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"
30 #include "fold-const.h"
31 #include "stor-layout.h"
35 #include "internal-fn.h"
37 #include "gimple-iterator.h"
38 #include "gimplify-me.h"
40 #include "insn-config.h"
50 #include "tree-pass.h"
51 #include "langhooks.h"
54 #include "tree-data-ref.h"
55 #include "gimple-pretty-print.h"
56 #include "insn-codes.h"
58 #include "tree-scalar-evolution.h"
59 #include "tree-inline.h"
62 static unsigned int tree_ssa_phiopt_worker (bool, bool);
63 static bool conditional_replacement (basic_block
, basic_block
,
64 edge
, edge
, gphi
*, tree
, tree
);
65 static bool factor_out_conditional_conversion (edge
, edge
, gphi
*, 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 cond_store_replacement (basic_block
, basic_block
, edge
, edge
,
74 static bool cond_if_else_store_replacement (basic_block
, basic_block
, basic_block
);
75 static hash_set
<tree
> * get_non_trapping ();
76 static void replace_phi_edge_with_variable (basic_block
, edge
, gimple
, tree
);
77 static void hoist_adjacent_loads (basic_block
, basic_block
,
78 basic_block
, basic_block
);
79 static bool gate_hoist_loads (void);
81 /* This pass tries to transform conditional stores into unconditional
82 ones, enabling further simplifications with the simpler then and else
83 blocks. In particular it replaces this:
86 if (cond) goto bb2; else goto bb1;
94 if (cond) goto bb1; else goto bb2;
98 condtmp = PHI <RHS, condtmp'>
101 This transformation can only be done under several constraints,
102 documented below. It also replaces:
105 if (cond) goto bb2; else goto bb1;
116 if (cond) goto bb3; else goto bb1;
119 condtmp = PHI <RHS1, RHS2>
123 tree_ssa_cs_elim (void)
126 /* ??? We are not interested in loop related info, but the following
127 will create it, ICEing as we didn't init loops with pre-headers.
128 An interfacing issue of find_data_references_in_bb. */
129 loop_optimizer_init (LOOPS_NORMAL
);
131 todo
= tree_ssa_phiopt_worker (true, false);
133 loop_optimizer_finalize ();
137 /* Return the singleton PHI in the SEQ of PHIs for edges E0 and E1. */
140 single_non_singleton_phi_for_edges (gimple_seq seq
, edge e0
, edge e1
)
142 gimple_stmt_iterator i
;
144 if (gimple_seq_singleton_p (seq
))
145 return as_a
<gphi
*> (gsi_stmt (gsi_start (seq
)));
146 for (i
= gsi_start (seq
); !gsi_end_p (i
); gsi_next (&i
))
148 gphi
*p
= as_a
<gphi
*> (gsi_stmt (i
));
149 /* If the PHI arguments are equal then we can skip this PHI. */
150 if (operand_equal_for_phi_arg_p (gimple_phi_arg_def (p
, e0
->dest_idx
),
151 gimple_phi_arg_def (p
, e1
->dest_idx
)))
154 /* If we already have a PHI that has the two edge arguments are
155 different, then return it is not a singleton for these PHIs. */
164 /* The core routine of conditional store replacement and normal
165 phi optimizations. Both share much of the infrastructure in how
166 to match applicable basic block patterns. DO_STORE_ELIM is true
167 when we want to do conditional store replacement, false otherwise.
168 DO_HOIST_LOADS is true when we want to hoist adjacent loads out
169 of diamond control flow patterns, false otherwise. */
171 tree_ssa_phiopt_worker (bool do_store_elim
, bool do_hoist_loads
)
174 basic_block
*bb_order
;
176 bool cfgchanged
= false;
177 hash_set
<tree
> *nontrap
= 0;
180 /* Calculate the set of non-trapping memory accesses. */
181 nontrap
= get_non_trapping ();
183 /* Search every basic block for COND_EXPR we may be able to optimize.
185 We walk the blocks in order that guarantees that a block with
186 a single predecessor is processed before the predecessor.
187 This ensures that we collapse inner ifs before visiting the
188 outer ones, and also that we do not try to visit a removed
190 bb_order
= single_pred_before_succ_order ();
191 n
= n_basic_blocks_for_fn (cfun
) - NUM_FIXED_BLOCKS
;
193 for (i
= 0; i
< n
; i
++)
197 basic_block bb1
, bb2
;
203 cond_stmt
= last_stmt (bb
);
204 /* Check to see if the last statement is a GIMPLE_COND. */
206 || gimple_code (cond_stmt
) != GIMPLE_COND
)
209 e1
= EDGE_SUCC (bb
, 0);
211 e2
= EDGE_SUCC (bb
, 1);
214 /* We cannot do the optimization on abnormal edges. */
215 if ((e1
->flags
& EDGE_ABNORMAL
) != 0
216 || (e2
->flags
& EDGE_ABNORMAL
) != 0)
219 /* If either bb1's succ or bb2 or bb2's succ is non NULL. */
220 if (EDGE_COUNT (bb1
->succs
) == 0
222 || EDGE_COUNT (bb2
->succs
) == 0)
225 /* Find the bb which is the fall through to the other. */
226 if (EDGE_SUCC (bb1
, 0)->dest
== bb2
)
228 else if (EDGE_SUCC (bb2
, 0)->dest
== bb1
)
230 std::swap (bb1
, bb2
);
233 else if (do_store_elim
234 && EDGE_SUCC (bb1
, 0)->dest
== EDGE_SUCC (bb2
, 0)->dest
)
236 basic_block bb3
= EDGE_SUCC (bb1
, 0)->dest
;
238 if (!single_succ_p (bb1
)
239 || (EDGE_SUCC (bb1
, 0)->flags
& EDGE_FALLTHRU
) == 0
240 || !single_succ_p (bb2
)
241 || (EDGE_SUCC (bb2
, 0)->flags
& EDGE_FALLTHRU
) == 0
242 || EDGE_COUNT (bb3
->preds
) != 2)
244 if (cond_if_else_store_replacement (bb1
, bb2
, bb3
))
248 else if (do_hoist_loads
249 && EDGE_SUCC (bb1
, 0)->dest
== EDGE_SUCC (bb2
, 0)->dest
)
251 basic_block bb3
= EDGE_SUCC (bb1
, 0)->dest
;
253 if (!FLOAT_TYPE_P (TREE_TYPE (gimple_cond_lhs (cond_stmt
)))
254 && single_succ_p (bb1
)
255 && single_succ_p (bb2
)
256 && single_pred_p (bb1
)
257 && single_pred_p (bb2
)
258 && EDGE_COUNT (bb
->succs
) == 2
259 && EDGE_COUNT (bb3
->preds
) == 2
260 /* If one edge or the other is dominant, a conditional move
261 is likely to perform worse than the well-predicted branch. */
262 && !predictable_edge_p (EDGE_SUCC (bb
, 0))
263 && !predictable_edge_p (EDGE_SUCC (bb
, 1)))
264 hoist_adjacent_loads (bb
, bb1
, bb2
, bb3
);
270 e1
= EDGE_SUCC (bb1
, 0);
272 /* Make sure that bb1 is just a fall through. */
273 if (!single_succ_p (bb1
)
274 || (e1
->flags
& EDGE_FALLTHRU
) == 0)
277 /* Also make sure that bb1 only have one predecessor and that it
279 if (!single_pred_p (bb1
)
280 || single_pred (bb1
) != bb
)
285 /* bb1 is the middle block, bb2 the join block, bb the split block,
286 e1 the fallthrough edge from bb1 to bb2. We can't do the
287 optimization if the join block has more than two predecessors. */
288 if (EDGE_COUNT (bb2
->preds
) > 2)
290 if (cond_store_replacement (bb1
, bb2
, e1
, e2
, nontrap
))
295 gimple_seq phis
= phi_nodes (bb2
);
296 gimple_stmt_iterator gsi
;
297 bool candorest
= true;
299 /* Value replacement can work with more than one PHI
300 so try that first. */
301 for (gsi
= gsi_start (phis
); !gsi_end_p (gsi
); gsi_next (&gsi
))
303 phi
= as_a
<gphi
*> (gsi_stmt (gsi
));
304 arg0
= gimple_phi_arg_def (phi
, e1
->dest_idx
);
305 arg1
= gimple_phi_arg_def (phi
, e2
->dest_idx
);
306 if (value_replacement (bb
, bb1
, e1
, e2
, phi
, arg0
, arg1
) == 2)
317 phi
= single_non_singleton_phi_for_edges (phis
, e1
, e2
);
321 arg0
= gimple_phi_arg_def (phi
, e1
->dest_idx
);
322 arg1
= gimple_phi_arg_def (phi
, e2
->dest_idx
);
324 /* Something is wrong if we cannot find the arguments in the PHI
326 gcc_assert (arg0
!= NULL
&& arg1
!= NULL
);
328 if (factor_out_conditional_conversion (e1
, e2
, phi
, arg0
, arg1
))
330 /* factor_out_conditional_conversion may create a new PHI in
331 BB2 and eliminate an existing PHI in BB2. Recompute values
332 that may be affected by that change. */
333 phis
= phi_nodes (bb2
);
334 phi
= single_non_singleton_phi_for_edges (phis
, e1
, e2
);
336 arg0
= gimple_phi_arg_def (phi
, e1
->dest_idx
);
337 arg1
= gimple_phi_arg_def (phi
, e2
->dest_idx
);
338 gcc_assert (arg0
!= NULL
&& arg1
!= NULL
);
341 /* Do the replacement of conditional if it can be done. */
342 if (conditional_replacement (bb
, bb1
, e1
, e2
, phi
, arg0
, arg1
))
344 else if (abs_replacement (bb
, bb1
, e1
, e2
, phi
, arg0
, arg1
))
346 else if (minmax_replacement (bb
, bb1
, e1
, e2
, phi
, arg0
, arg1
))
355 /* If the CFG has changed, we should cleanup the CFG. */
356 if (cfgchanged
&& do_store_elim
)
358 /* In cond-store replacement we have added some loads on edges
359 and new VOPS (as we moved the store, and created a load). */
360 gsi_commit_edge_inserts ();
361 return TODO_cleanup_cfg
| TODO_update_ssa_only_virtuals
;
364 return TODO_cleanup_cfg
;
368 /* Replace PHI node element whose edge is E in block BB with variable NEW.
