1 /* Optimization of PHI nodes by converting them into straightline code.
2 Copyright (C) 2004-2015 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"
27 #include "fold-const.h"
28 #include "stor-layout.h"
32 #include "hard-reg-set.h"
34 #include "dominance.h"
37 #include "basic-block.h"
38 #include "tree-ssa-alias.h"
39 #include "internal-fn.h"
40 #include "gimple-expr.h"
43 #include "gimple-iterator.h"
44 #include "gimplify-me.h"
45 #include "gimple-ssa.h"
47 #include "tree-phinodes.h"
48 #include "ssa-iterators.h"
49 #include "stringpool.h"
50 #include "tree-ssanames.h"
52 #include "insn-config.h"
62 #include "tree-pass.h"
63 #include "langhooks.h"
66 #include "tree-data-ref.h"
67 #include "gimple-pretty-print.h"
68 #include "insn-codes.h"
70 #include "tree-scalar-evolution.h"
71 #include "tree-inline.h"
73 static unsigned int tree_ssa_phiopt_worker (bool, bool);
74 static bool conditional_replacement (basic_block
, basic_block
,
75 edge
, edge
, gphi
*, tree
, tree
);
76 static int value_replacement (basic_block
, basic_block
,
77 edge
, edge
, gimple
, tree
, tree
);
78 static bool minmax_replacement (basic_block
, basic_block
,
79 edge
, edge
, gimple
, tree
, tree
);
80 static bool abs_replacement (basic_block
, basic_block
,
81 edge
, edge
, gimple
, tree
, tree
);
82 static bool cond_store_replacement (basic_block
, basic_block
, edge
, edge
,
84 static bool cond_if_else_store_replacement (basic_block
, basic_block
, basic_block
);
85 static hash_set
<tree
> * get_non_trapping ();
86 static void replace_phi_edge_with_variable (basic_block
, edge
, gimple
, tree
);
87 static void hoist_adjacent_loads (basic_block
, basic_block
,
88 basic_block
, basic_block
);
89 static bool gate_hoist_loads (void);
91 /* This pass tries to transform conditional stores into unconditional
92 ones, enabling further simplifications with the simpler then and else
93 blocks. In particular it replaces this:
96 if (cond) goto bb2; else goto bb1;
104 if (cond) goto bb1; else goto bb2;
108 condtmp = PHI <RHS, condtmp'>
111 This transformation can only be done under several constraints,
112 documented below. It also replaces:
115 if (cond) goto bb2; else goto bb1;
126 if (cond) goto bb3; else goto bb1;
129 condtmp = PHI <RHS1, RHS2>
133 tree_ssa_cs_elim (void)
136 /* ??? We are not interested in loop related info, but the following
137 will create it, ICEing as we didn't init loops with pre-headers.
138 An interfacing issue of find_data_references_in_bb. */
139 loop_optimizer_init (LOOPS_NORMAL
);
141 todo
= tree_ssa_phiopt_worker (true, false);
143 loop_optimizer_finalize ();
147 /* Return the singleton PHI in the SEQ of PHIs for edges E0 and E1. */
150 single_non_singleton_phi_for_edges (gimple_seq seq
, edge e0
, edge e1
)
152 gimple_stmt_iterator i
;
154 if (gimple_seq_singleton_p (seq
))
155 return as_a
<gphi
*> (gsi_stmt (gsi_start (seq
)));
156 for (i
= gsi_start (seq
); !gsi_end_p (i
); gsi_next (&i
))
158 gphi
*p
= as_a
<gphi
*> (gsi_stmt (i
));
159 /* If the PHI arguments are equal then we can skip this PHI. */
160 if (operand_equal_for_phi_arg_p (gimple_phi_arg_def (p
, e0
->dest_idx
),
161 gimple_phi_arg_def (p
, e1
->dest_idx
)))
164 /* If we already have a PHI that has the two edge arguments are
165 different, then return it is not a singleton for these PHIs. */
174 /* The core routine of conditional store replacement and normal
175 phi optimizations. Both share much of the infrastructure in how
176 to match applicable basic block patterns. DO_STORE_ELIM is true
177 when we want to do conditional store replacement, false otherwise.
178 DO_HOIST_LOADS is true when we want to hoist adjacent loads out
179 of diamond control flow patterns, false otherwise. */
181 tree_ssa_phiopt_worker (bool do_store_elim
, bool do_hoist_loads
)
184 basic_block
*bb_order
;
186 bool cfgchanged
= false;
187 hash_set
<tree
> *nontrap
= 0;
190 /* Calculate the set of non-trapping memory accesses. */
191 nontrap
= get_non_trapping ();
193 /* Search every basic block for COND_EXPR we may be able to optimize.
195 We walk the blocks in order that guarantees that a block with
196 a single predecessor is processed before the predecessor.
197 This ensures that we collapse inner ifs before visiting the
198 outer ones, and also that we do not try to visit a removed
200 bb_order
= single_pred_before_succ_order ();
201 n
= n_basic_blocks_for_fn (cfun
) - NUM_FIXED_BLOCKS
;
203 for (i
= 0; i
< n
; i
++)
207 basic_block bb1
, bb2
;
213 cond_stmt
= last_stmt (bb
);
214 /* Check to see if the last statement is a GIMPLE_COND. */
216 || gimple_code (cond_stmt
) != GIMPLE_COND
)
219 e1
= EDGE_SUCC (bb
, 0);
221 e2
= EDGE_SUCC (bb
, 1);
224 /* We cannot do the optimization on abnormal edges. */
225 if ((e1
->flags
& EDGE_ABNORMAL
) != 0
226 || (e2
->flags
& EDGE_ABNORMAL
) != 0)
229 /* If either bb1's succ or bb2 or bb2's succ is non NULL. */
230 if (EDGE_COUNT (bb1
->succs
) == 0
232 || EDGE_COUNT (bb2
->succs
) == 0)
235 /* Find the bb which is the fall through to the other. */
236 if (EDGE_SUCC (bb1
, 0)->dest
== bb2
)
238 else if (EDGE_SUCC (bb2
, 0)->dest
== bb1
)
240 std::swap (bb1
, bb2
);
243 else if (do_store_elim
244 && EDGE_SUCC (bb1
, 0)->dest
== EDGE_SUCC (bb2
, 0)->dest
)
246 basic_block bb3
= EDGE_SUCC (bb1
, 0)->dest
;
248 if (!single_succ_p (bb1
)
249 || (EDGE_SUCC (bb1
, 0)->flags
& EDGE_FALLTHRU
) == 0
250 || !single_succ_p (bb2
)
251 || (EDGE_SUCC (bb2
, 0)->flags
& EDGE_FALLTHRU
) == 0
252 || EDGE_COUNT (bb3
->preds
) != 2)
254 if (cond_if_else_store_replacement (bb1
, bb2
, bb3
))
258 else if (do_hoist_loads
259 && EDGE_SUCC (bb1
, 0)->dest
== EDGE_SUCC (bb2
, 0)->dest
)
261 basic_block bb3
= EDGE_SUCC (bb1
, 0)->dest
;
263 if (!FLOAT_TYPE_P (TREE_TYPE (gimple_cond_lhs (cond_stmt
)))
264 && single_succ_p (bb1
)
265 && single_succ_p (bb2
)
266 && single_pred_p (bb1
)
267 && single_pred_p (bb2
)
268 && EDGE_COUNT (bb
->succs
) == 2
269 && EDGE_COUNT (bb3
->preds
) == 2
270 /* If one edge or the other is dominant, a conditional move
271 is likely to perform worse than the well-predicted branch. */
272 && !predictable_edge_p (EDGE_SUCC (bb
, 0))
273 && !predictable_edge_p (EDGE_SUCC (bb
, 1)))
274 hoist_adjacent_loads (bb
, bb1
, bb2
, bb3
);
280 e1
= EDGE_SUCC (bb1
, 0);
282 /* Make sure that bb1 is just a fall through. */
283 if (!single_succ_p (bb1
)
284 || (e1
->flags
& EDGE_FALLTHRU
) == 0)
287 /* Also make sure that bb1 only have one predecessor and that it
289 if (!single_pred_p (bb1
)
290 || single_pred (bb1
) != bb
)
295 /* bb1 is the middle block, bb2 the join block, bb the split block,
296 e1 the fallthrough edge from bb1 to bb2. We can't do the
297 optimization if the join block has more than two predecessors. */
298 if (EDGE_COUNT (bb2
->preds
) > 2)
300 if (cond_store_replacement (bb1
, bb2
, e1
, e2
, nontrap
))
305 gimple_seq phis
= phi_nodes (bb2
);
306 gimple_stmt_iterator gsi
;
307 bool candorest
= true;
309 /* Value replacement can work with more than one PHI
310 so try that first. */
311 for (gsi
= gsi_start (phis
); !gsi_end_p (gsi
); gsi_next (&gsi
))
313 phi
= as_a
<gphi
*> (gsi_stmt (gsi
));
314 arg0
= gimple_phi_arg_def (phi
, e1
->dest_idx
);
315 arg1
= gimple_phi_arg_def (phi
, e2
->dest_idx
);
316 if (value_replacement (bb
, bb1
, e1
, e2
, phi
, arg0
, arg1
) == 2)
327 phi
= single_non_singleton_phi_for_edges (phis
, e1
, e2
);
331 arg0
= gimple_phi_arg_def (phi
, e1
->dest_idx
);
332 arg1
= gimple_phi_arg_def (phi
, e2
->dest_idx
);
334 /* Something is wrong if we cannot find the arguments in the PHI
336 gcc_assert (arg0
!= NULL
&& arg1
!= NULL
);
338 /* Do the replacement of conditional if it can be done. */
339 if (conditional_replacement (bb
, bb1
, e1
, e2
, phi
, arg0
, arg1
))
341 else if (abs_replacement (bb
, bb1
, e1
, e2
, phi
, arg0
, arg1
))
343 else if (minmax_replacement (bb
, bb1
, e1
, e2
, phi
, arg0
, arg1
))
352 /* If the CFG has changed, we should cleanup the CFG. */
353 if (cfgchanged
&& do_store_elim
)
355 /* In cond-store replacement we have added some loads on edges
356 and new VOPS (as we moved the store, and created a load). */
357 gsi_commit_edge_inserts ();
358 return TODO_cleanup_cfg
| TODO_update_ssa_only_virtuals
;
361 return TODO_cleanup_cfg
;
365 /* Replace PHI node element whose edge is E in block BB with variable NEW.
