2 Copyright (C) 2005-2013 Free Software Foundation, Inc.
3 Contributed by Jeff Law <law@redhat.com>
5 This file is part of GCC.
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)
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.
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/>. */
23 #include "coretypes.h"
28 #include "basic-block.h"
34 #include "gimple-ssa.h"
36 #include "tree-phinodes.h"
37 #include "ssa-iterators.h"
38 #include "tree-ssanames.h"
39 #include "tree-ssa-propagate.h"
40 #include "tree-ssa-threadupdate.h"
41 #include "langhooks.h"
43 #include "tree-ssa-threadedge.h"
45 /* To avoid code explosion due to jump threading, we limit the
46 number of statements we are going to copy. This variable
47 holds the number of statements currently seen that we'll have
48 to copy as part of the jump threading process. */
49 static int stmt_count
;
51 /* Array to record value-handles per SSA_NAME. */
52 vec
<tree
> ssa_name_values
;
54 /* Set the value for the SSA name NAME to VALUE. */
57 set_ssa_name_value (tree name
, tree value
)
59 if (SSA_NAME_VERSION (name
) >= ssa_name_values
.length ())
60 ssa_name_values
.safe_grow_cleared (SSA_NAME_VERSION (name
) + 1);
61 ssa_name_values
[SSA_NAME_VERSION (name
)] = value
;
64 /* Initialize the per SSA_NAME value-handles array. Returns it. */
66 threadedge_initialize_values (void)
68 gcc_assert (!ssa_name_values
.exists ());
69 ssa_name_values
.create (num_ssa_names
);
72 /* Free the per SSA_NAME value-handle array. */
74 threadedge_finalize_values (void)
76 ssa_name_values
.release ();
79 /* Return TRUE if we may be able to thread an incoming edge into
80 BB to an outgoing edge from BB. Return FALSE otherwise. */
83 potentially_threadable_block (basic_block bb
)
85 gimple_stmt_iterator gsi
;
87 /* If BB has a single successor or a single predecessor, then
88 there is no threading opportunity. */
89 if (single_succ_p (bb
) || single_pred_p (bb
))
92 /* If BB does not end with a conditional, switch or computed goto,
93 then there is no threading opportunity. */
94 gsi
= gsi_last_bb (bb
);
97 || (gimple_code (gsi_stmt (gsi
)) != GIMPLE_COND
98 && gimple_code (gsi_stmt (gsi
)) != GIMPLE_GOTO
99 && gimple_code (gsi_stmt (gsi
)) != GIMPLE_SWITCH
))
105 /* Return the LHS of any ASSERT_EXPR where OP appears as the first
106 argument to the ASSERT_EXPR and in which the ASSERT_EXPR dominates
107 BB. If no such ASSERT_EXPR is found, return OP. */
110 lhs_of_dominating_assert (tree op
, basic_block bb
, gimple stmt
)
112 imm_use_iterator imm_iter
;
116 FOR_EACH_IMM_USE_FAST (use_p
, imm_iter
, op
)
118 use_stmt
= USE_STMT (use_p
);
120 && gimple_assign_single_p (use_stmt
)
121 && TREE_CODE (gimple_assign_rhs1 (use_stmt
)) == ASSERT_EXPR
122 && TREE_OPERAND (gimple_assign_rhs1 (use_stmt
), 0) == op
123 && dominated_by_p (CDI_DOMINATORS
, bb
, gimple_bb (use_stmt
)))
125 return gimple_assign_lhs (use_stmt
);
131 /* We record temporary equivalences created by PHI nodes or
132 statements within the target block. Doing so allows us to
133 identify more jump threading opportunities, even in blocks
136 We keep track of those temporary equivalences in a stack
137 structure so that we can unwind them when we're done processing
138 a particular edge. This routine handles unwinding the data
142 remove_temporary_equivalences (vec
<tree
> *stack
)
144 while (stack
->length () > 0)
146 tree prev_value
, dest
;
148 dest
= stack
->pop ();
150 /* A NULL value indicates we should stop unwinding, otherwise
151 pop off the next entry as they're recorded in pairs. */
155 prev_value
= stack
->pop ();
156 set_ssa_name_value (dest
, prev_value
);
160 /* Record a temporary equivalence, saving enough information so that
161 we can restore the state of recorded equivalences when we're
162 done processing the current edge. */
165 record_temporary_equivalence (tree x
, tree y
, vec
<tree
> *stack
)
167 tree prev_x
= SSA_NAME_VALUE (x
);
169 if (TREE_CODE (y
) == SSA_NAME
)
171 tree tmp
= SSA_NAME_VALUE (y
);
175 set_ssa_name_value (x
, y
);
177 stack
->quick_push (prev_x
);
178 stack
->quick_push (x
);
181 /* Record temporary equivalences created by PHIs at the target of the
182 edge E. Record unwind information for the equivalences onto STACK.
