1 /* Routines for discovering and unpropagating edge equivalences.
2 Copyright (C) 2005, 2007, 2008 Free Software Foundation, Inc.
4 This file is part of GCC.
6 GCC is free software; you can redistribute it and/or modify
7 it under the terms of the GNU General Public License as published by
8 the Free Software Foundation; either version 3, or (at your option)
11 GCC is distributed in the hope that it will be useful,
12 but WITHOUT ANY WARRANTY; without even the implied warranty of
13 MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
14 GNU General Public License for more details.
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"
29 #include "basic-block.h"
33 #include "diagnostic.h"
35 #include "tree-dump.h"
36 #include "tree-flow.h"
39 #include "tree-pass.h"
40 #include "tree-ssa-propagate.h"
41 #include "langhooks.h"
43 /* The basic structure describing an equivalency created by traversing
44 an edge. Traversing the edge effectively means that we can assume
45 that we've seen an assignment LHS = RHS. */
46 struct edge_equivalency
52 /* This routine finds and records edge equivalences for every edge
55 When complete, each edge that creates an equivalency will have an
56 EDGE_EQUIVALENCY structure hanging off the edge's AUX field.
57 The caller is responsible for freeing the AUX fields. */
60 associate_equivalences_with_edges (void)
64 /* Walk over each block. If the block ends with a control statement,
65 then it might create a useful equivalence. */
68 gimple_stmt_iterator gsi
= gsi_last_bb (bb
);
71 /* If the block does not end with a COND_EXPR or SWITCH_EXPR
72 then there is nothing to do. */
76 stmt
= gsi_stmt (gsi
);
81 /* A COND_EXPR may create an equivalency in a variety of different
83 if (gimple_code (stmt
) == GIMPLE_COND
)
87 struct edge_equivalency
*equivalency
;
88 enum tree_code code
= gimple_cond_code (stmt
);
90 extract_true_false_edges_from_block (bb
, &true_edge
, &false_edge
);
92 /* Equality tests may create one or two equivalences. */
93 if (code
== EQ_EXPR
|| code
== NE_EXPR
)
95 tree op0
= gimple_cond_lhs (stmt
);
96 tree op1
= gimple_cond_rhs (stmt
);
98 /* Special case comparing booleans against a constant as we
99 know the value of OP0 on both arms of the branch. i.e., we
100 can record an equivalence for OP0 rather than COND. */
101 if (TREE_CODE (op0
) == SSA_NAME
102 && !SSA_NAME_OCCURS_IN_ABNORMAL_PHI (op0
)
103 && TREE_CODE (TREE_TYPE (op0
)) == BOOLEAN_TYPE
104 && is_gimple_min_invariant (op1
))
108 equivalency
= XNEW (struct edge_equivalency
);
109 equivalency
->lhs
= op0
;
110 equivalency
->rhs
= (integer_zerop (op1
)
112 : boolean_true_node
);
113 true_edge
->aux
= equivalency
;
115 equivalency
= XNEW (struct edge_equivalency
);
116 equivalency
->lhs
= op0
;
117 equivalency
->rhs
= (integer_zerop (op1
)
119 : boolean_false_node
);
120 false_edge
->aux
= equivalency
;
124 equivalency
= XNEW (struct edge_equivalency
);
125 equivalency
->lhs
= op0
;
126 equivalency
->rhs
= (integer_zerop (op1
)
128 : boolean_false_node
);
129 true_edge
->aux
= equivalency
;
131 equivalency
= XNEW (struct edge_equivalency
);
132 equivalency
->lhs
= op0
;
133 equivalency
->rhs
= (integer_zerop (op1
)
135 : boolean_true_node
);
136 false_edge
->aux
= equivalency
;
140 else if (TREE_CODE (op0
) == SSA_NAME
141 && !SSA_NAME_OCCURS_IN_ABNORMAL_PHI (op0
)
142 && (is_gimple_min_invariant (op1
)
143 || (TREE_CODE (op1
) == SSA_NAME
144 && !SSA_NAME_OCCURS_IN_ABNORMAL_PHI (op1
))))
146 /* For IEEE, -0.0 == 0.0, so we don't necessarily know
147 the sign of a variable compared against zero. If