366 Remove the edge from COND_BLOCK which does not lead to BB (COND_BLOCK
367 is known to have two edges, one of which must reach BB). */
370 replace_phi_edge_with_variable (basic_block cond_block
,
371 edge e
, gimple phi
, tree new_tree
)
373 basic_block bb
= gimple_bb (phi
);
374 basic_block block_to_remove
;
375 gimple_stmt_iterator gsi
;
377 /* Change the PHI argument to new. */
378 SET_USE (PHI_ARG_DEF_PTR (phi
, e
->dest_idx
), new_tree
);
380 /* Remove the empty basic block. */
381 if (EDGE_SUCC (cond_block
, 0)->dest
== bb
)
383 EDGE_SUCC (cond_block
, 0)->flags
|= EDGE_FALLTHRU
;
384 EDGE_SUCC (cond_block
, 0)->flags
&= ~(EDGE_TRUE_VALUE
| EDGE_FALSE_VALUE
);
385 EDGE_SUCC (cond_block
, 0)->probability
= REG_BR_PROB_BASE
;
386 EDGE_SUCC (cond_block
, 0)->count
+= EDGE_SUCC (cond_block
, 1)->count
;
388 block_to_remove
= EDGE_SUCC (cond_block
, 1)->dest
;
392 EDGE_SUCC (cond_block
, 1)->flags
|= EDGE_FALLTHRU
;
393 EDGE_SUCC (cond_block
, 1)->flags
394 &= ~(EDGE_TRUE_VALUE
| EDGE_FALSE_VALUE
);
395 EDGE_SUCC (cond_block
, 1)->probability
= REG_BR_PROB_BASE
;
396 EDGE_SUCC (cond_block
, 1)->count
+= EDGE_SUCC (cond_block
, 0)->count
;
398 block_to_remove
= EDGE_SUCC (cond_block
, 0)->dest
;
400 delete_basic_block (block_to_remove
);
402 /* Eliminate the COND_EXPR at the end of COND_BLOCK. */
403 gsi
= gsi_last_bb (cond_block
);
404 gsi_remove (&gsi
, true);
406 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
408 "COND_EXPR in block %d and PHI in block %d converted to straightline code.\n",
413 /* The function conditional_replacement does the main work of doing the
414 conditional replacement. Return true if the replacement is done.
415 Otherwise return false.
416 BB is the basic block where the replacement is going to be done on. ARG0
417 is argument 0 from PHI. Likewise for ARG1. */
420 conditional_replacement (basic_block cond_bb
, basic_block middle_bb
,
421 edge e0
, edge e1
, gphi
*phi
,
422 tree arg0
, tree arg1
)
428 gimple_stmt_iterator gsi
;
429 edge true_edge
, false_edge
;
430 tree new_var
, new_var2
;
433 /* FIXME: Gimplification of complex type is too hard for now. */
434 /* We aren't prepared to handle vectors either (and it is a question
435 if it would be worthwhile anyway). */
436 if (!(INTEGRAL_TYPE_P (TREE_TYPE (arg0
))
437 || POINTER_TYPE_P (TREE_TYPE (arg0
)))
438 || !(INTEGRAL_TYPE_P (TREE_TYPE (arg1
))
439 || POINTER_TYPE_P (TREE_TYPE (arg1
))))
442 /* The PHI arguments have the constants 0 and 1, or 0 and -1, then
443 convert it to the conditional. */
444 if ((integer_zerop (arg0
) && integer_onep (arg1
))
445 || (integer_zerop (arg1
) && integer_onep (arg0
)))
447 else if ((integer_zerop (arg0
) && integer_all_onesp (arg1
))
448 || (integer_zerop (arg1
) && integer_all_onesp (arg0
)))
453 if (!empty_block_p (middle_bb
))
456 /* At this point we know we have a GIMPLE_COND with two successors.
457 One successor is BB, the other successor is an empty block which
458 falls through into BB.
460 There is a single PHI node at the join point (BB) and its arguments
461 are constants (0, 1) or (0, -1).
463 So, given the condition COND, and the two PHI arguments, we can
464 rewrite this PHI into non-branching code:
466 dest = (COND) or dest = COND'
468 We use the condition as-is if the argument associated with the
469 true edge has the value one or the argument associated with the
470 false edge as the value zero. Note that those conditions are not
471 the same since only one of the outgoing edges from the GIMPLE_COND
472 will directly reach BB and thus be associated with an argument. */
474 stmt
= last_stmt (cond_bb
);
475 result
= PHI_RESULT (phi
);
477 /* To handle special cases like floating point comparison, it is easier and
478 less error-prone to build a tree and gimplify it on the fly though it is
480 cond
= fold_build2_loc (gimple_location (stmt
),
481 gimple_cond_code (stmt
), boolean_type_node
,
482 gimple_cond_lhs (stmt
), gimple_cond_rhs (stmt
));
484 /* We need to know which is the true edge and which is the false
485 edge so that we know when to invert the condition below. */
486 extract_true_false_edges_from_block (cond_bb
, &true_edge
, &false_edge
);
487 if ((e0
== true_edge
&& integer_zerop (arg0
))
488 || (e0
== false_edge
&& !integer_zerop (arg0
))
489 || (e1
== true_edge
&& integer_zerop (arg1
))
490 || (e1
== false_edge
&& !integer_zerop (arg1
)))
491 cond
= fold_build1_loc (gimple_location (stmt
),
492 TRUTH_NOT_EXPR
, TREE_TYPE (cond
), cond
);
496 cond
= fold_convert_loc (gimple_location (stmt
),
497 TREE_TYPE (result
), cond
);
498 cond
= fold_build1_loc (gimple_location (stmt
),
499 NEGATE_EXPR
, TREE_TYPE (cond
), cond
);
502 /* Insert our new statements at the end of conditional block before the
504 gsi
= gsi_for_stmt (stmt
);
505 new_var
= force_gimple_operand_gsi (&gsi
, cond
, true, NULL
, true,
508 if (!useless_type_conversion_p (TREE_TYPE (result
), TREE_TYPE (new_var
)))
510 source_location locus_0
, locus_1
;
512 new_var2
= make_ssa_name (TREE_TYPE (result
));
513 new_stmt
= gimple_build_assign (new_var2
, CONVERT_EXPR
, new_var
);
514 gsi_insert_before (&gsi
, new_stmt
, GSI_SAME_STMT
);
517 /* Set the locus to the first argument, unless is doesn't have one. */
518 locus_0
= gimple_phi_arg_location (phi
, 0);
519 locus_1
= gimple_phi_arg_location (phi
, 1);
520 if (locus_0
== UNKNOWN_LOCATION
)
522 gimple_set_location (new_stmt
, locus_0
);
525 replace_phi_edge_with_variable (cond_bb
, e1
, phi
, new_var
);
527 /* Note that we optimized this PHI. */
531 /* Update *ARG which is defined in STMT so that it contains the
532 computed value if that seems profitable. Return true if the
533 statement is made dead by that rewriting. */
536 jump_function_from_stmt (tree
*arg
, gimple stmt
)
538 enum tree_code code
= gimple_assign_rhs_code (stmt
);
539 if (code
== ADDR_EXPR
)
541 /* For arg = &p->i transform it to p, if possible. */
542 tree rhs1
= gimple_assign_rhs1 (stmt
);
543 HOST_WIDE_INT offset
;
544 tree tem
= get_addr_base_and_unit_offset (TREE_OPERAND (rhs1
, 0),
547 && TREE_CODE (tem
) == MEM_REF
548 && (mem_ref_offset (tem
) + offset
) == 0)
550 *arg
= TREE_OPERAND (tem
, 0);
554 /* TODO: Much like IPA-CP jump-functions we want to handle constant
555 additions symbolically here, and we'd need to update the comparison
556 code that compares the arg + cst tuples in our caller. For now the
557 code above exactly handles the VEC_BASE pattern from vec.h. */
561 /* RHS is a source argument in a BIT_AND_EXPR which feeds a conditional
562 of the form SSA_NAME NE 0.
