* trans-mem.c (execute_tm_mark): Release bb_regions.
[official-gcc.git] / gcc / tree-ssa-threadedge.c
blob40c9d44d92417dbdc63841e65bccd7fbd5d3f737
1 /* SSA Jump Threading
2 Copyright (C) 2005, 2006, 2007, 2008, 2009, 2010, 2011, 2012
3 Free Software Foundation, Inc.
4 Contributed by Jeff Law <law@redhat.com>
6 This file is part of GCC.
8 GCC is free software; you can redistribute it and/or modify
9 it under the terms of the GNU General Public License as published by
10 the Free Software Foundation; either version 3, or (at your option)
11 any later version.
13 GCC is distributed in the hope that it will be useful,
14 but WITHOUT ANY WARRANTY; without even the implied warranty of
15 MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
16 GNU General Public License for more details.
18 You should have received a copy of the GNU General Public License
19 along with GCC; see the file COPYING3. If not see
20 <http://www.gnu.org/licenses/>. */
22 #include "config.h"
23 #include "system.h"
24 #include "coretypes.h"
25 #include "tm.h"
26 #include "tree.h"
27 #include "flags.h"
28 #include "tm_p.h"
29 #include "basic-block.h"
30 #include "cfgloop.h"
31 #include "function.h"
32 #include "timevar.h"
33 #include "dumpfile.h"
34 #include "tree-flow.h"
35 #include "tree-ssa-propagate.h"
36 #include "langhooks.h"
37 #include "params.h"
39 /* To avoid code explosion due to jump threading, we limit the
40 number of statements we are going to copy. This variable
41 holds the number of statements currently seen that we'll have
42 to copy as part of the jump threading process. */
43 static int stmt_count;
45 /* Array to record value-handles per SSA_NAME. */
46 vec<tree> ssa_name_values;
48 /* Set the value for the SSA name NAME to VALUE. */
50 void
51 set_ssa_name_value (tree name, tree value)
53 if (SSA_NAME_VERSION (name) >= ssa_name_values.length ())
54 ssa_name_values.safe_grow_cleared (SSA_NAME_VERSION (name) + 1);
55 ssa_name_values[SSA_NAME_VERSION (name)] = value;
58 /* Initialize the per SSA_NAME value-handles array. Returns it. */
59 void
60 threadedge_initialize_values (void)
62 gcc_assert (!ssa_name_values.exists ());
63 ssa_name_values.create (num_ssa_names);
66 /* Free the per SSA_NAME value-handle array. */
67 void
68 threadedge_finalize_values (void)
70 ssa_name_values.release ();
73 /* Return TRUE if we may be able to thread an incoming edge into
74 BB to an outgoing edge from BB. Return FALSE otherwise. */
76 bool
77 potentially_threadable_block (basic_block bb)
79 gimple_stmt_iterator gsi;
81 /* If BB has a single successor or a single predecessor, then
82 there is no threading opportunity. */
83 if (single_succ_p (bb) || single_pred_p (bb))
84 return false;
86 /* If BB does not end with a conditional, switch or computed goto,
87 then there is no threading opportunity. */
88 gsi = gsi_last_bb (bb);
89 if (gsi_end_p (gsi)
90 || ! gsi_stmt (gsi)
91 || (gimple_code (gsi_stmt (gsi)) != GIMPLE_COND
92 && gimple_code (gsi_stmt (gsi)) != GIMPLE_GOTO
93 && gimple_code (gsi_stmt (gsi)) != GIMPLE_SWITCH))
94 return false;
96 return true;
99 /* Return the LHS of any ASSERT_EXPR where OP appears as the first
100 argument to the ASSERT_EXPR and in which the ASSERT_EXPR dominates
101 BB. If no such ASSERT_EXPR is found, return OP. */
103 static tree
104 lhs_of_dominating_assert (tree op, basic_block bb, gimple stmt)
106 imm_use_iterator imm_iter;
107 gimple use_stmt;
108 use_operand_p use_p;
110 FOR_EACH_IMM_USE_FAST (use_p, imm_iter, op)
112 use_stmt = USE_STMT (use_p);
113 if (use_stmt != stmt
114 && gimple_assign_single_p (use_stmt)
115 && TREE_CODE (gimple_assign_rhs1 (use_stmt)) == ASSERT_EXPR
116 && TREE_OPERAND (gimple_assign_rhs1 (use_stmt), 0) == op
117 && dominated_by_p (CDI_DOMINATORS, bb, gimple_bb (use_stmt)))
119 return gimple_assign_lhs (use_stmt);
122 return op;
125 /* We record temporary equivalences created by PHI nodes or
126 statements within the target block. Doing so allows us to
127 identify more jump threading opportunities, even in blocks
128 with side effects.
