1 /* Optimization of PHI nodes by converting them into straightline code.
2 Copyright (C) 2004-2017 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"
24 #include "insn-codes.h"
29 #include "tree-pass.h"
31 #include "optabs-tree.h"
32 #include "insn-config.h"
33 #include "gimple-pretty-print.h"
34 #include "fold-const.h"
35 #include "stor-layout.h"
38 #include "gimple-iterator.h"
39 #include "gimplify-me.h"
44 #include "tree-data-ref.h"
45 #include "tree-scalar-evolution.h"
46 #include "tree-inline.h"
49 static unsigned int tree_ssa_phiopt_worker (bool, bool);
50 static bool conditional_replacement (basic_block
, basic_block
,
51 edge
, edge
, gphi
*, tree
, tree
);
52 static gphi
*factor_out_conditional_conversion (edge
, edge
, gphi
*, tree
, tree
,
54 static int value_replacement (basic_block
, basic_block
,
55 edge
, edge
, gimple
*, tree
, tree
);
56 static bool minmax_replacement (basic_block
, basic_block
,
57 edge
, edge
, gimple
*, tree
, tree
);
58 static bool abs_replacement (basic_block
, basic_block
,
59 edge
, edge
, gimple
*, tree
, tree
);
60 static bool cond_store_replacement (basic_block
, basic_block
, edge
, edge
,
62 static bool cond_if_else_store_replacement (basic_block
, basic_block
, basic_block
);
63 static hash_set
<tree
> * get_non_trapping ();
64 static void replace_phi_edge_with_variable (basic_block
, edge
, gimple
*, tree
);
65 static void hoist_adjacent_loads (basic_block
, basic_block
,
66 basic_block
, basic_block
);
67 static bool gate_hoist_loads (void);
69 /* This pass tries to transform conditional stores into unconditional
70 ones, enabling further simplifications with the simpler then and else
71 blocks. In particular it replaces this:
74 if (cond) goto bb2; else goto bb1;
82 if (cond) goto bb1; else goto bb2;
86 condtmp = PHI <RHS, condtmp'>
89 This transformation can only be done under several constraints,
90 documented below. It also replaces:
93 if (cond) goto bb2; else goto bb1;
104 if (cond) goto bb3; else goto bb1;
107 condtmp = PHI <RHS1, RHS2>
111 tree_ssa_cs_elim (void)
114 /* ??? We are not interested in loop related info, but the following
115 will create it, ICEing as we didn't init loops with pre-headers.
116 An interfacing issue of find_data_references_in_bb. */
117 loop_optimizer_init (LOOPS_NORMAL
);
119 todo
= tree_ssa_phiopt_worker (true, false);
121 loop_optimizer_finalize ();
125 /* Return the singleton PHI in the SEQ of PHIs for edges E0 and E1. */
128 single_non_singleton_phi_for_edges (gimple_seq seq
, edge e0
, edge e1
)
130 gimple_stmt_iterator i
;
132 if (gimple_seq_singleton_p (seq
))
133 return as_a
<gphi
*> (gsi_stmt (gsi_start (seq
)));
134 for (i
= gsi_start (seq
); !gsi_end_p (i
); gsi_next (&i
))
136 gphi
*p
= as_a
<gphi
*> (gsi_stmt (i
));
137 /* If the PHI arguments are equal then we can skip this PHI. */
138 if (operand_equal_for_phi_arg_p (gimple_phi_arg_def (p
, e0
->dest_idx
),
139 gimple_phi_arg_def (p
, e1
->dest_idx
)))
142 /* If we already have a PHI that has the two edge arguments are
143 different, then return it is not a singleton for these PHIs. */
152 /* The core routine of conditional store replacement and normal
153 phi optimizations. Both share much of the infrastructure in how
154 to match applicable basic block patterns. DO_STORE_ELIM is true
155 when we want to do conditional store replacement, false otherwise.
156 DO_HOIST_LOADS is true when we want to hoist adjacent loads out
157 of diamond control flow patterns, false otherwise. */
159 tree_ssa_phiopt_worker (bool do_store_elim
, bool do_hoist_loads
)
162 basic_block
*bb_order
;
164 bool cfgchanged
= false;
165 hash_set
<tree
> *nontrap
= 0;
168 /* Calculate the set of non-trapping memory accesses. */
169 nontrap
= get_non_trapping ();
171 /* Search every basic block for COND_EXPR we may be able to optimize.
173 We walk the blocks in order that guarantees that a block with
174 a single predecessor is processed before the predecessor.
175 This ensures that we collapse inner ifs before visiting the
176 outer ones, and also that we do not try to visit a removed
178 bb_order
= single_pred_before_succ_order ();
179 n
= n_basic_blocks_for_fn (cfun
) - NUM_FIXED_BLOCKS
;
181 for (i
= 0; i
< n
; i
++)
185 basic_block bb1
, bb2
;
191 cond_stmt
= last_stmt (bb
);
192 /* Check to see if the last statement is a GIMPLE_COND. */
194 || gimple_code (cond_stmt
) != GIMPLE_COND
)
197 e1
= EDGE_SUCC (bb
, 0);
199 e2
= EDGE_SUCC (bb
, 1);
202 /* We cannot do the optimization on abnormal edges. */
203 if ((e1
->flags
& EDGE_ABNORMAL
) != 0
204 || (e2
->flags
& EDGE_ABNORMAL
) != 0)
207 /* If either bb1's succ or bb2 or bb2's succ is non NULL. */
208 if (EDGE_COUNT (bb1
->succs
) == 0
210 || EDGE_COUNT (bb2
->succs
) == 0)
213 /* Find the bb which is the fall through to the other. */
214 if (EDGE_SUCC (bb1
, 0)->dest
== bb2
)
216 else if (EDGE_SUCC (bb2
, 0)->dest
== bb1
)
218 std::swap (bb1
, bb2
);
221 else if (do_store_elim
222 && EDGE_SUCC (bb1
, 0)->dest
== EDGE_SUCC (bb2
, 0)->dest
)
224 basic_block bb3
= EDGE_SUCC (bb1
, 0)->dest
;
226 if (!single_succ_p (bb1
)
227 || (EDGE_SUCC (bb1
, 0)->flags
& EDGE_FALLTHRU
) == 0
228 || !single_succ_p (bb2
)
229 || (EDGE_SUCC (bb2
, 0)->flags
& EDGE_FALLTHRU
) == 0
230 || EDGE_COUNT (bb3
->preds
) != 2)
232 if (cond_if_else_store_replacement (bb1
, bb2
, bb3
))
236 else if (do_hoist_loads
237 && EDGE_SUCC (bb1
, 0)->dest
== EDGE_SUCC (bb2
, 0)->dest
)
239 basic_block bb3
= EDGE_SUCC (bb1
, 0)->dest
;
241 if (!FLOAT_TYPE_P (TREE_TYPE (gimple_cond_lhs (cond_stmt
)))
242 && single_succ_p (bb1
)
243 && single_succ_p (bb2
)
244 && single_pred_p (bb1
)
245 && single_pred_p (bb2
)
246 && EDGE_COUNT (bb
->succs
) == 2
247 && EDGE_COUNT (bb3
->preds
) == 2
248 /* If one edge or the other is dominant, a conditional move
249 is likely to perform worse than the well-predicted branch. */
250 && !predictable_edge_p (EDGE_SUCC (bb
, 0))
251 && !predictable_edge_p (EDGE_SUCC (bb
, 1)))
252 hoist_adjacent_loads (bb
, bb1
, bb2
, bb3
);
258 e1
= EDGE_SUCC (bb1
, 0);
260 /* Make sure that bb1 is just a fall through. */
261 if (!single_succ_p (bb1
)
262 || (e1
->flags
& EDGE_FALLTHRU
) == 0)
265 /* Also make sure that bb1 only have one predecessor and that it
267 if (!single_pred_p (bb1
)
268 || single_pred (bb1
) != bb
)
273 /* bb1 is the middle block, bb2 the join block, bb the split block,
274 e1 the fallthrough edge from bb1 to bb2. We can't do the
275 optimization if the join block has more than two predecessors. */
276 if (EDGE_COUNT (bb2
->preds
) > 2)
278 if (cond_store_replacement (bb1
, bb2
, e1
, e2
, nontrap
))
283 gimple_seq phis
= phi_nodes (bb2
);
284 gimple_stmt_iterator gsi
;
285 bool candorest
= true;
287 /* Value replacement can work with more than one PHI
288 so try that first. */
289 for (gsi
= gsi_start (phis
); !gsi_end_p (gsi
); gsi_next (&gsi
))
291 phi
= as_a
<gphi
*> (gsi_stmt (gsi
));
292 arg0
= gimple_phi_arg_def (phi
, e1
->dest_idx
);
293 arg1
= gimple_phi_arg_def (phi
, e2
->dest_idx
);
294 if (value_replacement (bb
, bb1
, e1
, e2
, phi
, arg0
, arg1
) == 2)
305 phi
= single_non_singleton_phi_for_edges (phis
, e1
, e2
);
309 arg0
= gimple_phi_arg_def (phi
, e1
->dest_idx
);
310 arg1
= gimple_phi_arg_def (phi
, e2
->dest_idx
);
312 /* Something is wrong if we cannot find the arguments in the PHI
314 gcc_assert (arg0
!= NULL_TREE
&& arg1
!= NULL_TREE
);
316 gphi
*newphi
= factor_out_conditional_conversion (e1
, e2
, phi
,
322 /* factor_out_conditional_conversion may create a new PHI in
323 BB2 and eliminate an existing PHI in BB2. Recompute values
324 that may be affected by that change. */
325 arg0
= gimple_phi_arg_def (phi
, e1
->dest_idx
);
326 arg1
= gimple_phi_arg_def (phi
, e2
->dest_idx
);
327 gcc_assert (arg0
!= NULL_TREE
&& arg1
!= NULL_TREE
);
330 /* Do the replacement of conditional if it can be done. */
331 if (conditional_replacement (bb
, bb1
, e1
, e2
, phi
, arg0
, arg1
))
333 else if (abs_replacement (bb
, bb1
, e1
, e2
, phi
, arg0
, arg1
))
335 else if (minmax_replacement (bb
, bb1
, e1
, e2
, phi
, arg0
, arg1
))
344 /* If the CFG has changed, we should cleanup the CFG. */
345 if (cfgchanged
&& do_store_elim
)
347 /* In cond-store replacement we have added some loads on edges
348 and new VOPS (as we moved the store, and created a load). */
349 gsi_commit_edge_inserts ();
350 return TODO_cleanup_cfg
| TODO_update_ssa_only_virtuals
;
353 return TODO_cleanup_cfg
;
357 /* Replace PHI node element whose edge is E in block BB with variable NEW.
