1 /* Optimization of PHI nodes by converting them into straightline code.
2 Copyright (C) 2004-2014 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"
26 #include "stor-layout.h"
29 #include "basic-block.h"
30 #include "pointer-set.h"
31 #include "tree-ssa-alias.h"
32 #include "internal-fn.h"
33 #include "gimple-expr.h"
37 #include "gimple-iterator.h"
38 #include "gimplify-me.h"
39 #include "gimple-ssa.h"
41 #include "tree-phinodes.h"
42 #include "ssa-iterators.h"
43 #include "stringpool.h"
44 #include "tree-ssanames.h"
47 #include "tree-pass.h"
48 #include "langhooks.h"
51 #include "tree-data-ref.h"
52 #include "gimple-pretty-print.h"
53 #include "insn-config.h"
56 #include "tree-scalar-evolution.h"
57 #include "tree-inline.h"
59 #ifndef HAVE_conditional_move
60 #define HAVE_conditional_move (0)
63 static unsigned int tree_ssa_phiopt_worker (bool, bool);
64 static bool conditional_replacement (basic_block
, basic_block
,
65 edge
, edge
, gimple
, tree
, tree
);
66 static int value_replacement (basic_block
, basic_block
,
67 edge
, edge
, gimple
, tree
, tree
);
68 static bool minmax_replacement (basic_block
, basic_block
,
69 edge
, edge
, gimple
, tree
, tree
);
70 static bool abs_replacement (basic_block
, basic_block
,
71 edge
, edge
, gimple
, tree
, tree
);
72 static bool neg_replacement (basic_block
, basic_block
,
73 edge
, edge
, gimple
, tree
, tree
);
74 static bool cond_store_replacement (basic_block
, basic_block
, edge
, edge
,
75 struct pointer_set_t
*);
76 static bool cond_if_else_store_replacement (basic_block
, basic_block
, basic_block
);
77 static struct pointer_set_t
* get_non_trapping (void);
78 static void replace_phi_edge_with_variable (basic_block
, edge
, gimple
, tree
);
79 static void hoist_adjacent_loads (basic_block
, basic_block
,
80 basic_block
, basic_block
);
81 static bool gate_hoist_loads (void);
83 /* This pass tries to transform conditional stores into unconditional
84 ones, enabling further simplifications with the simpler then and else
85 blocks. In particular it replaces this:
88 if (cond) goto bb2; else goto bb1;
96 if (cond) goto bb1; else goto bb2;
100 condtmp = PHI <RHS, condtmp'>
103 This transformation can only be done under several constraints,
104 documented below. It also replaces:
107 if (cond) goto bb2; else goto bb1;
118 if (cond) goto bb3; else goto bb1;
121 condtmp = PHI <RHS1, RHS2>
125 tree_ssa_cs_elim (void)
128 /* ??? We are not interested in loop related info, but the following
129 will create it, ICEing as we didn't init loops with pre-headers.
130 An interfacing issue of find_data_references_in_bb. */
131 loop_optimizer_init (LOOPS_NORMAL
);
133 todo
= tree_ssa_phiopt_worker (true, false);
135 loop_optimizer_finalize ();
139 /* Return the singleton PHI in the SEQ of PHIs for edges E0 and E1. */
142 single_non_singleton_phi_for_edges (gimple_seq seq
, edge e0
, edge e1
)
144 gimple_stmt_iterator i
;
146 if (gimple_seq_singleton_p (seq
))
147 return gsi_stmt (gsi_start (seq
));
148 for (i
= gsi_start (seq
); !gsi_end_p (i
); gsi_next (&i
))
150 gimple p
= gsi_stmt (i
);
151 /* If the PHI arguments are equal then we can skip this PHI. */
152 if (operand_equal_for_phi_arg_p (gimple_phi_arg_def (p
, e0
->dest_idx
),
153 gimple_phi_arg_def (p
, e1
->dest_idx
)))
156 /* If we already have a PHI that has the two edge arguments are
157 different, then return it is not a singleton for these PHIs. */
166 /* The core routine of conditional store replacement and normal
167 phi optimizations. Both share much of the infrastructure in how
168 to match applicable basic block patterns. DO_STORE_ELIM is true
169 when we want to do conditional store replacement, false otherwise.
170 DO_HOIST_LOADS is true when we want to hoist adjacent loads out
171 of diamond control flow patterns, false otherwise. */
173 tree_ssa_phiopt_worker (bool do_store_elim
, bool do_hoist_loads
)
176 basic_block
*bb_order
;
178 bool cfgchanged
= false;
179 struct pointer_set_t
*nontrap
= 0;
182 /* Calculate the set of non-trapping memory accesses. */
183 nontrap
= get_non_trapping ();
185 /* The replacement of conditional negation with a non-branching
186 sequence is really only a win when optimizing for speed and we
187 can avoid transformations by gimple if-conversion that result
188 in poor RTL generation.
190 Ideally either gimple if-conversion or the RTL expanders will
191 be improved and the code to emit branchless conditional negation
193 bool replace_conditional_negation
= false;
195 replace_conditional_negation
196 = ((!optimize_size
&& optimize
>= 2)
197 || (((flag_tree_loop_vectorize
|| cfun
->has_force_vectorize_loops
)
198 && flag_tree_loop_if_convert
!= 0)
199 || flag_tree_loop_if_convert
== 1
200 || flag_tree_loop_if_convert_stores
== 1));
202 /* Search every basic block for COND_EXPR we may be able to optimize.
204 We walk the blocks in order that guarantees that a block with
205 a single predecessor is processed before the predecessor.
206 This ensures that we collapse inner ifs before visiting the
207 outer ones, and also that we do not try to visit a removed
209 bb_order
= single_pred_before_succ_order ();
210 n
= n_basic_blocks_for_fn (cfun
) - NUM_FIXED_BLOCKS
;
212 for (i
= 0; i
< n
; i
++)
214 gimple cond_stmt
, phi
;
215 basic_block bb1
, bb2
;
221 cond_stmt
= last_stmt (bb
);
222 /* Check to see if the last statement is a GIMPLE_COND. */
224 || gimple_code (cond_stmt
) != GIMPLE_COND
)
227 e1
= EDGE_SUCC (bb
, 0);
229 e2
= EDGE_SUCC (bb
, 1);
232 /* We cannot do the optimization on abnormal edges. */
233 if ((e1
->flags
& EDGE_ABNORMAL
) != 0
234 || (e2
->flags
& EDGE_ABNORMAL
) != 0)
237 /* If either bb1's succ or bb2 or bb2's succ is non NULL. */
238 if (EDGE_COUNT (bb1
->succs
) == 0
240 || EDGE_COUNT (bb2
->succs
) == 0)
243 /* Find the bb which is the fall through to the other. */
244 if (EDGE_SUCC (bb1
, 0)->dest
== bb2
)
246 else if (EDGE_SUCC (bb2
, 0)->dest
== bb1
)
248 basic_block bb_tmp
= bb1
;
255 else if (do_store_elim
256 && EDGE_SUCC (bb1
, 0)->dest
== EDGE_SUCC (bb2
, 0)->dest
)
258 basic_block bb3
= EDGE_SUCC (bb1
, 0)->dest
;
260 if (!single_succ_p (bb1
)
261 || (EDGE_SUCC (bb1
, 0)->flags
& EDGE_FALLTHRU
) == 0
262 || !single_succ_p (bb2
)
263 || (EDGE_SUCC (bb2
, 0)->flags
& EDGE_FALLTHRU
) == 0
264 || EDGE_COUNT (bb3
->preds
) != 2)
266 if (cond_if_else_store_replacement (bb1
, bb2
, bb3
))
270 else if (do_hoist_loads
271 && EDGE_SUCC (bb1
, 0)->dest
== EDGE_SUCC (bb2
, 0)->dest
)
273 basic_block bb3
= EDGE_SUCC (bb1
, 0)->dest
;
275 if (!FLOAT_TYPE_P (TREE_TYPE (gimple_cond_lhs (cond_stmt
)))
276 && single_succ_p (bb1
)
277 && single_succ_p (bb2
)
278 && single_pred_p (bb1
)
279 && single_pred_p (bb2
)
280 && EDGE_COUNT (bb
->succs
) == 2
281 && EDGE_COUNT (bb3
->preds
) == 2
282 /* If one edge or the other is dominant, a conditional move
283 is likely to perform worse than the well-predicted branch. */
284 && !predictable_edge_p (EDGE_SUCC (bb
, 0))
285 && !predictable_edge_p (EDGE_SUCC (bb
, 1)))
286 hoist_adjacent_loads (bb
, bb1
, bb2
, bb3
);
292 e1
= EDGE_SUCC (bb1
, 0);
294 /* Make sure that bb1 is just a fall through. */
295 if (!single_succ_p (bb1
)
296 || (e1
->flags
& EDGE_FALLTHRU
) == 0)
299 /* Also make sure that bb1 only have one predecessor and that it
301 if (!single_pred_p (bb1
)
302 || single_pred (bb1
) != bb
)
307 /* bb1 is the middle block, bb2 the join block, bb the split block,
308 e1 the fallthrough edge from bb1 to bb2. We can't do the
309 optimization if the join block has more than two predecessors. */
310 if (EDGE_COUNT (bb2
->preds
) > 2)
312 if (cond_store_replacement (bb1
, bb2
, e1
, e2
, nontrap
))
317 gimple_seq phis
= phi_nodes (bb2
);
318 gimple_stmt_iterator gsi
;
319 bool candorest
= true;
321 /* Value replacement can work with more than one PHI
322 so try that first. */
323 for (gsi
= gsi_start (phis
); !gsi_end_p (gsi
); gsi_next (&gsi
))
325 phi
= gsi_stmt (gsi
);
326 arg0
= gimple_phi_arg_def (phi
, e1
->dest_idx
);
327 arg1
= gimple_phi_arg_def (phi
, e2
->dest_idx
);
328 if (value_replacement (bb
, bb1
, e1
, e2
, phi
, arg0
, arg1
) == 2)
339 phi
= single_non_singleton_phi_for_edges (phis
, e1
, e2
);
343 arg0
= gimple_phi_arg_def (phi
, e1
->dest_idx
);
344 arg1
= gimple_phi_arg_def (phi
, e2
->dest_idx
);
346 /* Something is wrong if we cannot find the arguments in the PHI
348 gcc_assert (arg0
!= NULL
&& arg1
!= NULL
);
350 /* Do the replacement of conditional if it can be done. */
351 if (conditional_replacement (bb
, bb1
, e1
, e2
, phi
, arg0
, arg1
))
353 else if (abs_replacement (bb
, bb1
, e1
, e2
, phi
, arg0
, arg1
))
355 else if (replace_conditional_negation
356 && neg_replacement (bb
, bb1
, e1
, e2
, phi
, arg0
, arg1
))
358 else if (minmax_replacement (bb
, bb1
, e1
, e2
, phi
, arg0
, arg1
))
366 pointer_set_destroy (nontrap
);
367 /* If the CFG has changed, we should cleanup the CFG. */
368 if (cfgchanged
&& do_store_elim
)
370 /* In cond-store replacement we have added some loads on edges
371 and new VOPS (as we moved the store, and created a load). */
372 gsi_commit_edge_inserts ();
373 return TODO_cleanup_cfg
| TODO_update_ssa_only_virtuals
;
376 return TODO_cleanup_cfg
;
380 /* Replace PHI node element whose edge is E in block BB with variable NEW.
