1 /* Optimization of PHI nodes by converting them into straightline code.
2 Copyright (C) 2004-2018 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"
48 #include "case-cfn-macros.h"
50 static unsigned int tree_ssa_phiopt_worker (bool, bool, bool);
51 static bool conditional_replacement (basic_block
, basic_block
,
52 edge
, edge
, gphi
*, tree
, tree
);
53 static gphi
*factor_out_conditional_conversion (edge
, edge
, gphi
*, tree
, tree
,
55 static int value_replacement (basic_block
, basic_block
,
56 edge
, edge
, gimple
*, tree
, tree
);
57 static bool minmax_replacement (basic_block
, basic_block
,
58 edge
, edge
, gimple
*, tree
, tree
);
59 static bool abs_replacement (basic_block
, basic_block
,
60 edge
, edge
, gimple
*, tree
, tree
);
61 static bool cond_removal_in_popcount_pattern (basic_block
, basic_block
,
62 edge
, edge
, gimple
*, tree
, tree
);
63 static bool cond_store_replacement (basic_block
, basic_block
, edge
, edge
,
65 static bool cond_if_else_store_replacement (basic_block
, basic_block
, basic_block
);
66 static hash_set
<tree
> * get_non_trapping ();
67 static void replace_phi_edge_with_variable (basic_block
, edge
, gimple
*, tree
);
68 static void hoist_adjacent_loads (basic_block
, basic_block
,
69 basic_block
, basic_block
);
70 static bool gate_hoist_loads (void);
72 /* This pass tries to transform conditional stores into unconditional
73 ones, enabling further simplifications with the simpler then and else
74 blocks. In particular it replaces this:
77 if (cond) goto bb2; else goto bb1;
85 if (cond) goto bb1; else goto bb2;
89 condtmp = PHI <RHS, condtmp'>
92 This transformation can only be done under several constraints,
93 documented below. It also replaces:
96 if (cond) goto bb2; else goto bb1;
107 if (cond) goto bb3; else goto bb1;
110 condtmp = PHI <RHS1, RHS2>
114 tree_ssa_cs_elim (void)
117 /* ??? We are not interested in loop related info, but the following
118 will create it, ICEing as we didn't init loops with pre-headers.
119 An interfacing issue of find_data_references_in_bb. */
120 loop_optimizer_init (LOOPS_NORMAL
);
122 todo
= tree_ssa_phiopt_worker (true, false, false);
124 loop_optimizer_finalize ();
128 /* Return the singleton PHI in the SEQ of PHIs for edges E0 and E1. */
131 single_non_singleton_phi_for_edges (gimple_seq seq
, edge e0
, edge e1
)
133 gimple_stmt_iterator i
;
135 if (gimple_seq_singleton_p (seq
))
136 return as_a
<gphi
*> (gsi_stmt (gsi_start (seq
)));
137 for (i
= gsi_start (seq
); !gsi_end_p (i
); gsi_next (&i
))
139 gphi
*p
= as_a
<gphi
*> (gsi_stmt (i
));
140 /* If the PHI arguments are equal then we can skip this PHI. */
141 if (operand_equal_for_phi_arg_p (gimple_phi_arg_def (p
, e0
->dest_idx
),
142 gimple_phi_arg_def (p
, e1
->dest_idx
)))
145 /* If we already have a PHI that has the two edge arguments are
146 different, then return it is not a singleton for these PHIs. */
155 /* The core routine of conditional store replacement and normal
156 phi optimizations. Both share much of the infrastructure in how
157 to match applicable basic block patterns. DO_STORE_ELIM is true
158 when we want to do conditional store replacement, false otherwise.
159 DO_HOIST_LOADS is true when we want to hoist adjacent loads out
160 of diamond control flow patterns, false otherwise. */
162 tree_ssa_phiopt_worker (bool do_store_elim
, bool do_hoist_loads
, bool early_p
)
165 basic_block
*bb_order
;
167 bool cfgchanged
= false;
168 hash_set
<tree
> *nontrap
= 0;
171 /* Calculate the set of non-trapping memory accesses. */
172 nontrap
= get_non_trapping ();
174 /* Search every basic block for COND_EXPR we may be able to optimize.
176 We walk the blocks in order that guarantees that a block with
177 a single predecessor is processed before the predecessor.
178 This ensures that we collapse inner ifs before visiting the
179 outer ones, and also that we do not try to visit a removed
181 bb_order
= single_pred_before_succ_order ();
182 n
= n_basic_blocks_for_fn (cfun
) - NUM_FIXED_BLOCKS
;
184 for (i
= 0; i
< n
; i
++)
188 basic_block bb1
, bb2
;
194 cond_stmt
= last_stmt (bb
);
195 /* Check to see if the last statement is a GIMPLE_COND. */
197 || gimple_code (cond_stmt
) != GIMPLE_COND
)
200 e1
= EDGE_SUCC (bb
, 0);
202 e2
= EDGE_SUCC (bb
, 1);
205 /* We cannot do the optimization on abnormal edges. */
206 if ((e1
->flags
& EDGE_ABNORMAL
) != 0
207 || (e2
->flags
& EDGE_ABNORMAL
) != 0)
210 /* If either bb1's succ or bb2 or bb2's succ is non NULL. */
211 if (EDGE_COUNT (bb1
->succs
) == 0
213 || EDGE_COUNT (bb2
->succs
) == 0)
216 /* Find the bb which is the fall through to the other. */
217 if (EDGE_SUCC (bb1
, 0)->dest
== bb2
)
219 else if (EDGE_SUCC (bb2
, 0)->dest
== bb1
)
221 std::swap (bb1
, bb2
);
224 else if (do_store_elim
225 && EDGE_SUCC (bb1
, 0)->dest
== EDGE_SUCC (bb2
, 0)->dest
)
227 basic_block bb3
= EDGE_SUCC (bb1
, 0)->dest
;
229 if (!single_succ_p (bb1
)
230 || (EDGE_SUCC (bb1
, 0)->flags
& EDGE_FALLTHRU
) == 0
231 || !single_succ_p (bb2
)
232 || (EDGE_SUCC (bb2
, 0)->flags
& EDGE_FALLTHRU
) == 0
233 || EDGE_COUNT (bb3
->preds
) != 2)
235 if (cond_if_else_store_replacement (bb1
, bb2
, bb3
))
239 else if (do_hoist_loads
240 && EDGE_SUCC (bb1
, 0)->dest
== EDGE_SUCC (bb2
, 0)->dest
)
242 basic_block bb3
= EDGE_SUCC (bb1
, 0)->dest
;
244 if (!FLOAT_TYPE_P (TREE_TYPE (gimple_cond_lhs (cond_stmt
)))
245 && single_succ_p (bb1
)
246 && single_succ_p (bb2
)
247 && single_pred_p (bb1
)
248 && single_pred_p (bb2
)
249 && EDGE_COUNT (bb
->succs
) == 2
250 && EDGE_COUNT (bb3
->preds
) == 2
251 /* If one edge or the other is dominant, a conditional move
252 is likely to perform worse than the well-predicted branch. */
253 && !predictable_edge_p (EDGE_SUCC (bb
, 0))
254 && !predictable_edge_p (EDGE_SUCC (bb
, 1)))
255 hoist_adjacent_loads (bb
, bb1
, bb2
, bb3
);
261 e1
= EDGE_SUCC (bb1
, 0);
263 /* Make sure that bb1 is just a fall through. */
264 if (!single_succ_p (bb1
)
265 || (e1
->flags
& EDGE_FALLTHRU
) == 0)
268 /* Also make sure that bb1 only have one predecessor and that it
270 if (!single_pred_p (bb1
)
271 || single_pred (bb1
) != bb
)
276 /* bb1 is the middle block, bb2 the join block, bb the split block,
277 e1 the fallthrough edge from bb1 to bb2. We can't do the
278 optimization if the join block has more than two predecessors. */
279 if (EDGE_COUNT (bb2
->preds
) > 2)
281 if (cond_store_replacement (bb1
, bb2
, e1
, e2
, nontrap
))
286 gimple_seq phis
= phi_nodes (bb2
);
287 gimple_stmt_iterator gsi
;
288 bool candorest
= true;
290 /* Value replacement can work with more than one PHI
291 so try that first. */
293 for (gsi
= gsi_start (phis
); !gsi_end_p (gsi
); gsi_next (&gsi
))
295 phi
= as_a
<gphi
*> (gsi_stmt (gsi
));
296 arg0
= gimple_phi_arg_def (phi
, e1
->dest_idx
);
297 arg1
= gimple_phi_arg_def (phi
, e2
->dest_idx
);
298 if (value_replacement (bb
, bb1
, e1
, e2
, phi
, arg0
, arg1
) == 2)
309 phi
= single_non_singleton_phi_for_edges (phis
, e1
, e2
);
313 arg0
= gimple_phi_arg_def (phi
, e1
->dest_idx
);
314 arg1
= gimple_phi_arg_def (phi
, e2
->dest_idx
);
316 /* Something is wrong if we cannot find the arguments in the PHI
318 gcc_assert (arg0
!= NULL_TREE
&& arg1
!= NULL_TREE
);
320 gphi
*newphi
= factor_out_conditional_conversion (e1
, e2
, phi
,
326 /* factor_out_conditional_conversion may create a new PHI in
327 BB2 and eliminate an existing PHI in BB2. Recompute values
328 that may be affected by that change. */
329 arg0
= gimple_phi_arg_def (phi
, e1
->dest_idx
);
330 arg1
= gimple_phi_arg_def (phi
, e2
->dest_idx
);
331 gcc_assert (arg0
!= NULL_TREE
&& arg1
!= NULL_TREE
);
334 /* Do the replacement of conditional if it can be done. */
336 && conditional_replacement (bb
, bb1
, e1
, e2
, phi
, arg0
, arg1
))
338 else if (abs_replacement (bb
, bb1
, e1
, e2
, phi
, arg0
, arg1
))
341 && cond_removal_in_popcount_pattern (bb
, bb1
, e1
, e2
,
344 else if (minmax_replacement (bb
, bb1
, e1
, e2
, phi
, arg0
, arg1
))
353 /* If the CFG has changed, we should cleanup the CFG. */
354 if (cfgchanged
&& do_store_elim
)
356 /* In cond-store replacement we have added some loads on edges
357 and new VOPS (as we moved the store, and created a load). */
358 gsi_commit_edge_inserts ();
359 return TODO_cleanup_cfg
| TODO_update_ssa_only_virtuals
;
362 return TODO_cleanup_cfg
;
366 /* Replace PHI node element whose edge is E in block BB with variable NEW.
367 Remove the edge from COND_BLOCK which does not lead to BB (COND_BLOCK
368 is known to have two edges, one of which must reach BB). */
371 replace_phi_edge_with_variable (basic_block cond_block
,
372 edge e
, gimple
*phi
, tree new_tree
)
374 basic_block bb
= gimple_bb (phi
);
375 basic_block block_to_remove
;
376 gimple_stmt_iterator gsi
;
378 /* Change the PHI argument to new. */
379 SET_USE (PHI_ARG_DEF_PTR (phi
, e
->dest_idx
), new_tree
);
381 /* Remove the empty basic block. */
382 if (EDGE_SUCC (cond_block
, 0)->dest
== bb
)
384 EDGE_SUCC (cond_block
, 0)->flags
|= EDGE_FALLTHRU
;
385 EDGE_SUCC (cond_block
, 0)->flags
&= ~(EDGE_TRUE_VALUE
| EDGE_FALSE_VALUE
);
386 EDGE_SUCC (cond_block
, 0)->probability
= profile_probability::always ();
388 block_to_remove
= EDGE_SUCC (cond_block
, 1)->dest
;
392 EDGE_SUCC (cond_block
, 1)->flags
|= EDGE_FALLTHRU
;
393 EDGE_SUCC (cond_block
, 1)->flags
394 &= ~(EDGE_TRUE_VALUE
| EDGE_FALSE_VALUE
);
395 EDGE_SUCC (cond_block
, 1)->probability
= profile_probability::always ();
397 block_to_remove
= EDGE_SUCC (cond_block
, 0)->dest
;
399 delete_basic_block (block_to_remove
);
401 /* Eliminate the COND_EXPR at the end of COND_BLOCK. */
402 gsi
= gsi_last_bb (cond_block
);
403 gsi_remove (&gsi
, true);
405 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
407 "COND_EXPR in block %d and PHI in block %d converted to straightline code.\n",
412 /* PR66726: Factor conversion out of COND_EXPR. If the arguments of the PHI
413 stmt are CONVERT_STMT, factor out the conversion and perform the conversion
414 to the result of PHI stmt. COND_STMT is the controlling predicate.
