1 /* Optimization of PHI nodes by converting them into straightline code.
2 Copyright (C) 2004-2017 Free Software Foundation, Inc.
4 This file is part of GCC.
6 GCC is free software; you can redistribute it and/or modify it
7 under the terms of the GNU General Public License as published by the
8 Free Software Foundation; either version 3, or (at your option) any
11 GCC is distributed in the hope that it will be useful, but WITHOUT
12 ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
13 FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
16 You should have received a copy of the GNU General Public License
17 along with GCC; see the file COPYING3. If not see
18 <http://www.gnu.org/licenses/>. */
22 #include "coretypes.h"
24 #include "insn-codes.h"
29 #include "tree-pass.h"
31 #include "optabs-tree.h"
32 #include "insn-config.h"
33 #include "gimple-pretty-print.h"
34 #include "fold-const.h"
35 #include "stor-layout.h"
38 #include "gimple-iterator.h"
39 #include "gimplify-me.h"
44 #include "tree-data-ref.h"
45 #include "tree-scalar-evolution.h"
46 #include "tree-inline.h"
49 static unsigned int tree_ssa_phiopt_worker (bool, bool);
50 static bool conditional_replacement (basic_block
, basic_block
,
51 edge
, edge
, gphi
*, tree
, tree
);
52 static gphi
*factor_out_conditional_conversion (edge
, edge
, gphi
*, tree
, tree
,
54 static int value_replacement (basic_block
, basic_block
,
55 edge
, edge
, gimple
*, tree
, tree
);
56 static bool minmax_replacement (basic_block
, basic_block
,
57 edge
, edge
, gimple
*, tree
, tree
);
58 static bool abs_replacement (basic_block
, basic_block
,
59 edge
, edge
, gimple
*, tree
, tree
);
60 static bool cond_store_replacement (basic_block
, basic_block
, edge
, edge
,
62 static bool cond_if_else_store_replacement (basic_block
, basic_block
, basic_block
);
63 static hash_set
<tree
> * get_non_trapping ();
64 static void replace_phi_edge_with_variable (basic_block
, edge
, gimple
*, tree
);
65 static void hoist_adjacent_loads (basic_block
, basic_block
,
66 basic_block
, basic_block
);
67 static bool gate_hoist_loads (void);
69 /* This pass tries to transform conditional stores into unconditional
70 ones, enabling further simplifications with the simpler then and else
71 blocks. In particular it replaces this:
74 if (cond) goto bb2; else goto bb1;
82 if (cond) goto bb1; else goto bb2;
86 condtmp = PHI <RHS, condtmp'>
89 This transformation can only be done under several constraints,
90 documented below. It also replaces:
93 if (cond) goto bb2; else goto bb1;
104 if (cond) goto bb3; else goto bb1;
107 condtmp = PHI <RHS1, RHS2>
111 tree_ssa_cs_elim (void)
114 /* ??? We are not interested in loop related info, but the following
115 will create it, ICEing as we didn't init loops with pre-headers.
116 An interfacing issue of find_data_references_in_bb. */
117 loop_optimizer_init (LOOPS_NORMAL
);
119 todo
= tree_ssa_phiopt_worker (true, false);
121 loop_optimizer_finalize ();
125 /* Return the singleton PHI in the SEQ of PHIs for edges E0 and E1. */
128 single_non_singleton_phi_for_edges (gimple_seq seq
, edge e0
, edge e1
)
130 gimple_stmt_iterator i
;
132 if (gimple_seq_singleton_p (seq
))
133 return as_a
<gphi
*> (gsi_stmt (gsi_start (seq
)));
134 for (i
= gsi_start (seq
); !gsi_end_p (i
); gsi_next (&i
))
136 gphi
*p
= as_a
<gphi
*> (gsi_stmt (i
));
137 /* If the PHI arguments are equal then we can skip this PHI. */
138 if (operand_equal_for_phi_arg_p (gimple_phi_arg_def (p
, e0
->dest_idx
),
139 gimple_phi_arg_def (p
, e1
->dest_idx
)))
142 /* If we already have a PHI that has the two edge arguments are
143 different, then return it is not a singleton for these PHIs. */
152 /* The core routine of conditional store replacement and normal
153 phi optimizations. Both share much of the infrastructure in how
154 to match applicable basic block patterns. DO_STORE_ELIM is true
155 when we want to do conditional store replacement, false otherwise.
156 DO_HOIST_LOADS is true when we want to hoist adjacent loads out
157 of diamond control flow patterns, false otherwise. */
159 tree_ssa_phiopt_worker (bool do_store_elim
, bool do_hoist_loads
)
162 basic_block
*bb_order
;
164 bool cfgchanged
= false;
165 hash_set
<tree
> *nontrap
= 0;
168 /* Calculate the set of non-trapping memory accesses. */
169 nontrap
= get_non_trapping ();
171 /* Search every basic block for COND_EXPR we may be able to optimize.
173 We walk the blocks in order that guarantees that a block with
174 a single predecessor is processed before the predecessor.
175 This ensures that we collapse inner ifs before visiting the
176 outer ones, and also that we do not try to visit a removed
178 bb_order
= single_pred_before_succ_order ();
179 n
= n_basic_blocks_for_fn (cfun
) - NUM_FIXED_BLOCKS
;
181 for (i
= 0; i
< n
; i
++)
185 basic_block bb1
, bb2
;
191 cond_stmt
= last_stmt (bb
);
192 /* Check to see if the last statement is a GIMPLE_COND. */
194 || gimple_code (cond_stmt
) != GIMPLE_COND
)
197 e1
= EDGE_SUCC (bb
, 0);
199 e2
= EDGE_SUCC (bb
, 1);
202 /* We cannot do the optimization on abnormal edges. */
203 if ((e1
->flags
& EDGE_ABNORMAL
) != 0
204 || (e2
->flags
& EDGE_ABNORMAL
) != 0)
207 /* If either bb1's succ or bb2 or bb2's succ is non NULL. */
208 if (EDGE_COUNT (bb1
->succs
) == 0
210 || EDGE_COUNT (bb2
->succs
) == 0)
213 /* Find the bb which is the fall through to the other. */
214 if (EDGE_SUCC (bb1
, 0)->dest
== bb2
)
216 else if (EDGE_SUCC (bb2
, 0)->dest
== bb1
)
218 std::swap (bb1
, bb2
);
221 else if (do_store_elim
222 && EDGE_SUCC (bb1
, 0)->dest
== EDGE_SUCC (bb2
, 0)->dest
)
224 basic_block bb3
= EDGE_SUCC (bb1
, 0)->dest
;
226 if (!single_succ_p (bb1
)
227 || (EDGE_SUCC (bb1
, 0)->flags
& EDGE_FALLTHRU
) == 0
228 || !single_succ_p (bb2
)
229 || (EDGE_SUCC (bb2
, 0)->flags
& EDGE_FALLTHRU
) == 0
230 || EDGE_COUNT (bb3
->preds
) != 2)
232 if (cond_if_else_store_replacement (bb1
, bb2
, bb3
))
236 else if (do_hoist_loads
237 && EDGE_SUCC (bb1
, 0)->dest
== EDGE_SUCC (bb2
, 0)->dest
)
239 basic_block bb3
= EDGE_SUCC (bb1
, 0)->dest
;
241 if (!FLOAT_TYPE_P (TREE_TYPE (gimple_cond_lhs (cond_stmt
)))
242 && single_succ_p (bb1
)
243 && single_succ_p (bb2
)
244 && single_pred_p (bb1
)
245 && single_pred_p (bb2
)
246 && EDGE_COUNT (bb
->succs
) == 2
247 && EDGE_COUNT (bb3
->preds
) == 2
248 /* If one edge or the other is dominant, a conditional move
249 is likely to perform worse than the well-predicted branch. */
250 && !predictable_edge_p (EDGE_SUCC (bb
, 0))
251 && !predictable_edge_p (EDGE_SUCC (bb
, 1)))
252 hoist_adjacent_loads (bb
, bb1
, bb2
, bb3
);
258 e1
= EDGE_SUCC (bb1
, 0);
260 /* Make sure that bb1 is just a fall through. */
261 if (!single_succ_p (bb1
)
262 || (e1
->flags
& EDGE_FALLTHRU
) == 0)
265 /* Also make sure that bb1 only have one predecessor and that it
267 if (!single_pred_p (bb1
)
268 || single_pred (bb1
) != bb
)
273 /* bb1 is the middle block, bb2 the join block, bb the split block,
274 e1 the fallthrough edge from bb1 to bb2. We can't do the
275 optimization if the join block has more than two predecessors. */
276 if (EDGE_COUNT (bb2
->preds
) > 2)
278 if (cond_store_replacement (bb1
, bb2
, e1
, e2
, nontrap
))
283 gimple_seq phis
= phi_nodes (bb2
);
284 gimple_stmt_iterator gsi
;
285 bool candorest
= true;
287 /* Value replacement can work with more than one PHI
288 so try that first. */
289 for (gsi
= gsi_start (phis
); !gsi_end_p (gsi
); gsi_next (&gsi
))
291 phi
= as_a
<gphi
*> (gsi_stmt (gsi
));
292 arg0
= gimple_phi_arg_def (phi
, e1
->dest_idx
);
293 arg1
= gimple_phi_arg_def (phi
, e2
->dest_idx
);
294 if (value_replacement (bb
, bb1
, e1
, e2
, phi
, arg0
, arg1
) == 2)
305 phi
= single_non_singleton_phi_for_edges (phis
, e1
, e2
);
309 arg0
= gimple_phi_arg_def (phi
, e1
->dest_idx
);
310 arg1
= gimple_phi_arg_def (phi
, e2
->dest_idx
);
312 /* Something is wrong if we cannot find the arguments in the PHI
314 gcc_assert (arg0
!= NULL_TREE
&& arg1
!= NULL_TREE
);
316 gphi
*newphi
= factor_out_conditional_conversion (e1
, e2
, phi
,
322 /* factor_out_conditional_conversion may create a new PHI in
323 BB2 and eliminate an existing PHI in BB2. Recompute values
324 that may be affected by that change. */
325 arg0
= gimple_phi_arg_def (phi
, e1
->dest_idx
);
326 arg1
= gimple_phi_arg_def (phi
, e2
->dest_idx
);
327 gcc_assert (arg0
!= NULL_TREE
&& arg1
!= NULL_TREE
);
330 /* Do the replacement of conditional if it can be done. */
331 if (conditional_replacement (bb
, bb1
, e1
, e2
, phi
, arg0
, arg1
))
333 else if (abs_replacement (bb
, bb1
, e1
, e2
, phi
, arg0
, arg1
))
335 else if (minmax_replacement (bb
, bb1
, e1
, e2
, phi
, arg0
, arg1
))
344 /* If the CFG has changed, we should cleanup the CFG. */
345 if (cfgchanged
&& do_store_elim
)
347 /* In cond-store replacement we have added some loads on edges
348 and new VOPS (as we moved the store, and created a load). */
349 gsi_commit_edge_inserts ();
350 return TODO_cleanup_cfg
| TODO_update_ssa_only_virtuals
;
353 return TODO_cleanup_cfg
;
357 /* Replace PHI node element whose edge is E in block BB with variable NEW.
