1 /* Optimization of PHI nodes by converting them into straightline code.
2 Copyright (C) 2004-2019 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 two_value_replacement (basic_block
, basic_block
, edge
, gphi
*,
53 static bool conditional_replacement (basic_block
, basic_block
,
54 edge
, edge
, gphi
*, tree
, tree
);
55 static gphi
*factor_out_conditional_conversion (edge
, edge
, gphi
*, tree
, tree
,
57 static int value_replacement (basic_block
, basic_block
,
58 edge
, edge
, gimple
*, tree
, tree
);
59 static bool minmax_replacement (basic_block
, basic_block
,
60 edge
, edge
, gimple
*, tree
, tree
);
61 static bool abs_replacement (basic_block
, basic_block
,
62 edge
, edge
, gimple
*, tree
, tree
);
63 static bool cond_removal_in_popcount_pattern (basic_block
, basic_block
,
64 edge
, edge
, gimple
*, tree
, tree
);
65 static bool cond_store_replacement (basic_block
, basic_block
, edge
, edge
,
67 static bool cond_if_else_store_replacement (basic_block
, basic_block
, basic_block
);
68 static hash_set
<tree
> * get_non_trapping ();
69 static void replace_phi_edge_with_variable (basic_block
, edge
, gimple
*, tree
);
70 static void hoist_adjacent_loads (basic_block
, basic_block
,
71 basic_block
, basic_block
);
72 static bool gate_hoist_loads (void);
74 /* This pass tries to transform conditional stores into unconditional
75 ones, enabling further simplifications with the simpler then and else
76 blocks. In particular it replaces this:
79 if (cond) goto bb2; else goto bb1;
87 if (cond) goto bb1; else goto bb2;
91 condtmp = PHI <RHS, condtmp'>
94 This transformation can only be done under several constraints,
95 documented below. It also replaces:
98 if (cond) goto bb2; else goto bb1;
109 if (cond) goto bb3; else goto bb1;
112 condtmp = PHI <RHS1, RHS2>
116 tree_ssa_cs_elim (void)
119 /* ??? We are not interested in loop related info, but the following
120 will create it, ICEing as we didn't init loops with pre-headers.
121 An interfacing issue of find_data_references_in_bb. */
122 loop_optimizer_init (LOOPS_NORMAL
);
124 todo
= tree_ssa_phiopt_worker (true, false, false);
126 loop_optimizer_finalize ();
130 /* Return the singleton PHI in the SEQ of PHIs for edges E0 and E1. */
133 single_non_singleton_phi_for_edges (gimple_seq seq
, edge e0
, edge e1
)
135 gimple_stmt_iterator i
;
137 if (gimple_seq_singleton_p (seq
))
138 return as_a
<gphi
*> (gsi_stmt (gsi_start (seq
)));
139 for (i
= gsi_start (seq
); !gsi_end_p (i
); gsi_next (&i
))
141 gphi
*p
= as_a
<gphi
*> (gsi_stmt (i
));
142 /* If the PHI arguments are equal then we can skip this PHI. */
143 if (operand_equal_for_phi_arg_p (gimple_phi_arg_def (p
, e0
->dest_idx
),
144 gimple_phi_arg_def (p
, e1
->dest_idx
)))
147 /* If we already have a PHI that has the two edge arguments are
148 different, then return it is not a singleton for these PHIs. */
157 /* The core routine of conditional store replacement and normal
158 phi optimizations. Both share much of the infrastructure in how
159 to match applicable basic block patterns. DO_STORE_ELIM is true
160 when we want to do conditional store replacement, false otherwise.
161 DO_HOIST_LOADS is true when we want to hoist adjacent loads out
162 of diamond control flow patterns, false otherwise. */
164 tree_ssa_phiopt_worker (bool do_store_elim
, bool do_hoist_loads
, bool early_p
)
167 basic_block
*bb_order
;
169 bool cfgchanged
= false;
170 hash_set
<tree
> *nontrap
= 0;
173 /* Calculate the set of non-trapping memory accesses. */
174 nontrap
= get_non_trapping ();
176 /* Search every basic block for COND_EXPR we may be able to optimize.
178 We walk the blocks in order that guarantees that a block with
179 a single predecessor is processed before the predecessor.
180 This ensures that we collapse inner ifs before visiting the
181 outer ones, and also that we do not try to visit a removed
183 bb_order
= single_pred_before_succ_order ();
184 n
= n_basic_blocks_for_fn (cfun
) - NUM_FIXED_BLOCKS
;
186 for (i
= 0; i
< n
; i
++)
190 basic_block bb1
, bb2
;
196 cond_stmt
= last_stmt (bb
);
197 /* Check to see if the last statement is a GIMPLE_COND. */
199 || gimple_code (cond_stmt
) != GIMPLE_COND
)
202 e1
= EDGE_SUCC (bb
, 0);
204 e2
= EDGE_SUCC (bb
, 1);
207 /* We cannot do the optimization on abnormal edges. */
208 if ((e1
->flags
& EDGE_ABNORMAL
) != 0
209 || (e2
->flags
& EDGE_ABNORMAL
) != 0)
212 /* If either bb1's succ or bb2 or bb2's succ is non NULL. */
213 if (EDGE_COUNT (bb1
->succs
) == 0
215 || EDGE_COUNT (bb2
->succs
) == 0)
218 /* Find the bb which is the fall through to the other. */
219 if (EDGE_SUCC (bb1
, 0)->dest
== bb2
)
221 else if (EDGE_SUCC (bb2
, 0)->dest
== bb1
)
223 std::swap (bb1
, bb2
);
226 else if (do_store_elim
227 && EDGE_SUCC (bb1
, 0)->dest
== EDGE_SUCC (bb2
, 0)->dest
)
229 basic_block bb3
= EDGE_SUCC (bb1
, 0)->dest
;
231 if (!single_succ_p (bb1
)
232 || (EDGE_SUCC (bb1
, 0)->flags
& EDGE_FALLTHRU
) == 0
233 || !single_succ_p (bb2
)
234 || (EDGE_SUCC (bb2
, 0)->flags
& EDGE_FALLTHRU
) == 0
235 || EDGE_COUNT (bb3
->preds
) != 2)
237 if (cond_if_else_store_replacement (bb1
, bb2
, bb3
))
241 else if (do_hoist_loads
242 && EDGE_SUCC (bb1
, 0)->dest
== EDGE_SUCC (bb2
, 0)->dest
)
244 basic_block bb3
= EDGE_SUCC (bb1
, 0)->dest
;
246 if (!FLOAT_TYPE_P (TREE_TYPE (gimple_cond_lhs (cond_stmt
)))
247 && single_succ_p (bb1
)
248 && single_succ_p (bb2
)
249 && single_pred_p (bb1
)
250 && single_pred_p (bb2
)
251 && EDGE_COUNT (bb
->succs
) == 2
252 && EDGE_COUNT (bb3
->preds
) == 2
253 /* If one edge or the other is dominant, a conditional move
254 is likely to perform worse than the well-predicted branch. */
255 && !predictable_edge_p (EDGE_SUCC (bb
, 0))
256 && !predictable_edge_p (EDGE_SUCC (bb
, 1)))
257 hoist_adjacent_loads (bb
, bb1
, bb2
, bb3
);
263 e1
= EDGE_SUCC (bb1
, 0);
265 /* Make sure that bb1 is just a fall through. */
266 if (!single_succ_p (bb1
)
267 || (e1
->flags
& EDGE_FALLTHRU
) == 0)
270 /* Also make sure that bb1 only have one predecessor and that it
272 if (!single_pred_p (bb1
)
273 || single_pred (bb1
) != bb
)
278 /* bb1 is the middle block, bb2 the join block, bb the split block,
279 e1 the fallthrough edge from bb1 to bb2. We can't do the
280 optimization if the join block has more than two predecessors. */
281 if (EDGE_COUNT (bb2
->preds
) > 2)
283 if (cond_store_replacement (bb1
, bb2
, e1
, e2
, nontrap
))
288 gimple_seq phis
= phi_nodes (bb2
);
289 gimple_stmt_iterator gsi
;
290 bool candorest
= true;
292 /* Value replacement can work with more than one PHI
293 so try that first. */
295 for (gsi
= gsi_start (phis
); !gsi_end_p (gsi
); gsi_next (&gsi
))
297 phi
= as_a
<gphi
*> (gsi_stmt (gsi
));
298 arg0
= gimple_phi_arg_def (phi
, e1
->dest_idx
);
299 arg1
= gimple_phi_arg_def (phi
, e2
->dest_idx
);
300 if (value_replacement (bb
, bb1
, e1
, e2
, phi
, arg0
, arg1
) == 2)
311 phi
= single_non_singleton_phi_for_edges (phis
, e1
, e2
);
315 arg0
= gimple_phi_arg_def (phi
, e1
->dest_idx
);
316 arg1
= gimple_phi_arg_def (phi
, e2
->dest_idx
);
318 /* Something is wrong if we cannot find the arguments in the PHI
320 gcc_assert (arg0
!= NULL_TREE
&& arg1
!= NULL_TREE
);
322 gphi
*newphi
= factor_out_conditional_conversion (e1
, e2
, phi
,
328 /* factor_out_conditional_conversion may create a new PHI in
329 BB2 and eliminate an existing PHI in BB2. Recompute values
330 that may be affected by that change. */
331 arg0
= gimple_phi_arg_def (phi
, e1
->dest_idx
);
332 arg1
= gimple_phi_arg_def (phi
, e2
->dest_idx
);
333 gcc_assert (arg0
!= NULL_TREE
&& arg1
!= NULL_TREE
);
336 /* Do the replacement of conditional if it can be done. */
337 if (two_value_replacement (bb
, bb1
, e2
, phi
, arg0
, arg1
))
340 && conditional_replacement (bb
, bb1
, e1
, e2
, phi
,
343 else if (abs_replacement (bb
, bb1
, e1
, e2
, phi
, arg0
, arg1
))
346 && cond_removal_in_popcount_pattern (bb
, bb1
, e1
, e2
,
349 else if (minmax_replacement (bb
, bb1
, e1
, e2
, phi
, arg0
, arg1
))
358 /* If the CFG has changed, we should cleanup the CFG. */
359 if (cfgchanged
&& do_store_elim
)
361 /* In cond-store replacement we have added some loads on edges
362 and new VOPS (as we moved the store, and created a load). */
363 gsi_commit_edge_inserts ();
364 return TODO_cleanup_cfg
| TODO_update_ssa_only_virtuals
;
367 return TODO_cleanup_cfg
;
371 /* Replace PHI node element whose edge is E in block BB with variable NEW.
372 Remove the edge from COND_BLOCK which does not lead to BB (COND_BLOCK
373 is known to have two edges, one of which must reach BB). */
376 replace_phi_edge_with_variable (basic_block cond_block
,
377 edge e
, gimple
*phi
, tree new_tree
)
379 basic_block bb
= gimple_bb (phi
);
380 basic_block block_to_remove
;
381 gimple_stmt_iterator gsi
;
383 /* Change the PHI argument to new. */
384 SET_USE (PHI_ARG_DEF_PTR (phi
, e
->dest_idx
), new_tree
);
386 /* Remove the empty basic block. */
387 if (EDGE_SUCC (cond_block
, 0)->dest
== bb
)
389 EDGE_SUCC (cond_block
, 0)->flags
|= EDGE_FALLTHRU
;
390 EDGE_SUCC (cond_block
, 0)->flags
&= ~(EDGE_TRUE_VALUE
| EDGE_FALSE_VALUE
);
391 EDGE_SUCC (cond_block
, 0)->probability
= profile_probability::always ();
393 block_to_remove
= EDGE_SUCC (cond_block
, 1)->dest
;
397 EDGE_SUCC (cond_block
, 1)->flags
|= EDGE_FALLTHRU
;
398 EDGE_SUCC (cond_block
, 1)->flags
399 &= ~(EDGE_TRUE_VALUE
| EDGE_FALSE_VALUE
);
400 EDGE_SUCC (cond_block
, 1)->probability
= profile_probability::always ();
402 block_to_remove
= EDGE_SUCC (cond_block
, 0)->dest
;
404 delete_basic_block (block_to_remove
);
406 /* Eliminate the COND_EXPR at the end of COND_BLOCK. */
407 gsi
= gsi_last_bb (cond_block
);
408 gsi_remove (&gsi
, true);
410 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
412 "COND_EXPR in block %d and PHI in block %d converted to straightline code.\n",
417 /* PR66726: Factor conversion out of COND_EXPR. If the arguments of the PHI
418 stmt are CONVERT_STMT, factor out the conversion and perform the conversion
419 to the result of PHI stmt. COND_STMT is the controlling predicate.