184 If a PHI which prevents threading is encountered, then return FALSE
185 indicating we should not thread this edge, else return TRUE. */
188 record_temporary_equivalences_from_phis (edge e
, vec
<tree
> *stack
)
190 gimple_stmt_iterator gsi
;
192 /* Each PHI creates a temporary equivalence, record them.
193 These are context sensitive equivalences and will be removed
195 for (gsi
= gsi_start_phis (e
->dest
); !gsi_end_p (gsi
); gsi_next (&gsi
))
197 gimple phi
= gsi_stmt (gsi
);
198 tree src
= PHI_ARG_DEF_FROM_EDGE (phi
, e
);
199 tree dst
= gimple_phi_result (phi
);
201 /* If the desired argument is not the same as this PHI's result
202 and it is set by a PHI in E->dest, then we can not thread
205 && TREE_CODE (src
) == SSA_NAME
206 && gimple_code (SSA_NAME_DEF_STMT (src
)) == GIMPLE_PHI
207 && gimple_bb (SSA_NAME_DEF_STMT (src
)) == e
->dest
)
210 /* We consider any non-virtual PHI as a statement since it
211 count result in a constant assignment or copy operation. */
212 if (!virtual_operand_p (dst
))
215 record_temporary_equivalence (dst
, src
, stack
);
220 /* Fold the RHS of an assignment statement and return it as a tree.
221 May return NULL_TREE if no simplification is possible. */
224 fold_assignment_stmt (gimple stmt
)
226 enum tree_code subcode
= gimple_assign_rhs_code (stmt
);
228 switch (get_gimple_rhs_class (subcode
))
230 case GIMPLE_SINGLE_RHS
:
231 return fold (gimple_assign_rhs1 (stmt
));
233 case GIMPLE_UNARY_RHS
:
235 tree lhs
= gimple_assign_lhs (stmt
);
236 tree op0
= gimple_assign_rhs1 (stmt
);
237 return fold_unary (subcode
, TREE_TYPE (lhs
), op0
);
240 case GIMPLE_BINARY_RHS
:
242 tree lhs
= gimple_assign_lhs (stmt
);
243 tree op0
= gimple_assign_rhs1 (stmt
);
244 tree op1
= gimple_assign_rhs2 (stmt
);
245 return fold_binary (subcode
, TREE_TYPE (lhs
), op0
, op1
);
248 case GIMPLE_TERNARY_RHS
:
250 tree lhs
= gimple_assign_lhs (stmt
);
251 tree op0
= gimple_assign_rhs1 (stmt
);
252 tree op1
= gimple_assign_rhs2 (stmt
);
253 tree op2
= gimple_assign_rhs3 (stmt
);
255 /* Sadly, we have to handle conditional assignments specially
256 here, because fold expects all the operands of an expression
257 to be folded before the expression itself is folded, but we
258 can't just substitute the folded condition here. */
259 if (gimple_assign_rhs_code (stmt
) == COND_EXPR
)
262 return fold_ternary (subcode
, TREE_TYPE (lhs
), op0
, op1
, op2
);
270 /* Try to simplify each statement in E->dest, ultimately leading to
271 a simplification of the COND_EXPR at the end of E->dest.
273 Record unwind information for temporary equivalences onto STACK.
275 Use SIMPLIFY (a pointer to a callback function) to further simplify
276 statements using pass specific information.
278 We might consider marking just those statements which ultimately
279 feed the COND_EXPR. It's not clear if the overhead of bookkeeping
280 would be recovered by trying to simplify fewer statements.