358 Remove the edge from COND_BLOCK which does not lead to BB (COND_BLOCK
359 is known to have two edges, one of which must reach BB). */
362 replace_phi_edge_with_variable (basic_block cond_block
,
363 edge e
, gimple
*phi
, tree new_tree
)
365 basic_block bb
= gimple_bb (phi
);
366 basic_block block_to_remove
;
367 gimple_stmt_iterator gsi
;
369 /* Change the PHI argument to new. */
370 SET_USE (PHI_ARG_DEF_PTR (phi
, e
->dest_idx
), new_tree
);
372 /* Remove the empty basic block. */
373 if (EDGE_SUCC (cond_block
, 0)->dest
== bb
)
375 EDGE_SUCC (cond_block
, 0)->flags
|= EDGE_FALLTHRU
;
376 EDGE_SUCC (cond_block
, 0)->flags
&= ~(EDGE_TRUE_VALUE
| EDGE_FALSE_VALUE
);
377 EDGE_SUCC (cond_block
, 0)->probability
= profile_probability::always ();
378 EDGE_SUCC (cond_block
, 0)->count
+= EDGE_SUCC (cond_block
, 1)->count
;
380 block_to_remove
= EDGE_SUCC (cond_block
, 1)->dest
;
384 EDGE_SUCC (cond_block
, 1)->flags
|= EDGE_FALLTHRU
;
385 EDGE_SUCC (cond_block
, 1)->flags
386 &= ~(EDGE_TRUE_VALUE
| EDGE_FALSE_VALUE
);
387 EDGE_SUCC (cond_block
, 1)->probability
= profile_probability::always ();
388 EDGE_SUCC (cond_block
, 1)->count
+= EDGE_SUCC (cond_block
, 0)->count
;
390 block_to_remove
= EDGE_SUCC (cond_block
, 0)->dest
;
392 delete_basic_block (block_to_remove
);
394 /* Eliminate the COND_EXPR at the end of COND_BLOCK. */
395 gsi
= gsi_last_bb (cond_block
);
396 gsi_remove (&gsi
, true);
398 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
400 "COND_EXPR in block %d and PHI in block %d converted to straightline code.\n",
405 /* PR66726: Factor conversion out of COND_EXPR. If the arguments of the PHI
406 stmt are CONVERT_STMT, factor out the conversion and perform the conversion
407 to the result of PHI stmt. COND_STMT is the controlling predicate.
408 Return the newly-created PHI, if any. */
411 factor_out_conditional_conversion (edge e0
, edge e1
, gphi
*phi
,
412 tree arg0
, tree arg1
, gimple
*cond_stmt
)
414 gimple
*arg0_def_stmt
= NULL
, *arg1_def_stmt
= NULL
, *new_stmt
;
415 tree new_arg0
= NULL_TREE
, new_arg1
= NULL_TREE
;
418 gimple_stmt_iterator gsi
, gsi_for_def
;
419 source_location locus
= gimple_location (phi
);
420 enum tree_code convert_code
;
422 /* Handle only PHI statements with two arguments. TODO: If all
423 other arguments to PHI are INTEGER_CST or if their defining
424 statement have the same unary operation, we can handle more
425 than two arguments too. */
426 if (gimple_phi_num_args (phi
) != 2)
429 /* First canonicalize to simplify tests. */
430 if (TREE_CODE (arg0
) != SSA_NAME
)
432 std::swap (arg0
, arg1
);
436 if (TREE_CODE (arg0
) != SSA_NAME
437 || (TREE_CODE (arg1
) != SSA_NAME
438 && TREE_CODE (arg1
) != INTEGER_CST
))
441 /* Check if arg0 is an SSA_NAME and the stmt which defines arg0 is
443 arg0_def_stmt
= SSA_NAME_DEF_STMT (arg0
);
444 if (!gimple_assign_cast_p (arg0_def_stmt
))
447 /* Use the RHS as new_arg0. */
448 convert_code
= gimple_assign_rhs_code (arg0_def_stmt
);
449 new_arg0
= gimple_assign_rhs1 (arg0_def_stmt
);
450 if (convert_code
== VIEW_CONVERT_EXPR
)
452 new_arg0
= TREE_OPERAND (new_arg0
, 0);
453 if (!is_gimple_reg_type (TREE_TYPE (new_arg0
)))
457 if (TREE_CODE (arg1
) == SSA_NAME
)
459 /* Check if arg1 is an SSA_NAME and the stmt which defines arg1
461 arg1_def_stmt
= SSA_NAME_DEF_STMT (arg1
);
462 if (!is_gimple_assign (arg1_def_stmt
)
463 || gimple_assign_rhs_code (arg1_def_stmt
) != convert_code
)
466 /* Use the RHS as new_arg1. */
467 new_arg1
= gimple_assign_rhs1 (arg1_def_stmt
);
468 if (convert_code
== VIEW_CONVERT_EXPR
)
469 new_arg1
= TREE_OPERAND (new_arg1
, 0);
473 /* If arg1 is an INTEGER_CST, fold it to new type. */
474 if (INTEGRAL_TYPE_P (TREE_TYPE (new_arg0
))
475 && int_fits_type_p (arg1
, TREE_TYPE (new_arg0
)))
477 if (gimple_assign_cast_p (arg0_def_stmt
))
479 /* For the INTEGER_CST case, we are just moving the
480 conversion from one place to another, which can often
481 hurt as the conversion moves further away from the
482 statement that computes the value. So, perform this
483 only if new_arg0 is an operand of COND_STMT, or
484 if arg0_def_stmt is the only non-debug stmt in
485 its basic block, because then it is possible this
486 could enable further optimizations (minmax replacement
487 etc.). See PR71016. */
488 if (new_arg0
!= gimple_cond_lhs (cond_stmt
)
489 && new_arg0
!= gimple_cond_rhs (cond_stmt
)
490 && gimple_bb (arg0_def_stmt
) == e0
->src
)
492 gsi
= gsi_for_stmt (arg0_def_stmt
);
493 gsi_prev_nondebug (&gsi
);
494 if (!gsi_end_p (gsi
))
496 gsi
= gsi_for_stmt (arg0_def_stmt
);
497 gsi_next_nondebug (&gsi
);
498 if (!gsi_end_p (gsi
))
501 new_arg1
= fold_convert (TREE_TYPE (new_arg0
), arg1
);
510 /* If arg0/arg1 have > 1 use, then this transformation actually increases
511 the number of expressions evaluated at runtime. */
512 if (!has_single_use (arg0
)
513 || (arg1_def_stmt
&& !has_single_use (arg1
)))
516 /* If types of new_arg0 and new_arg1 are different bailout. */
517 if (!types_compatible_p (TREE_TYPE (new_arg0
), TREE_TYPE (new_arg1
)))
520 /* Create a new PHI stmt. */
521 result
= PHI_RESULT (phi
);
522 temp
= make_ssa_name (TREE_TYPE (new_arg0
), NULL
);
523 newphi
= create_phi_node (temp
, gimple_bb (phi
));
525 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
527 fprintf (dump_file
, "PHI ");
528 print_generic_expr (dump_file
, gimple_phi_result (phi
));
530 " changed to factor conversion out from COND_EXPR.\n");
531 fprintf (dump_file
, "New stmt with CAST that defines ");
532 print_generic_expr (dump_file
, result
);
533 fprintf (dump_file
, ".\n");
536 /* Remove the old cast(s) that has single use. */
537 gsi_for_def
= gsi_for_stmt (arg0_def_stmt
);
538 gsi_remove (&gsi_for_def
, true);
539 release_defs (arg0_def_stmt
);
543 gsi_for_def
= gsi_for_stmt (arg1_def_stmt
);
544 gsi_remove (&gsi_for_def
, true);
545 release_defs (arg1_def_stmt
);
548 add_phi_arg (newphi
, new_arg0
, e0
, locus
);
549 add_phi_arg (newphi
, new_arg1
, e1
, locus
);
551 /* Create the conversion stmt and insert it. */
552 if (convert_code
== VIEW_CONVERT_EXPR
)
553 temp
= fold_build1 (VIEW_CONVERT_EXPR
, TREE_TYPE (result
), temp
);
554 new_stmt
= gimple_build_assign (result
, convert_code
, temp
);
555 gsi
= gsi_after_labels (gimple_bb (phi
));
556 gsi_insert_before (&gsi
, new_stmt
, GSI_SAME_STMT
);
558 /* Remove the original PHI stmt. */
559 gsi
= gsi_for_stmt (phi
);
560 gsi_remove (&gsi
, true);
564 /* The function conditional_replacement does the main work of doing the
565 conditional replacement. Return true if the replacement is done.
566 Otherwise return false.
567 BB is the basic block where the replacement is going to be done on. ARG0
568 is argument 0 from PHI. Likewise for ARG1. */
571 conditional_replacement (basic_block cond_bb
, basic_block middle_bb
,
572 edge e0
, edge e1
, gphi
*phi
,
573 tree arg0
, tree arg1
)
579 gimple_stmt_iterator gsi
;
580 edge true_edge
, false_edge
;
581 tree new_var
, new_var2
;
584 /* FIXME: Gimplification of complex type is too hard for now. */
585 /* We aren't prepared to handle vectors either (and it is a question
586 if it would be worthwhile anyway). */
587 if (!(INTEGRAL_TYPE_P (TREE_TYPE (arg0
))
588 || POINTER_TYPE_P (TREE_TYPE (arg0
)))
589 || !(INTEGRAL_TYPE_P (TREE_TYPE (arg1
))
590 || POINTER_TYPE_P (TREE_TYPE (arg1
))))
593 /* The PHI arguments have the constants 0 and 1, or 0 and -1, then
594 convert it to the conditional. */
595 if ((integer_zerop (arg0
) && integer_onep (arg1
))
596 || (integer_zerop (arg1
) && integer_onep (arg0
)))
598 else if ((integer_zerop (arg0
) && integer_all_onesp (arg1
))
599 || (integer_zerop (arg1
) && integer_all_onesp (arg0
)))
604 if (!empty_block_p (middle_bb
))
607 /* At this point we know we have a GIMPLE_COND with two successors.
608 One successor is BB, the other successor is an empty block which
609 falls through into BB.
611 There is a single PHI node at the join point (BB) and its arguments
612 are constants (0, 1) or (0, -1).