369 Remove the edge from COND_BLOCK which does not lead to BB (COND_BLOCK
370 is known to have two edges, one of which must reach BB). */
373 replace_phi_edge_with_variable (basic_block cond_block
,
374 edge e
, gimple phi
, tree new_tree
)
376 basic_block bb
= gimple_bb (phi
);
377 basic_block block_to_remove
;
378 gimple_stmt_iterator gsi
;
380 /* Change the PHI argument to new. */
381 SET_USE (PHI_ARG_DEF_PTR (phi
, e
->dest_idx
), new_tree
);
383 /* Remove the empty basic block. */
384 if (EDGE_SUCC (cond_block
, 0)->dest
== bb
)
386 EDGE_SUCC (cond_block
, 0)->flags
|= EDGE_FALLTHRU
;
387 EDGE_SUCC (cond_block
, 0)->flags
&= ~(EDGE_TRUE_VALUE
| EDGE_FALSE_VALUE
);
388 EDGE_SUCC (cond_block
, 0)->probability
= REG_BR_PROB_BASE
;
389 EDGE_SUCC (cond_block
, 0)->count
+= EDGE_SUCC (cond_block
, 1)->count
;
391 block_to_remove
= EDGE_SUCC (cond_block
, 1)->dest
;
395 EDGE_SUCC (cond_block
, 1)->flags
|= EDGE_FALLTHRU
;
396 EDGE_SUCC (cond_block
, 1)->flags
397 &= ~(EDGE_TRUE_VALUE
| EDGE_FALSE_VALUE
);
398 EDGE_SUCC (cond_block
, 1)->probability
= REG_BR_PROB_BASE
;
399 EDGE_SUCC (cond_block
, 1)->count
+= EDGE_SUCC (cond_block
, 0)->count
;
401 block_to_remove
= EDGE_SUCC (cond_block
, 0)->dest
;
403 delete_basic_block (block_to_remove
);
405 /* Eliminate the COND_EXPR at the end of COND_BLOCK. */
406 gsi
= gsi_last_bb (cond_block
);
407 gsi_remove (&gsi
, true);
409 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
411 "COND_EXPR in block %d and PHI in block %d converted to straightline code.\n",
416 /* PR66726: Factor conversion out of COND_EXPR. If the arguments of the PHI
417 stmt are CONVERT_STMT, factor out the conversion and perform the conversion
418 to the result of PHI stmt. */
421 factor_out_conditional_conversion (edge e0
, edge e1
, gphi
*phi
,
422 tree arg0
, tree arg1
)
424 gimple arg0_def_stmt
= NULL
, arg1_def_stmt
= NULL
, new_stmt
;
425 tree new_arg0
= NULL_TREE
, new_arg1
= NULL_TREE
;
428 gimple_stmt_iterator gsi
, gsi_for_def
;
429 source_location locus
= gimple_location (phi
);
430 enum tree_code convert_code
;
432 /* Handle only PHI statements with two arguments. TODO: If all
433 other arguments to PHI are INTEGER_CST or if their defining
434 statement have the same unary operation, we can handle more
435 than two arguments too. */
436 if (gimple_phi_num_args (phi
) != 2)
439 /* First canonicalize to simplify tests. */
440 if (TREE_CODE (arg0
) != SSA_NAME
)
442 std::swap (arg0
, arg1
);
446 if (TREE_CODE (arg0
) != SSA_NAME
447 || (TREE_CODE (arg1
) != SSA_NAME
448 && TREE_CODE (arg1
) != INTEGER_CST
))
451 /* Check if arg0 is an SSA_NAME and the stmt which defines arg0 is
453 arg0_def_stmt
= SSA_NAME_DEF_STMT (arg0
);
454 if (!is_gimple_assign (arg0_def_stmt
)
455 || !gimple_assign_cast_p (arg0_def_stmt
))
458 /* Use the RHS as new_arg0. */
459 convert_code
= gimple_assign_rhs_code (arg0_def_stmt
);
460 new_arg0
= gimple_assign_rhs1 (arg0_def_stmt
);
461 if (convert_code
== VIEW_CONVERT_EXPR
)
462 new_arg0
= TREE_OPERAND (new_arg0
, 0);
464 if (TREE_CODE (arg1
) == SSA_NAME
)
466 /* Check if arg1 is an SSA_NAME and the stmt which defines arg1
468 arg1_def_stmt
= SSA_NAME_DEF_STMT (arg1
);
469 if (!is_gimple_assign (arg1_def_stmt
)
470 || gimple_assign_rhs_code (arg1_def_stmt
) != convert_code
)
473 /* Use the RHS as new_arg1. */
474 new_arg1
= gimple_assign_rhs1 (arg1_def_stmt
);
475 if (convert_code
== VIEW_CONVERT_EXPR
)
476 new_arg1
= TREE_OPERAND (new_arg1
, 0);
480 /* If arg1 is an INTEGER_CST, fold it to new type. */
481 if (INTEGRAL_TYPE_P (TREE_TYPE (new_arg0
))
482 && int_fits_type_p (arg1
, TREE_TYPE (new_arg0
)))
484 if (gimple_assign_cast_p (arg0_def_stmt
))
485 new_arg1
= fold_convert (TREE_TYPE (new_arg0
), arg1
);
493 /* If arg0/arg1 have > 1 use, then this transformation actually increases
494 the number of expressions evaluated at runtime. */
495 if (!has_single_use (arg0
)
496 || (arg1_def_stmt
&& !has_single_use (arg1
)))
499 /* If types of new_arg0 and new_arg1 are different bailout. */
500 if (!types_compatible_p (TREE_TYPE (new_arg0
), TREE_TYPE (new_arg1
)))
503 /* Create a new PHI stmt. */
504 result
= PHI_RESULT (phi
);
505 temp
= make_ssa_name (TREE_TYPE (new_arg0
), NULL
);
506 newphi
= create_phi_node (temp
, gimple_bb (phi
));
508 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
510 fprintf (dump_file
, "PHI ");
511 print_generic_expr (dump_file
, gimple_phi_result (phi
), 0);
513 " changed to factor conversion out from COND_EXPR.\n");
514 fprintf (dump_file
, "New stmt with CAST that defines ");
515 print_generic_expr (dump_file
, result
, 0);
516 fprintf (dump_file
, ".\n");
519 /* Remove the old cast(s) that has single use. */
520 gsi_for_def
= gsi_for_stmt (arg0_def_stmt
);
521 gsi_remove (&gsi_for_def
, true);
524 gsi_for_def
= gsi_for_stmt (arg1_def_stmt
);
525 gsi_remove (&gsi_for_def
, true);
528 add_phi_arg (newphi
, new_arg0
, e0
, locus
);
529 add_phi_arg (newphi
, new_arg1
, e1
, locus
);
531 /* Create the conversion stmt and insert it. */
532 if (convert_code
== VIEW_CONVERT_EXPR
)
533 temp
= fold_build1 (VIEW_CONVERT_EXPR
, TREE_TYPE (result
), temp
);
534 new_stmt
= gimple_build_assign (result
, convert_code
, temp
);
535 gsi
= gsi_after_labels (gimple_bb (phi
));
536 gsi_insert_before (&gsi
, new_stmt
, GSI_SAME_STMT
);
538 /* Remove he original PHI stmt. */
539 gsi
= gsi_for_stmt (phi
);
540 gsi_remove (&gsi
, true);
544 /* The function conditional_replacement does the main work of doing the
545 conditional replacement. Return true if the replacement is done.
546 Otherwise return false.
547 BB is the basic block where the replacement is going to be done on. ARG0
548 is argument 0 from PHI. Likewise for ARG1. */
551 conditional_replacement (basic_block cond_bb
, basic_block middle_bb
,
552 edge e0
, edge e1
, gphi
*phi
,
553 tree arg0
, tree arg1
)
559 gimple_stmt_iterator gsi
;
560 edge true_edge
, false_edge
;
561 tree new_var
, new_var2
;
564 /* FIXME: Gimplification of complex type is too hard for now. */
565 /* We aren't prepared to handle vectors either (and it is a question
566 if it would be worthwhile anyway). */
567 if (!(INTEGRAL_TYPE_P (TREE_TYPE (arg0
))
568 || POINTER_TYPE_P (TREE_TYPE (arg0
)))
569 || !(INTEGRAL_TYPE_P (TREE_TYPE (arg1
))
570 || POINTER_TYPE_P (TREE_TYPE (arg1
))))
573 /* The PHI arguments have the constants 0 and 1, or 0 and -1, then
574 convert it to the conditional. */
575 if ((integer_zerop (arg0
) && integer_onep (arg1
))
576 || (integer_zerop (arg1
) && integer_onep (arg0
)))
578 else if ((integer_zerop (arg0
) && integer_all_onesp (arg1
))
579 || (integer_zerop (arg1
) && integer_all_onesp (arg0
)))
584 if (!empty_block_p (middle_bb
))
587 /* At this point we know we have a GIMPLE_COND with two successors.
588 One successor is BB, the other successor is an empty block which
589 falls through into BB.
591 There is a single PHI node at the join point (BB) and its arguments
592 are constants (0, 1) or (0, -1).