564 If RHS is fed by a simple EQ_EXPR comparison of two values, see if
565 the two input values of the EQ_EXPR match arg0 and arg1.
567 If so update *code and return TRUE. Otherwise return FALSE. */
570 rhs_is_fed_for_value_replacement (const_tree arg0
, const_tree arg1
,
571 enum tree_code
*code
, const_tree rhs
)
573 /* Obviously if RHS is not an SSA_NAME, we can't look at the defining
575 if (TREE_CODE (rhs
) == SSA_NAME
)
577 gimple def1
= SSA_NAME_DEF_STMT (rhs
);
579 /* Verify the defining statement has an EQ_EXPR on the RHS. */
580 if (is_gimple_assign (def1
) && gimple_assign_rhs_code (def1
) == EQ_EXPR
)
582 /* Finally verify the source operands of the EQ_EXPR are equal
584 tree op0
= gimple_assign_rhs1 (def1
);
585 tree op1
= gimple_assign_rhs2 (def1
);
586 if ((operand_equal_for_phi_arg_p (arg0
, op0
)
587 && operand_equal_for_phi_arg_p (arg1
, op1
))
588 || (operand_equal_for_phi_arg_p (arg0
, op1
)
589 && operand_equal_for_phi_arg_p (arg1
, op0
)))
591 /* We will perform the optimization. */
592 *code
= gimple_assign_rhs_code (def1
);
600 /* Return TRUE if arg0/arg1 are equal to the rhs/lhs or lhs/rhs of COND.
602 Also return TRUE if arg0/arg1 are equal to the source arguments of a
603 an EQ comparison feeding a BIT_AND_EXPR which feeds COND.
605 Return FALSE otherwise. */
608 operand_equal_for_value_replacement (const_tree arg0
, const_tree arg1
,
609 enum tree_code
*code
, gimple cond
)
612 tree lhs
= gimple_cond_lhs (cond
);
613 tree rhs
= gimple_cond_rhs (cond
);
615 if ((operand_equal_for_phi_arg_p (arg0
, lhs
)
616 && operand_equal_for_phi_arg_p (arg1
, rhs
))
617 || (operand_equal_for_phi_arg_p (arg1
, lhs
)
618 && operand_equal_for_phi_arg_p (arg0
, rhs
)))
621 /* Now handle more complex case where we have an EQ comparison
622 which feeds a BIT_AND_EXPR which feeds COND.
624 First verify that COND is of the form SSA_NAME NE 0. */
625 if (*code
!= NE_EXPR
|| !integer_zerop (rhs
)
626 || TREE_CODE (lhs
) != SSA_NAME
)
629 /* Now ensure that SSA_NAME is set by a BIT_AND_EXPR. */
630 def
= SSA_NAME_DEF_STMT (lhs
);
631 if (!is_gimple_assign (def
) || gimple_assign_rhs_code (def
) != BIT_AND_EXPR
)
634 /* Now verify arg0/arg1 correspond to the source arguments of an
635 EQ comparison feeding the BIT_AND_EXPR. */
637 tree tmp
= gimple_assign_rhs1 (def
);
638 if (rhs_is_fed_for_value_replacement (arg0
, arg1
, code
, tmp
))
641 tmp
= gimple_assign_rhs2 (def
);
642 if (rhs_is_fed_for_value_replacement (arg0
, arg1
, code
, tmp
))
648 /* Returns true if ARG is a neutral element for operation CODE
649 on the RIGHT side. */
652 neutral_element_p (tree_code code
, tree arg
, bool right
)
659 return integer_zerop (arg
);
666 case POINTER_PLUS_EXPR
:
667 return right
&& integer_zerop (arg
);
670 return integer_onep (arg
);
677 return right
&& integer_onep (arg
);
680 return integer_all_onesp (arg
);
687 /* Returns true if ARG is an absorbing element for operation CODE. */
690 absorbing_element_p (tree_code code
, tree arg
)
695 return integer_all_onesp (arg
);
699 return integer_zerop (arg
);
706 /* The function value_replacement does the main work of doing the value
707 replacement. Return non-zero if the replacement is done. Otherwise return
708 0. If we remove the middle basic block, return 2.
709 BB is the basic block where the replacement is going to be done on. ARG0
710 is argument 0 from the PHI. Likewise for ARG1. */
713 value_replacement (basic_block cond_bb
, basic_block middle_bb
,
714 edge e0
, edge e1
, gimple phi
,
715 tree arg0
, tree arg1
)
717 gimple_stmt_iterator gsi
;
719 edge true_edge
, false_edge
;
721 bool emtpy_or_with_defined_p
= true;
723 /* If the type says honor signed zeros we cannot do this
725 if (HONOR_SIGNED_ZEROS (arg1
))
728 /* If there is a statement in MIDDLE_BB that defines one of the PHI
729 arguments, then adjust arg0 or arg1. */
730 gsi
= gsi_start_nondebug_after_labels_bb (middle_bb
);
731 while (!gsi_end_p (gsi
))
733 gimple stmt
= gsi_stmt (gsi
);
735 gsi_next_nondebug (&gsi
);
736 if (!is_gimple_assign (stmt
))
738 emtpy_or_with_defined_p
= false;
741 /* Now try to adjust arg0 or arg1 according to the computation
743 lhs
= gimple_assign_lhs (stmt
);
745 && jump_function_from_stmt (&arg0
, stmt
))
747 && jump_function_from_stmt (&arg1
, stmt
)))
748 emtpy_or_with_defined_p
= false;
751 cond
= last_stmt (cond_bb
);
752 code
= gimple_cond_code (cond
);
754 /* This transformation is only valid for equality comparisons. */
755 if (code
!= NE_EXPR
&& code
!= EQ_EXPR
)
758 /* We need to know which is the true edge and which is the false
759 edge so that we know if have abs or negative abs. */
760 extract_true_false_edges_from_block (cond_bb
, &true_edge
, &false_edge
);
762 /* At this point we know we have a COND_EXPR with two successors.
763 One successor is BB, the other successor is an empty block which
764 falls through into BB.