130 We keep track of those temporary equivalences in a stack
131 structure so that we can unwind them when we're done processing
132 a particular edge. This routine handles unwinding the data
133 structures. */
135 static void
136 remove_temporary_equivalences (vec<tree> *stack)
138 while (stack->length () > 0)
140 tree prev_value, dest;
142 dest = stack->pop ();
144 /* A NULL value indicates we should stop unwinding, otherwise
145 pop off the next entry as they're recorded in pairs. */
146 if (dest == NULL)
147 break;
149 prev_value = stack->pop ();
150 set_ssa_name_value (dest, prev_value);
154 /* Record a temporary equivalence, saving enough information so that
155 we can restore the state of recorded equivalences when we're
156 done processing the current edge. */
158 static void
159 record_temporary_equivalence (tree x, tree y, vec<tree> *stack)
161 tree prev_x = SSA_NAME_VALUE (x);
163 if (TREE_CODE (y) == SSA_NAME)
165 tree tmp = SSA_NAME_VALUE (y);
166 y = tmp ? tmp : y;
169 set_ssa_name_value (x, y);
170 stack->reserve (2);
171 stack->quick_push (prev_x);
172 stack->quick_push (x);
175 /* Record temporary equivalences created by PHIs at the target of the
176 edge E. Record unwind information for the equivalences onto STACK.
178 If a PHI which prevents threading is encountered, then return FALSE
179 indicating we should not thread this edge, else return TRUE. */
181 static bool
182 record_temporary_equivalences_from_phis (edge e, vec<tree> *stack)
184 gimple_stmt_iterator gsi;
186 /* Each PHI creates a temporary equivalence, record them.
187 These are context sensitive equivalences and will be removed
188 later. */
189 for (gsi = gsi_start_phis (e->dest); !gsi_end_p (gsi); gsi_next (&gsi))
191 gimple phi = gsi_stmt (gsi);
192 tree src = PHI_ARG_DEF_FROM_EDGE (phi, e);
193 tree dst = gimple_phi_result (phi);
195 /* If the desired argument is not the same as this PHI's result
196 and it is set by a PHI in E->dest, then we can not thread
197 through E->dest. */
198 if (src != dst
199 && TREE_CODE (src) == SSA_NAME
200 && gimple_code (SSA_NAME_DEF_STMT (src)) == GIMPLE_PHI
201 && gimple_bb (SSA_NAME_DEF_STMT (src)) == e->dest)
202 return false;
204 /* We consider any non-virtual PHI as a statement since it
205 count result in a constant assignment or copy operation. */
206 if (!virtual_operand_p (dst))
207 stmt_count++;
209 record_temporary_equivalence (dst, src, stack);
211 return true;
214 /* Fold the RHS of an assignment statement and return it as a tree.
215 May return NULL_TREE if no simplification is possible. */
217 static tree
218 fold_assignment_stmt (gimple stmt)
220 enum tree_code subcode = gimple_assign_rhs_code (stmt);
222 switch (get_gimple_rhs_class (subcode))
224 case GIMPLE_SINGLE_RHS:
225 return fold (gimple_assign_rhs1 (stmt));
227 case GIMPLE_UNARY_RHS:
229 tree lhs = gimple_assign_lhs (stmt);
230 tree op0 = gimple_assign_rhs1 (stmt);
231 return fold_unary (subcode, TREE_TYPE (lhs), op0);
234 case GIMPLE_BINARY_RHS:
236 tree lhs = gimple_assign_lhs (stmt);
237 tree op0 = gimple_assign_rhs1 (stmt);
238 tree op1 = gimple_assign_rhs2 (stmt);
239 return fold_binary (subcode, TREE_TYPE (lhs), op0, op1);
242 case GIMPLE_TERNARY_RHS:
244 tree lhs = gimple_assign_lhs (stmt);
245 tree op0 = gimple_assign_rhs1 (stmt);
246 tree op1 = gimple_assign_rhs2 (stmt);
247 tree op2 = gimple_assign_rhs3 (stmt);
249 /* Sadly, we have to handle conditional assignments specially
250 here, because fold expects all the operands of an expression
251 to be folded before the expression itself is folded, but we
252 can't just substitute the folded condition here. */
253 if (gimple_assign_rhs_code (stmt) == COND_EXPR)
254 op0 = fold (op0);
256 return fold_ternary (subcode, TREE_TYPE (lhs), op0, op1, op2);
259 default:
260 gcc_unreachable ();
264 /* Try to simplify each statement in E->dest, ultimately leading to
265 a simplification of the COND_EXPR at the end of E->dest.