358 Remove the edge from COND_BLOCK which does not lead to BB (COND_BLOCK
359 is known to have two edges, one of which must reach BB). */
362 replace_phi_edge_with_variable (basic_block cond_block
,
363 edge e
, gimple
*phi
, tree new_tree
)
365 basic_block bb
= gimple_bb (phi
);
366 basic_block block_to_remove
;
367 gimple_stmt_iterator gsi
;
369 /* Change the PHI argument to new. */
370 SET_USE (PHI_ARG_DEF_PTR (phi
, e
->dest_idx
), new_tree
);
372 /* Remove the empty basic block. */
373 if (EDGE_SUCC (cond_block
, 0)->dest
== bb
)
375 EDGE_SUCC (cond_block
, 0)->flags
|= EDGE_FALLTHRU
;
376 EDGE_SUCC (cond_block
, 0)->flags
&= ~(EDGE_TRUE_VALUE
| EDGE_FALSE_VALUE
);
377 EDGE_SUCC (cond_block
, 0)->probability
= profile_probability::always ();
379 block_to_remove
= EDGE_SUCC (cond_block
, 1)->dest
;
383 EDGE_SUCC (cond_block
, 1)->flags
|= EDGE_FALLTHRU
;
384 EDGE_SUCC (cond_block
, 1)->flags
385 &= ~(EDGE_TRUE_VALUE
| EDGE_FALSE_VALUE
);
386 EDGE_SUCC (cond_block
, 1)->probability
= profile_probability::always ();
388 block_to_remove
= EDGE_SUCC (cond_block
, 0)->dest
;
390 delete_basic_block (block_to_remove
);
392 /* Eliminate the COND_EXPR at the end of COND_BLOCK. */
393 gsi
= gsi_last_bb (cond_block
);
394 gsi_remove (&gsi
, true);
396 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
398 "COND_EXPR in block %d and PHI in block %d converted to straightline code.\n",
403 /* PR66726: Factor conversion out of COND_EXPR. If the arguments of the PHI
404 stmt are CONVERT_STMT, factor out the conversion and perform the conversion
405 to the result of PHI stmt. COND_STMT is the controlling predicate.
406 Return the newly-created PHI, if any. */
409 factor_out_conditional_conversion (edge e0
, edge e1
, gphi
*phi
,
410 tree arg0
, tree arg1
, gimple
*cond_stmt
)
412 gimple
*arg0_def_stmt
= NULL
, *arg1_def_stmt
= NULL
, *new_stmt
;
413 tree new_arg0
= NULL_TREE
, new_arg1
= NULL_TREE
;
416 gimple_stmt_iterator gsi
, gsi_for_def
;
417 source_location locus
= gimple_location (phi
);
418 enum tree_code convert_code
;
420 /* Handle only PHI statements with two arguments. TODO: If all
421 other arguments to PHI are INTEGER_CST or if their defining
422 statement have the same unary operation, we can handle more
423 than two arguments too. */
424 if (gimple_phi_num_args (phi
) != 2)
427 /* First canonicalize to simplify tests. */
428 if (TREE_CODE (arg0
) != SSA_NAME
)
430 std::swap (arg0
, arg1
);
434 if (TREE_CODE (arg0
) != SSA_NAME
435 || (TREE_CODE (arg1
) != SSA_NAME
436 && TREE_CODE (arg1
) != INTEGER_CST
))
439 /* Check if arg0 is an SSA_NAME and the stmt which defines arg0 is
441 arg0_def_stmt
= SSA_NAME_DEF_STMT (arg0
);
442 if (!gimple_assign_cast_p (arg0_def_stmt
))
445 /* Use the RHS as new_arg0. */
446 convert_code
= gimple_assign_rhs_code (arg0_def_stmt
);
447 new_arg0
= gimple_assign_rhs1 (arg0_def_stmt
);
448 if (convert_code
== VIEW_CONVERT_EXPR
)
450 new_arg0
= TREE_OPERAND (new_arg0
, 0);
451 if (!is_gimple_reg_type (TREE_TYPE (new_arg0
)))
455 if (TREE_CODE (arg1
) == SSA_NAME
)
457 /* Check if arg1 is an SSA_NAME and the stmt which defines arg1
459 arg1_def_stmt
= SSA_NAME_DEF_STMT (arg1
);
460 if (!is_gimple_assign (arg1_def_stmt
)
461 || gimple_assign_rhs_code (arg1_def_stmt
) != convert_code
)
464 /* Use the RHS as new_arg1. */
465 new_arg1
= gimple_assign_rhs1 (arg1_def_stmt
);
466 if (convert_code
== VIEW_CONVERT_EXPR
)
467 new_arg1
= TREE_OPERAND (new_arg1
, 0);
471 /* If arg1 is an INTEGER_CST, fold it to new type. */
472 if (INTEGRAL_TYPE_P (TREE_TYPE (new_arg0
))
473 && int_fits_type_p (arg1
, TREE_TYPE (new_arg0
)))
475 if (gimple_assign_cast_p (arg0_def_stmt
))
477 /* For the INTEGER_CST case, we are just moving the
478 conversion from one place to another, which can often
479 hurt as the conversion moves further away from the
480 statement that computes the value. So, perform this
481 only if new_arg0 is an operand of COND_STMT, or
482 if arg0_def_stmt is the only non-debug stmt in
483 its basic block, because then it is possible this
484 could enable further optimizations (minmax replacement
485 etc.). See PR71016. */
486 if (new_arg0
!= gimple_cond_lhs (cond_stmt
)
487 && new_arg0
!= gimple_cond_rhs (cond_stmt
)
488 && gimple_bb (arg0_def_stmt
) == e0
->src
)
490 gsi
= gsi_for_stmt (arg0_def_stmt
);
491 gsi_prev_nondebug (&gsi
);
492 if (!gsi_end_p (gsi
))
494 gsi
= gsi_for_stmt (arg0_def_stmt
);
495 gsi_next_nondebug (&gsi
);
496 if (!gsi_end_p (gsi
))
499 new_arg1
= fold_convert (TREE_TYPE (new_arg0
), arg1
);
508 /* If arg0/arg1 have > 1 use, then this transformation actually increases
509 the number of expressions evaluated at runtime. */
510 if (!has_single_use (arg0
)
511 || (arg1_def_stmt
&& !has_single_use (arg1
)))
514 /* If types of new_arg0 and new_arg1 are different bailout. */
515 if (!types_compatible_p (TREE_TYPE (new_arg0
), TREE_TYPE (new_arg1
)))
518 /* Create a new PHI stmt. */
519 result
= PHI_RESULT (phi
);
520 temp
= make_ssa_name (TREE_TYPE (new_arg0
), NULL
);
521 newphi
= create_phi_node (temp
, gimple_bb (phi
));
523 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
525 fprintf (dump_file
, "PHI ");
526 print_generic_expr (dump_file
, gimple_phi_result (phi
));
528 " changed to factor conversion out from COND_EXPR.\n");
529 fprintf (dump_file
, "New stmt with CAST that defines ");
530 print_generic_expr (dump_file
, result
);
531 fprintf (dump_file
, ".\n");
534 /* Remove the old cast(s) that has single use. */
535 gsi_for_def
= gsi_for_stmt (arg0_def_stmt
);
536 gsi_remove (&gsi_for_def
, true);
537 release_defs (arg0_def_stmt
);
541 gsi_for_def
= gsi_for_stmt (arg1_def_stmt
);
542 gsi_remove (&gsi_for_def
, true);
543 release_defs (arg1_def_stmt
);
546 add_phi_arg (newphi
, new_arg0
, e0
, locus
);
547 add_phi_arg (newphi
, new_arg1
, e1
, locus
);
549 /* Create the conversion stmt and insert it. */
550 if (convert_code
== VIEW_CONVERT_EXPR
)
551 temp
= fold_build1 (VIEW_CONVERT_EXPR
, TREE_TYPE (result
), temp
);
552 new_stmt
= gimple_build_assign (result
, convert_code
, temp
);
553 gsi
= gsi_after_labels (gimple_bb (phi
));
554 gsi_insert_before (&gsi
, new_stmt
, GSI_SAME_STMT
);
556 /* Remove the original PHI stmt. */
557 gsi
= gsi_for_stmt (phi
);
558 gsi_remove (&gsi
, true);
562 /* The function conditional_replacement does the main work of doing the
563 conditional replacement. Return true if the replacement is done.
564 Otherwise return false.
565 BB is the basic block where the replacement is going to be done on. ARG0
566 is argument 0 from PHI. Likewise for ARG1. */
569 conditional_replacement (basic_block cond_bb
, basic_block middle_bb
,
570 edge e0
, edge e1
, gphi
*phi
,
571 tree arg0
, tree arg1
)
577 gimple_stmt_iterator gsi
;
578 edge true_edge
, false_edge
;
579 tree new_var
, new_var2
;
582 /* FIXME: Gimplification of complex type is too hard for now. */
583 /* We aren't prepared to handle vectors either (and it is a question
584 if it would be worthwhile anyway). */
585 if (!(INTEGRAL_TYPE_P (TREE_TYPE (arg0
))
586 || POINTER_TYPE_P (TREE_TYPE (arg0
)))
587 || !(INTEGRAL_TYPE_P (TREE_TYPE (arg1
))
588 || POINTER_TYPE_P (TREE_TYPE (arg1
))))
591 /* The PHI arguments have the constants 0 and 1, or 0 and -1, then
592 convert it to the conditional. */
593 if ((integer_zerop (arg0
) && integer_onep (arg1
))
594 || (integer_zerop (arg1
) && integer_onep (arg0
)))
596 else if ((integer_zerop (arg0
) && integer_all_onesp (arg1
))
597 || (integer_zerop (arg1
) && integer_all_onesp (arg0
)))
602 if (!empty_block_p (middle_bb
))
605 /* At this point we know we have a GIMPLE_COND with two successors.
606 One successor is BB, the other successor is an empty block which
607 falls through into BB.
609 There is a single PHI node at the join point (BB) and its arguments
610 are constants (0, 1) or (0, -1).
