1 /* Optimization of PHI nodes by converting them into straightline code.
2 Copyright (C) 2004-2013 Free Software Foundation, Inc.
4 This file is part of GCC.
6 GCC is free software; you can redistribute it and/or modify it
7 under the terms of the GNU General Public License as published by the
8 Free Software Foundation; either version 3, or (at your option) any
11 GCC is distributed in the hope that it will be useful, but WITHOUT
12 ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
13 FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
16 You should have received a copy of the GNU General Public License
17 along with GCC; see the file COPYING3. If not see
18 <http://www.gnu.org/licenses/>. */
22 #include "coretypes.h"
23 #include "hash-table.h"
29 #include "basic-block.h"
31 #include "gimple-ssa.h"
33 #include "tree-phinodes.h"
34 #include "ssa-iterators.h"
35 #include "tree-ssanames.h"
37 #include "tree-pass.h"
38 #include "langhooks.h"
39 #include "pointer-set.h"
42 #include "tree-data-ref.h"
43 #include "gimple-pretty-print.h"
44 #include "insn-config.h"
47 #include "tree-scalar-evolution.h"
49 #ifndef HAVE_conditional_move
50 #define HAVE_conditional_move (0)
53 static unsigned int tree_ssa_phiopt (void);
54 static unsigned int tree_ssa_phiopt_worker (bool, bool);
55 static bool conditional_replacement (basic_block
, basic_block
,
56 edge
, edge
, gimple
, tree
, tree
);
57 static int value_replacement (basic_block
, basic_block
,
58 edge
, edge
, gimple
, tree
, tree
);
59 static bool minmax_replacement (basic_block
, basic_block
,
60 edge
, edge
, gimple
, tree
, tree
);
61 static bool abs_replacement (basic_block
, basic_block
,
62 edge
, edge
, gimple
, tree
, tree
);
63 static bool cond_store_replacement (basic_block
, basic_block
, edge
, edge
,
64 struct pointer_set_t
*);
65 static bool cond_if_else_store_replacement (basic_block
, basic_block
, basic_block
);
66 static struct pointer_set_t
* get_non_trapping (void);
67 static void replace_phi_edge_with_variable (basic_block
, edge
, gimple
, tree
);
68 static void hoist_adjacent_loads (basic_block
, basic_block
,
69 basic_block
, basic_block
);
70 static bool gate_hoist_loads (void);
72 /* This pass tries to replaces an if-then-else block with an
73 assignment. We have four kinds of transformations. Some of these
74 transformations are also performed by the ifcvt RTL optimizer.
76 Conditional Replacement
77 -----------------------
79 This transformation, implemented in conditional_replacement,
83 if (cond) goto bb2; else goto bb1;
86 x = PHI <0 (bb1), 1 (bb0), ...>;
94 x = PHI <x' (bb0), ...>;
96 We remove bb1 as it becomes unreachable. This occurs often due to
97 gimplification of conditionals.
102 This transformation, implemented in value_replacement, replaces
105 if (a != b) goto bb2; else goto bb1;
108 x = PHI <a (bb1), b (bb0), ...>;
114 x = PHI <b (bb0), ...>;
116 This opportunity can sometimes occur as a result of other
120 Another case caught by value replacement looks like this:
126 if (t3 != 0) goto bb1; else goto bb2;
142 This transformation, implemented in abs_replacement, replaces
145 if (a >= 0) goto bb2; else goto bb1;
149 x = PHI <x (bb1), a (bb0), ...>;
156 x = PHI <x' (bb0), ...>;
161 This transformation, minmax_replacement replaces
164 if (a <= b) goto bb2; else goto bb1;
167 x = PHI <b (bb1), a (bb0), ...>;
174 x = PHI <x' (bb0), ...>;
176 A similar transformation is done for MAX_EXPR.
179 This pass also performs a fifth transformation of a slightly different
182 Adjacent Load Hoisting
183 ----------------------
185 This transformation replaces
188 if (...) goto bb2; else goto bb1;
190 x1 = (<expr>).field1;
193 x2 = (<expr>).field2;
200 x1 = (<expr>).field1;
201 x2 = (<expr>).field2;
202 if (...) goto bb2; else goto bb1;
209 The purpose of this transformation is to enable generation of conditional
210 move instructions such as Intel CMOVE or PowerPC ISEL. Because one of
211 the loads is speculative, the transformation is restricted to very
212 specific cases to avoid introducing a page fault. We are looking for
220 where left and right are typically adjacent pointers in a tree structure. */
223 tree_ssa_phiopt (void)
225 return tree_ssa_phiopt_worker (false, gate_hoist_loads ());
228 /* This pass tries to transform conditional stores into unconditional
229 ones, enabling further simplifications with the simpler then and else
230 blocks. In particular it replaces this:
233 if (cond) goto bb2; else goto bb1;
241 if (cond) goto bb1; else goto bb2;
245 condtmp = PHI <RHS, condtmp'>
248 This transformation can only be done under several constraints,
249 documented below. It also replaces:
252 if (cond) goto bb2; else goto bb1;
263 if (cond) goto bb3; else goto bb1;
266 condtmp = PHI <RHS1, RHS2>
270 tree_ssa_cs_elim (void)
273 /* ??? We are not interested in loop related info, but the following
274 will create it, ICEing as we didn't init loops with pre-headers.
275 An interfacing issue of find_data_references_in_bb. */
276 loop_optimizer_init (LOOPS_NORMAL
);
278 todo
= tree_ssa_phiopt_worker (true, false);
280 loop_optimizer_finalize ();
284 /* Return the singleton PHI in the SEQ of PHIs for edges E0 and E1. */
287 single_non_singleton_phi_for_edges (gimple_seq seq
, edge e0
, edge e1
)
289 gimple_stmt_iterator i
;
291 if (gimple_seq_singleton_p (seq
))
292 return gsi_stmt (gsi_start (seq
));
293 for (i
= gsi_start (seq
); !gsi_end_p (i
); gsi_next (&i
))
295 gimple p
= gsi_stmt (i
);
296 /* If the PHI arguments are equal then we can skip this PHI. */
297 if (operand_equal_for_phi_arg_p (gimple_phi_arg_def (p
, e0
->dest_idx
),
298 gimple_phi_arg_def (p
, e1
->dest_idx
)))
301 /* If we already have a PHI that has the two edge arguments are
302 different, then return it is not a singleton for these PHIs. */
311 /* The core routine of conditional store replacement and normal
312 phi optimizations. Both share much of the infrastructure in how
313 to match applicable basic block patterns. DO_STORE_ELIM is true
314 when we want to do conditional store replacement, false otherwise.
315 DO_HOIST_LOADS is true when we want to hoist adjacent loads out
316 of diamond control flow patterns, false otherwise. */
318 tree_ssa_phiopt_worker (bool do_store_elim
, bool do_hoist_loads
)
321 basic_block
*bb_order
;
323 bool cfgchanged
= false;
324 struct pointer_set_t
*nontrap
= 0;
327 /* Calculate the set of non-trapping memory accesses. */
328 nontrap
= get_non_trapping ();
330 /* Search every basic block for COND_EXPR we may be able to optimize.
332 We walk the blocks in order that guarantees that a block with
333 a single predecessor is processed before the predecessor.