282 If we are able to simplify a statement into the form
283 SSA_NAME = (SSA_NAME | gimple invariant), then we can record
284 a context sensitive equivalence which may help us simplify
285 later statements in E->dest. */
288 record_temporary_equivalences_from_stmts_at_dest (edge e
,
290 tree (*simplify
) (gimple
,
294 gimple_stmt_iterator gsi
;
297 max_stmt_count
= PARAM_VALUE (PARAM_MAX_JUMP_THREAD_DUPLICATION_STMTS
);
299 /* Walk through each statement in the block recording equivalences
300 we discover. Note any equivalences we discover are context
301 sensitive (ie, are dependent on traversing E) and must be unwound
302 when we're finished processing E. */
303 for (gsi
= gsi_start_bb (e
->dest
); !gsi_end_p (gsi
); gsi_next (&gsi
))
305 tree cached_lhs
= NULL
;
307 stmt
= gsi_stmt (gsi
);
309 /* Ignore empty statements and labels. */
310 if (gimple_code (stmt
) == GIMPLE_NOP
311 || gimple_code (stmt
) == GIMPLE_LABEL
312 || is_gimple_debug (stmt
))
315 /* If the statement has volatile operands, then we assume we
316 can not thread through this block. This is overly
317 conservative in some ways. */
318 if (gimple_code (stmt
) == GIMPLE_ASM
&& gimple_asm_volatile_p (stmt
))
321 /* If duplicating this block is going to cause too much code
322 expansion, then do not thread through this block. */
324 if (stmt_count
> max_stmt_count
)
327 /* If this is not a statement that sets an SSA_NAME to a new
328 value, then do not try to simplify this statement as it will
329 not simplify in any way that is helpful for jump threading. */
330 if ((gimple_code (stmt
) != GIMPLE_ASSIGN
331 || TREE_CODE (gimple_assign_lhs (stmt
)) != SSA_NAME
)
332 && (gimple_code (stmt
) != GIMPLE_CALL
333 || gimple_call_lhs (stmt
) == NULL_TREE
334 || TREE_CODE (gimple_call_lhs (stmt
)) != SSA_NAME
))
337 /* The result of __builtin_object_size depends on all the arguments
338 of a phi node. Temporarily using only one edge produces invalid
347 r = PHI <&w[2].a[1](2), &a.a[6](3)>
348 __builtin_object_size (r, 0)
350 The result of __builtin_object_size is defined to be the maximum of
351 remaining bytes. If we use only one edge on the phi, the result will
352 change to be the remaining bytes for the corresponding phi argument.
354 Similarly for __builtin_constant_p:
357 __builtin_constant_p (r)
359 Both PHI arguments are constant, but x ? 1 : 2 is still not
362 if (is_gimple_call (stmt
))
364 tree fndecl
= gimple_call_fndecl (stmt
);
366 && (DECL_FUNCTION_CODE (fndecl
) == BUILT_IN_OBJECT_SIZE
367 || DECL_FUNCTION_CODE (fndecl
) == BUILT_IN_CONSTANT_P
))
371 /* At this point we have a statement which assigns an RHS to an
372 SSA_VAR on the LHS. We want to try and simplify this statement
373 to expose more context sensitive equivalences which in turn may
374 allow us to simplify the condition at the end of the loop.
376 Handle simple copy operations as well as implied copies from
378 if (gimple_assign_single_p (stmt
)
379 && TREE_CODE (gimple_assign_rhs1 (stmt
)) == SSA_NAME
)
380 cached_lhs
= gimple_assign_rhs1 (stmt
);
381 else if (gimple_assign_single_p (stmt
)
382 && TREE_CODE (gimple_assign_rhs1 (stmt
)) == ASSERT_EXPR
)
383 cached_lhs
= TREE_OPERAND (gimple_assign_rhs1 (stmt
), 0);
386 /* A statement that is not a trivial copy or ASSERT_EXPR.