381 Remove the edge from COND_BLOCK which does not lead to BB (COND_BLOCK
382 is known to have two edges, one of which must reach BB). */
385 replace_phi_edge_with_variable (basic_block cond_block
,
386 edge e
, gimple phi
, tree new_tree
)
388 basic_block bb
= gimple_bb (phi
);
389 basic_block block_to_remove
;
390 gimple_stmt_iterator gsi
;
392 /* Change the PHI argument to new. */
393 SET_USE (PHI_ARG_DEF_PTR (phi
, e
->dest_idx
), new_tree
);
395 /* Remove the empty basic block. */
396 if (EDGE_SUCC (cond_block
, 0)->dest
== bb
)
398 EDGE_SUCC (cond_block
, 0)->flags
|= EDGE_FALLTHRU
;
399 EDGE_SUCC (cond_block
, 0)->flags
&= ~(EDGE_TRUE_VALUE
| EDGE_FALSE_VALUE
);
400 EDGE_SUCC (cond_block
, 0)->probability
= REG_BR_PROB_BASE
;
401 EDGE_SUCC (cond_block
, 0)->count
+= EDGE_SUCC (cond_block
, 1)->count
;
403 block_to_remove
= EDGE_SUCC (cond_block
, 1)->dest
;
407 EDGE_SUCC (cond_block
, 1)->flags
|= EDGE_FALLTHRU
;
408 EDGE_SUCC (cond_block
, 1)->flags
409 &= ~(EDGE_TRUE_VALUE
| EDGE_FALSE_VALUE
);
410 EDGE_SUCC (cond_block
, 1)->probability
= REG_BR_PROB_BASE
;
411 EDGE_SUCC (cond_block
, 1)->count
+= EDGE_SUCC (cond_block
, 0)->count
;
413 block_to_remove
= EDGE_SUCC (cond_block
, 0)->dest
;
415 delete_basic_block (block_to_remove
);
417 /* Eliminate the COND_EXPR at the end of COND_BLOCK. */
418 gsi
= gsi_last_bb (cond_block
);
419 gsi_remove (&gsi
, true);
421 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
423 "COND_EXPR in block %d and PHI in block %d converted to straightline code.\n",
428 /* The function conditional_replacement does the main work of doing the
429 conditional replacement. Return true if the replacement is done.
430 Otherwise return false.
431 BB is the basic block where the replacement is going to be done on. ARG0
432 is argument 0 from PHI. Likewise for ARG1. */
435 conditional_replacement (basic_block cond_bb
, basic_block middle_bb
,
436 edge e0
, edge e1
, gimple phi
,
437 tree arg0
, tree arg1
)
440 gimple stmt
, new_stmt
;
442 gimple_stmt_iterator gsi
;
443 edge true_edge
, false_edge
;
444 tree new_var
, new_var2
;
447 /* FIXME: Gimplification of complex type is too hard for now. */
448 /* We aren't prepared to handle vectors either (and it is a question
449 if it would be worthwhile anyway). */
450 if (!(INTEGRAL_TYPE_P (TREE_TYPE (arg0
))
451 || POINTER_TYPE_P (TREE_TYPE (arg0
)))
452 || !(INTEGRAL_TYPE_P (TREE_TYPE (arg1
))
453 || POINTER_TYPE_P (TREE_TYPE (arg1
))))
456 /* The PHI arguments have the constants 0 and 1, or 0 and -1, then
457 convert it to the conditional. */
458 if ((integer_zerop (arg0
) && integer_onep (arg1
))
459 || (integer_zerop (arg1
) && integer_onep (arg0
)))
461 else if ((integer_zerop (arg0
) && integer_all_onesp (arg1
))
462 || (integer_zerop (arg1
) && integer_all_onesp (arg0
)))
467 if (!empty_block_p (middle_bb
))
470 /* At this point we know we have a GIMPLE_COND with two successors.
471 One successor is BB, the other successor is an empty block which
472 falls through into BB.
474 There is a single PHI node at the join point (BB) and its arguments
475 are constants (0, 1) or (0, -1).
477 So, given the condition COND, and the two PHI arguments, we can
478 rewrite this PHI into non-branching code:
480 dest = (COND) or dest = COND'
482 We use the condition as-is if the argument associated with the
483 true edge has the value one or the argument associated with the
484 false edge as the value zero. Note that those conditions are not
485 the same since only one of the outgoing edges from the GIMPLE_COND
486 will directly reach BB and thus be associated with an argument. */
488 stmt
= last_stmt (cond_bb
);
489 result
= PHI_RESULT (phi
);
491 /* To handle special cases like floating point comparison, it is easier and
492 less error-prone to build a tree and gimplify it on the fly though it is
494 cond
= fold_build2_loc (gimple_location (stmt
),
495 gimple_cond_code (stmt
), boolean_type_node
,
496 gimple_cond_lhs (stmt
), gimple_cond_rhs (stmt
));
498 /* We need to know which is the true edge and which is the false
499 edge so that we know when to invert the condition below. */
500 extract_true_false_edges_from_block (cond_bb
, &true_edge
, &false_edge
);
501 if ((e0
== true_edge
&& integer_zerop (arg0
))
502 || (e0
== false_edge
&& !integer_zerop (arg0
))
503 || (e1
== true_edge
&& integer_zerop (arg1
))
504 || (e1
== false_edge
&& !integer_zerop (arg1
)))
505 cond
= fold_build1_loc (gimple_location (stmt
),
506 TRUTH_NOT_EXPR
, TREE_TYPE (cond
), cond
);
510 cond
= fold_convert_loc (gimple_location (stmt
),
511 TREE_TYPE (result
), cond
);
512 cond
= fold_build1_loc (gimple_location (stmt
),
513 NEGATE_EXPR
, TREE_TYPE (cond
), cond
);
516 /* Insert our new statements at the end of conditional block before the
518 gsi
= gsi_for_stmt (stmt
);
519 new_var
= force_gimple_operand_gsi (&gsi
, cond
, true, NULL
, true,
522 if (!useless_type_conversion_p (TREE_TYPE (result
), TREE_TYPE (new_var
)))
524 source_location locus_0
, locus_1
;
526 new_var2
= make_ssa_name (TREE_TYPE (result
), NULL
);
527 new_stmt
= gimple_build_assign_with_ops (CONVERT_EXPR
, new_var2
,
529 gsi_insert_before (&gsi
, new_stmt
, GSI_SAME_STMT
);
532 /* Set the locus to the first argument, unless is doesn't have one. */
533 locus_0
= gimple_phi_arg_location (phi
, 0);
534 locus_1
= gimple_phi_arg_location (phi
, 1);
535 if (locus_0
== UNKNOWN_LOCATION
)
537 gimple_set_location (new_stmt
, locus_0
);
540 replace_phi_edge_with_variable (cond_bb
, e1
, phi
, new_var
);
542 /* Note that we optimized this PHI. */
546 /* Update *ARG which is defined in STMT so that it contains the
547 computed value if that seems profitable. Return true if the
548 statement is made dead by that rewriting. */
551 jump_function_from_stmt (tree
*arg
, gimple stmt
)
553 enum tree_code code
= gimple_assign_rhs_code (stmt
);
554 if (code
== ADDR_EXPR
)
556 /* For arg = &p->i transform it to p, if possible. */
557 tree rhs1
= gimple_assign_rhs1 (stmt
);
558 HOST_WIDE_INT offset
;
559 tree tem
= get_addr_base_and_unit_offset (TREE_OPERAND (rhs1
, 0),
562 && TREE_CODE (tem
) == MEM_REF
563 && (mem_ref_offset (tem
) + offset
) == 0)
565 *arg
= TREE_OPERAND (tem
, 0);
569 /* TODO: Much like IPA-CP jump-functions we want to handle constant
570 additions symbolically here, and we'd need to update the comparison
571 code that compares the arg + cst tuples in our caller. For now the
572 code above exactly handles the VEC_BASE pattern from vec.h. */
576 /* RHS is a source argument in a BIT_AND_EXPR which feeds a conditional
577 of the form SSA_NAME NE 0.
579 If RHS is fed by a simple EQ_EXPR comparison of two values, see if
580 the two input values of the EQ_EXPR match arg0 and arg1.