334 This ensures that we collapse inner ifs before visiting the
335 outer ones, and also that we do not try to visit a removed
337 bb_order
= single_pred_before_succ_order ();
338 n
= n_basic_blocks
- NUM_FIXED_BLOCKS
;
340 for (i
= 0; i
< n
; i
++)
342 gimple cond_stmt
, phi
;
343 basic_block bb1
, bb2
;
349 cond_stmt
= last_stmt (bb
);
350 /* Check to see if the last statement is a GIMPLE_COND. */
352 || gimple_code (cond_stmt
) != GIMPLE_COND
)
355 e1
= EDGE_SUCC (bb
, 0);
357 e2
= EDGE_SUCC (bb
, 1);
360 /* We cannot do the optimization on abnormal edges. */
361 if ((e1
->flags
& EDGE_ABNORMAL
) != 0
362 || (e2
->flags
& EDGE_ABNORMAL
) != 0)
365 /* If either bb1's succ or bb2 or bb2's succ is non NULL. */
366 if (EDGE_COUNT (bb1
->succs
) == 0
368 || EDGE_COUNT (bb2
->succs
) == 0)
371 /* Find the bb which is the fall through to the other. */
372 if (EDGE_SUCC (bb1
, 0)->dest
== bb2
)
374 else if (EDGE_SUCC (bb2
, 0)->dest
== bb1
)
376 basic_block bb_tmp
= bb1
;
383 else if (do_store_elim
384 && EDGE_SUCC (bb1
, 0)->dest
== EDGE_SUCC (bb2
, 0)->dest
)
386 basic_block bb3
= EDGE_SUCC (bb1
, 0)->dest
;
388 if (!single_succ_p (bb1
)
389 || (EDGE_SUCC (bb1
, 0)->flags
& EDGE_FALLTHRU
) == 0
390 || !single_succ_p (bb2
)
391 || (EDGE_SUCC (bb2
, 0)->flags
& EDGE_FALLTHRU
) == 0
392 || EDGE_COUNT (bb3
->preds
) != 2)
394 if (cond_if_else_store_replacement (bb1
, bb2
, bb3
))
398 else if (do_hoist_loads
399 && EDGE_SUCC (bb1
, 0)->dest
== EDGE_SUCC (bb2
, 0)->dest
)
401 basic_block bb3
= EDGE_SUCC (bb1
, 0)->dest
;
403 if (!FLOAT_TYPE_P (TREE_TYPE (gimple_cond_lhs (cond_stmt
)))
404 && single_succ_p (bb1
)
405 && single_succ_p (bb2
)
406 && single_pred_p (bb1
)
407 && single_pred_p (bb2
)
408 && EDGE_COUNT (bb
->succs
) == 2
409 && EDGE_COUNT (bb3
->preds
) == 2
410 /* If one edge or the other is dominant, a conditional move
411 is likely to perform worse than the well-predicted branch. */
412 && !predictable_edge_p (EDGE_SUCC (bb
, 0))
413 && !predictable_edge_p (EDGE_SUCC (bb
, 1)))
414 hoist_adjacent_loads (bb
, bb1
, bb2
, bb3
);
420 e1
= EDGE_SUCC (bb1
, 0);
422 /* Make sure that bb1 is just a fall through. */
423 if (!single_succ_p (bb1
)
424 || (e1
->flags
& EDGE_FALLTHRU
) == 0)
427 /* Also make sure that bb1 only have one predecessor and that it
429 if (!single_pred_p (bb1
)
430 || single_pred (bb1
) != bb
)
435 /* bb1 is the middle block, bb2 the join block, bb the split block,
436 e1 the fallthrough edge from bb1 to bb2. We can't do the
437 optimization if the join block has more than two predecessors. */
438 if (EDGE_COUNT (bb2
->preds
) > 2)
440 if (cond_store_replacement (bb1
, bb2
, e1
, e2
, nontrap
))
445 gimple_seq phis
= phi_nodes (bb2
);
446 gimple_stmt_iterator gsi
;
447 bool candorest
= true;
449 /* Value replacement can work with more than one PHI
450 so try that first. */
451 for (gsi
= gsi_start (phis
); !gsi_end_p (gsi
); gsi_next (&gsi
))
453 phi
= gsi_stmt (gsi
);
454 arg0
= gimple_phi_arg_def (phi
, e1
->dest_idx
);
455 arg1
= gimple_phi_arg_def (phi
, e2
->dest_idx
);
456 if (value_replacement (bb
, bb1
, e1
, e2
, phi
, arg0
, arg1
) == 2)
467 phi
= single_non_singleton_phi_for_edges (phis
, e1
, e2
);
471 arg0
= gimple_phi_arg_def (phi
, e1
->dest_idx
);
472 arg1
= gimple_phi_arg_def (phi
, e2
->dest_idx
);
474 /* Something is wrong if we cannot find the arguments in the PHI
476 gcc_assert (arg0
!= NULL
&& arg1
!= NULL
);
478 /* Do the replacement of conditional if it can be done. */
479 if (conditional_replacement (bb
, bb1
, e1
, e2
, phi
, arg0
, arg1
))
481 else if (abs_replacement (bb
, bb1
, e1
, e2
, phi
, arg0
, arg1
))
483 else if (minmax_replacement (bb
, bb1
, e1
, e2
, phi
, arg0
, arg1
))
491 pointer_set_destroy (nontrap
);
492 /* If the CFG has changed, we should cleanup the CFG. */
493 if (cfgchanged
&& do_store_elim
)
495 /* In cond-store replacement we have added some loads on edges
496 and new VOPS (as we moved the store, and created a load). */
497 gsi_commit_edge_inserts ();
498 return TODO_cleanup_cfg
| TODO_update_ssa_only_virtuals
;
501 return TODO_cleanup_cfg
;
505 /* Replace PHI node element whose edge is E in block BB with variable NEW.
506 Remove the edge from COND_BLOCK which does not lead to BB (COND_BLOCK
507 is known to have two edges, one of which must reach BB). */
510 replace_phi_edge_with_variable (basic_block cond_block
,
511 edge e
, gimple phi
, tree new_tree
)
513 basic_block bb
= gimple_bb (phi
);
514 basic_block block_to_remove
;
515 gimple_stmt_iterator gsi
;
517 /* Change the PHI argument to new. */
518 SET_USE (PHI_ARG_DEF_PTR (phi
, e
->dest_idx
), new_tree
);
520 /* Remove the empty basic block. */
521 if (EDGE_SUCC (cond_block
, 0)->dest
== bb
)
523 EDGE_SUCC (cond_block
, 0)->flags
|= EDGE_FALLTHRU
;
524 EDGE_SUCC (cond_block
, 0)->flags
&= ~(EDGE_TRUE_VALUE
| EDGE_FALSE_VALUE
);
525 EDGE_SUCC (cond_block
, 0)->probability
= REG_BR_PROB_BASE
;
526 EDGE_SUCC (cond_block
, 0)->count
+= EDGE_SUCC (cond_block
, 1)->count
;
528 block_to_remove
= EDGE_SUCC (cond_block
, 1)->dest
;
532 EDGE_SUCC (cond_block
, 1)->flags
|= EDGE_FALLTHRU
;
533 EDGE_SUCC (cond_block
, 1)->flags
534 &= ~(EDGE_TRUE_VALUE
| EDGE_FALSE_VALUE
);
535 EDGE_SUCC (cond_block
, 1)->probability
= REG_BR_PROB_BASE
;
536 EDGE_SUCC (cond_block
, 1)->count
+= EDGE_SUCC (cond_block
, 0)->count
;
538 block_to_remove
= EDGE_SUCC (cond_block
, 0)->dest
;
540 delete_basic_block (block_to_remove
);
542 /* Eliminate the COND_EXPR at the end of COND_BLOCK. */
543 gsi
= gsi_last_bb (cond_block
);
544 gsi_remove (&gsi
, true);
546 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
548 "COND_EXPR in block %d and PHI in block %d converted to straightline code.\n",
553 /* The function conditional_replacement does the main work of doing the
554 conditional replacement. Return true if the replacement is done.
555 Otherwise return false.
556 BB is the basic block where the replacement is going to be done on. ARG0
557 is argument 0 from PHI. Likewise for ARG1. */
560 conditional_replacement (basic_block cond_bb
, basic_block middle_bb
,
561 edge e0
, edge e1
, gimple phi
,
562 tree arg0
, tree arg1
)
565 gimple stmt
, new_stmt
;
567 gimple_stmt_iterator gsi
;
568 edge true_edge
, false_edge
;
569 tree new_var
, new_var2
;
572 /* FIXME: Gimplification of complex type is too hard for now. */
573 /* We aren't prepared to handle vectors either (and it is a question
574 if it would be worthwhile anyway). */
575 if (!(INTEGRAL_TYPE_P (TREE_TYPE (arg0
))
576 || POINTER_TYPE_P (TREE_TYPE (arg0
)))
577 || !(INTEGRAL_TYPE_P (TREE_TYPE (arg1
))
578 || POINTER_TYPE_P (TREE_TYPE (arg1
))))
581 /* The PHI arguments have the constants 0 and 1, or 0 and -1, then
582 convert it to the conditional. */
583 if ((integer_zerop (arg0
) && integer_onep (arg1
))
584 || (integer_zerop (arg1
) && integer_onep (arg0
)))
586 else if ((integer_zerop (arg0
) && integer_all_onesp (arg1
))
587 || (integer_zerop (arg1
) && integer_all_onesp (arg0
)))
592 if (!empty_block_p (middle_bb
))
595 /* At this point we know we have a GIMPLE_COND with two successors.
596 One successor is BB, the other successor is an empty block which
597 falls through into BB.
599 There is a single PHI node at the join point (BB) and its arguments
600 are constants (0, 1) or (0, -1).