766 The condition for the COND_EXPR is known to be NE_EXPR or EQ_EXPR.
768 There is a single PHI node at the join point (BB) with two arguments.
770 We now need to verify that the two arguments in the PHI node match
771 the two arguments to the equality comparison. */
773 if (operand_equal_for_value_replacement (arg0
, arg1
, &code
, cond
))
778 /* For NE_EXPR, we want to build an assignment result = arg where
779 arg is the PHI argument associated with the true edge. For
780 EQ_EXPR we want the PHI argument associated with the false edge. */
781 e
= (code
== NE_EXPR
? true_edge
: false_edge
);
783 /* Unfortunately, E may not reach BB (it may instead have gone to
784 OTHER_BLOCK). If that is the case, then we want the single outgoing
785 edge from OTHER_BLOCK which reaches BB and represents the desired
786 path from COND_BLOCK. */
787 if (e
->dest
== middle_bb
)
788 e
= single_succ_edge (e
->dest
);
790 /* Now we know the incoming edge to BB that has the argument for the
791 RHS of our new assignment statement. */
797 /* If the middle basic block was empty or is defining the
798 PHI arguments and this is a single phi where the args are different
799 for the edges e0 and e1 then we can remove the middle basic block. */
800 if (emtpy_or_with_defined_p
801 && single_non_singleton_phi_for_edges (phi_nodes (gimple_bb (phi
)),
804 replace_phi_edge_with_variable (cond_bb
, e1
, phi
, arg
);
805 /* Note that we optimized this PHI. */
810 /* Replace the PHI arguments with arg. */
811 SET_PHI_ARG_DEF (phi
, e0
->dest_idx
, arg
);
812 SET_PHI_ARG_DEF (phi
, e1
->dest_idx
, arg
);
813 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
815 fprintf (dump_file
, "PHI ");
816 print_generic_expr (dump_file
, gimple_phi_result (phi
), 0);
817 fprintf (dump_file
, " reduced for COND_EXPR in block %d to ",
819 print_generic_expr (dump_file
, arg
, 0);
820 fprintf (dump_file
, ".\n");
827 /* Now optimize (x != 0) ? x + y : y to just y.
828 The following condition is too restrictive, there can easily be another
829 stmt in middle_bb, for instance a CONVERT_EXPR for the second argument. */
830 gimple assign
= last_and_only_stmt (middle_bb
);
831 if (!assign
|| gimple_code (assign
) != GIMPLE_ASSIGN
832 || gimple_assign_rhs_class (assign
) != GIMPLE_BINARY_RHS
833 || (!INTEGRAL_TYPE_P (TREE_TYPE (arg0
))
834 && !POINTER_TYPE_P (TREE_TYPE (arg0
))))
837 /* Punt if there are (degenerate) PHIs in middle_bb, there should not be. */
838 if (!gimple_seq_empty_p (phi_nodes (middle_bb
)))
841 /* Only transform if it removes the condition. */
842 if (!single_non_singleton_phi_for_edges (phi_nodes (gimple_bb (phi
)), e0
, e1
))
845 /* Size-wise, this is always profitable. */
846 if (optimize_bb_for_speed_p (cond_bb
)
847 /* The special case is useless if it has a low probability. */
848 && profile_status_for_fn (cfun
) != PROFILE_ABSENT
849 && EDGE_PRED (middle_bb
, 0)->probability
< PROB_EVEN
850 /* If assign is cheap, there is no point avoiding it. */
851 && estimate_num_insns (assign
, &eni_time_weights
)
852 >= 3 * estimate_num_insns (cond
, &eni_time_weights
))
855 tree lhs
= gimple_assign_lhs (assign
);
856 tree rhs1
= gimple_assign_rhs1 (assign
);
857 tree rhs2
= gimple_assign_rhs2 (assign
);
858 enum tree_code code_def
= gimple_assign_rhs_code (assign
);
859 tree cond_lhs
= gimple_cond_lhs (cond
);
860 tree cond_rhs
= gimple_cond_rhs (cond
);
862 if (((code
== NE_EXPR
&& e1
== false_edge
)
863 || (code
== EQ_EXPR
&& e1
== true_edge
))
866 && operand_equal_for_phi_arg_p (rhs2
, cond_lhs
)
867 && neutral_element_p (code_def
, cond_rhs
, true))
869 && operand_equal_for_phi_arg_p (rhs1
, cond_lhs
)
870 && neutral_element_p (code_def
, cond_rhs
, false))
871 || (operand_equal_for_phi_arg_p (arg1
, cond_rhs
)
872 && (operand_equal_for_phi_arg_p (rhs2
, cond_lhs
)
873 || operand_equal_for_phi_arg_p (rhs1
, cond_lhs
))
874 && absorbing_element_p (code_def
, cond_rhs
))))
876 gsi
= gsi_for_stmt (cond
);
877 if (INTEGRAL_TYPE_P (TREE_TYPE (lhs
)))
879 /* Moving ASSIGN might change VR of lhs, e.g. when moving u_6
887 # RANGE [0, 4294967294]
888 u_6 = n_5 + 4294967295;
891 # u_3 = PHI <u_6(3), 4294967295(2)> */
892 SSA_NAME_RANGE_INFO (lhs
) = NULL
;
893 SSA_NAME_ANTI_RANGE_P (lhs
) = 0;
894 /* If available, we can use VR of phi result at least. */
895 tree phires
= gimple_phi_result (phi
);
896 struct range_info_def
*phires_range_info
897 = SSA_NAME_RANGE_INFO (phires
);
898 if (phires_range_info
)
899 duplicate_ssa_name_range_info (lhs
, SSA_NAME_RANGE_TYPE (phires
),
902 gimple_stmt_iterator gsi_from
= gsi_for_stmt (assign
);
903 gsi_move_before (&gsi_from
, &gsi
);
904 replace_phi_edge_with_variable (cond_bb
, e1
, phi
, lhs
);
911 /* The function minmax_replacement does the main work of doing the minmax
912 replacement. Return true if the replacement is done. Otherwise return
914 BB is the basic block where the replacement is going to be done on. ARG0
915 is argument 0 from the PHI. Likewise for ARG1. */
918 minmax_replacement (basic_block cond_bb
, basic_block middle_bb
,
919 edge e0
, edge e1
, gimple phi
,
920 tree arg0
, tree arg1
)
925 edge true_edge
, false_edge
;
926 enum tree_code cmp
, minmax
, ass_code
;
927 tree smaller
, larger
, arg_true
, arg_false
;
928 gimple_stmt_iterator gsi
, gsi_from
;
930 type
= TREE_TYPE (PHI_RESULT (phi
));
932 /* The optimization may be unsafe due to NaNs. */
933 if (HONOR_NANS (type
))
936 cond
= as_a
<gcond
*> (last_stmt (cond_bb
));
937 cmp
= gimple_cond_code (cond
);
939 /* This transformation is only valid for order comparisons. Record which
940 operand is smaller/larger if the result of the comparison is true. */
941 if (cmp
== LT_EXPR
|| cmp
== LE_EXPR
)
943 smaller
= gimple_cond_lhs (cond
);
944 larger
= gimple_cond_rhs (cond
);
946 else if (cmp
== GT_EXPR
|| cmp
== GE_EXPR
)
948 smaller
= gimple_cond_rhs (cond
);
949 larger
= gimple_cond_lhs (cond
);
954 /* We need to know which is the true edge and which is the false
955 edge so that we know if have abs or negative abs. */
956 extract_true_false_edges_from_block (cond_bb
, &true_edge
, &false_edge
);
958 /* Forward the edges over the middle basic block. */
959 if (true_edge
->dest
== middle_bb
)
960 true_edge
= EDGE_SUCC (true_edge
->dest
, 0);
961 if (false_edge
->dest
== middle_bb
)
962 false_edge
= EDGE_SUCC (false_edge
->dest
, 0);
966 gcc_assert (false_edge
== e1
);
972 gcc_assert (false_edge
== e0
);
973 gcc_assert (true_edge
== e1
);
978 if (empty_block_p (middle_bb
))
980 if (operand_equal_for_phi_arg_p (arg_true
, smaller
)
981 && operand_equal_for_phi_arg_p (arg_false
, larger
))
985 if (smaller < larger)
991 else if (operand_equal_for_phi_arg_p (arg_false
, smaller
)
992 && operand_equal_for_phi_arg_p (arg_true
, larger
))
999 /* Recognize the following case, assuming d <= u:
1005 This is equivalent to
1010 gimple assign
= last_and_only_stmt (middle_bb
);
1011 tree lhs
, op0
, op1
, bound
;
1014 || gimple_code (assign
) != GIMPLE_ASSIGN
)
1017 lhs
= gimple_assign_lhs (assign
);
1018 ass_code
= gimple_assign_rhs_code (assign
);
1019 if (ass_code
!= MAX_EXPR
&& ass_code
!= MIN_EXPR
)
1021 op0
= gimple_assign_rhs1 (assign
);
1022 op1
= gimple_assign_rhs2 (assign
);
1024 if (true_edge
->src