267 Record unwind information for temporary equivalences onto STACK.
269 Use SIMPLIFY (a pointer to a callback function) to further simplify
270 statements using pass specific information.
272 We might consider marking just those statements which ultimately
273 feed the COND_EXPR. It's not clear if the overhead of bookkeeping
274 would be recovered by trying to simplify fewer statements.
276 If we are able to simplify a statement into the form
277 SSA_NAME = (SSA_NAME | gimple invariant), then we can record
278 a context sensitive equivalence which may help us simplify
279 later statements in E->dest. */
281 static gimple
282 record_temporary_equivalences_from_stmts_at_dest (edge e,
283 vec<tree> *stack,
284 tree (*simplify) (gimple,
285 gimple))
287 gimple stmt = NULL;
288 gimple_stmt_iterator gsi;
289 int max_stmt_count;
291 max_stmt_count = PARAM_VALUE (PARAM_MAX_JUMP_THREAD_DUPLICATION_STMTS);
293 /* Walk through each statement in the block recording equivalences
294 we discover. Note any equivalences we discover are context
295 sensitive (ie, are dependent on traversing E) and must be unwound
296 when we're finished processing E. */
297 for (gsi = gsi_start_bb (e->dest); !gsi_end_p (gsi); gsi_next (&gsi))
299 tree cached_lhs = NULL;
301 stmt = gsi_stmt (gsi);
303 /* Ignore empty statements and labels. */
304 if (gimple_code (stmt) == GIMPLE_NOP
305 || gimple_code (stmt) == GIMPLE_LABEL
306 || is_gimple_debug (stmt))
307 continue;
309 /* If the statement has volatile operands, then we assume we
310 can not thread through this block. This is overly
311 conservative in some ways. */
312 if (gimple_code (stmt) == GIMPLE_ASM && gimple_asm_volatile_p (stmt))
313 return NULL;
315 /* If duplicating this block is going to cause too much code
316 expansion, then do not thread through this block. */
317 stmt_count++;
318 if (stmt_count > max_stmt_count)
319 return NULL;
321 /* If this is not a statement that sets an SSA_NAME to a new
322 value, then do not try to simplify this statement as it will
323 not simplify in any way that is helpful for jump threading. */
324 if ((gimple_code (stmt) != GIMPLE_ASSIGN
325 || TREE_CODE (gimple_assign_lhs (stmt)) != SSA_NAME)
326 && (gimple_code (stmt) != GIMPLE_CALL
327 || gimple_call_lhs (stmt) == NULL_TREE
328 || TREE_CODE (gimple_call_lhs (stmt)) != SSA_NAME))
329 continue;
331 /* The result of __builtin_object_size depends on all the arguments
332 of a phi node. Temporarily using only one edge produces invalid
333 results. For example
335 if (x < 6)
336 goto l;
337 else
338 goto l;
341 r = PHI <&w[2].a[1](2), &a.a[6](3)>
342 __builtin_object_size (r, 0)
344 The result of __builtin_object_size is defined to be the maximum of
345 remaining bytes. If we use only one edge on the phi, the result will
346 change to be the remaining bytes for the corresponding phi argument.
348 Similarly for __builtin_constant_p:
350 r = PHI <1(2), 2(3)>
351 __builtin_constant_p (r)
353 Both PHI arguments are constant, but x ? 1 : 2 is still not
354 constant. */
356 if (is_gimple_call (stmt))
358 tree fndecl = gimple_call_fndecl (stmt);
359 if (fndecl
360 && (DECL_FUNCTION_CODE (fndecl) == BUILT_IN_OBJECT_SIZE
361 || DECL_FUNCTION_CODE (fndecl) == BUILT_IN_CONSTANT_P))
362 continue;
365 /* At this point we have a statement which assigns an RHS to an
366 SSA_VAR on the LHS. We want to try and simplify this statement
367 to expose more context sensitive equivalences which in turn may
368 allow us to simplify the condition at the end of the loop.