582 If so update *code and return TRUE. Otherwise return FALSE. */
585 rhs_is_fed_for_value_replacement (const_tree arg0
, const_tree arg1
,
586 enum tree_code
*code
, const_tree rhs
)
588 /* Obviously if RHS is not an SSA_NAME, we can't look at the defining
590 if (TREE_CODE (rhs
) == SSA_NAME
)
592 gimple def1
= SSA_NAME_DEF_STMT (rhs
);
594 /* Verify the defining statement has an EQ_EXPR on the RHS. */
595 if (is_gimple_assign (def1
) && gimple_assign_rhs_code (def1
) == EQ_EXPR
)
597 /* Finally verify the source operands of the EQ_EXPR are equal
599 tree op0
= gimple_assign_rhs1 (def1
);
600 tree op1
= gimple_assign_rhs2 (def1
);
601 if ((operand_equal_for_phi_arg_p (arg0
, op0
)
602 && operand_equal_for_phi_arg_p (arg1
, op1
))
603 || (operand_equal_for_phi_arg_p (arg0
, op1
)
604 && operand_equal_for_phi_arg_p (arg1
, op0
)))
606 /* We will perform the optimization. */
607 *code
= gimple_assign_rhs_code (def1
);
615 /* Return TRUE if arg0/arg1 are equal to the rhs/lhs or lhs/rhs of COND.
617 Also return TRUE if arg0/arg1 are equal to the source arguments of a
618 an EQ comparison feeding a BIT_AND_EXPR which feeds COND.
620 Return FALSE otherwise. */
623 operand_equal_for_value_replacement (const_tree arg0
, const_tree arg1
,
624 enum tree_code
*code
, gimple cond
)
627 tree lhs
= gimple_cond_lhs (cond
);
628 tree rhs
= gimple_cond_rhs (cond
);
630 if ((operand_equal_for_phi_arg_p (arg0
, lhs
)
631 && operand_equal_for_phi_arg_p (arg1
, rhs
))
632 || (operand_equal_for_phi_arg_p (arg1
, lhs
)
633 && operand_equal_for_phi_arg_p (arg0
, rhs
)))
636 /* Now handle more complex case where we have an EQ comparison
637 which feeds a BIT_AND_EXPR which feeds COND.
639 First verify that COND is of the form SSA_NAME NE 0. */
640 if (*code
!= NE_EXPR
|| !integer_zerop (rhs
)
641 || TREE_CODE (lhs
) != SSA_NAME
)
644 /* Now ensure that SSA_NAME is set by a BIT_AND_EXPR. */
645 def
= SSA_NAME_DEF_STMT (lhs
);
646 if (!is_gimple_assign (def
) || gimple_assign_rhs_code (def
) != BIT_AND_EXPR
)
649 /* Now verify arg0/arg1 correspond to the source arguments of an
650 EQ comparison feeding the BIT_AND_EXPR. */
652 tree tmp
= gimple_assign_rhs1 (def
);
653 if (rhs_is_fed_for_value_replacement (arg0
, arg1
, code
, tmp
))
656 tmp
= gimple_assign_rhs2 (def
);
657 if (rhs_is_fed_for_value_replacement (arg0
, arg1
, code
, tmp
))
663 /* Returns true if ARG is a neutral element for operation CODE
664 on the RIGHT side. */
667 neutral_element_p (tree_code code
, tree arg
, bool right
)
674 return integer_zerop (arg
);
681 case POINTER_PLUS_EXPR
:
682 return right
&& integer_zerop (arg
);
685 return integer_onep (arg
);
692 return right
&& integer_onep (arg
);
695 return integer_all_onesp (arg
);
702 /* Returns true if ARG is an absorbing element for operation CODE. */
705 absorbing_element_p (tree_code code
, tree arg
)
710 return integer_all_onesp (arg
);
714 return integer_zerop (arg
);
721 /* The function value_replacement does the main work of doing the value
722 replacement. Return non-zero if the replacement is done. Otherwise return
723 0. If we remove the middle basic block, return 2.
724 BB is the basic block where the replacement is going to be done on. ARG0
725 is argument 0 from the PHI. Likewise for ARG1. */
728 value_replacement (basic_block cond_bb
, basic_block middle_bb
,
729 edge e0
, edge e1
, gimple phi
,
730 tree arg0
, tree arg1
)
732 gimple_stmt_iterator gsi
;
734 edge true_edge
, false_edge
;
736 bool emtpy_or_with_defined_p
= true;
738 /* If the type says honor signed zeros we cannot do this
740 if (HONOR_SIGNED_ZEROS (TYPE_MODE (TREE_TYPE (arg1
))))
743 /* If there is a statement in MIDDLE_BB that defines one of the PHI
744 arguments, then adjust arg0 or arg1. */
745 gsi
= gsi_start_nondebug_after_labels_bb (middle_bb
);
746 while (!gsi_end_p (gsi
))
748 gimple stmt
= gsi_stmt (gsi
);
750 gsi_next_nondebug (&gsi
);
751 if (!is_gimple_assign (stmt
))
753 emtpy_or_with_defined_p
= false;
756 /* Now try to adjust arg0 or arg1 according to the computation
758 lhs
= gimple_assign_lhs (stmt
);
760 && jump_function_from_stmt (&arg0
, stmt
))
762 && jump_function_from_stmt (&arg1
, stmt
)))
763 emtpy_or_with_defined_p
= false;
766 cond
= last_stmt (cond_bb
);
767 code
= gimple_cond_code (cond
);
769 /* This transformation is only valid for equality comparisons. */
770 if (code
!= NE_EXPR
&& code
!= EQ_EXPR
)
773 /* We need to know which is the true edge and which is the false
774 edge so that we know if have abs or negative abs. */
775 extract_true_false_edges_from_block (cond_bb
, &true_edge
, &false_edge
);
777 /* At this point we know we have a COND_EXPR with two successors.
778 One successor is BB, the other successor is an empty block which
779 falls through into BB.
781 The condition for the COND_EXPR is known to be NE_EXPR or EQ_EXPR.
783 There is a single PHI node at the join point (BB) with two arguments.
785 We now need to verify that the two arguments in the PHI node match
786 the two arguments to the equality comparison. */
788 if (operand_equal_for_value_replacement (arg0
, arg1
, &code
, cond
))
793 /* For NE_EXPR, we want to build an assignment result = arg where
794 arg is the PHI argument associated with the true edge. For
795 EQ_EXPR we want the PHI argument associated with the false edge. */
796 e
= (code
== NE_EXPR
? true_edge
: false_edge
);
798 /* Unfortunately, E may not reach BB (it may instead have gone to
799 OTHER_BLOCK). If that is the case, then we want the single outgoing
800 edge from OTHER_BLOCK which reaches BB and represents the desired
801 path from COND_BLOCK. */
802 if (e
->dest
== middle_bb
)
803 e
= single_succ_edge (e
->dest
);
805 /* Now we know the incoming edge to BB that has the argument for the
806 RHS of our new assignment statement. */
812 /* If the middle basic block was empty or is defining the
813 PHI arguments and this is a single phi where the args are different
814 for the edges e0 and e1 then we can remove the middle basic block. */
815 if (emtpy_or_with_defined_p
816 && single_non_singleton_phi_for_edges (phi_nodes (gimple_bb (phi
)),
819 replace_phi_edge_with_variable (cond_bb
, e1
, phi
, arg
);
820 /* Note that we optimized this PHI. */
825 /* Replace the PHI arguments with arg. */
826 SET_PHI_ARG_DEF (phi
, e0
->dest_idx
, arg
);
827 SET_PHI_ARG_DEF (phi
, e1
->dest_idx
, arg
);
828 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
830 fprintf (dump_file
, "PHI ");
831 print_generic_expr (dump_file
, gimple_phi_result (phi
), 0);
832 fprintf (dump_file
, " reduced for COND_EXPR in block %d to ",
834 print_generic_expr (dump_file
, arg
, 0);
835 fprintf (dump_file
, ".\n");
842 /* Now optimize (x != 0) ? x + y : y to just y.