594 So, given the condition COND, and the two PHI arguments, we can
595 rewrite this PHI into non-branching code:
597 dest = (COND) or dest = COND'
599 We use the condition as-is if the argument associated with the
600 true edge has the value one or the argument associated with the
601 false edge as the value zero. Note that those conditions are not
602 the same since only one of the outgoing edges from the GIMPLE_COND
603 will directly reach BB and thus be associated with an argument. */
605 stmt
= last_stmt (cond_bb
);
606 result
= PHI_RESULT (phi
);
608 /* To handle special cases like floating point comparison, it is easier and
609 less error-prone to build a tree and gimplify it on the fly though it is
611 cond
= fold_build2_loc (gimple_location (stmt
),
612 gimple_cond_code (stmt
), boolean_type_node
,
613 gimple_cond_lhs (stmt
), gimple_cond_rhs (stmt
));
615 /* We need to know which is the true edge and which is the false
616 edge so that we know when to invert the condition below. */
617 extract_true_false_edges_from_block (cond_bb
, &true_edge
, &false_edge
);
618 if ((e0
== true_edge
&& integer_zerop (arg0
))
619 || (e0
== false_edge
&& !integer_zerop (arg0
))
620 || (e1
== true_edge
&& integer_zerop (arg1
))
621 || (e1
== false_edge
&& !integer_zerop (arg1
)))
622 cond
= fold_build1_loc (gimple_location (stmt
),
623 TRUTH_NOT_EXPR
, TREE_TYPE (cond
), cond
);
627 cond
= fold_convert_loc (gimple_location (stmt
),
628 TREE_TYPE (result
), cond
);
629 cond
= fold_build1_loc (gimple_location (stmt
),
630 NEGATE_EXPR
, TREE_TYPE (cond
), cond
);
633 /* Insert our new statements at the end of conditional block before the
635 gsi
= gsi_for_stmt (stmt
);
636 new_var
= force_gimple_operand_gsi (&gsi
, cond
, true, NULL
, true,
639 if (!useless_type_conversion_p (TREE_TYPE (result
), TREE_TYPE (new_var
)))
641 source_location locus_0
, locus_1
;
643 new_var2
= make_ssa_name (TREE_TYPE (result
));
644 new_stmt
= gimple_build_assign (new_var2
, CONVERT_EXPR
, new_var
);
645 gsi_insert_before (&gsi
, new_stmt
, GSI_SAME_STMT
);
648 /* Set the locus to the first argument, unless is doesn't have one. */
649 locus_0
= gimple_phi_arg_location (phi
, 0);
650 locus_1
= gimple_phi_arg_location (phi
, 1);
651 if (locus_0
== UNKNOWN_LOCATION
)
653 gimple_set_location (new_stmt
, locus_0
);
656 replace_phi_edge_with_variable (cond_bb
, e1
, phi
, new_var
);
658 /* Note that we optimized this PHI. */
662 /* Update *ARG which is defined in STMT so that it contains the
663 computed value if that seems profitable. Return true if the
664 statement is made dead by that rewriting. */
667 jump_function_from_stmt (tree
*arg
, gimple stmt
)
669 enum tree_code code
= gimple_assign_rhs_code (stmt
);
670 if (code
== ADDR_EXPR
)
672 /* For arg = &p->i transform it to p, if possible. */
673 tree rhs1
= gimple_assign_rhs1 (stmt
);
674 HOST_WIDE_INT offset
;
675 tree tem
= get_addr_base_and_unit_offset (TREE_OPERAND (rhs1
, 0),
678 && TREE_CODE (tem
) == MEM_REF
679 && (mem_ref_offset (tem
) + offset
) == 0)
681 *arg
= TREE_OPERAND (tem
, 0);
685 /* TODO: Much like IPA-CP jump-functions we want to handle constant
686 additions symbolically here, and we'd need to update the comparison
687 code that compares the arg + cst tuples in our caller. For now the
688 code above exactly handles the VEC_BASE pattern from vec.h. */
692 /* RHS is a source argument in a BIT_AND_EXPR which feeds a conditional
693 of the form SSA_NAME NE 0.
695 If RHS is fed by a simple EQ_EXPR comparison of two values, see if
696 the two input values of the EQ_EXPR match arg0 and arg1.
698 If so update *code and return TRUE. Otherwise return FALSE. */
701 rhs_is_fed_for_value_replacement (const_tree arg0
, const_tree arg1
,
702 enum tree_code
*code
, const_tree rhs
)
704 /* Obviously if RHS is not an SSA_NAME, we can't look at the defining
706 if (TREE_CODE (rhs
) == SSA_NAME
)
708 gimple def1
= SSA_NAME_DEF_STMT (rhs
);
710 /* Verify the defining statement has an EQ_EXPR on the RHS. */
711 if (is_gimple_assign (def1
) && gimple_assign_rhs_code (def1
) == EQ_EXPR
)
713 /* Finally verify the source operands of the EQ_EXPR are equal
715 tree op0
= gimple_assign_rhs1 (def1
);
716 tree op1
= gimple_assign_rhs2 (def1
);
717 if ((operand_equal_for_phi_arg_p (arg0
, op0
)
718 && operand_equal_for_phi_arg_p (arg1
, op1
))
719 || (operand_equal_for_phi_arg_p (arg0
, op1
)
720 && operand_equal_for_phi_arg_p (arg1
, op0
)))
722 /* We will perform the optimization. */
723 *code
= gimple_assign_rhs_code (def1
);
731 /* Return TRUE if arg0/arg1 are equal to the rhs/lhs or lhs/rhs of COND.
733 Also return TRUE if arg0/arg1 are equal to the source arguments of a
734 an EQ comparison feeding a BIT_AND_EXPR which feeds COND.
736 Return FALSE otherwise. */
739 operand_equal_for_value_replacement (const_tree arg0
, const_tree arg1
,
740 enum tree_code
*code
, gimple cond
)
743 tree lhs
= gimple_cond_lhs (cond
);
744 tree rhs
= gimple_cond_rhs (cond
);
746 if ((operand_equal_for_phi_arg_p (arg0
, lhs
)
747 && operand_equal_for_phi_arg_p (arg1
, rhs
))
748 || (operand_equal_for_phi_arg_p (arg1
, lhs
)
749 && operand_equal_for_phi_arg_p (arg0
, rhs
)))
752 /* Now handle more complex case where we have an EQ comparison
753 which feeds a BIT_AND_EXPR which feeds COND.
755 First verify that COND is of the form SSA_NAME NE 0. */
756 if (*code
!= NE_EXPR
|| !integer_zerop (rhs
)
757 || TREE_CODE (lhs
) != SSA_NAME
)
760 /* Now ensure that SSA_NAME is set by a BIT_AND_EXPR. */
761 def
= SSA_NAME_DEF_STMT (lhs
);
762 if (!is_gimple_assign (def
) || gimple_assign_rhs_code (def
) != BIT_AND_EXPR
)
765 /* Now verify arg0/arg1 correspond to the source arguments of an
766 EQ comparison feeding the BIT_AND_EXPR. */
768 tree tmp
= gimple_assign_rhs1 (def
);
769 if (rhs_is_fed_for_value_replacement (arg0
, arg1
, code
, tmp
))
772 tmp
= gimple_assign_rhs2 (def
);
773 if (rhs_is_fed_for_value_replacement (arg0
, arg1
, code
, tmp
))
779 /* Returns true if ARG is a neutral element for operation CODE
780 on the RIGHT side. */
783 neutral_element_p (tree_code code
, tree arg
, bool right
)
790 return integer_zerop (arg
);
797 case POINTER_PLUS_EXPR
:
798 return right
&& integer_zerop (arg
);
801 return integer_onep (arg
);
808 return right
&& integer_onep (arg
);
811 return integer_all_onesp (arg
);
818 /* Returns true if ARG is an absorbing element for operation CODE. */
821 absorbing_element_p (tree_code code
, tree arg
)
826 return integer_all_onesp (arg
);
830 return integer_zerop (arg
);
837 /* The function value_replacement does the main work of doing the value
838 replacement. Return non-zero if the replacement is done. Otherwise return
839 0. If we remove the middle basic block, return 2.
840 BB is the basic block where the replacement is going to be done on. ARG0
841 is argument 0 from the PHI. Likewise for ARG1. */
844 value_replacement (basic_block cond_bb
, basic_block middle_bb
,
845 edge e0
, edge e1
, gimple phi
,
846 tree arg0
, tree arg1
)
848 gimple_stmt_iterator gsi
;
850 edge true_edge
, false_edge
;
852 bool emtpy_or_with_defined_p
= true;
854 /* If the type says honor signed zeros we cannot do this
856 if (HONOR_SIGNED_ZEROS (arg1
))
859 /* If there is a statement in MIDDLE_BB that defines one of the PHI
860 arguments, then adjust arg0 or arg1. */
861 gsi
= gsi_start_nondebug_after_labels_bb (middle_bb
);
862 while (!gsi_end_p (gsi
))
864 gimple stmt
= gsi_stmt (gsi
);
866 gsi_next_nondebug (&gsi
);
867 if (!is_gimple_assign (stmt
))
869 emtpy_or_with_defined_p
= false;
872 /* Now try to adjust arg0 or arg1 according to the computation
874 lhs
= gimple_assign_lhs (stmt
);
876 && jump_function_from_stmt (&arg0
, stmt
))
878 && jump_function_from_stmt (&arg1
, stmt
)))
879 emtpy_or_with_defined_p
= false;
882 cond
= last_stmt (cond_bb
);
883 code
= gimple_cond_code (cond
);
885 /* This transformation is only valid for equality comparisons. */
886 if (code
!= NE_EXPR
&& code
!= EQ_EXPR
)
889 /* We need to know which is the true edge and which is the false
890 edge so that we know if have abs or negative abs. */
891 extract_true_false_edges_from_block (cond_bb
, &true_edge
, &false_edge
);
893 /* At this point we know we have a COND_EXPR with two successors.
894 One successor is BB, the other successor is an empty block which
895 falls through into BB.