614 So, given the condition COND, and the two PHI arguments, we can
615 rewrite this PHI into non-branching code:
617 dest = (COND) or dest = COND'
619 We use the condition as-is if the argument associated with the
620 true edge has the value one or the argument associated with the
621 false edge as the value zero. Note that those conditions are not
622 the same since only one of the outgoing edges from the GIMPLE_COND
623 will directly reach BB and thus be associated with an argument. */
625 stmt
= last_stmt (cond_bb
);
626 result
= PHI_RESULT (phi
);
628 /* To handle special cases like floating point comparison, it is easier and
629 less error-prone to build a tree and gimplify it on the fly though it is
631 cond
= fold_build2_loc (gimple_location (stmt
),
632 gimple_cond_code (stmt
), boolean_type_node
,
633 gimple_cond_lhs (stmt
), gimple_cond_rhs (stmt
));
635 /* We need to know which is the true edge and which is the false
636 edge so that we know when to invert the condition below. */
637 extract_true_false_edges_from_block (cond_bb
, &true_edge
, &false_edge
);
638 if ((e0
== true_edge
&& integer_zerop (arg0
))
639 || (e0
== false_edge
&& !integer_zerop (arg0
))
640 || (e1
== true_edge
&& integer_zerop (arg1
))
641 || (e1
== false_edge
&& !integer_zerop (arg1
)))
642 cond
= fold_build1_loc (gimple_location (stmt
),
643 TRUTH_NOT_EXPR
, TREE_TYPE (cond
), cond
);
647 cond
= fold_convert_loc (gimple_location (stmt
),
648 TREE_TYPE (result
), cond
);
649 cond
= fold_build1_loc (gimple_location (stmt
),
650 NEGATE_EXPR
, TREE_TYPE (cond
), cond
);
653 /* Insert our new statements at the end of conditional block before the
655 gsi
= gsi_for_stmt (stmt
);
656 new_var
= force_gimple_operand_gsi (&gsi
, cond
, true, NULL
, true,
659 if (!useless_type_conversion_p (TREE_TYPE (result
), TREE_TYPE (new_var
)))
661 source_location locus_0
, locus_1
;
663 new_var2
= make_ssa_name (TREE_TYPE (result
));
664 new_stmt
= gimple_build_assign (new_var2
, CONVERT_EXPR
, new_var
);
665 gsi_insert_before (&gsi
, new_stmt
, GSI_SAME_STMT
);
668 /* Set the locus to the first argument, unless is doesn't have one. */
669 locus_0
= gimple_phi_arg_location (phi
, 0);
670 locus_1
= gimple_phi_arg_location (phi
, 1);
671 if (locus_0
== UNKNOWN_LOCATION
)
673 gimple_set_location (new_stmt
, locus_0
);
676 replace_phi_edge_with_variable (cond_bb
, e1
, phi
, new_var
);
677 reset_flow_sensitive_info_in_bb (cond_bb
);
679 /* Note that we optimized this PHI. */
683 /* Update *ARG which is defined in STMT so that it contains the
684 computed value if that seems profitable. Return true if the
685 statement is made dead by that rewriting. */
688 jump_function_from_stmt (tree
*arg
, gimple
*stmt
)
690 enum tree_code code
= gimple_assign_rhs_code (stmt
);
691 if (code
== ADDR_EXPR
)
693 /* For arg = &p->i transform it to p, if possible. */
694 tree rhs1
= gimple_assign_rhs1 (stmt
);
695 HOST_WIDE_INT offset
;
696 tree tem
= get_addr_base_and_unit_offset (TREE_OPERAND (rhs1
, 0),
699 && TREE_CODE (tem
) == MEM_REF
700 && (mem_ref_offset (tem
) + offset
) == 0)
702 *arg
= TREE_OPERAND (tem
, 0);
706 /* TODO: Much like IPA-CP jump-functions we want to handle constant
707 additions symbolically here, and we'd need to update the comparison
708 code that compares the arg + cst tuples in our caller. For now the
709 code above exactly handles the VEC_BASE pattern from vec.h. */
713 /* RHS is a source argument in a BIT_AND_EXPR which feeds a conditional
714 of the form SSA_NAME NE 0.
716 If RHS is fed by a simple EQ_EXPR comparison of two values, see if
717 the two input values of the EQ_EXPR match arg0 and arg1.
719 If so update *code and return TRUE. Otherwise return FALSE. */
722 rhs_is_fed_for_value_replacement (const_tree arg0
, const_tree arg1
,
723 enum tree_code
*code
, const_tree rhs
)
725 /* Obviously if RHS is not an SSA_NAME, we can't look at the defining
727 if (TREE_CODE (rhs
) == SSA_NAME
)
729 gimple
*def1
= SSA_NAME_DEF_STMT (rhs
);
731 /* Verify the defining statement has an EQ_EXPR on the RHS. */
732 if (is_gimple_assign (def1
) && gimple_assign_rhs_code (def1
) == EQ_EXPR
)
734 /* Finally verify the source operands of the EQ_EXPR are equal
736 tree op0
= gimple_assign_rhs1 (def1
);
737 tree op1
= gimple_assign_rhs2 (def1
);
738 if ((operand_equal_for_phi_arg_p (arg0
, op0
)
739 && operand_equal_for_phi_arg_p (arg1
, op1
))
740 || (operand_equal_for_phi_arg_p (arg0
, op1
)
741 && operand_equal_for_phi_arg_p (arg1
, op0
)))
743 /* We will perform the optimization. */
744 *code
= gimple_assign_rhs_code (def1
);
752 /* Return TRUE if arg0/arg1 are equal to the rhs/lhs or lhs/rhs of COND.
754 Also return TRUE if arg0/arg1 are equal to the source arguments of a
755 an EQ comparison feeding a BIT_AND_EXPR which feeds COND.
757 Return FALSE otherwise. */
760 operand_equal_for_value_replacement (const_tree arg0
, const_tree arg1
,
761 enum tree_code
*code
, gimple
*cond
)
764 tree lhs
= gimple_cond_lhs (cond
);
765 tree rhs
= gimple_cond_rhs (cond
);
767 if ((operand_equal_for_phi_arg_p (arg0
, lhs
)
768 && operand_equal_for_phi_arg_p (arg1
, rhs
))
769 || (operand_equal_for_phi_arg_p (arg1
, lhs
)
770 && operand_equal_for_phi_arg_p (arg0
, rhs
)))
773 /* Now handle more complex case where we have an EQ comparison
774 which feeds a BIT_AND_EXPR which feeds COND.
776 First verify that COND is of the form SSA_NAME NE 0. */
777 if (*code
!= NE_EXPR
|| !integer_zerop (rhs
)
778 || TREE_CODE (lhs
) != SSA_NAME
)
781 /* Now ensure that SSA_NAME is set by a BIT_AND_EXPR. */
782 def
= SSA_NAME_DEF_STMT (lhs
);
783 if (!is_gimple_assign (def
) || gimple_assign_rhs_code (def
) != BIT_AND_EXPR
)
786 /* Now verify arg0/arg1 correspond to the source arguments of an
787 EQ comparison feeding the BIT_AND_EXPR. */
789 tree tmp
= gimple_assign_rhs1 (def
);
790 if (rhs_is_fed_for_value_replacement (arg0
, arg1
, code
, tmp
))
793 tmp
= gimple_assign_rhs2 (def
);
794 if (rhs_is_fed_for_value_replacement (arg0
, arg1
, code
, tmp
))
800 /* Returns true if ARG is a neutral element for operation CODE
801 on the RIGHT side. */
804 neutral_element_p (tree_code code
, tree arg
, bool right
)
811 return integer_zerop (arg
);
818 case POINTER_PLUS_EXPR
:
819 return right
&& integer_zerop (arg
);
822 return integer_onep (arg
);
829 return right
&& integer_onep (arg
);
832 return integer_all_onesp (arg
);
839 /* Returns true if ARG is an absorbing element for operation CODE. */
842 absorbing_element_p (tree_code code
, tree arg
, bool right
, tree rval
)
847 return integer_all_onesp (arg
);
851 return integer_zerop (arg
);
857 return !right
&& integer_zerop (arg
);
869 && integer_zerop (arg
)
870 && tree_single_nonzero_warnv_p (rval
, NULL
));
877 /* The function value_replacement does the main work of doing the value
878 replacement. Return non-zero if the replacement is done. Otherwise return
879 0. If we remove the middle basic block, return 2.
880 BB is the basic block where the replacement is going to be done on. ARG0
881 is argument 0 from the PHI. Likewise for ARG1. */
884 value_replacement (basic_block cond_bb
, basic_block middle_bb
,
885 edge e0
, edge e1
, gimple
*phi
,
886 tree arg0
, tree arg1
)
888 gimple_stmt_iterator gsi
;
890 edge true_edge
, false_edge
;
892 bool emtpy_or_with_defined_p
= true;
894 /* If the type says honor signed zeros we cannot do this
896 if (HONOR_SIGNED_ZEROS (arg1
))
899 /* If there is a statement in MIDDLE_BB that defines one of the PHI
900 arguments, then adjust arg0 or arg1. */
901 gsi
= gsi_start_nondebug_after_labels_bb (middle_bb
);
902 while (!gsi_end_p (gsi
))
904 gimple
*stmt
= gsi_stmt (gsi
);
906 gsi_next_nondebug (&gsi
);
907 if (!is_gimple_assign (stmt
))
909 emtpy_or_with_defined_p
= false;
912 /* Now try to adjust arg0 or arg1 according to the computation
914 lhs
= gimple_assign_lhs (stmt
);
916 && jump_function_from_stmt (&arg0
, stmt
))
918 && jump_function_from_stmt (&arg1
, stmt
)))
919 emtpy_or_with_defined_p
= false;
922 cond
= last_stmt (cond_bb
);
923 code
= gimple_cond_code (cond
);
925 /* This transformation is only valid for equality comparisons. */
926 if (code
!= NE_EXPR
&& code
!= EQ_EXPR
)
929 /* We need to know which is the true edge and which is the false
930 edge so that we know if have abs or negative abs. */
931 extract_true_false_edges_from_block (cond_bb
, &true_edge
, &false_edge
);
933 /* At this point we know we have a COND_EXPR with two successors.
934 One successor is BB, the other successor is an empty block which
935 falls through into BB.