843 The following condition is too restrictive, there can easily be another
844 stmt in middle_bb, for instance a CONVERT_EXPR for the second argument. */
845 gimple assign
= last_and_only_stmt (middle_bb
);
846 if (!assign
|| gimple_code (assign
) != GIMPLE_ASSIGN
847 || gimple_assign_rhs_class (assign
) != GIMPLE_BINARY_RHS
848 || (!INTEGRAL_TYPE_P (TREE_TYPE (arg0
))
849 && !POINTER_TYPE_P (TREE_TYPE (arg0
))))
852 /* Only transform if it removes the condition. */
853 if (!single_non_singleton_phi_for_edges (phi_nodes (gimple_bb (phi
)), e0
, e1
))
856 /* Size-wise, this is always profitable. */
857 if (optimize_bb_for_speed_p (cond_bb
)
858 /* The special case is useless if it has a low probability. */
859 && profile_status_for_fn (cfun
) != PROFILE_ABSENT
860 && EDGE_PRED (middle_bb
, 0)->probability
< PROB_EVEN
861 /* If assign is cheap, there is no point avoiding it. */
862 && estimate_num_insns (assign
, &eni_time_weights
)
863 >= 3 * estimate_num_insns (cond
, &eni_time_weights
))
866 tree lhs
= gimple_assign_lhs (assign
);
867 tree rhs1
= gimple_assign_rhs1 (assign
);
868 tree rhs2
= gimple_assign_rhs2 (assign
);
869 enum tree_code code_def
= gimple_assign_rhs_code (assign
);
870 tree cond_lhs
= gimple_cond_lhs (cond
);
871 tree cond_rhs
= gimple_cond_rhs (cond
);
873 if (((code
== NE_EXPR
&& e1
== false_edge
)
874 || (code
== EQ_EXPR
&& e1
== true_edge
))
877 && operand_equal_for_phi_arg_p (rhs2
, cond_lhs
)
878 && neutral_element_p (code_def
, cond_rhs
, true))
880 && operand_equal_for_phi_arg_p (rhs1
, cond_lhs
)
881 && neutral_element_p (code_def
, cond_rhs
, false))
882 || (operand_equal_for_phi_arg_p (arg1
, cond_rhs
)
883 && (operand_equal_for_phi_arg_p (rhs2
, cond_lhs
)
884 || operand_equal_for_phi_arg_p (rhs1
, cond_lhs
))
885 && absorbing_element_p (code_def
, cond_rhs
))))
887 gsi
= gsi_for_stmt (cond
);
888 gimple_stmt_iterator gsi_from
= gsi_for_stmt (assign
);
889 gsi_move_before (&gsi_from
, &gsi
);
890 replace_phi_edge_with_variable (cond_bb
, e1
, phi
, lhs
);
897 /* The function minmax_replacement does the main work of doing the minmax
898 replacement. Return true if the replacement is done. Otherwise return
900 BB is the basic block where the replacement is going to be done on. ARG0
901 is argument 0 from the PHI. Likewise for ARG1. */
904 minmax_replacement (basic_block cond_bb
, basic_block middle_bb
,
905 edge e0
, edge e1
, gimple phi
,
906 tree arg0
, tree arg1
)
909 gimple cond
, new_stmt
;
910 edge true_edge
, false_edge
;
911 enum tree_code cmp
, minmax
, ass_code
;
912 tree smaller
, larger
, arg_true
, arg_false
;
913 gimple_stmt_iterator gsi
, gsi_from
;
915 type
= TREE_TYPE (PHI_RESULT (phi
));
917 /* The optimization may be unsafe due to NaNs. */
918 if (HONOR_NANS (TYPE_MODE (type
)))
921 cond
= last_stmt (cond_bb
);
922 cmp
= gimple_cond_code (cond
);
924 /* This transformation is only valid for order comparisons. Record which
925 operand is smaller/larger if the result of the comparison is true. */
926 if (cmp
== LT_EXPR
|| cmp
== LE_EXPR
)
928 smaller
= gimple_cond_lhs (cond
);
929 larger
= gimple_cond_rhs (cond
);
931 else if (cmp
== GT_EXPR
|| cmp
== GE_EXPR
)
933 smaller
= gimple_cond_rhs (cond
);
934 larger
= gimple_cond_lhs (cond
);
939 /* We need to know which is the true edge and which is the false
940 edge so that we know if have abs or negative abs. */
941 extract_true_false_edges_from_block (cond_bb
, &true_edge
, &false_edge
);
943 /* Forward the edges over the middle basic block. */
944 if (true_edge
->dest
== middle_bb
)
945 true_edge
= EDGE_SUCC (true_edge
->dest
, 0);
946 if (false_edge
->dest
== middle_bb
)
947 false_edge
= EDGE_SUCC (false_edge
->dest
, 0);
951 gcc_assert (false_edge
== e1
);
957 gcc_assert (false_edge
== e0
);
958 gcc_assert (true_edge
== e1
);
963 if (empty_block_p (middle_bb
))
965 if (operand_equal_for_phi_arg_p (arg_true
, smaller
)
966 && operand_equal_for_phi_arg_p (arg_false
, larger
))
970 if (smaller < larger)
976 else if (operand_equal_for_phi_arg_p (arg_false
, smaller
)
977 && operand_equal_for_phi_arg_p (arg_true
, larger
))
984 /* Recognize the following case, assuming d <= u:
990 This is equivalent to
995 gimple assign
= last_and_only_stmt (middle_bb
);
996 tree lhs
, op0
, op1
, bound
;
999 || gimple_code (assign
) != GIMPLE_ASSIGN
)
1002 lhs
= gimple_assign_lhs (assign
);
1003 ass_code
= gimple_assign_rhs_code (assign
);
1004 if (ass_code
!= MAX_EXPR
&& ass_code
!= MIN_EXPR
)
1006 op0
= gimple_assign_rhs1 (assign
);
1007 op1
= gimple_assign_rhs2 (assign
);
1009 if (true_edge
->src
== middle_bb
)
1011 /* We got here if the condition is true, i.e., SMALLER < LARGER. */
1012 if (!operand_equal_for_phi_arg_p (lhs
, arg_true
))
1015 if (operand_equal_for_phi_arg_p (arg_false
, larger
))
1019 if (smaller < larger)
1021 r' = MAX_EXPR (smaller, bound)
1023 r = PHI <r', larger> --> to be turned to MIN_EXPR. */
1024 if (ass_code
!= MAX_EXPR
)
1028 if (operand_equal_for_phi_arg_p (op0
, smaller
))
1030 else if (operand_equal_for_phi_arg_p (op1
, smaller
))
1035 /* We need BOUND <= LARGER. */
1036 if (!integer_nonzerop (fold_build2 (LE_EXPR
, boolean_type_node
,
1040 else if (operand_equal_for_phi_arg_p (arg_false
, smaller
))
1044 if (smaller < larger)
1046 r' = MIN_EXPR (larger, bound)
1048 r = PHI <r', smaller> --> to be turned to MAX_EXPR. */
1049 if (ass_code
!= MIN_EXPR
)
1053 if (operand_equal_for_phi_arg_p (op0
, larger
))
1055 else if (operand_equal_for_phi_arg_p (op1
, larger
))
1060 /* We need BOUND >= SMALLER. */
1061 if (!integer_nonzerop (fold_build2 (GE_EXPR
, boolean_type_node
,
1070 /* We got here if the condition is false, i.e., SMALLER > LARGER. */
1071 if (!operand_equal_for_phi_arg_p (lhs
, arg_false
))
1074 if (operand_equal_for_phi_arg_p (arg_true
, larger
))
1078 if (smaller > larger)
1080 r' = MIN_EXPR (smaller, bound)
1082 r = PHI <r', larger> --> to be turned to MAX_EXPR. */
1083 if (ass_code
!= MIN_EXPR
)
1087 if (operand_equal_for_phi_arg_p (op0
, smaller
))
1089 else if (operand_equal_for_phi_arg_p (op1
, smaller
))
1094 /* We need BOUND >= LARGER. */
1095 if (!integer_nonzerop (fold_build2 (GE_EXPR
, boolean_type_node
,
1099 else if (operand_equal_for_phi_arg_p (arg_true
, smaller
))
1103 if (smaller > larger)
1105 r' = MAX_EXPR (larger, bound)
1107 r = PHI <r', smaller> --> to be turned to MIN_EXPR. */
1108 if (ass_code
!= MAX_EXPR
)
1112 if (operand_equal_for_phi_arg_p (op0
, larger
))
1114 else if (operand_equal_for_phi_arg_p (op1
, larger
))
1119 /* We need BOUND <= SMALLER. */
1120 if (!integer_nonzerop (fold_build2 (LE_EXPR
, boolean_type_node
,
1128 /* Move the statement from the middle block. */
1129 gsi
= gsi_last_bb (cond_bb
);
1130 gsi_from
= gsi_last_nondebug_bb (middle_bb
);
1131 gsi_move_before (&gsi_from
, &gsi
);
1134 /* Emit the statement to compute min/max. */
1135 result
= duplicate_ssa_name (PHI_RESULT (phi
), NULL
);
1136 new_stmt
= gimple_build_assign_with_ops (minmax
, result
, arg0
, arg1
);
1137 gsi
= gsi_last_bb (cond_bb
);
1138 gsi_insert_before (&gsi
, new_stmt
, GSI_NEW_STMT
);
1140 replace_phi_edge_with_variable (cond_bb
, e1
, phi
, result
);
1144 /* The function absolute_replacement does the main work of doing the absolute
1145 replacement. Return true if the replacement is done. Otherwise return
1147 bb is the basic block where the replacement is going to be done on. arg0
1148 is argument 0 from the phi. Likewise for arg1. */
1151 abs_replacement (basic_block cond_bb
, basic_block middle_bb
,
1152 edge e0 ATTRIBUTE_UNUSED
, edge e1
,
1153 gimple phi
, tree arg0
, tree arg1
)
1156 gimple new_stmt
, cond
;
1157 gimple_stmt_iterator gsi
;
1158 edge true_edge
, false_edge
;
1163 enum tree_code cond_code
;
1165 /* If the type says honor signed zeros we cannot do this
1167 if (HONOR_SIGNED_ZEROS (TYPE_MODE (TREE_TYPE (arg1
))))
1170 /* OTHER_BLOCK must have only one executable statement which must have the
1171 form arg0 = -arg1 or arg1 = -arg0. */
1173 assign
= last_and_only_stmt (middle_bb
);
1174 /* If we did not find the proper negation assignment, then we can not
1179 /* If we got here, then we have found the only executable statement
1180 in OTHER_BLOCK. If it is anything other than arg = -arg1 or
1181 arg1 = -arg0, then we can not optimize. */
1182 if (gimple_code (assign
) != GIMPLE_ASSIGN
)
1185 lhs
= gimple_assign_lhs (assign
);
1187 if (gimple_assign_rhs_code (assign
) != NEGATE_EXPR
)
1190 rhs
= gimple_assign_rhs1 (assign
);
1192 /* The assignment has to be arg0 = -arg1 or arg1 = -arg0. */
1193 if (!(lhs
== arg0
&& rhs
== arg1
)
1194 && !(lhs
== arg1
&& rhs
== arg0
))
1197 cond
= last_stmt (cond_bb
);
1198 result
= PHI_RESULT (phi
);
1200 /* Only relationals comparing arg[01] against zero are interesting. */
1201 cond_code
= gimple_cond_code (cond
);
1202 if (cond_code
!= GT_EXPR
&& cond_code
!= GE_EXPR
1203 && cond_code
!= LT_EXPR
&& cond_code
!= LE_EXPR
)
1206 /* Make sure the conditional is arg[01] OP y. */
1207 if (gimple_cond_lhs (cond
) != rhs
)
1210 if (FLOAT_TYPE_P (TREE_TYPE (gimple_cond_rhs (cond
)))
1211 ? real_zerop (gimple_cond_rhs (cond
))
1212 : integer_zerop (gimple_cond_rhs (cond
)))
1217 /* We need to know which is the true edge and which is the false
1218 edge so that we know if have abs or negative abs. */
1219 extract_true_false_edges_from_block (cond_bb
, &true_edge
, &false_edge
);
1221 /* For GT_EXPR/GE_EXPR, if the true edge goes to OTHER_BLOCK, then we
1222 will need to negate the result. Similarly for LT_EXPR/LE_EXPR if
1223 the false edge goes to OTHER_BLOCK. */
1224 if (cond_code
== GT_EXPR
|| cond_code
== GE_EXPR
)
1229 if (e
->dest
== middle_bb
)
1234 result
= duplicate_ssa_name (result
, NULL
);
1237 lhs
= make_ssa_name (TREE_TYPE (result
), NULL
);
1241 /* Build the modify expression with abs expression. */
1242 new_stmt
= gimple_build_assign_with_ops (ABS_EXPR
, lhs
, rhs
, NULL
);
1244 gsi
= gsi_last_bb (cond_bb
);
1245 gsi_insert_before (&gsi
, new_stmt
, GSI_NEW_STMT
);
1249 /* Get the right GSI. We want to insert after the recently
1250 added ABS_EXPR statement (which we know is the first statement
1252 new_stmt
= gimple_build_assign_with_ops (NEGATE_EXPR
, result
, lhs
, NULL
);
1254 gsi_insert_after (&gsi
, new_stmt
, GSI_NEW_STMT
);
1257 replace_phi_edge_with_variable (cond_bb
, e1
, phi
, result
);
1259 /* Note that we optimized this PHI. */
1263 /* The function neg_replacement replaces conditional negation with
1264 equivalent straight line code. Returns TRUE if replacement is done,
1265 otherwise returns FALSE.