602 So, given the condition COND, and the two PHI arguments, we can
603 rewrite this PHI into non-branching code:
605 dest = (COND) or dest = COND'
607 We use the condition as-is if the argument associated with the
608 true edge has the value one or the argument associated with the
609 false edge as the value zero. Note that those conditions are not
610 the same since only one of the outgoing edges from the GIMPLE_COND
611 will directly reach BB and thus be associated with an argument. */
613 stmt
= last_stmt (cond_bb
);
614 result
= PHI_RESULT (phi
);
616 /* To handle special cases like floating point comparison, it is easier and
617 less error-prone to build a tree and gimplify it on the fly though it is
619 cond
= fold_build2_loc (gimple_location (stmt
),
620 gimple_cond_code (stmt
), boolean_type_node
,
621 gimple_cond_lhs (stmt
), gimple_cond_rhs (stmt
));
623 /* We need to know which is the true edge and which is the false
624 edge so that we know when to invert the condition below. */
625 extract_true_false_edges_from_block (cond_bb
, &true_edge
, &false_edge
);
626 if ((e0
== true_edge
&& integer_zerop (arg0
))
627 || (e0
== false_edge
&& !integer_zerop (arg0
))
628 || (e1
== true_edge
&& integer_zerop (arg1
))
629 || (e1
== false_edge
&& !integer_zerop (arg1
)))
630 cond
= fold_build1_loc (gimple_location (stmt
),
631 TRUTH_NOT_EXPR
, TREE_TYPE (cond
), cond
);
635 cond
= fold_convert_loc (gimple_location (stmt
),
636 TREE_TYPE (result
), cond
);
637 cond
= fold_build1_loc (gimple_location (stmt
),
638 NEGATE_EXPR
, TREE_TYPE (cond
), cond
);
641 /* Insert our new statements at the end of conditional block before the
643 gsi
= gsi_for_stmt (stmt
);
644 new_var
= force_gimple_operand_gsi (&gsi
, cond
, true, NULL
, true,
647 if (!useless_type_conversion_p (TREE_TYPE (result
), TREE_TYPE (new_var
)))
649 source_location locus_0
, locus_1
;
651 new_var2
= make_ssa_name (TREE_TYPE (result
), NULL
);
652 new_stmt
= gimple_build_assign_with_ops (CONVERT_EXPR
, new_var2
,
654 gsi_insert_before (&gsi
, new_stmt
, GSI_SAME_STMT
);
657 /* Set the locus to the first argument, unless is doesn't have one. */
658 locus_0
= gimple_phi_arg_location (phi
, 0);
659 locus_1
= gimple_phi_arg_location (phi
, 1);
660 if (locus_0
== UNKNOWN_LOCATION
)
662 gimple_set_location (new_stmt
, locus_0
);
665 replace_phi_edge_with_variable (cond_bb
, e1
, phi
, new_var
);
667 /* Note that we optimized this PHI. */
671 /* Update *ARG which is defined in STMT so that it contains the
672 computed value if that seems profitable. Return true if the
673 statement is made dead by that rewriting. */
676 jump_function_from_stmt (tree
*arg
, gimple stmt
)
678 enum tree_code code
= gimple_assign_rhs_code (stmt
);
679 if (code
== ADDR_EXPR
)
681 /* For arg = &p->i transform it to p, if possible. */
682 tree rhs1
= gimple_assign_rhs1 (stmt
);
683 HOST_WIDE_INT offset
;
684 tree tem
= get_addr_base_and_unit_offset (TREE_OPERAND (rhs1
, 0),
687 && TREE_CODE (tem
) == MEM_REF
688 && (mem_ref_offset (tem
) + double_int::from_shwi (offset
)).is_zero ())
690 *arg
= TREE_OPERAND (tem
, 0);
694 /* TODO: Much like IPA-CP jump-functions we want to handle constant
695 additions symbolically here, and we'd need to update the comparison
696 code that compares the arg + cst tuples in our caller. For now the
697 code above exactly handles the VEC_BASE pattern from vec.h. */
701 /* RHS is a source argument in a BIT_AND_EXPR which feeds a conditional
702 of the form SSA_NAME NE 0.
704 If RHS is fed by a simple EQ_EXPR comparison of two values, see if
705 the two input values of the EQ_EXPR match arg0 and arg1.
707 If so update *code and return TRUE. Otherwise return FALSE. */
710 rhs_is_fed_for_value_replacement (const_tree arg0
, const_tree arg1
,
711 enum tree_code
*code
, const_tree rhs
)
713 /* Obviously if RHS is not an SSA_NAME, we can't look at the defining
715 if (TREE_CODE (rhs
) == SSA_NAME
)
717 gimple def1
= SSA_NAME_DEF_STMT (rhs
);
719 /* Verify the defining statement has an EQ_EXPR on the RHS. */
720 if (is_gimple_assign (def1
) && gimple_assign_rhs_code (def1
) == EQ_EXPR
)
722 /* Finally verify the source operands of the EQ_EXPR are equal
724 tree op0
= gimple_assign_rhs1 (def1
);
725 tree op1
= gimple_assign_rhs2 (def1
);
726 if ((operand_equal_for_phi_arg_p (arg0
, op0
)
727 && operand_equal_for_phi_arg_p (arg1
, op1
))
728 || (operand_equal_for_phi_arg_p (arg0
, op1
)
729 && operand_equal_for_phi_arg_p (arg1
, op0
)))
731 /* We will perform the optimization. */
732 *code
= gimple_assign_rhs_code (def1
);
740 /* Return TRUE if arg0/arg1 are equal to the rhs/lhs or lhs/rhs of COND.
742 Also return TRUE if arg0/arg1 are equal to the source arguments of a
743 an EQ comparison feeding a BIT_AND_EXPR which feeds COND.
745 Return FALSE otherwise. */
748 operand_equal_for_value_replacement (const_tree arg0
, const_tree arg1
,
749 enum tree_code
*code
, gimple cond
)
752 tree lhs
= gimple_cond_lhs (cond
);
753 tree rhs
= gimple_cond_rhs (cond
);
755 if ((operand_equal_for_phi_arg_p (arg0
, lhs
)
756 && operand_equal_for_phi_arg_p (arg1
, rhs
))
757 || (operand_equal_for_phi_arg_p (arg1
, lhs
)
758 && operand_equal_for_phi_arg_p (arg0
, rhs
)))
761 /* Now handle more complex case where we have an EQ comparison
762 which feeds a BIT_AND_EXPR which feeds COND.
764 First verify that COND is of the form SSA_NAME NE 0. */
765 if (*code
!= NE_EXPR
|| !integer_zerop (rhs
)
766 || TREE_CODE (lhs
) != SSA_NAME
)
769 /* Now ensure that SSA_NAME is set by a BIT_AND_EXPR. */
770 def
= SSA_NAME_DEF_STMT (lhs
);
771 if (!is_gimple_assign (def
) || gimple_assign_rhs_code (def
) != BIT_AND_EXPR
)
774 /* Now verify arg0/arg1 correspond to the source arguments of an
775 EQ comparison feeding the BIT_AND_EXPR. */
777 tree tmp
= gimple_assign_rhs1 (def
);
778 if (rhs_is_fed_for_value_replacement (arg0
, arg1
, code
, tmp
))
781 tmp
= gimple_assign_rhs2 (def
);
782 if (rhs_is_fed_for_value_replacement (arg0
, arg1
, code
, tmp
))
788 /* The function value_replacement does the main work of doing the value
789 replacement. Return non-zero if the replacement is done. Otherwise return
790 0. If we remove the middle basic block, return 2.
791 BB is the basic block where the replacement is going to be done on. ARG0
792 is argument 0 from the PHI. Likewise for ARG1. */
795 value_replacement (basic_block cond_bb
, basic_block middle_bb
,
796 edge e0
, edge e1
, gimple phi
,
797 tree arg0
, tree arg1
)
799 gimple_stmt_iterator gsi
;
801 edge true_edge
, false_edge
;
803 bool emtpy_or_with_defined_p
= true;
805 /* If the type says honor signed zeros we cannot do this
807 if (HONOR_SIGNED_ZEROS (TYPE_MODE (TREE_TYPE (arg1
))))
810 /* If there is a statement in MIDDLE_BB that defines one of the PHI
811 arguments, then adjust arg0 or arg1. */
812 gsi
= gsi_after_labels (middle_bb
);
813 if (!gsi_end_p (gsi
) && is_gimple_debug (gsi_stmt (gsi
)))
814 gsi_next_nondebug (&gsi
);
815 while (!gsi_end_p (gsi
))
817 gimple stmt
= gsi_stmt (gsi
);
819 gsi_next_nondebug (&gsi
);
820 if (!is_gimple_assign (stmt
))
822 emtpy_or_with_defined_p
= false;
825 /* Now try to adjust arg0 or arg1 according to the computation
827 lhs
= gimple_assign_lhs (stmt
);
829 && jump_function_from_stmt (&arg0
, stmt
))
831 && jump_function_from_stmt (&arg1
, stmt
)))
832 emtpy_or_with_defined_p
= false;
835 cond
= last_stmt (cond_bb
);
836 code
= gimple_cond_code (cond
);
838 /* This transformation is only valid for equality comparisons. */
839 if (code
!= NE_EXPR
&& code
!= EQ_EXPR
)
842 /* We need to know which is the true edge and which is the false
843 edge so that we know if have abs or negative abs. */
844 extract_true_false_edges_from_block (cond_bb
, &true_edge
, &false_edge
);
846 /* At this point we know we have a COND_EXPR with two successors.