420 Return the newly-created PHI, if any. */
423 factor_out_conditional_conversion (edge e0
, edge e1
, gphi
*phi
,
424 tree arg0
, tree arg1
, gimple
*cond_stmt
)
426 gimple
*arg0_def_stmt
= NULL
, *arg1_def_stmt
= NULL
, *new_stmt
;
427 tree new_arg0
= NULL_TREE
, new_arg1
= NULL_TREE
;
430 gimple_stmt_iterator gsi
, gsi_for_def
;
431 location_t locus
= gimple_location (phi
);
432 enum tree_code convert_code
;
434 /* Handle only PHI statements with two arguments. TODO: If all
435 other arguments to PHI are INTEGER_CST or if their defining
436 statement have the same unary operation, we can handle more
437 than two arguments too. */
438 if (gimple_phi_num_args (phi
) != 2)
441 /* First canonicalize to simplify tests. */
442 if (TREE_CODE (arg0
) != SSA_NAME
)
444 std::swap (arg0
, arg1
);
448 if (TREE_CODE (arg0
) != SSA_NAME
449 || (TREE_CODE (arg1
) != SSA_NAME
450 && TREE_CODE (arg1
) != INTEGER_CST
))
453 /* Check if arg0 is an SSA_NAME and the stmt which defines arg0 is
455 arg0_def_stmt
= SSA_NAME_DEF_STMT (arg0
);
456 if (!gimple_assign_cast_p (arg0_def_stmt
))
459 /* Use the RHS as new_arg0. */
460 convert_code
= gimple_assign_rhs_code (arg0_def_stmt
);
461 new_arg0
= gimple_assign_rhs1 (arg0_def_stmt
);
462 if (convert_code
== VIEW_CONVERT_EXPR
)
464 new_arg0
= TREE_OPERAND (new_arg0
, 0);
465 if (!is_gimple_reg_type (TREE_TYPE (new_arg0
)))
469 if (TREE_CODE (arg1
) == SSA_NAME
)
471 /* Check if arg1 is an SSA_NAME and the stmt which defines arg1
473 arg1_def_stmt
= SSA_NAME_DEF_STMT (arg1
);
474 if (!is_gimple_assign (arg1_def_stmt
)
475 || gimple_assign_rhs_code (arg1_def_stmt
) != convert_code
)
478 /* Use the RHS as new_arg1. */
479 new_arg1
= gimple_assign_rhs1 (arg1_def_stmt
);
480 if (convert_code
== VIEW_CONVERT_EXPR
)
481 new_arg1
= TREE_OPERAND (new_arg1
, 0);
485 /* If arg1 is an INTEGER_CST, fold it to new type. */
486 if (INTEGRAL_TYPE_P (TREE_TYPE (new_arg0
))
487 && int_fits_type_p (arg1
, TREE_TYPE (new_arg0
)))
489 if (gimple_assign_cast_p (arg0_def_stmt
))
491 /* For the INTEGER_CST case, we are just moving the
492 conversion from one place to another, which can often
493 hurt as the conversion moves further away from the
494 statement that computes the value. So, perform this
495 only if new_arg0 is an operand of COND_STMT, or
496 if arg0_def_stmt is the only non-debug stmt in
497 its basic block, because then it is possible this
498 could enable further optimizations (minmax replacement
499 etc.). See PR71016. */
500 if (new_arg0
!= gimple_cond_lhs (cond_stmt
)
501 && new_arg0
!= gimple_cond_rhs (cond_stmt
)
502 && gimple_bb (arg0_def_stmt
) == e0
->src
)
504 gsi
= gsi_for_stmt (arg0_def_stmt
);
505 gsi_prev_nondebug (&gsi
);
506 if (!gsi_end_p (gsi
))
509 = dyn_cast
<gassign
*> (gsi_stmt (gsi
)))
511 tree lhs
= gimple_assign_lhs (assign
);
512 enum tree_code ass_code
513 = gimple_assign_rhs_code (assign
);
514 if (ass_code
!= MAX_EXPR
&& ass_code
!= MIN_EXPR
)
516 if (lhs
!= gimple_assign_rhs1 (arg0_def_stmt
))
518 gsi_prev_nondebug (&gsi
);
519 if (!gsi_end_p (gsi
))
525 gsi
= gsi_for_stmt (arg0_def_stmt
);
526 gsi_next_nondebug (&gsi
);
527 if (!gsi_end_p (gsi
))
530 new_arg1
= fold_convert (TREE_TYPE (new_arg0
), arg1
);
539 /* If arg0/arg1 have > 1 use, then this transformation actually increases
540 the number of expressions evaluated at runtime. */
541 if (!has_single_use (arg0
)
542 || (arg1_def_stmt
&& !has_single_use (arg1
)))
545 /* If types of new_arg0 and new_arg1 are different bailout. */
546 if (!types_compatible_p (TREE_TYPE (new_arg0
), TREE_TYPE (new_arg1
)))
549 /* Create a new PHI stmt. */
550 result
= PHI_RESULT (phi
);
551 temp
= make_ssa_name (TREE_TYPE (new_arg0
), NULL
);
552 newphi
= create_phi_node (temp
, gimple_bb (phi
));
554 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
556 fprintf (dump_file
, "PHI ");
557 print_generic_expr (dump_file
, gimple_phi_result (phi
));
559 " changed to factor conversion out from COND_EXPR.\n");
560 fprintf (dump_file
, "New stmt with CAST that defines ");
561 print_generic_expr (dump_file
, result
);
562 fprintf (dump_file
, ".\n");
565 /* Remove the old cast(s) that has single use. */
566 gsi_for_def
= gsi_for_stmt (arg0_def_stmt
);
567 gsi_remove (&gsi_for_def
, true);
568 release_defs (arg0_def_stmt
);
572 gsi_for_def
= gsi_for_stmt (arg1_def_stmt
);
573 gsi_remove (&gsi_for_def
, true);
574 release_defs (arg1_def_stmt
);
577 add_phi_arg (newphi
, new_arg0
, e0
, locus
);
578 add_phi_arg (newphi
, new_arg1
, e1
, locus
);
580 /* Create the conversion stmt and insert it. */
581 if (convert_code
== VIEW_CONVERT_EXPR
)
583 temp
= fold_build1 (VIEW_CONVERT_EXPR
, TREE_TYPE (result
), temp
);
584 new_stmt
= gimple_build_assign (result
, temp
);
587 new_stmt
= gimple_build_assign (result
, convert_code
, temp
);
588 gsi
= gsi_after_labels (gimple_bb (phi
));
589 gsi_insert_before (&gsi
, new_stmt
, GSI_SAME_STMT
);
591 /* Remove the original PHI stmt. */
592 gsi
= gsi_for_stmt (phi
);
593 gsi_remove (&gsi
, true);
598 # x_5 in range [cst1, cst2] where cst2 = cst1 + 1
599 if (x_5 op cstN) # where op is == or != and N is 1 or 2
605 # r_6 = PHI<cst3(2), cst4(3)> # where cst3 == cst4 + 1 or cst4 == cst3 + 1
607 to r_6 = x_5 + (min (cst3, cst4) - cst1) or
608 r_6 = (min (cst3, cst4) + cst1) - x_5 depending on op, N and which
609 of cst3 and cst4 is smaller. */
612 two_value_replacement (basic_block cond_bb
, basic_block middle_bb
,
613 edge e1
, gphi
*phi
, tree arg0
, tree arg1
)
615 /* Only look for adjacent integer constants. */
616 if (!INTEGRAL_TYPE_P (TREE_TYPE (arg0
))
617 || !INTEGRAL_TYPE_P (TREE_TYPE (arg1
))
618 || TREE_CODE (arg0
) != INTEGER_CST
619 || TREE_CODE (arg1
) != INTEGER_CST
620 || (tree_int_cst_lt (arg0
, arg1
)
621 ? wi::to_widest (arg0
) + 1 != wi::to_widest (arg1
)
622 : wi::to_widest (arg1
) + 1 != wi::to_widest (arg0
)))
625 if (!empty_block_p (middle_bb
))
628 gimple
*stmt
= last_stmt (cond_bb
);
629 tree lhs
= gimple_cond_lhs (stmt
);
630 tree rhs
= gimple_cond_rhs (stmt
);
632 if (TREE_CODE (lhs
) != SSA_NAME
633 || !INTEGRAL_TYPE_P (TREE_TYPE (lhs
))
634 || TREE_CODE (TREE_TYPE (lhs
)) == BOOLEAN_TYPE
635 || TREE_CODE (rhs
) != INTEGER_CST
)
638 switch (gimple_cond_code (stmt
))
648 if (get_range_info (lhs
, &min
, &max
) != VR_RANGE
650 || (wi::to_wide (rhs
) != min
651 && wi::to_wide (rhs
) != max
))
654 /* We need to know which is the true edge and which is the false
655 edge so that we know when to invert the condition below. */
656 edge true_edge
, false_edge
;
657 extract_true_false_edges_from_block (cond_bb
, &true_edge
, &false_edge
);
658 if ((gimple_cond_code (stmt
) == EQ_EXPR
)
659 ^ (wi::to_wide (rhs
) == max
)
660 ^ (e1
== false_edge
))
661 std::swap (arg0
, arg1
);
664 if (TYPE_PRECISION (TREE_TYPE (lhs
)) == TYPE_PRECISION (TREE_TYPE (arg0
)))
666 /* Avoid performing the arithmetics in bool type which has different
667 semantics, otherwise prefer unsigned types from the two with
668 the same precision. */
669 if (TREE_CODE (TREE_TYPE (arg0
)) == BOOLEAN_TYPE
670 || !TYPE_UNSIGNED (TREE_TYPE (arg0
)))
671 type
= TREE_TYPE (lhs
);
673 type
= TREE_TYPE (arg0
);
675 else if (TYPE_PRECISION (TREE_TYPE (lhs
)) > TYPE_PRECISION (TREE_TYPE (arg0
)))
676 type
= TREE_TYPE (lhs
);
678 type
= TREE_TYPE (arg0
);
680 min
= wide_int::from (min
, TYPE_PRECISION (type
),
681 TYPE_SIGN (TREE_TYPE (lhs
)));
682 wide_int a
= wide_int::from (wi::to_wide (arg0
), TYPE_PRECISION (type
),
683 TYPE_SIGN (TREE_TYPE (arg0
)));
685 wi::overflow_type ovf
;
686 if (tree_int_cst_lt (arg0
, arg1
))
690 if (!TYPE_UNSIGNED (type
))
692 /* lhs is known to be in range [min, min+1] and we want to add a
693 to it. Check if that operation can overflow for those 2 values
694 and if yes, force unsigned type. */
695 wi::add (min
+ (wi::neg_p (a
) ? 0 : 1), a
, SIGNED
, &ovf
);
697 type
= unsigned_type_for (type
);
704 if (!TYPE_UNSIGNED (type
))
706 /* lhs is known to be in range [min, min+1] and we want to subtract
707 it from a. Check if that operation can overflow for those 2
708 values and if yes, force unsigned type. */
709 wi::sub (a
, min
+ (wi::neg_p (min
) ? 0 : 1), SIGNED
, &ovf
);
711 type
= unsigned_type_for (type
);
715 tree arg
= wide_int_to_tree (type
, a
);
716 gimple_stmt_iterator gsi
= gsi_for_stmt (stmt
);
717 if (!useless_type_conversion_p (type
, TREE_TYPE (lhs
)))
718 lhs
= gimplify_build1 (&gsi
, NOP_EXPR
, type
, lhs
);
720 if (code
== PLUS_EXPR
)
721 new_rhs
= gimplify_build2 (&gsi
, PLUS_EXPR
, type
, lhs
, arg
);
723 new_rhs
= gimplify_build2 (&gsi
, MINUS_EXPR
, type
, arg
, lhs
);
724 if (!useless_type_conversion_p (TREE_TYPE (arg0
), type
))
725 new_rhs
= gimplify_build1 (&gsi
, NOP_EXPR
, TREE_TYPE (arg0
), new_rhs
);
727 replace_phi_edge_with_variable (cond_bb
, e1
, phi
, new_rhs
);
729 /* Note that we optimized this PHI. */
733 /* The function conditional_replacement does the main work of doing the
734 conditional replacement. Return true if the replacement is done.