1267 COND_BB branches around negation occuring in MIDDLE_BB.
1269 E0 and E1 are edges out of COND_BB. E0 reaches MIDDLE_BB and
1270 E1 reaches the other successor which should contain PHI with
1271 arguments ARG0 and ARG1.
1273 Assuming negation is to occur when the condition is true,
1274 then the non-branching sequence is:
1276 result = (rhs ^ -cond) + cond
1278 Inverting the condition or its result gives us negation
1279 when the original condition is false. */
1282 neg_replacement (basic_block cond_bb
, basic_block middle_bb
,
1283 edge e0 ATTRIBUTE_UNUSED
, edge e1
,
1284 gimple phi
, tree arg0
, tree arg1
)
1286 gimple new_stmt
, cond
;
1287 gimple_stmt_iterator gsi
;
1289 edge true_edge
, false_edge
;
1291 enum tree_code cond_code
;
1292 bool invert
= false;
1294 /* This transformation performs logical operations on the
1295 incoming arguments. So force them to be integral types. */
1296 if (!INTEGRAL_TYPE_P (TREE_TYPE (arg0
)))
1299 /* OTHER_BLOCK must have only one executable statement which must have the
1300 form arg0 = -arg1 or arg1 = -arg0. */
1302 assign
= last_and_only_stmt (middle_bb
);
1303 /* If we did not find the proper negation assignment, then we can not
1308 /* If we got here, then we have found the only executable statement
1309 in OTHER_BLOCK. If it is anything other than arg0 = -arg1 or
1310 arg1 = -arg0, then we can not optimize. */
1311 if (gimple_code (assign
) != GIMPLE_ASSIGN
)
1314 lhs
= gimple_assign_lhs (assign
);
1316 if (gimple_assign_rhs_code (assign
) != NEGATE_EXPR
)
1319 rhs
= gimple_assign_rhs1 (assign
);
1321 /* The assignment has to be arg0 = -arg1 or arg1 = -arg0. */
1322 if (!(lhs
== arg0
&& rhs
== arg1
)
1323 && !(lhs
== arg1
&& rhs
== arg0
))
1326 /* The basic sequence assumes we negate when the condition is true.
1327 If we need the opposite, then we will either need to invert the
1328 condition or its result. */
1329 extract_true_false_edges_from_block (cond_bb
, &true_edge
, &false_edge
);
1330 invert
= false_edge
->dest
== middle_bb
;
1332 /* Unlike abs_replacement, we can handle arbitrary conditionals here. */
1333 cond
= last_stmt (cond_bb
);
1334 cond_code
= gimple_cond_code (cond
);
1336 /* If inversion is needed, first try to invert the test since
1341 = HONOR_NANS (TYPE_MODE (TREE_TYPE (gimple_cond_lhs (cond
))));
1342 enum tree_code new_code
= invert_tree_comparison (cond_code
, honor_nans
);
1344 /* If invert_tree_comparison was successful, then use its return
1345 value as the new code and note that inversion is no longer
1347 if (new_code
!= ERROR_MARK
)
1349 cond_code
= new_code
;
1354 tree cond_val
= make_ssa_name (boolean_type_node
, NULL
);
1355 new_stmt
= gimple_build_assign_with_ops (cond_code
, cond_val
,
1356 gimple_cond_lhs (cond
),
1357 gimple_cond_rhs (cond
));
1358 gsi
= gsi_last_bb (cond_bb
);
1359 gsi_insert_before (&gsi
, new_stmt
, GSI_NEW_STMT
);
1361 /* If we still need inversion, then invert the result of the
1365 tree tmp
= make_ssa_name (boolean_type_node
, NULL
);
1366 new_stmt
= gimple_build_assign_with_ops (BIT_XOR_EXPR
, tmp
,
1367 cond_val
, boolean_true_node
);
1368 gsi_insert_after (&gsi
, new_stmt
, GSI_NEW_STMT
);
1372 /* Get the condition in the right type so that we can perform
1373 logical and arithmetic operations on it. */
1374 tree cond_val_converted
= make_ssa_name (TREE_TYPE (rhs
), NULL
);
1375 new_stmt
= gimple_build_assign_with_ops (NOP_EXPR
, cond_val_converted
,
1376 cond_val
, NULL_TREE
);
1377 gsi_insert_after (&gsi
, new_stmt
, GSI_NEW_STMT
);
1379 tree neg_cond_val_converted
= make_ssa_name (TREE_TYPE (rhs
), NULL
);
1380 new_stmt
= gimple_build_assign_with_ops (NEGATE_EXPR
, neg_cond_val_converted
,
1381 cond_val_converted
, NULL_TREE
);
1382 gsi_insert_after (&gsi
, new_stmt
, GSI_NEW_STMT
);
1384 tree tmp
= make_ssa_name (TREE_TYPE (rhs
), NULL
);
1385 new_stmt
= gimple_build_assign_with_ops (BIT_XOR_EXPR
, tmp
,
1386 rhs
, neg_cond_val_converted
);
1387 gsi_insert_after (&gsi
, new_stmt
, GSI_NEW_STMT
);
1389 tree new_lhs
= make_ssa_name (TREE_TYPE (rhs
), NULL
);
1390 new_stmt
= gimple_build_assign_with_ops (PLUS_EXPR
, new_lhs
,
1391 tmp
, cond_val_converted
);
1392 gsi_insert_after (&gsi
, new_stmt
, GSI_NEW_STMT
);
1394 replace_phi_edge_with_variable (cond_bb
, e1
, phi
, new_lhs
);
1396 /* Note that we optimized this PHI. */
1400 /* Auxiliary functions to determine the set of memory accesses which
1401 can't trap because they are preceded by accesses to the same memory
1402 portion. We do that for MEM_REFs, so we only need to track
1403 the SSA_NAME of the pointer indirectly referenced. The algorithm
1404 simply is a walk over all instructions in dominator order. When
1405 we see an MEM_REF we determine if we've already seen a same
1406 ref anywhere up to the root of the dominator tree. If we do the
1407 current access can't trap. If we don't see any dominating access
1408 the current access might trap, but might also make later accesses
1409 non-trapping, so we remember it. We need to be careful with loads
1410 or stores, for instance a load might not trap, while a store would,
1411 so if we see a dominating read access this doesn't mean that a later
1412 write access would not trap. Hence we also need to differentiate the
1413 type of access(es) seen.
1415 ??? We currently are very conservative and assume that a load might
1416 trap even if a store doesn't (write-only memory). This probably is
1417 overly conservative. */
1419 /* A hash-table of SSA_NAMEs, and in which basic block an MEM_REF
1420 through it was seen, which would constitute a no-trap region for
1424 unsigned int ssa_name_ver
;
1427 HOST_WIDE_INT offset
, size
;
1431 /* Hashtable helpers. */
1433 struct ssa_names_hasher
: typed_free_remove
<name_to_bb
>
1435 typedef name_to_bb value_type
;
1436 typedef name_to_bb compare_type
;
1437 static inline hashval_t
hash (const value_type
*);
1438 static inline bool equal (const value_type
*, const compare_type
*);
1441 /* Used for quick clearing of the hash-table when we see calls.
1442 Hash entries with phase < nt_call_phase are invalid. */
1443 static unsigned int nt_call_phase
;
1445 /* The hash function. */
1448 ssa_names_hasher::hash (const value_type
*n
)
1450 return n
->ssa_name_ver
^ (((hashval_t
) n
->store
) << 31)
1451 ^ (n
->offset
<< 6) ^ (n
->size
<< 3);
1454 /* The equality function of *P1 and *P2. */
1457 ssa_names_hasher::equal (const value_type
*n1
, const compare_type
*n2
)
1459 return n1
->ssa_name_ver
== n2
->ssa_name_ver
1460 && n1
->store
== n2
->store
1461 && n1
->offset
== n2
->offset
1462 && n1
->size
== n2
->size
;
1465 /* The hash table for remembering what we've seen. */
1466 static hash_table
<ssa_names_hasher
> seen_ssa_names
;
1468 /* We see the expression EXP in basic block BB. If it's an interesting
1469 expression (an MEM_REF through an SSA_NAME) possibly insert the
1470 expression into the set NONTRAP or the hash table of seen expressions.