370 Handle simple copy operations as well as implied copies from
371 ASSERT_EXPRs. */
372 if (gimple_assign_single_p (stmt)
373 && TREE_CODE (gimple_assign_rhs1 (stmt)) == SSA_NAME)
374 cached_lhs = gimple_assign_rhs1 (stmt);
375 else if (gimple_assign_single_p (stmt)
376 && TREE_CODE (gimple_assign_rhs1 (stmt)) == ASSERT_EXPR)
377 cached_lhs = TREE_OPERAND (gimple_assign_rhs1 (stmt), 0);
378 else
380 /* A statement that is not a trivial copy or ASSERT_EXPR.
381 We're going to temporarily copy propagate the operands
382 and see if that allows us to simplify this statement. */
383 tree *copy;
384 ssa_op_iter iter;
385 use_operand_p use_p;
386 unsigned int num, i = 0;
388 num = NUM_SSA_OPERANDS (stmt, (SSA_OP_USE | SSA_OP_VUSE));
389 copy = XCNEWVEC (tree, num);
391 /* Make a copy of the uses & vuses into USES_COPY, then cprop into
392 the operands. */
393 FOR_EACH_SSA_USE_OPERAND (use_p, stmt, iter, SSA_OP_USE | SSA_OP_VUSE)
395 tree tmp = NULL;
396 tree use = USE_FROM_PTR (use_p);
398 copy[i++] = use;
399 if (TREE_CODE (use) == SSA_NAME)
400 tmp = SSA_NAME_VALUE (use);
401 if (tmp)
402 SET_USE (use_p, tmp);
405 /* Try to fold/lookup the new expression. Inserting the
406 expression into the hash table is unlikely to help. */
407 if (is_gimple_call (stmt))
408 cached_lhs = fold_call_stmt (stmt, false);
409 else
410 cached_lhs = fold_assignment_stmt (stmt);
412 if (!cached_lhs
413 || (TREE_CODE (cached_lhs) != SSA_NAME
414 && !is_gimple_min_invariant (cached_lhs)))
415 cached_lhs = (*simplify) (stmt, stmt);
417 /* Restore the statement's original uses/defs. */
418 i = 0;
419 FOR_EACH_SSA_USE_OPERAND (use_p, stmt, iter, SSA_OP_USE | SSA_OP_VUSE)
420 SET_USE (use_p, copy[i++]);
422 free (copy);
425 /* Record the context sensitive equivalence if we were able
426 to simplify this statement. */
427 if (cached_lhs
428 && (TREE_CODE (cached_lhs) == SSA_NAME
429 || is_gimple_min_invariant (cached_lhs)))
430 record_temporary_equivalence (gimple_get_lhs (stmt), cached_lhs, stack);
432 return stmt;
435 /* Simplify the control statement at the end of the block E->dest.
437 To avoid allocating memory unnecessarily, a scratch GIMPLE_COND
438 is available to use/clobber in DUMMY_COND.
440 Use SIMPLIFY (a pointer to a callback function) to further simplify
441 a condition using pass specific information.
443 Return the simplified condition or NULL if simplification could
444 not be performed. */
446 static tree
447 simplify_control_stmt_condition (edge e,
448 gimple stmt,
449 gimple dummy_cond,
450 tree (*simplify) (gimple, gimple),
451 bool handle_dominating_asserts)
453 tree cond, cached_lhs;
454 enum gimple_code code = gimple_code (stmt);
456 /* For comparisons, we have to update both operands, then try
457 to simplify the comparison. */
458 if (code == GIMPLE_COND)
460 tree op0, op1;
461 enum tree_code cond_code;
463 op0 = gimple_cond_lhs (stmt);
464 op1 = gimple_cond_rhs (stmt);
465 cond_code = gimple_cond_code (stmt);
467 /* Get the current value of both operands. */
468 if (TREE_CODE (op0) == SSA_NAME)
470 tree tmp = SSA_NAME_VALUE (op0);
471 if (tmp)
472 op0 = tmp;
475 if (TREE_CODE (op1) == SSA_NAME)
477 tree tmp = SSA_NAME_VALUE (op1);
478 if (tmp)
479 op1 = tmp;
482 if (handle_dominating_asserts)
484 /* Now see if the operand was consumed by an ASSERT_EXPR
485 which dominates E->src. If so, we want to replace the
486 operand with the LHS of the ASSERT_EXPR. */
487 if (TREE_CODE (op0) == SSA_NAME)
488 op0 = lhs_of_dominating_assert (op0, e->src, stmt);
490 if (TREE_CODE (op1) == SSA_NAME)
491 op1 = lhs_of_dominating_assert (op1, e->src, stmt);
494 /* We may need to canonicalize the comparison. For
495 example, op0 might be a constant while op1 is an
496 SSA_NAME. Failure to canonicalize will cause us to
497 miss threading opportunities. */
498 if (tree_swap_operands_p (op0, op1, false))
500 tree tmp;
501 cond_code = swap_tree_comparison (cond_code);
502 tmp = op0;
503 op0 = op1;