415 Return the newly-created PHI, if any. */
418 factor_out_conditional_conversion (edge e0
, edge e1
, gphi
*phi
,
419 tree arg0
, tree arg1
, gimple
*cond_stmt
)
421 gimple
*arg0_def_stmt
= NULL
, *arg1_def_stmt
= NULL
, *new_stmt
;
422 tree new_arg0
= NULL_TREE
, new_arg1
= NULL_TREE
;
425 gimple_stmt_iterator gsi
, gsi_for_def
;
426 source_location locus
= gimple_location (phi
);
427 enum tree_code convert_code
;
429 /* Handle only PHI statements with two arguments. TODO: If all
430 other arguments to PHI are INTEGER_CST or if their defining
431 statement have the same unary operation, we can handle more
432 than two arguments too. */
433 if (gimple_phi_num_args (phi
) != 2)
436 /* First canonicalize to simplify tests. */
437 if (TREE_CODE (arg0
) != SSA_NAME
)
439 std::swap (arg0
, arg1
);
443 if (TREE_CODE (arg0
) != SSA_NAME
444 || (TREE_CODE (arg1
) != SSA_NAME
445 && TREE_CODE (arg1
) != INTEGER_CST
))
448 /* Check if arg0 is an SSA_NAME and the stmt which defines arg0 is
450 arg0_def_stmt
= SSA_NAME_DEF_STMT (arg0
);
451 if (!gimple_assign_cast_p (arg0_def_stmt
))
454 /* Use the RHS as new_arg0. */
455 convert_code
= gimple_assign_rhs_code (arg0_def_stmt
);
456 new_arg0
= gimple_assign_rhs1 (arg0_def_stmt
);
457 if (convert_code
== VIEW_CONVERT_EXPR
)
459 new_arg0
= TREE_OPERAND (new_arg0
, 0);
460 if (!is_gimple_reg_type (TREE_TYPE (new_arg0
)))
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
))
486 /* For the INTEGER_CST case, we are just moving the
487 conversion from one place to another, which can often
488 hurt as the conversion moves further away from the
489 statement that computes the value. So, perform this
490 only if new_arg0 is an operand of COND_STMT, or
491 if arg0_def_stmt is the only non-debug stmt in
492 its basic block, because then it is possible this
493 could enable further optimizations (minmax replacement
494 etc.). See PR71016. */
495 if (new_arg0
!= gimple_cond_lhs (cond_stmt
)
496 && new_arg0
!= gimple_cond_rhs (cond_stmt
)
497 && gimple_bb (arg0_def_stmt
) == e0
->src
)
499 gsi
= gsi_for_stmt (arg0_def_stmt
);
500 gsi_prev_nondebug (&gsi
);
501 if (!gsi_end_p (gsi
))
503 gsi
= gsi_for_stmt (arg0_def_stmt
);
504 gsi_next_nondebug (&gsi
);
505 if (!gsi_end_p (gsi
))
508 new_arg1
= fold_convert (TREE_TYPE (new_arg0
), arg1
);
517 /* If arg0/arg1 have > 1 use, then this transformation actually increases
518 the number of expressions evaluated at runtime. */
519 if (!has_single_use (arg0
)
520 || (arg1_def_stmt
&& !has_single_use (arg1
)))
523 /* If types of new_arg0 and new_arg1 are different bailout. */
524 if (!types_compatible_p (TREE_TYPE (new_arg0
), TREE_TYPE (new_arg1
)))
527 /* Create a new PHI stmt. */
528 result
= PHI_RESULT (phi
);
529 temp
= make_ssa_name (TREE_TYPE (new_arg0
), NULL
);
530 newphi
= create_phi_node (temp
, gimple_bb (phi
));
532 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
534 fprintf (dump_file
, "PHI ");
535 print_generic_expr (dump_file
, gimple_phi_result (phi
));
537 " changed to factor conversion out from COND_EXPR.\n");
538 fprintf (dump_file
, "New stmt with CAST that defines ");
539 print_generic_expr (dump_file
, result
);
540 fprintf (dump_file
, ".\n");
543 /* Remove the old cast(s) that has single use. */
544 gsi_for_def
= gsi_for_stmt (arg0_def_stmt
);
545 gsi_remove (&gsi_for_def
, true);
546 release_defs (arg0_def_stmt
);
550 gsi_for_def
= gsi_for_stmt (arg1_def_stmt
);
551 gsi_remove (&gsi_for_def
, true);
552 release_defs (arg1_def_stmt
);
555 add_phi_arg (newphi
, new_arg0
, e0
, locus
);
556 add_phi_arg (newphi
, new_arg1
, e1
, locus
);
558 /* Create the conversion stmt and insert it. */
559 if (convert_code
== VIEW_CONVERT_EXPR
)
561 temp
= fold_build1 (VIEW_CONVERT_EXPR
, TREE_TYPE (result
), temp
);
562 new_stmt
= gimple_build_assign (result
, temp
);
565 new_stmt
= gimple_build_assign (result
, convert_code
, temp
);
566 gsi
= gsi_after_labels (gimple_bb (phi
));
567 gsi_insert_before (&gsi
, new_stmt
, GSI_SAME_STMT
);
569 /* Remove the original PHI stmt. */
570 gsi
= gsi_for_stmt (phi
);
571 gsi_remove (&gsi
, true);
575 /* The function conditional_replacement does the main work of doing the
576 conditional replacement. Return true if the replacement is done.
577 Otherwise return false.
578 BB is the basic block where the replacement is going to be done on. ARG0
579 is argument 0 from PHI. Likewise for ARG1. */
582 conditional_replacement (basic_block cond_bb
, basic_block middle_bb
,
583 edge e0
, edge e1
, gphi
*phi
,
584 tree arg0
, tree arg1
)
590 gimple_stmt_iterator gsi
;
591 edge true_edge
, false_edge
;
592 tree new_var
, new_var2
;
595 /* FIXME: Gimplification of complex type is too hard for now. */
596 /* We aren't prepared to handle vectors either (and it is a question
597 if it would be worthwhile anyway). */
598 if (!(INTEGRAL_TYPE_P (TREE_TYPE (arg0
))
599 || POINTER_TYPE_P (TREE_TYPE (arg0
)))
600 || !(INTEGRAL_TYPE_P (TREE_TYPE (arg1
))
601 || POINTER_TYPE_P (TREE_TYPE (arg1
))))
604 /* The PHI arguments have the constants 0 and 1, or 0 and -1, then
605 convert it to the conditional. */
606 if ((integer_zerop (arg0
) && integer_onep (arg1
))
607 || (integer_zerop (arg1
) && integer_onep (arg0
)))
609 else if ((integer_zerop (arg0
) && integer_all_onesp (arg1
))
610 || (integer_zerop (arg1
) && integer_all_onesp (arg0
)))
615 if (!empty_block_p (middle_bb
))
618 /* At this point we know we have a GIMPLE_COND with two successors.
619 One successor is BB, the other successor is an empty block which
620 falls through into BB.
622 There is a single PHI node at the join point (BB) and its arguments
623 are constants (0, 1) or (0, -1).
625 So, given the condition COND, and the two PHI arguments, we can
626 rewrite this PHI into non-branching code:
628 dest = (COND) or dest = COND'
630 We use the condition as-is if the argument associated with the
631 true edge has the value one or the argument associated with the
632 false edge as the value zero. Note that those conditions are not
633 the same since only one of the outgoing edges from the GIMPLE_COND
634 will directly reach BB and thus be associated with an argument. */
636 stmt
= last_stmt (cond_bb
);
637 result
= PHI_RESULT (phi
);
639 /* To handle special cases like floating point comparison, it is easier and
640 less error-prone to build a tree and gimplify it on the fly though it is
642 cond
= fold_build2_loc (gimple_location (stmt
),
643 gimple_cond_code (stmt
), boolean_type_node
,
644 gimple_cond_lhs (stmt
), gimple_cond_rhs (stmt
));
646 /* We need to know which is the true edge and which is the false
647 edge so that we know when to invert the condition below. */
648 extract_true_false_edges_from_block (cond_bb
, &true_edge
, &false_edge
);
649 if ((e0
== true_edge
&& integer_zerop (arg0
))
650 || (e0
== false_edge
&& !integer_zerop (arg0
))
651 || (e1
== true_edge
&& integer_zerop (arg1
))
652 || (e1
== false_edge
&& !integer_zerop (arg1
)))
653 cond
= fold_build1_loc (gimple_location (stmt
),
654 TRUTH_NOT_EXPR
, TREE_TYPE (cond
), cond
);
658 cond
= fold_convert_loc (gimple_location (stmt
),
659 TREE_TYPE (result
), cond
);
660 cond
= fold_build1_loc (gimple_location (stmt
),
661 NEGATE_EXPR
, TREE_TYPE (cond
), cond
);
664 /* Insert our new statements at the end of conditional block before the
666 gsi
= gsi_for_stmt (stmt
);
667 new_var
= force_gimple_operand_gsi (&gsi
, cond
, true, NULL
, true,
670 if (!useless_type_conversion_p (TREE_TYPE (result
), TREE_TYPE (new_var
)))
672 source_location locus_0
, locus_1
;
674 new_var2
= make_ssa_name (TREE_TYPE (result
));
675 new_stmt
= gimple_build_assign (new_var2
, CONVERT_EXPR
, new_var
);
676 gsi_insert_before (&gsi
, new_stmt
, GSI_SAME_STMT
);
679 /* Set the locus to the first argument, unless is doesn't have one. */
680 locus_0
= gimple_phi_arg_location (phi
, 0);
681 locus_1
= gimple_phi_arg_location (phi
, 1);
682 if (locus_0
== UNKNOWN_LOCATION
)
684 gimple_set_location (new_stmt
, locus_0
);
687 replace_phi_edge_with_variable (cond_bb
, e1
, phi
, new_var
);
689 /* Note that we optimized this PHI. */
693 /* Update *ARG which is defined in STMT so that it contains the
694 computed value if that seems profitable. Return true if the
695 statement is made dead by that rewriting. */
698 jump_function_from_stmt (tree
*arg
, gimple
*stmt
)
700 enum tree_code code
= gimple_assign_rhs_code (stmt
);
701 if (code
== ADDR_EXPR
)
703 /* For arg = &p->i transform it to p, if possible. */
704 tree rhs1
= gimple_assign_rhs1 (stmt
);
706 tree tem
= get_addr_base_and_unit_offset (TREE_OPERAND (rhs1
, 0),
709 && TREE_CODE (tem
) == MEM_REF
710 && known_eq (mem_ref_offset (tem
) + offset
, 0))
712 *arg
= TREE_OPERAND (tem
, 0);
716 /* TODO: Much like IPA-CP jump-functions we want to handle constant
717 additions symbolically here, and we'd need to update the comparison
718 code that compares the arg + cst tuples in our caller. For now the
719 code above exactly handles the VEC_BASE pattern from vec.h. */
723 /* RHS is a source argument in a BIT_AND_EXPR which feeds a conditional
724 of the form SSA_NAME NE 0.