387 We're going to temporarily copy propagate the operands
388 and see if that allows us to simplify this statement. */
392 unsigned int num
, i
= 0;
394 num
= NUM_SSA_OPERANDS (stmt
, (SSA_OP_USE
| SSA_OP_VUSE
));
395 copy
= XCNEWVEC (tree
, num
);
397 /* Make a copy of the uses & vuses into USES_COPY, then cprop into
399 FOR_EACH_SSA_USE_OPERAND (use_p
, stmt
, iter
, SSA_OP_USE
| SSA_OP_VUSE
)
402 tree use
= USE_FROM_PTR (use_p
);
405 if (TREE_CODE (use
) == SSA_NAME
)
406 tmp
= SSA_NAME_VALUE (use
);
408 SET_USE (use_p
, tmp
);
411 /* Try to fold/lookup the new expression. Inserting the
412 expression into the hash table is unlikely to help. */
413 if (is_gimple_call (stmt
))
414 cached_lhs
= fold_call_stmt (stmt
, false);
416 cached_lhs
= fold_assignment_stmt (stmt
);
419 || (TREE_CODE (cached_lhs
) != SSA_NAME
420 && !is_gimple_min_invariant (cached_lhs
)))
421 cached_lhs
= (*simplify
) (stmt
, stmt
);
423 /* Restore the statement's original uses/defs. */
425 FOR_EACH_SSA_USE_OPERAND (use_p
, stmt
, iter
, SSA_OP_USE
| SSA_OP_VUSE
)
426 SET_USE (use_p
, copy
[i
++]);
431 /* Record the context sensitive equivalence if we were able
432 to simplify this statement. */
434 && (TREE_CODE (cached_lhs
) == SSA_NAME
435 || is_gimple_min_invariant (cached_lhs
)))
436 record_temporary_equivalence (gimple_get_lhs (stmt
), cached_lhs
, stack
);
441 /* Simplify the control statement at the end of the block E->dest.
443 To avoid allocating memory unnecessarily, a scratch GIMPLE_COND
444 is available to use/clobber in DUMMY_COND.
446 Use SIMPLIFY (a pointer to a callback function) to further simplify
447 a condition using pass specific information.
449 Return the simplified condition or NULL if simplification could
453 simplify_control_stmt_condition (edge e
,
456 tree (*simplify
) (gimple
, gimple
),
457 bool handle_dominating_asserts
)
459 tree cond
, cached_lhs
;
460 enum gimple_code code
= gimple_code (stmt
);
462 /* For comparisons, we have to update both operands, then try
463 to simplify the comparison. */
464 if (code
== GIMPLE_COND
)
467 enum tree_code cond_code
;
469 op0
= gimple_cond_lhs (stmt
);
470 op1
= gimple_cond_rhs (stmt
);
471 cond_code
= gimple_cond_code (stmt
);
473 /* Get the current value of both operands. */
474 if (TREE_CODE (op0
) == SSA_NAME
)
476 tree tmp
= SSA_NAME_VALUE (op0
);
481 if (TREE_CODE (op1
) == SSA_NAME
)
483 tree tmp
= SSA_NAME_VALUE (op1
);
488 if (handle_dominating_asserts
)
490 /* Now see if the operand was consumed by an ASSERT_EXPR
491 which dominates E->src. If so, we want to replace the
492 operand with the LHS of the ASSERT_EXPR. */
493 if (TREE_CODE (op0
) == SSA_NAME
)
494 op0
= lhs_of_dominating_assert (op0
, e
->src
, stmt
);
496 if (TREE_CODE (op1
) == SSA_NAME
)
497 op1
= lhs_of_dominating_assert (op1
, e
->src
, stmt
);
500 /* We may need to canonicalize the comparison. For
501 example, op0 might be a constant while op1 is an
502 SSA_NAME. Failure to canonicalize will cause us to
503 miss threading opportunities. */
504 if (tree_swap_operands_p (op0
, op1
, false))
507 cond_code
= swap_tree_comparison (cond_code
);
513 /* Stuff the operator and operands into our dummy conditional
515 gimple_cond_set_code (dummy_cond
, cond_code
);
516 gimple_cond_set_lhs (dummy_cond
, op0
);
517 gimple_cond_set_rhs (dummy_cond
, op1
);
519 /* We absolutely do not care about any type conversions
520 we only care about a zero/nonzero value. */
521 fold_defer_overflow_warnings ();
523 cached_lhs
= fold_binary (cond_code
, boolean_type_node
, op0
, op1
);
525 while (CONVERT_EXPR_P (cached_lhs
))
526 cached_lhs
= TREE_OPERAND (cached_lhs
, 0);
528 fold_undefer_overflow_warnings ((cached_lhs
529 && is_gimple_min_invariant (cached_lhs
)),
530 stmt
, WARN_STRICT_OVERFLOW_CONDITIONAL
);
532 /* If we have not simplified the condition down to an invariant,
533 then use the pass specific callback to simplify the condition. */
535 || !is_gimple_min_invariant (cached_lhs
))
536 cached_lhs
= (*simplify
) (dummy_cond
, stmt
);
541 if (code
== GIMPLE_SWITCH
)
542 cond
= gimple_switch_index (stmt
);
543 else if (code
== GIMPLE_GOTO
)
544 cond
= gimple_goto_dest (stmt
);
548 /* We can have conditionals which just test the state of a variable
549 rather than use a relational operator. These are simpler to handle. */
550 if (TREE_CODE (cond
) == SSA_NAME
)
554 /* Get the variable's current value from the equivalence chains.