504 op1 = tmp;
507 /* Stuff the operator and operands into our dummy conditional
508 expression. */
509 gimple_cond_set_code (dummy_cond, cond_code);
510 gimple_cond_set_lhs (dummy_cond, op0);
511 gimple_cond_set_rhs (dummy_cond, op1);
513 /* We absolutely do not care about any type conversions
514 we only care about a zero/nonzero value. */
515 fold_defer_overflow_warnings ();
517 cached_lhs = fold_binary (cond_code, boolean_type_node, op0, op1);
518 if (cached_lhs)
519 while (CONVERT_EXPR_P (cached_lhs))
520 cached_lhs = TREE_OPERAND (cached_lhs, 0);
522 fold_undefer_overflow_warnings ((cached_lhs
523 && is_gimple_min_invariant (cached_lhs)),
524 stmt, WARN_STRICT_OVERFLOW_CONDITIONAL);
526 /* If we have not simplified the condition down to an invariant,
527 then use the pass specific callback to simplify the condition. */
528 if (!cached_lhs
529 || !is_gimple_min_invariant (cached_lhs))
530 cached_lhs = (*simplify) (dummy_cond, stmt);
532 return cached_lhs;
535 if (code == GIMPLE_SWITCH)
536 cond = gimple_switch_index (stmt);
537 else if (code == GIMPLE_GOTO)
538 cond = gimple_goto_dest (stmt);
539 else
540 gcc_unreachable ();
542 /* We can have conditionals which just test the state of a variable
543 rather than use a relational operator. These are simpler to handle. */
544 if (TREE_CODE (cond) == SSA_NAME)
546 cached_lhs = cond;
548 /* Get the variable's current value from the equivalence chains.
550 It is possible to get loops in the SSA_NAME_VALUE chains
551 (consider threading the backedge of a loop where we have
552 a loop invariant SSA_NAME used in the condition. */
553 if (cached_lhs
554 && TREE_CODE (cached_lhs) == SSA_NAME
555 && SSA_NAME_VALUE (cached_lhs))
556 cached_lhs = SSA_NAME_VALUE (cached_lhs);
558 /* If we're dominated by a suitable ASSERT_EXPR, then
559 update CACHED_LHS appropriately. */
560 if (handle_dominating_asserts && TREE_CODE (cached_lhs) == SSA_NAME)
561 cached_lhs = lhs_of_dominating_assert (cached_lhs, e->src, stmt);
563 /* If we haven't simplified to an invariant yet, then use the
564 pass specific callback to try and simplify it further. */
565 if (cached_lhs && ! is_gimple_min_invariant (cached_lhs))
566 cached_lhs = (*simplify) (stmt, stmt);
568 else
569 cached_lhs = NULL;
571 return cached_lhs;
574 /* Return TRUE if the statement at the end of e->dest depends on
575 the output of any statement in BB. Otherwise return FALSE.
577 This is used when we are threading a backedge and need to ensure
578 that temporary equivalences from BB do not affect the condition
579 in e->dest. */
581 static bool
582 cond_arg_set_in_bb (edge e, basic_block bb)
584 ssa_op_iter iter;
585 use_operand_p use_p;
586 gimple last = last_stmt (e->dest);
588 /* E->dest does not have to end with a control transferring
589 instruction. This can occurr when we try to extend a jump
590 threading opportunity deeper into the CFG. In that case
591 it is safe for this check to return false. */
592 if (!last)
593 return false;
595 if (gimple_code (last) != GIMPLE_COND
596 && gimple_code (last) != GIMPLE_GOTO
597 && gimple_code (last) != GIMPLE_SWITCH)
598 return false;
600 FOR_EACH_SSA_USE_OPERAND (use_p, last, iter, SSA_OP_USE | SSA_OP_VUSE)
602 tree use = USE_FROM_PTR (use_p);
604 if (TREE_CODE (use) == SSA_NAME
605 && gimple_code (SSA_NAME_DEF_STMT (use)) != GIMPLE_PHI
606 && gimple_bb (SSA_NAME_DEF_STMT (use)) == bb)
607 return true;
609 return false;
612 /* Copy debug stmts from DEST's chain of single predecessors up to
613 SRC, so that we don't lose the bindings as PHI nodes are introduced
614 when DEST gains new predecessors. */
615 void
616 propagate_threaded_block_debug_into (basic_block dest, basic_block src)
618 if (!MAY_HAVE_DEBUG_STMTS)
619 return;
621 if (!single_pred_p (dest))
622 return;
624 gcc_checking_assert (dest != src);
626 gimple_stmt_iterator gsi = gsi_after_labels (dest);
627 int i = 0;
628 const int alloc_count = 16; // ?? Should this be a PARAM?