897 The condition for the COND_EXPR is known to be NE_EXPR or EQ_EXPR.
899 There is a single PHI node at the join point (BB) with two arguments.
901 We now need to verify that the two arguments in the PHI node match
902 the two arguments to the equality comparison. */
904 if (operand_equal_for_value_replacement (arg0
, arg1
, &code
, cond
))
909 /* For NE_EXPR, we want to build an assignment result = arg where
910 arg is the PHI argument associated with the true edge. For
911 EQ_EXPR we want the PHI argument associated with the false edge. */
912 e
= (code
== NE_EXPR
? true_edge
: false_edge
);
914 /* Unfortunately, E may not reach BB (it may instead have gone to
915 OTHER_BLOCK). If that is the case, then we want the single outgoing
916 edge from OTHER_BLOCK which reaches BB and represents the desired
917 path from COND_BLOCK. */
918 if (e
->dest
== middle_bb
)
919 e
= single_succ_edge (e
->dest
);
921 /* Now we know the incoming edge to BB that has the argument for the
922 RHS of our new assignment statement. */
928 /* If the middle basic block was empty or is defining the
929 PHI arguments and this is a single phi where the args are different
930 for the edges e0 and e1 then we can remove the middle basic block. */
931 if (emtpy_or_with_defined_p
932 && single_non_singleton_phi_for_edges (phi_nodes (gimple_bb (phi
)),
935 replace_phi_edge_with_variable (cond_bb
, e1
, phi
, arg
);
936 /* Note that we optimized this PHI. */
941 /* Replace the PHI arguments with arg. */
942 SET_PHI_ARG_DEF (phi
, e0
->dest_idx
, arg
);
943 SET_PHI_ARG_DEF (phi
, e1
->dest_idx
, arg
);
944 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
946 fprintf (dump_file
, "PHI ");
947 print_generic_expr (dump_file
, gimple_phi_result (phi
), 0);
948 fprintf (dump_file
, " reduced for COND_EXPR in block %d to ",
950 print_generic_expr (dump_file
, arg
, 0);
951 fprintf (dump_file
, ".\n");
958 /* Now optimize (x != 0) ? x + y : y to just y.
959 The following condition is too restrictive, there can easily be another
960 stmt in middle_bb, for instance a CONVERT_EXPR for the second argument. */
961 gimple assign
= last_and_only_stmt (middle_bb
);
962 if (!assign
|| gimple_code (assign
) != GIMPLE_ASSIGN
963 || gimple_assign_rhs_class (assign
) != GIMPLE_BINARY_RHS
964 || (!INTEGRAL_TYPE_P (TREE_TYPE (arg0
))
965 && !POINTER_TYPE_P (TREE_TYPE (arg0
))))
968 /* Punt if there are (degenerate) PHIs in middle_bb, there should not be. */
969 if (!gimple_seq_empty_p (phi_nodes (middle_bb
)))
972 /* Only transform if it removes the condition. */
973 if (!single_non_singleton_phi_for_edges (phi_nodes (gimple_bb (phi
)), e0
, e1
))
976 /* Size-wise, this is always profitable. */
977 if (optimize_bb_for_speed_p (cond_bb
)
978 /* The special case is useless if it has a low probability. */
979 && profile_status_for_fn (cfun
) != PROFILE_ABSENT
980 && EDGE_PRED (middle_bb
, 0)->probability
< PROB_EVEN
981 /* If assign is cheap, there is no point avoiding it. */
982 && estimate_num_insns (assign
, &eni_time_weights
)
983 >= 3 * estimate_num_insns (cond
, &eni_time_weights
))
986 tree lhs
= gimple_assign_lhs (assign
);
987 tree rhs1
= gimple_assign_rhs1 (assign
);
988 tree rhs2
= gimple_assign_rhs2 (assign
);
989 enum tree_code code_def
= gimple_assign_rhs_code (assign
);
990 tree cond_lhs
= gimple_cond_lhs (cond
);
991 tree cond_rhs
= gimple_cond_rhs (cond
);
993 if (((code
== NE_EXPR
&& e1
== false_edge
)
994 || (code
== EQ_EXPR
&& e1
== true_edge
))
997 && operand_equal_for_phi_arg_p (rhs2
, cond_lhs
)
998 && neutral_element_p (code_def
, cond_rhs
, true))
1000 && operand_equal_for_phi_arg_p (rhs1
, cond_lhs
)
1001 && neutral_element_p (code_def
, cond_rhs
, false))
1002 || (operand_equal_for_phi_arg_p (arg1
, cond_rhs
)
1003 && (operand_equal_for_phi_arg_p (rhs2
, cond_lhs
)
1004 || operand_equal_for_phi_arg_p (rhs1
, cond_lhs
))
1005 && absorbing_element_p (code_def
, cond_rhs
))))
1007 gsi
= gsi_for_stmt (cond
);
1008 if (INTEGRAL_TYPE_P (TREE_TYPE (lhs
)))
1010 /* Moving ASSIGN might change VR of lhs, e.g. when moving u_6
1018 # RANGE [0, 4294967294]
1019 u_6 = n_5 + 4294967295;
1022 # u_3 = PHI <u_6(3), 4294967295(2)> */
1023 SSA_NAME_RANGE_INFO (lhs
) = NULL
;
1024 SSA_NAME_ANTI_RANGE_P (lhs
) = 0;
1025 /* If available, we can use VR of phi result at least. */
1026 tree phires
= gimple_phi_result (phi
);
1027 struct range_info_def
*phires_range_info
1028 = SSA_NAME_RANGE_INFO (phires
);
1029 if (phires_range_info
)
1030 duplicate_ssa_name_range_info (lhs
, SSA_NAME_RANGE_TYPE (phires
),
1033 gimple_stmt_iterator gsi_from
= gsi_for_stmt (assign
);
1034 gsi_move_before (&gsi_from
, &gsi
);
1035 replace_phi_edge_with_variable (cond_bb
, e1
, phi
, lhs
);
1042 /* The function minmax_replacement does the main work of doing the minmax
1043 replacement. Return true if the replacement is done. Otherwise return
1045 BB is the basic block where the replacement is going to be done on. ARG0
1046 is argument 0 from the PHI. Likewise for ARG1. */
1049 minmax_replacement (basic_block cond_bb
, basic_block middle_bb
,
1050 edge e0
, edge e1
, gimple phi
,
1051 tree arg0
, tree arg1
)
1056 edge true_edge
, false_edge
;
1057 enum tree_code cmp
, minmax
, ass_code
;
1058 tree smaller
, larger
, arg_true
, arg_false
;
1059 gimple_stmt_iterator gsi
, gsi_from
;
1061 type
= TREE_TYPE (PHI_RESULT (phi
));
1063 /* The optimization may be unsafe due to NaNs. */
1064 if (HONOR_NANS (type
))
1067 cond
= as_a
<gcond
*> (last_stmt (cond_bb
));
1068 cmp
= gimple_cond_code (cond
);
1070 /* This transformation is only valid for order comparisons. Record which
1071 operand is smaller/larger if the result of the comparison is true. */
1072 if (cmp
== LT_EXPR
|| cmp
== LE_EXPR
)
1074 smaller
= gimple_cond_lhs (cond
);
1075 larger
= gimple_cond_rhs (cond
);
1077 else if (cmp
== GT_EXPR
|| cmp
== GE_EXPR
)
1079 smaller
= gimple_cond_rhs (cond
);
1080 larger
= gimple_cond_lhs (cond
);
1085 /* We need to know which is the true edge and which is the false
1086 edge so that we know if have abs or negative abs. */
1087 extract_true_false_edges_from_block (cond_bb
, &true_edge
, &false_edge
);
1089 /* Forward the edges over the middle basic block. */
1090 if (true_edge
->dest
== middle_bb
)
1091 true_edge
= EDGE_SUCC (true_edge
->dest
, 0);
1092 if (false_edge
->dest
== middle_bb
)
1093 false_edge
= EDGE_SUCC (false_edge
->dest
, 0);
1095 if (true_edge
== e0
)
1097 gcc_assert (false_edge
== e1
);
1103 gcc_assert (false_edge
== e0
);
1104 gcc_assert (true_edge
== e1
);
1109 if (empty_block_p (middle_bb
))
1111 if (operand_equal_for_phi_arg_p (arg_true
, smaller
)
1112 && operand_equal_for_phi_arg_p (arg_false
, larger
))
1116 if (smaller < larger)
1122 else if (operand_equal_for_phi_arg_p (arg_false
, smaller
)
1123 && operand_equal_for_phi_arg_p (arg_true
, larger
))
1130 /* Recognize the following case, assuming d <= u:
1136 This is equivalent to
1141 gimple assign
= last_and_only_stmt (middle_bb
);
1142 tree lhs
, op0
, op1
, bound
;
1145 || gimple_code (assign
) != GIMPLE_ASSIGN
)
1148 lhs
= gimple_assign_lhs (assign
);
1149 ass_code
= gimple_assign_rhs_code (assign
);
1150 if (ass_code
!= MAX_EXPR
&& ass_code
!= MIN_EXPR
)
1152 op0
= gimple_assign_rhs1 (assign
);
1153 op1
= gimple_assign_rhs2 (assign
);
1155 if (true_edge
->src
== middle_bb
)
1157 /* We got here if the condition is true, i.e., SMALLER < LARGER. */
1158 if (!operand_equal_for_phi_arg_p (lhs
, arg_true
))
1161 if (operand_equal_for_phi_arg_p (arg_false
, larger
))
1165 if (smaller < larger)
1167 r' = MAX_EXPR (smaller, bound)
1169 r = PHI <r', larger> --> to be turned to MIN_EXPR. */
1170 if (ass_code
!= MAX_EXPR
)
1174 if (operand_equal_for_phi_arg_p (op0
, smaller
))
1176 else if (operand_equal_for_phi_arg_p (op1
, smaller
))
1181 /* We need BOUND <= LARGER. */
1182 if (!integer_nonzerop (fold_build2 (LE_EXPR