726 If RHS is fed by a simple EQ_EXPR comparison of two values, see if
727 the two input values of the EQ_EXPR match arg0 and arg1.
729 If so update *code and return TRUE. Otherwise return FALSE. */
732 rhs_is_fed_for_value_replacement (const_tree arg0
, const_tree arg1
,
733 enum tree_code
*code
, const_tree rhs
)
735 /* Obviously if RHS is not an SSA_NAME, we can't look at the defining
737 if (TREE_CODE (rhs
) == SSA_NAME
)
739 gimple
*def1
= SSA_NAME_DEF_STMT (rhs
);
741 /* Verify the defining statement has an EQ_EXPR on the RHS. */
742 if (is_gimple_assign (def1
) && gimple_assign_rhs_code (def1
) == EQ_EXPR
)
744 /* Finally verify the source operands of the EQ_EXPR are equal
746 tree op0
= gimple_assign_rhs1 (def1
);
747 tree op1
= gimple_assign_rhs2 (def1
);
748 if ((operand_equal_for_phi_arg_p (arg0
, op0
)
749 && operand_equal_for_phi_arg_p (arg1
, op1
))
750 || (operand_equal_for_phi_arg_p (arg0
, op1
)
751 && operand_equal_for_phi_arg_p (arg1
, op0
)))
753 /* We will perform the optimization. */
754 *code
= gimple_assign_rhs_code (def1
);
762 /* Return TRUE if arg0/arg1 are equal to the rhs/lhs or lhs/rhs of COND.
764 Also return TRUE if arg0/arg1 are equal to the source arguments of a
765 an EQ comparison feeding a BIT_AND_EXPR which feeds COND.
767 Return FALSE otherwise. */
770 operand_equal_for_value_replacement (const_tree arg0
, const_tree arg1
,
771 enum tree_code
*code
, gimple
*cond
)
774 tree lhs
= gimple_cond_lhs (cond
);
775 tree rhs
= gimple_cond_rhs (cond
);
777 if ((operand_equal_for_phi_arg_p (arg0
, lhs
)
778 && operand_equal_for_phi_arg_p (arg1
, rhs
))
779 || (operand_equal_for_phi_arg_p (arg1
, lhs
)
780 && operand_equal_for_phi_arg_p (arg0
, rhs
)))
783 /* Now handle more complex case where we have an EQ comparison
784 which feeds a BIT_AND_EXPR which feeds COND.
786 First verify that COND is of the form SSA_NAME NE 0. */
787 if (*code
!= NE_EXPR
|| !integer_zerop (rhs
)
788 || TREE_CODE (lhs
) != SSA_NAME
)
791 /* Now ensure that SSA_NAME is set by a BIT_AND_EXPR. */
792 def
= SSA_NAME_DEF_STMT (lhs
);
793 if (!is_gimple_assign (def
) || gimple_assign_rhs_code (def
) != BIT_AND_EXPR
)
796 /* Now verify arg0/arg1 correspond to the source arguments of an
797 EQ comparison feeding the BIT_AND_EXPR. */
799 tree tmp
= gimple_assign_rhs1 (def
);
800 if (rhs_is_fed_for_value_replacement (arg0
, arg1
, code
, tmp
))
803 tmp
= gimple_assign_rhs2 (def
);
804 if (rhs_is_fed_for_value_replacement (arg0
, arg1
, code
, tmp
))
810 /* Returns true if ARG is a neutral element for operation CODE
811 on the RIGHT side. */
814 neutral_element_p (tree_code code
, tree arg
, bool right
)
821 return integer_zerop (arg
);
828 case POINTER_PLUS_EXPR
:
829 return right
&& integer_zerop (arg
);
832 return integer_onep (arg
);
839 return right
&& integer_onep (arg
);
842 return integer_all_onesp (arg
);
849 /* Returns true if ARG is an absorbing element for operation CODE. */
852 absorbing_element_p (tree_code code
, tree arg
, bool right
, tree rval
)
857 return integer_all_onesp (arg
);
861 return integer_zerop (arg
);
867 return !right
&& integer_zerop (arg
);
879 && integer_zerop (arg
)
880 && tree_single_nonzero_warnv_p (rval
, NULL
));
887 /* The function value_replacement does the main work of doing the value
888 replacement. Return non-zero if the replacement is done. Otherwise return
889 0. If we remove the middle basic block, return 2.
890 BB is the basic block where the replacement is going to be done on. ARG0
891 is argument 0 from the PHI. Likewise for ARG1. */
894 value_replacement (basic_block cond_bb
, basic_block middle_bb
,
895 edge e0
, edge e1
, gimple
*phi
,
896 tree arg0
, tree arg1
)
898 gimple_stmt_iterator gsi
;
900 edge true_edge
, false_edge
;
902 bool emtpy_or_with_defined_p
= true;
904 /* If the type says honor signed zeros we cannot do this
906 if (HONOR_SIGNED_ZEROS (arg1
))
909 /* If there is a statement in MIDDLE_BB that defines one of the PHI
910 arguments, then adjust arg0 or arg1. */
911 gsi
= gsi_start_nondebug_after_labels_bb (middle_bb
);
912 while (!gsi_end_p (gsi
))
914 gimple
*stmt
= gsi_stmt (gsi
);
916 gsi_next_nondebug (&gsi
);
917 if (!is_gimple_assign (stmt
))
919 if (gimple_code (stmt
) != GIMPLE_PREDICT
920 && gimple_code (stmt
) != GIMPLE_NOP
)
921 emtpy_or_with_defined_p
= false;
924 /* Now try to adjust arg0 or arg1 according to the computation
926 lhs
= gimple_assign_lhs (stmt
);
928 && jump_function_from_stmt (&arg0
, stmt
))
930 && jump_function_from_stmt (&arg1
, stmt
)))
931 emtpy_or_with_defined_p
= false;
934 cond
= last_stmt (cond_bb
);
935 code
= gimple_cond_code (cond
);
937 /* This transformation is only valid for equality comparisons. */
938 if (code
!= NE_EXPR
&& code
!= EQ_EXPR
)
941 /* We need to know which is the true edge and which is the false
942 edge so that we know if have abs or negative abs. */
943 extract_true_false_edges_from_block (cond_bb
, &true_edge
, &false_edge
);
945 /* At this point we know we have a COND_EXPR with two successors.
946 One successor is BB, the other successor is an empty block which
947 falls through into BB.
949 The condition for the COND_EXPR is known to be NE_EXPR or EQ_EXPR.
951 There is a single PHI node at the join point (BB) with two arguments.
953 We now need to verify that the two arguments in the PHI node match
954 the two arguments to the equality comparison. */
956 if (operand_equal_for_value_replacement (arg0
, arg1
, &code
, cond
))
961 /* For NE_EXPR, we want to build an assignment result = arg where
962 arg is the PHI argument associated with the true edge. For
963 EQ_EXPR we want the PHI argument associated with the false edge. */
964 e
= (code
== NE_EXPR
? true_edge
: false_edge
);
966 /* Unfortunately, E may not reach BB (it may instead have gone to
967 OTHER_BLOCK). If that is the case, then we want the single outgoing
968 edge from OTHER_BLOCK which reaches BB and represents the desired
969 path from COND_BLOCK. */
970 if (e
->dest
== middle_bb
)
971 e
= single_succ_edge (e
->dest
);
973 /* Now we know the incoming edge to BB that has the argument for the
974 RHS of our new assignment statement. */
980 /* If the middle basic block was empty or is defining the
981 PHI arguments and this is a single phi where the args are different
982 for the edges e0 and e1 then we can remove the middle basic block. */
983 if (emtpy_or_with_defined_p
984 && single_non_singleton_phi_for_edges (phi_nodes (gimple_bb (phi
)),
987 replace_phi_edge_with_variable (cond_bb
, e1
, phi
, arg
);
988 /* Note that we optimized this PHI. */
993 /* Replace the PHI arguments with arg. */
994 SET_PHI_ARG_DEF (phi
, e0
->dest_idx
, arg
);
995 SET_PHI_ARG_DEF (phi
, e1
->dest_idx
, arg
);
996 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
998 fprintf (dump_file
, "PHI ");
999 print_generic_expr (dump_file
, gimple_phi_result (phi
));
1000 fprintf (dump_file
, " reduced for COND_EXPR in block %d to ",
1002 print_generic_expr (dump_file
, arg
);
1003 fprintf (dump_file
, ".\n");
1010 /* Now optimize (x != 0) ? x + y : y to just x + y. */
1011 gsi
= gsi_last_nondebug_bb (middle_bb
);
1012 if (gsi_end_p (gsi
))
1015 gimple
*assign
= gsi_stmt (gsi
);
1016 if (!is_gimple_assign (assign
)
1017 || gimple_assign_rhs_class (assign
) != GIMPLE_BINARY_RHS
1018 || (!INTEGRAL_TYPE_P (TREE_TYPE (arg0
))
1019 && !POINTER_TYPE_P (TREE_TYPE (arg0
))))
1022 /* Punt if there are (degenerate) PHIs in middle_bb, there should not be. */
1023 if (!gimple_seq_empty_p (phi_nodes (middle_bb
)))
1026 /* Allow up to 2 cheap preparation statements that prepare argument
1034 iftmp.0_6 = x_5(D) r<< _1;
1036 # iftmp.0_2 = PHI <iftmp.0_6(3), x_5(D)(2)>
1047 # _2 = PHI <x_5(D)(2), _6(3)> */
1048 gimple
*prep_stmt
[2] = { NULL
, NULL
};
1050 for (prep_cnt
= 0; ; prep_cnt
++)
1052 gsi_prev_nondebug (&gsi
);
1053 if (gsi_end_p (gsi
))
1056 gimple
*g
= gsi_stmt (gsi
);
1057 if (gimple_code (g
) == GIMPLE_LABEL
)
1060 if (prep_cnt
== 2 || !is_gimple_assign (g
))
1063 tree lhs
= gimple_assign_lhs (g
);
1064 tree rhs1
= gimple_assign_rhs1 (g
);
1065 use_operand_p use_p
;
1067 if (TREE_CODE (lhs
) != SSA_NAME
1068 || TREE_CODE (rhs1
) != SSA_NAME
1069 || !INTEGRAL_TYPE_P (TREE_TYPE (lhs
))
1070 || !INTEGRAL_TYPE_P (TREE_TYPE (rhs1
))
1071 || !single_imm_use (lhs
, &use_p
, &use_stmt
)
1072 || use_stmt
!= (prep_cnt
? prep_stmt
[prep_cnt
- 1] : assign
))
1074 switch (gimple_assign_rhs_code (g
))
1082 if (TREE_CODE (gimple_assign_rhs2 (g
)) != INTEGER_CST
)
1088 prep_stmt
[prep_cnt
] = g
;
1091 /* Only transform if it removes the condition. */
1092 if (!single_non_singleton_phi_for_edges (phi_nodes (gimple_bb (phi
)), e0
, e1
))
1095 /* Size-wise, this is always profitable. */
1096 if (optimize_bb_for_speed_p (cond_bb
)
1097 /* The special case is useless if it has a low probability. */
1098 && profile_status_for_fn (cfun
) != PROFILE_ABSENT
1099 && EDGE_PRED (middle_bb
, 0)->probability
< profile_probability::even ()
1100 /* If assign is cheap, there is no point avoiding it. */