735 Otherwise return false.
736 BB is the basic block where the replacement is going to be done on. ARG0
737 is argument 0 from PHI. Likewise for ARG1. */
740 conditional_replacement (basic_block cond_bb
, basic_block middle_bb
,
741 edge e0
, edge e1
, gphi
*phi
,
742 tree arg0
, tree arg1
)
748 gimple_stmt_iterator gsi
;
749 edge true_edge
, false_edge
;
750 tree new_var
, new_var2
;
753 /* FIXME: Gimplification of complex type is too hard for now. */
754 /* We aren't prepared to handle vectors either (and it is a question
755 if it would be worthwhile anyway). */
756 if (!(INTEGRAL_TYPE_P (TREE_TYPE (arg0
))
757 || POINTER_TYPE_P (TREE_TYPE (arg0
)))
758 || !(INTEGRAL_TYPE_P (TREE_TYPE (arg1
))
759 || POINTER_TYPE_P (TREE_TYPE (arg1
))))
762 /* The PHI arguments have the constants 0 and 1, or 0 and -1, then
763 convert it to the conditional. */
764 if ((integer_zerop (arg0
) && integer_onep (arg1
))
765 || (integer_zerop (arg1
) && integer_onep (arg0
)))
767 else if ((integer_zerop (arg0
) && integer_all_onesp (arg1
))
768 || (integer_zerop (arg1
) && integer_all_onesp (arg0
)))
773 if (!empty_block_p (middle_bb
))
776 /* At this point we know we have a GIMPLE_COND with two successors.
777 One successor is BB, the other successor is an empty block which
778 falls through into BB.
780 There is a single PHI node at the join point (BB) and its arguments
781 are constants (0, 1) or (0, -1).
783 So, given the condition COND, and the two PHI arguments, we can
784 rewrite this PHI into non-branching code:
786 dest = (COND) or dest = COND'
788 We use the condition as-is if the argument associated with the
789 true edge has the value one or the argument associated with the
790 false edge as the value zero. Note that those conditions are not
791 the same since only one of the outgoing edges from the GIMPLE_COND
792 will directly reach BB and thus be associated with an argument. */
794 stmt
= last_stmt (cond_bb
);
795 result
= PHI_RESULT (phi
);
797 /* To handle special cases like floating point comparison, it is easier and
798 less error-prone to build a tree and gimplify it on the fly though it is
800 cond
= fold_build2_loc (gimple_location (stmt
),
801 gimple_cond_code (stmt
), boolean_type_node
,
802 gimple_cond_lhs (stmt
), gimple_cond_rhs (stmt
));
804 /* We need to know which is the true edge and which is the false
805 edge so that we know when to invert the condition below. */
806 extract_true_false_edges_from_block (cond_bb
, &true_edge
, &false_edge
);
807 if ((e0
== true_edge
&& integer_zerop (arg0
))
808 || (e0
== false_edge
&& !integer_zerop (arg0
))
809 || (e1
== true_edge
&& integer_zerop (arg1
))
810 || (e1
== false_edge
&& !integer_zerop (arg1
)))
811 cond
= fold_build1_loc (gimple_location (stmt
),
812 TRUTH_NOT_EXPR
, TREE_TYPE (cond
), cond
);
816 cond
= fold_convert_loc (gimple_location (stmt
),
817 TREE_TYPE (result
), cond
);
818 cond
= fold_build1_loc (gimple_location (stmt
),
819 NEGATE_EXPR
, TREE_TYPE (cond
), cond
);
822 /* Insert our new statements at the end of conditional block before the
824 gsi
= gsi_for_stmt (stmt
);
825 new_var
= force_gimple_operand_gsi (&gsi
, cond
, true, NULL
, true,
828 if (!useless_type_conversion_p (TREE_TYPE (result
), TREE_TYPE (new_var
)))
830 location_t locus_0
, locus_1
;
832 new_var2
= make_ssa_name (TREE_TYPE (result
));
833 new_stmt
= gimple_build_assign (new_var2
, CONVERT_EXPR
, new_var
);
834 gsi_insert_before (&gsi
, new_stmt
, GSI_SAME_STMT
);
837 /* Set the locus to the first argument, unless is doesn't have one. */
838 locus_0
= gimple_phi_arg_location (phi
, 0);
839 locus_1
= gimple_phi_arg_location (phi
, 1);
840 if (locus_0
== UNKNOWN_LOCATION
)
842 gimple_set_location (new_stmt
, locus_0
);
845 replace_phi_edge_with_variable (cond_bb
, e1
, phi
, new_var
);
847 /* Note that we optimized this PHI. */
851 /* Update *ARG which is defined in STMT so that it contains the
852 computed value if that seems profitable. Return true if the
853 statement is made dead by that rewriting. */
856 jump_function_from_stmt (tree
*arg
, gimple
*stmt
)
858 enum tree_code code
= gimple_assign_rhs_code (stmt
);
859 if (code
== ADDR_EXPR
)
861 /* For arg = &p->i transform it to p, if possible. */
862 tree rhs1
= gimple_assign_rhs1 (stmt
);
864 tree tem
= get_addr_base_and_unit_offset (TREE_OPERAND (rhs1
, 0),
867 && TREE_CODE (tem
) == MEM_REF
868 && known_eq (mem_ref_offset (tem
) + offset
, 0))
870 *arg
= TREE_OPERAND (tem
, 0);
874 /* TODO: Much like IPA-CP jump-functions we want to handle constant
875 additions symbolically here, and we'd need to update the comparison
876 code that compares the arg + cst tuples in our caller. For now the
877 code above exactly handles the VEC_BASE pattern from vec.h. */
881 /* RHS is a source argument in a BIT_AND_EXPR which feeds a conditional
882 of the form SSA_NAME NE 0.
884 If RHS is fed by a simple EQ_EXPR comparison of two values, see if
885 the two input values of the EQ_EXPR match arg0 and arg1.
887 If so update *code and return TRUE. Otherwise return FALSE. */
890 rhs_is_fed_for_value_replacement (const_tree arg0
, const_tree arg1
,
891 enum tree_code
*code
, const_tree rhs
)
893 /* Obviously if RHS is not an SSA_NAME, we can't look at the defining
895 if (TREE_CODE (rhs
) == SSA_NAME
)
897 gimple
*def1
= SSA_NAME_DEF_STMT (rhs
);
899 /* Verify the defining statement has an EQ_EXPR on the RHS. */
900 if (is_gimple_assign (def1
) && gimple_assign_rhs_code (def1
) == EQ_EXPR
)
902 /* Finally verify the source operands of the EQ_EXPR are equal
904 tree op0
= gimple_assign_rhs1 (def1
);
905 tree op1
= gimple_assign_rhs2 (def1
);
906 if ((operand_equal_for_phi_arg_p (arg0
, op0
)
907 && operand_equal_for_phi_arg_p (arg1
, op1
))
908 || (operand_equal_for_phi_arg_p (arg0
, op1
)
909 && operand_equal_for_phi_arg_p (arg1
, op0
)))
911 /* We will perform the optimization. */
912 *code
= gimple_assign_rhs_code (def1
);
920 /* Return TRUE if arg0/arg1 are equal to the rhs/lhs or lhs/rhs of COND.
922 Also return TRUE if arg0/arg1 are equal to the source arguments of a
923 an EQ comparison feeding a BIT_AND_EXPR which feeds COND.
925 Return FALSE otherwise. */
928 operand_equal_for_value_replacement (const_tree arg0
, const_tree arg1
,
929 enum tree_code
*code
, gimple
*cond
)
932 tree lhs
= gimple_cond_lhs (cond
);
933 tree rhs
= gimple_cond_rhs (cond
);
935 if ((operand_equal_for_phi_arg_p (arg0
, lhs
)
936 && operand_equal_for_phi_arg_p (arg1
, rhs
))
937 || (operand_equal_for_phi_arg_p (arg1
, lhs
)
938 && operand_equal_for_phi_arg_p (arg0
, rhs
)))
941 /* Now handle more complex case where we have an EQ comparison
942 which feeds a BIT_AND_EXPR which feeds COND.
944 First verify that COND is of the form SSA_NAME NE 0. */
945 if (*code
!= NE_EXPR
|| !integer_zerop (rhs
)
946 || TREE_CODE (lhs
) != SSA_NAME
)
949 /* Now ensure that SSA_NAME is set by a BIT_AND_EXPR. */
950 def
= SSA_NAME_DEF_STMT (lhs
);
951 if (!is_gimple_assign (def
) || gimple_assign_rhs_code (def
) != BIT_AND_EXPR
)
954 /* Now verify arg0/arg1 correspond to the source arguments of an
955 EQ comparison feeding the BIT_AND_EXPR. */
957 tree tmp
= gimple_assign_rhs1 (def
);
958 if (rhs_is_fed_for_value_replacement (arg0
, arg1
, code
, tmp
))
961 tmp
= gimple_assign_rhs2 (def
);
962 if (rhs_is_fed_for_value_replacement (arg0
, arg1
, code
, tmp
))
968 /* Returns true if ARG is a neutral element for operation CODE
969 on the RIGHT side. */
972 neutral_element_p (tree_code code
, tree arg
, bool right
)
979 return integer_zerop (arg
);
986 case POINTER_PLUS_EXPR
:
987 return right
&& integer_zerop (arg
);
990 return integer_onep (arg
);
997 return right
&& integer_onep (arg
);
1000 return integer_all_onesp (arg
);
1007 /* Returns true if ARG is an absorbing element for operation CODE. */
1010 absorbing_element_p (tree_code code
, tree arg
, bool right
, tree rval
)
1015 return integer_all_onesp (arg
);
1019 return integer_zerop (arg
);
1025 return !right
&& integer_zerop (arg
);
1027 case TRUNC_DIV_EXPR
:
1029 case FLOOR_DIV_EXPR
:
1030 case ROUND_DIV_EXPR
:
1031 case EXACT_DIV_EXPR
:
1032 case TRUNC_MOD_EXPR
:
1034 case FLOOR_MOD_EXPR
:
1035 case ROUND_MOD_EXPR
:
1037 && integer_zerop (arg
)
1038 && tree_single_nonzero_warnv_p (rval
, NULL
));
1045 /* The function value_replacement does the main work of doing the value
1046 replacement. Return non-zero if the replacement is done. Otherwise return
1047 0. If we remove the middle basic block, return 2.
1048 BB is the basic block where the replacement is going to be done on. ARG0
1049 is argument 0 from the PHI. Likewise for ARG1. */
1052 value_replacement (basic_block cond_bb
, basic_block middle_bb
,
1053 edge e0
, edge e1
, gimple
*phi
,
1054 tree arg0
, tree arg1
)
1056 gimple_stmt_iterator gsi
;
1058 edge true_edge
, false_edge
;
1059 enum tree_code code
;
1060 bool emtpy_or_with_defined_p
= true;
1062 /* If the type says honor signed zeros we cannot do this
1064 if (HONOR_SIGNED_ZEROS (arg1
))
1067 /* If there is a statement in MIDDLE_BB that defines one of the PHI
1068 arguments, then adjust arg0 or arg1. */
1069 gsi
= gsi_start_nondebug_after_labels_bb (middle_bb
);
1070 while (!gsi_end_p (gsi
))
1072 gimple
*stmt
= gsi_stmt (gsi
);
1074 gsi_next_nondebug (&gsi
);
1075 if (!is_gimple_assign (stmt
))
1077 if (gimple_code (stmt
) != GIMPLE_PREDICT
1078 && gimple_code (stmt
) != GIMPLE_NOP
)
1079 emtpy_or_with_defined_p
= false;
1082 /* Now try to adjust arg0 or arg1 according to the computation
1083 in the statement. */
1084 lhs
= gimple_assign_lhs (stmt
);
1086 && jump_function_from_stmt (&arg0
, stmt
))
1088 && jump_function_from_stmt (&arg1
, stmt
)))
1089 emtpy_or_with_defined_p
= false;
1092 cond
= last_stmt (cond_bb
);
1093 code
= gimple_cond_code (cond
);
1095 /* This transformation is only valid for equality comparisons. */
1096 if (code
!= NE_EXPR
&& code
!= EQ_EXPR
)
1099 /* We need to know which is the true edge and which is the false
1100 edge so that we know if have abs or negative abs. */
1101 extract_true_false_edges_from_block (cond_bb
, &true_edge
, &false_edge
);
1103 /* At this point we know we have a COND_EXPR with two successors.
1104 One successor is BB, the other successor is an empty block which
1105 falls through into BB.