937 The condition for the COND_EXPR is known to be NE_EXPR or EQ_EXPR.
939 There is a single PHI node at the join point (BB) with two arguments.
941 We now need to verify that the two arguments in the PHI node match
942 the two arguments to the equality comparison. */
944 if (operand_equal_for_value_replacement (arg0
, arg1
, &code
, cond
))
949 /* For NE_EXPR, we want to build an assignment result = arg where
950 arg is the PHI argument associated with the true edge. For
951 EQ_EXPR we want the PHI argument associated with the false edge. */
952 e
= (code
== NE_EXPR
? true_edge
: false_edge
);
954 /* Unfortunately, E may not reach BB (it may instead have gone to
955 OTHER_BLOCK). If that is the case, then we want the single outgoing
956 edge from OTHER_BLOCK which reaches BB and represents the desired
957 path from COND_BLOCK. */
958 if (e
->dest
== middle_bb
)
959 e
= single_succ_edge (e
->dest
);
961 /* Now we know the incoming edge to BB that has the argument for the
962 RHS of our new assignment statement. */
968 /* If the middle basic block was empty or is defining the
969 PHI arguments and this is a single phi where the args are different
970 for the edges e0 and e1 then we can remove the middle basic block. */
971 if (emtpy_or_with_defined_p
972 && single_non_singleton_phi_for_edges (phi_nodes (gimple_bb (phi
)),
975 replace_phi_edge_with_variable (cond_bb
, e1
, phi
, arg
);
976 /* Note that we optimized this PHI. */
981 /* Replace the PHI arguments with arg. */
982 SET_PHI_ARG_DEF (phi
, e0
->dest_idx
, arg
);
983 SET_PHI_ARG_DEF (phi
, e1
->dest_idx
, arg
);
984 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
986 fprintf (dump_file
, "PHI ");
987 print_generic_expr (dump_file
, gimple_phi_result (phi
));
988 fprintf (dump_file
, " reduced for COND_EXPR in block %d to ",
990 print_generic_expr (dump_file
, arg
);
991 fprintf (dump_file
, ".\n");
998 /* Now optimize (x != 0) ? x + y : y to just x + y. */
999 gsi
= gsi_last_nondebug_bb (middle_bb
);
1000 if (gsi_end_p (gsi
))
1003 gimple
*assign
= gsi_stmt (gsi
);
1004 if (!is_gimple_assign (assign
)
1005 || gimple_assign_rhs_class (assign
) != GIMPLE_BINARY_RHS
1006 || (!INTEGRAL_TYPE_P (TREE_TYPE (arg0
))
1007 && !POINTER_TYPE_P (TREE_TYPE (arg0
))))
1010 /* Punt if there are (degenerate) PHIs in middle_bb, there should not be. */
1011 if (!gimple_seq_empty_p (phi_nodes (middle_bb
)))
1014 /* Allow up to 2 cheap preparation statements that prepare argument
1022 iftmp.0_6 = x_5(D) r<< _1;
1024 # iftmp.0_2 = PHI <iftmp.0_6(3), x_5(D)(2)>
1035 # _2 = PHI <x_5(D)(2), _6(3)> */
1036 gimple
*prep_stmt
[2] = { NULL
, NULL
};
1038 for (prep_cnt
= 0; ; prep_cnt
++)
1040 gsi_prev_nondebug (&gsi
);
1041 if (gsi_end_p (gsi
))
1044 gimple
*g
= gsi_stmt (gsi
);
1045 if (gimple_code (g
) == GIMPLE_LABEL
)
1048 if (prep_cnt
== 2 || !is_gimple_assign (g
))
1051 tree lhs
= gimple_assign_lhs (g
);
1052 tree rhs1
= gimple_assign_rhs1 (g
);
1053 use_operand_p use_p
;
1055 if (TREE_CODE (lhs
) != SSA_NAME
1056 || TREE_CODE (rhs1
) != SSA_NAME
1057 || !INTEGRAL_TYPE_P (TREE_TYPE (lhs
))
1058 || !INTEGRAL_TYPE_P (TREE_TYPE (rhs1
))
1059 || !single_imm_use (lhs
, &use_p
, &use_stmt
)
1060 || use_stmt
!= (prep_cnt
? prep_stmt
[prep_cnt
- 1] : assign
))
1062 switch (gimple_assign_rhs_code (g
))
1070 if (TREE_CODE (gimple_assign_rhs2 (g
)) != INTEGER_CST
)
1076 prep_stmt
[prep_cnt
] = g
;
1079 /* Only transform if it removes the condition. */
1080 if (!single_non_singleton_phi_for_edges (phi_nodes (gimple_bb (phi
)), e0
, e1
))
1083 /* Size-wise, this is always profitable. */
1084 if (optimize_bb_for_speed_p (cond_bb
)
1085 /* The special case is useless if it has a low probability. */
1086 && profile_status_for_fn (cfun
) != PROFILE_ABSENT
1087 && EDGE_PRED (middle_bb
, 0)->probability
< profile_probability::even ()
1088 /* If assign is cheap, there is no point avoiding it. */
1089 && estimate_num_insns (bb_seq (middle_bb
), &eni_time_weights
)
1090 >= 3 * estimate_num_insns (cond
, &eni_time_weights
))
1093 tree lhs
= gimple_assign_lhs (assign
);
1094 tree rhs1
= gimple_assign_rhs1 (assign
);
1095 tree rhs2
= gimple_assign_rhs2 (assign
);
1096 enum tree_code code_def
= gimple_assign_rhs_code (assign
);
1097 tree cond_lhs
= gimple_cond_lhs (cond
);
1098 tree cond_rhs
= gimple_cond_rhs (cond
);
1100 /* Propagate the cond_rhs constant through preparation stmts,
1101 make sure UB isn't invoked while doing that. */
1102 for (int i
= prep_cnt
- 1; i
>= 0; --i
)
1104 gimple
*g
= prep_stmt
[i
];
1105 tree grhs1
= gimple_assign_rhs1 (g
);
1106 if (!operand_equal_for_phi_arg_p (cond_lhs
, grhs1
))
1108 cond_lhs
= gimple_assign_lhs (g
);
1109 cond_rhs
= fold_convert (TREE_TYPE (grhs1
), cond_rhs
);
1110 if (TREE_CODE (cond_rhs
) != INTEGER_CST
1111 || TREE_OVERFLOW (cond_rhs
))
1113 if (gimple_assign_rhs_class (g
) == GIMPLE_BINARY_RHS
)
1115 cond_rhs
= int_const_binop (gimple_assign_rhs_code (g
), cond_rhs
,
1116 gimple_assign_rhs2 (g
));
1117 if (TREE_OVERFLOW (cond_rhs
))
1120 cond_rhs
= fold_convert (TREE_TYPE (cond_lhs
), cond_rhs
);
1121 if (TREE_CODE (cond_rhs
) != INTEGER_CST
1122 || TREE_OVERFLOW (cond_rhs
))
1126 if (((code
== NE_EXPR
&& e1
== false_edge
)
1127 || (code
== EQ_EXPR
&& e1
== true_edge
))
1130 && operand_equal_for_phi_arg_p (rhs2
, cond_lhs
)
1131 && neutral_element_p (code_def
, cond_rhs
, true))
1133 && operand_equal_for_phi_arg_p (rhs1
, cond_lhs
)
1134 && neutral_element_p (code_def
, cond_rhs
, false))
1135 || (operand_equal_for_phi_arg_p (arg1
, cond_rhs
)
1136 && ((operand_equal_for_phi_arg_p (rhs2
, cond_lhs
)
1137 && absorbing_element_p (code_def
, cond_rhs
, true, rhs2
))
1138 || (operand_equal_for_phi_arg_p (rhs1
, cond_lhs
)
1139 && absorbing_element_p (code_def
,
1140 cond_rhs
, false, rhs2
))))))
1142 gsi
= gsi_for_stmt (cond
);
1143 if (INTEGRAL_TYPE_P (TREE_TYPE (lhs
)))
1145 /* Moving ASSIGN might change VR of lhs, e.g. when moving u_6
1153 # RANGE [0, 4294967294]
1154 u_6 = n_5 + 4294967295;
1157 # u_3 = PHI <u_6(3), 4294967295(2)> */
1158 SSA_NAME_RANGE_INFO (lhs
) = NULL
;
1159 /* If available, we can use VR of phi result at least. */
1160 tree phires
= gimple_phi_result (phi
);
1161 struct range_info_def
*phires_range_info
1162 = SSA_NAME_RANGE_INFO (phires
);
1163 if (phires_range_info
)
1164 duplicate_ssa_name_range_info (lhs
, SSA_NAME_RANGE_TYPE (phires
),
1167 gimple_stmt_iterator gsi_from
;
1168 for (int i
= prep_cnt
- 1; i
>= 0; --i
)
1170 tree plhs
= gimple_assign_lhs (prep_stmt
[i
]);
1171 SSA_NAME_RANGE_INFO (plhs
) = NULL
;
1172 gsi_from
= gsi_for_stmt (prep_stmt
[i
]);
1173 gsi_move_before (&gsi_from
, &gsi
);
1175 gsi_from
= gsi_for_stmt (assign
);
1176 gsi_move_before (&gsi_from
, &gsi
);
1177 replace_phi_edge_with_variable (cond_bb
, e1
, phi
, lhs
);
1184 /* The function minmax_replacement does the main work of doing the minmax
1185 replacement. Return true if the replacement is done. Otherwise return
1187 BB is the basic block where the replacement is going to be done on. ARG0
1188 is argument 0 from the PHI. Likewise for ARG1. */
1191 minmax_replacement (basic_block cond_bb
, basic_block middle_bb
,
1192 edge e0
, edge e1
, gimple
*phi
,
1193 tree arg0
, tree arg1
)
1198 edge true_edge
, false_edge
;
1199 enum tree_code cmp
, minmax
, ass_code
;
1200 tree smaller
, alt_smaller
, larger
, alt_larger
, arg_true
, arg_false
;
1201 gimple_stmt_iterator gsi
, gsi_from
;
1203 type
= TREE_TYPE (PHI_RESULT (phi
));
1205 /* The optimization may be unsafe due to NaNs. */
1206 if (HONOR_NANS (type
) || HONOR_SIGNED_ZEROS (type
))
1209 cond
= as_a
<gcond
*> (last_stmt (cond_bb
));
1210 cmp
= gimple_cond_code (cond
);
1212 /* This transformation is only valid for order comparisons. Record which
1213 operand is smaller/larger if the result of the comparison is true. */
1214 alt_smaller
= NULL_TREE
;
1215 alt_larger
= NULL_TREE
;
1216 if (cmp
== LT_EXPR
|| cmp
== LE_EXPR
)
1218 smaller
= gimple_cond_lhs (cond
);
1219 larger
= gimple_cond_rhs (cond
);
1220 /* If we have smaller < CST it is equivalent to smaller <= CST-1.
1221 Likewise smaller <= CST is equivalent to smaller < CST+1. */
1222 if (TREE_CODE (larger
) == INTEGER_CST
)
1227 wide_int alt
= wi::sub (wi::to_wide (larger
), 1,
1228 TYPE_SIGN (TREE_TYPE (larger
)),
1231 alt_larger
= wide_int_to_tree (TREE_TYPE (larger
), alt
);
1236 wide_int alt
= wi::add (wi::to_wide (larger
), 1,
1237 TYPE_SIGN (TREE_TYPE (larger
)),
1240 alt_larger
= wide_int_to_tree (TREE_TYPE (larger
), alt
);
1244 else if (cmp
== GT_EXPR
|| cmp
== GE_EXPR
)
1246 smaller
= gimple_cond_rhs (cond
);
1247 larger
= gimple_cond_lhs (cond
);
1248 /* If we have larger > CST it is equivalent to larger >= CST+1.
