (visited, taken_edge->dest->index);
807 *backedge_seen_p |= ((taken_edge->flags & EDGE_DFS_BACK) != 0);
808 return thread_around_empty_blocks (taken_edge,
809 dummy_cond,
810 handle_dominating_asserts,
811 simplify,
812 visited,
813 path,
814 backedge_seen_p);
815 }
816 }
817
818 /* We have a block with no statements, but multiple successors? */
819 return false;
820 }
821
822 /* The only real statements this block can have are a control
823 flow altering statement. Anything else stops the thread. */
824 stmt = gsi_stmt (gsi);
825 if (gimple_code (stmt) != GIMPLE_COND
826 && gimple_code (stmt) != GIMPLE_GOTO
827 && gimple_code (stmt) != GIMPLE_SWITCH)
828 return false;
829
830 /* Extract and simplify the condition. */
831 cond = simplify_control_stmt_condition (taken_edge, stmt, dummy_cond,
832 simplify, handle_dominating_asserts);
833
834 /* If the condition can be statically computed and we have not already
835 visited the destination edge, then add the taken edge to our thread
836 path. */
837 if (cond && is_gimple_min_invariant (cond))
838 {
839 taken_edge = find_taken_edge (bb, cond);
840
841 if (bitmap_bit_p (visited, taken_edge->dest->index))
842 return false;
843 bitmap_set_bit (visited, taken_edge->dest->index);
844
845 jump_thread_edge *x
846 = new jump_thread_edge (taken_edge, EDGE_NO_COPY_SRC_BLOCK);
847 path->safe_push (x);
848 *backedge_seen_p |= ((taken_edge->flags & EDGE_DFS_BACK) != 0);
849
850 thread_around_empty_blocks (taken_edge,
851 dummy_cond,
852 handle_dominating_asserts,
853 simplify,
854 visited,
855 path,
856 backedge_seen_p);
857 return true;
858 }
859
860 return false;
861 }
862
863 /* We are exiting E->src, see if E->dest ends with a conditional
864 jump which has a known value when reached via E.
865
866 E->dest can have arbitrary side effects which, if threading is
867 successful, will be maintained.
868
869 Special care is necessary if E is a back edge in the CFG as we
870 may have already recorded equivalences for E->dest into our
871 various tables, including the result of the conditional at
872 the end of E->dest. Threading opportunities are severely
873 limited in that case to avoid short-circuiting the loop
874 incorrectly.
875
876 DUMMY_COND is a shared cond_expr used by condition simplification as scratch,
877 to avoid allocating memory.
878
879 HANDLE_DOMINATING_ASSERTS is true if we should try to replace operands of
880 the simplified condition with left-hand sides of ASSERT_EXPRs they are
881 used in.
882
883 STACK is used to undo temporary equivalences created during the walk of
884 E->dest.
885
886 SIMPLIFY is a pass-specific function used to simplify statements.
887
888 Our caller is responsible for restoring the state of the expression
889 and const_and_copies stacks. */
890
891 static bool
892 thread_through_normal_block (edge e,
893 gimple dummy_cond,
894 bool handle_dominating_asserts,
895 vec<tree> *stack,
896 tree (*simplify) (gimple, gimple),
897 vec<jump_thread_edge *> *path,
898 bitmap visited,
899 bool *backedge_seen_p)
900 {
901 /* If we have crossed a backedge, then we want to verify that the COND_EXPR,
902 SWITCH_EXPR or GOTO_EXPR at the end of e->dest is not affected
903 by any statements in e->dest. If it is affected, then it is not
904 safe to thread this edge. */
905 if (*backedge_seen_p
906 && cond_arg_set_in_bb (e, e->dest))
907 return false;
908
909 /* PHIs create temporary equivalences. */
910 if (!record_temporary_equivalences_from_phis (e, stack))
911 return false;
912
913 /* Now walk each statement recording any context sensitive
914 temporary equivalences we can detect. */
915 gimple stmt
916 = record_temporary_equivalences_from_stmts_at_dest (e, stack, simplify);
917 if (!stmt)
918 return false;
919
920 /* If we stopped at a COND_EXPR or SWITCH_EXPR, see if we know which arm
921 will be taken. */
922 if (gimple_code (stmt) == GIMPLE_COND
923 || gimple_code (stmt) == GIMPLE_GOTO
924 || gimple_code (stmt) == GIMPLE_SWITCH)
925 {
926 tree cond;
927
928 /* Extract and simplify the condition. */
929 cond = simplify_control_stmt_condition (e, stmt, dummy_cond, simplify,