556 It is possible to get loops in the SSA_NAME_VALUE chains
557 (consider threading the backedge of a loop where we have
558 a loop invariant SSA_NAME used in the condition. */
560 && TREE_CODE (cached_lhs
) == SSA_NAME
561 && SSA_NAME_VALUE (cached_lhs
))
562 cached_lhs
= SSA_NAME_VALUE (cached_lhs
);
564 /* If we're dominated by a suitable ASSERT_EXPR, then
565 update CACHED_LHS appropriately. */
566 if (handle_dominating_asserts
&& TREE_CODE (cached_lhs
) == SSA_NAME
)
567 cached_lhs
= lhs_of_dominating_assert (cached_lhs
, e
->src
, stmt
);
569 /* If we haven't simplified to an invariant yet, then use the
570 pass specific callback to try and simplify it further. */
571 if (cached_lhs
&& ! is_gimple_min_invariant (cached_lhs
))
572 cached_lhs
= (*simplify
) (stmt
, stmt
);
580 /* Return TRUE if the statement at the end of e->dest depends on
581 the output of any statement in BB. Otherwise return FALSE.
583 This is used when we are threading a backedge and need to ensure
584 that temporary equivalences from BB do not affect the condition
588 cond_arg_set_in_bb (edge e
, basic_block bb
)
592 gimple last
= last_stmt (e
->dest
);
594 /* E->dest does not have to end with a control transferring
595 instruction. This can occur when we try to extend a jump
596 threading opportunity deeper into the CFG. In that case
597 it is safe for this check to return false. */
601 if (gimple_code (last
) != GIMPLE_COND
602 && gimple_code (last
) != GIMPLE_GOTO
603 && gimple_code (last
) != GIMPLE_SWITCH
)
606 FOR_EACH_SSA_USE_OPERAND (use_p
, last
, iter
, SSA_OP_USE
| SSA_OP_VUSE
)
608 tree use
= USE_FROM_PTR (use_p
);
610 if (TREE_CODE (use
) == SSA_NAME
611 && gimple_code (SSA_NAME_DEF_STMT (use
)) != GIMPLE_PHI
612 && gimple_bb (SSA_NAME_DEF_STMT (use
)) == bb
)
618 /* Copy debug stmts from DEST's chain of single predecessors up to
619 SRC, so that we don't lose the bindings as PHI nodes are introduced
620 when DEST gains new predecessors. */
622 propagate_threaded_block_debug_into (basic_block dest
, basic_block src
)
624 if (!MAY_HAVE_DEBUG_STMTS
)
627 if (!single_pred_p (dest
))
630 gcc_checking_assert (dest
!= src
);
632 gimple_stmt_iterator gsi
= gsi_after_labels (dest
);
634 const int alloc_count
= 16; // ?? Should this be a PARAM?