== middle_bb
)
1026 /* We got here if the condition is true, i.e., SMALLER < LARGER. */
1027 if (!operand_equal_for_phi_arg_p (lhs
, arg_true
))
1030 if (operand_equal_for_phi_arg_p (arg_false
, larger
))
1034 if (smaller < larger)
1036 r' = MAX_EXPR (smaller, bound)
1038 r = PHI <r', larger> --> to be turned to MIN_EXPR. */
1039 if (ass_code
!= MAX_EXPR
)
1043 if (operand_equal_for_phi_arg_p (op0
, smaller
))
1045 else if (operand_equal_for_phi_arg_p (op1
, smaller
))
1050 /* We need BOUND <= LARGER. */
1051 if (!integer_nonzerop (fold_build2 (LE_EXPR
, boolean_type_node
,
1055 else if (operand_equal_for_phi_arg_p (arg_false
, smaller
))
1059 if (smaller < larger)
1061 r' = MIN_EXPR (larger, bound)
1063 r = PHI <r', smaller> --> to be turned to MAX_EXPR. */
1064 if (ass_code
!= MIN_EXPR
)
1068 if (operand_equal_for_phi_arg_p (op0
, larger
))
1070 else if (operand_equal_for_phi_arg_p (op1
, larger
))
1075 /* We need BOUND >= SMALLER. */
1076 if (!integer_nonzerop (fold_build2 (GE_EXPR
, boolean_type_node
,
1085 /* We got here if the condition is false, i.e., SMALLER > LARGER. */
1086 if (!operand_equal_for_phi_arg_p (lhs
, arg_false
))
1089 if (operand_equal_for_phi_arg_p (arg_true
, larger
))
1093 if (smaller > larger)
1095 r' = MIN_EXPR (smaller, bound)
1097 r = PHI <r', larger> --> to be turned to MAX_EXPR. */
1098 if (ass_code
!= MIN_EXPR
)
1102 if (operand_equal_for_phi_arg_p (op0
, smaller
))
1104 else if (operand_equal_for_phi_arg_p (op1
, smaller
))
1109 /* We need BOUND >= LARGER. */
1110 if (!integer_nonzerop (fold_build2 (GE_EXPR
, boolean_type_node
,
1114 else if (operand_equal_for_phi_arg_p (arg_true
, smaller
))
1118 if (smaller > larger)
1120 r' = MAX_EXPR (larger, bound)
1122 r = PHI <r', smaller> --> to be turned to MIN_EXPR. */
1123 if (ass_code
!= MAX_EXPR
)
1127 if (operand_equal_for_phi_arg_p (op0
, larger
))
1129 else if (operand_equal_for_phi_arg_p (op1
, larger
))
1134 /* We need BOUND <= SMALLER. */
1135 if (!integer_nonzerop (fold_build2 (LE_EXPR
, boolean_type_node
,
1143 /* Move the statement from the middle block. */
1144 gsi
= gsi_last_bb (cond_bb
);
1145 gsi_from
= gsi_last_nondebug_bb (middle_bb
);
1146 gsi_move_before (&gsi_from
, &gsi
);
1149 /* Emit the statement to compute min/max. */
1150 result
= duplicate_ssa_name (PHI_RESULT (phi
), NULL
);
1151 new_stmt
= gimple_build_assign (result
, minmax
, arg0
, arg1
);
1152 gsi
= gsi_last_bb (cond_bb
);
1153 gsi_insert_before (&gsi
, new_stmt
, GSI_NEW_STMT
);
1155 replace_phi_edge_with_variable (cond_bb
, e1
, phi
, result
);
1159 /* The function absolute_replacement does the main work of doing the absolute
1160 replacement. Return true if the replacement is done. Otherwise return
1162 bb is the basic block where the replacement is going to be done on. arg0
1163 is argument 0 from the phi. Likewise for arg1. */
1166 abs_replacement (basic_block cond_bb
, basic_block middle_bb
,
1167 edge e0 ATTRIBUTE_UNUSED
, edge e1
,
1168 gimple phi
, tree arg0
, tree arg1
)
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 (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
));
1257 /* Build the modify expression with abs expression. */
1258 new_stmt
= gimple_build_assign (lhs
, ABS_EXPR
, rhs
);
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 (result
, NEGATE_EXPR
, lhs
);
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 name_to_bb
*);
1317 static inline bool equal (const name_to_bb
*, const name_to_bb
*);
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 name_to_bb
*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 name_to_bb
*n1
, const name_to_bb
*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 class nontrapping_dom_walker
: public dom_walker
1347 nontrapping_dom_walker (cdi_direction direction
, hash_set
<tree
> *ps
)
1348 : dom_walker (direction
), m_nontrapping (ps
), m_seen_ssa_names (128) {}
1350 virtual void before_dom_children (basic_block
);
1351 virtual void after_dom_children (basic_block
);
1355 /* We see the expression EXP in basic block BB. If it's an interesting
1356 expression (an MEM_REF through an SSA_NAME) possibly insert the
1357 expression into the set NONTRAP or the hash table of seen expressions.
1358 STORE is true if this expression is on the LHS, otherwise it's on
1360 void add_or_mark_expr (basic_block
, tree
, bool);
1362 hash_set
<tree
> *m_nontrapping
;
1364 /* The hash table for remembering what we've seen. */
1365 hash_table
<ssa_names_hasher
> m_seen_ssa_names
;
1368 /* Called by walk_dominator_tree, when entering the block BB. */
1370 nontrapping_dom_walker::before_dom_children (basic_block bb
)
1374 gimple_stmt_iterator gsi
;
1376 /* If we haven't seen all our predecessors, clear the hash-table. */
1377 FOR_EACH_EDGE (e
, ei
, bb
->preds
)
1378 if ((((size_t)e
->src
->aux
) & 2) == 0)
1384 /* Mark this BB as being on the path to dominator root and as visited. */
1385 bb
->aux
= (void*)(1 | 2);
1387 /* And walk the statements in order. */
1388 for (gsi
= gsi_start_bb (bb
); !gsi_end_p (gsi
); gsi_next (&gsi
))
1390 gimple stmt
= gsi_stmt (gsi
);
1392 if (is_gimple_call (stmt
) && !nonfreeing_call_p (stmt
))
1394 else if (gimple_assign_single_p (stmt
) && !gimple_has_volatile_ops (stmt
))
1396 add_or_mark_expr (bb
, gimple_assign_lhs (stmt
), true);
1397 add_or_mark_expr (bb
, gimple_assign_rhs1 (stmt
), false);
1402 /* Called by walk_dominator_tree, when basic block BB is exited. */
1404 nontrapping_dom_walker::after_dom_children (basic_block bb
)
1406 /* This BB isn't on the path to dominator root anymore. */
1410 /* We see the expression EXP in basic block BB. If it's an interesting
1411 expression (an MEM_REF through an SSA_NAME) possibly insert the
1412 expression into the set NONTRAP or the hash table of seen expressions.
1413 STORE is true if this expression is on the LHS, otherwise it's on
1416 nontrapping_dom_walker::add_or_mark_expr (basic_block bb
, tree exp
, bool store
)
1420 if (TREE_CODE (exp
) == MEM_REF
1421 && TREE_CODE (TREE_OPERAND (exp
, 0)) == SSA_NAME
1422 && tree_fits_shwi_p (TREE_OPERAND (exp
, 1))
1423 && (size
= int_size_in_bytes (TREE_TYPE (exp
))) > 0)
1425 tree name
= TREE_OPERAND (exp
, 0);
1426 struct name_to_bb map
;
1428 struct name_to_bb
*n2bb
;
1429 basic_block found_bb
= 0;
1431 /* Try to find the last seen MEM_REF through the same
1432 SSA_NAME, which can trap. */
1433 map
.ssa_name_ver
= SSA_NAME_VERSION (name
);
1437 map
.offset
= tree_to_shwi (TREE_OPERAND (exp
, 1));
1440 slot
= m_seen_ssa_names
.find_slot (&map
, INSERT
);
1442 if (n2bb
&& n2bb
->phase
>= nt_call_phase
)
1443 found_bb
= n2bb
->bb
;
1445 /* If we've found a trapping MEM_REF, _and_ it dominates EXP
1446 (it's in a basic block on the path from us to the dominator root)
1447 then we can't trap. */
1448 if (found_bb
&& (((size_t)found_bb
->aux
) & 1) == 1)
1450 m_nontrapping
->add (exp
);
1454 /* EXP might trap, so insert it into the hash table. */
1457 n2bb
->phase
= nt_call_phase
;
1462 n2bb
= XNEW (struct name_to_bb
);
1463 n2bb
->ssa_name_ver
= SSA_NAME_VERSION (name
);
1464 n2bb
->phase
= nt_call_phase
;
1466 n2bb
->store
= store
;
1467 n2bb
->offset
= map
.offset
;
1475 /* This is the entry point of gathering non trapping memory accesses.