930 handle_dominating_asserts);
931
932 if (cond && is_gimple_min_invariant (cond))
933 {
934 edge taken_edge = find_taken_edge (e->dest, cond);
935 basic_block dest = (taken_edge ? taken_edge->dest : NULL);
936
937 /* DEST could be NULL for a computed jump to an absolute
938 address. */
939 if (dest == NULL
940 || dest == e->dest
941 || bitmap_bit_p (visited, dest->index))
942 return false;
943
944 /* Only push the EDGE_START_JUMP_THREAD marker if this is
945 first edge on the path. */
946 if (path->length () == 0)
947 {
948 jump_thread_edge *x
949 = new jump_thread_edge (e, EDGE_START_JUMP_THREAD);
950 path->safe_push (x);
951 *backedge_seen_p |= ((e->flags & EDGE_DFS_BACK) != 0);
952 }
953
954 jump_thread_edge *x
955 = new jump_thread_edge (taken_edge, EDGE_COPY_SRC_BLOCK);
956 path->safe_push (x);
957 *backedge_seen_p |= ((taken_edge->flags & EDGE_DFS_BACK) != 0);
958
959 /* See if we can thread through DEST as well, this helps capture
960 secondary effects of threading without having to re-run DOM or
961 VRP. */
962 if (!*backedge_seen_p
963 || ! cond_arg_set_in_bb (taken_edge, e->dest))
964 {
965 /* We don't want to thread back to a block we have already
966 visited. This may be overly conservative. */
967 bitmap_set_bit (visited, dest->index);
968 bitmap_set_bit (visited, e->dest->index);
969 thread_around_empty_blocks (taken_edge,
970 dummy_cond,
971 handle_dominating_asserts,
972 simplify,
973 visited,
974 path,
975 backedge_seen_p);
976 }
977 return true;
978 }
979 }
980 return false;
981 }
982
983 /* We are exiting E->src, see if E->dest ends with a conditional
984 jump which has a known value when reached via E.
985
986 Special care is necessary if E is a back edge in the CFG as we
987 may have already recorded equivalences for E->dest into our
988 various tables, including the result of the conditional at
989 the end of E->dest. Threading opportunities are severely
990 limited in that case to avoid short-circuiting the loop
991 incorrectly.
992
993 Note it is quite common for the first block inside a loop to
994 end with a conditional which is either always true or always
995 false when reached via the loop backedge. Thus we do not want
996 to blindly disable threading across a loop backedge.
997
998 DUMMY_COND is a shared cond_expr used by condition simplification as scratch,
999 to avoid allocating memory.
1000
1001 HANDLE_DOMINATING_ASSERTS is true if we should try to replace operands of
1002 the simplified condition with left-hand sides of ASSERT_EXPRs they are
1003 used in.
1004
1005 STACK is used to undo temporary equivalences created during the walk of
1006 E->dest.
1007
1008 SIMPLIFY is a pass-specific function used to simplify statements. */
1009
1010 void
1011 thread_across_edge (gimple dummy_cond,
1012 edge e,
1013 bool handle_dominating_asserts,
1014 vec<tree> *stack,
1015 tree (*simplify) (gimple, gimple))
1016 {
1017 bitmap visited = BITMAP_ALLOC (NULL);
1018 bool backedge_seen;
1019
1020 stmt_count = 0;
1021
1022 vec<jump_thread_edge *> *path = new vec<jump_thread_edge *> ();
1023 bitmap_clear (visited);
1024 bitmap_set_bit (visited, e->src->index);
1025 bitmap_set_bit (visited, e->dest->index);
1026 backedge_seen = ((e->flags & EDGE_DFS_BACK) != 0);
1027 if (thread_through_normal_block (e, dummy_cond, handle_dominating_asserts,
1028 stack, simplify, path, visited,
1029 &backedge_seen))
1030 {
1031 propagate_threaded_block_debug_into (path->last ()->e->dest,
1032 e->dest);
1033 remove_temporary_equivalences (stack);
1034 BITMAP_FREE (visited);
1035 register_jump_thread (path);
1036 return;
1037 }
1038 else
1039 {
1040 /* There should be no edges on the path, so no need to walk through
1041 the vector entries. */
1042 gcc_assert (path->length () == 0);
1043 path->release ();
1044 }
1045
1046 /* We were unable to determine what out edge from E->dest is taken. However,
1047 we might still be able to thread through successors of E->dest. This
1048 often occurs when E->dest is a joiner block which then fans back out
1049 based on redundant tests.