1249 Likewise larger >= CST is equivalent to larger > CST-1. */
1250 if (TREE_CODE (smaller
) == INTEGER_CST
)
1255 wide_int alt
= wi::add (wi::to_wide (smaller
), 1,
1256 TYPE_SIGN (TREE_TYPE (smaller
)),
1259 alt_smaller
= wide_int_to_tree (TREE_TYPE (smaller
), alt
);
1264 wide_int alt
= wi::sub (wi::to_wide (smaller
), 1,
1265 TYPE_SIGN (TREE_TYPE (smaller
)),
1268 alt_smaller
= wide_int_to_tree (TREE_TYPE (smaller
), alt
);
1275 /* We need to know which is the true edge and which is the false
1276 edge so that we know if have abs or negative abs. */
1277 extract_true_false_edges_from_block (cond_bb
, &true_edge
, &false_edge
);
1279 /* Forward the edges over the middle basic block. */
1280 if (true_edge
->dest
== middle_bb
)
1281 true_edge
= EDGE_SUCC (true_edge
->dest
, 0);
1282 if (false_edge
->dest
== middle_bb
)
1283 false_edge
= EDGE_SUCC (false_edge
->dest
, 0);
1285 if (true_edge
== e0
)
1287 gcc_assert (false_edge
== e1
);
1293 gcc_assert (false_edge
== e0
);
1294 gcc_assert (true_edge
== e1
);
1299 if (empty_block_p (middle_bb
))
1301 if ((operand_equal_for_phi_arg_p (arg_true
, smaller
)
1303 && operand_equal_for_phi_arg_p (arg_true
, alt_smaller
)))
1304 && (operand_equal_for_phi_arg_p (arg_false
, larger
)
1306 && operand_equal_for_phi_arg_p (arg_true
, alt_larger
))))
1310 if (smaller < larger)
1316 else if ((operand_equal_for_phi_arg_p (arg_false
, smaller
)
1318 && operand_equal_for_phi_arg_p (arg_false
, alt_smaller
)))
1319 && (operand_equal_for_phi_arg_p (arg_true
, larger
)
1321 && operand_equal_for_phi_arg_p (arg_true
, alt_larger
))))
1328 /* Recognize the following case, assuming d <= u:
1334 This is equivalent to
1339 gimple
*assign
= last_and_only_stmt (middle_bb
);
1340 tree lhs
, op0
, op1
, bound
;
1343 || gimple_code (assign
) != GIMPLE_ASSIGN
)
1346 lhs
= gimple_assign_lhs (assign
);
1347 ass_code
= gimple_assign_rhs_code (assign
);
1348 if (ass_code
!= MAX_EXPR
&& ass_code
!= MIN_EXPR
)
1350 op0
= gimple_assign_rhs1 (assign
);
1351 op1
= gimple_assign_rhs2 (assign
);
1353 if (true_edge
->src
== middle_bb
)
1355 /* We got here if the condition is true, i.e., SMALLER < LARGER. */
1356 if (!operand_equal_for_phi_arg_p (lhs
, arg_true
))
1359 if (operand_equal_for_phi_arg_p (arg_false
, larger
)
1361 && operand_equal_for_phi_arg_p (arg_false
, alt_larger
)))
1365 if (smaller < larger)
1367 r' = MAX_EXPR (smaller, bound)
1369 r = PHI <r', larger> --> to be turned to MIN_EXPR. */
1370 if (ass_code
!= MAX_EXPR
)
1374 if (operand_equal_for_phi_arg_p (op0
, smaller
)
1376 && operand_equal_for_phi_arg_p (op0
, alt_smaller
)))
1378 else if (operand_equal_for_phi_arg_p (op1
, smaller
)
1380 && operand_equal_for_phi_arg_p (op1
, alt_smaller
)))
1385 /* We need BOUND <= LARGER. */
1386 if (!integer_nonzerop (fold_build2 (LE_EXPR
, boolean_type_node
,
1390 else if (operand_equal_for_phi_arg_p (arg_false
, smaller
)
1392 && operand_equal_for_phi_arg_p (arg_false
, alt_smaller
)))
1396 if (smaller < larger)
1398 r' = MIN_EXPR (larger, bound)
1400 r = PHI <r', smaller> --> to be turned to MAX_EXPR. */
1401 if (ass_code
!= MIN_EXPR
)
1405 if (operand_equal_for_phi_arg_p (op0
, larger
)
1407 && operand_equal_for_phi_arg_p (op0
, alt_larger
)))
1409 else if (operand_equal_for_phi_arg_p (op1
, larger
)
1411 && operand_equal_for_phi_arg_p (op1
, alt_larger
)))
1416 /* We need BOUND >= SMALLER. */
1417 if (!integer_nonzerop (fold_build2 (GE_EXPR
, boolean_type_node
,
1426 /* We got here if the condition is false, i.e., SMALLER > LARGER. */
1427 if (!operand_equal_for_phi_arg_p (lhs
, arg_false
))
1430 if (operand_equal_for_phi_arg_p (arg_true
, larger
)
1432 && operand_equal_for_phi_arg_p (arg_true
, alt_larger
)))
1436 if (smaller > larger)
1438 r' = MIN_EXPR (smaller, bound)
1440 r = PHI <r', larger> --> to be turned to MAX_EXPR. */
1441 if (ass_code
!= MIN_EXPR
)
1445 if (operand_equal_for_phi_arg_p (op0
, smaller
)
1447 && operand_equal_for_phi_arg_p (op0
, alt_smaller
)))
1449 else if (operand_equal_for_phi_arg_p (op1
, smaller
)
1451 && operand_equal_for_phi_arg_p (op1
, alt_smaller
)))
1456 /* We need BOUND >= LARGER. */
1457 if (!integer_nonzerop (fold_build2 (GE_EXPR
, boolean_type_node
,
1461 else if (operand_equal_for_phi_arg_p (arg_true
, smaller
)
1463 && operand_equal_for_phi_arg_p (arg_true
, alt_smaller
)))
1467 if (smaller > larger)
1469 r' = MAX_EXPR (larger, bound)
1471 r = PHI <r', smaller> --> to be turned to MIN_EXPR. */
1472 if (ass_code
!= MAX_EXPR
)
1476 if (operand_equal_for_phi_arg_p (op0
, larger
))
1478 else if (operand_equal_for_phi_arg_p (op1
, larger
))
1483 /* We need BOUND <= SMALLER. */
1484 if (!integer_nonzerop (fold_build2 (LE_EXPR
, boolean_type_node
,
1492 /* Move the statement from the middle block. */
1493 gsi
= gsi_last_bb (cond_bb
);
1494 gsi_from
= gsi_last_nondebug_bb (middle_bb
);
1495 gsi_move_before (&gsi_from
, &gsi
);
1498 /* Create an SSA var to hold the min/max result. If we're the only
1499 things setting the target PHI, then we can clone the PHI
1500 variable. Otherwise we must create a new one. */
1501 result
= PHI_RESULT (phi
);
1502 if (EDGE_COUNT (gimple_bb (phi
)->preds
) == 2)
1503 result
= duplicate_ssa_name (result
, NULL
);
1505 result
= make_ssa_name (TREE_TYPE (result
));
1507 /* Emit the statement to compute min/max. */
1508 new_stmt
= gimple_build_assign (result
, minmax
, arg0
, arg1
);
1509 gsi
= gsi_last_bb (cond_bb
);
1510 gsi_insert_before (&gsi
, new_stmt
, GSI_NEW_STMT
);
1512 replace_phi_edge_with_variable (cond_bb
, e1
, phi
, result
);
1513 reset_flow_sensitive_info_in_bb (cond_bb
);
1518 /* The function absolute_replacement does the main work of doing the absolute
1519 replacement. Return true if the replacement is done. Otherwise return
1521 bb is the basic block where the replacement is going to be done on. arg0
1522 is argument 0 from the phi. Likewise for arg1. */
1525 abs_replacement (basic_block cond_bb
, basic_block middle_bb
,
1526 edge e0 ATTRIBUTE_UNUSED
, edge e1
,
1527 gimple
*phi
, tree arg0
, tree arg1
)
1532 gimple_stmt_iterator gsi
;
1533 edge true_edge
, false_edge
;
1538 enum tree_code cond_code
;
1540 /* If the type says honor signed zeros we cannot do this
1542 if (HONOR_SIGNED_ZEROS (arg1
))
1545 /* OTHER_BLOCK must have only one executable statement which must have the
1546 form arg0 = -arg1 or arg1 = -arg0. */
1548 assign
= last_and_only_stmt (middle_bb
);
1549 /* If we did not find the proper negation assignment, then we can not
1554 /* If we got here, then we have found the only executable statement
1555 in OTHER_BLOCK. If it is anything other than arg = -arg1 or
1556 arg1 = -arg0, then we can not optimize. */
1557 if (gimple_code (assign
) != GIMPLE_ASSIGN
)
1560 lhs
= gimple_assign_lhs (assign
);
1562 if (gimple_assign_rhs_code (assign
) != NEGATE_EXPR
)
1565 rhs
= gimple_assign_rhs1 (assign
);
1567 /* The assignment has to be arg0 = -arg1 or arg1 = -arg0. */
1568 if (!(lhs
== arg0
&& rhs
== arg1
)
1569 && !(lhs
== arg1
&& rhs
== arg0
))
1572 cond
= last_stmt (cond_bb
);
1573 result
= PHI_RESULT (phi
);
1575 /* Only relationals comparing arg[01] against zero are interesting. */
1576 cond_code
= gimple_cond_code (cond
);
1577 if (cond_code
!= GT_EXPR
&& cond_code
!= GE_EXPR
1578 && cond_code
!= LT_EXPR
&& cond_code
!= LE_EXPR
)
1581 /* Make sure the conditional is arg[01] OP y. */
1582 if (gimple_cond_lhs (cond
) != rhs
)
1585 if (FLOAT_TYPE_P (TREE_TYPE (gimple_cond_rhs (cond
)))
1586 ? real_zerop (gimple_cond_rhs (cond
))
1587 : integer_zerop (gimple_cond_rhs (cond
)))
1592 /* We need to know which is the true edge and which is the false
1593 edge so that we know if have abs or negative abs. */
1594 extract_true_false_edges_from_block (cond_bb
, &true_edge
, &false_edge
);
1596 /* For GT_EXPR/GE_EXPR, if the true edge goes to OTHER_BLOCK, then we
1597 will need to negate the result. Similarly for LT_EXPR/LE_EXPR if
1598 the false edge goes to OTHER_BLOCK. */
1599 if (cond_code
== GT_EXPR
|| cond_code
== GE_EXPR
)
1604 if (e
->dest
== middle_bb
)
1609 /* If the code negates only iff positive then make sure to not
1610 introduce undefined behavior when negating or computing the absolute.
1611 ??? We could use range info if present to check for arg1 == INT_MIN. */
1613 && (ANY_INTEGRAL_TYPE_P (TREE_TYPE (arg1
))
1614 && ! TYPE_OVERFLOW_WRAPS (TREE_TYPE (arg1
))))
1617 result
= duplicate_ssa_name (result
, NULL
);
1620 lhs
= make_ssa_name (TREE_TYPE (result
));
1624 /* Build the modify expression with abs expression. */
1625 new_stmt
= gimple_build_assign (lhs
, ABS_EXPR
, rhs
);
1627 gsi
= gsi_last_bb (cond_bb
);
1628 gsi_insert_before (&gsi
, new_stmt
, GSI_NEW_STMT
);
1632 /* Get the right GSI. We want to insert after the recently
1633 added ABS_EXPR statement (which we know is the first statement
1635 new_stmt
= gimple_build_assign (result
, NEGATE_EXPR
, lhs
);
1637 gsi_insert_after (&gsi
, new_stmt
, GSI_NEW_STMT
);
1640 replace_phi_edge_with_variable (cond_bb
, e1
, phi
, result
);
1641 reset_flow_sensitive_info_in_bb (cond_bb
);
1643 /* Note that we optimized this PHI. */
1647 /* Auxiliary functions to determine the set of memory accesses which
1648 can't trap because they are preceded by accesses to the same memory
1649 portion. We do that for MEM_REFs, so we only need to track
1650 the SSA_NAME of the pointer indirectly referenced. The algorithm
1651 simply is a walk over all instructions in dominator order. When
1652 we see an MEM_REF we determine if we've already seen a same
1653 ref anywhere up to the root of the dominator tree. If we do the
1654 current access can't trap. If we don't see any dominating access
1655 the current access might trap, but might also make later accesses
1656 non-trapping, so we remember it. We need to be careful with loads
1657 or stores, for instance a load might not trap, while a store would,
1658 so if we see a dominating read access this doesn't mean that a later
1659 write access would not trap. Hence we also need to differentiate the
1660 type of access(es) seen.