1101 && estimate_num_insns (bb_seq (middle_bb
), &eni_time_weights
)
1102 >= 3 * estimate_num_insns (cond
, &eni_time_weights
))
1105 tree lhs
= gimple_assign_lhs (assign
);
1106 tree rhs1
= gimple_assign_rhs1 (assign
);
1107 tree rhs2
= gimple_assign_rhs2 (assign
);
1108 enum tree_code code_def
= gimple_assign_rhs_code (assign
);
1109 tree cond_lhs
= gimple_cond_lhs (cond
);
1110 tree cond_rhs
= gimple_cond_rhs (cond
);
1112 /* Propagate the cond_rhs constant through preparation stmts,
1113 make sure UB isn't invoked while doing that. */
1114 for (int i
= prep_cnt
- 1; i
>= 0; --i
)
1116 gimple
*g
= prep_stmt
[i
];
1117 tree grhs1
= gimple_assign_rhs1 (g
);
1118 if (!operand_equal_for_phi_arg_p (cond_lhs
, grhs1
))
1120 cond_lhs
= gimple_assign_lhs (g
);
1121 cond_rhs
= fold_convert (TREE_TYPE (grhs1
), cond_rhs
);
1122 if (TREE_CODE (cond_rhs
) != INTEGER_CST
1123 || TREE_OVERFLOW (cond_rhs
))
1125 if (gimple_assign_rhs_class (g
) == GIMPLE_BINARY_RHS
)
1127 cond_rhs
= int_const_binop (gimple_assign_rhs_code (g
), cond_rhs
,
1128 gimple_assign_rhs2 (g
));
1129 if (TREE_OVERFLOW (cond_rhs
))
1132 cond_rhs
= fold_convert (TREE_TYPE (cond_lhs
), cond_rhs
);
1133 if (TREE_CODE (cond_rhs
) != INTEGER_CST
1134 || TREE_OVERFLOW (cond_rhs
))
1138 if (((code
== NE_EXPR
&& e1
== false_edge
)
1139 || (code
== EQ_EXPR
&& e1
== true_edge
))
1142 && operand_equal_for_phi_arg_p (rhs2
, cond_lhs
)
1143 && neutral_element_p (code_def
, cond_rhs
, true))
1145 && operand_equal_for_phi_arg_p (rhs1
, cond_lhs
)
1146 && neutral_element_p (code_def
, cond_rhs
, false))
1147 || (operand_equal_for_phi_arg_p (arg1
, cond_rhs
)
1148 && ((operand_equal_for_phi_arg_p (rhs2
, cond_lhs
)
1149 && absorbing_element_p (code_def
, cond_rhs
, true, rhs2
))
1150 || (operand_equal_for_phi_arg_p (rhs1
, cond_lhs
)
1151 && absorbing_element_p (code_def
,
1152 cond_rhs
, false, rhs2
))))))
1154 gsi
= gsi_for_stmt (cond
);
1155 /* Moving ASSIGN might change VR of lhs, e.g. when moving u_6
1163 # RANGE [0, 4294967294]
1164 u_6 = n_5 + 4294967295;
1167 # u_3 = PHI <u_6(3), 4294967295(2)> */
1168 reset_flow_sensitive_info (lhs
);
1169 if (INTEGRAL_TYPE_P (TREE_TYPE (lhs
)))
1171 /* If available, we can use VR of phi result at least. */
1172 tree phires
= gimple_phi_result (phi
);
1173 struct range_info_def
*phires_range_info
1174 = SSA_NAME_RANGE_INFO (phires
);
1175 if (phires_range_info
)
1176 duplicate_ssa_name_range_info (lhs
, SSA_NAME_RANGE_TYPE (phires
),
1179 gimple_stmt_iterator gsi_from
;
1180 for (int i
= prep_cnt
- 1; i
>= 0; --i
)
1182 tree plhs
= gimple_assign_lhs (prep_stmt
[i
]);
1183 reset_flow_sensitive_info (plhs
);
1184 gsi_from
= gsi_for_stmt (prep_stmt
[i
]);
1185 gsi_move_before (&gsi_from
, &gsi
);
1187 gsi_from
= gsi_for_stmt (assign
);
1188 gsi_move_before (&gsi_from
, &gsi
);
1189 replace_phi_edge_with_variable (cond_bb
, e1
, phi
, lhs
);
1196 /* The function minmax_replacement does the main work of doing the minmax
1197 replacement. Return true if the replacement is done. Otherwise return
1199 BB is the basic block where the replacement is going to be done on. ARG0
1200 is argument 0 from the PHI. Likewise for ARG1. */
1203 minmax_replacement (basic_block cond_bb
, basic_block middle_bb
,
1204 edge e0
, edge e1
, gimple
*phi
,
1205 tree arg0
, tree arg1
)
1207 tree result
, type
, rhs
;
1210 edge true_edge
, false_edge
;
1211 enum tree_code cmp
, minmax
, ass_code
;
1212 tree smaller
, alt_smaller
, larger
, alt_larger
, arg_true
, arg_false
;
1213 gimple_stmt_iterator gsi
, gsi_from
;
1215 type
= TREE_TYPE (PHI_RESULT (phi
));
1217 /* The optimization may be unsafe due to NaNs. */
1218 if (HONOR_NANS (type
) || HONOR_SIGNED_ZEROS (type
))
1221 cond
= as_a
<gcond
*> (last_stmt (cond_bb
));
1222 cmp
= gimple_cond_code (cond
);
1223 rhs
= gimple_cond_rhs (cond
);
1225 /* Turn EQ/NE of extreme values to order comparisons. */
1226 if ((cmp
== NE_EXPR
|| cmp
== EQ_EXPR
)
1227 && TREE_CODE (rhs
) == INTEGER_CST
)
1229 if (wi::eq_p (wi::to_wide (rhs
), wi::min_value (TREE_TYPE (rhs
))))
1231 cmp
= (cmp
== EQ_EXPR
) ? LT_EXPR
: GE_EXPR
;
1232 rhs
= wide_int_to_tree (TREE_TYPE (rhs
),
1233 wi::min_value (TREE_TYPE (rhs
)) + 1);
1235 else if (wi::eq_p (wi::to_wide (rhs
), wi::max_value (TREE_TYPE (rhs
))))
1237 cmp
= (cmp
== EQ_EXPR
) ? GT_EXPR
: LE_EXPR
;
1238 rhs
= wide_int_to_tree (TREE_TYPE (rhs
),
1239 wi::max_value (TREE_TYPE (rhs
)) - 1);
1243 /* This transformation is only valid for order comparisons. Record which
1244 operand is smaller/larger if the result of the comparison is true. */
1245 alt_smaller
= NULL_TREE
;
1246 alt_larger
= NULL_TREE
;
1247 if (cmp
== LT_EXPR
|| cmp
== LE_EXPR
)
1249 smaller
= gimple_cond_lhs (cond
);
1251 /* If we have smaller < CST it is equivalent to smaller <= CST-1.
1252 Likewise smaller <= CST is equivalent to smaller < CST+1. */
1253 if (TREE_CODE (larger
) == INTEGER_CST
)
1257 wi::overflow_type overflow
;
1258 wide_int alt
= wi::sub (wi::to_wide (larger
), 1,
1259 TYPE_SIGN (TREE_TYPE (larger
)),
1262 alt_larger
= wide_int_to_tree (TREE_TYPE (larger
), alt
);
1266 wi::overflow_type overflow
;
1267 wide_int alt
= wi::add (wi::to_wide (larger
), 1,
1268 TYPE_SIGN (TREE_TYPE (larger
)),
1271 alt_larger
= wide_int_to_tree (TREE_TYPE (larger
), alt
);
1275 else if (cmp
== GT_EXPR
|| cmp
== GE_EXPR
)
1278 larger
= gimple_cond_lhs (cond
);
1279 /* If we have larger > CST it is equivalent to larger >= CST+1.
1280 Likewise larger >= CST is equivalent to larger > CST-1. */
1281 if (TREE_CODE (smaller
) == INTEGER_CST
)
1283 wi::overflow_type overflow
;
1286 wide_int alt
= wi::add (wi::to_wide (smaller
), 1,
1287 TYPE_SIGN (TREE_TYPE (smaller
)),
1290 alt_smaller
= wide_int_to_tree (TREE_TYPE (smaller
), alt
);
1294 wide_int alt
= wi::sub (wi::to_wide (smaller
), 1,
1295 TYPE_SIGN (TREE_TYPE (smaller
)),
1298 alt_smaller
= wide_int_to_tree (TREE_TYPE (smaller
), alt
);
1305 /* We need to know which is the true edge and which is the false
1306 edge so that we know if have abs or negative abs. */
1307 extract_true_false_edges_from_block (cond_bb
, &true_edge
, &false_edge
);
1309 /* Forward the edges over the middle basic block. */
1310 if (true_edge
->dest
== middle_bb
)
1311 true_edge
= EDGE_SUCC (true_edge
->dest
, 0);
1312 if (false_edge
->dest
== middle_bb
)
1313 false_edge
= EDGE_SUCC (false_edge
->dest
, 0);
1315 if (true_edge
== e0
)
1317 gcc_assert (false_edge
== e1
);
1323 gcc_assert (false_edge
== e0
);
1324 gcc_assert (true_edge
== e1
);
1329 if (empty_block_p (middle_bb
))
1331 if ((operand_equal_for_phi_arg_p (arg_true
, smaller
)
1333 && operand_equal_for_phi_arg_p (arg_true
, alt_smaller
)))
1334 && (operand_equal_for_phi_arg_p (arg_false
, larger
)
1336 && operand_equal_for_phi_arg_p (arg_true
, alt_larger
))))
1340 if (smaller < larger)
1346 else if ((operand_equal_for_phi_arg_p (arg_false
, smaller
)
1348 && operand_equal_for_phi_arg_p (arg_false
, alt_smaller
)))
1349 && (operand_equal_for_phi_arg_p (arg_true
, larger
)
1351 && operand_equal_for_phi_arg_p (arg_true
, alt_larger
))))
1358 /* Recognize the following case, assuming d <= u:
1364 This is equivalent to
1369 gimple
*assign
= last_and_only_stmt (middle_bb
);
1370 tree lhs
, op0
, op1
, bound
;
1373 || gimple_code (assign
) != GIMPLE_ASSIGN
)
1376 lhs
= gimple_assign_lhs (assign
);
1377 ass_code
= gimple_assign_rhs_code (assign
);
1378 if (ass_code
!= MAX_EXPR
&& ass_code
!= MIN_EXPR
)
1380 op0
= gimple_assign_rhs1 (assign
);
1381 op1
= gimple_assign_rhs2 (assign
);
1383 if (true_edge
->src
== middle_bb
)
1385 /* We got here if the condition is true, i.e., SMALLER < LARGER. */
1386 if (!operand_equal_for_phi_arg_p (lhs
, arg_true
))
1389 if (operand_equal_for_phi_arg_p (arg_false
, larger
)
1391 && operand_equal_for_phi_arg_p (arg_false
, alt_larger
)))
1395 if (smaller < larger)
1397 r' = MAX_EXPR (smaller, bound)
1399 r = PHI <r', larger> --> to be turned to MIN_EXPR. */
1400 if (ass_code
!= MAX_EXPR
)
1404 if (operand_equal_for_phi_arg_p (op0
, smaller
)
1406 && operand_equal_for_phi_arg_p (op0
, alt_smaller
)))
1408 else if (operand_equal_for_phi_arg_p (op1
, smaller
)
1410 && operand_equal_for_phi_arg_p (op1
, alt_smaller
)))
1415 /* We need BOUND <= LARGER. */
1416 if (!integer_nonzerop (fold_build2 (LE_EXPR
, boolean_type_node
,
1420 else if (operand_equal_for_phi_arg_p (arg_false
, smaller
)
1422 && operand_equal_for_phi_arg_p (arg_false
, alt_smaller
)))
1426 if (smaller < larger)
1428 r' = MIN_EXPR (larger, bound)
1430 r = PHI <r', smaller> --> to be turned to MAX_EXPR. */
1431 if (ass_code
!= MIN_EXPR
)
1435 if (operand_equal_for_phi_arg_p (op0
, larger
)
1437 && operand_equal_for_phi_arg_p (op0
, alt_larger
)))
1439 else if (operand_equal_for_phi_arg_p (op1
, larger
)
1441 && operand_equal_for_phi_arg_p (op1
, alt_larger
)))
1446 /* We need BOUND >= SMALLER. */
1447 if (!integer_nonzerop (fold_build2 (GE_EXPR
, boolean_type_node
,
1456 /* We got here if the condition is false, i.e., SMALLER > LARGER. */
1457 if (!operand_equal_for_phi_arg_p (lhs
, arg_false
))