630 /* Estimate the number of debug vars overridden in the beginning of
631 DEST, to tell how many we're going to need to begin with. */
632 for (gimple_stmt_iterator si = gsi;
633 i * 4 <= alloc_count * 3 && !gsi_end_p (si); gsi_next (&si))
635 gimple stmt = gsi_stmt (si);
636 if (!is_gimple_debug (stmt))
637 break;
638 i++;
641 vec<tree, va_stack> fewvars = vec<tree, va_stack>();
642 pointer_set_t *vars = NULL;
644 /* If we're already starting with 3/4 of alloc_count, go for a
645 pointer_set, otherwise start with an unordered stack-allocated
646 VEC. */
647 if (i * 4 > alloc_count * 3)
648 vars = pointer_set_create ();
649 else if (alloc_count)
650 vec_stack_alloc (tree, fewvars, alloc_count);
652 /* Now go through the initial debug stmts in DEST again, this time
653 actually inserting in VARS or FEWVARS. Don't bother checking for
654 duplicates in FEWVARS. */
655 for (gimple_stmt_iterator si = gsi; !gsi_end_p (si); gsi_next (&si))
657 gimple stmt = gsi_stmt (si);
658 if (!is_gimple_debug (stmt))
659 break;
661 tree var;
663 if (gimple_debug_bind_p (stmt))
664 var = gimple_debug_bind_get_var (stmt);
665 else if (gimple_debug_source_bind_p (stmt))
666 var = gimple_debug_source_bind_get_var (stmt);
667 else
668 gcc_unreachable ();
670 if (vars)
671 pointer_set_insert (vars, var);
672 else
673 fewvars.quick_push (var);
676 basic_block bb = dest;
680 bb = single_pred (bb);
681 for (gimple_stmt_iterator si = gsi_last_bb (bb);
682 !gsi_end_p (si); gsi_prev (&si))
684 gimple stmt = gsi_stmt (si);
685 if (!is_gimple_debug (stmt))
686 continue;
688 tree var;
690 if (gimple_debug_bind_p (stmt))
691 var = gimple_debug_bind_get_var (stmt);
692 else if (gimple_debug_source_bind_p (stmt))
693 var = gimple_debug_source_bind_get_var (stmt);
694 else
695 gcc_unreachable ();
697 /* Discard debug bind overlaps. ??? Unlike stmts from src,
698 copied into a new block that will precede BB, debug bind
699 stmts in bypassed BBs may actually be discarded if
700 they're overwritten by subsequent debug bind stmts, which
701 might be a problem once we introduce stmt frontier notes
702 or somesuch. Adding `&& bb == src' to the condition
703 below will preserve all potentially relevant debug
704 notes. */
705 if (vars && pointer_set_insert (vars, var))
706 continue;
707 else if (!vars)
709 int i = fewvars.length ();
710 while (i--)
711 if (fewvars[i] == var)
712 break;
713 if (i >= 0)
714 continue;
716 if (fewvars.length () < alloc_count)
717 fewvars.quick_push (var);
718 else
720 vars = pointer_set_create ();
721 for (i = 0; i < alloc_count; i++)
722 pointer_set_insert (vars, fewvars[i]);
723 fewvars.release ();
724 pointer_set_insert (vars, var);
728 stmt = gimple_copy (stmt);
729 /* ??? Should we drop the location of the copy to denote
730 they're artificial bindings? */
731 gsi_insert_before (&gsi, stmt, GSI_NEW_STMT);
734 while (bb != src && single_pred_p (bb));
736 if (vars)
737 pointer_set_destroy (vars);
738 else if (fewvars.exists ())
739 fewvars.release ();
742 /* TAKEN_EDGE represents the an edge taken as a result of jump threading.
743 See if we can thread around TAKEN_EDGE->dest as well. If so, return
744 the edge out of TAKEN_EDGE->dest that we can statically compute will be
745 traversed.
747 We are much more restrictive as to the contents of TAKEN_EDGE->dest
748 as the path isolation code in tree-ssa-threadupdate.c isn't prepared
749 to handle copying intermediate blocks on a threaded path.