1107 The condition for the COND_EXPR is known to be NE_EXPR or EQ_EXPR.
1109 There is a single PHI node at the join point (BB) with two arguments.
1111 We now need to verify that the two arguments in the PHI node match
1112 the two arguments to the equality comparison. */
1114 if (operand_equal_for_value_replacement (arg0
, arg1
, &code
, cond
))
1119 /* For NE_EXPR, we want to build an assignment result = arg where
1120 arg is the PHI argument associated with the true edge. For
1121 EQ_EXPR we want the PHI argument associated with the false edge. */
1122 e
= (code
== NE_EXPR
? true_edge
: false_edge
);
1124 /* Unfortunately, E may not reach BB (it may instead have gone to
1125 OTHER_BLOCK). If that is the case, then we want the single outgoing
1126 edge from OTHER_BLOCK which reaches BB and represents the desired
1127 path from COND_BLOCK. */
1128 if (e
->dest
== middle_bb
)
1129 e
= single_succ_edge (e
->dest
);
1131 /* Now we know the incoming edge to BB that has the argument for the
1132 RHS of our new assignment statement. */
1138 /* If the middle basic block was empty or is defining the
1139 PHI arguments and this is a single phi where the args are different
1140 for the edges e0 and e1 then we can remove the middle basic block. */
1141 if (emtpy_or_with_defined_p
1142 && single_non_singleton_phi_for_edges (phi_nodes (gimple_bb (phi
)),
1145 replace_phi_edge_with_variable (cond_bb
, e1
, phi
, arg
);
1146 /* Note that we optimized this PHI. */
1151 /* Replace the PHI arguments with arg. */
1152 SET_PHI_ARG_DEF (phi
, e0
->dest_idx
, arg
);
1153 SET_PHI_ARG_DEF (phi
, e1
->dest_idx
, arg
);
1154 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
1156 fprintf (dump_file
, "PHI ");
1157 print_generic_expr (dump_file
, gimple_phi_result (phi
));
1158 fprintf (dump_file
, " reduced for COND_EXPR in block %d to ",
1160 print_generic_expr (dump_file
, arg
);
1161 fprintf (dump_file
, ".\n");
1168 /* Now optimize (x != 0) ? x + y : y to just x + y. */
1169 gsi
= gsi_last_nondebug_bb (middle_bb
);
1170 if (gsi_end_p (gsi
))
1173 gimple
*assign
= gsi_stmt (gsi
);
1174 if (!is_gimple_assign (assign
)
1175 || gimple_assign_rhs_class (assign
) != GIMPLE_BINARY_RHS
1176 || (!INTEGRAL_TYPE_P (TREE_TYPE (arg0
))
1177 && !POINTER_TYPE_P (TREE_TYPE (arg0
))))
1180 /* Punt if there are (degenerate) PHIs in middle_bb, there should not be. */
1181 if (!gimple_seq_empty_p (phi_nodes (middle_bb
)))
1184 /* Allow up to 2 cheap preparation statements that prepare argument
1192 iftmp.0_6 = x_5(D) r<< _1;
1194 # iftmp.0_2 = PHI <iftmp.0_6(3), x_5(D)(2)>
1205 # _2 = PHI <x_5(D)(2), _6(3)> */
1206 gimple
*prep_stmt
[2] = { NULL
, NULL
};
1208 for (prep_cnt
= 0; ; prep_cnt
++)
1210 gsi_prev_nondebug (&gsi
);
1211 if (gsi_end_p (gsi
))
1214 gimple
*g
= gsi_stmt (gsi
);
1215 if (gimple_code (g
) == GIMPLE_LABEL
)
1218 if (prep_cnt
== 2 || !is_gimple_assign (g
))
1221 tree lhs
= gimple_assign_lhs (g
);
1222 tree rhs1
= gimple_assign_rhs1 (g
);
1223 use_operand_p use_p
;
1225 if (TREE_CODE (lhs
) != SSA_NAME
1226 || TREE_CODE (rhs1
) != SSA_NAME
1227 || !INTEGRAL_TYPE_P (TREE_TYPE (lhs
))
1228 || !INTEGRAL_TYPE_P (TREE_TYPE (rhs1
))
1229 || !single_imm_use (lhs
, &use_p
, &use_stmt
)
1230 || use_stmt
!= (prep_cnt
? prep_stmt
[prep_cnt
- 1] : assign
))
1232 switch (gimple_assign_rhs_code (g
))
1240 if (TREE_CODE (gimple_assign_rhs2 (g
)) != INTEGER_CST
)
1246 prep_stmt
[prep_cnt
] = g
;
1249 /* Only transform if it removes the condition. */
1250 if (!single_non_singleton_phi_for_edges (phi_nodes (gimple_bb (phi
)), e0
, e1
))
1253 /* Size-wise, this is always profitable. */
1254 if (optimize_bb_for_speed_p (cond_bb
)
1255 /* The special case is useless if it has a low probability. */
1256 && profile_status_for_fn (cfun
) != PROFILE_ABSENT
1257 && EDGE_PRED (middle_bb
, 0)->probability
< profile_probability::even ()
1258 /* If assign is cheap, there is no point avoiding it. */
1259 && estimate_num_insns (bb_seq (middle_bb
), &eni_time_weights
)
1260 >= 3 * estimate_num_insns (cond
, &eni_time_weights
))
1263 tree lhs
= gimple_assign_lhs (assign
);
1264 tree rhs1
= gimple_assign_rhs1 (assign
);
1265 tree rhs2
= gimple_assign_rhs2 (assign
);
1266 enum tree_code code_def
= gimple_assign_rhs_code (assign
);
1267 tree cond_lhs
= gimple_cond_lhs (cond
);
1268 tree cond_rhs
= gimple_cond_rhs (cond
);
1270 /* Propagate the cond_rhs constant through preparation stmts,
1271 make sure UB isn't invoked while doing that. */
1272 for (int i
= prep_cnt
- 1; i
>= 0; --i
)
1274 gimple
*g
= prep_stmt
[i
];
1275 tree grhs1
= gimple_assign_rhs1 (g
);
1276 if (!operand_equal_for_phi_arg_p (cond_lhs
, grhs1
))
1278 cond_lhs
= gimple_assign_lhs (g
);
1279 cond_rhs
= fold_convert (TREE_TYPE (grhs1
), cond_rhs
);
1280 if (TREE_CODE (cond_rhs
) != INTEGER_CST
1281 || TREE_OVERFLOW (cond_rhs
))
1283 if (gimple_assign_rhs_class (g
) == GIMPLE_BINARY_RHS
)
1285 cond_rhs
= int_const_binop (gimple_assign_rhs_code (g
), cond_rhs
,
1286 gimple_assign_rhs2 (g
));
1287 if (TREE_OVERFLOW (cond_rhs
))
1290 cond_rhs
= fold_convert (TREE_TYPE (cond_lhs
), cond_rhs
);
1291 if (TREE_CODE (cond_rhs
) != INTEGER_CST
1292 || TREE_OVERFLOW (cond_rhs
))
1296 if (((code
== NE_EXPR
&& e1
== false_edge
)
1297 || (code
== EQ_EXPR
&& e1
== true_edge
))
1300 && operand_equal_for_phi_arg_p (rhs2
, cond_lhs
)
1301 && neutral_element_p (code_def
, cond_rhs
, true))
1303 && operand_equal_for_phi_arg_p (rhs1
, cond_lhs
)
1304 && neutral_element_p (code_def
, cond_rhs
, false))
1305 || (operand_equal_for_phi_arg_p (arg1
, cond_rhs
)
1306 && ((operand_equal_for_phi_arg_p (rhs2
, cond_lhs
)
1307 && absorbing_element_p (code_def
, cond_rhs
, true, rhs2
))
1308 || (operand_equal_for_phi_arg_p (rhs1
, cond_lhs
)
1309 && absorbing_element_p (code_def
,
1310 cond_rhs
, false, rhs2
))))))
1312 gsi
= gsi_for_stmt (cond
);
1313 /* Moving ASSIGN might change VR of lhs, e.g. when moving u_6
1321 # RANGE [0, 4294967294]
1322 u_6 = n_5 + 4294967295;
1325 # u_3 = PHI <u_6(3), 4294967295(2)> */
1326 reset_flow_sensitive_info (lhs
);
1327 if (INTEGRAL_TYPE_P (TREE_TYPE (lhs
)))
1329 /* If available, we can use VR of phi result at least. */
1330 tree phires
= gimple_phi_result (phi
);
1331 struct range_info_def
*phires_range_info
1332 = SSA_NAME_RANGE_INFO (phires
);
1333 if (phires_range_info
)
1334 duplicate_ssa_name_range_info (lhs
, SSA_NAME_RANGE_TYPE (phires
),
1337 gimple_stmt_iterator gsi_from
;
1338 for (int i
= prep_cnt
- 1; i
>= 0; --i
)
1340 tree plhs
= gimple_assign_lhs (prep_stmt
[i
]);
1341 reset_flow_sensitive_info (plhs
);
1342 gsi_from
= gsi_for_stmt (prep_stmt
[i
]);
1343 gsi_move_before (&gsi_from
, &gsi
);
1345 gsi_from
= gsi_for_stmt (assign
);
1346 gsi_move_before (&gsi_from
, &gsi
);
1347 replace_phi_edge_with_variable (cond_bb
, e1
, phi
, lhs
);
1354 /* The function minmax_replacement does the main work of doing the minmax
1355 replacement. Return true if the replacement is done. Otherwise return
1357 BB is the basic block where the replacement is going to be done on. ARG0
1358 is argument 0 from the PHI. Likewise for ARG1. */
1361 minmax_replacement (basic_block cond_bb
, basic_block middle_bb
,
1362 edge e0
, edge e1
, gimple
*phi
,
1363 tree arg0
, tree arg1
)
1365 tree result
, type
, rhs
;
1368 edge true_edge
, false_edge
;
1369 enum tree_code cmp
, minmax
, ass_code
;
1370 tree smaller
, alt_smaller
, larger
, alt_larger
, arg_true
, arg_false
;
1371 gimple_stmt_iterator gsi
, gsi_from
;
1373 type
= TREE_TYPE (PHI_RESULT (phi
));
1375 /* The optimization may be unsafe due to NaNs. */
1376 if (HONOR_NANS (type
) || HONOR_SIGNED_ZEROS (type
))
1379 cond
= as_a
<gcond
*> (last_stmt (cond_bb
));
1380 cmp
= gimple_cond_code (cond
);
1381 rhs
= gimple_cond_rhs (cond
);
1383 /* Turn EQ/NE of extreme values to order comparisons. */
1384 if ((cmp
== NE_EXPR
|| cmp
== EQ_EXPR
)
1385 && TREE_CODE (rhs
) == INTEGER_CST
)
1387 if (wi::eq_p (wi::to_wide (rhs
), wi::min_value (TREE_TYPE (rhs
))))
1389 cmp
= (cmp
== EQ_EXPR
) ? LT_EXPR
: GE_EXPR
;
1390 rhs
= wide_int_to_tree (TREE_TYPE (rhs
),
1391 wi::min_value (TREE_TYPE (rhs
)) + 1);
1393 else if (wi::eq_p (wi::to_wide (rhs
), wi::max_value (TREE_TYPE (rhs
))))
1395 cmp
= (cmp
== EQ_EXPR
) ? GT_EXPR
: LE_EXPR
;
1396 rhs
= wide_int_to_tree (TREE_TYPE (rhs
),
1397 wi::max_value (TREE_TYPE (rhs
)) - 1);
1401 /* This transformation is only valid for order comparisons. Record which
1402 operand is smaller/larger if the result of the comparison is true. */
1403 alt_smaller
= NULL_TREE
;
1404 alt_larger
= NULL_TREE
;
1405 if (cmp
== LT_EXPR
|| cmp
== LE_EXPR
)
1407 smaller
= gimple_cond_lhs (cond
);
1409 /* If we have smaller < CST it is equivalent to smaller <= CST-1.
1410 Likewise smaller <= CST is equivalent to smaller < CST+1. */
1411 if (TREE_CODE (larger
) == INTEGER_CST
)
1415 wi::overflow_type overflow
;
1416 wide_int alt
= wi::sub (wi::to_wide (larger
), 1,
1417 TYPE_SIGN (TREE_TYPE (larger
)),
1420 alt_larger
= wide_int_to_tree (TREE_TYPE (larger
), alt
);
1424 wi::overflow_type overflow
;
1425 wide_int alt
= wi::add (wi::to_wide (larger
), 1,
1426 TYPE_SIGN (TREE_TYPE (larger
)),
1429 alt_larger
= wide_int_to_tree (TREE_TYPE (larger
), alt
);
1433 else if (cmp
== GT_EXPR
|| cmp
== GE_EXPR
)
1436 larger
= gimple_cond_lhs (cond
);
1437 /* If we have larger > CST it is equivalent to larger >= CST+1.