612 So, given the condition COND, and the two PHI arguments, we can
613 rewrite this PHI into non-branching code:
615 dest = (COND) or dest = COND'
617 We use the condition as-is if the argument associated with the
618 true edge has the value one or the argument associated with the
619 false edge as the value zero. Note that those conditions are not
620 the same since only one of the outgoing edges from the GIMPLE_COND
621 will directly reach BB and thus be associated with an argument. */
623 stmt
= last_stmt (cond_bb
);
624 result
= PHI_RESULT (phi
);
626 /* To handle special cases like floating point comparison, it is easier and
627 less error-prone to build a tree and gimplify it on the fly though it is
629 cond
= fold_build2_loc (gimple_location (stmt
),
630 gimple_cond_code (stmt
), boolean_type_node
,
631 gimple_cond_lhs (stmt
), gimple_cond_rhs (stmt
));
633 /* We need to know which is the true edge and which is the false
634 edge so that we know when to invert the condition below. */
635 extract_true_false_edges_from_block (cond_bb
, &true_edge
, &false_edge
);
636 if ((e0
== true_edge
&& integer_zerop (arg0
))
637 || (e0
== false_edge
&& !integer_zerop (arg0
))
638 || (e1
== true_edge
&& integer_zerop (arg1
))
639 || (e1
== false_edge
&& !integer_zerop (arg1
)))
640 cond
= fold_build1_loc (gimple_location (stmt
),
641 TRUTH_NOT_EXPR
, TREE_TYPE (cond
), cond
);
645 cond
= fold_convert_loc (gimple_location (stmt
),
646 TREE_TYPE (result
), cond
);
647 cond
= fold_build1_loc (gimple_location (stmt
),
648 NEGATE_EXPR
, TREE_TYPE (cond
), cond
);
651 /* Insert our new statements at the end of conditional block before the
653 gsi
= gsi_for_stmt (stmt
);
654 new_var
= force_gimple_operand_gsi (&gsi
, cond
, true, NULL
, true,
657 if (!useless_type_conversion_p (TREE_TYPE (result
), TREE_TYPE (new_var
)))
659 source_location locus_0
, locus_1
;
661 new_var2
= make_ssa_name (TREE_TYPE (result
));
662 new_stmt
= gimple_build_assign (new_var2
, CONVERT_EXPR
, new_var
);
663 gsi_insert_before (&gsi
, new_stmt
, GSI_SAME_STMT
);
666 /* Set the locus to the first argument, unless is doesn't have one. */
667 locus_0
= gimple_phi_arg_location (phi
, 0);
668 locus_1
= gimple_phi_arg_location (phi
, 1);
669 if (locus_0
== UNKNOWN_LOCATION
)
671 gimple_set_location (new_stmt
, locus_0
);
674 replace_phi_edge_with_variable (cond_bb
, e1
, phi
, new_var
);
675 reset_flow_sensitive_info_in_bb (cond_bb
);
677 /* Note that we optimized this PHI. */
681 /* Update *ARG which is defined in STMT so that it contains the
682 computed value if that seems profitable. Return true if the
683 statement is made dead by that rewriting. */
686 jump_function_from_stmt (tree
*arg
, gimple
*stmt
)
688 enum tree_code code
= gimple_assign_rhs_code (stmt
);
689 if (code
== ADDR_EXPR
)
691 /* For arg = &p->i transform it to p, if possible. */
692 tree rhs1
= gimple_assign_rhs1 (stmt
);
693 HOST_WIDE_INT offset
;
694 tree tem
= get_addr_base_and_unit_offset (TREE_OPERAND (rhs1
, 0),
697 && TREE_CODE (tem
) == MEM_REF
698 && (mem_ref_offset (tem
) + offset
) == 0)
700 *arg
= TREE_OPERAND (tem
, 0);
704 /* TODO: Much like IPA-CP jump-functions we want to handle constant
705 additions symbolically here, and we'd need to update the comparison
706 code that compares the arg + cst tuples in our caller. For now the
707 code above exactly handles the VEC_BASE pattern from vec.h. */
711 /* RHS is a source argument in a BIT_AND_EXPR which feeds a conditional
712 of the form SSA_NAME NE 0.
714 If RHS is fed by a simple EQ_EXPR comparison of two values, see if
715 the two input values of the EQ_EXPR match arg0 and arg1.
717 If so update *code and return TRUE. Otherwise return FALSE. */
720 rhs_is_fed_for_value_replacement (const_tree arg0
, const_tree arg1
,
721 enum tree_code
*code
, const_tree rhs
)
723 /* Obviously if RHS is not an SSA_NAME, we can't look at the defining
725 if (TREE_CODE (rhs
) == SSA_NAME
)
727 gimple
*def1
= SSA_NAME_DEF_STMT (rhs
);
729 /* Verify the defining statement has an EQ_EXPR on the RHS. */
730 if (is_gimple_assign (def1
) && gimple_assign_rhs_code (def1
) == EQ_EXPR
)
732 /* Finally verify the source operands of the EQ_EXPR are equal
734 tree op0
= gimple_assign_rhs1 (def1
);
735 tree op1
= gimple_assign_rhs2 (def1
);
736 if ((operand_equal_for_phi_arg_p (arg0
, op0
)
737 && operand_equal_for_phi_arg_p (arg1
, op1
))
738 || (operand_equal_for_phi_arg_p (arg0
, op1
)
739 && operand_equal_for_phi_arg_p (arg1
, op0
)))
741 /* We will perform the optimization. */
742 *code
= gimple_assign_rhs_code (def1
);
750 /* Return TRUE if arg0/arg1 are equal to the rhs/lhs or lhs/rhs of COND.
752 Also return TRUE if arg0/arg1 are equal to the source arguments of a
753 an EQ comparison feeding a BIT_AND_EXPR which feeds COND.
755 Return FALSE otherwise. */
758 operand_equal_for_value_replacement (const_tree arg0
, const_tree arg1
,
759 enum tree_code
*code
, gimple
*cond
)
762 tree lhs
= gimple_cond_lhs (cond
);
763 tree rhs
= gimple_cond_rhs (cond
);
765 if ((operand_equal_for_phi_arg_p (arg0
, lhs
)
766 && operand_equal_for_phi_arg_p (arg1
, rhs
))
767 || (operand_equal_for_phi_arg_p (arg1
, lhs
)
768 && operand_equal_for_phi_arg_p (arg0
, rhs
)))
771 /* Now handle more complex case where we have an EQ comparison
772 which feeds a BIT_AND_EXPR which feeds COND.
774 First verify that COND is of the form SSA_NAME NE 0. */
775 if (*code
!= NE_EXPR
|| !integer_zerop (rhs
)
776 || TREE_CODE (lhs
) != SSA_NAME
)
779 /* Now ensure that SSA_NAME is set by a BIT_AND_EXPR. */
780 def
= SSA_NAME_DEF_STMT (lhs
);
781 if (!is_gimple_assign (def
) || gimple_assign_rhs_code (def
) != BIT_AND_EXPR
)
784 /* Now verify arg0/arg1 correspond to the source arguments of an
785 EQ comparison feeding the BIT_AND_EXPR. */
787 tree tmp
= gimple_assign_rhs1 (def
);
788 if (rhs_is_fed_for_value_replacement (arg0
, arg1
, code
, tmp
))
791 tmp
= gimple_assign_rhs2 (def
);
792 if (rhs_is_fed_for_value_replacement (arg0
, arg1
, code
, tmp
))
798 /* Returns true if ARG is a neutral element for operation CODE
799 on the RIGHT side. */
802 neutral_element_p (tree_code code
, tree arg
, bool right
)
809 return integer_zerop (arg
);
816 case POINTER_PLUS_EXPR
:
817 return right
&& integer_zerop (arg
);
820 return integer_onep (arg
);
827 return right
&& integer_onep (arg
);
830 return integer_all_onesp (arg
);
837 /* Returns true if ARG is an absorbing element for operation CODE. */
840 absorbing_element_p (tree_code code
, tree arg
, bool right
, tree rval
)
845 return integer_all_onesp (arg
);
849 return integer_zerop (arg
);
855 return !right
&& integer_zerop (arg
);
867 && integer_zerop (arg
)
868 && tree_single_nonzero_warnv_p (rval
, NULL
));
875 /* The function value_replacement does the main work of doing the value
876 replacement. Return non-zero if the replacement is done. Otherwise return
877 0. If we remove the middle basic block, return 2.
878 BB is the basic block where the replacement is going to be done on. ARG0
879 is argument 0 from the PHI. Likewise for ARG1. */
882 value_replacement (basic_block cond_bb
, basic_block middle_bb
,
883 edge e0
, edge e1
, gimple
*phi
,
884 tree arg0
, tree arg1
)
886 gimple_stmt_iterator gsi
;
888 edge true_edge
, false_edge
;
890 bool emtpy_or_with_defined_p
= true;
892 /* If the type says honor signed zeros we cannot do this
894 if (HONOR_SIGNED_ZEROS (arg1
))
897 /* If there is a statement in MIDDLE_BB that defines one of the PHI
898 arguments, then adjust arg0 or arg1. */
899 gsi
= gsi_start_nondebug_after_labels_bb (middle_bb
);
900 while (!gsi_end_p (gsi
))
902 gimple
*stmt
= gsi_stmt (gsi
);
904 gsi_next_nondebug (&gsi
);
905 if (!is_gimple_assign (stmt
))
907 emtpy_or_with_defined_p
= false;
910 /* Now try to adjust arg0 or arg1 according to the computation
912 lhs
= gimple_assign_lhs (stmt
);
914 && jump_function_from_stmt (&arg0
, stmt
))
916 && jump_function_from_stmt (&arg1
, stmt
)))
917 emtpy_or_with_defined_p
= false;
920 cond
= last_stmt (cond_bb
);
921 code
= gimple_cond_code (cond
);
923 /* This transformation is only valid for equality comparisons. */
924 if (code
!= NE_EXPR
&& code
!= EQ_EXPR
)
927 /* We need to know which is the true edge and which is the false
928 edge so that we know if have abs or negative abs. */
929 extract_true_false_edges_from_block (cond_bb
, &true_edge
, &false_edge
);
931 /* At this point we know we have a COND_EXPR with two successors.
932 One successor is BB, the other successor is an empty block which
933 falls through into BB.