1476 It will do a dominator walk over the whole function, and it will
1477 make use of the bb->aux pointers. It returns a set of trees
1478 (the MEM_REFs itself) which can't trap. */
1479 static hash_set
<tree
> *
1480 get_non_trapping (void)
1483 hash_set
<tree
> *nontrap
= new hash_set
<tree
>;
1484 /* We're going to do a dominator walk, so ensure that we have
1485 dominance information. */
1486 calculate_dominance_info (CDI_DOMINATORS
);
1488 nontrapping_dom_walker (CDI_DOMINATORS
, nontrap
)
1489 .walk (cfun
->cfg
->x_entry_block_ptr
);
1491 clear_aux_for_blocks ();
1495 /* Do the main work of conditional store replacement. We already know
1496 that the recognized pattern looks like so:
1499 if (cond) goto MIDDLE_BB; else goto JOIN_BB (edge E1)
1502 fallthrough (edge E0)
1506 We check that MIDDLE_BB contains only one store, that that store
1507 doesn't trap (not via NOTRAP, but via checking if an access to the same
1508 memory location dominates us) and that the store has a "simple" RHS. */
1511 cond_store_replacement (basic_block middle_bb
, basic_block join_bb
,
1512 edge e0
, edge e1
, hash_set
<tree
> *nontrap
)
1514 gimple assign
= last_and_only_stmt (middle_bb
);
1515 tree lhs
, rhs
, name
, name2
;
1518 gimple_stmt_iterator gsi
;
1519 source_location locus
;
1521 /* Check if middle_bb contains of only one store. */
1523 || !gimple_assign_single_p (assign
)
1524 || gimple_has_volatile_ops (assign
))
1527 locus
= gimple_location (assign
);
1528 lhs
= gimple_assign_lhs (assign
);
1529 rhs
= gimple_assign_rhs1 (assign
);
1530 if (TREE_CODE (lhs
) != MEM_REF
1531 || TREE_CODE (TREE_OPERAND (lhs
, 0)) != SSA_NAME
1532 || !is_gimple_reg_type (TREE_TYPE (lhs
)))
1535 /* Prove that we can move the store down. We could also check
1536 TREE_THIS_NOTRAP here, but in that case we also could move stores,
1537 whose value is not available readily, which we want to avoid. */
1538 if (!nontrap
->contains (lhs
))
1541 /* Now we've checked the constraints, so do the transformation:
1542 1) Remove the single store. */
1543 gsi
= gsi_for_stmt (assign
);
1544 unlink_stmt_vdef (assign
);
1545 gsi_remove (&gsi
, true);
1546 release_defs (assign
);
1548 /* 2) Insert a load from the memory of the store to the temporary
1549 on the edge which did not contain the store. */
1550 lhs
= unshare_expr (lhs
);
1551 name
= make_temp_ssa_name (TREE_TYPE (lhs
), NULL
, "cstore");
1552 new_stmt
= gimple_build_assign (name
, lhs
);
1553 gimple_set_location (new_stmt
, locus
);
1554 gsi_insert_on_edge (e1
, new_stmt
);
1556 /* 3) Create a PHI node at the join block, with one argument
1557 holding the old RHS, and the other holding the temporary
1558 where we stored the old memory contents. */
1559 name2
= make_temp_ssa_name (TREE_TYPE (lhs
), NULL
, "cstore");
1560 newphi
= create_phi_node (name2
, join_bb
);
1561 add_phi_arg (newphi
, rhs
, e0
, locus
);
1562 add_phi_arg (newphi
, name
, e1
, locus
);
1564 lhs
= unshare_expr (lhs
);
1565 new_stmt
= gimple_build_assign (lhs
, PHI_RESULT (newphi
));
1567 /* 4) Insert that PHI node. */
1568 gsi
= gsi_after_labels (join_bb
);
1569 if (gsi_end_p (gsi
))
1571 gsi
= gsi_last_bb (join_bb
);
1572 gsi_insert_after (&gsi
, new_stmt
, GSI_NEW_STMT
);
1575 gsi_insert_before (&gsi
, new_stmt
, GSI_NEW_STMT
);
1580 /* Do the main work of conditional store replacement. */
1583 cond_if_else_store_replacement_1 (basic_block then_bb
, basic_block else_bb
,
1584 basic_block join_bb
, gimple then_assign
,
1587 tree lhs_base
, lhs
, then_rhs
, else_rhs
, name
;
1588 source_location then_locus
, else_locus
;
1589 gimple_stmt_iterator gsi
;
1593 if (then_assign
== NULL
1594 || !gimple_assign_single_p (then_assign
)
1595 || gimple_clobber_p (then_assign
)
1596 || gimple_has_volatile_ops (then_assign
)
1597 || else_assign
== NULL
1598 || !gimple_assign_single_p (else_assign
)
1599 || gimple_clobber_p (else_assign
)
1600 || gimple_has_volatile_ops (else_assign
))
1603 lhs
= gimple_assign_lhs (then_assign
);
1604 if (!is_gimple_reg_type (TREE_TYPE (lhs
))
1605 || !operand_equal_p (lhs
, gimple_assign_lhs (else_assign
), 0))
1608 lhs_base
= get_base_address (lhs
);
1609 if (lhs_base
== NULL_TREE
1610 || (!DECL_P (lhs_base
) && TREE_CODE (lhs_base
) != MEM_REF
))
1613 then_rhs
= gimple_assign_rhs1 (then_assign
);
1614 else_rhs
= gimple_assign_rhs1 (else_assign
);
1615 then_locus
= gimple_location (then_assign
);
1616 else_locus
= gimple_location (else_assign
);
1618 /* Now we've checked the constraints, so do the transformation:
1619 1) Remove the stores. */
1620 gsi
= gsi_for_stmt (then_assign
);
1621 unlink_stmt_vdef (then_assign
);
1622 gsi_remove (&gsi
, true);
1623 release_defs (then_assign
);
1625 gsi
= gsi_for_stmt (else_assign
);
1626 unlink_stmt_vdef (else_assign
);
1627 gsi_remove (&gsi
, true);
1628 release_defs (else_assign
);
1630 /* 2) Create a PHI node at the join block, with one argument
1631 holding the old RHS, and the other holding the temporary
1632 where we stored the old memory contents. */
1633 name
= make_temp_ssa_name (TREE_TYPE (lhs
), NULL
, "cstore");
1634 newphi
= create_phi_node (name
, join_bb
);
1635 add_phi_arg (newphi
, then_rhs
, EDGE_SUCC (then_bb
, 0), then_locus
);
1636 add_phi_arg (newphi
, else_rhs
, EDGE_SUCC (else_bb
, 0), else_locus
);
1638 new_stmt
= gimple_build_assign (lhs
, PHI_RESULT (newphi
));
1640 /* 3) Insert that PHI node. */
1641 gsi
= gsi_after_labels (join_bb
);
1642 if (gsi_end_p (gsi
))
1644 gsi
= gsi_last_bb (join_bb
);
1645 gsi_insert_after (&gsi
, new_stmt
, GSI_NEW_STMT
);
1648 gsi_insert_before (&gsi
, new_stmt
, GSI_NEW_STMT
);
1653 /* Conditional store replacement. We already know
1654 that the recognized pattern looks like so:
1657 if (cond) goto THEN_BB; else goto ELSE_BB (edge E1)
1667 fallthrough (edge E0)
1671 We check that it is safe to sink the store to JOIN_BB by verifying that
1672 there are no read-after-write or write-after-write dependencies in
1673 THEN_BB and ELSE_BB. */
1676 cond_if_else_store_replacement (basic_block then_bb
, basic_block else_bb
,
1677 basic_block join_bb
)
1679 gimple then_assign
= last_and_only_stmt (then_bb
);
1680 gimple else_assign
= last_and_only_stmt (else_bb
);
1681 vec
<data_reference_p
> then_datarefs
, else_datarefs
;
1682 vec
<ddr_p
> then_ddrs
, else_ddrs
;
1683 gimple then_store
, else_store
;
1684 bool found
, ok
= false, res
;
1685 struct data_dependence_relation
*ddr
;
1686 data_reference_p then_dr
, else_dr
;
1688 tree then_lhs
, else_lhs
;
1689 basic_block blocks
[3];
1691 if (MAX_STORES_TO_SINK
== 0)
1694 /* Handle the case with single statement in THEN_BB and ELSE_BB. */
1695 if (then_assign
&& else_assign
)
1696 return cond_if_else_store_replacement_1 (then_bb
, else_bb
, join_bb
,
1697 then_assign
, else_assign
);
1699 /* Find data references. */
1700 then_datarefs
.create (1);
1701 else_datarefs
.create (1);
1702 if ((find_data_references_in_bb (NULL
, then_bb
, &then_datarefs
)
1704 || !then_datarefs
.length ()
1705 || (find_data_references_in_bb (NULL
, else_bb
, &else_datarefs
)
1707 || !else_datarefs
.length ())
1709 free_data_refs (then_datarefs
);
1710 free_data_refs (else_datarefs
);
1714 /* Find pairs of stores with equal LHS. */
1715 auto_vec
<gimple
, 1> then_stores
, else_stores
;
1716 FOR_EACH_VEC_ELT (then_datarefs
, i
, then_dr
)
1718 if (DR_IS_READ (then_dr
))
1721 then_store
= DR_STMT (then_dr
);
1722 then_lhs
= gimple_get_lhs (then_store
);
1723 if (then_lhs
== NULL_TREE
)
1727 FOR_EACH_VEC_ELT (else_datarefs
, j
, else_dr
)
1729 if (DR_IS_READ (else_dr
))
1732 else_store
= DR_STMT (else_dr
);
1733 else_lhs
= gimple_get_lhs (else_store
);
1734 if (else_lhs
== NULL_TREE
)
1737 if (operand_equal_p (then_lhs
, else_lhs
, 0))
1747 then_stores
.safe_push (then_store
);
1748 else_stores
.safe_push (else_store
);
1751 /* No pairs of stores found. */
1752 if (!then_stores
.length ()
1753 || then_stores
.length () > (unsigned) MAX_STORES_TO_SINK
)
1755 free_data_refs (then_datarefs
);
1756 free_data_refs (else_datarefs
);
1760 /* Compute and check data dependencies in both basic blocks. */
1761 then_ddrs
.create (1);
1762 else_ddrs
.create (1);
1763 if (!compute_all_dependences (then_datarefs
, &then_ddrs
,
1765 || !compute_all_dependences (else_datarefs
, &else_ddrs
,
1768 free_dependence_relations (then_ddrs
);
1769 free_dependence_relations (else_ddrs
);
1770 free_data_refs (then_datarefs
);
1771 free_data_refs (else_datarefs
);
1774 blocks
[0] = then_bb
;
1775 blocks
[1] = else_bb
;
1776 blocks
[2] = join_bb
;
1777 renumber_gimple_stmt_uids_in_blocks (blocks
, 3);
1779 /* Check that there are no read-after-write or write-after-write dependencies
1781 FOR_EACH_VEC_ELT (then_ddrs
, i
, ddr
)
1783 struct data_reference
*dra
= DDR_A (ddr
);
1784 struct data_reference
*drb
= DDR_B (ddr
);
1786 if (DDR_ARE_DEPENDENT (ddr
) != chrec_known
1787 && ((DR_IS_READ (dra
) && DR_IS_WRITE (drb
)
1788 && gimple_uid (DR_STMT (dra
)) > gimple_uid (DR_STMT (drb
)))
1789 || (DR_IS_READ (drb
) && DR_IS_WRITE (dra
)
1790 && gimple_uid (DR_STMT (drb
)) > gimple_uid (DR_STMT (dra
)))
1791 || (DR_IS_WRITE (dra
) && DR_IS_WRITE (drb
))))
1793 free_dependence_relations (then_ddrs
);
1794 free_dependence_relations (else_ddrs
);
1795 free_data_refs (then_datarefs
);
1796 free_data_refs (else_datarefs
);
1801 /* Check that there are no read-after-write or write-after-write dependencies
1803 FOR_EACH_VEC_ELT (else_ddrs
, i
, ddr
)
1805 struct data_reference
*dra
= DDR_A (ddr
);
1806 struct data_reference
*drb
= DDR_B (ddr
);
1808 if (DDR_ARE_DEPENDENT (ddr
) != chrec_known
1809 && ((DR_IS_READ (dra
) && DR_IS_WRITE (drb
)
1810 && gimple_uid (DR_STMT (dra
)) > gimple_uid (DR_STMT (drb
)))
1811 || (DR_IS_READ (drb
) && DR_IS_WRITE (dra
)
1812 && gimple_uid (DR_STMT (drb
)) > gimple_uid (DR_STMT (dra
)))
1813 || (DR_IS_WRITE (dra
) && DR_IS_WRITE (drb
))))
1815 free_dependence_relations (then_ddrs
);
1816 free_dependence_relations (else_ddrs
);
1817 free_data_refs (then_datarefs
);
1818 free_data_refs (else_datarefs
);
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
);
1840 /* Return TRUE if STMT has a VUSE whose corresponding VDEF is in BB. */
1843 local_mem_dependence (gimple stmt
, basic_block bb
)
1845 tree vuse
= gimple_vuse (stmt
);
1851 def
= SSA_NAME_DEF_STMT (vuse
);
1852 return (def
&& gimple_bb (def
) == bb
);
1855 /* Given a "diamond" control-flow pattern where BB0 tests a condition,
1856 BB1 and BB2 are "then" and "else" blocks dependent on this test,
1857 and BB3 rejoins control flow following BB1 and BB2, look for
1858 opportunities to hoist loads as follows. If BB3 contains a PHI of
1859 two loads, one each occurring in BB1 and BB2, and the loads are
1860 provably of adjacent fields in the same structure, then move both
1861 loads into BB0. Of course this can only be done if there are no
1862 dependencies preventing such motion.
1864 One of the hoisted loads will always be speculative, so the
1865 transformation is currently conservative:
1867 - The fields must be strictly adjacent.
1868 - The two fields must occupy a single memory block that is
1869 guaranteed to not cross a page boundary.
1871 The last is difficult to prove, as such memory blocks should be
1872 aligned on the minimum of the stack alignment boundary and the
1873 alignment guaranteed by heap allocation interfaces. Thus we rely
1874 on a parameter for the alignment value.