1050
1051 If so, we'll copy E->dest and redirect the appropriate predecessor to
1052 the copy. Within the copy of E->dest, we'll thread one or more edges
1053 to points deeper in the CFG.
1054
1055 This is a stopgap until we have a more structured approach to path
1056 isolation. */
1057 {
1058 edge taken_edge;
1059 edge_iterator ei;
1060 bool found;
1061
1062 /* If E->dest has abnormal outgoing edges, then there's no guarantee
1063 we can safely redirect any of the edges. Just punt those cases. */
1064 FOR_EACH_EDGE (taken_edge, ei, e->dest->succs)
1065 if (taken_edge->flags & EDGE_ABNORMAL)
1066 {
1067 remove_temporary_equivalences (stack);
1068 BITMAP_FREE (visited);
1069 return;
1070 }
1071
1072 /* Look at each successor of E->dest to see if we can thread through it. */
1073 FOR_EACH_EDGE (taken_edge, ei, e->dest->succs)
1074 {
1075 /* Push a fresh marker so we can unwind the equivalences created
1076 for each of E->dest's successors. */
1077 stack->safe_push (NULL_TREE);
1078
1079 /* Avoid threading to any block we have already visited. */
1080 bitmap_clear (visited);
1081 bitmap_set_bit (visited, taken_edge->dest->index);
1082 bitmap_set_bit (visited, e->dest->index);
1083 vec<jump_thread_edge *> *path = new vec<jump_thread_edge *> ();
1084
1085 /* Record whether or not we were able to thread through a successor
1086 of E->dest. */
1087 jump_thread_edge *x = new jump_thread_edge (e, EDGE_START_JUMP_THREAD);
1088 path->safe_push (x);
1089
1090 x = new jump_thread_edge (taken_edge, EDGE_COPY_SRC_JOINER_BLOCK);
1091 path->safe_push (x);
1092 found = false;
1093 backedge_seen = ((e->flags & EDGE_DFS_BACK) != 0);
1094 backedge_seen |= ((taken_edge->flags & EDGE_DFS_BACK) != 0);
1095 if (!backedge_seen
1096 || ! cond_arg_set_in_bb (path->last ()->e, e->dest))
1097 found = thread_around_empty_blocks (taken_edge,
1098 dummy_cond,
1099 handle_dominating_asserts,
1100 simplify,
1101 visited,
1102 path,
1103 &backedge_seen);
1104
1105 if (!found
1106 && (!backedge_seen
1107 || ! cond_arg_set_in_bb (path->last ()->e, e->dest)))
1108 found = thread_through_normal_block (path->last ()->e, dummy_cond,
1109 handle_dominating_asserts,
1110 stack, simplify, path, visited,
1111 &backedge_seen);
1112
1113 /* If we were able to thread through a successor of E->dest, then
1114 record the jump threading opportunity. */
1115 if (found)
1116 {
1117 propagate_threaded_block_debug_into (path->last ()->e->dest,
1118 taken_edge->dest);
1119 register_jump_thread (path);
1120 }
1121 else
1122 {
1123 delete_jump_thread_path (path);
1124 }
1125
1126 /* And unwind the equivalence table. */
1127 remove_temporary_equivalences (stack);
1128 }
1129 BITMAP_FREE (visited);
1130 }
1131
1132 remove_temporary_equivalences (stack);
1133 }