751 Long term a more consistent and structured approach to path isolation
752 would be a huge help. */
753 static edge
754 thread_around_empty_block (edge taken_edge,
755 gimple dummy_cond,
756 bool handle_dominating_asserts,
757 tree (*simplify) (gimple, gimple),
758 bitmap visited)
760 basic_block bb = taken_edge->dest;
761 gimple_stmt_iterator gsi;
762 gimple stmt;
763 tree cond;
765 /* This block must have a single predecessor (E->dest). */
766 if (!single_pred_p (bb))
767 return NULL;
769 /* This block must have more than one successor. */
770 if (single_succ_p (bb))
771 return NULL;
773 /* This block can have no PHI nodes. This is overly conservative. */
774 if (!gsi_end_p (gsi_start_phis (bb)))
775 return NULL;
777 /* Skip over DEBUG statements at the start of the block. */
778 gsi = gsi_start_nondebug_bb (bb);
780 if (gsi_end_p (gsi))
781 return NULL;
783 /* This block can have no statements other than its control altering
784 statement. This is overly conservative. */
785 stmt = gsi_stmt (gsi);
786 if (gimple_code (stmt) != GIMPLE_COND
787 && gimple_code (stmt) != GIMPLE_GOTO
788 && gimple_code (stmt) != GIMPLE_SWITCH)
789 return NULL;
791 /* Extract and simplify the condition. */
792 cond = simplify_control_stmt_condition (taken_edge, stmt, dummy_cond,
793 simplify, handle_dominating_asserts);
795 /* If the condition can be statically computed and we have not already
796 visited the destination edge, then add the taken edge to our thread
797 path. */
798 if (cond && is_gimple_min_invariant (cond))
800 edge taken_edge = find_taken_edge (bb, cond);
802 if (bitmap_bit_p (visited, taken_edge->dest->index))
803 return NULL;
804 bitmap_set_bit (visited, taken_edge->dest->index);
805 return taken_edge;
808 return NULL;
811 /* E1 and E2 are edges into the same basic block. Return TRUE if the
812 PHI arguments associated with those edges are equal or there are no
813 PHI arguments, otherwise return FALSE. */
815 static bool
816 phi_args_equal_on_edges (edge e1, edge e2)
818 gimple_stmt_iterator gsi;
819 int indx1 = e1->dest_idx;
820 int indx2 = e2->dest_idx;
822 for (gsi = gsi_start_phis (e1->dest); !gsi_end_p (gsi); gsi_next (&gsi))
824 gimple phi = gsi_stmt (gsi);
826 if (!operand_equal_p (gimple_phi_arg_def (phi, indx1),
827 gimple_phi_arg_def (phi, indx2), 0))
828 return false;
830 return true;
833 /* We are exiting E->src, see if E->dest ends with a conditional
834 jump which has a known value when reached via E.
836 Special care is necessary if E is a back edge in the CFG as we
837 may have already recorded equivalences for E->dest into our
838 various tables, including the result of the conditional at
839 the end of E->dest. Threading opportunities are severely
840 limited in that case to avoid short-circuiting the loop
841 incorrectly.
843 Note it is quite common for the first block inside a loop to
844 end with a conditional which is either always true or always
845 false when reached via the loop backedge. Thus we do not want
846 to blindly disable threading across a loop backedge.
848 DUMMY_COND is a shared cond_expr used by condition simplification as scratch,
849 to avoid allocating memory.
851 HANDLE_DOMINATING_ASSERTS is true if we should try to replace operands of
852 the simplified condition with left-hand sides of ASSERT_EXPRs they are
853 used in.