, boolean_type_node
,
1186 else if (operand_equal_for_phi_arg_p (arg_false
, smaller
))
1190 if (smaller < larger)
1192 r' = MIN_EXPR (larger, bound)
1194 r = PHI <r', smaller> --> to be turned to MAX_EXPR. */
1195 if (ass_code
!= MIN_EXPR
)
1199 if (operand_equal_for_phi_arg_p (op0
, larger
))
1201 else if (operand_equal_for_phi_arg_p (op1
, larger
))
1206 /* We need BOUND >= SMALLER. */
1207 if (!integer_nonzerop (fold_build2 (GE_EXPR
, boolean_type_node
,
1216 /* We got here if the condition is false, i.e., SMALLER > LARGER. */
1217 if (!operand_equal_for_phi_arg_p (lhs
, arg_false
))
1220 if (operand_equal_for_phi_arg_p (arg_true
, larger
))
1224 if (smaller > larger)
1226 r' = MIN_EXPR (smaller, bound)
1228 r = PHI <r', larger> --> to be turned to MAX_EXPR. */
1229 if (ass_code
!= MIN_EXPR
)
1233 if (operand_equal_for_phi_arg_p (op0
, smaller
))
1235 else if (operand_equal_for_phi_arg_p (op1
, smaller
))
1240 /* We need BOUND >= LARGER. */
1241 if (!integer_nonzerop (fold_build2 (GE_EXPR
, boolean_type_node
,
1245 else if (operand_equal_for_phi_arg_p (arg_true
, smaller
))
1249 if (smaller > larger)
1251 r' = MAX_EXPR (larger, bound)
1253 r = PHI <r', smaller> --> to be turned to MIN_EXPR. */
1254 if (ass_code
!= MAX_EXPR
)
1258 if (operand_equal_for_phi_arg_p (op0
, larger
))
1260 else if (operand_equal_for_phi_arg_p (op1
, larger
))
1265 /* We need BOUND <= SMALLER. */
1266 if (!integer_nonzerop (fold_build2 (LE_EXPR
, boolean_type_node
,
1274 /* Move the statement from the middle block. */
1275 gsi
= gsi_last_bb (cond_bb
);
1276 gsi_from
= gsi_last_nondebug_bb (middle_bb
);
1277 gsi_move_before (&gsi_from
, &gsi
);
1280 /* Emit the statement to compute min/max. */
1281 result
= duplicate_ssa_name (PHI_RESULT (phi
), NULL
);
1282 new_stmt
= gimple_build_assign (result
, minmax
, arg0
, arg1
);
1283 gsi
= gsi_last_bb (cond_bb
);
1284 gsi_insert_before (&gsi
, new_stmt
, GSI_NEW_STMT
);
1286 replace_phi_edge_with_variable (cond_bb
, e1
, phi
, result
);
1290 /* The function absolute_replacement does the main work of doing the absolute
1291 replacement. Return true if the replacement is done. Otherwise return
1293 bb is the basic block where the replacement is going to be done on. arg0
1294 is argument 0 from the phi. Likewise for arg1. */
1297 abs_replacement (basic_block cond_bb
, basic_block middle_bb
,
1298 edge e0 ATTRIBUTE_UNUSED
, edge e1
,
1299 gimple phi
, tree arg0
, tree arg1
)
1304 gimple_stmt_iterator gsi
;
1305 edge true_edge
, false_edge
;
1310 enum tree_code cond_code
;
1312 /* If the type says honor signed zeros we cannot do this
1314 if (HONOR_SIGNED_ZEROS (arg1
))
1317 /* OTHER_BLOCK must have only one executable statement which must have the
1318 form arg0 = -arg1 or arg1 = -arg0. */
1320 assign
= last_and_only_stmt (middle_bb
);
1321 /* If we did not find the proper negation assignment, then we can not
1326 /* If we got here, then we have found the only executable statement
1327 in OTHER_BLOCK. If it is anything other than arg = -arg1 or
1328 arg1 = -arg0, then we can not optimize. */
1329 if (gimple_code (assign
) != GIMPLE_ASSIGN
)
1332 lhs
= gimple_assign_lhs (assign
);
1334 if (gimple_assign_rhs_code (assign
) != NEGATE_EXPR
)
1337 rhs
= gimple_assign_rhs1 (assign
);
1339 /* The assignment has to be arg0 = -arg1 or arg1 = -arg0. */
1340 if (!(lhs
== arg0
&& rhs
== arg1
)
1341 && !(lhs
== arg1
&& rhs
== arg0
))
1344 cond
= last_stmt (cond_bb
);
1345 result
= PHI_RESULT (phi
);
1347 /* Only relationals comparing arg[01] against zero are interesting. */
1348 cond_code
= gimple_cond_code (cond
);
1349 if (cond_code
!= GT_EXPR
&& cond_code
!= GE_EXPR
1350 && cond_code
!= LT_EXPR
&& cond_code
!= LE_EXPR
)
1353 /* Make sure the conditional is arg[01] OP y. */
1354 if (gimple_cond_lhs (cond
) != rhs
)
1357 if (FLOAT_TYPE_P (TREE_TYPE (gimple_cond_rhs (cond
)))
1358 ? real_zerop (gimple_cond_rhs (cond
))
1359 : integer_zerop (gimple_cond_rhs (cond
)))
1364 /* We need to know which is the true edge and which is the false
1365 edge so that we know if have abs or negative abs. */
1366 extract_true_false_edges_from_block (cond_bb
, &true_edge
, &false_edge
);
1368 /* For GT_EXPR/GE_EXPR, if the true edge goes to OTHER_BLOCK, then we
1369 will need to negate the result. Similarly for LT_EXPR/LE_EXPR if
1370 the false edge goes to OTHER_BLOCK. */
1371 if (cond_code
== GT_EXPR
|| cond_code
== GE_EXPR
)
1376 if (e
->dest
== middle_bb
)
1381 result
= duplicate_ssa_name (result
, NULL
);
1384 lhs
= make_ssa_name (TREE_TYPE (result
));
1388 /* Build the modify expression with abs expression. */
1389 new_stmt
= gimple_build_assign (lhs
, ABS_EXPR
, rhs
);
1391 gsi
= gsi_last_bb (cond_bb
);
1392 gsi_insert_before (&gsi
, new_stmt
, GSI_NEW_STMT
);
1396 /* Get the right GSI. We want to insert after the recently
1397 added ABS_EXPR statement (which we know is the first statement
1399 new_stmt
= gimple_build_assign (result
, NEGATE_EXPR
, lhs
);
1401 gsi_insert_after (&gsi
, new_stmt
, GSI_NEW_STMT
);
1404 replace_phi_edge_with_variable (cond_bb
, e1
, phi
, result
);
1406 /* Note that we optimized this PHI. */
1410 /* Auxiliary functions to determine the set of memory accesses which
1411 can't trap because they are preceded by accesses to the same memory
1412 portion. We do that for MEM_REFs, so we only need to track
1413 the SSA_NAME of the pointer indirectly referenced. The algorithm
1414 simply is a walk over all instructions in dominator order. When
1415 we see an MEM_REF we determine if we've already seen a same
1416 ref anywhere up to the root of the dominator tree. If we do the
1417 current access can't trap. If we don't see any dominating access
1418 the current access might trap, but might also make later accesses
1419 non-trapping, so we remember it. We need to be careful with loads
1420 or stores, for instance a load might not trap, while a store would,
1421 so if we see a dominating read access this doesn't mean that a later
1422 write access would not trap. Hence we also need to differentiate the
1423 type of access(es) seen.
1425 ??? We currently are very conservative and assume that a load might
1426 trap even if a store doesn't (write-only memory). This probably is
1427 overly conservative. */
1429 /* A hash-table of SSA_NAMEs, and in which basic block an MEM_REF
1430 through it was seen, which would constitute a no-trap region for
1434 unsigned int ssa_name_ver
;
1437 HOST_WIDE_INT offset
, size
;
1441 /* Hashtable helpers. */
1443 struct ssa_names_hasher
: free_ptr_hash
<name_to_bb
>
1445 static inline hashval_t
hash (const name_to_bb
*);
1446 static inline bool equal (const name_to_bb
*, const name_to_bb
*);
1449 /* Used for quick clearing of the hash-table when we see calls.
1450 Hash entries with phase < nt_call_phase are invalid. */
1451 static unsigned int nt_call_phase
;
1453 /* The hash function. */
1456 ssa_names_hasher::hash (const name_to_bb
*n
)
1458 return n
->ssa_name_ver
^ (((hashval_t
) n
->store
) << 31)
1459 ^ (n
->offset
<< 6) ^ (n
->size
<< 3);
1462 /* The equality function of *P1 and *P2. */
1465 ssa_names_hasher::equal (const name_to_bb
*n1
, const name_to_bb
*n2
)
1467 return n1
->ssa_name_ver
== n2
->ssa_name_ver
1468 && n1
->store
== n2
->store
1469 && n1
->offset
== n2
->offset
1470 && n1
->size
== n2
->size
;
1473 class nontrapping_dom_walker
: public dom_walker
1476 nontrapping_dom_walker (cdi_direction direction
, hash_set
<tree
> *ps
)
1477 : dom_walker (direction
), m_nontrapping (ps
), m_seen_ssa_names (128) {}
1479 virtual void before_dom_children (basic_block
);
1480 virtual void after_dom_children (basic_block
);
1484 /* We see the expression EXP in basic block BB. If it's an interesting
1485 expression (an MEM_REF through an SSA_NAME) possibly insert the
1486 expression into the set NONTRAP or the hash table of seen expressions.