1460 if (operand_equal_for_phi_arg_p (arg_true
, larger
)
1462 && operand_equal_for_phi_arg_p (arg_true
, alt_larger
)))
1466 if (smaller > larger)
1468 r' = MIN_EXPR (smaller, bound)
1470 r = PHI <r', larger> --> to be turned to MAX_EXPR. */
1471 if (ass_code
!= MIN_EXPR
)
1475 if (operand_equal_for_phi_arg_p (op0
, smaller
)
1477 && operand_equal_for_phi_arg_p (op0
, alt_smaller
)))
1479 else if (operand_equal_for_phi_arg_p (op1
, smaller
)
1481 && operand_equal_for_phi_arg_p (op1
, alt_smaller
)))
1486 /* We need BOUND >= LARGER. */
1487 if (!integer_nonzerop (fold_build2 (GE_EXPR
, boolean_type_node
,
1491 else if (operand_equal_for_phi_arg_p (arg_true
, smaller
)
1493 && operand_equal_for_phi_arg_p (arg_true
, alt_smaller
)))
1497 if (smaller > larger)
1499 r' = MAX_EXPR (larger, bound)
1501 r = PHI <r', smaller> --> to be turned to MIN_EXPR. */
1502 if (ass_code
!= MAX_EXPR
)
1506 if (operand_equal_for_phi_arg_p (op0
, larger
))
1508 else if (operand_equal_for_phi_arg_p (op1
, larger
))
1513 /* We need BOUND <= SMALLER. */
1514 if (!integer_nonzerop (fold_build2 (LE_EXPR
, boolean_type_node
,
1522 /* Move the statement from the middle block. */
1523 gsi
= gsi_last_bb (cond_bb
);
1524 gsi_from
= gsi_last_nondebug_bb (middle_bb
);
1525 reset_flow_sensitive_info (SINGLE_SSA_TREE_OPERAND (gsi_stmt (gsi_from
),
1527 gsi_move_before (&gsi_from
, &gsi
);
1530 /* Create an SSA var to hold the min/max result. If we're the only
1531 things setting the target PHI, then we can clone the PHI
1532 variable. Otherwise we must create a new one. */
1533 result
= PHI_RESULT (phi
);
1534 if (EDGE_COUNT (gimple_bb (phi
)->preds
) == 2)
1535 result
= duplicate_ssa_name (result
, NULL
);
1537 result
= make_ssa_name (TREE_TYPE (result
));
1539 /* Emit the statement to compute min/max. */
1540 new_stmt
= gimple_build_assign (result
, minmax
, arg0
, arg1
);
1541 gsi
= gsi_last_bb (cond_bb
);
1542 gsi_insert_before (&gsi
, new_stmt
, GSI_NEW_STMT
);
1544 replace_phi_edge_with_variable (cond_bb
, e1
, phi
, result
);
1558 _2 = (unsigned long) b_4(D);
1559 _9 = __builtin_popcountl (_2);
1561 _9 = __builtin_popcountl (b_4(D));
1564 c_12 = PHI <0(2), _9(3)>
1568 _2 = (unsigned long) b_4(D);
1569 _9 = __builtin_popcountl (_2);
1571 _9 = __builtin_popcountl (b_4(D));
1578 cond_removal_in_popcount_pattern (basic_block cond_bb
, basic_block middle_bb
,
1580 gimple
*phi
, tree arg0
, tree arg1
)
1583 gimple_stmt_iterator gsi
, gsi_from
;
1585 gimple
*cast
= NULL
;
1589 _2 = (unsigned long) b_4(D);
1590 _9 = __builtin_popcountl (_2);
1592 _9 = __builtin_popcountl (b_4(D));
1593 are the only stmts in the middle_bb. */
1595 gsi
= gsi_start_nondebug_after_labels_bb (middle_bb
);
1596 if (gsi_end_p (gsi
))
1598 cast
= gsi_stmt (gsi
);
1599 gsi_next_nondebug (&gsi
);
1600 if (!gsi_end_p (gsi
))
1602 popcount
= gsi_stmt (gsi
);
1603 gsi_next_nondebug (&gsi
);
1604 if (!gsi_end_p (gsi
))
1613 /* Check that we have a popcount builtin. */
1614 if (!is_gimple_call (popcount
))
1616 combined_fn cfn
= gimple_call_combined_fn (popcount
);
1625 arg
= gimple_call_arg (popcount
, 0);
1626 lhs
= gimple_get_lhs (popcount
);
1630 /* We have a cast stmt feeding popcount builtin. */
1631 /* Check that we have a cast prior to that. */
1632 if (gimple_code (cast
) != GIMPLE_ASSIGN
1633 || !CONVERT_EXPR_CODE_P (gimple_assign_rhs_code (cast
)))
1635 /* Result of the cast stmt is the argument to the builtin. */
1636 if (arg
!= gimple_assign_lhs (cast
))
1638 arg
= gimple_assign_rhs1 (cast
);
1641 cond
= last_stmt (cond_bb
);
1643 /* Cond_bb has a check for b_4 [!=|==] 0 before calling the popcount
1645 if (gimple_code (cond
) != GIMPLE_COND
1646 || (gimple_cond_code (cond
) != NE_EXPR
1647 && gimple_cond_code (cond
) != EQ_EXPR
)
1648 || !integer_zerop (gimple_cond_rhs (cond
))
1649 || arg
!= gimple_cond_lhs (cond
))
1653 if ((e2
->flags
& EDGE_TRUE_VALUE
1654 && gimple_cond_code (cond
) == NE_EXPR
)
1655 || (e1
->flags
& EDGE_TRUE_VALUE
1656 && gimple_cond_code (cond
) == EQ_EXPR
))
1658 std::swap (arg0
, arg1
);
1662 /* Check PHI arguments. */
1663 if (lhs
!= arg0
|| !integer_zerop (arg1
))
1666 /* And insert the popcount builtin and cast stmt before the cond_bb. */
1667 gsi
= gsi_last_bb (cond_bb
);
1670 gsi_from
= gsi_for_stmt (cast
);
1671 gsi_move_before (&gsi_from
, &gsi
);
1672 reset_flow_sensitive_info (gimple_get_lhs (cast
));
1674 gsi_from
= gsi_for_stmt (popcount
);
1675 gsi_move_before (&gsi_from
, &gsi
);
1676 reset_flow_sensitive_info (gimple_get_lhs (popcount
));
1678 /* Now update the PHI and remove unneeded bbs. */
1679 replace_phi_edge_with_variable (cond_bb
, e2
, phi
, lhs
);
1683 /* The function absolute_replacement does the main work of doing the absolute
1684 replacement. Return true if the replacement is done. Otherwise return
1686 bb is the basic block where the replacement is going to be done on. arg0
1687 is argument 0 from the phi. Likewise for arg1. */
1690 abs_replacement (basic_block cond_bb
, basic_block middle_bb
,
1691 edge e0 ATTRIBUTE_UNUSED
, edge e1
,
1692 gimple
*phi
, tree arg0
, tree arg1
)
1697 gimple_stmt_iterator gsi
;
1698 edge true_edge
, false_edge
;
1703 enum tree_code cond_code
;
1705 /* If the type says honor signed zeros we cannot do this
1707 if (HONOR_SIGNED_ZEROS (arg1
))
1710 /* OTHER_BLOCK must have only one executable statement which must have the
1711 form arg0 = -arg1 or arg1 = -arg0. */
1713 assign
= last_and_only_stmt (middle_bb
);
1714 /* If we did not find the proper negation assignment, then we can not
1719 /* If we got here, then we have found the only executable statement
1720 in OTHER_BLOCK. If it is anything other than arg = -arg1 or
1721 arg1 = -arg0, then we can not optimize. */
1722 if (gimple_code (assign
) != GIMPLE_ASSIGN
)
1725 lhs
= gimple_assign_lhs (assign
);
1727 if (gimple_assign_rhs_code (assign
) != NEGATE_EXPR
)
1730 rhs
= gimple_assign_rhs1 (assign
);
1732 /* The assignment has to be arg0 = -arg1 or arg1 = -arg0. */
1733 if (!(lhs
== arg0
&& rhs
== arg1
)
1734 && !(lhs
== arg1
&& rhs
== arg0
))
1737 cond
= last_stmt (cond_bb
);
1738 result
= PHI_RESULT (phi
);
1740 /* Only relationals comparing arg[01] against zero are interesting. */
1741 cond_code
= gimple_cond_code (cond
);
1742 if (cond_code
!= GT_EXPR
&& cond_code
!= GE_EXPR
1743 && cond_code
!= LT_EXPR
&& cond_code
!= LE_EXPR
)
1746 /* Make sure the conditional is arg[01] OP y. */
1747 if (gimple_cond_lhs (cond
) != rhs
)
1750 if (FLOAT_TYPE_P (TREE_TYPE (gimple_cond_rhs (cond
)))
1751 ? real_zerop (gimple_cond_rhs (cond
))
1752 : integer_zerop (gimple_cond_rhs (cond
)))
1757 /* We need to know which is the true edge and which is the false
1758 edge so that we know if have abs or negative abs. */
1759 extract_true_false_edges_from_block (cond_bb
, &true_edge
, &false_edge
);
1761 /* For GT_EXPR/GE_EXPR, if the true edge goes to OTHER_BLOCK, then we
1762 will need to negate the result. Similarly for LT_EXPR/LE_EXPR if
1763 the false edge goes to OTHER_BLOCK. */
1764 if (cond_code
== GT_EXPR
|| cond_code
== GE_EXPR
)
1769 if (e
->dest
== middle_bb
)
1774 /* If the code negates only iff positive then make sure to not
1775 introduce undefined behavior when negating or computing the absolute.
1776 ??? We could use range info if present to check for arg1 == INT_MIN. */
1778 && (ANY_INTEGRAL_TYPE_P (TREE_TYPE (arg1
))
1779 && ! TYPE_OVERFLOW_WRAPS (TREE_TYPE (arg1
))))
1782 result
= duplicate_ssa_name (result
, NULL
);
1785 lhs
= make_ssa_name (TREE_TYPE (result
));
1789 /* Build the modify expression with abs expression. */
1790 new_stmt
= gimple_build_assign (lhs
, ABS_EXPR
, rhs
);
1792 gsi
= gsi_last_bb (cond_bb
);
1793 gsi_insert_before (&gsi
, new_stmt
, GSI_NEW_STMT
);
1797 /* Get the right GSI. We want to insert after the recently
1798 added ABS_EXPR statement (which we know is the first statement
1800 new_stmt
= gimple_build_assign (result
, NEGATE_EXPR
, lhs
);
1802 gsi_insert_after (&gsi
, new_stmt
, GSI_NEW_STMT
);
1805 replace_phi_edge_with_variable (cond_bb
, e1
, phi
, result
);
1807 /* Note that we optimized this PHI. */
1811 /* Auxiliary functions to determine the set of memory accesses which
1812 can't trap because they are preceded by accesses to the same memory
1813 portion. We do that for MEM_REFs, so we only need to track
1814 the SSA_NAME of the pointer indirectly referenced. The algorithm
1815 simply is a walk over all instructions in dominator order. When
1816 we see an MEM_REF we determine if we've already seen a same
1817 ref anywhere up to the root of the dominator tree. If we do the
1818 current access can't trap. If we don't see any dominating access
1819 the current access might trap, but might also make later accesses
1820 non-trapping, so we remember it. We need to be careful with loads
1821 or stores, for instance a load might not trap, while a store would,
1822 so if we see a dominating read access this doesn't mean that a later
1823 write access would not trap. Hence we also need to differentiate the
1824 type of access(es) seen.