935 The condition for the COND_EXPR is known to be NE_EXPR or EQ_EXPR.
937 There is a single PHI node at the join point (BB) with two arguments.
939 We now need to verify that the two arguments in the PHI node match
940 the two arguments to the equality comparison. */
942 if (operand_equal_for_value_replacement (arg0
, arg1
, &code
, cond
))
947 /* For NE_EXPR, we want to build an assignment result = arg where
948 arg is the PHI argument associated with the true edge. For
949 EQ_EXPR we want the PHI argument associated with the false edge. */
950 e
= (code
== NE_EXPR
? true_edge
: false_edge
);
952 /* Unfortunately, E may not reach BB (it may instead have gone to
953 OTHER_BLOCK). If that is the case, then we want the single outgoing
954 edge from OTHER_BLOCK which reaches BB and represents the desired
955 path from COND_BLOCK. */
956 if (e
->dest
== middle_bb
)
957 e
= single_succ_edge (e
->dest
);
959 /* Now we know the incoming edge to BB that has the argument for the
960 RHS of our new assignment statement. */
966 /* If the middle basic block was empty or is defining the
967 PHI arguments and this is a single phi where the args are different
968 for the edges e0 and e1 then we can remove the middle basic block. */
969 if (emtpy_or_with_defined_p
970 && single_non_singleton_phi_for_edges (phi_nodes (gimple_bb (phi
)),
973 replace_phi_edge_with_variable (cond_bb
, e1
, phi
, arg
);
974 /* Note that we optimized this PHI. */
979 /* Replace the PHI arguments with arg. */
980 SET_PHI_ARG_DEF (phi
, e0
->dest_idx
, arg
);
981 SET_PHI_ARG_DEF (phi
, e1
->dest_idx
, arg
);
982 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
984 fprintf (dump_file
, "PHI ");
985 print_generic_expr (dump_file
, gimple_phi_result (phi
));
986 fprintf (dump_file
, " reduced for COND_EXPR in block %d to ",
988 print_generic_expr (dump_file
, arg
);
989 fprintf (dump_file
, ".\n");
996 /* Now optimize (x != 0) ? x + y : y to just x + y. */
997 gsi
= gsi_last_nondebug_bb (middle_bb
);
1001 gimple
*assign
= gsi_stmt (gsi
);
1002 if (!is_gimple_assign (assign
)
1003 || gimple_assign_rhs_class (assign
) != GIMPLE_BINARY_RHS
1004 || (!INTEGRAL_TYPE_P (TREE_TYPE (arg0
))
1005 && !POINTER_TYPE_P (TREE_TYPE (arg0
))))
1008 /* Punt if there are (degenerate) PHIs in middle_bb, there should not be. */
1009 if (!gimple_seq_empty_p (phi_nodes (middle_bb
)))
1012 /* Allow up to 2 cheap preparation statements that prepare argument
1020 iftmp.0_6 = x_5(D) r<< _1;
1022 # iftmp.0_2 = PHI <iftmp.0_6(3), x_5(D)(2)>
1033 # _2 = PHI <x_5(D)(2), _6(3)> */
1034 gimple
*prep_stmt
[2] = { NULL
, NULL
};
1036 for (prep_cnt
= 0; ; prep_cnt
++)
1038 gsi_prev_nondebug (&gsi
);
1039 if (gsi_end_p (gsi
))
1042 gimple
*g
= gsi_stmt (gsi
);
1043 if (gimple_code (g
) == GIMPLE_LABEL
)
1046 if (prep_cnt
== 2 || !is_gimple_assign (g
))
1049 tree lhs
= gimple_assign_lhs (g
);
1050 tree rhs1
= gimple_assign_rhs1 (g
);
1051 use_operand_p use_p
;
1053 if (TREE_CODE (lhs
) != SSA_NAME
1054 || TREE_CODE (rhs1
) != SSA_NAME
1055 || !INTEGRAL_TYPE_P (TREE_TYPE (lhs
))
1056 || !INTEGRAL_TYPE_P (TREE_TYPE (rhs1
))
1057 || !single_imm_use (lhs
, &use_p
, &use_stmt
)
1058 || use_stmt
!= (prep_cnt
? prep_stmt
[prep_cnt
- 1] : assign
))
1060 switch (gimple_assign_rhs_code (g
))
1068 if (TREE_CODE (gimple_assign_rhs2 (g
)) != INTEGER_CST
)
1074 prep_stmt
[prep_cnt
] = g
;
1077 /* Only transform if it removes the condition. */
1078 if (!single_non_singleton_phi_for_edges (phi_nodes (gimple_bb (phi
)), e0
, e1
))
1081 /* Size-wise, this is always profitable. */
1082 if (optimize_bb_for_speed_p (cond_bb
)
1083 /* The special case is useless if it has a low probability. */
1084 && profile_status_for_fn (cfun
) != PROFILE_ABSENT
1085 && EDGE_PRED (middle_bb
, 0)->probability
< profile_probability::even ()
1086 /* If assign is cheap, there is no point avoiding it. */
1087 && estimate_num_insns (bb_seq (middle_bb
), &eni_time_weights
)
1088 >= 3 * estimate_num_insns (cond
, &eni_time_weights
))
1091 tree lhs
= gimple_assign_lhs (assign
);
1092 tree rhs1
= gimple_assign_rhs1 (assign
);
1093 tree rhs2
= gimple_assign_rhs2 (assign
);
1094 enum tree_code code_def
= gimple_assign_rhs_code (assign
);
1095 tree cond_lhs
= gimple_cond_lhs (cond
);
1096 tree cond_rhs
= gimple_cond_rhs (cond
);
1098 /* Propagate the cond_rhs constant through preparation stmts,
1099 make sure UB isn't invoked while doing that. */
1100 for (int i
= prep_cnt
- 1; i
>= 0; --i
)
1102 gimple
*g
= prep_stmt
[i
];
1103 tree grhs1
= gimple_assign_rhs1 (g
);
1104 if (!operand_equal_for_phi_arg_p (cond_lhs
, grhs1
))
1106 cond_lhs
= gimple_assign_lhs (g
);
1107 cond_rhs
= fold_convert (TREE_TYPE (grhs1
), cond_rhs
);
1108 if (TREE_CODE (cond_rhs
) != INTEGER_CST
1109 || TREE_OVERFLOW (cond_rhs
))
1111 if (gimple_assign_rhs_class (g
) == GIMPLE_BINARY_RHS
)
1113 cond_rhs
= int_const_binop (gimple_assign_rhs_code (g
), cond_rhs
,
1114 gimple_assign_rhs2 (g
));
1115 if (TREE_OVERFLOW (cond_rhs
))
1118 cond_rhs
= fold_convert (TREE_TYPE (cond_lhs
), cond_rhs
);
1119 if (TREE_CODE (cond_rhs
) != INTEGER_CST
1120 || TREE_OVERFLOW (cond_rhs
))
1124 if (((code
== NE_EXPR
&& e1
== false_edge
)
1125 || (code
== EQ_EXPR
&& e1
== true_edge
))
1128 && operand_equal_for_phi_arg_p (rhs2
, cond_lhs
)
1129 && neutral_element_p (code_def
, cond_rhs
, true))
1131 && operand_equal_for_phi_arg_p (rhs1
, cond_lhs
)
1132 && neutral_element_p (code_def
, cond_rhs
, false))
1133 || (operand_equal_for_phi_arg_p (arg1
, cond_rhs
)
1134 && ((operand_equal_for_phi_arg_p (rhs2
, cond_lhs
)
1135 && absorbing_element_p (code_def
, cond_rhs
, true, rhs2
))
1136 || (operand_equal_for_phi_arg_p (rhs1
, cond_lhs
)
1137 && absorbing_element_p (code_def
,
1138 cond_rhs
, false, rhs2
))))))
1140 gsi
= gsi_for_stmt (cond
);
1141 if (INTEGRAL_TYPE_P (TREE_TYPE (lhs
)))
1143 /* Moving ASSIGN might change VR of lhs, e.g. when moving u_6
1151 # RANGE [0, 4294967294]
1152 u_6 = n_5 + 4294967295;
1155 # u_3 = PHI <u_6(3), 4294967295(2)> */
1156 SSA_NAME_RANGE_INFO (lhs
) = NULL
;
1157 /* If available, we can use VR of phi result at least. */
1158 tree phires
= gimple_phi_result (phi
);
1159 struct range_info_def
*phires_range_info
1160 = SSA_NAME_RANGE_INFO (phires
);
1161 if (phires_range_info
)
1162 duplicate_ssa_name_range_info (lhs
, SSA_NAME_RANGE_TYPE (phires
),
1165 gimple_stmt_iterator gsi_from
;
1166 for (int i
= prep_cnt
- 1; i
>= 0; --i
)
1168 tree plhs
= gimple_assign_lhs (prep_stmt
[i
]);
1169 SSA_NAME_RANGE_INFO (plhs
) = NULL
;
1170 gsi_from
= gsi_for_stmt (prep_stmt
[i
]);
1171 gsi_move_before (&gsi_from
, &gsi
);
1173 gsi_from
= gsi_for_stmt (assign
);
1174 gsi_move_before (&gsi_from
, &gsi
);
1175 replace_phi_edge_with_variable (cond_bb
, e1
, phi
, lhs
);
1182 /* The function minmax_replacement does the main work of doing the minmax
1183 replacement. Return true if the replacement is done. Otherwise return
1185 BB is the basic block where the replacement is going to be done on. ARG0
1186 is argument 0 from the PHI. Likewise for ARG1. */
1189 minmax_replacement (basic_block cond_bb
, basic_block middle_bb
,
1190 edge e0
, edge e1
, gimple
*phi
,
1191 tree arg0
, tree arg1
)
1196 edge true_edge
, false_edge
;
1197 enum tree_code cmp
, minmax
, ass_code
;
1198 tree smaller
, alt_smaller
, larger
, alt_larger
, arg_true
, arg_false
;
1199 gimple_stmt_iterator gsi
, gsi_from
;
1201 type
= TREE_TYPE (PHI_RESULT (phi
));
1203 /* The optimization may be unsafe due to NaNs. */
1204 if (HONOR_NANS (type
) || HONOR_SIGNED_ZEROS (type
))
1207 cond
= as_a
<gcond
*> (last_stmt (cond_bb
));
1208 cmp
= gimple_cond_code (cond
);
1210 /* This transformation is only valid for order comparisons. Record which
1211 operand is smaller/larger if the result of the comparison is true. */
1212 alt_smaller
= NULL_TREE
;
1213 alt_larger
= NULL_TREE
;
1214 if (cmp
== LT_EXPR
|| cmp
== LE_EXPR
)
1216 smaller
= gimple_cond_lhs (cond
);
1217 larger
= gimple_cond_rhs (cond
);
1218 /* If we have smaller < CST it is equivalent to smaller <= CST-1.
1219 Likewise smaller <= CST is equivalent to smaller < CST+1. */
1220 if (TREE_CODE (larger
) == INTEGER_CST
)
1225 wide_int alt
= wi::sub (wi::to_wide (larger
), 1,
1226 TYPE_SIGN (TREE_TYPE (larger
)),
1229 alt_larger
= wide_int_to_tree (TREE_TYPE (larger
), alt
);
1234 wide_int alt
= wi::add (wi::to_wide (larger
), 1,
1235 TYPE_SIGN (TREE_TYPE (larger
)),
1238 alt_larger
= wide_int_to_tree (TREE_TYPE (larger
), alt
);
1242 else if (cmp
== GT_EXPR
|| cmp
== GE_EXPR
)
1244 smaller
= gimple_cond_rhs (cond
);
1245 larger
= gimple_cond_lhs (cond
);
1246 /* If we have larger > CST it is equivalent to larger >= CST+1.