855 STACK is used to undo temporary equivalences created during the walk of
856 E->dest.
858 SIMPLIFY is a pass-specific function used to simplify statements. */
860 void
861 thread_across_edge (gimple dummy_cond,
862 edge e,
863 bool handle_dominating_asserts,
864 vec<tree> *stack,
865 tree (*simplify) (gimple, gimple))
867 gimple stmt;
869 /* If E is a backedge, then we want to verify that the COND_EXPR,
870 SWITCH_EXPR or GOTO_EXPR at the end of e->dest is not affected
871 by any statements in e->dest. If it is affected, then it is not
872 safe to thread this edge. */
873 if (e->flags & EDGE_DFS_BACK)
875 if (cond_arg_set_in_bb (e, e->dest))
876 goto fail;
879 stmt_count = 0;
881 /* PHIs create temporary equivalences. */
882 if (!record_temporary_equivalences_from_phis (e, stack))
883 goto fail;
885 /* Now walk each statement recording any context sensitive
886 temporary equivalences we can detect. */
887 stmt = record_temporary_equivalences_from_stmts_at_dest (e, stack, simplify);
888 if (!stmt)
889 goto fail;
891 /* If we stopped at a COND_EXPR or SWITCH_EXPR, see if we know which arm
892 will be taken. */
893 if (gimple_code (stmt) == GIMPLE_COND
894 || gimple_code (stmt) == GIMPLE_GOTO
895 || gimple_code (stmt) == GIMPLE_SWITCH)
897 tree cond;
899 /* Extract and simplify the condition. */
900 cond = simplify_control_stmt_condition (e, stmt, dummy_cond, simplify,
901 handle_dominating_asserts);
903 if (cond && is_gimple_min_invariant (cond))
905 edge taken_edge = find_taken_edge (e->dest, cond);
906 basic_block dest = (taken_edge ? taken_edge->dest : NULL);
907 bitmap visited;
908 edge e2;
910 if (dest == e->dest)
911 goto fail;
913 /* DEST could be null for a computed jump to an absolute
914 address. If DEST is not null, then see if we can thread
915 through it as well, this helps capture secondary effects
916 of threading without having to re-run DOM or VRP. */
917 if (dest
918 && ((e->flags & EDGE_DFS_BACK) == 0
919 || ! cond_arg_set_in_bb (taken_edge, e->dest)))
921 /* We don't want to thread back to a block we have already
922 visited. This may be overly conservative. */
923 visited = BITMAP_ALLOC (NULL);
924 bitmap_set_bit (visited, dest->index);
925 bitmap_set_bit (visited, e->dest->index);
928 e2 = thread_around_empty_block (taken_edge,
929 dummy_cond,
930 handle_dominating_asserts,
931 simplify,
932 visited);
933 if (e2)
934 taken_edge = e2;
936 while (e2);
937 BITMAP_FREE (visited);
940 remove_temporary_equivalences (stack);
941 if (!taken_edge)
942 return;
943 propagate_threaded_block_debug_into (taken_edge->dest, e->dest);
944 register_jump_thread (e, taken_edge, NULL);
945 return;
949 /* We were unable to determine what out edge from E->dest is taken. However,
950 we might still be able to thread through successors of E->dest. This
951 often occurs when E->dest is a joiner block which then fans back out
952 based on redundant tests.
954 If so, we'll copy E->dest and redirect the appropriate predecessor to
955 the copy. Within the copy of E->dest, we'll thread one or more edges
956 to points deeper in the CFG.
958 This is a stopgap until we have a more structured approach to path
959 isolation. */
961 edge e2, e3, taken_edge;
962 edge_iterator ei;
963 bool found = false;
964 bitmap visited = BITMAP_ALLOC (NULL);
966 /* Look at each successor of E->dest to see if we can thread through it. */
967 FOR_EACH_EDGE (taken_edge, ei, e->dest->succs)
969 /* Avoid threading to any block we have already visited. */
970 bitmap_clear (visited);
971 bitmap_set_bit (visited, taken_edge->dest->index);
972 bitmap_set_bit (visited, e->dest->index);
974 /* Record whether or not we were able to thread through a successor
975 of E->dest. */
976 found = false;
977 e3 = taken_edge;
980 if ((e->flags & EDGE_DFS_BACK) == 0
981 || ! cond_arg_set_in_bb (e3, e->dest))
982 e2 = thread_around_empty_block (e3,
983 dummy_cond,
984 handle_dominating_asserts,
985 simplify,
986 visited);
987 else
988 e2 = NULL;
990 if (e2)
992 e3 = e2;
993 found = true;
996 while (e2);
998 /* If we were able to thread through a successor of E->dest, then
999 record the jump threading opportunity. */
1000 if (found)
1002 edge tmp;
1003 /* If there is already an edge from the block to be duplicated
1004 (E2->src) to the final target (E3->dest), then make sure that
1005 the PHI args associated with the edges E2 and E3 are the
1006 same. */
1007 tmp = find_edge (taken_edge->src, e3->dest);
1008 if (!tmp || phi_args_equal_on_edges (tmp, e3))
1010 propagate_threaded_block_debug_into (e3->dest,
1011 taken_edge->dest);
1012 register_jump_thread (e, taken_edge, e3);
1017 BITMAP_FREE (visited);
1020 fail:
1021 remove_temporary_equivalences (stack);