1826 ??? We currently are very conservative and assume that a load might
1827 trap even if a store doesn't (write-only memory). This probably is
1828 overly conservative. */
1830 /* A hash-table of SSA_NAMEs, and in which basic block an MEM_REF
1831 through it was seen, which would constitute a no-trap region for
1835 unsigned int ssa_name_ver
;
1838 HOST_WIDE_INT offset
, size
;
1842 /* Hashtable helpers. */
1844 struct ssa_names_hasher
: free_ptr_hash
<name_to_bb
>
1846 static inline hashval_t
hash (const name_to_bb
*);
1847 static inline bool equal (const name_to_bb
*, const name_to_bb
*);
1850 /* Used for quick clearing of the hash-table when we see calls.
1851 Hash entries with phase < nt_call_phase are invalid. */
1852 static unsigned int nt_call_phase
;
1854 /* The hash function. */
1857 ssa_names_hasher::hash (const name_to_bb
*n
)
1859 return n
->ssa_name_ver
^ (((hashval_t
) n
->store
) << 31)
1860 ^ (n
->offset
<< 6) ^ (n
->size
<< 3);
1863 /* The equality function of *P1 and *P2. */
1866 ssa_names_hasher::equal (const name_to_bb
*n1
, const name_to_bb
*n2
)
1868 return n1
->ssa_name_ver
== n2
->ssa_name_ver
1869 && n1
->store
== n2
->store
1870 && n1
->offset
== n2
->offset
1871 && n1
->size
== n2
->size
;
1874 class nontrapping_dom_walker
: public dom_walker
1877 nontrapping_dom_walker (cdi_direction direction
, hash_set
<tree
> *ps
)
1878 : dom_walker (direction
), m_nontrapping (ps
), m_seen_ssa_names (128) {}
1880 virtual edge
before_dom_children (basic_block
);
1881 virtual void after_dom_children (basic_block
);
1885 /* We see the expression EXP in basic block BB. If it's an interesting
1886 expression (an MEM_REF through an SSA_NAME) possibly insert the
1887 expression into the set NONTRAP or the hash table of seen expressions.
1888 STORE is true if this expression is on the LHS, otherwise it's on
1890 void add_or_mark_expr (basic_block
, tree
, bool);
1892 hash_set
<tree
> *m_nontrapping
;
1894 /* The hash table for remembering what we've seen. */
1895 hash_table
<ssa_names_hasher
> m_seen_ssa_names
;
1898 /* Called by walk_dominator_tree, when entering the block BB. */
1900 nontrapping_dom_walker::before_dom_children (basic_block bb
)
1904 gimple_stmt_iterator gsi
;
1906 /* If we haven't seen all our predecessors, clear the hash-table. */
1907 FOR_EACH_EDGE (e
, ei
, bb
->preds
)
1908 if ((((size_t)e
->src
->aux
) & 2) == 0)
1914 /* Mark this BB as being on the path to dominator root and as visited. */
1915 bb
->aux
= (void*)(1 | 2);
1917 /* And walk the statements in order. */
1918 for (gsi
= gsi_start_bb (bb
); !gsi_end_p (gsi
); gsi_next (&gsi
))
1920 gimple
*stmt
= gsi_stmt (gsi
);
1922 if ((gimple_code (stmt
) == GIMPLE_ASM
&& gimple_vdef (stmt
))
1923 || (is_gimple_call (stmt
)
1924 && (!nonfreeing_call_p (stmt
) || !nonbarrier_call_p (stmt
))))
1926 else if (gimple_assign_single_p (stmt
) && !gimple_has_volatile_ops (stmt
))
1928 add_or_mark_expr (bb
, gimple_assign_lhs (stmt
), true);
1929 add_or_mark_expr (bb
, gimple_assign_rhs1 (stmt
), false);
1935 /* Called by walk_dominator_tree, when basic block BB is exited. */
1937 nontrapping_dom_walker::after_dom_children (basic_block bb
)
1939 /* This BB isn't on the path to dominator root anymore. */
1943 /* We see the expression EXP in basic block BB. If it's an interesting
1944 expression (an MEM_REF through an SSA_NAME) possibly insert the
1945 expression into the set NONTRAP or the hash table of seen expressions.
1946 STORE is true if this expression is on the LHS, otherwise it's on
1949 nontrapping_dom_walker::add_or_mark_expr (basic_block bb
, tree exp
, bool store
)
1953 if (TREE_CODE (exp
) == MEM_REF
1954 && TREE_CODE (TREE_OPERAND (exp
, 0)) == SSA_NAME
1955 && tree_fits_shwi_p (TREE_OPERAND (exp
, 1))
1956 && (size
= int_size_in_bytes (TREE_TYPE (exp
))) > 0)
1958 tree name
= TREE_OPERAND (exp
, 0);
1959 struct name_to_bb map
;
1961 struct name_to_bb
*n2bb
;
1962 basic_block found_bb
= 0;
1964 /* Try to find the last seen MEM_REF through the same
1965 SSA_NAME, which can trap. */
1966 map
.ssa_name_ver
= SSA_NAME_VERSION (name
);
1970 map
.offset
= tree_to_shwi (TREE_OPERAND (exp
, 1));
1973 slot
= m_seen_ssa_names
.find_slot (&map
, INSERT
);
1975 if (n2bb
&& n2bb
->phase
>= nt_call_phase
)
1976 found_bb
= n2bb
->bb
;
1978 /* If we've found a trapping MEM_REF, _and_ it dominates EXP
1979 (it's in a basic block on the path from us to the dominator root)
1980 then we can't trap. */
1981 if (found_bb
&& (((size_t)found_bb
->aux
) & 1) == 1)
1983 m_nontrapping
->add (exp
);
1987 /* EXP might trap, so insert it into the hash table. */
1990 n2bb
->phase
= nt_call_phase
;
1995 n2bb
= XNEW (struct name_to_bb
);
1996 n2bb
->ssa_name_ver
= SSA_NAME_VERSION (name
);
1997 n2bb
->phase
= nt_call_phase
;
1999 n2bb
->store
= store
;
2000 n2bb
->offset
= map
.offset
;
2008 /* This is the entry point of gathering non trapping memory accesses.
2009 It will do a dominator walk over the whole function, and it will
2010 make use of the bb->aux pointers. It returns a set of trees
2011 (the MEM_REFs itself) which can't trap. */
2012 static hash_set
<tree
> *
2013 get_non_trapping (void)
2016 hash_set
<tree
> *nontrap
= new hash_set
<tree
>;
2017 /* We're going to do a dominator walk, so ensure that we have
2018 dominance information. */
2019 calculate_dominance_info (CDI_DOMINATORS
);
2021 nontrapping_dom_walker (CDI_DOMINATORS
, nontrap
)
2022 .walk (cfun
->cfg
->x_entry_block_ptr
);
2024 clear_aux_for_blocks ();
2028 /* Do the main work of conditional store replacement. We already know
2029 that the recognized pattern looks like so:
2032 if (cond) goto MIDDLE_BB; else goto JOIN_BB (edge E1)
2035 fallthrough (edge E0)
2039 We check that MIDDLE_BB contains only one store, that that store
2040 doesn't trap (not via NOTRAP, but via checking if an access to the same
2041 memory location dominates us) and that the store has a "simple" RHS. */
2044 cond_store_replacement (basic_block middle_bb
, basic_block join_bb
,
2045 edge e0
, edge e1
, hash_set
<tree
> *nontrap
)
2047 gimple
*assign
= last_and_only_stmt (middle_bb
);
2048 tree lhs
, rhs
, name
, name2
;
2051 gimple_stmt_iterator gsi
;
2052 source_location locus
;
2054 /* Check if middle_bb contains of only one store. */
2056 || !gimple_assign_single_p (assign
)
2057 || gimple_has_volatile_ops (assign
))
2060 locus
= gimple_location (assign
);
2061 lhs
= gimple_assign_lhs (assign
);
2062 rhs
= gimple_assign_rhs1 (assign
);
2063 if (TREE_CODE (lhs
) != MEM_REF
2064 || TREE_CODE (TREE_OPERAND (lhs
, 0)) != SSA_NAME
2065 || !is_gimple_reg_type (TREE_TYPE (lhs
)))
2068 /* Prove that we can move the store down. We could also check
2069 TREE_THIS_NOTRAP here, but in that case we also could move stores,
2070 whose value is not available readily, which we want to avoid. */
2071 if (!nontrap
->contains (lhs
))
2074 /* Now we've checked the constraints, so do the transformation:
2075 1) Remove the single store. */
2076 gsi
= gsi_for_stmt (assign
);
2077 unlink_stmt_vdef (assign
);
2078 gsi_remove (&gsi
, true);
2079 release_defs (assign
);
2081 /* Make both store and load use alias-set zero as we have to
2082 deal with the case of the store being a conditional change
2083 of the dynamic type. */
2084 lhs
= unshare_expr (lhs
);
2086 while (handled_component_p (*basep
))
2087 basep
= &TREE_OPERAND (*basep
, 0);
2088 if (TREE_CODE (*basep
) == MEM_REF
2089 || TREE_CODE (*basep
) == TARGET_MEM_REF
)
2090 TREE_OPERAND (*basep
, 1)
2091 = fold_convert (ptr_type_node
, TREE_OPERAND (*basep
, 1));
2093 *basep
= build2 (MEM_REF
, TREE_TYPE (*basep
),
2094 build_fold_addr_expr (*basep
),
2095 build_zero_cst (ptr_type_node
));
2097 /* 2) Insert a load from the memory of the store to the temporary
2098 on the edge which did not contain the store. */
2099 name
= make_temp_ssa_name (TREE_TYPE (lhs
), NULL
, "cstore");
2100 new_stmt
= gimple_build_assign (name
, lhs
);
2101 gimple_set_location (new_stmt
, locus
);
2102 gsi_insert_on_edge (e1
, new_stmt
);
2104 /* 3) Create a PHI node at the join block, with one argument
2105 holding the old RHS, and the other holding the temporary
2106 where we stored the old memory contents. */
2107 name2
= make_temp_ssa_name (TREE_TYPE (lhs
), NULL
, "cstore");
2108 newphi
= create_phi_node (name2
, join_bb
);
2109 add_phi_arg (newphi
, rhs
, e0
, locus
);
2110 add_phi_arg (newphi
, name
, e1
, locus
);
2112 lhs
= unshare_expr (lhs
);
2113 new_stmt
= gimple_build_assign (lhs
, PHI_RESULT (newphi
));
2115 /* 4) Insert that PHI node. */
2116 gsi
= gsi_after_labels (join_bb
);
2117 if (gsi_end_p (gsi
))
2119 gsi
= gsi_last_bb (join_bb
);
2120 gsi_insert_after (&gsi
, new_stmt
, GSI_NEW_STMT
);
2123 gsi_insert_before (&gsi
, new_stmt
, GSI_NEW_STMT
);
2128 /* Do the main work of conditional store replacement. */
2131 cond_if_else_store_replacement_1 (basic_block then_bb
, basic_block else_bb
,
2132 basic_block join_bb
, gimple
*then_assign
,
2133 gimple
*else_assign
)
2135 tree lhs_base
, lhs
, then_rhs
, else_rhs
, name
;
2136 source_location then_locus
, else_locus