1487 STORE is true if this expression is on the LHS, otherwise it's on
1489 void add_or_mark_expr (basic_block
, tree
, bool);
1491 hash_set
<tree
> *m_nontrapping
;
1493 /* The hash table for remembering what we've seen. */
1494 hash_table
<ssa_names_hasher
> m_seen_ssa_names
;
1497 /* Called by walk_dominator_tree, when entering the block BB. */
1499 nontrapping_dom_walker::before_dom_children (basic_block bb
)
1503 gimple_stmt_iterator gsi
;
1505 /* If we haven't seen all our predecessors, clear the hash-table. */
1506 FOR_EACH_EDGE (e
, ei
, bb
->preds
)
1507 if ((((size_t)e
->src
->aux
) & 2) == 0)
1513 /* Mark this BB as being on the path to dominator root and as visited. */
1514 bb
->aux
= (void*)(1 | 2);
1516 /* And walk the statements in order. */
1517 for (gsi
= gsi_start_bb (bb
); !gsi_end_p (gsi
); gsi_next (&gsi
))
1519 gimple stmt
= gsi_stmt (gsi
);
1521 if (is_gimple_call (stmt
) && !nonfreeing_call_p (stmt
))
1523 else if (gimple_assign_single_p (stmt
) && !gimple_has_volatile_ops (stmt
))
1525 add_or_mark_expr (bb
, gimple_assign_lhs (stmt
), true);
1526 add_or_mark_expr (bb
, gimple_assign_rhs1 (stmt
), false);
1531 /* Called by walk_dominator_tree, when basic block BB is exited. */
1533 nontrapping_dom_walker::after_dom_children (basic_block bb
)
1535 /* This BB isn't on the path to dominator root anymore. */
1539 /* We see the expression EXP in basic block BB. If it's an interesting
1540 expression (an MEM_REF through an SSA_NAME) possibly insert the
1541 expression into the set NONTRAP or the hash table of seen expressions.
1542 STORE is true if this expression is on the LHS, otherwise it's on
1545 nontrapping_dom_walker::add_or_mark_expr (basic_block bb
, tree exp
, bool store
)
1549 if (TREE_CODE (exp
) == MEM_REF
1550 && TREE_CODE (TREE_OPERAND (exp
, 0)) == SSA_NAME
1551 && tree_fits_shwi_p (TREE_OPERAND (exp
, 1))
1552 && (size
= int_size_in_bytes (TREE_TYPE (exp
))) > 0)
1554 tree name
= TREE_OPERAND (exp
, 0);
1555 struct name_to_bb map
;
1557 struct name_to_bb
*n2bb
;
1558 basic_block found_bb
= 0;
1560 /* Try to find the last seen MEM_REF through the same
1561 SSA_NAME, which can trap. */
1562 map
.ssa_name_ver
= SSA_NAME_VERSION (name
);
1566 map
.offset
= tree_to_shwi (TREE_OPERAND (exp
, 1));
1569 slot
= m_seen_ssa_names
.find_slot (&map
, INSERT
);
1571 if (n2bb
&& n2bb
->phase
>= nt_call_phase
)
1572 found_bb
= n2bb
->bb
;
1574 /* If we've found a trapping MEM_REF, _and_ it dominates EXP
1575 (it's in a basic block on the path from us to the dominator root)
1576 then we can't trap. */
1577 if (found_bb
&& (((size_t)found_bb
->aux
) & 1) == 1)
1579 m_nontrapping
->add (exp
);
1583 /* EXP might trap, so insert it into the hash table. */
1586 n2bb
->phase
= nt_call_phase
;
1591 n2bb
= XNEW (struct name_to_bb
);
1592 n2bb
->ssa_name_ver
= SSA_NAME_VERSION (name
);
1593 n2bb
->phase
= nt_call_phase
;
1595 n2bb
->store
= store
;
1596 n2bb
->offset
= map
.offset
;
1604 /* This is the entry point of gathering non trapping memory accesses.
1605 It will do a dominator walk over the whole function, and it will
1606 make use of the bb->aux pointers. It returns a set of trees
1607 (the MEM_REFs itself) which can't trap. */
1608 static hash_set
<tree
> *
1609 get_non_trapping (void)
1612 hash_set
<tree
> *nontrap
= new hash_set
<tree
>;
1613 /* We're going to do a dominator walk, so ensure that we have
1614 dominance information. */
1615 calculate_dominance_info (CDI_DOMINATORS
);
1617 nontrapping_dom_walker (CDI_DOMINATORS
, nontrap
)
1618 .walk (cfun
->cfg
->x_entry_block_ptr
);
1620 clear_aux_for_blocks ();
1624 /* Do the main work of conditional store replacement. We already know
1625 that the recognized pattern looks like so:
1628 if (cond) goto MIDDLE_BB; else goto JOIN_BB (edge E1)
1631 fallthrough (edge E0)
1635 We check that MIDDLE_BB contains only one store, that that store
1636 doesn't trap (not via NOTRAP, but via checking if an access to the same
1637 memory location dominates us) and that the store has a "simple" RHS. */
1640 cond_store_replacement (basic_block middle_bb
, basic_block join_bb
,
1641 edge e0
, edge e1
, hash_set
<tree
> *nontrap
)
1643 gimple assign
= last_and_only_stmt (middle_bb
);
1644 tree lhs
, rhs
, name
, name2
;
1647 gimple_stmt_iterator gsi
;
1648 source_location locus
;
1650 /* Check if middle_bb contains of only one store. */
1652 || !gimple_assign_single_p (assign
)
1653 || gimple_has_volatile_ops (assign
))
1656 locus
= gimple_location (assign
);
1657 lhs
= gimple_assign_lhs (assign
);
1658 rhs
= gimple_assign_rhs1 (assign
);
1659 if (TREE_CODE (lhs
) != MEM_REF
1660 || TREE_CODE (TREE_OPERAND (lhs
, 0)) != SSA_NAME
1661 || !is_gimple_reg_type (TREE_TYPE (lhs
)))
1664 /* Prove that we can move the store down. We could also check
1665 TREE_THIS_NOTRAP here, but in that case we also could move stores,
1666 whose value is not available readily, which we want to avoid. */
1667 if (!nontrap
->contains (lhs
))
1670 /* Now we've checked the constraints, so do the transformation:
1671 1) Remove the single store. */
1672 gsi
= gsi_for_stmt (assign
);
1673 unlink_stmt_vdef (assign
);
1674 gsi_remove (&gsi
, true);
1675 release_defs (assign
);
1677 /* 2) Insert a load from the memory of the store to the temporary
1678 on the edge which did not contain the store. */
1679 lhs
= unshare_expr (lhs
);
1680 name
= make_temp_ssa_name (TREE_TYPE (lhs
), NULL
, "cstore");
1681 new_stmt
= gimple_build_assign (name
, lhs
);
1682 gimple_set_location (new_stmt
, locus
);
1683 gsi_insert_on_edge (e1
, new_stmt
);
1685 /* 3) Create a PHI node at the join block, with one argument
1686 holding the old RHS, and the other holding the temporary
1687 where we stored the old memory contents. */
1688 name2
= make_temp_ssa_name (TREE_TYPE (lhs
), NULL
, "cstore");
1689 newphi
= create_phi_node (name2
, join_bb
);
1690 add_phi_arg (newphi
, rhs
, e0
, locus
);
1691 add_phi_arg (newphi
, name
, e1
, locus
);
1693 lhs
= unshare_expr (lhs
);
1694 new_stmt
= gimple_build_assign (lhs
, PHI_RESULT (newphi
));
1696 /* 4) Insert that PHI node. */
1697 gsi
= gsi_after_labels (join_bb
);
1698 if (gsi_end_p (gsi
))
1700 gsi
= gsi_last_bb (join_bb
);
1701 gsi_insert_after (&gsi
, new_stmt
, GSI_NEW_STMT
);
1704 gsi_insert_before (&gsi
, new_stmt
, GSI_NEW_STMT
);
1709 /* Do the main work of conditional store replacement. */
1712 cond_if_else_store_replacement_1 (basic_block then_bb
, basic_block else_bb
,
1713 basic_block join_bb
, gimple then_assign
,
1716 tree lhs_base
, lhs
, then_rhs
, else_rhs
, name
;
1717 source_location then_locus
, else_locus
;
1718 gimple_stmt_iterator gsi
;
1722 if (then_assign
== NULL
1723 || !gimple_assign_single_p (then_assign
)
1724 || gimple_clobber_p (then_assign
)
1725 || gimple_has_volatile_ops (then_assign
)
1726 || else_assign
== NULL
1727 || !gimple_assign_single_p (else_assign
)
1728 || gimple_clobber_p (else_assign
)
1729 || gimple_has_volatile_ops (else_assign
))
1732 lhs
= gimple_assign_lhs (then_assign
);
1733 if (!is_gimple_reg_type (TREE_TYPE (lhs
))
1734 || !operand_equal_p (lhs
, gimple_assign_lhs (else_assign
), 0))
1737 lhs_base
= get_base_address (lhs
);
1738 if (lhs_base
== NULL_TREE
1739 || (!DECL_P (lhs_base
) && TREE_CODE (lhs_base
) != MEM_REF
))
1742 then_rhs
= gimple_assign_rhs1 (then_assign
);
1743 else_rhs
= gimple_assign_rhs1 (else_assign
);
1744 then_locus