636 /* Estimate the number of debug vars overridden in the beginning of
637 DEST, to tell how many we're going to need to begin with. */
638 for (gimple_stmt_iterator si
= gsi
;
639 i
* 4 <= alloc_count
* 3 && !gsi_end_p (si
); gsi_next (&si
))
641 gimple stmt
= gsi_stmt (si
);
642 if (!is_gimple_debug (stmt
))
647 vec
<tree
, va_stack
> fewvars
= vNULL
;
648 pointer_set_t
*vars
= NULL
;
650 /* If we're already starting with 3/4 of alloc_count, go for a
651 pointer_set, otherwise start with an unordered stack-allocated
653 if (i
* 4 > alloc_count
* 3)
654 vars
= pointer_set_create ();
655 else if (alloc_count
)
656 vec_stack_alloc (tree
, fewvars
, alloc_count
);
658 /* Now go through the initial debug stmts in DEST again, this time
659 actually inserting in VARS or FEWVARS. Don't bother checking for
660 duplicates in FEWVARS. */
661 for (gimple_stmt_iterator si
= gsi
; !gsi_end_p (si
); gsi_next (&si
))
663 gimple stmt
= gsi_stmt (si
);
664 if (!is_gimple_debug (stmt
))
669 if (gimple_debug_bind_p (stmt
))
670 var
= gimple_debug_bind_get_var (stmt
);
671 else if (gimple_debug_source_bind_p (stmt
))
672 var
= gimple_debug_source_bind_get_var (stmt
);
677 pointer_set_insert (vars
, var
);
679 fewvars
.quick_push (var
);
682 basic_block bb
= dest
;
686 bb
= single_pred (bb
);
687 for (gimple_stmt_iterator si
= gsi_last_bb (bb
);
688 !gsi_end_p (si
); gsi_prev (&si
))
690 gimple stmt
= gsi_stmt (si
);
691 if (!is_gimple_debug (stmt
))
696 if (gimple_debug_bind_p (stmt
))
697 var
= gimple_debug_bind_get_var (stmt
);
698 else if (gimple_debug_source_bind_p (stmt
))
699 var
= gimple_debug_source_bind_get_var (stmt
);
703 /* Discard debug bind overlaps. ??? Unlike stmts from src,
704 copied into a new block that will precede BB, debug bind
705 stmts in bypassed BBs may actually be discarded if
706 they're overwritten by subsequent debug bind stmts, which
707 might be a problem once we introduce stmt frontier notes
708 or somesuch. Adding `&& bb == src' to the condition
709 below will preserve all potentially relevant debug
711 if (vars
&& pointer_set_insert (vars
, var
))
715 int i
= fewvars
.length ();
717 if (fewvars
[i
] == var
)
722 if (fewvars
.length () < (unsigned) alloc_count
)
723 fewvars
.quick_push (var
);
726 vars
= pointer_set_create ();
727 for (i
= 0; i
< alloc_count
; i
++)
728 pointer_set_insert (vars
, fewvars
[i
]);
730 pointer_set_insert (vars
, var
);
734 stmt
= gimple_copy (stmt
);
735 /* ??? Should we drop the location of the copy to denote
736 they're artificial bindings? */
737 gsi_insert_before (&gsi
, stmt
, GSI_NEW_STMT
);
740 while (bb
!= src
&& single_pred_p (bb
));
743 pointer_set_destroy (vars
);
744 else if (fewvars
.exists ())
748 /* See if TAKEN_EDGE->dest is a threadable block with no side effecs (ie, it
749 need not be duplicated as part of the CFG/SSA updating process).
751 If it is threadable, add it to PATH and VISITED and recurse, ultimately
752 returning TRUE from the toplevel call. Otherwise do nothing and
755 DUMMY_COND, HANDLE_DOMINATING_ASSERTS and SIMPLIFY are used to
756 try and simplify the condition at the end of TAKEN_EDGE->dest. */
758 thread_around_empty_blocks (edge taken_edge
,
760 bool handle_dominating_asserts
,
761 tree (*simplify
) (gimple
, gimple
),
763 vec
<jump_thread_edge
*> *path
)
765 basic_block bb
= taken_edge
->dest
;
766 gimple_stmt_iterator gsi
;
770 /* The key property of these blocks is that they need not be duplicated
771 when threading. Thus they can not have visible side effects such
773 if (!gsi_end_p (gsi_start_phis (bb
)))
776 /* Skip over DEBUG statements at the start of the block. */
777 gsi
= gsi_start_nondebug_bb (bb
);
779 /* If the block has no statements, but does have a single successor, then
780 it's just a forwarding block and we can thread through it trivially.
782 However, note that just threading through empty blocks with single
783 successors is not inherently profitable. For the jump thread to
784 be profitable, we must avoid a runtime conditional.
786 By taking the return value from the recursive call, we get the
787 desired effect of returning TRUE when we found a profitable jump
788 threading opportunity and FALSE otherwise.