;
2137 gimple_stmt_iterator gsi
;
2141 if (then_assign
== NULL
2142 || !gimple_assign_single_p (then_assign
)
2143 || gimple_clobber_p (then_assign
)
2144 || gimple_has_volatile_ops (then_assign
)
2145 || else_assign
== NULL
2146 || !gimple_assign_single_p (else_assign
)
2147 || gimple_clobber_p (else_assign
)
2148 || gimple_has_volatile_ops (else_assign
))
2151 lhs
= gimple_assign_lhs (then_assign
);
2152 if (!is_gimple_reg_type (TREE_TYPE (lhs
))
2153 || !operand_equal_p (lhs
, gimple_assign_lhs (else_assign
), 0))
2156 lhs_base
= get_base_address (lhs
);
2157 if (lhs_base
== NULL_TREE
2158 || (!DECL_P (lhs_base
) && TREE_CODE (lhs_base
) != MEM_REF
))
2161 then_rhs
= gimple_assign_rhs1 (then_assign
);
2162 else_rhs
= gimple_assign_rhs1 (else_assign
);
2163 then_locus
= gimple_location (then_assign
);
2164 else_locus
= gimple_location (else_assign
);
2166 /* Now we've checked the constraints, so do the transformation:
2167 1) Remove the stores. */
2168 gsi
= gsi_for_stmt (then_assign
);
2169 unlink_stmt_vdef (then_assign
);
2170 gsi_remove (&gsi
, true);
2171 release_defs (then_assign
);
2173 gsi
= gsi_for_stmt (else_assign
);
2174 unlink_stmt_vdef (else_assign
);
2175 gsi_remove (&gsi
, true);
2176 release_defs (else_assign
);
2178 /* 2) Create a PHI node at the join block, with one argument
2179 holding the old RHS, and the other holding the temporary
2180 where we stored the old memory contents. */
2181 name
= make_temp_ssa_name (TREE_TYPE (lhs
), NULL
, "cstore");
2182 newphi
= create_phi_node (name
, join_bb
);
2183 add_phi_arg (newphi
, then_rhs
, EDGE_SUCC (then_bb
, 0), then_locus
);
2184 add_phi_arg (newphi
, else_rhs
, EDGE_SUCC (else_bb
, 0), else_locus
);
2186 new_stmt
= gimple_build_assign (lhs
, PHI_RESULT (newphi
));
2188 /* 3) Insert that PHI node. */
2189 gsi
= gsi_after_labels (join_bb
);
2190 if (gsi_end_p (gsi
))
2192 gsi
= gsi_last_bb (join_bb
);
2193 gsi_insert_after (&gsi
, new_stmt
, GSI_NEW_STMT
);
2196 gsi_insert_before (&gsi
, new_stmt
, GSI_NEW_STMT
);
2201 /* Return the single store in BB with VDEF or NULL if there are
2202 other stores in the BB or loads following the store. */
2205 single_trailing_store_in_bb (basic_block bb
, tree vdef
)
2207 if (SSA_NAME_IS_DEFAULT_DEF (vdef
))
2209 gimple
*store
= SSA_NAME_DEF_STMT (vdef
);
2210 if (gimple_bb (store
) != bb
2211 || gimple_code (store
) == GIMPLE_PHI
)
2214 /* Verify there is no other store in this BB. */
2215 if (!SSA_NAME_IS_DEFAULT_DEF (gimple_vuse (store
))
2216 && gimple_bb (SSA_NAME_DEF_STMT (gimple_vuse (store
))) == bb
2217 && gimple_code (SSA_NAME_DEF_STMT (gimple_vuse (store
))) != GIMPLE_PHI
)
2220 /* Verify there is no load or store after the store. */
2221 use_operand_p use_p
;
2222 imm_use_iterator imm_iter
;
2223 FOR_EACH_IMM_USE_FAST (use_p
, imm_iter
, gimple_vdef (store
))
2224 if (USE_STMT (use_p
) != store
2225 && gimple_bb (USE_STMT (use_p
)) == bb
)
2231 /* Conditional store replacement. We already know
2232 that the recognized pattern looks like so:
2235 if (cond) goto THEN_BB; else goto ELSE_BB (edge E1)
2245 fallthrough (edge E0)
2249 We check that it is safe to sink the store to JOIN_BB by verifying that
2250 there are no read-after-write or write-after-write dependencies in
2251 THEN_BB and ELSE_BB. */
2254 cond_if_else_store_replacement (basic_block then_bb
, basic_block else_bb
,
2255 basic_block join_bb
)
2257 vec
<data_reference_p
> then_datarefs
, else_datarefs
;
2258 vec
<ddr_p
> then_ddrs
, else_ddrs
;
2259 gimple
*then_store
, *else_store
;
2260 bool found
, ok
= false, res
;
2261 struct data_dependence_relation
*ddr
;
2262 data_reference_p then_dr
, else_dr
;
2264 tree then_lhs
, else_lhs
;
2265 basic_block blocks
[3];
2267 /* Handle the case with single store in THEN_BB and ELSE_BB. That is
2268 cheap enough to always handle as it allows us to elide dependence
2271 for (gphi_iterator si
= gsi_start_phis (join_bb
); !gsi_end_p (si
);
2273 if (virtual_operand_p (gimple_phi_result (si
.phi ())))
2280 tree then_vdef
= PHI_ARG_DEF_FROM_EDGE (vphi
, single_succ_edge (then_bb
));
2281 tree else_vdef
= PHI_ARG_DEF_FROM_EDGE (vphi
, single_succ_edge (else_bb
));
2282 gimple
*then_assign
= single_trailing_store_in_bb (then_bb
, then_vdef
);
2285 gimple
*else_assign
= single_trailing_store_in_bb (else_bb
, else_vdef
);
2287 return cond_if_else_store_replacement_1 (then_bb
, else_bb
, join_bb
,
2288 then_assign
, else_assign
);
2291 if (MAX_STORES_TO_SINK
== 0)
2294 /* Find data references. */
2295 then_datarefs
.create (1);
2296 else_datarefs
.create (1);
2297 if ((find_data_references_in_bb (NULL
, then_bb
, &then_datarefs
)
2299 || !then_datarefs
.length ()
2300 || (find_data_references_in_bb (NULL
, else_bb
, &else_datarefs
)
2302 || !else_datarefs
.length ())
2304 free_data_refs (then_datarefs
);
2305 free_data_refs (else_datarefs
);
2309 /* Find pairs of stores with equal LHS. */
2310 auto_vec
<gimple
*, 1> then_stores
, else_stores
;
2311 FOR_EACH_VEC_ELT (then_datarefs
, i
, then_dr
)
2313 if (DR_IS_READ (then_dr
))
2316 then_store
= DR_STMT (then_dr
);
2317 then_lhs
= gimple_get_lhs (then_store
);
2318 if (then_lhs
== NULL_TREE
)
2322 FOR_EACH_VEC_ELT (else_datarefs
, j
, else_dr
)
2324 if (DR_IS_READ (else_dr
))
2327 else_store
= DR_STMT (else_dr
);
2328 else_lhs
= gimple_get_lhs (else_store
);
2329 if (else_lhs
== NULL_TREE
)
2332 if (operand_equal_p (then_lhs
, else_lhs
, 0))
2342 then_stores
.safe_push (then_store
);
2343 else_stores
.safe_push (else_store
);
2346 /* No pairs of stores found. */
2347 if (!then_stores
.length ()
2348 || then_stores
.length () > (unsigned) MAX_STORES_TO_SINK
)
2350 free_data_refs (then_datarefs
);
2351 free_data_refs (else_datarefs
);
2355 /* Compute and check data dependencies in both basic blocks. */
2356 then_ddrs
.create (1);
2357 else_ddrs
.create (1);
2358 if (!compute_all_dependences (then_datarefs
, &then_ddrs
,
2360 || !compute_all_dependences (else_datarefs
, &else_ddrs
,
2363 free_dependence_relations (then_ddrs
);
2364 free_dependence_relations (else_ddrs
);
2365 free_data_refs (then_datarefs
);
2366 free_data_refs (else_datarefs
);
2369 blocks
[0] = then_bb
;
2370 blocks
[1] = else_bb
;
2371 blocks
[2] = join_bb
;
2372 renumber_gimple_stmt_uids_in_blocks (blocks
, 3);
2374 /* Check that there are no read-after-write or write-after-write dependencies
2376 FOR_EACH_VEC_ELT (then_ddrs
, i
, ddr
)
2378 struct data_reference
*dra
= DDR_A (ddr
);
2379 struct data_reference
*drb
= DDR_B (ddr
);
2381 if (DDR_ARE_DEPENDENT (ddr
) != chrec_known
2382 && ((DR_IS_READ (dra
) && DR_IS_WRITE (drb
)
2383 && gimple_uid (DR_STMT (dra
)) > gimple_uid (DR_STMT (drb
)))
2384 || (DR_IS_READ (drb
) && DR_IS_WRITE (dra
)
2385 && gimple_uid (DR_STMT (drb
)) > gimple_uid (DR_STMT (dra
)))
2386 || (DR_IS_WRITE (dra
) && DR_IS_WRITE (drb
))))
2388 free_dependence_relations (then_ddrs
);
2389 free_dependence_relations (else_ddrs
);
2390 free_data_refs (then_datarefs
);
2391 free_data_refs (else_datarefs
);
2396 /* Check that there are no read-after-write or write-after-write dependencies
2398 FOR_EACH_VEC_ELT (else_ddrs
, i
, ddr
)
2400 struct data_reference
*dra
= DDR_A (ddr
);
2401 struct data_reference
*drb
= DDR_B (ddr
);
2403 if (DDR_ARE_DEPENDENT (ddr
) != chrec_known
2404 && ((DR_IS_READ (dra
) && DR_IS_WRITE (drb
)
2405 && gimple_uid (DR_STMT (dra
)) > gimple_uid (DR_STMT (drb
)))
2406 || (DR_IS_READ (drb
) && DR_IS_WRITE (dra
)
2407 && gimple_uid (DR_STMT (drb
)) > gimple_uid (DR_STMT (dra
)))
2408 || (DR_IS_WRITE (dra
) && DR_IS_WRITE (drb
))))
2410 free_dependence_relations (then_ddrs
);
2411 free_dependence_relations (else_ddrs
);
2412 free_data_refs (then_datarefs
);
2413 free_data_refs (else_datarefs
);
2418 /* Sink stores with same LHS. */
2419 FOR_EACH_VEC_ELT (then_stores
, i
, then_store
)
2421 else_store
= else_stores
[i
];
2422 res
= cond_if_else_store_replacement_1 (then_bb
, else_bb
, join_bb
,
2423 then_store
, else_store
);
2427 free_dependence_relations (then_ddrs
);
2428 free_dependence_relations (else_ddrs
);
2429 free_data_refs (then_datarefs
);
2430 free_data_refs (else_datarefs
);
2435 /* Return TRUE if STMT has a VUSE whose corresponding VDEF is in BB. */
2438 local_mem_dependence (gimple
*stmt
, basic_block bb
)
2440 tree vuse
= gimple_vuse (stmt
);
2446 def
= SSA_NAME_DEF_STMT (vuse
);
2447 return (def
&& gimple_bb (def
) == bb
);
2450 /* Given a "diamond" control-flow pattern where BB0 tests a condition,
2451 BB1 and BB2 are "then" and "else" blocks dependent on this test,
2452 and BB3 rejoins control flow following BB1 and BB2, look for
2453 opportunities to hoist loads as follows. If BB3 contains a PHI of
2454 two loads, one each occurring in BB1 and BB2, and the loads are
2455 provably of adjacent fields in the same structure, then move both
2456 loads into BB0. Of course this can only be done if there are no
2457 dependencies preventing such motion.
2459 One of the hoisted loads will always be speculative, so the
2460 transformation is currently conservative:
2462 - The fields must be strictly adjacent.