1247 Likewise larger >= CST is equivalent to larger > CST-1. */
1248 if (TREE_CODE (smaller
) == INTEGER_CST
)
1253 wide_int alt
= wi::add (wi::to_wide (smaller
), 1,
1254 TYPE_SIGN (TREE_TYPE (smaller
)),
1257 alt_smaller
= wide_int_to_tree (TREE_TYPE (smaller
), alt
);
1262 wide_int alt
= wi::sub (wi::to_wide (smaller
), 1,
1263 TYPE_SIGN (TREE_TYPE (smaller
)),
1266 alt_smaller
= wide_int_to_tree (TREE_TYPE (smaller
), alt
);
1273 /* We need to know which is the true edge and which is the false
1274 edge so that we know if have abs or negative abs. */
1275 extract_true_false_edges_from_block (cond_bb
, &true_edge
, &false_edge
);
1277 /* Forward the edges over the middle basic block. */
1278 if (true_edge
->dest
== middle_bb
)
1279 true_edge
= EDGE_SUCC (true_edge
->dest
, 0);
1280 if (false_edge
->dest
== middle_bb
)
1281 false_edge
= EDGE_SUCC (false_edge
->dest
, 0);
1283 if (true_edge
== e0
)
1285 gcc_assert (false_edge
== e1
);
1291 gcc_assert (false_edge
== e0
);
1292 gcc_assert (true_edge
== e1
);
1297 if (empty_block_p (middle_bb
))
1299 if ((operand_equal_for_phi_arg_p (arg_true
, smaller
)
1301 && operand_equal_for_phi_arg_p (arg_true
, alt_smaller
)))
1302 && (operand_equal_for_phi_arg_p (arg_false
, larger
)
1304 && operand_equal_for_phi_arg_p (arg_true
, alt_larger
))))
1308 if (smaller < larger)
1314 else if ((operand_equal_for_phi_arg_p (arg_false
, smaller
)
1316 && operand_equal_for_phi_arg_p (arg_false
, alt_smaller
)))
1317 && (operand_equal_for_phi_arg_p (arg_true
, larger
)
1319 && operand_equal_for_phi_arg_p (arg_true
, alt_larger
))))
1326 /* Recognize the following case, assuming d <= u:
1332 This is equivalent to
1337 gimple
*assign
= last_and_only_stmt (middle_bb
);
1338 tree lhs
, op0
, op1
, bound
;
1341 || gimple_code (assign
) != GIMPLE_ASSIGN
)
1344 lhs
= gimple_assign_lhs (assign
);
1345 ass_code
= gimple_assign_rhs_code (assign
);
1346 if (ass_code
!= MAX_EXPR
&& ass_code
!= MIN_EXPR
)
1348 op0
= gimple_assign_rhs1 (assign
);
1349 op1
= gimple_assign_rhs2 (assign
);
1351 if (true_edge
->src
== middle_bb
)
1353 /* We got here if the condition is true, i.e., SMALLER < LARGER. */
1354 if (!operand_equal_for_phi_arg_p (lhs
, arg_true
))
1357 if (operand_equal_for_phi_arg_p (arg_false
, larger
)
1359 && operand_equal_for_phi_arg_p (arg_false
, alt_larger
)))
1363 if (smaller < larger)
1365 r' = MAX_EXPR (smaller, bound)
1367 r = PHI <r', larger> --> to be turned to MIN_EXPR. */
1368 if (ass_code
!= MAX_EXPR
)
1372 if (operand_equal_for_phi_arg_p (op0
, smaller
)
1374 && operand_equal_for_phi_arg_p (op0
, alt_smaller
)))
1376 else if (operand_equal_for_phi_arg_p (op1
, smaller
)
1378 && operand_equal_for_phi_arg_p (op1
, alt_smaller
)))
1383 /* We need BOUND <= LARGER. */
1384 if (!integer_nonzerop (fold_build2 (LE_EXPR
, boolean_type_node
,
1388 else if (operand_equal_for_phi_arg_p (arg_false
, smaller
)
1390 && operand_equal_for_phi_arg_p (arg_false
, alt_smaller
)))
1394 if (smaller < larger)
1396 r' = MIN_EXPR (larger, bound)
1398 r = PHI <r', smaller> --> to be turned to MAX_EXPR. */
1399 if (ass_code
!= MIN_EXPR
)
1403 if (operand_equal_for_phi_arg_p (op0
, larger
)
1405 && operand_equal_for_phi_arg_p (op0
, alt_larger
)))
1407 else if (operand_equal_for_phi_arg_p (op1
, larger
)
1409 && operand_equal_for_phi_arg_p (op1
, alt_larger
)))
1414 /* We need BOUND >= SMALLER. */
1415 if (!integer_nonzerop (fold_build2 (GE_EXPR
, boolean_type_node
,
1424 /* We got here if the condition is false, i.e., SMALLER > LARGER. */
1425 if (!operand_equal_for_phi_arg_p (lhs
, arg_false
))
1428 if (operand_equal_for_phi_arg_p (arg_true
, larger
)
1430 && operand_equal_for_phi_arg_p (arg_true
, alt_larger
)))
1434 if (smaller > larger)
1436 r' = MIN_EXPR (smaller, bound)
1438 r = PHI <r', larger> --> to be turned to MAX_EXPR. */
1439 if (ass_code
!= MIN_EXPR
)
1443 if (operand_equal_for_phi_arg_p (op0
, smaller
)
1445 && operand_equal_for_phi_arg_p (op0
, alt_smaller
)))
1447 else if (operand_equal_for_phi_arg_p (op1
, smaller
)
1449 && operand_equal_for_phi_arg_p (op1
, alt_smaller
)))
1454 /* We need BOUND >= LARGER. */
1455 if (!integer_nonzerop (fold_build2 (GE_EXPR
, boolean_type_node
,
1459 else if (operand_equal_for_phi_arg_p (arg_true
, smaller
)
1461 && operand_equal_for_phi_arg_p (arg_true
, alt_smaller
)))
1465 if (smaller > larger)
1467 r' = MAX_EXPR (larger, bound)
1469 r = PHI <r', smaller> --> to be turned to MIN_EXPR. */
1470 if (ass_code
!= MAX_EXPR
)
1474 if (operand_equal_for_phi_arg_p (op0
, larger
))
1476 else if (operand_equal_for_phi_arg_p (op1
, larger
))
1481 /* We need BOUND <= SMALLER. */
1482 if (!integer_nonzerop (fold_build2 (LE_EXPR
, boolean_type_node
,
1490 /* Move the statement from the middle block. */
1491 gsi
= gsi_last_bb (cond_bb
);
1492 gsi_from
= gsi_last_nondebug_bb (middle_bb
);
1493 gsi_move_before (&gsi_from
, &gsi
);
1496 /* Create an SSA var to hold the min/max result. If we're the only
1497 things setting the target PHI, then we can clone the PHI
1498 variable. Otherwise we must create a new one. */
1499 result
= PHI_RESULT (phi
);
1500 if (EDGE_COUNT (gimple_bb (phi
)->preds
) == 2)
1501 result
= duplicate_ssa_name (result
, NULL
);
1503 result
= make_ssa_name (TREE_TYPE (result
));
1505 /* Emit the statement to compute min/max. */
1506 new_stmt
= gimple_build_assign (result
, minmax
, arg0
, arg1
);
1507 gsi
= gsi_last_bb (cond_bb
);
1508 gsi_insert_before (&gsi
, new_stmt
, GSI_NEW_STMT
);
1510 replace_phi_edge_with_variable (cond_bb
, e1
, phi
, result
);
1511 reset_flow_sensitive_info_in_bb (cond_bb
);
1516 /* The function absolute_replacement does the main work of doing the absolute
1517 replacement. Return true if the replacement is done. Otherwise return
1519 bb is the basic block where the replacement is going to be done on. arg0
1520 is argument 0 from the phi. Likewise for arg1. */
1523 abs_replacement (basic_block cond_bb
, basic_block middle_bb
,
1524 edge e0 ATTRIBUTE_UNUSED
, edge e1
,
1525 gimple
*phi
, tree arg0
, tree arg1
)
1530 gimple_stmt_iterator gsi
;
1531 edge true_edge
, false_edge
;
1536 enum tree_code cond_code
;
1538 /* If the type says honor signed zeros we cannot do this
1540 if (HONOR_SIGNED_ZEROS (arg1
))
1543 /* OTHER_BLOCK must have only one executable statement which must have the
1544 form arg0 = -arg1 or arg1 = -arg0. */
1546 assign
= last_and_only_stmt (middle_bb
);
1547 /* If we did not find the proper negation assignment, then we can not
1552 /* If we got here, then we have found the only executable statement
1553 in OTHER_BLOCK. If it is anything other than arg = -arg1 or
1554 arg1 = -arg0, then we can not optimize. */
1555 if (gimple_code (assign
) != GIMPLE_ASSIGN
)
1558 lhs
= gimple_assign_lhs (assign
);
1560 if (gimple_assign_rhs_code (assign
) != NEGATE_EXPR
)
1563 rhs
= gimple_assign_rhs1 (assign
);
1565 /* The assignment has to be arg0 = -arg1 or arg1 = -arg0. */
1566 if (!(lhs
== arg0
&& rhs
== arg1
)
1567 && !(lhs
== arg1
&& rhs
== arg0
))
1570 cond
= last_stmt (cond_bb
);
1571 result
= PHI_RESULT (phi
);
1573 /* Only relationals comparing arg[01] against zero are interesting. */
1574 cond_code
= gimple_cond_code (cond
);
1575 if (cond_code
!= GT_EXPR
&& cond_code
!= GE_EXPR
1576 && cond_code
!= LT_EXPR
&& cond_code
!= LE_EXPR
)
1579 /* Make sure the conditional is arg[01] OP y. */
1580 if (gimple_cond_lhs (cond
) != rhs
)
1583 if (FLOAT_TYPE_P (TREE_TYPE (gimple_cond_rhs (cond
)))
1584 ? real_zerop (gimple_cond_rhs (cond
))
1585 : integer_zerop (gimple_cond_rhs (cond
)))
1590 /* We need to know which is the true edge and which is the false
1591 edge so that we know if have abs or negative abs. */
1592 extract_true_false_edges_from_block (cond_bb
, &true_edge
, &false_edge
);
1594 /* For GT_EXPR/GE_EXPR, if the true edge goes to OTHER_BLOCK, then we
1595 will need to negate the result. Similarly for LT_EXPR/LE_EXPR if
1596 the false edge goes to OTHER_BLOCK. */
1597 if (cond_code
== GT_EXPR
|| cond_code
== GE_EXPR
)
1602 if (e
->dest
== middle_bb
)
1607 /* If the code negates only iff positive then make sure to not
1608 introduce undefined behavior when negating or computing the absolute.
1609 ??? We could use range info if present to check for arg1 == INT_MIN. */
1611 && (ANY_INTEGRAL_TYPE_P (TREE_TYPE (arg1
))
1612 && ! TYPE_OVERFLOW_WRAPS (TREE_TYPE (arg1
))))
1615 result
= duplicate_ssa_name (result
, NULL
);
1618 lhs
= make_ssa_name (TREE_TYPE (result
));
1622 /* Build the modify expression with abs expression. */
1623 new_stmt
= gimple_build_assign (lhs
, ABS_EXPR
, rhs
);
1625 gsi
= gsi_last_bb (cond_bb
);
1626 gsi_insert_before (&gsi
, new_stmt
, GSI_NEW_STMT
);
1630 /* Get the right GSI. We want to insert after the recently
1631 added ABS_EXPR statement (which we know is the first statement
1633 new_stmt
= gimple_build_assign (result
, NEGATE_EXPR
, lhs
);
1635 gsi_insert_after (&gsi
, new_stmt
, GSI_NEW_STMT
);
1638 replace_phi_edge_with_variable (cond_bb
, e1
, phi
, result
);
1639 reset_flow_sensitive_info_in_bb (cond_bb
);
1641 /* Note that we optimized this PHI. */
1645 /* Auxiliary functions to determine the set of memory accesses which
1646 can't trap because they are preceded by accesses to the same memory
1647 portion. We do that for MEM_REFs, so we only need to track
1648 the SSA_NAME of the pointer indirectly referenced. The algorithm
1649 simply is a walk over all instructions in dominator order. When
1650 we see an MEM_REF we determine if we've already seen a same
1651 ref anywhere up to the root of the dominator tree. If we do the
1652 current access can't trap. If we don't see any dominating access
1653 the current access might trap, but might also make later accesses
1654 non-trapping, so we remember it. We need to be careful with loads
1655 or stores, for instance a load might not trap, while a store would,
1656 so if we see a dominating read access this doesn't mean that a later
1657 write access would not trap. Hence we also need to differentiate the
1658 type of access(es) seen.