1662 ??? We currently are very conservative and assume that a load might
1663 trap even if a store doesn't (write-only memory). This probably is
1664 overly conservative. */
1666 /* A hash-table of SSA_NAMEs, and in which basic block an MEM_REF
1667 through it was seen, which would constitute a no-trap region for
1671 unsigned int ssa_name_ver
;
1674 HOST_WIDE_INT offset
, size
;
1678 /* Hashtable helpers. */
1680 struct ssa_names_hasher
: free_ptr_hash
<name_to_bb
>
1682 static inline hashval_t
hash (const name_to_bb
*);
1683 static inline bool equal (const name_to_bb
*, const name_to_bb
*);
1686 /* Used for quick clearing of the hash-table when we see calls.
1687 Hash entries with phase < nt_call_phase are invalid. */
1688 static unsigned int nt_call_phase
;
1690 /* The hash function. */
1693 ssa_names_hasher::hash (const name_to_bb
*n
)
1695 return n
->ssa_name_ver
^ (((hashval_t
) n
->store
) << 31)
1696 ^ (n
->offset
<< 6) ^ (n
->size
<< 3);
1699 /* The equality function of *P1 and *P2. */
1702 ssa_names_hasher::equal (const name_to_bb
*n1
, const name_to_bb
*n2
)
1704 return n1
->ssa_name_ver
== n2
->ssa_name_ver
1705 && n1
->store
== n2
->store
1706 && n1
->offset
== n2
->offset
1707 && n1
->size
== n2
->size
;
1710 class nontrapping_dom_walker
: public dom_walker
1713 nontrapping_dom_walker (cdi_direction direction
, hash_set
<tree
> *ps
)
1714 : dom_walker (direction
), m_nontrapping (ps
), m_seen_ssa_names (128) {}
1716 virtual edge
before_dom_children (basic_block
);
1717 virtual void after_dom_children (basic_block
);
1721 /* We see the expression EXP in basic block BB. If it's an interesting
1722 expression (an MEM_REF through an SSA_NAME) possibly insert the
1723 expression into the set NONTRAP or the hash table of seen expressions.
1724 STORE is true if this expression is on the LHS, otherwise it's on
1726 void add_or_mark_expr (basic_block
, tree
, bool);
1728 hash_set
<tree
> *m_nontrapping
;
1730 /* The hash table for remembering what we've seen. */
1731 hash_table
<ssa_names_hasher
> m_seen_ssa_names
;
1734 /* Called by walk_dominator_tree, when entering the block BB. */
1736 nontrapping_dom_walker::before_dom_children (basic_block bb
)
1740 gimple_stmt_iterator gsi
;
1742 /* If we haven't seen all our predecessors, clear the hash-table. */
1743 FOR_EACH_EDGE (e
, ei
, bb
->preds
)
1744 if ((((size_t)e
->src
->aux
) & 2) == 0)
1750 /* Mark this BB as being on the path to dominator root and as visited. */
1751 bb
->aux
= (void*)(1 | 2);
1753 /* And walk the statements in order. */
1754 for (gsi
= gsi_start_bb (bb
); !gsi_end_p (gsi
); gsi_next (&gsi
))
1756 gimple
*stmt
= gsi_stmt (gsi
);
1758 if ((gimple_code (stmt
) == GIMPLE_ASM
&& gimple_vdef (stmt
))
1759 || (is_gimple_call (stmt
)
1760 && (!nonfreeing_call_p (stmt
) || !nonbarrier_call_p (stmt
))))
1762 else if (gimple_assign_single_p (stmt
) && !gimple_has_volatile_ops (stmt
))
1764 add_or_mark_expr (bb
, gimple_assign_lhs (stmt
), true);
1765 add_or_mark_expr (bb
, gimple_assign_rhs1 (stmt
), false);
1771 /* Called by walk_dominator_tree, when basic block BB is exited. */
1773 nontrapping_dom_walker::after_dom_children (basic_block bb
)
1775 /* This BB isn't on the path to dominator root anymore. */
1779 /* We see the expression EXP in basic block BB. If it's an interesting
1780 expression (an MEM_REF through an SSA_NAME) possibly insert the
1781 expression into the set NONTRAP or the hash table of seen expressions.
1782 STORE is true if this expression is on the LHS, otherwise it's on
1785 nontrapping_dom_walker::add_or_mark_expr (basic_block bb
, tree exp
, bool store
)
1789 if (TREE_CODE (exp
) == MEM_REF
1790 && TREE_CODE (TREE_OPERAND (exp
, 0)) == SSA_NAME
1791 && tree_fits_shwi_p (TREE_OPERAND (exp
, 1))
1792 && (size
= int_size_in_bytes (TREE_TYPE (exp
))) > 0)
1794 tree name
= TREE_OPERAND (exp
, 0);
1795 struct name_to_bb map
;
1797 struct name_to_bb
*n2bb
;
1798 basic_block found_bb
= 0;
1800 /* Try to find the last seen MEM_REF through the same
1801 SSA_NAME, which can trap. */
1802 map
.ssa_name_ver
= SSA_NAME_VERSION (name
);
1806 map
.offset
= tree_to_shwi (TREE_OPERAND (exp
, 1));
1809 slot
= m_seen_ssa_names
.find_slot (&map
, INSERT
);
1811 if (n2bb
&& n2bb
->phase
>= nt_call_phase
)
1812 found_bb
= n2bb
->bb
;
1814 /* If we've found a trapping MEM_REF, _and_ it dominates EXP
1815 (it's in a basic block on the path from us to the dominator root)
1816 then we can't trap. */
1817 if (found_bb
&& (((size_t)found_bb
->aux
) & 1) == 1)
1819 m_nontrapping
->add (exp
);
1823 /* EXP might trap, so insert it into the hash table. */
1826 n2bb
->phase
= nt_call_phase
;
1831 n2bb
= XNEW (struct name_to_bb
);
1832 n2bb
->ssa_name_ver
= SSA_NAME_VERSION (name
);
1833 n2bb
->phase
= nt_call_phase
;
1835 n2bb
->store
= store
;
1836 n2bb
->offset
= map
.offset
;
1844 /* This is the entry point of gathering non trapping memory accesses.
1845 It will do a dominator walk over the whole function, and it will
1846 make use of the bb->aux pointers. It returns a set of trees
1847 (the MEM_REFs itself) which can't trap. */
1848 static hash_set
<tree
> *
1849 get_non_trapping (void)
1852 hash_set
<tree
> *nontrap
= new hash_set
<tree
>;
1853 /* We're going to do a dominator walk, so ensure that we have
1854 dominance information. */
1855 calculate_dominance_info (CDI_DOMINATORS
);
1857 nontrapping_dom_walker (CDI_DOMINATORS
, nontrap
)
1858 .walk (cfun
->cfg
->x_entry_block_ptr
);
1860 clear_aux_for_blocks ();
1864 /* Do the main work of conditional store replacement. We already know
1865 that the recognized pattern looks like so:
1868 if (cond) goto MIDDLE_BB; else goto JOIN_BB (edge E1)
1871 fallthrough (edge E0)
1875 We check that MIDDLE_BB contains only one store, that that store
1876 doesn't trap (not via NOTRAP, but via checking if an access to the same
1877 memory location dominates us) and that the store has a "simple" RHS. */
1880 cond_store_replacement (basic_block middle_bb
, basic_block join_bb
,
1881 edge e0
, edge e1
, hash_set
<tree
> *nontrap
)
1883 gimple
*assign
= last_and_only_stmt (middle_bb
);
1884 tree lhs
, rhs
, name
, name2
;
1887 gimple_stmt_iterator gsi
;
1888 source_location locus
;
1890 /* Check if middle_bb contains of only one store. */
1892 || !gimple_assign_single_p (assign
)
1893 || gimple_has_volatile_ops (assign
))
1896 locus
= gimple_location (assign
);
1897 lhs
= gimple_assign_lhs (assign
);
1898 rhs
= gimple_assign_rhs1 (assign
);
1899 if (TREE_CODE (lhs
) != MEM_REF
1900 || TREE_CODE (TREE_OPERAND (lhs
, 0)) != SSA_NAME
1901 || !is_gimple_reg_type (TREE_TYPE (lhs
)))
1904 /* Prove that we can move the store down. We could also check
1905 TREE_THIS_NOTRAP here, but in that case we also could move stores,
1906 whose value is not available readily, which we want to avoid. */
1907 if (!nontrap
->contains (lhs
))
1910 /* Now we've checked the constraints, so do the transformation:
1911 1) Remove the single store. */
1912 gsi
= gsi_for_stmt (assign
);
1913 unlink_stmt_vdef (assign
);
1914 gsi_remove (&gsi
, true);
1915 release_defs (assign
);
1917 /* 2) Insert a load from the memory of the store to the temporary
1918 on the edge which did not contain the store. */
1919 lhs
= unshare_expr (lhs
);
1920 name
= make_temp_ssa_name (TREE_TYPE (lhs
), NULL
, "cstore");
1921 new_stmt
= gimple_build_assign (name
, lhs
);
1922 gimple_set_location (new_stmt
, locus
);
1923 gsi_insert_on_edge (e1
, new_stmt
);
1925 /* 3) Create a PHI node at the join block, with one argument
1926 holding the old RHS, and the other holding the temporary
1927 where we stored the old memory contents. */
1928 name2
= make_temp_ssa_name (TREE_TYPE (lhs
), NULL
, "cstore");
1929 newphi
= create_phi_node (name2
, join_bb
);
1930 add_phi_arg (newphi
, rhs
, e0
, locus
);
1931 add_phi_arg (newphi
, name
, e1
, locus
);
1933 lhs
= unshare_expr (lhs
);
1934 new_stmt
= gimple_build_assign (lhs
, PHI_RESULT (newphi
));
1936 /* 4) Insert that PHI node. */
1937 gsi
= gsi_after_labels (join_bb
);
1938 if (gsi_end_p (gsi
))
1940 gsi
= gsi_last_bb (join_bb
);
1941 gsi_insert_after (&gsi
, new_stmt
, GSI_NEW_STMT
);
1944 gsi_insert_before (&gsi
, new_stmt
, GSI_NEW_STMT
);
1949 /* Do the main work of conditional store replacement. */
1952 cond_if_else_store_replacement_1 (basic_block then_bb
, basic_block else_bb