1438 Likewise larger >= CST is equivalent to larger > CST-1. */
1439 if (TREE_CODE (smaller
) == INTEGER_CST
)
1441 wi::overflow_type overflow
;
1444 wide_int alt
= wi::add (wi::to_wide (smaller
), 1,
1445 TYPE_SIGN (TREE_TYPE (smaller
)),
1448 alt_smaller
= wide_int_to_tree (TREE_TYPE (smaller
), alt
);
1452 wide_int alt
= wi::sub (wi::to_wide (smaller
), 1,
1453 TYPE_SIGN (TREE_TYPE (smaller
)),
1456 alt_smaller
= wide_int_to_tree (TREE_TYPE (smaller
), alt
);
1463 /* We need to know which is the true edge and which is the false
1464 edge so that we know if have abs or negative abs. */
1465 extract_true_false_edges_from_block (cond_bb
, &true_edge
, &false_edge
);
1467 /* Forward the edges over the middle basic block. */
1468 if (true_edge
->dest
== middle_bb
)
1469 true_edge
= EDGE_SUCC (true_edge
->dest
, 0);
1470 if (false_edge
->dest
== middle_bb
)
1471 false_edge
= EDGE_SUCC (false_edge
->dest
, 0);
1473 if (true_edge
== e0
)
1475 gcc_assert (false_edge
== e1
);
1481 gcc_assert (false_edge
== e0
);
1482 gcc_assert (true_edge
== e1
);
1487 if (empty_block_p (middle_bb
))
1489 if ((operand_equal_for_phi_arg_p (arg_true
, smaller
)
1491 && operand_equal_for_phi_arg_p (arg_true
, alt_smaller
)))
1492 && (operand_equal_for_phi_arg_p (arg_false
, larger
)
1494 && operand_equal_for_phi_arg_p (arg_true
, alt_larger
))))
1498 if (smaller < larger)
1504 else if ((operand_equal_for_phi_arg_p (arg_false
, smaller
)
1506 && operand_equal_for_phi_arg_p (arg_false
, alt_smaller
)))
1507 && (operand_equal_for_phi_arg_p (arg_true
, larger
)
1509 && operand_equal_for_phi_arg_p (arg_true
, alt_larger
))))
1516 /* Recognize the following case, assuming d <= u:
1522 This is equivalent to
1527 gimple
*assign
= last_and_only_stmt (middle_bb
);
1528 tree lhs
, op0
, op1
, bound
;
1531 || gimple_code (assign
) != GIMPLE_ASSIGN
)
1534 lhs
= gimple_assign_lhs (assign
);
1535 ass_code
= gimple_assign_rhs_code (assign
);
1536 if (ass_code
!= MAX_EXPR
&& ass_code
!= MIN_EXPR
)
1538 op0
= gimple_assign_rhs1 (assign
);
1539 op1
= gimple_assign_rhs2 (assign
);
1541 if (true_edge
->src
== middle_bb
)
1543 /* We got here if the condition is true, i.e., SMALLER < LARGER. */
1544 if (!operand_equal_for_phi_arg_p (lhs
, arg_true
))
1547 if (operand_equal_for_phi_arg_p (arg_false
, larger
)
1549 && operand_equal_for_phi_arg_p (arg_false
, alt_larger
)))
1553 if (smaller < larger)
1555 r' = MAX_EXPR (smaller, bound)
1557 r = PHI <r', larger> --> to be turned to MIN_EXPR. */
1558 if (ass_code
!= MAX_EXPR
)
1562 if (operand_equal_for_phi_arg_p (op0
, smaller
)
1564 && operand_equal_for_phi_arg_p (op0
, alt_smaller
)))
1566 else if (operand_equal_for_phi_arg_p (op1
, smaller
)
1568 && operand_equal_for_phi_arg_p (op1
, alt_smaller
)))
1573 /* We need BOUND <= LARGER. */
1574 if (!integer_nonzerop (fold_build2 (LE_EXPR
, boolean_type_node
,
1578 else if (operand_equal_for_phi_arg_p (arg_false
, smaller
)
1580 && operand_equal_for_phi_arg_p (arg_false
, alt_smaller
)))
1584 if (smaller < larger)
1586 r' = MIN_EXPR (larger, bound)
1588 r = PHI <r', smaller> --> to be turned to MAX_EXPR. */
1589 if (ass_code
!= MIN_EXPR
)
1593 if (operand_equal_for_phi_arg_p (op0
, larger
)
1595 && operand_equal_for_phi_arg_p (op0
, alt_larger
)))
1597 else if (operand_equal_for_phi_arg_p (op1
, larger
)
1599 && operand_equal_for_phi_arg_p (op1
, alt_larger
)))
1604 /* We need BOUND >= SMALLER. */
1605 if (!integer_nonzerop (fold_build2 (GE_EXPR
, boolean_type_node
,
1614 /* We got here if the condition is false, i.e., SMALLER > LARGER. */
1615 if (!operand_equal_for_phi_arg_p (lhs
, arg_false
))
1618 if (operand_equal_for_phi_arg_p (arg_true
, larger
)
1620 && operand_equal_for_phi_arg_p (arg_true
, alt_larger
)))
1624 if (smaller > larger)
1626 r' = MIN_EXPR (smaller, bound)
1628 r = PHI <r', larger> --> to be turned to MAX_EXPR. */
1629 if (ass_code
!= MIN_EXPR
)
1633 if (operand_equal_for_phi_arg_p (op0
, smaller
)
1635 && operand_equal_for_phi_arg_p (op0
, alt_smaller
)))
1637 else if (operand_equal_for_phi_arg_p (op1
, smaller
)
1639 && operand_equal_for_phi_arg_p (op1
, alt_smaller
)))
1644 /* We need BOUND >= LARGER. */
1645 if (!integer_nonzerop (fold_build2 (GE_EXPR
, boolean_type_node
,
1649 else if (operand_equal_for_phi_arg_p (arg_true
, smaller
)
1651 && operand_equal_for_phi_arg_p (arg_true
, alt_smaller
)))
1655 if (smaller > larger)
1657 r' = MAX_EXPR (larger, bound)
1659 r = PHI <r', smaller> --> to be turned to MIN_EXPR. */
1660 if (ass_code
!= MAX_EXPR
)
1664 if (operand_equal_for_phi_arg_p (op0
, larger
))
1666 else if (operand_equal_for_phi_arg_p (op1
, larger
))
1671 /* We need BOUND <= SMALLER. */
1672 if (!integer_nonzerop (fold_build2 (LE_EXPR
, boolean_type_node
,
1680 /* Move the statement from the middle block. */
1681 gsi
= gsi_last_bb (cond_bb
);
1682 gsi_from
= gsi_last_nondebug_bb (middle_bb
);
1683 reset_flow_sensitive_info (SINGLE_SSA_TREE_OPERAND (gsi_stmt (gsi_from
),
1685 gsi_move_before (&gsi_from
, &gsi
);
1688 /* Create an SSA var to hold the min/max result. If we're the only
1689 things setting the target PHI, then we can clone the PHI
1690 variable. Otherwise we must create a new one. */
1691 result
= PHI_RESULT (phi
);
1692 if (EDGE_COUNT (gimple_bb (phi
)->preds
) == 2)
1693 result
= duplicate_ssa_name (result
, NULL
);
1695 result
= make_ssa_name (TREE_TYPE (result
));
1697 /* Emit the statement to compute min/max. */
1698 new_stmt
= gimple_build_assign (result
, minmax
, arg0
, arg1
);
1699 gsi
= gsi_last_bb (cond_bb
);
1700 gsi_insert_before (&gsi
, new_stmt
, GSI_NEW_STMT
);
1702 replace_phi_edge_with_variable (cond_bb
, e1
, phi
, result
);
1716 _2 = (unsigned long) b_4(D);
1717 _9 = __builtin_popcountl (_2);
1719 _9 = __builtin_popcountl (b_4(D));
1722 c_12 = PHI <0(2), _9(3)>
1726 _2 = (unsigned long) b_4(D);
1727 _9 = __builtin_popcountl (_2);
1729 _9 = __builtin_popcountl (b_4(D));
1736 cond_removal_in_popcount_pattern (basic_block cond_bb
, basic_block middle_bb
,
1738 gimple
*phi
, tree arg0
, tree arg1
)
1741 gimple_stmt_iterator gsi
, gsi_from
;
1743 gimple
*cast
= NULL
;
1747 _2 = (unsigned long) b_4(D);
1748 _9 = __builtin_popcountl (_2);
1750 _9 = __builtin_popcountl (b_4(D));
1751 are the only stmts in the middle_bb. */
1753 gsi
= gsi_start_nondebug_after_labels_bb (middle_bb
);
1754 if (gsi_end_p (gsi
))
1756 cast
= gsi_stmt (gsi
);
1757 gsi_next_nondebug (&gsi
);
1758 if (!gsi_end_p (gsi
))
1760 popcount
= gsi_stmt (gsi
);
1761 gsi_next_nondebug (&gsi
);
1762 if (!gsi_end_p (gsi
))
1771 /* Check that we have a popcount builtin. */
1772 if (!is_gimple_call (popcount
))
1774 combined_fn cfn
= gimple_call_combined_fn (popcount
);
1783 arg
= gimple_call_arg (popcount
, 0);
1784 lhs
= gimple_get_lhs (popcount
);
1788 /* We have a cast stmt feeding popcount builtin. */
1789 /* Check that we have a cast prior to that. */
1790 if (gimple_code (cast
) != GIMPLE_ASSIGN
1791 || !CONVERT_EXPR_CODE_P (gimple_assign_rhs_code (cast
)))
1793 /* Result of the cast stmt is the argument to the builtin. */
1794 if (arg
!= gimple_assign_lhs (cast
))
1796 arg
= gimple_assign_rhs1 (cast
);
1799 cond
= last_stmt (cond_bb
);
1801 /* Cond_bb has a check for b_4 [!=|==] 0 before calling the popcount
1803 if (gimple_code (cond
) != GIMPLE_COND
1804 || (gimple_cond_code (cond
) != NE_EXPR
1805 && gimple_cond_code (cond
) != EQ_EXPR
)
1806 || !integer_zerop (gimple_cond_rhs (cond
))
1807 || arg
!= gimple_cond_lhs (cond
))
1811 if ((e2
->flags
& EDGE_TRUE_VALUE
1812 && gimple_cond_code (cond
) == NE_EXPR
)
1813 || (e1
->flags
& EDGE_TRUE_VALUE
1814 && gimple_cond_code (cond
) == EQ_EXPR
))
1816 std::swap (arg0
, arg1
);
1820 /* Check PHI arguments. */
1821 if (lhs
!= arg0
|| !integer_zerop (arg1
))
1824 /* And insert the popcount builtin and cast stmt before the cond_bb. */
1825 gsi
= gsi_last_bb (cond_bb
);
1828 gsi_from
= gsi_for_stmt (cast
);
1829 gsi_move_before (&gsi_from
, &gsi
);
1830 reset_flow_sensitive_info (gimple_get_lhs (cast
));
1832 gsi_from
= gsi_for_stmt (popcount
);
1833 gsi_move_before (&gsi_from
, &gsi
);
1834 reset_flow_sensitive_info (gimple_get_lhs (popcount
));
1836 /* Now update the PHI and remove unneeded bbs. */
1837 replace_phi_edge_with_variable (cond_bb
, e2
, phi
, lhs
);
1841 /* The function absolute_replacement does the main work of doing the absolute
1842 replacement. Return true if the replacement is done. Otherwise return
1844 bb is the basic block where the replacement is going to be done on. arg0
1845 is argument 0 from the phi. Likewise for arg1. */
1848 abs_replacement (basic_block cond_bb
, basic_block middle_bb
,
1849 edge e0 ATTRIBUTE_UNUSED
, edge e1
,
1850 gimple
*phi
, tree arg0
, tree arg1
)
1855 gimple_stmt_iterator gsi
;
1856 edge true_edge
, false_edge
;
1861 enum tree_code cond_code
;
1863 /* If the type says honor signed zeros we cannot do this
1865 if (HONOR_SIGNED_ZEROS (arg1
))
1868 /* OTHER_BLOCK must have only one executable statement which must have the
1869 form arg0 = -arg1 or arg1 = -arg0. */
1871 assign
= last_and_only_stmt (middle_bb
);
1872 /* If we did not find the proper negation assignment, then we cannot
1877 /* If we got here, then we have found the only executable statement
1878 in OTHER_BLOCK. If it is anything other than arg = -arg1 or
1879 arg1 = -arg0, then we cannot optimize. */
1880 if (gimple_code (assign
) != GIMPLE_ASSIGN
)
1883 lhs
= gimple_assign_lhs (assign
);
1885 if (gimple_assign_rhs_code (assign
) != NEGATE_EXPR
)
1888 rhs
= gimple_assign_rhs1 (assign
);
1890 /* The assignment has to be arg0 = -arg1 or arg1 = -arg0. */
1891 if (!(lhs
== arg0
&& rhs
== arg1
)
1892 && !(lhs
== arg1
&& rhs
== arg0
))
1895 cond
= last_stmt (cond_bb
);
1896 result
= PHI_RESULT (phi
);
1898 /* Only relationals comparing arg[01] against zero are interesting. */
1899 cond_code
= gimple_cond_code (cond
);
1900 if (cond_code
!= GT_EXPR
&& cond_code
!= GE_EXPR
1901 && cond_code
!= LT_EXPR
&& cond_code
!= LE_EXPR
)
1904 /* Make sure the conditional is arg[01] OP y. */
1905 if (gimple_cond_lhs (cond
) != rhs
)
1908 if (FLOAT_TYPE_P (TREE_TYPE (gimple_cond_rhs (cond
)))
1909 ? real_zerop (gimple_cond_rhs (cond
))
1910 : integer_zerop (gimple_cond_rhs (cond
)))
1915 /* We need to know which is the true edge and which is the false
1916 edge so that we know if have abs or negative abs. */
1917 extract_true_false_edges_from_block (cond_bb
, &true_edge
, &false_edge
);
1919 /* For GT_EXPR/GE_EXPR, if the true edge goes to OTHER_BLOCK, then we
1920 will need to negate the result. Similarly for LT_EXPR/LE_EXPR if
1921 the false edge goes to OTHER_BLOCK. */
1922 if (cond_code
== GT_EXPR
|| cond_code
== GE_EXPR
)
1927 if (e
->dest
== middle_bb
)
1932 /* If the code negates only iff positive then make sure to not
1933 introduce undefined behavior when negating or computing the absolute.