1660 ??? We currently are very conservative and assume that a load might
1661 trap even if a store doesn't (write-only memory). This probably is
1662 overly conservative. */
1664 /* A hash-table of SSA_NAMEs, and in which basic block an MEM_REF
1665 through it was seen, which would constitute a no-trap region for
1669 unsigned int ssa_name_ver
;
1672 HOST_WIDE_INT offset
, size
;
1676 /* Hashtable helpers. */
1678 struct ssa_names_hasher
: free_ptr_hash
<name_to_bb
>
1680 static inline hashval_t
hash (const name_to_bb
*);
1681 static inline bool equal (const name_to_bb
*, const name_to_bb
*);
1684 /* Used for quick clearing of the hash-table when we see calls.
1685 Hash entries with phase < nt_call_phase are invalid. */
1686 static unsigned int nt_call_phase
;
1688 /* The hash function. */
1691 ssa_names_hasher::hash (const name_to_bb
*n
)
1693 return n
->ssa_name_ver
^ (((hashval_t
) n
->store
) << 31)
1694 ^ (n
->offset
<< 6) ^ (n
->size
<< 3);
1697 /* The equality function of *P1 and *P2. */
1700 ssa_names_hasher::equal (const name_to_bb
*n1
, const name_to_bb
*n2
)
1702 return n1
->ssa_name_ver
== n2
->ssa_name_ver
1703 && n1
->store
== n2
->store
1704 && n1
->offset
== n2
->offset
1705 && n1
->size
== n2
->size
;
1708 class nontrapping_dom_walker
: public dom_walker
1711 nontrapping_dom_walker (cdi_direction direction
, hash_set
<tree
> *ps
)
1712 : dom_walker (direction
), m_nontrapping (ps
), m_seen_ssa_names (128) {}
1714 virtual edge
before_dom_children (basic_block
);
1715 virtual void after_dom_children (basic_block
);
1719 /* We see the expression EXP in basic block BB. If it's an interesting
1720 expression (an MEM_REF through an SSA_NAME) possibly insert the
1721 expression into the set NONTRAP or the hash table of seen expressions.
1722 STORE is true if this expression is on the LHS, otherwise it's on
1724 void add_or_mark_expr (basic_block
, tree
, bool);
1726 hash_set
<tree
> *m_nontrapping
;
1728 /* The hash table for remembering what we've seen. */
1729 hash_table
<ssa_names_hasher
> m_seen_ssa_names
;
1732 /* Called by walk_dominator_tree, when entering the block BB. */
1734 nontrapping_dom_walker::before_dom_children (basic_block bb
)
1738 gimple_stmt_iterator gsi
;
1740 /* If we haven't seen all our predecessors, clear the hash-table. */
1741 FOR_EACH_EDGE (e
, ei
, bb
->preds
)
1742 if ((((size_t)e
->src
->aux
) & 2) == 0)
1748 /* Mark this BB as being on the path to dominator root and as visited. */
1749 bb
->aux
= (void*)(1 | 2);
1751 /* And walk the statements in order. */
1752 for (gsi
= gsi_start_bb (bb
); !gsi_end_p (gsi
); gsi_next (&gsi
))
1754 gimple
*stmt
= gsi_stmt (gsi
);
1756 if ((gimple_code (stmt
) == GIMPLE_ASM
&& gimple_vdef (stmt
))
1757 || (is_gimple_call (stmt
)
1758 && (!nonfreeing_call_p (stmt
) || !nonbarrier_call_p (stmt
))))
1760 else if (gimple_assign_single_p (stmt
) && !gimple_has_volatile_ops (stmt
))
1762 add_or_mark_expr (bb
, gimple_assign_lhs (stmt
), true);
1763 add_or_mark_expr (bb
, gimple_assign_rhs1 (stmt
), false);
1769 /* Called by walk_dominator_tree, when basic block BB is exited. */
1771 nontrapping_dom_walker::after_dom_children (basic_block bb
)
1773 /* This BB isn't on the path to dominator root anymore. */
1777 /* We see the expression EXP in basic block BB. If it's an interesting
1778 expression (an MEM_REF through an SSA_NAME) possibly insert the
1779 expression into the set NONTRAP or the hash table of seen expressions.
1780 STORE is true if this expression is on the LHS, otherwise it's on
1783 nontrapping_dom_walker::add_or_mark_expr (basic_block bb
, tree exp
, bool store
)
1787 if (TREE_CODE (exp
) == MEM_REF
1788 && TREE_CODE (TREE_OPERAND (exp
, 0)) == SSA_NAME
1789 && tree_fits_shwi_p (TREE_OPERAND (exp
, 1))
1790 && (size
= int_size_in_bytes (TREE_TYPE (exp
))) > 0)
1792 tree name
= TREE_OPERAND (exp
, 0);
1793 struct name_to_bb map
;
1795 struct name_to_bb
*n2bb
;
1796 basic_block found_bb
= 0;
1798 /* Try to find the last seen MEM_REF through the same
1799 SSA_NAME, which can trap. */
1800 map
.ssa_name_ver
= SSA_NAME_VERSION (name
);
1804 map
.offset
= tree_to_shwi (TREE_OPERAND (exp
, 1));
1807 slot
= m_seen_ssa_names
.find_slot (&map
, INSERT
);
1809 if (n2bb
&& n2bb
->phase
>= nt_call_phase
)
1810 found_bb
= n2bb
->bb
;
1812 /* If we've found a trapping MEM_REF, _and_ it dominates EXP
1813 (it's in a basic block on the path from us to the dominator root)
1814 then we can't trap. */
1815 if (found_bb
&& (((size_t)found_bb
->aux
) & 1) == 1)
1817 m_nontrapping
->add (exp
);
1821 /* EXP might trap, so insert it into the hash table. */
1824 n2bb
->phase
= nt_call_phase
;
1829 n2bb
= XNEW (struct name_to_bb
);
1830 n2bb
->ssa_name_ver
= SSA_NAME_VERSION (name
);
1831 n2bb
->phase
= nt_call_phase
;
1833 n2bb
->store
= store
;
1834 n2bb
->offset
= map
.offset
;
1842 /* This is the entry point of gathering non trapping memory accesses.
1843 It will do a dominator walk over the whole function, and it will
1844 make use of the bb->aux pointers. It returns a set of trees
1845 (the MEM_REFs itself) which can't trap. */
1846 static hash_set
<tree
> *
1847 get_non_trapping (void)
1850 hash_set
<tree
> *nontrap
= new hash_set
<tree
>;
1851 /* We're going to do a dominator walk, so ensure that we have
1852 dominance information. */
1853 calculate_dominance_info (CDI_DOMINATORS
);
1855 nontrapping_dom_walker (CDI_DOMINATORS
, nontrap
)
1856 .walk (cfun
->cfg
->x_entry_block_ptr
);
1858 clear_aux_for_blocks ();
1862 /* Do the main work of conditional store replacement. We already know
1863 that the recognized pattern looks like so:
1866 if (cond) goto MIDDLE_BB; else goto JOIN_BB (edge E1)
1869 fallthrough (edge E0)
1873 We check that MIDDLE_BB contains only one store, that that store
1874 doesn't trap (not via NOTRAP, but via checking if an access to the same
1875 memory location dominates us) and that the store has a "simple" RHS. */
1878 cond_store_replacement (basic_block middle_bb
, basic_block join_bb
,
1879 edge e0
, edge e1
, hash_set
<tree
> *nontrap
)
1881 gimple
*assign
= last_and_only_stmt (middle_bb
);
1882 tree lhs
, rhs
, name
, name2
;
1885 gimple_stmt_iterator gsi
;
1886 source_location locus
;
1888 /* Check if middle_bb contains of only one store. */
1890 || !gimple_assign_single_p (assign
)
1891 || gimple_has_volatile_ops (assign
))
1894 locus
= gimple_location (assign
);
1895 lhs
= gimple_assign_lhs (assign
);
1896 rhs
= gimple_assign_rhs1 (assign
);
1897 if (TREE_CODE (lhs
) != MEM_REF
1898 || TREE_CODE (TREE_OPERAND (lhs
, 0)) != SSA_NAME
1899 || !is_gimple_reg_type (TREE_TYPE (lhs
)))
1902 /* Prove that we can move the store down. We could also check
1903 TREE_THIS_NOTRAP here, but in that case we also could move stores,
1904 whose value is not available readily, which we want to avoid. */
1905 if (!nontrap
->contains (lhs
))
1908 /* Now we've checked the constraints, so do the transformation:
1909 1) Remove the single store. */
1910 gsi
= gsi_for_stmt (assign
);
1911 unlink_stmt_vdef (assign
);
1912 gsi_remove (&gsi
, true);
1913 release_defs (assign
);
1915 /* 2) Insert a load from the memory of the store to the temporary
1916 on the edge which did not contain the store. */
1917 lhs
= unshare_expr (lhs
);
1918 name
= make_temp_ssa_name (TREE_TYPE (lhs
), NULL
, "cstore");
1919 new_stmt
= gimple_build_assign (name
, lhs
);
1920 gimple_set_location (new_stmt
, locus
);
1921 gsi_insert_on_edge (e1
, new_stmt
);
1923 /* 3) Create a PHI node at the join block, with one argument
1924 holding the old RHS, and the other holding the temporary
1925 where we stored the old memory contents. */
1926 name2
= make_temp_ssa_name (TREE_TYPE (lhs
), NULL
, "cstore");
1927 newphi
= create_phi_node (name2
, join_bb
);
1928 add_phi_arg (newphi
, rhs
, e0
, locus
);
1929 add_phi_arg (newphi
, name
, e1
, locus
);
1931 lhs
= unshare_expr (lhs
);
1932 new_stmt
= gimple_build_assign (lhs
, PHI_RESULT (newphi
));
1934 /* 4) Insert that PHI node. */
1935 gsi
= gsi_after_labels (join_bb
);
1936 if (gsi_end_p (gsi
))
1938 gsi
= gsi_last_bb (join_bb
);
1939 gsi_insert_after (&gsi
, new_stmt
, GSI_NEW_STMT
);
1942 gsi_insert_before (&gsi
, new_stmt
, GSI_NEW_STMT
);
1947 /* Do the main work of conditional store replacement. */
1950 cond_if_else_store_replacement_1 (basic_block then_bb
, basic_block else_bb