,
1953 basic_block join_bb
, gimple
*then_assign
,
1954 gimple
*else_assign
)
1956 tree lhs_base
, lhs
, then_rhs
, else_rhs
, name
;
1957 source_location then_locus
, else_locus
;
1958 gimple_stmt_iterator gsi
;
1962 if (then_assign
== NULL
1963 || !gimple_assign_single_p (then_assign
)
1964 || gimple_clobber_p (then_assign
)
1965 || gimple_has_volatile_ops (then_assign
)
1966 || else_assign
== NULL
1967 || !gimple_assign_single_p (else_assign
)
1968 || gimple_clobber_p (else_assign
)
1969 || gimple_has_volatile_ops (else_assign
))
1972 lhs
= gimple_assign_lhs (then_assign
);
1973 if (!is_gimple_reg_type (TREE_TYPE (lhs
))
1974 || !operand_equal_p (lhs
, gimple_assign_lhs (else_assign
), 0))
1977 lhs_base
= get_base_address (lhs
);
1978 if (lhs_base
== NULL_TREE
1979 || (!DECL_P (lhs_base
) && TREE_CODE (lhs_base
) != MEM_REF
))
1982 then_rhs
= gimple_assign_rhs1 (then_assign
);
1983 else_rhs
= gimple_assign_rhs1 (else_assign
);
1984 then_locus
= gimple_location (then_assign
);
1985 else_locus
= gimple_location (else_assign
);
1987 /* Now we've checked the constraints, so do the transformation:
1988 1) Remove the stores. */
1989 gsi
= gsi_for_stmt (then_assign
);
1990 unlink_stmt_vdef (then_assign
);
1991 gsi_remove (&gsi
, true);
1992 release_defs (then_assign
);
1994 gsi
= gsi_for_stmt (else_assign
);
1995 unlink_stmt_vdef (else_assign
);
1996 gsi_remove (&gsi
, true);
1997 release_defs (else_assign
);
1999 /* 2) Create a PHI node at the join block, with one argument
2000 holding the old RHS, and the other holding the temporary
2001 where we stored the old memory contents. */
2002 name
= make_temp_ssa_name (TREE_TYPE (lhs
), NULL
, "cstore");
2003 newphi
= create_phi_node (name
, join_bb
);
2004 add_phi_arg (newphi
, then_rhs
, EDGE_SUCC (then_bb
, 0), then_locus
);
2005 add_phi_arg (newphi
, else_rhs
, EDGE_SUCC (else_bb
, 0), else_locus
);
2007 new_stmt
= gimple_build_assign (lhs
, PHI_RESULT (newphi
));
2009 /* 3) Insert that PHI node. */
2010 gsi
= gsi_after_labels (join_bb
);
2011 if (gsi_end_p (gsi
))
2013 gsi
= gsi_last_bb (join_bb
);
2014 gsi_insert_after (&gsi
, new_stmt
, GSI_NEW_STMT
);
2017 gsi_insert_before (&gsi
, new_stmt
, GSI_NEW_STMT
);
2022 /* Conditional store replacement. We already know
2023 that the recognized pattern looks like so:
2026 if (cond) goto THEN_BB; else goto ELSE_BB (edge E1)
2036 fallthrough (edge E0)
2040 We check that it is safe to sink the store to JOIN_BB by verifying that
2041 there are no read-after-write or write-after-write dependencies in
2042 THEN_BB and ELSE_BB. */
2045 cond_if_else_store_replacement (basic_block then_bb
, basic_block else_bb
,
2046 basic_block join_bb
)
2048 gimple
*then_assign
= last_and_only_stmt (then_bb
);
2049 gimple
*else_assign
= last_and_only_stmt (else_bb
);
2050 vec
<data_reference_p
> then_datarefs
, else_datarefs
;
2051 vec
<ddr_p
> then_ddrs
, else_ddrs
;
2052 gimple
*then_store
, *else_store
;
2053 bool found
, ok
= false, res
;
2054 struct data_dependence_relation
*ddr
;
2055 data_reference_p then_dr
, else_dr
;
2057 tree then_lhs
, else_lhs
;
2058 basic_block blocks
[3];
2060 if (MAX_STORES_TO_SINK
== 0)
2063 /* Handle the case with single statement in THEN_BB and ELSE_BB. */
2064 if (then_assign
&& else_assign
)
2065 return cond_if_else_store_replacement_1 (then_bb
, else_bb
, join_bb
,
2066 then_assign
, else_assign
);
2068 /* Find data references. */
2069 then_datarefs
.create (1);
2070 else_datarefs
.create (1);
2071 if ((find_data_references_in_bb (NULL
, then_bb
, &then_datarefs
)
2073 || !then_datarefs
.length ()
2074 || (find_data_references_in_bb (NULL
, else_bb
, &else_datarefs
)
2076 || !else_datarefs
.length ())
2078 free_data_refs (then_datarefs
);
2079 free_data_refs (else_datarefs
);
2083 /* Find pairs of stores with equal LHS. */
2084 auto_vec
<gimple
*, 1> then_stores
, else_stores
;
2085 FOR_EACH_VEC_ELT (then_datarefs
, i
, then_dr
)
2087 if (DR_IS_READ (then_dr
))
2090 then_store
= DR_STMT (then_dr
);
2091 then_lhs
= gimple_get_lhs (then_store
);
2092 if (then_lhs
== NULL_TREE
)
2096 FOR_EACH_VEC_ELT (else_datarefs
, j
, else_dr
)
2098 if (DR_IS_READ (else_dr
))
2101 else_store
= DR_STMT (else_dr
);
2102 else_lhs
= gimple_get_lhs (else_store
);
2103 if (else_lhs
== NULL_TREE
)
2106 if (operand_equal_p (then_lhs
, else_lhs
, 0))
2116 then_stores
.safe_push (then_store
);
2117 else_stores
.safe_push (else_store
);
2120 /* No pairs of stores found. */
2121 if (!then_stores
.length ()
2122 || then_stores
.length () > (unsigned) MAX_STORES_TO_SINK
)
2124 free_data_refs (then_datarefs
);
2125 free_data_refs (else_datarefs
);
2129 /* Compute and check data dependencies in both basic blocks. */
2130 then_ddrs
.create (1);
2131 else_ddrs
.create (1);
2132 if (!compute_all_dependences (then_datarefs
, &then_ddrs
,
2134 || !compute_all_dependences (else_datarefs
, &else_ddrs
,
2137 free_dependence_relations (then_ddrs
);
2138 free_dependence_relations (else_ddrs
);
2139 free_data_refs (then_datarefs
);
2140 free_data_refs (else_datarefs
);
2143 blocks
[0] = then_bb
;
2144 blocks
[1] = else_bb
;
2145 blocks
[2] = join_bb
;
2146 renumber_gimple_stmt_uids_in_blocks (blocks
, 3);
2148 /* Check that there are no read-after-write or write-after-write dependencies
2150 FOR_EACH_VEC_ELT (then_ddrs
, i
, ddr
)
2152 struct data_reference
*dra
= DDR_A (ddr
);
2153 struct data_reference
*drb
= DDR_B (ddr
);
2155 if (DDR_ARE_DEPENDENT (ddr
) != chrec_known
2156 && ((DR_IS_READ (dra
) && DR_IS_WRITE (drb
)
2157 && gimple_uid (DR_STMT (dra
)) > gimple_uid (DR_STMT (drb
)))
2158 || (DR_IS_READ (drb
) && DR_IS_WRITE (dra
)
2159 && gimple_uid (DR_STMT (drb
)) > gimple_uid (DR_STMT (dra
)))
2160 || (DR_IS_WRITE (dra
) && DR_IS_WRITE (drb
))))
2162 free_dependence_relations (then_ddrs
);
2163 free_dependence_relations (else_ddrs
);
2164 free_data_refs (then_datarefs
);
2165 free_data_refs (else_datarefs
);
2170 /* Check that there are no read-after-write or write-after-write dependencies
2172 FOR_EACH_VEC_ELT (else_ddrs
, i
, ddr
)
2174 struct data_reference
*dra
= DDR_A (ddr
);
2175 struct data_reference
*drb
= DDR_B (ddr
);
2177 if (DDR_ARE_DEPENDENT (ddr
) != chrec_known
2178 && ((DR_IS_READ (dra
) && DR_IS_WRITE (drb
)
2179 && gimple_uid (DR_STMT (dra
)) > gimple_uid (DR_STMT (drb
)))
2180 || (DR_IS_READ (drb
) && DR_IS_WRITE (dra
)
2181 && gimple_uid (DR_STMT (drb
)) > gimple_uid (DR_STMT (dra
)))
2182 || (DR_IS_WRITE (dra
) && DR_IS_WRITE (drb
))))
2184 free_dependence_relations (then_ddrs
);
2185 free_dependence_relations (else_ddrs
);
2186 free_data_refs (then_datarefs
);
2187 free_data_refs (else_datarefs
);
2192 /* Sink stores with same LHS. */
2193 FOR_EACH_VEC_ELT (then_stores
, i
, then_store
)
2195 else_store
= else_stores
[i
];
2196 res
= cond_if_else_store_replacement_1 (then_bb
, else_bb
, join_bb
,
2197 then_store
, else_store
);
2201 free_dependence_relations (then_ddrs
);
2202 free_dependence_relations (else_ddrs
);
2203 free_data_refs (then_datarefs
);
2204 free_data_refs (else_datarefs
);
2209 /* Return TRUE if STMT has a VUSE whose corresponding VDEF is in BB. */
2212 local_mem_dependence (gimple
*stmt
, basic_block bb
)
2214 tree vuse
= gimple_vuse (stmt
);
2220 def
= SSA_NAME_DEF_STMT (vuse
);
2221 return (def
&& gimple_bb (def
) == bb
);
2224 /* Given a "diamond" control-flow pattern where BB0 tests a condition,
2225 BB1 and BB2 are "then" and "else" blocks dependent on this test,
2226 and BB3 rejoins control flow following BB1 and BB2, look for
2227 opportunities to hoist loads as follows. If BB3 contains a PHI of
2228 two loads, one each occurring in BB1 and BB2, and the loads are
2229 provably of adjacent fields in the same structure, then move both
2230 loads into BB0. Of course this can only be done if there are no
2231 dependencies preventing such motion.
2233 One of the hoisted loads will always be speculative, so the
2234 transformation is currently conservative:
2236 - The fields must be strictly adjacent.
2237 - The two fields must occupy a single memory block that is
2238 guaranteed to not cross a page boundary.
2240 The last is difficult to prove, as such memory blocks should be
2241 aligned on the minimum of the stack alignment boundary and the
2242 alignment guaranteed by heap allocation interfaces. Thus we rely
2243 on a parameter for the alignment value.