1934 ??? We could use range info if present to check for arg1 == INT_MIN. */
1936 && (ANY_INTEGRAL_TYPE_P (TREE_TYPE (arg1
))
1937 && ! TYPE_OVERFLOW_WRAPS (TREE_TYPE (arg1
))))
1940 result
= duplicate_ssa_name (result
, NULL
);
1943 lhs
= make_ssa_name (TREE_TYPE (result
));
1947 /* Build the modify expression with abs expression. */
1948 new_stmt
= gimple_build_assign (lhs
, ABS_EXPR
, rhs
);
1950 gsi
= gsi_last_bb (cond_bb
);
1951 gsi_insert_before (&gsi
, new_stmt
, GSI_NEW_STMT
);
1955 /* Get the right GSI. We want to insert after the recently
1956 added ABS_EXPR statement (which we know is the first statement
1958 new_stmt
= gimple_build_assign (result
, NEGATE_EXPR
, lhs
);
1960 gsi_insert_after (&gsi
, new_stmt
, GSI_NEW_STMT
);
1963 replace_phi_edge_with_variable (cond_bb
, e1
, phi
, result
);
1965 /* Note that we optimized this PHI. */
1969 /* Auxiliary functions to determine the set of memory accesses which
1970 can't trap because they are preceded by accesses to the same memory
1971 portion. We do that for MEM_REFs, so we only need to track
1972 the SSA_NAME of the pointer indirectly referenced. The algorithm
1973 simply is a walk over all instructions in dominator order. When
1974 we see an MEM_REF we determine if we've already seen a same
1975 ref anywhere up to the root of the dominator tree. If we do the
1976 current access can't trap. If we don't see any dominating access
1977 the current access might trap, but might also make later accesses
1978 non-trapping, so we remember it. We need to be careful with loads
1979 or stores, for instance a load might not trap, while a store would,
1980 so if we see a dominating read access this doesn't mean that a later
1981 write access would not trap. Hence we also need to differentiate the
1982 type of access(es) seen.
1984 ??? We currently are very conservative and assume that a load might
1985 trap even if a store doesn't (write-only memory). This probably is
1986 overly conservative. */
1988 /* A hash-table of SSA_NAMEs, and in which basic block an MEM_REF
1989 through it was seen, which would constitute a no-trap region for
1993 unsigned int ssa_name_ver
;
1996 HOST_WIDE_INT offset
, size
;
2000 /* Hashtable helpers. */
2002 struct ssa_names_hasher
: free_ptr_hash
<name_to_bb
>
2004 static inline hashval_t
hash (const name_to_bb
*);
2005 static inline bool equal (const name_to_bb
*, const name_to_bb
*);
2008 /* Used for quick clearing of the hash-table when we see calls.
2009 Hash entries with phase < nt_call_phase are invalid. */
2010 static unsigned int nt_call_phase
;
2012 /* The hash function. */
2015 ssa_names_hasher::hash (const name_to_bb
*n
)
2017 return n
->ssa_name_ver
^ (((hashval_t
) n
->store
) << 31)
2018 ^ (n
->offset
<< 6) ^ (n
->size
<< 3);
2021 /* The equality function of *P1 and *P2. */
2024 ssa_names_hasher::equal (const name_to_bb
*n1
, const name_to_bb
*n2
)
2026 return n1
->ssa_name_ver
== n2
->ssa_name_ver
2027 && n1
->store
== n2
->store
2028 && n1
->offset
== n2
->offset
2029 && n1
->size
== n2
->size
;
2032 class nontrapping_dom_walker
: public dom_walker
2035 nontrapping_dom_walker (cdi_direction direction
, hash_set
<tree
> *ps
)
2036 : dom_walker (direction
), m_nontrapping (ps
), m_seen_ssa_names (128) {}
2038 virtual edge
before_dom_children (basic_block
);
2039 virtual void after_dom_children (basic_block
);
2043 /* We see the expression EXP in basic block BB. If it's an interesting
2044 expression (an MEM_REF through an SSA_NAME) possibly insert the
2045 expression into the set NONTRAP or the hash table of seen expressions.
2046 STORE is true if this expression is on the LHS, otherwise it's on
2048 void add_or_mark_expr (basic_block
, tree
, bool);
2050 hash_set
<tree
> *m_nontrapping
;
2052 /* The hash table for remembering what we've seen. */
2053 hash_table
<ssa_names_hasher
> m_seen_ssa_names
;
2056 /* Called by walk_dominator_tree, when entering the block BB. */
2058 nontrapping_dom_walker::before_dom_children (basic_block bb
)
2062 gimple_stmt_iterator gsi
;
2064 /* If we haven't seen all our predecessors, clear the hash-table. */
2065 FOR_EACH_EDGE (e
, ei
, bb
->preds
)
2066 if ((((size_t)e
->src
->aux
) & 2) == 0)
2072 /* Mark this BB as being on the path to dominator root and as visited. */
2073 bb
->aux
= (void*)(1 | 2);
2075 /* And walk the statements in order. */
2076 for (gsi
= gsi_start_bb (bb
); !gsi_end_p (gsi
); gsi_next (&gsi
))
2078 gimple
*stmt
= gsi_stmt (gsi
);
2080 if ((gimple_code (stmt
) == GIMPLE_ASM
&& gimple_vdef (stmt
))
2081 || (is_gimple_call (stmt
)
2082 && (!nonfreeing_call_p (stmt
) || !nonbarrier_call_p (stmt
))))
2084 else if (gimple_assign_single_p (stmt
) && !gimple_has_volatile_ops (stmt
))
2086 add_or_mark_expr (bb
, gimple_assign_lhs (stmt
), true);
2087 add_or_mark_expr (bb
, gimple_assign_rhs1 (stmt
), false);
2093 /* Called by walk_dominator_tree, when basic block BB is exited. */
2095 nontrapping_dom_walker::after_dom_children (basic_block bb
)
2097 /* This BB isn't on the path to dominator root anymore. */
2101 /* We see the expression EXP in basic block BB. If it's an interesting
2102 expression (an MEM_REF through an SSA_NAME) possibly insert the
2103 expression into the set NONTRAP or the hash table of seen expressions.
2104 STORE is true if this expression is on the LHS, otherwise it's on
2107 nontrapping_dom_walker::add_or_mark_expr (basic_block bb
, tree exp
, bool store
)
2111 if (TREE_CODE (exp
) == MEM_REF
2112 && TREE_CODE (TREE_OPERAND (exp
, 0)) == SSA_NAME
2113 && tree_fits_shwi_p (TREE_OPERAND (exp
, 1))
2114 && (size
= int_size_in_bytes (TREE_TYPE (exp
))) > 0)
2116 tree name
= TREE_OPERAND (exp
, 0);
2117 struct name_to_bb map
;
2119 struct name_to_bb
*n2bb
;
2120 basic_block found_bb
= 0;
2122 /* Try to find the last seen MEM_REF through the same
2123 SSA_NAME, which can trap. */
2124 map
.ssa_name_ver
= SSA_NAME_VERSION (name
);
2128 map
.offset
= tree_to_shwi (TREE_OPERAND (exp
, 1));
2131 slot
= m_seen_ssa_names
.find_slot (&map
, INSERT
);
2133 if (n2bb
&& n2bb
->phase
>= nt_call_phase
)
2134 found_bb
= n2bb
->bb
;
2136 /* If we've found a trapping MEM_REF, _and_ it dominates EXP
2137 (it's in a basic block on the path from us to the dominator root)
2138 then we can't trap. */
2139 if (found_bb
&& (((size_t)found_bb
->aux
) & 1) == 1)
2141 m_nontrapping
->add (exp
);
2145 /* EXP might trap, so insert it into the hash table. */
2148 n2bb
->phase
= nt_call_phase
;
2153 n2bb
= XNEW (struct name_to_bb
);
2154 n2bb
->ssa_name_ver
= SSA_NAME_VERSION (name
);
2155 n2bb
->phase
= nt_call_phase
;
2157 n2bb
->store
= store
;
2158 n2bb
->offset
= map
.offset
;
2166 /* This is the entry point of gathering non trapping memory accesses.
2167 It will do a dominator walk over the whole function, and it will
2168 make use of the bb->aux pointers. It returns a set of trees
2169 (the MEM_REFs itself) which can't trap. */
2170 static hash_set
<tree
> *
2171 get_non_trapping (void)
2174 hash_set
<tree
> *nontrap
= new hash_set
<tree
>;
2175 /* We're going to do a dominator walk, so ensure that we have
2176 dominance information. */
2177 calculate_dominance_info (CDI_DOMINATORS
);
2179 nontrapping_dom_walker (CDI_DOMINATORS
, nontrap
)
2180 .walk (cfun
->cfg
->x_entry_block_ptr
);
2182 clear_aux_for_blocks ();
2186 /* Do the main work of conditional store replacement. We already know
2187 that the recognized pattern looks like so:
2190 if (cond) goto MIDDLE_BB; else goto JOIN_BB (edge E1)
2193 fallthrough (edge E0)
2197 We check that MIDDLE_BB contains only one store, that that store
2198 doesn't trap (not via NOTRAP, but via checking if an access to the same
2199 memory location dominates us, or the store is to a local addressable
2200 object) and that the store has a "simple" RHS. */
2203 cond_store_replacement (basic_block middle_bb
, basic_block join_bb
,
2204 edge e0
, edge e1
, hash_set
<tree
> *nontrap
)
2206 gimple
*assign
= last_and_only_stmt (middle_bb
);
2207 tree lhs
, rhs
, name
, name2
;
2210 gimple_stmt_iterator gsi
;
2213 /* Check if middle_bb contains of only one store. */
2215 || !gimple_assign_single_p (assign
)
2216 || gimple_has_volatile_ops (assign
))
2219 /* And no PHI nodes so all uses in the single stmt are also
2220 available where we insert to. */
2221 if (!gimple_seq_empty_p (phi_nodes (middle_bb
)))
2224 locus
= gimple_location (assign
);
2225 lhs
= gimple_assign_lhs (assign
);
2226 rhs
= gimple_assign_rhs1 (assign
);
2227 if ((TREE_CODE (lhs
) != MEM_REF
2228 && TREE_CODE (lhs
) != ARRAY_REF
2229 && TREE_CODE (lhs
) != COMPONENT_REF
)
2230 || !is_gimple_reg_type (TREE_TYPE (lhs
)))
2233 /* Prove that we can move the store down. We could also check
2234 TREE_THIS_NOTRAP here, but in that case we also could move stores,
2235 whose value is not available readily, which we want to avoid. */
2236 if (!nontrap
->contains (lhs
))
2238 /* If LHS is a local variable without address-taken, we could
2239 always safely move down the store. */
2240 tree base
= get_base_address (lhs
);
2241 if (!auto_var_p (base
) || TREE_ADDRESSABLE (base
))
2245 /* Now we've checked the constraints, so do the transformation:
2246 1) Remove the single store. */
2247 gsi
= gsi_for_stmt (assign
);
2248 unlink_stmt_vdef (assign
);
2249 gsi_remove (&gsi
, true);
2250 release_defs (assign
);
2252 /* Make both store and load use alias-set zero as we have to
2253 deal with the case of the store being a conditional change
2254 of the dynamic type. */
2255 lhs
= unshare_expr (lhs
);
2257 while (handled_component_p (*basep
))
2258 basep
= &TREE_OPERAND (*basep
, 0);
2259 if (TREE_CODE (*basep
) == MEM_REF
2260 || TREE_CODE (*basep
) == TARGET_MEM_REF
)
2261 TREE_OPERAND (*basep
, 1)
2262 = fold_convert (ptr_type_node
, TREE_OPERAND (*basep
, 1));
2264 *basep
= build2 (MEM_REF
, TREE_TYPE (*basep
),
2265 build_fold_addr_expr (*basep
),
2266 build_zero_cst (ptr_type_node
));
2268 /* 2) Insert a load from the memory of the store to the temporary
2269 on the edge which did not contain the store. */
2270 name
= make_temp_ssa_name (TREE_TYPE (lhs
), NULL
, "cstore");
2271 new_stmt
= gimple_build_assign (name
, lhs
);
2272 gimple_set_location (new_stmt
, locus
);
2273 gsi_insert_on_edge (e1
, new_stmt
);
2275 /* 3) Create a PHI node at the join block, with one argument
2276 holding the old RHS, and the other holding the temporary