847 One successor is BB, the other successor is an empty block which
848 falls through into BB.
850 The condition for the COND_EXPR is known to be NE_EXPR or EQ_EXPR.
852 There is a single PHI node at the join point (BB) with two arguments.
854 We now need to verify that the two arguments in the PHI node match
855 the two arguments to the equality comparison. */
857 if (operand_equal_for_value_replacement (arg0
, arg1
, &code
, cond
))
862 /* For NE_EXPR, we want to build an assignment result = arg where
863 arg is the PHI argument associated with the true edge. For
864 EQ_EXPR we want the PHI argument associated with the false edge. */
865 e
= (code
== NE_EXPR
? true_edge
: false_edge
);
867 /* Unfortunately, E may not reach BB (it may instead have gone to
868 OTHER_BLOCK). If that is the case, then we want the single outgoing
869 edge from OTHER_BLOCK which reaches BB and represents the desired
870 path from COND_BLOCK. */
871 if (e
->dest
== middle_bb
)
872 e
= single_succ_edge (e
->dest
);
874 /* Now we know the incoming edge to BB that has the argument for the
875 RHS of our new assignment statement. */
881 /* If the middle basic block was empty or is defining the
882 PHI arguments and this is a single phi where the args are different
883 for the edges e0 and e1 then we can remove the middle basic block. */
884 if (emtpy_or_with_defined_p
885 && single_non_singleton_phi_for_edges (phi_nodes (gimple_bb (phi
)),
888 replace_phi_edge_with_variable (cond_bb
, e1
, phi
, arg
);
889 /* Note that we optimized this PHI. */
894 /* Replace the PHI arguments with arg. */
895 SET_PHI_ARG_DEF (phi
, e0
->dest_idx
, arg
);
896 SET_PHI_ARG_DEF (phi
, e1
->dest_idx
, arg
);
897 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
899 fprintf (dump_file
, "PHI ");
900 print_generic_expr (dump_file
, gimple_phi_result (phi
), 0);
901 fprintf (dump_file
, " reduced for COND_EXPR in block %d to ",
903 print_generic_expr (dump_file
, arg
, 0);
904 fprintf (dump_file
, ".\n");
913 /* The function minmax_replacement does the main work of doing the minmax
914 replacement. Return true if the replacement is done. Otherwise return
916 BB is the basic block where the replacement is going to be done on. ARG0
917 is argument 0 from the PHI. Likewise for ARG1. */
920 minmax_replacement (basic_block cond_bb
, basic_block middle_bb
,
921 edge e0
, edge e1
, gimple phi
,
922 tree arg0
, tree arg1
)
925 gimple cond
, new_stmt
;
926 edge true_edge
, false_edge
;
927 enum tree_code cmp
, minmax
, ass_code
;
928 tree smaller
, larger
, arg_true
, arg_false
;
929 gimple_stmt_iterator gsi
, gsi_from
;
931 type
= TREE_TYPE (PHI_RESULT (phi
));
933 /* The optimization may be unsafe due to NaNs. */
934 if (HONOR_NANS (TYPE_MODE (type
)))
937 cond
= last_stmt (cond_bb
);
938 cmp
= gimple_cond_code (cond
);
940 /* This transformation is only valid for order comparisons. Record which
941 operand is smaller/larger if the result of the comparison is true. */
942 if (cmp
== LT_EXPR
|| cmp
== LE_EXPR
)
944 smaller
= gimple_cond_lhs (cond
);
945 larger
= gimple_cond_rhs (cond
);
947 else if (cmp
== GT_EXPR
|| cmp
== GE_EXPR
)
949 smaller
= gimple_cond_rhs (cond
);
950 larger
= gimple_cond_lhs (cond
);
955 /* We need to know which is the true edge and which is the false
956 edge so that we know if have abs or negative abs. */
957 extract_true_false_edges_from_block (cond_bb
, &true_edge
, &false_edge
);
959 /* Forward the edges over the middle basic block. */
960 if (true_edge
->dest
== middle_bb
)
961 true_edge
= EDGE_SUCC (true_edge
->dest
, 0);
962 if (false_edge
->dest
== middle_bb
)
963 false_edge
= EDGE_SUCC (false_edge
->dest
, 0);
967 gcc_assert (false_edge
== e1
);
973 gcc_assert (false_edge
== e0
);
974 gcc_assert (true_edge
== e1
);
979 if (empty_block_p (middle_bb
))
981 if (operand_equal_for_phi_arg_p (arg_true
, smaller
)
982 && operand_equal_for_phi_arg_p (arg_false
, larger
))
986 if (smaller < larger)
992 else if (operand_equal_for_phi_arg_p (arg_false
, smaller
)
993 && operand_equal_for_phi_arg_p (arg_true
, larger
))
1000 /* Recognize the following case, assuming d <= u:
1006 This is equivalent to
1011 gimple assign
= last_and_only_stmt (middle_bb
);
1012 tree lhs
, op0
, op1
, bound
;
1015 || gimple_code (assign
) != GIMPLE_ASSIGN
)
1018 lhs
= gimple_assign_lhs (assign
);
1019 ass_code
= gimple_assign_rhs_code (assign
);
1020 if (ass_code
!= MAX_EXPR
&& ass_code
!= MIN_EXPR
)
1022 op0
= gimple_assign_rhs1 (assign
);
1023 op1
= gimple_assign_rhs2 (assign
);
1025 if (true_edge
->src
== middle_bb
)
1027 /* We got here if the condition is true, i.e., SMALLER < LARGER. */
1028 if (!operand_equal_for_phi_arg_p (lhs
, arg_true
))
1031 if (operand_equal_for_phi_arg_p (arg_false
, larger
))
1035 if (smaller < larger)
1037 r' = MAX_EXPR (smaller, bound)
1039 r = PHI <r', larger> --> to be turned to MIN_EXPR. */
1040 if (ass_code
!= MAX_EXPR
)
1044 if (operand_equal_for_phi_arg_p (op0
, smaller
))
1046 else if (operand_equal_for_phi_arg_p (op1
, smaller
))
1051 /* We need BOUND <= LARGER. */
1052 if (!integer_nonzerop (fold_build2 (LE_EXPR
, boolean_type_node
,
1056 else if (operand_equal_for_phi_arg_p (arg_false
, smaller
))
1060 if (smaller < larger)
1062 r' = MIN_EXPR (larger, bound)
1064 r = PHI <r', smaller> --> to be turned to MAX_EXPR. */
1065 if (ass_code
!= MIN_EXPR
)
1069 if (operand_equal_for_phi_arg_p (op0
, larger
))
1071 else if (operand_equal_for_phi_arg_p (op1
, larger
))
1076 /* We need BOUND >= SMALLER. */
1077 if (!integer_nonzerop (fold_build2 (GE_EXPR
, boolean_type_node
,
1086 /* We got here if the condition is false, i.e., SMALLER > LARGER. */
1087 if (!operand_equal_for_phi_arg_p (lhs
, arg_false
))
1090 if (operand_equal_for_phi_arg_p (arg_true
, larger
))
1094 if (smaller > larger)
1096 r' = MIN_EXPR (smaller, bound)
1098 r = PHI <r', larger> --> to be turned to MAX_EXPR. */
1099 if (ass_code
!= MIN_EXPR
)
1103 if (operand_equal_for_phi_arg_p (op0
, smaller
))
1105 else if (operand_equal_for_phi_arg_p (op1
, smaller
))
1110 /* We need BOUND >= LARGER. */
1111 if (!integer_nonzerop (fold_build2 (GE_EXPR
, boolean_type_node
,
1115 else if (operand_equal_for_phi_arg_p (arg_true
, smaller
))
1119 if (smaller > larger)
1121 r' = MAX_EXPR (larger, bound)
1123 r = PHI <r', smaller> --> to be turned to MIN_EXPR. */
1124 if (ass_code
!= MAX_EXPR
)
1128 if (operand_equal_for_phi_arg_p (op0
, larger
))
1130 else if (operand_equal_for_phi_arg_p (op1
, larger
))