1471 STORE is true if this expression is on the LHS, otherwise it's on
1474 add_or_mark_expr (basic_block bb
, tree exp
,
1475 struct pointer_set_t
*nontrap
, bool store
)
1479 if (TREE_CODE (exp
) == MEM_REF
1480 && TREE_CODE (TREE_OPERAND (exp
, 0)) == SSA_NAME
1481 && tree_fits_shwi_p (TREE_OPERAND (exp
, 1))
1482 && (size
= int_size_in_bytes (TREE_TYPE (exp
))) > 0)
1484 tree name
= TREE_OPERAND (exp
, 0);
1485 struct name_to_bb map
;
1487 struct name_to_bb
*n2bb
;
1488 basic_block found_bb
= 0;
1490 /* Try to find the last seen MEM_REF through the same
1491 SSA_NAME, which can trap. */
1492 map
.ssa_name_ver
= SSA_NAME_VERSION (name
);
1496 map
.offset
= tree_to_shwi (TREE_OPERAND (exp
, 1));
1499 slot
= seen_ssa_names
.find_slot (&map
, INSERT
);
1501 if (n2bb
&& n2bb
->phase
>= nt_call_phase
)
1502 found_bb
= n2bb
->bb
;
1504 /* If we've found a trapping MEM_REF, _and_ it dominates EXP
1505 (it's in a basic block on the path from us to the dominator root)
1506 then we can't trap. */
1507 if (found_bb
&& (((size_t)found_bb
->aux
) & 1) == 1)
1509 pointer_set_insert (nontrap
, exp
);
1513 /* EXP might trap, so insert it into the hash table. */
1516 n2bb
->phase
= nt_call_phase
;
1521 n2bb
= XNEW (struct name_to_bb
);
1522 n2bb
->ssa_name_ver
= SSA_NAME_VERSION (name
);
1523 n2bb
->phase
= nt_call_phase
;
1525 n2bb
->store
= store
;
1526 n2bb
->offset
= map
.offset
;
1534 class nontrapping_dom_walker
: public dom_walker
1537 nontrapping_dom_walker (cdi_direction direction
, pointer_set_t
*ps
)
1538 : dom_walker (direction
), m_nontrapping (ps
) {}
1540 virtual void before_dom_children (basic_block
);
1541 virtual void after_dom_children (basic_block
);
1544 pointer_set_t
*m_nontrapping
;
1547 /* Called by walk_dominator_tree, when entering the block BB. */
1549 nontrapping_dom_walker::before_dom_children (basic_block bb
)
1553 gimple_stmt_iterator gsi
;
1555 /* If we haven't seen all our predecessors, clear the hash-table. */
1556 FOR_EACH_EDGE (e
, ei
, bb
->preds
)
1557 if ((((size_t)e
->src
->aux
) & 2) == 0)
1563 /* Mark this BB as being on the path to dominator root and as visited. */
1564 bb
->aux
= (void*)(1 | 2);
1566 /* And walk the statements in order. */
1567 for (gsi
= gsi_start_bb (bb
); !gsi_end_p (gsi
); gsi_next (&gsi
))
1569 gimple stmt
= gsi_stmt (gsi
);
1571 if (is_gimple_call (stmt
) && !nonfreeing_call_p (stmt
))
1573 else if (gimple_assign_single_p (stmt
) && !gimple_has_volatile_ops (stmt
))
1575 add_or_mark_expr (bb
, gimple_assign_lhs (stmt
), m_nontrapping
, true);
1576 add_or_mark_expr (bb
, gimple_assign_rhs1 (stmt
), m_nontrapping
, false);
1581 /* Called by walk_dominator_tree, when basic block BB is exited. */
1583 nontrapping_dom_walker::after_dom_children (basic_block bb
)
1585 /* This BB isn't on the path to dominator root anymore. */
1589 /* This is the entry point of gathering non trapping memory accesses.
1590 It will do a dominator walk over the whole function, and it will
1591 make use of the bb->aux pointers. It returns a set of trees
1592 (the MEM_REFs itself) which can't trap. */
1593 static struct pointer_set_t
*
1594 get_non_trapping (void)
1597 pointer_set_t
*nontrap
= pointer_set_create ();
1598 seen_ssa_names
.create (128);
1599 /* We're going to do a dominator walk, so ensure that we have
1600 dominance information. */
1601 calculate_dominance_info (CDI_DOMINATORS
);
1603 nontrapping_dom_walker (CDI_DOMINATORS
, nontrap
)
1604 .walk (cfun
->cfg
->x_entry_block_ptr
);
1606 seen_ssa_names
.dispose ();
1608 clear_aux_for_blocks ();
1612 /* Do the main work of conditional store replacement. We already know
1613 that the recognized pattern looks like so:
1616 if (cond) goto MIDDLE_BB; else goto JOIN_BB (edge E1)
1619 fallthrough (edge E0)
1623 We check that MIDDLE_BB contains only one store, that that store
1624 doesn't trap (not via NOTRAP, but via checking if an access to the same
1625 memory location dominates us) and that the store has a "simple" RHS. */
1628 cond_store_replacement (basic_block middle_bb
, basic_block join_bb
,
1629 edge e0
, edge e1
, struct pointer_set_t
*nontrap
)
1631 gimple assign
= last_and_only_stmt (middle_bb
);
1632 tree lhs
, rhs
, name
, name2
;
1633 gimple newphi
, new_stmt
;
1634 gimple_stmt_iterator gsi
;
1635 source_location locus
;
1637 /* Check if middle_bb contains of only one store. */
1639 || !gimple_assign_single_p (assign
)
1640 || gimple_has_volatile_ops (assign
))
1643 locus
= gimple_location (assign
);
1644 lhs
= gimple_assign_lhs (assign
);
1645 rhs
= gimple_assign_rhs1 (assign
);
1646 if (TREE_CODE (lhs
) != MEM_REF
1647 || TREE_CODE (TREE_OPERAND (lhs
, 0)) != SSA_NAME
1648 || !is_gimple_reg_type (TREE_TYPE (lhs
)))
1651 /* Prove that we can move the store down. We could also check
1652 TREE_THIS_NOTRAP here, but in that case we also could move stores,
1653 whose value is not available readily, which we want to avoid. */
1654 if (!pointer_set_contains (nontrap
, lhs
))
1657 /* Now we've checked the constraints, so do the transformation:
1658 1) Remove the single store. */
1659 gsi
= gsi_for_stmt (assign
);
1660 unlink_stmt_vdef (assign
);
1661 gsi_remove (&gsi
, true);
1662 release_defs (assign
);
1664 /* 2) Insert a load from the memory of the store to the temporary
1665 on the edge which did not contain the store. */
1666 lhs
= unshare_expr (lhs
);
1667 name
= make_temp_ssa_name (TREE_TYPE (lhs
), NULL
, "cstore");
1668 new_stmt
= gimple_build_assign (name
, lhs
);
1669 gimple_set_location (new_stmt
, locus
);
1670 gsi_insert_on_edge (e1
, new_stmt
);
1672 /* 3) Create a PHI node at the join block, with one argument
1673 holding the old RHS, and the other holding the temporary
1674 where we stored the old memory contents. */
1675 name2
= make_temp_ssa_name (TREE_TYPE (lhs
), NULL
, "cstore");
1676 newphi
= create_phi_node (name2
, join_bb
);
1677 add_phi_arg (newphi
, rhs
, e0
, locus
);
1678 add_phi_arg (newphi
, name
, e1
, locus
);
1680 lhs
= unshare_expr (lhs
);
1681 new_stmt
= gimple_build_assign (lhs
, PHI_RESULT (newphi
));
1683 /* 4) Insert that PHI node. */
1684 gsi
= gsi_after_labels (join_bb
);
1685 if (gsi_end_p (gsi
))
1687 gsi
= gsi_last_bb (join_bb
);
1688 gsi_insert_after (&gsi
, new_stmt
, GSI_NEW_STMT
);
1691 gsi_insert_before (&gsi
, new_stmt
, GSI_NEW_STMT
);
1696 /* Do the main work of conditional store replacement. */
1699 cond_if_else_store_replacement_1 (basic_block then_bb
, basic_block else_bb
,
1700 basic_block join_bb
, gimple then_assign
,
1703 tree lhs_base
, lhs
, then_rhs
, else_rhs
, name
;
1704 source_location then_locus
, else_locus
;
1705 gimple_stmt_iterator gsi
;
1706 gimple newphi
, new_stmt
;
1708 if (then_assign
== NULL
1709 || !gimple_assign_single_p (then_assign
)
1710 || gimple_clobber_p (then_assign
)
1711 || gimple_has_volatile_ops (then_assign
)
1712 || else_assign
== NULL
1713 || !gimple_assign_single_p (else_assign
)
1714 || gimple_clobber_p (else_assign
)
1715 || gimple_has_volatile_ops (else_assign
))
1718 lhs
= gimple_assign_lhs (then_assign
);
1719 if (!is_gimple_reg_type (TREE_TYPE (lhs
))
1720 || !operand_equal_p (lhs
, gimple_assign_lhs (else_assign
), 0))
1723 lhs_base
= get_base_address (lhs
);
1724 if (lhs_base
== NULL_TREE
1725 || (!DECL_P (lhs_base
) && TREE_CODE (lhs_base
) != MEM_REF
))
1728 then_rhs
= gimple_assign_rhs1 (then_assign
);
1729 else_rhs
= gimple_assign_rhs1 (else_assign