,
1951 basic_block join_bb
, gimple
*then_assign
,
1952 gimple
*else_assign
)
1954 tree lhs_base
, lhs
, then_rhs
, else_rhs
, name
;
1955 source_location then_locus
, else_locus
;
1956 gimple_stmt_iterator gsi
;
1960 if (then_assign
== NULL
1961 || !gimple_assign_single_p (then_assign
)
1962 || gimple_clobber_p (then_assign
)
1963 || gimple_has_volatile_ops (then_assign
)
1964 || else_assign
== NULL
1965 || !gimple_assign_single_p (else_assign
)
1966 || gimple_clobber_p (else_assign
)
1967 || gimple_has_volatile_ops (else_assign
))
1970 lhs
= gimple_assign_lhs (then_assign
);
1971 if (!is_gimple_reg_type (TREE_TYPE (lhs
))
1972 || !operand_equal_p (lhs
, gimple_assign_lhs (else_assign
), 0))
1975 lhs_base
= get_base_address (lhs
);
1976 if (lhs_base
== NULL_TREE
1977 || (!DECL_P (lhs_base
) && TREE_CODE (lhs_base
) != MEM_REF
))
1980 then_rhs
= gimple_assign_rhs1 (then_assign
);
1981 else_rhs
= gimple_assign_rhs1 (else_assign
);
1982 then_locus
= gimple_location (then_assign
);
1983 else_locus
= gimple_location (else_assign
);
1985 /* Now we've checked the constraints, so do the transformation:
1986 1) Remove the stores. */
1987 gsi
= gsi_for_stmt (then_assign
);
1988 unlink_stmt_vdef (then_assign
);
1989 gsi_remove (&gsi
, true);
1990 release_defs (then_assign
);
1992 gsi
= gsi_for_stmt (else_assign
);
1993 unlink_stmt_vdef (else_assign
);
1994 gsi_remove (&gsi
, true);
1995 release_defs (else_assign
);
1997 /* 2) Create a PHI node at the join block, with one argument
1998 holding the old RHS, and the other holding the temporary
1999 where we stored the old memory contents. */
2000 name
= make_temp_ssa_name (TREE_TYPE (lhs
), NULL
, "cstore");
2001 newphi
= create_phi_node (name
, join_bb
);
2002 add_phi_arg (newphi
, then_rhs
, EDGE_SUCC (then_bb
, 0), then_locus
);
2003 add_phi_arg (newphi
, else_rhs
, EDGE_SUCC (else_bb
, 0), else_locus
);
2005 new_stmt
= gimple_build_assign (lhs
, PHI_RESULT (newphi
));
2007 /* 3) Insert that PHI node. */
2008 gsi
= gsi_after_labels (join_bb
);
2009 if (gsi_end_p (gsi
))
2011 gsi
= gsi_last_bb (join_bb
);
2012 gsi_insert_after (&gsi
, new_stmt
, GSI_NEW_STMT
);
2015 gsi_insert_before (&gsi
, new_stmt
, GSI_NEW_STMT
);
2020 /* Conditional store replacement. We already know
2021 that the recognized pattern looks like so:
2024 if (cond) goto THEN_BB; else goto ELSE_BB (edge E1)
2034 fallthrough (edge E0)
2038 We check that it is safe to sink the store to JOIN_BB by verifying that
2039 there are no read-after-write or write-after-write dependencies in
2040 THEN_BB and ELSE_BB. */
2043 cond_if_else_store_replacement (basic_block then_bb
, basic_block else_bb
,
2044 basic_block join_bb
)
2046 gimple
*then_assign
= last_and_only_stmt (then_bb
);
2047 gimple
*else_assign
= last_and_only_stmt (else_bb
);
2048 vec
<data_reference_p
> then_datarefs
, else_datarefs
;
2049 vec
<ddr_p
> then_ddrs
, else_ddrs
;
2050 gimple
*then_store
, *else_store
;
2051 bool found
, ok
= false, res
;
2052 struct data_dependence_relation
*ddr
;
2053 data_reference_p then_dr
, else_dr
;
2055 tree then_lhs
, else_lhs
;
2056 basic_block blocks
[3];
2058 if (MAX_STORES_TO_SINK
== 0)
2061 /* Handle the case with single statement in THEN_BB and ELSE_BB. */
2062 if (then_assign
&& else_assign
)
2063 return cond_if_else_store_replacement_1 (then_bb
, else_bb
, join_bb
,
2064 then_assign
, else_assign
);
2066 /* Find data references. */
2067 then_datarefs
.create (1);
2068 else_datarefs
.create (1);
2069 if ((find_data_references_in_bb (NULL
, then_bb
, &then_datarefs
)
2071 || !then_datarefs
.length ()
2072 || (find_data_references_in_bb (NULL
, else_bb
, &else_datarefs
)
2074 || !else_datarefs
.length ())
2076 free_data_refs (then_datarefs
);
2077 free_data_refs (else_datarefs
);
2081 /* Find pairs of stores with equal LHS. */
2082 auto_vec
<gimple
*, 1> then_stores
, else_stores
;
2083 FOR_EACH_VEC_ELT (then_datarefs
, i
, then_dr
)
2085 if (DR_IS_READ (then_dr
))
2088 then_store
= DR_STMT (then_dr
);
2089 then_lhs
= gimple_get_lhs (then_store
);
2090 if (then_lhs
== NULL_TREE
)
2094 FOR_EACH_VEC_ELT (else_datarefs
, j
, else_dr
)
2096 if (DR_IS_READ (else_dr
))
2099 else_store
= DR_STMT (else_dr
);
2100 else_lhs
= gimple_get_lhs (else_store
);
2101 if (else_lhs
== NULL_TREE
)
2104 if (operand_equal_p (then_lhs
, else_lhs
, 0))
2114 then_stores
.safe_push (then_store
);
2115 else_stores
.safe_push (else_store
);
2118 /* No pairs of stores found. */
2119 if (!then_stores
.length ()
2120 || then_stores
.length () > (unsigned) MAX_STORES_TO_SINK
)
2122 free_data_refs (then_datarefs
);
2123 free_data_refs (else_datarefs
);
2127 /* Compute and check data dependencies in both basic blocks. */
2128 then_ddrs
.create (1);
2129 else_ddrs
.create (1);
2130 if (!compute_all_dependences (then_datarefs
, &then_ddrs
,
2132 || !compute_all_dependences (else_datarefs
, &else_ddrs
,
2135 free_dependence_relations (then_ddrs
);
2136 free_dependence_relations (else_ddrs
);
2137 free_data_refs (then_datarefs
);
2138 free_data_refs (else_datarefs
);
2141 blocks
[0] = then_bb
;
2142 blocks
[1] = else_bb
;
2143 blocks
[2] = join_bb
;
2144 renumber_gimple_stmt_uids_in_blocks (blocks
, 3);
2146 /* Check that there are no read-after-write or write-after-write dependencies
2148 FOR_EACH_VEC_ELT (then_ddrs
, i
, ddr
)
2150 struct data_reference
*dra
= DDR_A (ddr
);
2151 struct data_reference
*drb
= DDR_B (ddr
);
2153 if (DDR_ARE_DEPENDENT (ddr
) != chrec_known
2154 && ((DR_IS_READ (dra
) && DR_IS_WRITE (drb
)
2155 && gimple_uid (DR_STMT (dra
)) > gimple_uid (DR_STMT (drb
)))
2156 || (DR_IS_READ (drb
) && DR_IS_WRITE (dra
)
2157 && gimple_uid (DR_STMT (drb
)) > gimple_uid (DR_STMT (dra
)))
2158 || (DR_IS_WRITE (dra
) && DR_IS_WRITE (drb
))))
2160 free_dependence_relations (then_ddrs
);
2161 free_dependence_relations (else_ddrs
);
2162 free_data_refs (then_datarefs
);
2163 free_data_refs (else_datarefs
);
2168 /* Check that there are no read-after-write or write-after-write dependencies
2170 FOR_EACH_VEC_ELT (else_ddrs
, i
, ddr
)
2172 struct data_reference
*dra
= DDR_A (ddr
);
2173 struct data_reference
*drb
= DDR_B (ddr
);
2175 if (DDR_ARE_DEPENDENT (ddr
) != chrec_known
2176 && ((DR_IS_READ (dra
) && DR_IS_WRITE (drb
)
2177 && gimple_uid (DR_STMT (dra
)) > gimple_uid (DR_STMT (drb
)))
2178 || (DR_IS_READ (drb
) && DR_IS_WRITE (dra
)
2179 && gimple_uid (DR_STMT (drb
)) > gimple_uid (DR_STMT (dra
)))
2180 || (DR_IS_WRITE (dra
) && DR_IS_WRITE (drb
))))
2182 free_dependence_relations (then_ddrs
);
2183 free_dependence_relations (else_ddrs
);
2184 free_data_refs (then_datarefs
);
2185 free_data_refs (else_datarefs
);
2190 /* Sink stores with same LHS. */
2191 FOR_EACH_VEC_ELT (then_stores
, i
, then_store
)
2193 else_store
= else_stores
[i
];
2194 res
= cond_if_else_store_replacement_1 (then_bb
, else_bb
, join_bb
,
2195 then_store
, else_store
);
2199 free_dependence_relations (then_ddrs
);
2200 free_dependence_relations (else_ddrs
);
2201 free_data_refs (then_datarefs
);
2202 free_data_refs (else_datarefs
);
2207 /* Return TRUE if STMT has a VUSE whose corresponding VDEF is in BB. */
2210 local_mem_dependence (gimple
*stmt
, basic_block bb
)
2212 tree vuse
= gimple_vuse (stmt
);
2218 def
= SSA_NAME_DEF_STMT (vuse
);
2219 return (def
&& gimple_bb (def
) == bb
);
2222 /* Given a "diamond" control-flow pattern where BB0 tests a condition,
2223 BB1 and BB2 are "then" and "else" blocks dependent on this test,
2224 and BB3 rejoins control flow following BB1 and BB2, look for
2225 opportunities to hoist loads as follows. If BB3 contains a PHI of
2226 two loads, one each occurring in BB1 and BB2, and the loads are
2227 provably of adjacent fields in the same structure, then move both
2228 loads into BB0. Of course this can only be done if there are no
2229 dependencies preventing such motion.
2231 One of the hoisted loads will always be speculative, so the
2232 transformation is currently conservative:
2234 - The fields must be strictly adjacent.
2235 - The two fields must occupy a single memory block that is
2236 guaranteed to not cross a page boundary.
2238 The last is difficult to prove, as such memory blocks should be
2239 aligned on the minimum of the stack alignment boundary and the
2240 alignment guaranteed by heap allocation interfaces. Thus we rely
2241 on a parameter for the alignment value.