790 This is particularly important when this routine is called after
791 processing a joiner block. Returning TRUE too aggressively in
792 that case results in pointless duplication of the joiner block. */
795 if (single_succ_p (bb
))
797 taken_edge
= single_succ_edge (bb
);
798 if ((taken_edge
->flags
& EDGE_DFS_BACK
) == 0
799 && !bitmap_bit_p (visited
, taken_edge
->dest
->index
))
802 = new jump_thread_edge (taken_edge
, EDGE_NO_COPY_SRC_BLOCK
);
804 bitmap_set_bit (visited
, taken_edge
->dest
->index
);
805 return thread_around_empty_blocks (taken_edge
,
807 handle_dominating_asserts
,
814 /* We have a block with no statements, but multiple successors? */
818 /* The only real statements this block can have are a control
819 flow altering statement. Anything else stops the thread. */
820 stmt
= gsi_stmt (gsi
);
821 if (gimple_code (stmt
) != GIMPLE_COND
822 && gimple_code (stmt
) != GIMPLE_GOTO
823 && gimple_code (stmt
) != GIMPLE_SWITCH
)
826 /* Extract and simplify the condition. */
827 cond
= simplify_control_stmt_condition (taken_edge
, stmt
, dummy_cond
,
828 simplify
, handle_dominating_asserts
);
830 /* If the condition can be statically computed and we have not already
831 visited the destination edge, then add the taken edge to our thread
833 if (cond
&& is_gimple_min_invariant (cond
))
835 taken_edge
= find_taken_edge (bb
, cond
);
837 if (bitmap_bit_p (visited
, taken_edge
->dest
->index
))
839 bitmap_set_bit (visited
, taken_edge
->dest
->index
);
842 = new jump_thread_edge (taken_edge
, EDGE_NO_COPY_SRC_BLOCK
);
845 thread_around_empty_blocks (taken_edge
,
847 handle_dominating_asserts
,
857 /* We are exiting E->src, see if E->dest ends with a conditional
858 jump which has a known value when reached via E.
860 E->dest can have arbitrary side effects which, if threading is
861 successful, will be maintained.
863 Special care is necessary if E is a back edge in the CFG as we
864 may have already recorded equivalences for E->dest into our
865 various tables, including the result of the conditional at
866 the end of E->dest. Threading opportunities are severely
867 limited in that case to avoid short-circuiting the loop
870 DUMMY_COND is a shared cond_expr used by condition simplification as scratch,
871 to avoid allocating memory.
873 HANDLE_DOMINATING_ASSERTS is true if we should try to replace operands of
874 the simplified condition with left-hand sides of ASSERT_EXPRs they are
877 STACK is used to undo temporary equivalences created during the walk of
880 SIMPLIFY is a pass-specific function used to simplify statements.
882 Our caller is responsible for restoring the state of the expression
883 and const_and_copies stacks. */
886 thread_through_normal_block (edge e
,
888 bool handle_dominating_asserts
,
890 tree (*simplify
) (gimple
, gimple
),
891 vec
<jump_thread_edge
*> *path
,
894 /* If E is a backedge, then we want to verify that the COND_EXPR,
895 SWITCH_EXPR or GOTO_EXPR at the end of e->dest is not affected
896 by any statements in e->dest. If it is affected, then it is not
897 safe to thread this edge. */
898 if (e
->flags
& EDGE_DFS_BACK
)
900 if (cond_arg_set_in_bb (e
, e
->dest
))
904 /* PHIs create temporary equivalences. */
905 if (!record_temporary_equivalences_from_phis (e
, stack
))
908 /* Now walk each statement recording any context sensitive
909 temporary equivalences we can detect. */
911 = record_temporary_equivalences_from_stmts_at_dest (e
, stack
, simplify
);
915 /* If we stopped at a COND_EXPR or SWITCH_EXPR, see if we know which arm
917 if (gimple_code (stmt
) == GIMPLE_COND
918 || gimple_code (stmt
) == GIMPLE_GOTO
919 || gimple_code (stmt
) == GIMPLE_SWITCH
)
923 /* Extract and simplify the condition. */
924 cond
= simplify_control_stmt_condition (e
, stmt
, dummy_cond
, simplify
,
925 handle_dominating_asserts
);
927 if (cond
&& is_gimple_min_invariant (cond
))
929 edge taken_edge
= find_taken_edge (e
->dest
, cond
);
930 basic_block dest
= (taken_edge
? taken_edge
->dest
: NULL
);
932 /* DEST could be NULL for a computed jump to an absolute
934 if (dest
== NULL
|| dest
== e
->dest
|| bitmap_bit_p (visited
, dest
->index
))
938 = new jump_thread_edge (e
, EDGE_START_JUMP_THREAD
);
941 x
= new jump_thread_edge (taken_edge
, EDGE_COPY_SRC_BLOCK
);
944 /* See if we can thread through DEST as well, this helps capture
945 secondary effects of threading without having to re-run DOM or
947 if ((e
->flags
& EDGE_DFS_BACK
) == 0
948 || ! cond_arg_set_in_bb (taken_edge
, e
->dest
))
950 /* We don't want to thread back to a block we have already
951 visited. This may be overly conservative. */
952 bitmap_set_bit (visited
, dest
->index
);
953 bitmap_set_bit (visited
, e
->dest
->index
);
954 thread_around_empty_blocks (taken_edge
,
956 handle_dominating_asserts
,
967 /* We are exiting E->src, see if E->dest ends with a conditional
968 jump which has a known value when reached via E.