2277 where we stored the old memory contents. */
2278 name2
= make_temp_ssa_name (TREE_TYPE (lhs
), NULL
, "cstore");
2279 newphi
= create_phi_node (name2
, join_bb
);
2280 add_phi_arg (newphi
, rhs
, e0
, locus
);
2281 add_phi_arg (newphi
, name
, e1
, locus
);
2283 lhs
= unshare_expr (lhs
);
2284 new_stmt
= gimple_build_assign (lhs
, PHI_RESULT (newphi
));
2286 /* 4) Insert that PHI node. */
2287 gsi
= gsi_after_labels (join_bb
);
2288 if (gsi_end_p (gsi
))
2290 gsi
= gsi_last_bb (join_bb
);
2291 gsi_insert_after (&gsi
, new_stmt
, GSI_NEW_STMT
);
2294 gsi_insert_before (&gsi
, new_stmt
, GSI_NEW_STMT
);
2296 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
2298 fprintf (dump_file
, "\nConditional store replacement happened!");
2299 fprintf (dump_file
, "\nReplaced the store with a load.");
2300 fprintf (dump_file
, "\nInserted a new PHI statement in joint block:\n");
2301 print_gimple_stmt (dump_file
, new_stmt
, 0, TDF_VOPS
|TDF_MEMSYMS
);
2307 /* Do the main work of conditional store replacement. */
2310 cond_if_else_store_replacement_1 (basic_block then_bb
, basic_block else_bb
,
2311 basic_block join_bb
, gimple
*then_assign
,
2312 gimple
*else_assign
)
2314 tree lhs_base
, lhs
, then_rhs
, else_rhs
, name
;
2315 location_t then_locus
, else_locus
;
2316 gimple_stmt_iterator gsi
;
2320 if (then_assign
== NULL
2321 || !gimple_assign_single_p (then_assign
)
2322 || gimple_clobber_p (then_assign
)
2323 || gimple_has_volatile_ops (then_assign
)
2324 || else_assign
== NULL
2325 || !gimple_assign_single_p (else_assign
)
2326 || gimple_clobber_p (else_assign
)
2327 || gimple_has_volatile_ops (else_assign
))
2330 lhs
= gimple_assign_lhs (then_assign
);
2331 if (!is_gimple_reg_type (TREE_TYPE (lhs
))
2332 || !operand_equal_p (lhs
, gimple_assign_lhs (else_assign
), 0))
2335 lhs_base
= get_base_address (lhs
);
2336 if (lhs_base
== NULL_TREE
2337 || (!DECL_P (lhs_base
) && TREE_CODE (lhs_base
) != MEM_REF
))
2340 then_rhs
= gimple_assign_rhs1 (then_assign
);
2341 else_rhs
= gimple_assign_rhs1 (else_assign
);
2342 then_locus
= gimple_location (then_assign
);
2343 else_locus
= gimple_location (else_assign
);
2345 /* Now we've checked the constraints, so do the transformation:
2346 1) Remove the stores. */
2347 gsi
= gsi_for_stmt (then_assign
);
2348 unlink_stmt_vdef (then_assign
);
2349 gsi_remove (&gsi
, true);
2350 release_defs (then_assign
);
2352 gsi
= gsi_for_stmt (else_assign
);
2353 unlink_stmt_vdef (else_assign
);
2354 gsi_remove (&gsi
, true);
2355 release_defs (else_assign
);
2357 /* 2) Create a PHI node at the join block, with one argument
2358 holding the old RHS, and the other holding the temporary
2359 where we stored the old memory contents. */
2360 name
= make_temp_ssa_name (TREE_TYPE (lhs
), NULL
, "cstore");
2361 newphi
= create_phi_node (name
, join_bb
);
2362 add_phi_arg (newphi
, then_rhs
, EDGE_SUCC (then_bb
, 0), then_locus
);
2363 add_phi_arg (newphi
, else_rhs
, EDGE_SUCC (else_bb
, 0), else_locus
);
2365 new_stmt
= gimple_build_assign (lhs
, PHI_RESULT (newphi
));
2367 /* 3) Insert that PHI node. */
2368 gsi
= gsi_after_labels (join_bb
);
2369 if (gsi_end_p (gsi
))
2371 gsi
= gsi_last_bb (join_bb
);
2372 gsi_insert_after (&gsi
, new_stmt
, GSI_NEW_STMT
);
2375 gsi_insert_before (&gsi
, new_stmt
, GSI_NEW_STMT
);
2380 /* Return the single store in BB with VDEF or NULL if there are
2381 other stores in the BB or loads following the store. */
2384 single_trailing_store_in_bb (basic_block bb
, tree vdef
)
2386 if (SSA_NAME_IS_DEFAULT_DEF (vdef
))
2388 gimple
*store
= SSA_NAME_DEF_STMT (vdef
);
2389 if (gimple_bb (store
) != bb
2390 || gimple_code (store
) == GIMPLE_PHI
)
2393 /* Verify there is no other store in this BB. */
2394 if (!SSA_NAME_IS_DEFAULT_DEF (gimple_vuse (store
))
2395 && gimple_bb (SSA_NAME_DEF_STMT (gimple_vuse (store
))) == bb
2396 && gimple_code (SSA_NAME_DEF_STMT (gimple_vuse (store
))) != GIMPLE_PHI
)
2399 /* Verify there is no load or store after the store. */
2400 use_operand_p use_p
;
2401 imm_use_iterator imm_iter
;
2402 FOR_EACH_IMM_USE_FAST (use_p
, imm_iter
, gimple_vdef (store
))
2403 if (USE_STMT (use_p
) != store
2404 && gimple_bb (USE_STMT (use_p
)) == bb
)
2410 /* Conditional store replacement. We already know
2411 that the recognized pattern looks like so:
2414 if (cond) goto THEN_BB; else goto ELSE_BB (edge E1)
2424 fallthrough (edge E0)
2428 We check that it is safe to sink the store to JOIN_BB by verifying that
2429 there are no read-after-write or write-after-write dependencies in
2430 THEN_BB and ELSE_BB. */
2433 cond_if_else_store_replacement (basic_block then_bb
, basic_block else_bb
,
2434 basic_block join_bb
)
2436 vec
<data_reference_p
> then_datarefs
, else_datarefs
;
2437 vec
<ddr_p
> then_ddrs
, else_ddrs
;
2438 gimple
*then_store
, *else_store
;
2439 bool found
, ok
= false, res
;
2440 struct data_dependence_relation
*ddr
;
2441 data_reference_p then_dr
, else_dr
;
2443 tree then_lhs
, else_lhs
;
2444 basic_block blocks
[3];
2446 /* Handle the case with single store in THEN_BB and ELSE_BB. That is
2447 cheap enough to always handle as it allows us to elide dependence
2450 for (gphi_iterator si
= gsi_start_phis (join_bb
); !gsi_end_p (si
);
2452 if (virtual_operand_p (gimple_phi_result (si
.phi ())))
2459 tree then_vdef
= PHI_ARG_DEF_FROM_EDGE (vphi
, single_succ_edge (then_bb
));
2460 tree else_vdef
= PHI_ARG_DEF_FROM_EDGE (vphi
, single_succ_edge (else_bb
));
2461 gimple
*then_assign
= single_trailing_store_in_bb (then_bb
, then_vdef
);
2464 gimple
*else_assign
= single_trailing_store_in_bb (else_bb
, else_vdef
);
2466 return cond_if_else_store_replacement_1 (then_bb
, else_bb
, join_bb
,
2467 then_assign
, else_assign
);
2470 /* If either vectorization or if-conversion is disabled then do
2471 not sink any stores. */
2472 if (param_max_stores_to_sink
== 0
2473 || (!flag_tree_loop_vectorize
&& !flag_tree_slp_vectorize
)
2474 || !flag_tree_loop_if_convert
)
2477 /* Find data references. */
2478 then_datarefs
.create (1);
2479 else_datarefs
.create (1);
2480 if ((find_data_references_in_bb (NULL
, then_bb
, &then_datarefs
)
2482 || !then_datarefs
.length ()
2483 || (find_data_references_in_bb (NULL
, else_bb
, &else_datarefs
)
2485 || !else_datarefs
.length ())
2487 free_data_refs (then_datarefs
);
2488 free_data_refs (else_datarefs
);
2492 /* Find pairs of stores with equal LHS. */
2493 auto_vec
<gimple
*, 1> then_stores
, else_stores
;
2494 FOR_EACH_VEC_ELT (then_datarefs
, i
, then_dr
)
2496 if (DR_IS_READ (then_dr
))
2499 then_store
= DR_STMT (then_dr
);
2500 then_lhs
= gimple_get_lhs (then_store
);
2501 if (then_lhs
== NULL_TREE
)
2505 FOR_EACH_VEC_ELT (else_datarefs
, j
, else_dr
)
2507 if (DR_IS_READ (else_dr
))
2510 else_store
= DR_STMT (else_dr
);
2511 else_lhs
= gimple_get_lhs (else_store
);
2512 if (else_lhs
== NULL_TREE
)
2515 if (operand_equal_p (then_lhs
, else_lhs
, 0))
2525 then_stores
.safe_push (then_store
);
2526 else_stores
.safe_push (else_store
);
2529 /* No pairs of stores found. */
2530 if (!then_stores
.length ()
2531 || then_stores
.length () > (unsigned) param_max_stores_to_sink
)
2533 free_data_refs (then_datarefs
);
2534 free_data_refs (else_datarefs
);
2538 /* Compute and check data dependencies in both basic blocks. */
2539 then_ddrs
.create (1);
2540 else_ddrs
.create (1);
2541 if (!compute_all_dependences (then_datarefs
, &then_ddrs
,
2543 || !compute_all_dependences (else_datarefs
, &else_ddrs
,
2546 free_dependence_relations (then_ddrs
);
2547 free_dependence_relations (else_ddrs
);
2548 free_data_refs (then_datarefs
);
2549 free_data_refs (else_datarefs
);
2552 blocks
[0] = then_bb
;
2553 blocks
[1] = else_bb
;
2554 blocks
[2] = join_bb
;
2555 renumber_gimple_stmt_uids_in_blocks (blocks
, 3);
2557 /* Check that there are no read-after-write or write-after-write dependencies
2559 FOR_EACH_VEC_ELT (then_ddrs
, i
, ddr
)
2561 struct data_reference
*dra
= DDR_A (ddr
);
2562 struct data_reference
*drb
= DDR_B (ddr
);
2564 if (DDR_ARE_DEPENDENT (ddr
) != chrec_known
2565 && ((DR_IS_READ (dra
) && DR_IS_WRITE (drb
)
2566 && gimple_uid (DR_STMT (dra
)) > gimple_uid (DR_STMT (drb
)))
2567 || (DR_IS_READ (drb
) && DR_IS_WRITE (dra
)
2568 && gimple_uid (DR_STMT (drb
)) > gimple_uid (DR_STMT (dra
)))
2569 || (DR_IS_WRITE (dra
) && DR_IS_WRITE (drb
))))
2571 free_dependence_relations (then_ddrs
);
2572 free_dependence_relations (else_ddrs
);
2573 free_data_refs (then_datarefs
);
2574 free_data_refs (else_datarefs
);
2579 /* Check that there are no read-after-write or write-after-write dependencies
2581 FOR_EACH_VEC_ELT (else_ddrs
, i
, ddr
)
2583 struct data_reference
*dra
= DDR_A (ddr
);
2584 struct data_reference
*drb
= DDR_B (ddr
);
2586 if (DDR_ARE_DEPENDENT (ddr
) != chrec_known
2587 && ((DR_IS_READ (dra
) && DR_IS_WRITE (drb
)
2588 && gimple_uid (DR_STMT (dra
)) > gimple_uid (DR_STMT (drb
)))
2589 || (DR_IS_READ (drb
) && DR_IS_WRITE (dra
)
2590 && gimple_uid (DR_STMT (drb
)) > gimple_uid (DR_STMT (dra
)))
2591 || (DR_IS_WRITE (dra
) && DR_IS_WRITE (drb
))))
2593 free_dependence_relations (then_ddrs
);
2594 free_dependence_relations (else_ddrs
);
2595 free_data_refs (then_datarefs
);
2596 free_data_refs (else_datarefs
);
2601 /* Sink stores with same LHS. */
2602 FOR_EACH_VEC_ELT (then_stores
, i
, then_store
)
2604 else_store
= else_stores
[i
];
2605 res
= cond_if_else_store_replacement_1 (then_bb
, else_bb
, join_bb
,
2606 then_store
, else_store
);
2610 free_dependence_relations (then_ddrs
);
2611 free_dependence_relations (else_ddrs
);
2612 free_data_refs (then_datarefs
);
2613 free_data_refs (else_datarefs
);
2618 /* Return TRUE if STMT has a VUSE whose corresponding VDEF is in BB. */
2621 local_mem_dependence (gimple
*stmt
, basic_block bb
)
2623 tree vuse
= gimple_vuse (stmt
);
2629 def
= SSA_NAME_DEF_STMT (vuse
);
2630 return (def
&& gimple_bb (def
) == bb
);
2633 /* Given a "diamond" control-flow pattern where BB0 tests a condition,
2634 BB1 and BB2 are "then" and "else" blocks dependent on this test,
2635 and BB3 rejoins control flow following BB1 and BB2, look for
2636 opportunities to hoist loads as follows. If BB3 contains a PHI of
2637 two loads, one each occurring in BB1 and BB2, and the loads are
2638 provably of adjacent fields in the same structure, then move both
2639 loads into BB0. Of course this can only be done if there are no
2640 dependencies preventing such motion.