);
1730 then_locus
= gimple_location (then_assign
);
1731 else_locus
= gimple_location (else_assign
);
1733 /* Now we've checked the constraints, so do the transformation:
1734 1) Remove the stores. */
1735 gsi
= gsi_for_stmt (then_assign
);
1736 unlink_stmt_vdef (then_assign
);
1737 gsi_remove (&gsi
, true);
1738 release_defs (then_assign
);
1740 gsi
= gsi_for_stmt (else_assign
);
1741 unlink_stmt_vdef (else_assign
);
1742 gsi_remove (&gsi
, true);
1743 release_defs (else_assign
);
1745 /* 2) Create a PHI node at the join block, with one argument
1746 holding the old RHS, and the other holding the temporary
1747 where we stored the old memory contents. */
1748 name
= make_temp_ssa_name (TREE_TYPE (lhs
), NULL
, "cstore");
1749 newphi
= create_phi_node (name
, join_bb
);
1750 add_phi_arg (newphi
, then_rhs
, EDGE_SUCC (then_bb
, 0), then_locus
);
1751 add_phi_arg (newphi
, else_rhs
, EDGE_SUCC (else_bb
, 0), else_locus
);
1753 new_stmt
= gimple_build_assign (lhs
, PHI_RESULT (newphi
));
1755 /* 3) Insert that PHI node. */
1756 gsi
= gsi_after_labels (join_bb
);
1757 if (gsi_end_p (gsi
))
1759 gsi
= gsi_last_bb (join_bb
);
1760 gsi_insert_after (&gsi
, new_stmt
, GSI_NEW_STMT
);
1763 gsi_insert_before (&gsi
, new_stmt
, GSI_NEW_STMT
);
1768 /* Conditional store replacement. We already know
1769 that the recognized pattern looks like so:
1772 if (cond) goto THEN_BB; else goto ELSE_BB (edge E1)
1782 fallthrough (edge E0)
1786 We check that it is safe to sink the store to JOIN_BB by verifying that
1787 there are no read-after-write or write-after-write dependencies in
1788 THEN_BB and ELSE_BB. */
1791 cond_if_else_store_replacement (basic_block then_bb
, basic_block else_bb
,
1792 basic_block join_bb
)
1794 gimple then_assign
= last_and_only_stmt (then_bb
);
1795 gimple else_assign
= last_and_only_stmt (else_bb
);
1796 vec
<data_reference_p
> then_datarefs
, else_datarefs
;
1797 vec
<ddr_p
> then_ddrs
, else_ddrs
;
1798 gimple then_store
, else_store
;
1799 bool found
, ok
= false, res
;
1800 struct data_dependence_relation
*ddr
;
1801 data_reference_p then_dr
, else_dr
;
1803 tree then_lhs
, else_lhs
;
1804 basic_block blocks
[3];
1806 if (MAX_STORES_TO_SINK
== 0)
1809 /* Handle the case with single statement in THEN_BB and ELSE_BB. */
1810 if (then_assign
&& else_assign
)
1811 return cond_if_else_store_replacement_1 (then_bb
, else_bb
, join_bb
,
1812 then_assign
, else_assign
);
1814 /* Find data references. */
1815 then_datarefs
.create (1);
1816 else_datarefs
.create (1);
1817 if ((find_data_references_in_bb (NULL
, then_bb
, &then_datarefs
)
1819 || !then_datarefs
.length ()
1820 || (find_data_references_in_bb (NULL
, else_bb
, &else_datarefs
)
1822 || !else_datarefs
.length ())
1824 free_data_refs (then_datarefs
);
1825 free_data_refs (else_datarefs
);
1829 /* Find pairs of stores with equal LHS. */
1830 auto_vec
<gimple
, 1> then_stores
, else_stores
;
1831 FOR_EACH_VEC_ELT (then_datarefs
, i
, then_dr
)
1833 if (DR_IS_READ (then_dr
))
1836 then_store
= DR_STMT (then_dr
);
1837 then_lhs
= gimple_get_lhs (then_store
);
1838 if (then_lhs
== NULL_TREE
)
1842 FOR_EACH_VEC_ELT (else_datarefs
, j
, else_dr
)
1844 if (DR_IS_READ (else_dr
))
1847 else_store
= DR_STMT (else_dr
);
1848 else_lhs
= gimple_get_lhs (else_store
);
1849 if (else_lhs
== NULL_TREE
)
1852 if (operand_equal_p (then_lhs
, else_lhs
, 0))
1862 then_stores
.safe_push (then_store
);
1863 else_stores
.safe_push (else_store
);
1866 /* No pairs of stores found. */
1867 if (!then_stores
.length ()
1868 || then_stores
.length () > (unsigned) MAX_STORES_TO_SINK
)
1870 free_data_refs (then_datarefs
);
1871 free_data_refs (else_datarefs
);
1875 /* Compute and check data dependencies in both basic blocks. */
1876 then_ddrs
.create (1);
1877 else_ddrs
.create (1);
1878 if (!compute_all_dependences (then_datarefs
, &then_ddrs
,
1880 || !compute_all_dependences (else_datarefs
, &else_ddrs
,
1883 free_dependence_relations (then_ddrs
);
1884 free_dependence_relations (else_ddrs
);
1885 free_data_refs (then_datarefs
);
1886 free_data_refs (else_datarefs
);
1889 blocks
[0] = then_bb
;
1890 blocks
[1] = else_bb
;
1891 blocks
[2] = join_bb
;
1892 renumber_gimple_stmt_uids_in_blocks (blocks
, 3);
1894 /* Check that there are no read-after-write or write-after-write dependencies
1896 FOR_EACH_VEC_ELT (then_ddrs
, i
, ddr
)
1898 struct data_reference
*dra
= DDR_A (ddr
);
1899 struct data_reference
*drb
= DDR_B (ddr
);
1901 if (DDR_ARE_DEPENDENT (ddr
) != chrec_known
1902 && ((DR_IS_READ (dra
) && DR_IS_WRITE (drb
)
1903 && gimple_uid (DR_STMT (dra
)) > gimple_uid (DR_STMT (drb
)))
1904 || (DR_IS_READ (drb
) && DR_IS_WRITE (dra
)
1905 && gimple_uid (DR_STMT (drb
)) > gimple_uid (DR_STMT (dra
)))
1906 || (DR_IS_WRITE (dra
) && DR_IS_WRITE (drb
))))
1908 free_dependence_relations (then_ddrs
);
1909 free_dependence_relations (else_ddrs
);
1910 free_data_refs (then_datarefs
);
1911 free_data_refs (else_datarefs
);
1916 /* Check that there are no read-after-write or write-after-write dependencies
1918 FOR_EACH_VEC_ELT (else_ddrs
, i
, ddr
)
1920 struct data_reference
*dra
= DDR_A (ddr
);
1921 struct data_reference
*drb
= DDR_B (ddr
);
1923 if (DDR_ARE_DEPENDENT (ddr
) != chrec_known
1924 && ((DR_IS_READ (dra
) && DR_IS_WRITE (drb
)
1925 && gimple_uid (DR_STMT (dra
)) > gimple_uid (DR_STMT (drb
)))
1926 || (DR_IS_READ (drb
) && DR_IS_WRITE (dra
)
1927 && gimple_uid (DR_STMT (drb
)) > gimple_uid (DR_STMT (dra
)))
1928 || (DR_IS_WRITE (dra
) && DR_IS_WRITE (drb
))))
1930 free_dependence_relations (then_ddrs
);
1931 free_dependence_relations (else_ddrs
);
1932 free_data_refs (then_datarefs
);
1933 free_data_refs (else_datarefs
);
1938 /* Sink stores with same LHS. */
1939 FOR_EACH_VEC_ELT (then_stores
, i
, then_store
)
1941 else_store
= else_stores
[i
];
1942 res
= cond_if_else_store_replacement_1 (then_bb
, else_bb
, join_bb
,
1943 then_store
, else_store
);
1947 free_dependence_relations (then_ddrs
);
1948 free_dependence_relations (else_ddrs
);
1949 free_data_refs (then_datarefs
);
1950 free_data_refs (else_datarefs
);
1955 /* Return TRUE if STMT has a VUSE whose corresponding VDEF is in BB. */
1958 local_mem_dependence (gimple stmt
, basic_block bb
)
1960 tree vuse
= gimple_vuse (stmt
);
1966 def
= SSA_NAME_DEF_STMT (vuse
);
1967 return (def
&& gimple_bb (def
) == bb
);
1970 /* Given a "diamond" control-flow pattern where BB0 tests a condition,
1971 BB1 and BB2 are "then" and "else" blocks dependent on this test,
1972 and BB3 rejoins control flow following BB1 and BB2, look for
1973 opportunities to hoist loads as follows. If BB3 contains a PHI of
1974 two loads, one each occurring in BB1 and BB2, and the loads are
1975 provably of adjacent fields in the same structure, then move both
1976 loads into BB0. Of course this can only be done if there are no
1977 dependencies preventing such motion.
1979 One of the hoisted loads will always be speculative, so the
1980 transformation is currently conservative:
1982 - The fields must be strictly adjacent.
1983 - The two fields must occupy a single memory block that is
1984 guaranteed to not cross a page boundary.
1986 The last is difficult to prove, as such memory blocks should be
1987 aligned on the minimum of the stack alignment boundary and the
1988 alignment guaranteed by heap allocation interfaces. Thus we rely
1989 on a parameter for the alignment value.