= gimple_location (then_assign
);
1745 else_locus
= gimple_location (else_assign
);
1747 /* Now we've checked the constraints, so do the transformation:
1748 1) Remove the stores. */
1749 gsi
= gsi_for_stmt (then_assign
);
1750 unlink_stmt_vdef (then_assign
);
1751 gsi_remove (&gsi
, true);
1752 release_defs (then_assign
);
1754 gsi
= gsi_for_stmt (else_assign
);
1755 unlink_stmt_vdef (else_assign
);
1756 gsi_remove (&gsi
, true);
1757 release_defs (else_assign
);
1759 /* 2) Create a PHI node at the join block, with one argument
1760 holding the old RHS, and the other holding the temporary
1761 where we stored the old memory contents. */
1762 name
= make_temp_ssa_name (TREE_TYPE (lhs
), NULL
, "cstore");
1763 newphi
= create_phi_node (name
, join_bb
);
1764 add_phi_arg (newphi
, then_rhs
, EDGE_SUCC (then_bb
, 0), then_locus
);
1765 add_phi_arg (newphi
, else_rhs
, EDGE_SUCC (else_bb
, 0), else_locus
);
1767 new_stmt
= gimple_build_assign (lhs
, PHI_RESULT (newphi
));
1769 /* 3) Insert that PHI node. */
1770 gsi
= gsi_after_labels (join_bb
);
1771 if (gsi_end_p (gsi
))
1773 gsi
= gsi_last_bb (join_bb
);
1774 gsi_insert_after (&gsi
, new_stmt
, GSI_NEW_STMT
);
1777 gsi_insert_before (&gsi
, new_stmt
, GSI_NEW_STMT
);
1782 /* Conditional store replacement. We already know
1783 that the recognized pattern looks like so:
1786 if (cond) goto THEN_BB; else goto ELSE_BB (edge E1)
1796 fallthrough (edge E0)
1800 We check that it is safe to sink the store to JOIN_BB by verifying that
1801 there are no read-after-write or write-after-write dependencies in
1802 THEN_BB and ELSE_BB. */
1805 cond_if_else_store_replacement (basic_block then_bb
, basic_block else_bb
,
1806 basic_block join_bb
)
1808 gimple then_assign
= last_and_only_stmt (then_bb
);
1809 gimple else_assign
= last_and_only_stmt (else_bb
);
1810 vec
<data_reference_p
> then_datarefs
, else_datarefs
;
1811 vec
<ddr_p
> then_ddrs
, else_ddrs
;
1812 gimple then_store
, else_store
;
1813 bool found
, ok
= false, res
;
1814 struct data_dependence_relation
*ddr
;
1815 data_reference_p then_dr
, else_dr
;
1817 tree then_lhs
, else_lhs
;
1818 basic_block blocks
[3];
1820 if (MAX_STORES_TO_SINK
== 0)
1823 /* Handle the case with single statement in THEN_BB and ELSE_BB. */
1824 if (then_assign
&& else_assign
)
1825 return cond_if_else_store_replacement_1 (then_bb
, else_bb
, join_bb
,
1826 then_assign
, else_assign
);
1828 /* Find data references. */
1829 then_datarefs
.create (1);
1830 else_datarefs
.create (1);
1831 if ((find_data_references_in_bb (NULL
, then_bb
, &then_datarefs
)
1833 || !then_datarefs
.length ()
1834 || (find_data_references_in_bb (NULL
, else_bb
, &else_datarefs
)
1836 || !else_datarefs
.length ())
1838 free_data_refs (then_datarefs
);
1839 free_data_refs (else_datarefs
);
1843 /* Find pairs of stores with equal LHS. */
1844 auto_vec
<gimple
, 1> then_stores
, else_stores
;
1845 FOR_EACH_VEC_ELT (then_datarefs
, i
, then_dr
)
1847 if (DR_IS_READ (then_dr
))
1850 then_store
= DR_STMT (then_dr
);
1851 then_lhs
= gimple_get_lhs (then_store
);
1852 if (then_lhs
== NULL_TREE
)
1856 FOR_EACH_VEC_ELT (else_datarefs
, j
, else_dr
)
1858 if (DR_IS_READ (else_dr
))
1861 else_store
= DR_STMT (else_dr
);
1862 else_lhs
= gimple_get_lhs (else_store
);
1863 if (else_lhs
== NULL_TREE
)
1866 if (operand_equal_p (then_lhs
, else_lhs
, 0))
1876 then_stores
.safe_push (then_store
);
1877 else_stores
.safe_push (else_store
);
1880 /* No pairs of stores found. */
1881 if (!then_stores
.length ()
1882 || then_stores
.length () > (unsigned) MAX_STORES_TO_SINK
)
1884 free_data_refs (then_datarefs
);
1885 free_data_refs (else_datarefs
);
1889 /* Compute and check data dependencies in both basic blocks. */
1890 then_ddrs
.create (1);
1891 else_ddrs
.create (1);
1892 if (!compute_all_dependences (then_datarefs
, &then_ddrs
,
1894 || !compute_all_dependences (else_datarefs
, &else_ddrs
,
1897 free_dependence_relations (then_ddrs
);
1898 free_dependence_relations (else_ddrs
);
1899 free_data_refs (then_datarefs
);
1900 free_data_refs (else_datarefs
);
1903 blocks
[0] = then_bb
;
1904 blocks
[1] = else_bb
;
1905 blocks
[2] = join_bb
;
1906 renumber_gimple_stmt_uids_in_blocks (blocks
, 3);
1908 /* Check that there are no read-after-write or write-after-write dependencies
1910 FOR_EACH_VEC_ELT (then_ddrs
, i
, ddr
)
1912 struct data_reference
*dra
= DDR_A (ddr
);
1913 struct data_reference
*drb
= DDR_B (ddr
);
1915 if (DDR_ARE_DEPENDENT (ddr
) != chrec_known
1916 && ((DR_IS_READ (dra
) && DR_IS_WRITE (drb
)
1917 && gimple_uid (DR_STMT (dra
)) > gimple_uid (DR_STMT (drb
)))
1918 || (DR_IS_READ (drb
) && DR_IS_WRITE (dra
)
1919 && gimple_uid (DR_STMT (drb
)) > gimple_uid (DR_STMT (dra
)))
1920 || (DR_IS_WRITE (dra
) && DR_IS_WRITE (drb
))))
1922 free_dependence_relations (then_ddrs
);
1923 free_dependence_relations (else_ddrs
);
1924 free_data_refs (then_datarefs
);
1925 free_data_refs (else_datarefs
);
1930 /* Check that there are no read-after-write or write-after-write dependencies
1932 FOR_EACH_VEC_ELT (else_ddrs
, i
, ddr
)
1934 struct data_reference
*dra
= DDR_A (ddr
);
1935 struct data_reference
*drb
= DDR_B (ddr
);
1937 if (DDR_ARE_DEPENDENT (ddr
) != chrec_known
1938 && ((DR_IS_READ (dra
) && DR_IS_WRITE (drb
)
1939 && gimple_uid (DR_STMT (dra
)) > gimple_uid (DR_STMT (drb
)))
1940 || (DR_IS_READ (drb
) && DR_IS_WRITE (dra
)
1941 && gimple_uid (DR_STMT (drb
)) > gimple_uid (DR_STMT (dra
)))
1942 || (DR_IS_WRITE (dra
) && DR_IS_WRITE (drb
))))
1944 free_dependence_relations (then_ddrs
);
1945 free_dependence_relations (else_ddrs
);
1946 free_data_refs (then_datarefs
);
1947 free_data_refs (else_datarefs
);
1952 /* Sink stores with same LHS. */
1953 FOR_EACH_VEC_ELT (then_stores
, i
, then_store
)
1955 else_store
= else_stores
[i
];
1956 res
= cond_if_else_store_replacement_1 (then_bb
, else_bb
, join_bb
,
1957 then_store
, else_store
);
1961 free_dependence_relations (then_ddrs
);
1962 free_dependence_relations (else_ddrs
);
1963 free_data_refs (then_datarefs
);
1964 free_data_refs (else_datarefs
);
1969 /* Return TRUE if STMT has a VUSE whose corresponding VDEF is in BB. */
1972 local_mem_dependence (gimple stmt
, basic_block bb
)
1974 tree vuse
= gimple_vuse (stmt
);
1980 def
= SSA_NAME_DEF_STMT (vuse
);
1981 return (def
&& gimple_bb (def
) == bb
);
1984 /* Given a "diamond" control-flow pattern where BB0 tests a condition,
1985 BB1 and BB2 are "then" and "else" blocks dependent on this test,
1986 and BB3 rejoins control flow following BB1 and BB2, look for
1987 opportunities to hoist loads as follows. If BB3 contains a PHI of
1988 two loads, one each occurring in BB1 and BB2, and the loads are
1989 provably of adjacent fields in the same structure, then move both
1990 loads into BB0. Of course this can only be done if there are no
1991 dependencies preventing such motion.
1993 One of the hoisted loads will always be speculative, so the
1994 transformation is currently conservative:
1996 - The fields must be strictly adjacent.
1997 - The two fields must occupy a single memory block that is
1998 guaranteed to not cross a page boundary.
2000 The last is difficult to prove, as such memory blocks should be
2001 aligned on the minimum of the stack alignment boundary and the
2002 alignment guaranteed by heap allocation interfaces. Thus we rely
2003 on a parameter for the alignment value.