2243 Provided a good value is used for the last case, the first
2244 restriction could possibly be relaxed. */
2247 hoist_adjacent_loads (basic_block bb0
, basic_block bb1
,
2248 basic_block bb2
, basic_block bb3
)
2250 int param_align
= PARAM_VALUE (PARAM_L1_CACHE_LINE_SIZE
);
2251 unsigned param_align_bits
= (unsigned) (param_align
* BITS_PER_UNIT
);
2254 /* Walk the phis in bb3 looking for an opportunity. We are looking
2255 for phis of two SSA names, one each of which is defined in bb1 and
2257 for (gsi
= gsi_start_phis (bb3
); !gsi_end_p (gsi
); gsi_next (&gsi
))
2259 gphi
*phi_stmt
= gsi
.phi ();
2260 gimple
*def1
, *def2
;
2261 tree arg1
, arg2
, ref1
, ref2
, field1
, field2
;
2262 tree tree_offset1
, tree_offset2
, tree_size2
, next
;
2263 int offset1
, offset2
, size2
;
2265 gimple_stmt_iterator gsi2
;
2266 basic_block bb_for_def1
, bb_for_def2
;
2268 if (gimple_phi_num_args (phi_stmt
) != 2
2269 || virtual_operand_p (gimple_phi_result (phi_stmt
)))
2272 arg1
= gimple_phi_arg_def (phi_stmt
, 0);
2273 arg2
= gimple_phi_arg_def (phi_stmt
, 1);
2275 if (TREE_CODE (arg1
) != SSA_NAME
2276 || TREE_CODE (arg2
) != SSA_NAME
2277 || SSA_NAME_IS_DEFAULT_DEF (arg1
)
2278 || SSA_NAME_IS_DEFAULT_DEF (arg2
))
2281 def1
= SSA_NAME_DEF_STMT (arg1
);
2282 def2
= SSA_NAME_DEF_STMT (arg2
);
2284 if ((gimple_bb (def1
) != bb1
|| gimple_bb (def2
) != bb2
)
2285 && (gimple_bb (def2
) != bb1
|| gimple_bb (def1
) != bb2
))
2288 /* Check the mode of the arguments to be sure a conditional move
2289 can be generated for it. */
2290 if (optab_handler (movcc_optab
, TYPE_MODE (TREE_TYPE (arg1
)))
2291 == CODE_FOR_nothing
)
2294 /* Both statements must be assignments whose RHS is a COMPONENT_REF. */
2295 if (!gimple_assign_single_p (def1
)
2296 || !gimple_assign_single_p (def2
)
2297 || gimple_has_volatile_ops (def1
)
2298 || gimple_has_volatile_ops (def2
))
2301 ref1
= gimple_assign_rhs1 (def1
);
2302 ref2
= gimple_assign_rhs1 (def2
);
2304 if (TREE_CODE (ref1
) != COMPONENT_REF
2305 || TREE_CODE (ref2
) != COMPONENT_REF
)
2308 /* The zeroth operand of the two component references must be
2309 identical. It is not sufficient to compare get_base_address of
2310 the two references, because this could allow for different
2311 elements of the same array in the two trees. It is not safe to
2312 assume that the existence of one array element implies the
2313 existence of a different one. */
2314 if (!operand_equal_p (TREE_OPERAND (ref1
, 0), TREE_OPERAND (ref2
, 0), 0))
2317 field1
= TREE_OPERAND (ref1
, 1);
2318 field2
= TREE_OPERAND (ref2
, 1);
2320 /* Check for field adjacency, and ensure field1 comes first. */
2321 for (next
= DECL_CHAIN (field1
);
2322 next
&& TREE_CODE (next
) != FIELD_DECL
;
2323 next
= DECL_CHAIN (next
))
2328 for (next
= DECL_CHAIN (field2
);
2329 next
&& TREE_CODE (next
) != FIELD_DECL
;
2330 next
= DECL_CHAIN (next
))
2336 std::swap (field1
, field2
);
2337 std::swap (def1
, def2
);
2340 bb_for_def1
= gimple_bb (def1
);
2341 bb_for_def2
= gimple_bb (def2
);
2343 /* Check for proper alignment of the first field. */
2344 tree_offset1
= bit_position (field1
);
2345 tree_offset2
= bit_position (field2
);
2346 tree_size2
= DECL_SIZE (field2
);
2348 if (!tree_fits_uhwi_p (tree_offset1
)
2349 || !tree_fits_uhwi_p (tree_offset2
)
2350 || !tree_fits_uhwi_p (tree_size2
))
2353 offset1
= tree_to_uhwi (tree_offset1
);
2354 offset2
= tree_to_uhwi (tree_offset2
);
2355 size2
= tree_to_uhwi (tree_size2
);
2356 align1
= DECL_ALIGN (field1
) % param_align_bits
;
2358 if (offset1
% BITS_PER_UNIT
!= 0)
2361 /* For profitability, the two field references should fit within
2362 a single cache line. */
2363 if (align1
+ offset2
- offset1
+ size2
> param_align_bits
)
2366 /* The two expressions cannot be dependent upon vdefs defined
2368 if (local_mem_dependence (def1
, bb_for_def1
)
2369 || local_mem_dependence (def2
, bb_for_def2
))
2372 /* The conditions are satisfied; hoist the loads from bb1 and bb2 into
2373 bb0. We hoist the first one first so that a cache miss is handled
2374 efficiently regardless of hardware cache-fill policy. */
2375 gsi2
= gsi_for_stmt (def1
);
2376 gsi_move_to_bb_end (&gsi2
, bb0
);
2377 gsi2
= gsi_for_stmt (def2
);
2378 gsi_move_to_bb_end (&gsi2
, bb0
);
2380 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
2383 "\nHoisting adjacent loads from %d and %d into %d: \n",
2384 bb_for_def1
->index
, bb_for_def2
->index
, bb0
->index
);
2385 print_gimple_stmt (dump_file
, def1
, 0, TDF_VOPS
|TDF_MEMSYMS
);
2386 print_gimple_stmt (dump_file
, def2
, 0, TDF_VOPS
|TDF_MEMSYMS
);
2391 /* Determine whether we should attempt to hoist adjacent loads out of
2392 diamond patterns in pass_phiopt. Always hoist loads if
2393 -fhoist-adjacent-loads is specified and the target machine has
2394 both a conditional move instruction and a defined cache line size. */
2397 gate_hoist_loads (void)
2399 return (flag_hoist_adjacent_loads
== 1
2400 && PARAM_VALUE (PARAM_L1_CACHE_LINE_SIZE
)
2401 && HAVE_conditional_move
);
2404 /* This pass tries to replaces an if-then-else block with an
2405 assignment. We have four kinds of transformations. Some of these
2406 transformations are also performed by the ifcvt RTL optimizer.
2408 Conditional Replacement
2409 -----------------------
2411 This transformation, implemented in conditional_replacement,
2415 if (cond) goto bb2; else goto bb1;
2418 x = PHI <0 (bb1), 1 (bb0), ...>;
2426 x = PHI <x' (bb0), ...>;
2428 We remove bb1 as it becomes unreachable. This occurs often due to
2429 gimplification of conditionals.
2434 This transformation, implemented in value_replacement, replaces
2437 if (a != b) goto bb2; else goto bb1;
2440 x = PHI <a (bb1), b (bb0), ...>;
2446 x = PHI <b (bb0), ...>;
2448 This opportunity can sometimes occur as a result of other
2452 Another case caught by value replacement looks like this:
2458 if (t3 != 0) goto bb1; else goto bb2;
2474 This transformation, implemented in abs_replacement, replaces
2477 if (a >= 0) goto bb2; else goto bb1;
2481 x = PHI <x (bb1), a (bb0), ...>;
2488 x = PHI <x' (bb0), ...>;
2493 This transformation, minmax_replacement replaces
2496 if (a <= b) goto bb2; else goto bb1;
2499 x = PHI <b (bb1), a (bb0), ...>;
2504 x' = MIN_EXPR (a, b)
2506 x = PHI <x' (bb0), ...>;
2508 A similar transformation is done for MAX_EXPR.
2511 This pass also performs a fifth transformation of a slightly different
2514 Factor conversion in COND_EXPR
2515 ------------------------------
2517 This transformation factors the conversion out of COND_EXPR with
2518 factor_out_conditional_conversion.
2521 if (a <= CST) goto <bb 3>; else goto <bb 4>;
2525 tmp = PHI <tmp, CST>
2528 if (a <= CST) goto <bb 3>; else goto <bb 4>;
2534 Adjacent Load Hoisting
2535 ----------------------
2537 This transformation replaces
2540 if (...) goto bb2; else goto bb1;
2542 x1 = (<expr>).field1;
2545 x2 = (<expr>).field2;
2552 x1 = (<expr>).field1;
2553 x2 = (<expr>).field2;
2554 if (...) goto bb2; else goto bb1;
2561 The purpose of this transformation is to enable generation of conditional
2562 move instructions such as Intel CMOVE or PowerPC ISEL. Because one of
2563 the loads is speculative, the transformation is restricted to very
2564 specific cases to avoid introducing a page fault. We are looking for
2572 where left and right are typically adjacent pointers in a tree structure. */
2576 const pass_data pass_data_phiopt
=
2578 GIMPLE_PASS
, /* type */
2579 "phiopt", /* name */
2580 OPTGROUP_NONE
, /* optinfo_flags */
2581 TV_TREE_PHIOPT
, /* tv_id */
2582 ( PROP_cfg
| PROP_ssa
), /* properties_required */
2583 0, /* properties_provided */
2584 0, /* properties_destroyed */
2585 0, /* todo_flags_start */
2586 0, /* todo_flags_finish */
2589 class pass_phiopt
: public gimple_opt_pass
2592 pass_phiopt (gcc::context
*ctxt
)
2593 : gimple_opt_pass (pass_data_phiopt
, ctxt
)
2596 /* opt_pass methods: */
2597 opt_pass
* clone () { return new pass_phiopt (m_ctxt
); }
2598 virtual bool gate (function
*) { return flag_ssa_phiopt
; }
2599 virtual unsigned int execute (function
*)
2601 return tree_ssa_phiopt_worker (false, gate_hoist_loads ());
2604 }; // class pass_phiopt
2609 make_pass_phiopt (gcc::context
*ctxt
)
2611 return new pass_phiopt (ctxt
);
2616 const pass_data pass_data_cselim
=
2618 GIMPLE_PASS
, /* type */
2619 "cselim", /* name */
2620 OPTGROUP_NONE
, /* optinfo_flags */
2621 TV_TREE_PHIOPT
, /* tv_id */
2622 ( PROP_cfg
| PROP_ssa
), /* properties_required */
2623 0, /* properties_provided */
2624 0, /* properties_destroyed */
2625 0, /* todo_flags_start */
2626 0, /* todo_flags_finish */
2629 class pass_cselim
: public gimple_opt_pass
2632 pass_cselim (gcc::context
*ctxt
)
2633 : gimple_opt_pass (pass_data_cselim
, ctxt
)
2636 /* opt_pass methods: */
2637 virtual bool gate (function
*) { return flag_tree_cselim
; }
2638 virtual unsigned int execute (function
*) { return tree_ssa_cs_elim (); }
2640 }; // class pass_cselim
2645 make_pass_cselim (gcc::context
*ctxt
)
2647 return new pass_cselim (ctxt
);