1135 /* We need BOUND <= SMALLER. */
1136 if (!integer_nonzerop (fold_build2 (LE_EXPR
, boolean_type_node
,
1144 /* Move the statement from the middle block. */
1145 gsi
= gsi_last_bb (cond_bb
);
1146 gsi_from
= gsi_last_nondebug_bb (middle_bb
);
1147 gsi_move_before (&gsi_from
, &gsi
);
1150 /* Emit the statement to compute min/max. */
1151 result
= duplicate_ssa_name (PHI_RESULT (phi
), NULL
);
1152 new_stmt
= gimple_build_assign_with_ops (minmax
, result
, arg0
, arg1
);
1153 gsi
= gsi_last_bb (cond_bb
);
1154 gsi_insert_before (&gsi
, new_stmt
, GSI_NEW_STMT
);
1156 replace_phi_edge_with_variable (cond_bb
, e1
, phi
, result
);
1160 /* The function absolute_replacement does the main work of doing the absolute
1161 replacement. Return true if the replacement is done. Otherwise return
1163 bb is the basic block where the replacement is going to be done on. arg0
1164 is argument 0 from the phi. Likewise for arg1. */
1167 abs_replacement (basic_block cond_bb
, basic_block middle_bb
,
1168 edge e0 ATTRIBUTE_UNUSED
, edge e1
,
1169 gimple phi
, tree arg0
, tree arg1
)
1172 gimple new_stmt
, cond
;
1173 gimple_stmt_iterator gsi
;
1174 edge true_edge
, false_edge
;
1179 enum tree_code cond_code
;
1181 /* If the type says honor signed zeros we cannot do this
1183 if (HONOR_SIGNED_ZEROS (TYPE_MODE (TREE_TYPE (arg1
))))
1186 /* OTHER_BLOCK must have only one executable statement which must have the
1187 form arg0 = -arg1 or arg1 = -arg0. */
1189 assign
= last_and_only_stmt (middle_bb
);
1190 /* If we did not find the proper negation assignment, then we can not
1195 /* If we got here, then we have found the only executable statement
1196 in OTHER_BLOCK. If it is anything other than arg = -arg1 or
1197 arg1 = -arg0, then we can not optimize. */
1198 if (gimple_code (assign
) != GIMPLE_ASSIGN
)
1201 lhs
= gimple_assign_lhs (assign
);
1203 if (gimple_assign_rhs_code (assign
) != NEGATE_EXPR
)
1206 rhs
= gimple_assign_rhs1 (assign
);
1208 /* The assignment has to be arg0 = -arg1 or arg1 = -arg0. */
1209 if (!(lhs
== arg0
&& rhs
== arg1
)
1210 && !(lhs
== arg1
&& rhs
== arg0
))
1213 cond
= last_stmt (cond_bb
);
1214 result
= PHI_RESULT (phi
);
1216 /* Only relationals comparing arg[01] against zero are interesting. */
1217 cond_code
= gimple_cond_code (cond
);
1218 if (cond_code
!= GT_EXPR
&& cond_code
!= GE_EXPR
1219 && cond_code
!= LT_EXPR
&& cond_code
!= LE_EXPR
)
1222 /* Make sure the conditional is arg[01] OP y. */
1223 if (gimple_cond_lhs (cond
) != rhs
)
1226 if (FLOAT_TYPE_P (TREE_TYPE (gimple_cond_rhs (cond
)))
1227 ? real_zerop (gimple_cond_rhs (cond
))
1228 : integer_zerop (gimple_cond_rhs (cond
)))
1233 /* We need to know which is the true edge and which is the false
1234 edge so that we know if have abs or negative abs. */
1235 extract_true_false_edges_from_block (cond_bb
, &true_edge
, &false_edge
);
1237 /* For GT_EXPR/GE_EXPR, if the true edge goes to OTHER_BLOCK, then we
1238 will need to negate the result. Similarly for LT_EXPR/LE_EXPR if
1239 the false edge goes to OTHER_BLOCK. */
1240 if (cond_code
== GT_EXPR
|| cond_code
== GE_EXPR
)
1245 if (e
->dest
== middle_bb
)
1250 result
= duplicate_ssa_name (result
, NULL
);
1253 lhs
= make_ssa_name (TREE_TYPE (result
), NULL
);
1257 /* Build the modify expression with abs expression. */
1258 new_stmt
= gimple_build_assign_with_ops (ABS_EXPR
, lhs
, rhs
, NULL
);
1260 gsi
= gsi_last_bb (cond_bb
);
1261 gsi_insert_before (&gsi
, new_stmt
, GSI_NEW_STMT
);
1265 /* Get the right GSI. We want to insert after the recently
1266 added ABS_EXPR statement (which we know is the first statement
1268 new_stmt
= gimple_build_assign_with_ops (NEGATE_EXPR
, result
, lhs
, NULL
);
1270 gsi_insert_after (&gsi
, new_stmt
, GSI_NEW_STMT
);
1273 replace_phi_edge_with_variable (cond_bb
, e1
, phi
, result
);
1275 /* Note that we optimized this PHI. */
1279 /* Auxiliary functions to determine the set of memory accesses which
1280 can't trap because they are preceded by accesses to the same memory
1281 portion. We do that for MEM_REFs, so we only need to track
1282 the SSA_NAME of the pointer indirectly referenced. The algorithm
1283 simply is a walk over all instructions in dominator order. When
1284 we see an MEM_REF we determine if we've already seen a same
1285 ref anywhere up to the root of the dominator tree. If we do the
1286 current access can't trap. If we don't see any dominating access
1287 the current access might trap, but might also make later accesses
1288 non-trapping, so we remember it. We need to be careful with loads
1289 or stores, for instance a load might not trap, while a store would,
1290 so if we see a dominating read access this doesn't mean that a later
1291 write access would not trap. Hence we also need to differentiate the
1292 type of access(es) seen.
1294 ??? We currently are very conservative and assume that a load might
1295 trap even if a store doesn't (write-only memory). This probably is
1296 overly conservative. */
1298 /* A hash-table of SSA_NAMEs, and in which basic block an MEM_REF
1299 through it was seen, which would constitute a no-trap region for
1303 unsigned int ssa_name_ver
;
1306 HOST_WIDE_INT offset
, size
;
1310 /* Hashtable helpers. */
1312 struct ssa_names_hasher
: typed_free_remove
<name_to_bb
>
1314 typedef name_to_bb value_type
;
1315 typedef name_to_bb compare_type
;
1316 static inline hashval_t
hash (const value_type
*);
1317 static inline bool equal (const value_type
*, const compare_type
*);
1320 /* Used for quick clearing of the hash-table when we see calls.
1321 Hash entries with phase < nt_call_phase are invalid. */
1322 static unsigned int nt_call_phase
;
1324 /* The hash function. */
1327 ssa_names_hasher::hash (const value_type
*n
)
1329 return n
->ssa_name_ver
^ (((hashval_t
) n
->store
) << 31)
1330 ^ (n
->offset
<< 6) ^ (n
->size
<< 3);
1333 /* The equality function of *P1 and *P2. */
1336 ssa_names_hasher::equal (const value_type
*n1
, const compare_type
*n2
)
1338 return n1
->ssa_name_ver
== n2
->ssa_name_ver
1339 && n1
->store
== n2
->store
1340 && n1
->offset
== n2
->offset
1341 && n1
->size
== n2
->size
;
1344 /* The hash table for remembering what we've seen. */
1345 static hash_table
<ssa_names_hasher
> seen_ssa_names
;
1347 /* We see the expression EXP in basic block BB. If it's an interesting
1348 expression (an MEM_REF through an SSA_NAME) possibly insert the
1349 expression into the set NONTRAP or the hash table of seen expressions.