2642 One of the hoisted loads will always be speculative, so the
2643 transformation is currently conservative:
2645 - The fields must be strictly adjacent.
2646 - The two fields must occupy a single memory block that is
2647 guaranteed to not cross a page boundary.
2649 The last is difficult to prove, as such memory blocks should be
2650 aligned on the minimum of the stack alignment boundary and the
2651 alignment guaranteed by heap allocation interfaces. Thus we rely
2652 on a parameter for the alignment value.
2654 Provided a good value is used for the last case, the first
2655 restriction could possibly be relaxed. */
2658 hoist_adjacent_loads (basic_block bb0
, basic_block bb1
,
2659 basic_block bb2
, basic_block bb3
)
2661 int param_align
= param_l1_cache_line_size
;
2662 unsigned param_align_bits
= (unsigned) (param_align
* BITS_PER_UNIT
);
2665 /* Walk the phis in bb3 looking for an opportunity. We are looking
2666 for phis of two SSA names, one each of which is defined in bb1 and
2668 for (gsi
= gsi_start_phis (bb3
); !gsi_end_p (gsi
); gsi_next (&gsi
))
2670 gphi
*phi_stmt
= gsi
.phi ();
2671 gimple
*def1
, *def2
;
2672 tree arg1
, arg2
, ref1
, ref2
, field1
, field2
;
2673 tree tree_offset1
, tree_offset2
, tree_size2
, next
;
2674 int offset1
, offset2
, size2
;
2676 gimple_stmt_iterator gsi2
;
2677 basic_block bb_for_def1
, bb_for_def2
;
2679 if (gimple_phi_num_args (phi_stmt
) != 2
2680 || virtual_operand_p (gimple_phi_result (phi_stmt
)))
2683 arg1
= gimple_phi_arg_def (phi_stmt
, 0);
2684 arg2
= gimple_phi_arg_def (phi_stmt
, 1);
2686 if (TREE_CODE (arg1
) != SSA_NAME
2687 || TREE_CODE (arg2
) != SSA_NAME
2688 || SSA_NAME_IS_DEFAULT_DEF (arg1
)
2689 || SSA_NAME_IS_DEFAULT_DEF (arg2
))
2692 def1
= SSA_NAME_DEF_STMT (arg1
);
2693 def2
= SSA_NAME_DEF_STMT (arg2
);
2695 if ((gimple_bb (def1
) != bb1
|| gimple_bb (def2
) != bb2
)
2696 && (gimple_bb (def2
) != bb1
|| gimple_bb (def1
) != bb2
))
2699 /* Check the mode of the arguments to be sure a conditional move
2700 can be generated for it. */
2701 if (optab_handler (movcc_optab
, TYPE_MODE (TREE_TYPE (arg1
)))
2702 == CODE_FOR_nothing
)
2705 /* Both statements must be assignments whose RHS is a COMPONENT_REF. */
2706 if (!gimple_assign_single_p (def1
)
2707 || !gimple_assign_single_p (def2
)
2708 || gimple_has_volatile_ops (def1
)
2709 || gimple_has_volatile_ops (def2
))
2712 ref1
= gimple_assign_rhs1 (def1
);
2713 ref2
= gimple_assign_rhs1 (def2
);
2715 if (TREE_CODE (ref1
) != COMPONENT_REF
2716 || TREE_CODE (ref2
) != COMPONENT_REF
)
2719 /* The zeroth operand of the two component references must be
2720 identical. It is not sufficient to compare get_base_address of
2721 the two references, because this could allow for different
2722 elements of the same array in the two trees. It is not safe to
2723 assume that the existence of one array element implies the
2724 existence of a different one. */
2725 if (!operand_equal_p (TREE_OPERAND (ref1
, 0), TREE_OPERAND (ref2
, 0), 0))
2728 field1
= TREE_OPERAND (ref1
, 1);
2729 field2
= TREE_OPERAND (ref2
, 1);
2731 /* Check for field adjacency, and ensure field1 comes first. */
2732 for (next
= DECL_CHAIN (field1
);
2733 next
&& TREE_CODE (next
) != FIELD_DECL
;
2734 next
= DECL_CHAIN (next
))
2739 for (next
= DECL_CHAIN (field2
);
2740 next
&& TREE_CODE (next
) != FIELD_DECL
;
2741 next
= DECL_CHAIN (next
))
2747 std::swap (field1
, field2
);
2748 std::swap (def1
, def2
);
2751 bb_for_def1
= gimple_bb (def1
);
2752 bb_for_def2
= gimple_bb (def2
);
2754 /* Check for proper alignment of the first field. */
2755 tree_offset1
= bit_position (field1
);
2756 tree_offset2
= bit_position (field2
);
2757 tree_size2
= DECL_SIZE (field2
);
2759 if (!tree_fits_uhwi_p (tree_offset1
)
2760 || !tree_fits_uhwi_p (tree_offset2
)
2761 || !tree_fits_uhwi_p (tree_size2
))
2764 offset1
= tree_to_uhwi (tree_offset1
);
2765 offset2
= tree_to_uhwi (tree_offset2
);
2766 size2
= tree_to_uhwi (tree_size2
);
2767 align1
= DECL_ALIGN (field1
) % param_align_bits
;
2769 if (offset1
% BITS_PER_UNIT
!= 0)
2772 /* For profitability, the two field references should fit within
2773 a single cache line. */
2774 if (align1
+ offset2
- offset1
+ size2
> param_align_bits
)
2777 /* The two expressions cannot be dependent upon vdefs defined
2779 if (local_mem_dependence (def1
, bb_for_def1
)
2780 || local_mem_dependence (def2
, bb_for_def2
))
2783 /* The conditions are satisfied; hoist the loads from bb1 and bb2 into
2784 bb0. We hoist the first one first so that a cache miss is handled
2785 efficiently regardless of hardware cache-fill policy. */
2786 gsi2
= gsi_for_stmt (def1
);
2787 gsi_move_to_bb_end (&gsi2
, bb0
);
2788 gsi2
= gsi_for_stmt (def2
);
2789 gsi_move_to_bb_end (&gsi2
, bb0
);
2791 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
2794 "\nHoisting adjacent loads from %d and %d into %d: \n",
2795 bb_for_def1
->index
, bb_for_def2
->index
, bb0
->index
);
2796 print_gimple_stmt (dump_file
, def1
, 0, TDF_VOPS
|TDF_MEMSYMS
);
2797 print_gimple_stmt (dump_file
, def2
, 0, TDF_VOPS
|TDF_MEMSYMS
);
2802 /* Determine whether we should attempt to hoist adjacent loads out of
2803 diamond patterns in pass_phiopt. Always hoist loads if
2804 -fhoist-adjacent-loads is specified and the target machine has
2805 both a conditional move instruction and a defined cache line size. */
2808 gate_hoist_loads (void)
2810 return (flag_hoist_adjacent_loads
== 1
2811 && param_l1_cache_line_size
2812 && HAVE_conditional_move
);
2815 /* This pass tries to replaces an if-then-else block with an
2816 assignment. We have four kinds of transformations. Some of these
2817 transformations are also performed by the ifcvt RTL optimizer.
2819 Conditional Replacement
2820 -----------------------
2822 This transformation, implemented in conditional_replacement,
2826 if (cond) goto bb2; else goto bb1;
2829 x = PHI <0 (bb1), 1 (bb0), ...>;
2837 x = PHI <x' (bb0), ...>;
2839 We remove bb1 as it becomes unreachable. This occurs often due to
2840 gimplification of conditionals.
2845 This transformation, implemented in value_replacement, replaces
2848 if (a != b) goto bb2; else goto bb1;
2851 x = PHI <a (bb1), b (bb0), ...>;
2857 x = PHI <b (bb0), ...>;
2859 This opportunity can sometimes occur as a result of other
2863 Another case caught by value replacement looks like this:
2869 if (t3 != 0) goto bb1; else goto bb2;
2885 This transformation, implemented in abs_replacement, replaces
2888 if (a >= 0) goto bb2; else goto bb1;
2892 x = PHI <x (bb1), a (bb0), ...>;
2899 x = PHI <x' (bb0), ...>;
2904 This transformation, minmax_replacement replaces
2907 if (a <= b) goto bb2; else goto bb1;
2910 x = PHI <b (bb1), a (bb0), ...>;
2915 x' = MIN_EXPR (a, b)
2917 x = PHI <x' (bb0), ...>;
2919 A similar transformation is done for MAX_EXPR.
2922 This pass also performs a fifth transformation of a slightly different
2925 Factor conversion in COND_EXPR
2926 ------------------------------
2928 This transformation factors the conversion out of COND_EXPR with
2929 factor_out_conditional_conversion.
2932 if (a <= CST) goto <bb 3>; else goto <bb 4>;
2936 tmp = PHI <tmp, CST>
2939 if (a <= CST) goto <bb 3>; else goto <bb 4>;
2945 Adjacent Load Hoisting
2946 ----------------------
2948 This transformation replaces
2951 if (...) goto bb2; else goto bb1;
2953 x1 = (<expr>).field1;
2956 x2 = (<expr>).field2;
2963 x1 = (<expr>).field1;
2964 x2 = (<expr>).field2;
2965 if (...) goto bb2; else goto bb1;
2972 The purpose of this transformation is to enable generation of conditional
2973 move instructions such as Intel CMOVE or PowerPC ISEL. Because one of
2974 the loads is speculative, the transformation is restricted to very
2975 specific cases to avoid introducing a page fault. We are looking for
2983 where left and right are typically adjacent pointers in a tree structure. */
2987 const pass_data pass_data_phiopt
=
2989 GIMPLE_PASS
, /* type */
2990 "phiopt", /* name */
2991 OPTGROUP_NONE
, /* optinfo_flags */
2992 TV_TREE_PHIOPT
, /* tv_id */
2993 ( PROP_cfg
| PROP_ssa
), /* properties_required */
2994 0, /* properties_provided */
2995 0, /* properties_destroyed */
2996 0, /* todo_flags_start */
2997 0, /* todo_flags_finish */
3000 class pass_phiopt
: public gimple_opt_pass
3003 pass_phiopt (gcc::context
*ctxt
)
3004 : gimple_opt_pass (pass_data_phiopt
, ctxt
), early_p (false)
3007 /* opt_pass methods: */
3008 opt_pass
* clone () { return new pass_phiopt (m_ctxt
); }
3009 void set_pass_param (unsigned n
, bool param
)
3011 gcc_assert (n
== 0);
3014 virtual bool gate (function
*) { return flag_ssa_phiopt
; }
3015 virtual unsigned int execute (function
*)
3017 return tree_ssa_phiopt_worker (false,
3018 !early_p
? gate_hoist_loads () : false,
3024 }; // class pass_phiopt
3029 make_pass_phiopt (gcc::context
*ctxt
)
3031 return new pass_phiopt (ctxt
);
3036 const pass_data pass_data_cselim
=
3038 GIMPLE_PASS
, /* type */
3039 "cselim", /* name */
3040 OPTGROUP_NONE
, /* optinfo_flags */
3041 TV_TREE_PHIOPT
, /* tv_id */
3042 ( PROP_cfg
| PROP_ssa
), /* properties_required */
3043 0, /* properties_provided */
3044 0, /* properties_destroyed */
3045 0, /* todo_flags_start */
3046 0, /* todo_flags_finish */
3049 class pass_cselim
: public gimple_opt_pass
3052 pass_cselim (gcc::context
*ctxt
)
3053 : gimple_opt_pass (pass_data_cselim
, ctxt
)
3056 /* opt_pass methods: */
3057 virtual bool gate (function
*) { return flag_tree_cselim
; }
3058 virtual unsigned int execute (function
*) { return tree_ssa_cs_elim (); }
3060 }; // class pass_cselim
3065 make_pass_cselim (gcc::context
*ctxt
)
3067 return new pass_cselim (ctxt
);