1991 Provided a good value is used for the last case, the first
1992 restriction could possibly be relaxed. */
1995 hoist_adjacent_loads (basic_block bb0
, basic_block bb1
,
1996 basic_block bb2
, basic_block bb3
)
1998 int param_align
= PARAM_VALUE (PARAM_L1_CACHE_LINE_SIZE
);
1999 unsigned param_align_bits
= (unsigned) (param_align
* BITS_PER_UNIT
);
2000 gimple_stmt_iterator gsi
;
2002 /* Walk the phis in bb3 looking for an opportunity. We are looking
2003 for phis of two SSA names, one each of which is defined in bb1 and
2005 for (gsi
= gsi_start_phis (bb3
); !gsi_end_p (gsi
); gsi_next (&gsi
))
2007 gimple phi_stmt
= gsi_stmt (gsi
);
2008 gimple def1
, def2
, defswap
;
2009 tree arg1
, arg2
, ref1
, ref2
, field1
, field2
, fieldswap
;
2010 tree tree_offset1
, tree_offset2
, tree_size2
, next
;
2011 int offset1
, offset2
, size2
;
2013 gimple_stmt_iterator gsi2
;
2014 basic_block bb_for_def1
, bb_for_def2
;
2016 if (gimple_phi_num_args (phi_stmt
) != 2
2017 || virtual_operand_p (gimple_phi_result (phi_stmt
)))
2020 arg1
= gimple_phi_arg_def (phi_stmt
, 0);
2021 arg2
= gimple_phi_arg_def (phi_stmt
, 1);
2023 if (TREE_CODE (arg1
) != SSA_NAME
2024 || TREE_CODE (arg2
) != SSA_NAME
2025 || SSA_NAME_IS_DEFAULT_DEF (arg1
)
2026 || SSA_NAME_IS_DEFAULT_DEF (arg2
))
2029 def1
= SSA_NAME_DEF_STMT (arg1
);
2030 def2
= SSA_NAME_DEF_STMT (arg2
);
2032 if ((gimple_bb (def1
) != bb1
|| gimple_bb (def2
) != bb2
)
2033 && (gimple_bb (def2
) != bb1
|| gimple_bb (def1
) != bb2
))
2036 /* Check the mode of the arguments to be sure a conditional move
2037 can be generated for it. */
2038 if (optab_handler (movcc_optab
, TYPE_MODE (TREE_TYPE (arg1
)))
2039 == CODE_FOR_nothing
)
2042 /* Both statements must be assignments whose RHS is a COMPONENT_REF. */
2043 if (!gimple_assign_single_p (def1
)
2044 || !gimple_assign_single_p (def2
)
2045 || gimple_has_volatile_ops (def1
)
2046 || gimple_has_volatile_ops (def2
))
2049 ref1
= gimple_assign_rhs1 (def1
);
2050 ref2
= gimple_assign_rhs1 (def2
);
2052 if (TREE_CODE (ref1
) != COMPONENT_REF
2053 || TREE_CODE (ref2
) != COMPONENT_REF
)
2056 /* The zeroth operand of the two component references must be
2057 identical. It is not sufficient to compare get_base_address of
2058 the two references, because this could allow for different
2059 elements of the same array in the two trees. It is not safe to
2060 assume that the existence of one array element implies the
2061 existence of a different one. */
2062 if (!operand_equal_p (TREE_OPERAND (ref1
, 0), TREE_OPERAND (ref2
, 0), 0))
2065 field1
= TREE_OPERAND (ref1
, 1);
2066 field2
= TREE_OPERAND (ref2
, 1);
2068 /* Check for field adjacency, and ensure field1 comes first. */
2069 for (next
= DECL_CHAIN (field1
);
2070 next
&& TREE_CODE (next
) != FIELD_DECL
;
2071 next
= DECL_CHAIN (next
))
2076 for (next
= DECL_CHAIN (field2
);
2077 next
&& TREE_CODE (next
) != FIELD_DECL
;
2078 next
= DECL_CHAIN (next
))
2092 bb_for_def1
= gimple_bb (def1
);
2093 bb_for_def2
= gimple_bb (def2
);
2095 /* Check for proper alignment of the first field. */
2096 tree_offset1
= bit_position (field1
);
2097 tree_offset2
= bit_position (field2
);
2098 tree_size2
= DECL_SIZE (field2
);
2100 if (!tree_fits_uhwi_p (tree_offset1
)
2101 || !tree_fits_uhwi_p (tree_offset2
)
2102 || !tree_fits_uhwi_p (tree_size2
))
2105 offset1
= tree_to_uhwi (tree_offset1
);
2106 offset2
= tree_to_uhwi (tree_offset2
);
2107 size2
= tree_to_uhwi (tree_size2
);
2108 align1
= DECL_ALIGN (field1
) % param_align_bits
;
2110 if (offset1
% BITS_PER_UNIT
!= 0)
2113 /* For profitability, the two field references should fit within
2114 a single cache line. */
2115 if (align1
+ offset2
- offset1
+ size2
> param_align_bits
)
2118 /* The two expressions cannot be dependent upon vdefs defined
2120 if (local_mem_dependence (def1
, bb_for_def1
)
2121 || local_mem_dependence (def2
, bb_for_def2
))
2124 /* The conditions are satisfied; hoist the loads from bb1 and bb2 into
2125 bb0. We hoist the first one first so that a cache miss is handled
2126 efficiently regardless of hardware cache-fill policy. */
2127 gsi2
= gsi_for_stmt (def1
);
2128 gsi_move_to_bb_end (&gsi2
, bb0
);
2129 gsi2
= gsi_for_stmt (def2
);
2130 gsi_move_to_bb_end (&gsi2
, bb0
);
2132 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
2135 "\nHoisting adjacent loads from %d and %d into %d: \n",
2136 bb_for_def1
->index
, bb_for_def2
->index
, bb0
->index
);
2137 print_gimple_stmt (dump_file
, def1
, 0, TDF_VOPS
|TDF_MEMSYMS
);
2138 print_gimple_stmt (dump_file
, def2
, 0, TDF_VOPS
|TDF_MEMSYMS
);
2143 /* Determine whether we should attempt to hoist adjacent loads out of
2144 diamond patterns in pass_phiopt. Always hoist loads if
2145 -fhoist-adjacent-loads is specified and the target machine has
2146 both a conditional move instruction and a defined cache line size. */
2149 gate_hoist_loads (void)
2151 return (flag_hoist_adjacent_loads
== 1
2152 && PARAM_VALUE (PARAM_L1_CACHE_LINE_SIZE
)
2153 && HAVE_conditional_move
);
2156 /* This pass tries to replaces an if-then-else block with an
2157 assignment. We have four kinds of transformations. Some of these
2158 transformations are also performed by the ifcvt RTL optimizer.
2160 Conditional Replacement
2161 -----------------------
2163 This transformation, implemented in conditional_replacement,
2167 if (cond) goto bb2; else goto bb1;
2170 x = PHI <0 (bb1), 1 (bb0), ...>;
2178 x = PHI <x' (bb0), ...>;
2180 We remove bb1 as it becomes unreachable. This occurs often due to
2181 gimplification of conditionals.
2186 This transformation, implemented in value_replacement, replaces
2189 if (a != b) goto bb2; else goto bb1;
2192 x = PHI <a (bb1), b (bb0), ...>;
2198 x = PHI <b (bb0), ...>;
2200 This opportunity can sometimes occur as a result of other
2204 Another case caught by value replacement looks like this:
2210 if (t3 != 0) goto bb1; else goto bb2;
2226 This transformation, implemented in abs_replacement, replaces
2229 if (a >= 0) goto bb2; else goto bb1;
2233 x = PHI <x (bb1), a (bb0), ...>;
2240 x = PHI <x' (bb0), ...>;
2245 This transformation, minmax_replacement replaces
2248 if (a <= b) goto bb2; else goto bb1;
2251 x = PHI <b (bb1), a (bb0), ...>;
2256 x' = MIN_EXPR (a, b)
2258 x = PHI <x' (bb0), ...>;
2260 A similar transformation is done for MAX_EXPR.
2263 This pass also performs a fifth transformation of a slightly different
2266 Adjacent Load Hoisting
2267 ----------------------
2269 This transformation replaces
2272 if (...) goto bb2; else goto bb1;
2274 x1 = (<expr>).field1;
2277 x2 = (<expr>).field2;
2284 x1 = (<expr>).field1;
2285 x2 = (<expr>).field2;
2286 if (...) goto bb2; else goto bb1;
2293 The purpose of this transformation is to enable generation of conditional
2294 move instructions such as Intel CMOVE or PowerPC ISEL. Because one of
2295 the loads is speculative, the transformation is restricted to very
2296 specific cases to avoid introducing a page fault. We are looking for
2304 where left and right are typically adjacent pointers in a tree structure. */
2308 const pass_data pass_data_phiopt
=
2310 GIMPLE_PASS
, /* type */
2311 "phiopt", /* name */
2312 OPTGROUP_NONE
, /* optinfo_flags */
2313 true, /* has_execute */
2314 TV_TREE_PHIOPT
, /* tv_id */
2315 ( PROP_cfg
| PROP_ssa
), /* properties_required */
2316 0, /* properties_provided */
2317 0, /* properties_destroyed */
2318 0, /* todo_flags_start */
2319 0, /* todo_flags_finish */
2322 class pass_phiopt
: public gimple_opt_pass
2325 pass_phiopt (gcc::context
*ctxt
)
2326 : gimple_opt_pass (pass_data_phiopt
, ctxt
)
2329 /* opt_pass methods: */
2330 opt_pass
* clone () { return new pass_phiopt (m_ctxt
); }
2331 virtual unsigned int execute (function
*)
2333 return tree_ssa_phiopt_worker (false, gate_hoist_loads ());
2336 }; // class pass_phiopt
2341 make_pass_phiopt (gcc::context
*ctxt
)
2343 return new pass_phiopt (ctxt
);
2348 const pass_data pass_data_cselim
=
2350 GIMPLE_PASS
, /* type */
2351 "cselim", /* name */
2352 OPTGROUP_NONE
, /* optinfo_flags */
2353 true, /* has_execute */
2354 TV_TREE_PHIOPT
, /* tv_id */
2355 ( PROP_cfg
| PROP_ssa
), /* properties_required */
2356 0, /* properties_provided */
2357 0, /* properties_destroyed */
2358 0, /* todo_flags_start */
2359 0, /* todo_flags_finish */
2362 class pass_cselim
: public gimple_opt_pass
2365 pass_cselim (gcc::context
*ctxt
)
2366 : gimple_opt_pass (pass_data_cselim
, ctxt
)
2369 /* opt_pass methods: */
2370 virtual bool gate (function
*) { return flag_tree_cselim
; }
2371 virtual unsigned int execute (function
*) { return tree_ssa_cs_elim (); }
2373 }; // class pass_cselim
2378 make_pass_cselim (gcc::context
*ctxt
)
2380 return new pass_cselim (ctxt
);