1350 STORE is true if this expression is on the LHS, otherwise it's on
1353 add_or_mark_expr (basic_block bb
, tree exp
,
1354 struct pointer_set_t
*nontrap
, bool store
)
1358 if (TREE_CODE (exp
) == MEM_REF
1359 && TREE_CODE (TREE_OPERAND (exp
, 0)) == SSA_NAME
1360 && host_integerp (TREE_OPERAND (exp
, 1), 0)
1361 && (size
= int_size_in_bytes (TREE_TYPE (exp
))) > 0)
1363 tree name
= TREE_OPERAND (exp
, 0);
1364 struct name_to_bb map
;
1366 struct name_to_bb
*n2bb
;
1367 basic_block found_bb
= 0;
1369 /* Try to find the last seen MEM_REF through the same
1370 SSA_NAME, which can trap. */
1371 map
.ssa_name_ver
= SSA_NAME_VERSION (name
);
1375 map
.offset
= tree_low_cst (TREE_OPERAND (exp
, 1), 0);
1378 slot
= seen_ssa_names
.find_slot (&map
, INSERT
);
1380 if (n2bb
&& n2bb
->phase
>= nt_call_phase
)
1381 found_bb
= n2bb
->bb
;
1383 /* If we've found a trapping MEM_REF, _and_ it dominates EXP
1384 (it's in a basic block on the path from us to the dominator root)
1385 then we can't trap. */
1386 if (found_bb
&& (((size_t)found_bb
->aux
) & 1) == 1)
1388 pointer_set_insert (nontrap
, exp
);
1392 /* EXP might trap, so insert it into the hash table. */
1395 n2bb
->phase
= nt_call_phase
;
1400 n2bb
= XNEW (struct name_to_bb
);
1401 n2bb
->ssa_name_ver
= SSA_NAME_VERSION (name
);
1402 n2bb
->phase
= nt_call_phase
;
1404 n2bb
->store
= store
;
1405 n2bb
->offset
= map
.offset
;
1413 class nontrapping_dom_walker
: public dom_walker
1416 nontrapping_dom_walker (cdi_direction direction
, pointer_set_t
*ps
)
1417 : dom_walker (direction
), m_nontrapping (ps
) {}
1419 virtual void before_dom_children (basic_block
);
1420 virtual void after_dom_children (basic_block
);
1423 pointer_set_t
*m_nontrapping
;
1426 /* Called by walk_dominator_tree, when entering the block BB. */
1428 nontrapping_dom_walker::before_dom_children (basic_block bb
)
1432 gimple_stmt_iterator gsi
;
1434 /* If we haven't seen all our predecessors, clear the hash-table. */
1435 FOR_EACH_EDGE (e
, ei
, bb
->preds
)
1436 if ((((size_t)e
->src
->aux
) & 2) == 0)
1442 /* Mark this BB as being on the path to dominator root and as visited. */
1443 bb
->aux
= (void*)(1 | 2);
1445 /* And walk the statements in order. */
1446 for (gsi
= gsi_start_bb (bb
); !gsi_end_p (gsi
); gsi_next (&gsi
))
1448 gimple stmt
= gsi_stmt (gsi
);
1450 if (is_gimple_call (stmt
) && !nonfreeing_call_p (stmt
))
1452 else if (gimple_assign_single_p (stmt
) && !gimple_has_volatile_ops (stmt
))
1454 add_or_mark_expr (bb
, gimple_assign_lhs (stmt
), m_nontrapping
, true);
1455 add_or_mark_expr (bb
, gimple_assign_rhs1 (stmt
), m_nontrapping
, false);
1460 /* Called by walk_dominator_tree, when basic block BB is exited. */
1462 nontrapping_dom_walker::after_dom_children (basic_block bb
)
1464 /* This BB isn't on the path to dominator root anymore. */
1468 /* This is the entry point of gathering non trapping memory accesses.
1469 It will do a dominator walk over the whole function, and it will
1470 make use of the bb->aux pointers. It returns a set of trees
1471 (the MEM_REFs itself) which can't trap. */
1472 static struct pointer_set_t
*
1473 get_non_trapping (void)
1476 pointer_set_t
*nontrap
= pointer_set_create ();
1477 seen_ssa_names
.create (128);
1478 /* We're going to do a dominator walk, so ensure that we have
1479 dominance information. */
1480 calculate_dominance_info (CDI_DOMINATORS
);
1482 nontrapping_dom_walker (CDI_DOMINATORS
, nontrap
)
1483 .walk (cfun
->cfg
->x_entry_block_ptr
);
1485 seen_ssa_names
.dispose ();
1487 clear_aux_for_blocks ();
1491 /* Do the main work of conditional store replacement. We already know
1492 that the recognized pattern looks like so:
1495 if (cond) goto MIDDLE_BB; else goto JOIN_BB (edge E1)
1498 fallthrough (edge E0)
1502 We check that MIDDLE_BB contains only one store, that that store
1503 doesn't trap (not via NOTRAP, but via checking if an access to the same
1504 memory location dominates us) and that the store has a "simple" RHS. */
1507 cond_store_replacement (basic_block middle_bb
, basic_block join_bb
,
1508 edge e0
, edge e1
, struct pointer_set_t
*nontrap
)
1510 gimple assign
= last_and_only_stmt (middle_bb
);
1511 tree lhs
, rhs
, name
, name2
;
1512 gimple newphi
, new_stmt
;
1513 gimple_stmt_iterator gsi
;
1514 source_location locus
;
1516 /* Check if middle_bb contains of only one store. */
1518 || !gimple_assign_single_p (assign
)
1519 || gimple_has_volatile_ops (assign
))
1522 locus
= gimple_location (assign
);
1523 lhs
= gimple_assign_lhs (assign
);
1524 rhs
= gimple_assign_rhs1 (assign
);
1525 if (TREE_CODE (lhs
) != MEM_REF
1526 || TREE_CODE (TREE_OPERAND (lhs
, 0)) != SSA_NAME
1527 || !is_gimple_reg_type (TREE_TYPE (lhs
)))
1530 /* Prove that we can move the store down. We could also check
1531 TREE_THIS_NOTRAP here, but in that case we also could move stores,
1532 whose value is not available readily, which we want to avoid. */
1533 if (!pointer_set_contains (nontrap
, lhs
))
1536 /* Now we've checked the constraints, so do the transformation:
1537 1) Remove the single store. */
1538 gsi
= gsi_for_stmt (assign
);
1539 unlink_stmt_vdef (assign
);
1540 gsi_remove (&gsi
, true);
1541 release_defs (assign
);
1543 /* 2) Insert a load from the memory of the store to the temporary
1544 on the edge which did not contain the store. */
1545 lhs
= unshare_expr (lhs
);
1546 name
= make_temp_ssa_name (TREE_TYPE (lhs
), NULL
, "cstore");
1547 new_stmt
= gimple_build_assign (name
, lhs
);
1548 gimple_set_location (new_stmt
, locus
);
1549 gsi_insert_on_edge (e1
, new_stmt
);
1551 /* 3) Create a PHI node at the join block, with one argument
1552 holding the old RHS, and the other holding the temporary
1553 where we stored the old memory contents. */
1554 name2
= make_temp_ssa_name (TREE_TYPE (lhs
), NULL
, "cstore");
1555 newphi
= create_phi_node (name2
, join_bb
);
1556 add_phi_arg (newphi
, rhs
, e0
, locus
);
1557 add_phi_arg (newphi
, name
, e1
, locus
);
1559 lhs
= unshare_expr (lhs
);
1560 new_stmt
= gimple_build_assign (lhs
, PHI_RESULT (newphi
));
1562 /* 4) Insert that PHI node. */
1563 gsi
= gsi_after_labels (join_bb
);
1564 if (gsi_end_p (gsi
))
1566 gsi
= gsi_last_bb (join_bb
);
1567 gsi_insert_after (&gsi
, new_stmt
, GSI_NEW_STMT
);
1570 gsi_insert_before (&gsi
, new_stmt
, GSI_NEW_STMT
);
1575 /* Do the main work of conditional store replacement. */
1578 cond_if_else_store_replacement_1 (basic_block then_bb
, basic_block else_bb
,
1579 basic_block join_bb
, gimple then_assign
,
1582 tree lhs_base
, lhs
, then_rhs
, else_rhs
, name
;
1583 source_location then_locus
, else_locus
;
1584 gimple_stmt_iterator gsi
;
1585 gimple newphi
, new_stmt
;
1587 if (then_assign
== NULL
1588 || !gimple_assign_single_p (then_assign
)
1589 || gimple_clobber_p (then_assign
)
1590 || gimple_has_volatile_ops (then_assign
)
1591 || else_assign
== NULL
1592 || !gimple_assign_single_p (else_assign
)
1593 || gimple_clobber_p (else_assign
)
1594 || gimple_has_volatile_ops (else_assign
))
1597 lhs
= gimple_assign_lhs (then_assign
);
1598 if (!is_gimple_reg_type (TREE_TYPE (lhs
))
1599 || !operand_equal_p (lhs
, gimple_assign_lhs (else_assign
), 0))
1602 lhs_base
= get_base_address (lhs
);
1603 if (lhs_base
== NULL_TREE
1604 || (!DECL_P (lhs_base
) && TREE_CODE (lhs_base
) != MEM_REF
))
1607 then_rhs
= gimple_assign_rhs1 (then_assign
);
1608 else_rhs
= gimple_assign_rhs1 (else_assign
);
1609 then_locus
= gimple_location (then_assign
);
1610 else_locus
= gimple_location (else_assign
);
1612 /* Now we've checked the constraints, so do the transformation:
1613 1) Remove the stores. */
1614 gsi
= gsi_for_stmt (then_assign
);
1615 unlink_stmt_vdef (then_assign
);
1616 gsi_remove (&gsi
, true);
1617 release_defs (then_assign
);
1619 gsi
= gsi_for_stmt (else_assign
);
1620 unlink_stmt_vdef (else_assign
);
1621 gsi_remove (&gsi
, true);
1622 release_defs (else_assign
);
1624 /* 2) Create a PHI node at the join block, with one argument
1625 holding the old RHS, and the other holding the temporary
1626 where we stored the old memory contents. */
1627 name
= make_temp_ssa_name (TREE_TYPE (lhs
), NULL
, "cstore");
1628 newphi
= create_phi_node (name
, join_bb
);
1629 add_phi_arg (newphi
, then_rhs
, EDGE_SUCC (then_bb
, 0), then_locus
);
1630 add_phi_arg (newphi
, else_rhs
, EDGE_SUCC (else_bb
, 0), else_locus
);
1632 new_stmt
= gimple_build_assign (lhs
, PHI_RESULT (newphi
));
1634 /* 3) Insert that PHI node. */
1635 gsi
= gsi_after_labels (join_bb
);
1636 if (gsi_end_p (gsi
))
1638 gsi
= gsi_last_bb (join_bb
);
1639 gsi_insert_after (&gsi
, new_stmt
, GSI_NEW_STMT
);
1642 gsi_insert_before (&gsi
, new_stmt
, GSI_NEW_STMT
);
1647 /* Conditional store replacement. We already know
1648 that the recognized pattern looks like so:
1651 if (cond) goto THEN_BB; else goto ELSE_BB (edge E1)
1661 fallthrough (edge E0)
1665 We check that it is safe to sink the store to JOIN_BB by verifying that
1666 there are no read-after-write or write-after-write dependencies in
1667 THEN_BB and ELSE_BB. */
1670 cond_if_else_store_replacement (basic_block then_bb
, basic_block else_bb
,
1671 basic_block join_bb
)
1673 gimple then_assign
= last_and_only_stmt (then_bb
);
1674 gimple else_assign
= last_and_only_stmt (else_bb
);
1675 vec
<data_reference_p
> then_datarefs
, else_datarefs
;
1676 vec
<ddr_p
> then_ddrs
, else_ddrs
;
1677 gimple then_store
, else_store
;
1678 bool found
, ok
= false, res
;
1679 struct data_dependence_relation
*ddr
;
1680 data_reference_p then_dr
, else_dr
;
1682 tree then_lhs
, else_lhs
;
1683 vec
<gimple
> then_stores
, else_stores
;
1684 basic_block blocks
[3];
1686 if (MAX_STORES_TO_SINK
== 0)
1689 /* Handle the case with single statement in THEN_BB and ELSE_BB. */
1690 if (then_assign
&& else_assign
)
1691 return cond_if_else_store_replacement_1 (then_bb
, else_bb
, join_bb
,
1692 then_assign
, else_assign
);
1694 /* Find data references. */
1695 then_datarefs
.create (1);
1696 else_datarefs
.create (1);
1697 if ((find_data_references_in_bb (NULL
, then_bb
, &then_datarefs
)
1699 || !then_datarefs
.length ()
1700 || (find_data_references_in_bb (NULL
, else_bb
, &else_datarefs
)
1702 || !else_datarefs
.length ())
1704 free_data_refs (then_datarefs
);
1705 free_data_refs (else_datarefs
);
1709 /* Find pairs of stores with equal LHS. */
1710 then_stores
.create (1);
1711 else_stores
.create (1);
1712 FOR_EACH_VEC_ELT (then_datarefs
, i
, then_dr
)
1714 if (DR_IS_READ (then_dr
))
1717 then_store
= DR_STMT (then_dr
);
1718 then_lhs
= gimple_get_lhs (then_store
);
1721 FOR_EACH_VEC_ELT (else_datarefs
, j
, else_dr
)
1723 if (DR_IS_READ (else_dr
))
1726 else_store
= DR_STMT (else_dr
);
1727 else_lhs
= gimple_get_lhs (else_store
);
1729 if (operand_equal_p (then_lhs
, else_lhs
, 0))
1739 then_stores
.safe_push (then_store
);
1740 else_stores
.safe_push (else_store
);
1743 /* No pairs of stores found. */
1744 if (!then_stores
.length ()
1745 || then_stores
.length () > (unsigned) MAX_STORES_TO_SINK
)
1747 free_data_refs (then_datarefs
);
1748 free_data_refs (else_datarefs
);
1749 then_stores
.release ();
1750 else_stores
.release ();
1754 /* Compute and check data dependencies in both basic blocks. */
1755 then_ddrs
.create (1);
1756 else_ddrs
.create (1);
1757 if (!compute_all_dependences (then_datarefs
, &then_ddrs
,
1759 || !compute_all_dependences (else_datarefs
, &else_ddrs
,
1762 free_dependence_relations (then_ddrs
);
1763 free_dependence_relations (else_ddrs
);
1764 free_data_refs (then_datarefs
);
1765 free_data_refs (else_datarefs
);
1766 then_stores
.release ();
1767 else_stores
.release ();
1770 blocks
[0] = then_bb
;
1771 blocks
[1] = else_bb
;
1772 blocks
[2] = join_bb
;
1773 renumber_gimple_stmt_uids_in_blocks (blocks
, 3);
1775 /* Check that there are no read-after-write or write-after-write dependencies
1777 FOR_EACH_VEC_ELT (then_ddrs
, i
, ddr
)
1779 struct data_reference
*dra
= DDR_A (ddr
);
1780 struct data_reference
*drb
= DDR_B (ddr
);
1782 if (DDR_ARE_DEPENDENT (ddr
) != chrec_known
1783 && ((DR_IS_READ (dra
) && DR_IS_WRITE (drb
)
1784 && gimple_uid (DR_STMT (dra
)) > gimple_uid (DR_STMT (drb
)))
1785 || (DR_IS_READ (drb
) && DR_IS_WRITE (dra
)
1786 && gimple_uid (DR_STMT (drb
)) > gimple_uid (DR_STMT (dra
)))
1787 || (DR_IS_WRITE (dra
) && DR_IS_WRITE (drb
))))
1789 free_dependence_relations (then_ddrs
);
1790 free_dependence_relations (else_ddrs
);
1791 free_data_refs (then_datarefs
);
1792 free_data_refs (else_datarefs
);
1793 then_stores
.release ();
1794 else_stores
.release ();
1799 /* Check that there are no read-after-write or write-after-write dependencies
1801 FOR_EACH_VEC_ELT (else_ddrs
, i
, ddr
)
1803 struct data_reference
*dra
= DDR_A (ddr
);
1804 struct data_reference
*drb
= DDR_B (ddr
);
1806 if (DDR_ARE_DEPENDENT (ddr
) != chrec_known
1807 && ((DR_IS_READ (dra
) && DR_IS_WRITE (drb
)
1808 && gimple_uid (DR_STMT (dra
)) > gimple_uid (DR_STMT (drb
)))
1809 || (DR_IS_READ (drb
) && DR_IS_WRITE (dra
)
1810 && gimple_uid (DR_STMT (drb
)) > gimple_uid (DR_STMT (dra
)))
1811 || (DR_IS_WRITE (dra
) && DR_IS_WRITE (drb
))))
1813 free_dependence_relations (then_ddrs
);
1814 free_dependence_relations (else_ddrs
);
1815 free_data_refs (then_datarefs
);
1816 free_data_refs (else_datarefs
);
1817 then_stores
.release ();
1818 else_stores
.release ();
1823 /* Sink stores with same LHS. */
1824 FOR_EACH_VEC_ELT (then_stores
, i
, then_store
)
1826 else_store
= else_stores
[i
];
1827 res
= cond_if_else_store_replacement_1 (then_bb
, else_bb
, join_bb
,
1828 then_store
, else_store
);
1832 free_dependence_relations (then_ddrs
);
1833 free_dependence_relations (else_ddrs
);
1834 free_data_refs (then_datarefs
);
1835 free_data_refs (else_datarefs
);
1836 then_stores
.release ();
1837 else_stores
.release ();
1842 /* Return TRUE if STMT has a VUSE whose corresponding VDEF is in BB. */
1845 local_mem_dependence (gimple stmt
, basic_block bb
)
1847 tree vuse
= gimple_vuse (stmt
);
1853 def
= SSA_NAME_DEF_STMT (vuse
);
1854 return (def
&& gimple_bb (def
) == bb
);
1857 /* Given a "diamond" control-flow pattern where BB0 tests a condition,
1858 BB1 and BB2 are "then" and "else" blocks dependent on this test,
1859 and BB3 rejoins control flow following BB1 and BB2, look for
1860 opportunities to hoist loads as follows. If BB3 contains a PHI of
1861 two loads, one each occurring in BB1 and BB2, and the loads are
1862 provably of adjacent fields in the same structure, then move both
1863 loads into BB0. Of course this can only be done if there are no
1864 dependencies preventing such motion.
1866 One of the hoisted loads will always be speculative, so the
1867 transformation is currently conservative:
1869 - The fields must be strictly adjacent.
1870 - The two fields must occupy a single memory block that is
1871 guaranteed to not cross a page boundary.
1873 The last is difficult to prove, as such memory blocks should be
1874 aligned on the minimum of the stack alignment boundary and the
1875 alignment guaranteed by heap allocation interfaces. Thus we rely
1876 on a parameter for the alignment value.