2245 Provided a good value is used for the last case, the first
2246 restriction could possibly be relaxed. */
2249 hoist_adjacent_loads (basic_block bb0
, basic_block bb1
,
2250 basic_block bb2
, basic_block bb3
)
2252 int param_align
= PARAM_VALUE (PARAM_L1_CACHE_LINE_SIZE
);
2253 unsigned param_align_bits
= (unsigned) (param_align
* BITS_PER_UNIT
);
2256 /* Walk the phis in bb3 looking for an opportunity. We are looking
2257 for phis of two SSA names, one each of which is defined in bb1 and
2259 for (gsi
= gsi_start_phis (bb3
); !gsi_end_p (gsi
); gsi_next (&gsi
))
2261 gphi
*phi_stmt
= gsi
.phi ();
2262 gimple
*def1
, *def2
;
2263 tree arg1
, arg2
, ref1
, ref2
, field1
, field2
;
2264 tree tree_offset1
, tree_offset2
, tree_size2
, next
;
2265 int offset1
, offset2
, size2
;
2267 gimple_stmt_iterator gsi2
;
2268 basic_block bb_for_def1
, bb_for_def2
;
2270 if (gimple_phi_num_args (phi_stmt
) != 2
2271 || virtual_operand_p (gimple_phi_result (phi_stmt
)))
2274 arg1
= gimple_phi_arg_def (phi_stmt
, 0);
2275 arg2
= gimple_phi_arg_def (phi_stmt
, 1);
2277 if (TREE_CODE (arg1
) != SSA_NAME
2278 || TREE_CODE (arg2
) != SSA_NAME
2279 || SSA_NAME_IS_DEFAULT_DEF (arg1
)
2280 || SSA_NAME_IS_DEFAULT_DEF (arg2
))
2283 def1
= SSA_NAME_DEF_STMT (arg1
);
2284 def2
= SSA_NAME_DEF_STMT (arg2
);
2286 if ((gimple_bb (def1
) != bb1
|| gimple_bb (def2
) != bb2
)
2287 && (gimple_bb (def2
) != bb1
|| gimple_bb (def1
) != bb2
))
2290 /* Check the mode of the arguments to be sure a conditional move
2291 can be generated for it. */
2292 if (optab_handler (movcc_optab
, TYPE_MODE (TREE_TYPE (arg1
)))
2293 == CODE_FOR_nothing
)
2296 /* Both statements must be assignments whose RHS is a COMPONENT_REF. */
2297 if (!gimple_assign_single_p (def1
)
2298 || !gimple_assign_single_p (def2
)
2299 || gimple_has_volatile_ops (def1
)
2300 || gimple_has_volatile_ops (def2
))
2303 ref1
= gimple_assign_rhs1 (def1
);
2304 ref2
= gimple_assign_rhs1 (def2
);
2306 if (TREE_CODE (ref1
) != COMPONENT_REF
2307 || TREE_CODE (ref2
) != COMPONENT_REF
)
2310 /* The zeroth operand of the two component references must be
2311 identical. It is not sufficient to compare get_base_address of
2312 the two references, because this could allow for different
2313 elements of the same array in the two trees. It is not safe to
2314 assume that the existence of one array element implies the
2315 existence of a different one. */
2316 if (!operand_equal_p (TREE_OPERAND (ref1
, 0), TREE_OPERAND (ref2
, 0), 0))
2319 field1
= TREE_OPERAND (ref1
, 1);
2320 field2
= TREE_OPERAND (ref2
, 1);
2322 /* Check for field adjacency, and ensure field1 comes first. */
2323 for (next
= DECL_CHAIN (field1
);
2324 next
&& TREE_CODE (next
) != FIELD_DECL
;
2325 next
= DECL_CHAIN (next
))
2330 for (next
= DECL_CHAIN (field2
);
2331 next
&& TREE_CODE (next
) != FIELD_DECL
;
2332 next
= DECL_CHAIN (next
))
2338 std::swap (field1
, field2
);
2339 std::swap (def1
, def2
);
2342 bb_for_def1
= gimple_bb (def1
);
2343 bb_for_def2
= gimple_bb (def2
);
2345 /* Check for proper alignment of the first field. */
2346 tree_offset1
= bit_position (field1
);
2347 tree_offset2
= bit_position (field2
);
2348 tree_size2
= DECL_SIZE (field2
);
2350 if (!tree_fits_uhwi_p (tree_offset1
)
2351 || !tree_fits_uhwi_p (tree_offset2
)
2352 || !tree_fits_uhwi_p (tree_size2
))
2355 offset1
= tree_to_uhwi (tree_offset1
);
2356 offset2
= tree_to_uhwi (tree_offset2
);
2357 size2
= tree_to_uhwi (tree_size2
);
2358 align1
= DECL_ALIGN (field1
) % param_align_bits
;
2360 if (offset1
% BITS_PER_UNIT
!= 0)
2363 /* For profitability, the two field references should fit within
2364 a single cache line. */
2365 if (align1
+ offset2
- offset1
+ size2
> param_align_bits
)
2368 /* The two expressions cannot be dependent upon vdefs defined
2370 if (local_mem_dependence (def1
, bb_for_def1
)
2371 || local_mem_dependence (def2
, bb_for_def2
))
2374 /* The conditions are satisfied; hoist the loads from bb1 and bb2 into
2375 bb0. We hoist the first one first so that a cache miss is handled
2376 efficiently regardless of hardware cache-fill policy. */
2377 gsi2
= gsi_for_stmt (def1
);
2378 gsi_move_to_bb_end (&gsi2
, bb0
);
2379 gsi2
= gsi_for_stmt (def2
);
2380 gsi_move_to_bb_end (&gsi2
, bb0
);
2382 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
2385 "\nHoisting adjacent loads from %d and %d into %d: \n",
2386 bb_for_def1
->index
, bb_for_def2
->index
, bb0
->index
);
2387 print_gimple_stmt (dump_file
, def1
, 0, TDF_VOPS
|TDF_MEMSYMS
);
2388 print_gimple_stmt (dump_file
, def2
, 0, TDF_VOPS
|TDF_MEMSYMS
);
2393 /* Determine whether we should attempt to hoist adjacent loads out of
2394 diamond patterns in pass_phiopt. Always hoist loads if
2395 -fhoist-adjacent-loads is specified and the target machine has
2396 both a conditional move instruction and a defined cache line size. */
2399 gate_hoist_loads (void)
2401 return (flag_hoist_adjacent_loads
== 1
2402 && PARAM_VALUE (PARAM_L1_CACHE_LINE_SIZE
)
2403 && HAVE_conditional_move
);
2406 /* This pass tries to replaces an if-then-else block with an
2407 assignment. We have four kinds of transformations. Some of these
2408 transformations are also performed by the ifcvt RTL optimizer.
2410 Conditional Replacement
2411 -----------------------
2413 This transformation, implemented in conditional_replacement,
2417 if (cond) goto bb2; else goto bb1;
2420 x = PHI <0 (bb1), 1 (bb0), ...>;
2428 x = PHI <x' (bb0), ...>;
2430 We remove bb1 as it becomes unreachable. This occurs often due to
2431 gimplification of conditionals.
2436 This transformation, implemented in value_replacement, replaces
2439 if (a != b) goto bb2; else goto bb1;
2442 x = PHI <a (bb1), b (bb0), ...>;
2448 x = PHI <b (bb0), ...>;
2450 This opportunity can sometimes occur as a result of other
2454 Another case caught by value replacement looks like this:
2460 if (t3 != 0) goto bb1; else goto bb2;
2476 This transformation, implemented in abs_replacement, replaces
2479 if (a >= 0) goto bb2; else goto bb1;
2483 x = PHI <x (bb1), a (bb0), ...>;
2490 x = PHI <x' (bb0), ...>;
2495 This transformation, minmax_replacement replaces
2498 if (a <= b) goto bb2; else goto bb1;
2501 x = PHI <b (bb1), a (bb0), ...>;
2506 x' = MIN_EXPR (a, b)
2508 x = PHI <x' (bb0), ...>;
2510 A similar transformation is done for MAX_EXPR.
2513 This pass also performs a fifth transformation of a slightly different
2516 Factor conversion in COND_EXPR
2517 ------------------------------
2519 This transformation factors the conversion out of COND_EXPR with
2520 factor_out_conditional_conversion.
2523 if (a <= CST) goto <bb 3>; else goto <bb 4>;
2527 tmp = PHI <tmp, CST>
2530 if (a <= CST) goto <bb 3>; else goto <bb 4>;
2536 Adjacent Load Hoisting
2537 ----------------------
2539 This transformation replaces
2542 if (...) goto bb2; else goto bb1;
2544 x1 = (<expr>).field1;
2547 x2 = (<expr>).field2;
2554 x1 = (<expr>).field1;
2555 x2 = (<expr>).field2;
2556 if (...) goto bb2; else goto bb1;
2563 The purpose of this transformation is to enable generation of conditional
2564 move instructions such as Intel CMOVE or PowerPC ISEL. Because one of
2565 the loads is speculative, the transformation is restricted to very
2566 specific cases to avoid introducing a page fault. We are looking for
2574 where left and right are typically adjacent pointers in a tree structure. */
2578 const pass_data pass_data_phiopt
=
2580 GIMPLE_PASS
, /* type */
2581 "phiopt", /* name */
2582 OPTGROUP_NONE
, /* optinfo_flags */
2583 TV_TREE_PHIOPT
, /* tv_id */
2584 ( PROP_cfg
| PROP_ssa
), /* properties_required */
2585 0, /* properties_provided */
2586 0, /* properties_destroyed */
2587 0, /* todo_flags_start */
2588 0, /* todo_flags_finish */
2591 class pass_phiopt
: public gimple_opt_pass
2594 pass_phiopt (gcc::context
*ctxt
)
2595 : gimple_opt_pass (pass_data_phiopt
, ctxt
)
2598 /* opt_pass methods: */
2599 opt_pass
* clone () { return new pass_phiopt (m_ctxt
); }
2600 virtual bool gate (function
*) { return flag_ssa_phiopt
; }
2601 virtual unsigned int execute (function
*)
2603 return tree_ssa_phiopt_worker (false, gate_hoist_loads ());
2606 }; // class pass_phiopt
2611 make_pass_phiopt (gcc::context
*ctxt
)
2613 return new pass_phiopt (ctxt
);
2618 const pass_data pass_data_cselim
=
2620 GIMPLE_PASS
, /* type */
2621 "cselim", /* name */
2622 OPTGROUP_NONE
, /* optinfo_flags */
2623 TV_TREE_PHIOPT
, /* tv_id */
2624 ( PROP_cfg
| PROP_ssa
), /* properties_required */
2625 0, /* properties_provided */
2626 0, /* properties_destroyed */
2627 0, /* todo_flags_start */
2628 0, /* todo_flags_finish */
2631 class pass_cselim
: public gimple_opt_pass
2634 pass_cselim (gcc::context
*ctxt
)
2635 : gimple_opt_pass (pass_data_cselim
, ctxt
)
2638 /* opt_pass methods: */
2639 virtual bool gate (function
*) { return flag_tree_cselim
; }
2640 virtual unsigned int execute (function
*) { return tree_ssa_cs_elim (); }
2642 }; // class pass_cselim
2647 make_pass_cselim (gcc::context
*ctxt
)
2649 return new pass_cselim (ctxt
);