2005 Provided a good value is used for the last case, the first
2006 restriction could possibly be relaxed. */
2009 hoist_adjacent_loads (basic_block bb0
, basic_block bb1
,
2010 basic_block bb2
, basic_block bb3
)
2012 int param_align
= PARAM_VALUE (PARAM_L1_CACHE_LINE_SIZE
);
2013 unsigned param_align_bits
= (unsigned) (param_align
* BITS_PER_UNIT
);
2016 /* Walk the phis in bb3 looking for an opportunity. We are looking
2017 for phis of two SSA names, one each of which is defined in bb1 and
2019 for (gsi
= gsi_start_phis (bb3
); !gsi_end_p (gsi
); gsi_next (&gsi
))
2021 gphi
*phi_stmt
= gsi
.phi ();
2023 tree arg1
, arg2
, ref1
, ref2
, field1
, field2
;
2024 tree tree_offset1
, tree_offset2
, tree_size2
, next
;
2025 int offset1
, offset2
, size2
;
2027 gimple_stmt_iterator gsi2
;
2028 basic_block bb_for_def1
, bb_for_def2
;
2030 if (gimple_phi_num_args (phi_stmt
) != 2
2031 || virtual_operand_p (gimple_phi_result (phi_stmt
)))
2034 arg1
= gimple_phi_arg_def (phi_stmt
, 0);
2035 arg2
= gimple_phi_arg_def (phi_stmt
, 1);
2037 if (TREE_CODE (arg1
) != SSA_NAME
2038 || TREE_CODE (arg2
) != SSA_NAME
2039 || SSA_NAME_IS_DEFAULT_DEF (arg1
)
2040 || SSA_NAME_IS_DEFAULT_DEF (arg2
))
2043 def1
= SSA_NAME_DEF_STMT (arg1
);
2044 def2
= SSA_NAME_DEF_STMT (arg2
);
2046 if ((gimple_bb (def1
) != bb1
|| gimple_bb (def2
) != bb2
)
2047 && (gimple_bb (def2
) != bb1
|| gimple_bb (def1
) != bb2
))
2050 /* Check the mode of the arguments to be sure a conditional move
2051 can be generated for it. */
2052 if (optab_handler (movcc_optab
, TYPE_MODE (TREE_TYPE (arg1
)))
2053 == CODE_FOR_nothing
)
2056 /* Both statements must be assignments whose RHS is a COMPONENT_REF. */
2057 if (!gimple_assign_single_p (def1
)
2058 || !gimple_assign_single_p (def2
)
2059 || gimple_has_volatile_ops (def1
)
2060 || gimple_has_volatile_ops (def2
))
2063 ref1
= gimple_assign_rhs1 (def1
);
2064 ref2
= gimple_assign_rhs1 (def2
);
2066 if (TREE_CODE (ref1
) != COMPONENT_REF
2067 || TREE_CODE (ref2
) != COMPONENT_REF
)
2070 /* The zeroth operand of the two component references must be
2071 identical. It is not sufficient to compare get_base_address of
2072 the two references, because this could allow for different
2073 elements of the same array in the two trees. It is not safe to
2074 assume that the existence of one array element implies the
2075 existence of a different one. */
2076 if (!operand_equal_p (TREE_OPERAND (ref1
, 0), TREE_OPERAND (ref2
, 0), 0))
2079 field1
= TREE_OPERAND (ref1
, 1);
2080 field2
= TREE_OPERAND (ref2
, 1);
2082 /* Check for field adjacency, and ensure field1 comes first. */
2083 for (next
= DECL_CHAIN (field1
);
2084 next
&& TREE_CODE (next
) != FIELD_DECL
;
2085 next
= DECL_CHAIN (next
))
2090 for (next
= DECL_CHAIN (field2
);
2091 next
&& TREE_CODE (next
) != FIELD_DECL
;
2092 next
= DECL_CHAIN (next
))
2098 std::swap (field1
, field2
);
2099 std::swap (def1
, def2
);
2102 bb_for_def1
= gimple_bb (def1
);
2103 bb_for_def2
= gimple_bb (def2
);
2105 /* Check for proper alignment of the first field. */
2106 tree_offset1
= bit_position (field1
);
2107 tree_offset2
= bit_position (field2
);
2108 tree_size2
= DECL_SIZE (field2
);
2110 if (!tree_fits_uhwi_p (tree_offset1
)
2111 || !tree_fits_uhwi_p (tree_offset2
)
2112 || !tree_fits_uhwi_p (tree_size2
))
2115 offset1
= tree_to_uhwi (tree_offset1
);
2116 offset2
= tree_to_uhwi (tree_offset2
);
2117 size2
= tree_to_uhwi (tree_size2
);
2118 align1
= DECL_ALIGN (field1
) % param_align_bits
;
2120 if (offset1
% BITS_PER_UNIT
!= 0)
2123 /* For profitability, the two field references should fit within
2124 a single cache line. */
2125 if (align1
+ offset2
- offset1
+ size2
> param_align_bits
)
2128 /* The two expressions cannot be dependent upon vdefs defined
2130 if (local_mem_dependence (def1
, bb_for_def1
)
2131 || local_mem_dependence (def2
, bb_for_def2
))
2134 /* The conditions are satisfied; hoist the loads from bb1 and bb2 into
2135 bb0. We hoist the first one first so that a cache miss is handled
2136 efficiently regardless of hardware cache-fill policy. */
2137 gsi2
= gsi_for_stmt (def1
);
2138 gsi_move_to_bb_end (&gsi2
, bb0
);
2139 gsi2
= gsi_for_stmt (def2
);
2140 gsi_move_to_bb_end (&gsi2
, bb0
);
2142 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
2145 "\nHoisting adjacent loads from %d and %d into %d: \n",
2146 bb_for_def1
->index
, bb_for_def2
->index
, bb0
->index
);
2147 print_gimple_stmt (dump_file
, def1
, 0, TDF_VOPS
|TDF_MEMSYMS
);
2148 print_gimple_stmt (dump_file
, def2
, 0, TDF_VOPS
|TDF_MEMSYMS
);
2153 /* Determine whether we should attempt to hoist adjacent loads out of
2154 diamond patterns in pass_phiopt. Always hoist loads if
2155 -fhoist-adjacent-loads is specified and the target machine has
2156 both a conditional move instruction and a defined cache line size. */
2159 gate_hoist_loads (void)
2161 return (flag_hoist_adjacent_loads
== 1
2162 && PARAM_VALUE (PARAM_L1_CACHE_LINE_SIZE
)
2163 && HAVE_conditional_move
);
2166 /* This pass tries to replaces an if-then-else block with an
2167 assignment. We have four kinds of transformations. Some of these
2168 transformations are also performed by the ifcvt RTL optimizer.
2170 Conditional Replacement
2171 -----------------------
2173 This transformation, implemented in conditional_replacement,
2177 if (cond) goto bb2; else goto bb1;
2180 x = PHI <0 (bb1), 1 (bb0), ...>;
2188 x = PHI <x' (bb0), ...>;
2190 We remove bb1 as it becomes unreachable. This occurs often due to
2191 gimplification of conditionals.
2196 This transformation, implemented in value_replacement, replaces
2199 if (a != b) goto bb2; else goto bb1;
2202 x = PHI <a (bb1), b (bb0), ...>;
2208 x = PHI <b (bb0), ...>;
2210 This opportunity can sometimes occur as a result of other
2214 Another case caught by value replacement looks like this:
2220 if (t3 != 0) goto bb1; else goto bb2;
2236 This transformation, implemented in abs_replacement, replaces
2239 if (a >= 0) goto bb2; else goto bb1;
2243 x = PHI <x (bb1), a (bb0), ...>;
2250 x = PHI <x' (bb0), ...>;
2255 This transformation, minmax_replacement replaces
2258 if (a <= b) goto bb2; else goto bb1;
2261 x = PHI <b (bb1), a (bb0), ...>;
2266 x' = MIN_EXPR (a, b)
2268 x = PHI <x' (bb0), ...>;
2270 A similar transformation is done for MAX_EXPR.
2273 This pass also performs a fifth transformation of a slightly different
2276 Factor conversion in COND_EXPR
2277 ------------------------------
2279 This transformation factors the conversion out of COND_EXPR with
2280 factor_out_conditional_conversion.
2283 if (a <= CST) goto <bb 3>; else goto <bb 4>;
2287 tmp = PHI <tmp, CST>
2290 if (a <= CST) goto <bb 3>; else goto <bb 4>;
2296 Adjacent Load Hoisting
2297 ----------------------
2299 This transformation replaces
2302 if (...) goto bb2; else goto bb1;
2304 x1 = (<expr>).field1;
2307 x2 = (<expr>).field2;
2314 x1 = (<expr>).field1;
2315 x2 = (<expr>).field2;
2316 if (...) goto bb2; else goto bb1;
2323 The purpose of this transformation is to enable generation of conditional
2324 move instructions such as Intel CMOVE or PowerPC ISEL. Because one of
2325 the loads is speculative, the transformation is restricted to very
2326 specific cases to avoid introducing a page fault. We are looking for
2334 where left and right are typically adjacent pointers in a tree structure. */
2338 const pass_data pass_data_phiopt
=
2340 GIMPLE_PASS
, /* type */
2341 "phiopt", /* name */
2342 OPTGROUP_NONE
, /* optinfo_flags */
2343 TV_TREE_PHIOPT
, /* tv_id */
2344 ( PROP_cfg
| PROP_ssa
), /* properties_required */
2345 0, /* properties_provided */
2346 0, /* properties_destroyed */
2347 0, /* todo_flags_start */
2348 0, /* todo_flags_finish */
2351 class pass_phiopt
: public gimple_opt_pass
2354 pass_phiopt (gcc::context
*ctxt
)
2355 : gimple_opt_pass (pass_data_phiopt
, ctxt
)
2358 /* opt_pass methods: */
2359 opt_pass
* clone () { return new pass_phiopt (m_ctxt
); }
2360 virtual bool gate (function
*) { return flag_ssa_phiopt
; }
2361 virtual unsigned int execute (function
*)
2363 return tree_ssa_phiopt_worker (false, gate_hoist_loads ());
2366 }; // class pass_phiopt
2371 make_pass_phiopt (gcc::context
*ctxt
)
2373 return new pass_phiopt (ctxt
);
2378 const pass_data pass_data_cselim
=
2380 GIMPLE_PASS
, /* type */
2381 "cselim", /* name */
2382 OPTGROUP_NONE
, /* optinfo_flags */
2383 TV_TREE_PHIOPT
, /* tv_id */
2384 ( PROP_cfg
| PROP_ssa
), /* properties_required */
2385 0, /* properties_provided */
2386 0, /* properties_destroyed */
2387 0, /* todo_flags_start */
2388 0, /* todo_flags_finish */
2391 class pass_cselim
: public gimple_opt_pass
2394 pass_cselim (gcc::context
*ctxt
)
2395 : gimple_opt_pass (pass_data_cselim
, ctxt
)
2398 /* opt_pass methods: */
2399 virtual bool gate (function
*) { return flag_tree_cselim
; }
2400 virtual unsigned int execute (function
*) { return tree_ssa_cs_elim (); }
2402 }; // class pass_cselim
2407 make_pass_cselim (gcc::context
*ctxt
)
2409 return new pass_cselim (ctxt
);