1878 Provided a good value is used for the last case, the first
1879 restriction could possibly be relaxed. */
1882 hoist_adjacent_loads (basic_block bb0
, basic_block bb1
,
1883 basic_block bb2
, basic_block bb3
)
1885 int param_align
= PARAM_VALUE (PARAM_L1_CACHE_LINE_SIZE
);
1886 unsigned param_align_bits
= (unsigned) (param_align
* BITS_PER_UNIT
);
1887 gimple_stmt_iterator gsi
;
1889 /* Walk the phis in bb3 looking for an opportunity. We are looking
1890 for phis of two SSA names, one each of which is defined in bb1 and
1892 for (gsi
= gsi_start_phis (bb3
); !gsi_end_p (gsi
); gsi_next (&gsi
))
1894 gimple phi_stmt
= gsi_stmt (gsi
);
1895 gimple def1
, def2
, defswap
;
1896 tree arg1
, arg2
, ref1
, ref2
, field1
, field2
, fieldswap
;
1897 tree tree_offset1
, tree_offset2
, tree_size2
, next
;
1898 int offset1
, offset2
, size2
;
1900 gimple_stmt_iterator gsi2
;
1901 basic_block bb_for_def1
, bb_for_def2
;
1903 if (gimple_phi_num_args (phi_stmt
) != 2
1904 || virtual_operand_p (gimple_phi_result (phi_stmt
)))
1907 arg1
= gimple_phi_arg_def (phi_stmt
, 0);
1908 arg2
= gimple_phi_arg_def (phi_stmt
, 1);
1910 if (TREE_CODE (arg1
) != SSA_NAME
1911 || TREE_CODE (arg2
) != SSA_NAME
1912 || SSA_NAME_IS_DEFAULT_DEF (arg1
)
1913 || SSA_NAME_IS_DEFAULT_DEF (arg2
))
1916 def1
= SSA_NAME_DEF_STMT (arg1
);
1917 def2
= SSA_NAME_DEF_STMT (arg2
);
1919 if ((gimple_bb (def1
) != bb1
|| gimple_bb (def2
) != bb2
)
1920 && (gimple_bb (def2
) != bb1
|| gimple_bb (def1
) != bb2
))
1923 /* Check the mode of the arguments to be sure a conditional move
1924 can be generated for it. */
1925 if (optab_handler (movcc_optab
, TYPE_MODE (TREE_TYPE (arg1
)))
1926 == CODE_FOR_nothing
)
1929 /* Both statements must be assignments whose RHS is a COMPONENT_REF. */
1930 if (!gimple_assign_single_p (def1
)
1931 || !gimple_assign_single_p (def2
)
1932 || gimple_has_volatile_ops (def1
)
1933 || gimple_has_volatile_ops (def2
))
1936 ref1
= gimple_assign_rhs1 (def1
);
1937 ref2
= gimple_assign_rhs1 (def2
);
1939 if (TREE_CODE (ref1
) != COMPONENT_REF
1940 || TREE_CODE (ref2
) != COMPONENT_REF
)
1943 /* The zeroth operand of the two component references must be
1944 identical. It is not sufficient to compare get_base_address of
1945 the two references, because this could allow for different
1946 elements of the same array in the two trees. It is not safe to
1947 assume that the existence of one array element implies the
1948 existence of a different one. */
1949 if (!operand_equal_p (TREE_OPERAND (ref1
, 0), TREE_OPERAND (ref2
, 0), 0))
1952 field1
= TREE_OPERAND (ref1
, 1);
1953 field2
= TREE_OPERAND (ref2
, 1);
1955 /* Check for field adjacency, and ensure field1 comes first. */
1956 for (next
= DECL_CHAIN (field1
);
1957 next
&& TREE_CODE (next
) != FIELD_DECL
;
1958 next
= DECL_CHAIN (next
))
1963 for (next
= DECL_CHAIN (field2
);
1964 next
&& TREE_CODE (next
) != FIELD_DECL
;
1965 next
= DECL_CHAIN (next
))
1979 bb_for_def1
= gimple_bb (def1
);
1980 bb_for_def2
= gimple_bb (def2
);
1982 /* Check for proper alignment of the first field. */
1983 tree_offset1
= bit_position (field1
);
1984 tree_offset2
= bit_position (field2
);
1985 tree_size2
= DECL_SIZE (field2
);
1987 if (!host_integerp (tree_offset1
, 1)
1988 || !host_integerp (tree_offset2
, 1)
1989 || !host_integerp (tree_size2
, 1))
1992 offset1
= TREE_INT_CST_LOW (tree_offset1
);
1993 offset2
= TREE_INT_CST_LOW (tree_offset2
);
1994 size2
= TREE_INT_CST_LOW (tree_size2
);
1995 align1
= DECL_ALIGN (field1
) % param_align_bits
;
1997 if (offset1
% BITS_PER_UNIT
!= 0)
2000 /* For profitability, the two field references should fit within
2001 a single cache line. */
2002 if (align1
+ offset2
- offset1
+ size2
> param_align_bits
)
2005 /* The two expressions cannot be dependent upon vdefs defined
2007 if (local_mem_dependence (def1
, bb_for_def1
)
2008 || local_mem_dependence (def2
, bb_for_def2
))
2011 /* The conditions are satisfied; hoist the loads from bb1 and bb2 into
2012 bb0. We hoist the first one first so that a cache miss is handled
2013 efficiently regardless of hardware cache-fill policy. */
2014 gsi2
= gsi_for_stmt (def1
);
2015 gsi_move_to_bb_end (&gsi2
, bb0
);
2016 gsi2
= gsi_for_stmt (def2
);
2017 gsi_move_to_bb_end (&gsi2
, bb0
);
2019 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
2022 "\nHoisting adjacent loads from %d and %d into %d: \n",
2023 bb_for_def1
->index
, bb_for_def2
->index
, bb0
->index
);
2024 print_gimple_stmt (dump_file
, def1
, 0, TDF_VOPS
|TDF_MEMSYMS
);
2025 print_gimple_stmt (dump_file
, def2
, 0, TDF_VOPS
|TDF_MEMSYMS
);
2030 /* Determine whether we should attempt to hoist adjacent loads out of
2031 diamond patterns in pass_phiopt. Always hoist loads if
2032 -fhoist-adjacent-loads is specified and the target machine has
2033 both a conditional move instruction and a defined cache line size. */
2036 gate_hoist_loads (void)
2038 return (flag_hoist_adjacent_loads
== 1
2039 && PARAM_VALUE (PARAM_L1_CACHE_LINE_SIZE
)
2040 && HAVE_conditional_move
);
2043 /* Always do these optimizations if we have SSA
2044 trees to work on. */
2053 const pass_data pass_data_phiopt
=
2055 GIMPLE_PASS
, /* type */
2056 "phiopt", /* name */
2057 OPTGROUP_NONE
, /* optinfo_flags */
2058 true, /* has_gate */
2059 true, /* has_execute */
2060 TV_TREE_PHIOPT
, /* tv_id */
2061 ( PROP_cfg
| PROP_ssa
), /* properties_required */
2062 0, /* properties_provided */
2063 0, /* properties_destroyed */
2064 0, /* todo_flags_start */
2065 ( TODO_verify_ssa
| TODO_verify_flow
2066 | TODO_verify_stmts
), /* todo_flags_finish */
2069 class pass_phiopt
: public gimple_opt_pass
2072 pass_phiopt (gcc::context
*ctxt
)
2073 : gimple_opt_pass (pass_data_phiopt
, ctxt
)
2076 /* opt_pass methods: */
2077 opt_pass
* clone () { return new pass_phiopt (m_ctxt
); }
2078 bool gate () { return gate_phiopt (); }
2079 unsigned int execute () { return tree_ssa_phiopt (); }
2081 }; // class pass_phiopt
2086 make_pass_phiopt (gcc::context
*ctxt
)
2088 return new pass_phiopt (ctxt
);
2094 return flag_tree_cselim
;
2099 const pass_data pass_data_cselim
=
2101 GIMPLE_PASS
, /* type */
2102 "cselim", /* name */
2103 OPTGROUP_NONE
, /* optinfo_flags */
2104 true, /* has_gate */
2105 true, /* has_execute */
2106 TV_TREE_PHIOPT
, /* tv_id */
2107 ( PROP_cfg
| PROP_ssa
), /* properties_required */
2108 0, /* properties_provided */
2109 0, /* properties_destroyed */
2110 0, /* todo_flags_start */
2111 ( TODO_verify_ssa
| TODO_verify_flow
2112 | TODO_verify_stmts
), /* todo_flags_finish */
2115 class pass_cselim
: public gimple_opt_pass
2118 pass_cselim (gcc::context
*ctxt
)
2119 : gimple_opt_pass (pass_data_cselim
, ctxt
)
2122 /* opt_pass methods: */
2123 bool gate () { return gate_cselim (); }
2124 unsigned int execute () { return tree_ssa_cs_elim (); }
2126 }; // class pass_cselim
2131 make_pass_cselim (gcc::context
*ctxt
)
2133 return new pass_cselim (ctxt
);