2463 - The two fields must occupy a single memory block that is
2464 guaranteed to not cross a page boundary.
2466 The last is difficult to prove, as such memory blocks should be
2467 aligned on the minimum of the stack alignment boundary and the
2468 alignment guaranteed by heap allocation interfaces. Thus we rely
2469 on a parameter for the alignment value.
2471 Provided a good value is used for the last case, the first
2472 restriction could possibly be relaxed. */
2475 hoist_adjacent_loads (basic_block bb0
, basic_block bb1
,
2476 basic_block bb2
, basic_block bb3
)
2478 int param_align
= PARAM_VALUE (PARAM_L1_CACHE_LINE_SIZE
);
2479 unsigned param_align_bits
= (unsigned) (param_align
* BITS_PER_UNIT
);
2482 /* Walk the phis in bb3 looking for an opportunity. We are looking
2483 for phis of two SSA names, one each of which is defined in bb1 and
2485 for (gsi
= gsi_start_phis (bb3
); !gsi_end_p (gsi
); gsi_next (&gsi
))
2487 gphi
*phi_stmt
= gsi
.phi ();
2488 gimple
*def1
, *def2
;
2489 tree arg1
, arg2
, ref1
, ref2
, field1
, field2
;
2490 tree tree_offset1
, tree_offset2
, tree_size2
, next
;
2491 int offset1
, offset2
, size2
;
2493 gimple_stmt_iterator gsi2
;
2494 basic_block bb_for_def1
, bb_for_def2
;
2496 if (gimple_phi_num_args (phi_stmt
) != 2
2497 || virtual_operand_p (gimple_phi_result (phi_stmt
)))
2500 arg1
= gimple_phi_arg_def (phi_stmt
, 0);
2501 arg2
= gimple_phi_arg_def (phi_stmt
, 1);
2503 if (TREE_CODE (arg1
) != SSA_NAME
2504 || TREE_CODE (arg2
) != SSA_NAME
2505 || SSA_NAME_IS_DEFAULT_DEF (arg1
)
2506 || SSA_NAME_IS_DEFAULT_DEF (arg2
))
2509 def1
= SSA_NAME_DEF_STMT (arg1
);
2510 def2
= SSA_NAME_DEF_STMT (arg2
);
2512 if ((gimple_bb (def1
) != bb1
|| gimple_bb (def2
) != bb2
)
2513 && (gimple_bb (def2
) != bb1
|| gimple_bb (def1
) != bb2
))
2516 /* Check the mode of the arguments to be sure a conditional move
2517 can be generated for it. */
2518 if (optab_handler (movcc_optab
, TYPE_MODE (TREE_TYPE (arg1
)))
2519 == CODE_FOR_nothing
)
2522 /* Both statements must be assignments whose RHS is a COMPONENT_REF. */
2523 if (!gimple_assign_single_p (def1
)
2524 || !gimple_assign_single_p (def2
)
2525 || gimple_has_volatile_ops (def1
)
2526 || gimple_has_volatile_ops (def2
))
2529 ref1
= gimple_assign_rhs1 (def1
);
2530 ref2
= gimple_assign_rhs1 (def2
);
2532 if (TREE_CODE (ref1
) != COMPONENT_REF
2533 || TREE_CODE (ref2
) != COMPONENT_REF
)
2536 /* The zeroth operand of the two component references must be
2537 identical. It is not sufficient to compare get_base_address of
2538 the two references, because this could allow for different
2539 elements of the same array in the two trees. It is not safe to
2540 assume that the existence of one array element implies the
2541 existence of a different one. */
2542 if (!operand_equal_p (TREE_OPERAND (ref1
, 0), TREE_OPERAND (ref2
, 0), 0))
2545 field1
= TREE_OPERAND (ref1
, 1);
2546 field2
= TREE_OPERAND (ref2
, 1);
2548 /* Check for field adjacency, and ensure field1 comes first. */
2549 for (next
= DECL_CHAIN (field1
);
2550 next
&& TREE_CODE (next
) != FIELD_DECL
;
2551 next
= DECL_CHAIN (next
))
2556 for (next
= DECL_CHAIN (field2
);
2557 next
&& TREE_CODE (next
) != FIELD_DECL
;
2558 next
= DECL_CHAIN (next
))
2564 std::swap (field1
, field2
);
2565 std::swap (def1
, def2
);
2568 bb_for_def1
= gimple_bb (def1
);
2569 bb_for_def2
= gimple_bb (def2
);
2571 /* Check for proper alignment of the first field. */
2572 tree_offset1
= bit_position (field1
);
2573 tree_offset2
= bit_position (field2
);
2574 tree_size2
= DECL_SIZE (field2
);
2576 if (!tree_fits_uhwi_p (tree_offset1
)
2577 || !tree_fits_uhwi_p (tree_offset2
)
2578 || !tree_fits_uhwi_p (tree_size2
))
2581 offset1
= tree_to_uhwi (tree_offset1
);
2582 offset2
= tree_to_uhwi (tree_offset2
);
2583 size2
= tree_to_uhwi (tree_size2
);
2584 align1
= DECL_ALIGN (field1
) % param_align_bits
;
2586 if (offset1
% BITS_PER_UNIT
!= 0)
2589 /* For profitability, the two field references should fit within
2590 a single cache line. */
2591 if (align1
+ offset2
- offset1
+ size2
> param_align_bits
)
2594 /* The two expressions cannot be dependent upon vdefs defined
2596 if (local_mem_dependence (def1
, bb_for_def1
)
2597 || local_mem_dependence (def2
, bb_for_def2
))
2600 /* The conditions are satisfied; hoist the loads from bb1 and bb2 into
2601 bb0. We hoist the first one first so that a cache miss is handled
2602 efficiently regardless of hardware cache-fill policy. */
2603 gsi2
= gsi_for_stmt (def1
);
2604 gsi_move_to_bb_end (&gsi2
, bb0
);
2605 gsi2
= gsi_for_stmt (def2
);
2606 gsi_move_to_bb_end (&gsi2
, bb0
);
2608 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
2611 "\nHoisting adjacent loads from %d and %d into %d: \n",
2612 bb_for_def1
->index
, bb_for_def2
->index
, bb0
->index
);
2613 print_gimple_stmt (dump_file
, def1
, 0, TDF_VOPS
|TDF_MEMSYMS
);
2614 print_gimple_stmt (dump_file
, def2
, 0, TDF_VOPS
|TDF_MEMSYMS
);
2619 /* Determine whether we should attempt to hoist adjacent loads out of
2620 diamond patterns in pass_phiopt. Always hoist loads if
2621 -fhoist-adjacent-loads is specified and the target machine has
2622 both a conditional move instruction and a defined cache line size. */
2625 gate_hoist_loads (void)
2627 return (flag_hoist_adjacent_loads
== 1
2628 && PARAM_VALUE (PARAM_L1_CACHE_LINE_SIZE
)
2629 && HAVE_conditional_move
);
2632 /* This pass tries to replaces an if-then-else block with an
2633 assignment. We have four kinds of transformations. Some of these
2634 transformations are also performed by the ifcvt RTL optimizer.
2636 Conditional Replacement
2637 -----------------------
2639 This transformation, implemented in conditional_replacement,
2643 if (cond) goto bb2; else goto bb1;
2646 x = PHI <0 (bb1), 1 (bb0), ...>;
2654 x = PHI <x' (bb0), ...>;
2656 We remove bb1 as it becomes unreachable. This occurs often due to
2657 gimplification of conditionals.
2662 This transformation, implemented in value_replacement, replaces
2665 if (a != b) goto bb2; else goto bb1;
2668 x = PHI <a (bb1), b (bb0), ...>;
2674 x = PHI <b (bb0), ...>;
2676 This opportunity can sometimes occur as a result of other
2680 Another case caught by value replacement looks like this:
2686 if (t3 != 0) goto bb1; else goto bb2;
2702 This transformation, implemented in abs_replacement, replaces
2705 if (a >= 0) goto bb2; else goto bb1;
2709 x = PHI <x (bb1), a (bb0), ...>;
2716 x = PHI <x' (bb0), ...>;
2721 This transformation, minmax_replacement replaces
2724 if (a <= b) goto bb2; else goto bb1;
2727 x = PHI <b (bb1), a (bb0), ...>;
2732 x' = MIN_EXPR (a, b)
2734 x = PHI <x' (bb0), ...>;
2736 A similar transformation is done for MAX_EXPR.
2739 This pass also performs a fifth transformation of a slightly different
2742 Factor conversion in COND_EXPR
2743 ------------------------------
2745 This transformation factors the conversion out of COND_EXPR with
2746 factor_out_conditional_conversion.
2749 if (a <= CST) goto <bb 3>; else goto <bb 4>;
2753 tmp = PHI <tmp, CST>
2756 if (a <= CST) goto <bb 3>; else goto <bb 4>;
2762 Adjacent Load Hoisting
2763 ----------------------
2765 This transformation replaces
2768 if (...) goto bb2; else goto bb1;
2770 x1 = (<expr>).field1;
2773 x2 = (<expr>).field2;
2780 x1 = (<expr>).field1;
2781 x2 = (<expr>).field2;
2782 if (...) goto bb2; else goto bb1;
2789 The purpose of this transformation is to enable generation of conditional
2790 move instructions such as Intel CMOVE or PowerPC ISEL. Because one of
2791 the loads is speculative, the transformation is restricted to very
2792 specific cases to avoid introducing a page fault. We are looking for
2800 where left and right are typically adjacent pointers in a tree structure. */
2804 const pass_data pass_data_phiopt
=
2806 GIMPLE_PASS
, /* type */
2807 "phiopt", /* name */
2808 OPTGROUP_NONE
, /* optinfo_flags */
2809 TV_TREE_PHIOPT
, /* tv_id */
2810 ( PROP_cfg
| PROP_ssa
), /* properties_required */
2811 0, /* properties_provided */
2812 0, /* properties_destroyed */
2813 0, /* todo_flags_start */
2814 0, /* todo_flags_finish */
2817 class pass_phiopt
: public gimple_opt_pass
2820 pass_phiopt (gcc::context
*ctxt
)
2821 : gimple_opt_pass (pass_data_phiopt
, ctxt
), early_p (false)
2824 /* opt_pass methods: */
2825 opt_pass
* clone () { return new pass_phiopt (m_ctxt
); }
2826 void set_pass_param (unsigned n
, bool param
)
2828 gcc_assert (n
== 0);
2831 virtual bool gate (function
*) { return flag_ssa_phiopt
; }
2832 virtual unsigned int execute (function
*)
2834 return tree_ssa_phiopt_worker (false,
2835 !early_p
? gate_hoist_loads () : false,
2841 }; // class pass_phiopt
2846 make_pass_phiopt (gcc::context
*ctxt
)
2848 return new pass_phiopt (ctxt
);
2853 const pass_data pass_data_cselim
=
2855 GIMPLE_PASS
, /* type */
2856 "cselim", /* name */
2857 OPTGROUP_NONE
, /* optinfo_flags */
2858 TV_TREE_PHIOPT
, /* tv_id */
2859 ( PROP_cfg
| PROP_ssa
), /* properties_required */
2860 0, /* properties_provided */
2861 0, /* properties_destroyed */
2862 0, /* todo_flags_start */
2863 0, /* todo_flags_finish */
2866 class pass_cselim
: public gimple_opt_pass
2869 pass_cselim (gcc::context
*ctxt
)
2870 : gimple_opt_pass (pass_data_cselim
, ctxt
)
2873 /* opt_pass methods: */
2874 virtual bool gate (function
*) { return flag_tree_cselim
; }
2875 virtual unsigned int execute (function
*) { return tree_ssa_cs_elim (); }
2877 }; // class pass_cselim
2882 make_pass_cselim (gcc::context
*ctxt
)
2884 return new pass_cselim (ctxt
);