970 Special care is necessary if E is a back edge in the CFG as we
971 may have already recorded equivalences for E->dest into our
972 various tables, including the result of the conditional at
973 the end of E->dest. Threading opportunities are severely
974 limited in that case to avoid short-circuiting the loop
977 Note it is quite common for the first block inside a loop to
978 end with a conditional which is either always true or always
979 false when reached via the loop backedge. Thus we do not want
980 to blindly disable threading across a loop backedge.
982 DUMMY_COND is a shared cond_expr used by condition simplification as scratch,
983 to avoid allocating memory.
985 HANDLE_DOMINATING_ASSERTS is true if we should try to replace operands of
986 the simplified condition with left-hand sides of ASSERT_EXPRs they are
989 STACK is used to undo temporary equivalences created during the walk of
992 SIMPLIFY is a pass-specific function used to simplify statements. */
995 thread_across_edge (gimple dummy_cond
,
997 bool handle_dominating_asserts
,
999 tree (*simplify
) (gimple
, gimple
))
1001 bitmap visited
= BITMAP_ALLOC (NULL
);
1005 vec
<jump_thread_edge
*> *path
= new vec
<jump_thread_edge
*> ();
1006 bitmap_clear (visited
);
1007 bitmap_set_bit (visited
, e
->src
->index
);
1008 bitmap_set_bit (visited
, e
->dest
->index
);
1009 if (thread_through_normal_block (e
, dummy_cond
, handle_dominating_asserts
,
1010 stack
, simplify
, path
, visited
))
1012 propagate_threaded_block_debug_into (path
->last ()->e
->dest
,
1014 remove_temporary_equivalences (stack
);
1015 BITMAP_FREE (visited
);
1016 register_jump_thread (path
);
1021 /* There should be no edges on the path, so no need to walk through
1022 the vector entries. */
1023 gcc_assert (path
->length () == 0);
1027 /* We were unable to determine what out edge from E->dest is taken. However,
1028 we might still be able to thread through successors of E->dest. This
1029 often occurs when E->dest is a joiner block which then fans back out
1030 based on redundant tests.
1032 If so, we'll copy E->dest and redirect the appropriate predecessor to
1033 the copy. Within the copy of E->dest, we'll thread one or more edges
1034 to points deeper in the CFG.
1036 This is a stopgap until we have a more structured approach to path
1043 /* Look at each successor of E->dest to see if we can thread through it. */
1044 FOR_EACH_EDGE (taken_edge
, ei
, e
->dest
->succs
)
1046 /* Avoid threading to any block we have already visited. */
1047 bitmap_clear (visited
);
1048 bitmap_set_bit (visited
, taken_edge
->dest
->index
);
1049 bitmap_set_bit (visited
, e
->dest
->index
);
1050 vec
<jump_thread_edge
*> *path
= new vec
<jump_thread_edge
*> ();
1052 /* Record whether or not we were able to thread through a successor
1054 jump_thread_edge
*x
= new jump_thread_edge (e
, EDGE_START_JUMP_THREAD
);
1055 path
->safe_push (x
);
1057 x
= new jump_thread_edge (taken_edge
, EDGE_COPY_SRC_JOINER_BLOCK
);
1058 path
->safe_push (x
);
1060 if ((e
->flags
& EDGE_DFS_BACK
) == 0
1061 || ! cond_arg_set_in_bb (path
->last ()->e
, e
->dest
))
1062 found
= thread_around_empty_blocks (taken_edge
,
1064 handle_dominating_asserts
,
1069 /* If we were able to thread through a successor of E->dest, then
1070 record the jump threading opportunity. */
1073 propagate_threaded_block_debug_into (path
->last ()->e
->dest
,
1075 register_jump_thread (path
);
1079 for (unsigned int i
= 0; i
< path
->length (); i
++)
1084 BITMAP_FREE (visited
);
1087 remove_temporary_equivalences (stack
);