1876 Provided a good value is used for the last case, the first
1877 restriction could possibly be relaxed. */
1880 hoist_adjacent_loads (basic_block bb0
, basic_block bb1
,
1881 basic_block bb2
, basic_block bb3
)
1883 int param_align
= PARAM_VALUE (PARAM_L1_CACHE_LINE_SIZE
);
1884 unsigned param_align_bits
= (unsigned) (param_align
* BITS_PER_UNIT
);
1887 /* Walk the phis in bb3 looking for an opportunity. We are looking
1888 for phis of two SSA names, one each of which is defined in bb1 and
1890 for (gsi
= gsi_start_phis (bb3
); !gsi_end_p (gsi
); gsi_next (&gsi
))
1892 gphi
*phi_stmt
= gsi
.phi ();
1894 tree arg1
, arg2
, ref1
, ref2
, field1
, field2
;
1895 tree tree_offset1
, tree_offset2
, tree_size2
, next
;
1896 int offset1
, offset2
, size2
;
1898 gimple_stmt_iterator gsi2
;
1899 basic_block bb_for_def1
, bb_for_def2
;
1901 if (gimple_phi_num_args (phi_stmt
) != 2
1902 || virtual_operand_p (gimple_phi_result (phi_stmt
)))
1905 arg1
= gimple_phi_arg_def (phi_stmt
, 0);
1906 arg2
= gimple_phi_arg_def (phi_stmt
, 1);
1908 if (TREE_CODE (arg1
) != SSA_NAME
1909 || TREE_CODE (arg2
) != SSA_NAME
1910 || SSA_NAME_IS_DEFAULT_DEF (arg1
)
1911 || SSA_NAME_IS_DEFAULT_DEF (arg2
))
1914 def1
= SSA_NAME_DEF_STMT (arg1
);
1915 def2
= SSA_NAME_DEF_STMT (arg2
);
1917 if ((gimple_bb (def1
) != bb1
|| gimple_bb (def2
) != bb2
)
1918 && (gimple_bb (def2
) != bb1
|| gimple_bb (def1
) != bb2
))
1921 /* Check the mode of the arguments to be sure a conditional move
1922 can be generated for it. */
1923 if (optab_handler (movcc_optab
, TYPE_MODE (TREE_TYPE (arg1
)))
1924 == CODE_FOR_nothing
)
1927 /* Both statements must be assignments whose RHS is a COMPONENT_REF. */
1928 if (!gimple_assign_single_p (def1
)
1929 || !gimple_assign_single_p (def2
)
1930 || gimple_has_volatile_ops (def1
)
1931 || gimple_has_volatile_ops (def2
))
1934 ref1
= gimple_assign_rhs1 (def1
);
1935 ref2
= gimple_assign_rhs1 (def2
);
1937 if (TREE_CODE (ref1
) != COMPONENT_REF
1938 || TREE_CODE (ref2
) != COMPONENT_REF
)
1941 /* The zeroth operand of the two component references must be
1942 identical. It is not sufficient to compare get_base_address of
1943 the two references, because this could allow for different
1944 elements of the same array in the two trees. It is not safe to
1945 assume that the existence of one array element implies the
1946 existence of a different one. */
1947 if (!operand_equal_p (TREE_OPERAND (ref1
, 0), TREE_OPERAND (ref2
, 0), 0))
1950 field1
= TREE_OPERAND (ref1
, 1);
1951 field2
= TREE_OPERAND (ref2
, 1);
1953 /* Check for field adjacency, and ensure field1 comes first. */
1954 for (next
= DECL_CHAIN (field1
);
1955 next
&& TREE_CODE (next
) != FIELD_DECL
;
1956 next
= DECL_CHAIN (next
))
1961 for (next
= DECL_CHAIN (field2
);
1962 next
&& TREE_CODE (next
) != FIELD_DECL
;
1963 next
= DECL_CHAIN (next
))
1969 std::swap (field1
, field2
);
1970 std::swap (def1
, def2
);
1973 bb_for_def1
= gimple_bb (def1
);
1974 bb_for_def2
= gimple_bb (def2
);
1976 /* Check for proper alignment of the first field. */
1977 tree_offset1
= bit_position (field1
);
1978 tree_offset2
= bit_position (field2
);
1979 tree_size2
= DECL_SIZE (field2
);
1981 if (!tree_fits_uhwi_p (tree_offset1
)
1982 || !tree_fits_uhwi_p (tree_offset2
)
1983 || !tree_fits_uhwi_p (tree_size2
))
1986 offset1
= tree_to_uhwi (tree_offset1
);
1987 offset2
= tree_to_uhwi (tree_offset2
);
1988 size2
= tree_to_uhwi (tree_size2
);
1989 align1
= DECL_ALIGN (field1
) % param_align_bits
;
1991 if (offset1
% BITS_PER_UNIT
!= 0)
1994 /* For profitability, the two field references should fit within
1995 a single cache line. */
1996 if (align1
+ offset2
- offset1
+ size2
> param_align_bits
)
1999 /* The two expressions cannot be dependent upon vdefs defined
2001 if (local_mem_dependence (def1
, bb_for_def1
)
2002 || local_mem_dependence (def2
, bb_for_def2
))
2005 /* The conditions are satisfied; hoist the loads from bb1 and bb2 into
2006 bb0. We hoist the first one first so that a cache miss is handled
2007 efficiently regardless of hardware cache-fill policy. */
2008 gsi2
= gsi_for_stmt (def1
);
2009 gsi_move_to_bb_end (&gsi2
, bb0
);
2010 gsi2
= gsi_for_stmt (def2
);
2011 gsi_move_to_bb_end (&gsi2
, bb0
);
2013 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
2016 "\nHoisting adjacent loads from %d and %d into %d: \n",
2017 bb_for_def1
->index
, bb_for_def2
->index
, bb0
->index
);
2018 print_gimple_stmt (dump_file
, def1
, 0, TDF_VOPS
|TDF_MEMSYMS
);
2019 print_gimple_stmt (dump_file
, def2
, 0, TDF_VOPS
|TDF_MEMSYMS
);
2024 /* Determine whether we should attempt to hoist adjacent loads out of
2025 diamond patterns in pass_phiopt. Always hoist loads if
2026 -fhoist-adjacent-loads is specified and the target machine has
2027 both a conditional move instruction and a defined cache line size. */
2030 gate_hoist_loads (void)
2032 return (flag_hoist_adjacent_loads
== 1
2033 && PARAM_VALUE (PARAM_L1_CACHE_LINE_SIZE
)
2034 && HAVE_conditional_move
);
2037 /* This pass tries to replaces an if-then-else block with an
2038 assignment. We have four kinds of transformations. Some of these
2039 transformations are also performed by the ifcvt RTL optimizer.
2041 Conditional Replacement
2042 -----------------------
2044 This transformation, implemented in conditional_replacement,
2048 if (cond) goto bb2; else goto bb1;
2051 x = PHI <0 (bb1), 1 (bb0), ...>;
2059 x = PHI <x' (bb0), ...>;
2061 We remove bb1 as it becomes unreachable. This occurs often due to
2062 gimplification of conditionals.
2067 This transformation, implemented in value_replacement, replaces
2070 if (a != b) goto bb2; else goto bb1;
2073 x = PHI <a (bb1), b (bb0), ...>;
2079 x = PHI <b (bb0), ...>;
2081 This opportunity can sometimes occur as a result of other
2085 Another case caught by value replacement looks like this:
2091 if (t3 != 0) goto bb1; else goto bb2;
2107 This transformation, implemented in abs_replacement, replaces
2110 if (a >= 0) goto bb2; else goto bb1;
2114 x = PHI <x (bb1), a (bb0), ...>;
2121 x = PHI <x' (bb0), ...>;
2126 This transformation, minmax_replacement replaces
2129 if (a <= b) goto bb2; else goto bb1;
2132 x = PHI <b (bb1), a (bb0), ...>;
2137 x' = MIN_EXPR (a, b)
2139 x = PHI <x' (bb0), ...>;
2141 A similar transformation is done for MAX_EXPR.
2144 This pass also performs a fifth transformation of a slightly different
2147 Adjacent Load Hoisting
2148 ----------------------
2150 This transformation replaces
2153 if (...) goto bb2; else goto bb1;
2155 x1 = (<expr>).field1;
2158 x2 = (<expr>).field2;
2165 x1 = (<expr>).field1;
2166 x2 = (<expr>).field2;
2167 if (...) goto bb2; else goto bb1;
2174 The purpose of this transformation is to enable generation of conditional
2175 move instructions such as Intel CMOVE or PowerPC ISEL. Because one of
2176 the loads is speculative, the transformation is restricted to very
2177 specific cases to avoid introducing a page fault. We are looking for
2185 where left and right are typically adjacent pointers in a tree structure. */
2189 const pass_data pass_data_phiopt
=
2191 GIMPLE_PASS
, /* type */
2192 "phiopt", /* name */
2193 OPTGROUP_NONE
, /* optinfo_flags */
2194 TV_TREE_PHIOPT
, /* tv_id */
2195 ( PROP_cfg
| PROP_ssa
), /* properties_required */
2196 0, /* properties_provided */
2197 0, /* properties_destroyed */
2198 0, /* todo_flags_start */
2199 0, /* todo_flags_finish */
2202 class pass_phiopt
: public gimple_opt_pass
2205 pass_phiopt (gcc::context
*ctxt
)
2206 : gimple_opt_pass (pass_data_phiopt
, ctxt
)
2209 /* opt_pass methods: */
2210 opt_pass
* clone () { return new pass_phiopt (m_ctxt
); }
2211 virtual bool gate (function
*) { return flag_ssa_phiopt
; }
2212 virtual unsigned int execute (function
*)
2214 return tree_ssa_phiopt_worker (false, gate_hoist_loads ());
2217 }; // class pass_phiopt
2222 make_pass_phiopt (gcc::context
*ctxt
)
2224 return new pass_phiopt (ctxt
);
2229 const pass_data pass_data_cselim
=
2231 GIMPLE_PASS
, /* type */
2232 "cselim", /* name */
2233 OPTGROUP_NONE
, /* optinfo_flags */
2234 TV_TREE_PHIOPT
, /* tv_id */
2235 ( PROP_cfg
| PROP_ssa
), /* properties_required */
2236 0, /* properties_provided */
2237 0, /* properties_destroyed */
2238 0, /* todo_flags_start */
2239 0, /* todo_flags_finish */
2242 class pass_cselim
: public gimple_opt_pass
2245 pass_cselim (gcc::context
*ctxt
)
2246 : gimple_opt_pass (pass_data_cselim
, ctxt
)
2249 /* opt_pass methods: */
2250 virtual bool gate (function
*) { return flag_tree_cselim
; }
2251 virtual unsigned int execute (function
*) { return tree_ssa_cs_elim (); }
2253 }; // class pass_cselim
2258 make_pass_cselim (gcc::context
*ctxt
)
2260 return new pass_cselim (ctxt
);