148 we're honoring signed zeros, then we cannot record
149 this value unless we know that the value is nonzero. */
150 if (HONOR_SIGNED_ZEROS (TYPE_MODE (TREE_TYPE (op0
)))
151 && (TREE_CODE (op1
) != REAL_CST
152 || REAL_VALUES_EQUAL (dconst0
, TREE_REAL_CST (op1
))))
155 equivalency
= XNEW (struct edge_equivalency
);
156 equivalency
->lhs
= op0
;
157 equivalency
->rhs
= op1
;
159 true_edge
->aux
= equivalency
;
161 false_edge
->aux
= equivalency
;
166 /* ??? TRUTH_NOT_EXPR can create an equivalence too. */
169 /* For a SWITCH_EXPR, a case label which represents a single
170 value and which is the only case label which reaches the
171 target block creates an equivalence. */
172 else if (gimple_code (stmt
) == GIMPLE_SWITCH
)
174 tree cond
= gimple_switch_index (stmt
);
176 if (TREE_CODE (cond
) == SSA_NAME
177 && !SSA_NAME_OCCURS_IN_ABNORMAL_PHI (cond
))
179 int i
, n_labels
= gimple_switch_num_labels (stmt
);
180 tree
*info
= XCNEWVEC (tree
, n_basic_blocks
);
182 /* Walk over the case label vector. Record blocks
183 which are reached by a single case label which represents
185 for (i
= 0; i
< n_labels
; i
++)
187 tree label
= gimple_switch_label (stmt
, i
);
188 basic_block bb
= label_to_block (CASE_LABEL (label
));
190 if (CASE_HIGH (label
)
193 info
[bb
->index
] = error_mark_node
;
195 info
[bb
->index
] = label
;
198 /* Now walk over the blocks to determine which ones were
199 marked as being reached by a useful case label. */
200 for (i
= 0; i
< n_basic_blocks
; i
++)
205 && node
!= error_mark_node
)
207 tree x
= fold_convert (TREE_TYPE (cond
), CASE_LOW (node
));
208 struct edge_equivalency
*equivalency
;
210 /* Record an equivalency on the edge from BB to basic
212 equivalency
= XNEW (struct edge_equivalency
);
213 equivalency
->rhs
= x
;
214 equivalency
->lhs
= cond
;
215 find_edge (bb
, BASIC_BLOCK (i
))->aux
= equivalency
;
226 /* Translating out of SSA sometimes requires inserting copies and
227 constant initializations on edges to eliminate PHI nodes.
229 In some cases those copies and constant initializations are
230 redundant because the target already has the value on the
231 RHS of the assignment.
233 We previously tried to catch these cases after translating
234 out of SSA form. However, that code often missed cases. Worse
235 yet, the cases it missed were also often missed by the RTL
236 optimizers. Thus the resulting code had redundant instructions.
238 This pass attempts to detect these situations before translating
241 The key concept that this pass is built upon is that these
242 redundant copies and constant initializations often occur
243 due to constant/copy propagating equivalences resulting from
244 COND_EXPRs and SWITCH_EXPRs.
246 We want to do those propagations as they can sometimes allow
247 the SSA optimizers to do a better job. However, in the cases
248 where such propagations do not result in further optimization,
249 we would like to "undo" the propagation to avoid the redundant
250 copies and constant initializations.
252 This pass works by first associating equivalences with edges in
253 the CFG. For example, the edge leading from a SWITCH_EXPR to
254 its associated CASE_LABEL will have an equivalency between
255 SWITCH_COND and the value in the case label.
257 Once we have found the edge equivalences, we proceed to walk
258 the CFG in dominator order. As we traverse edges we record
259 equivalences associated with those edges we traverse.
261 When we encounter a PHI node, we walk its arguments to see if we
262 have an equivalence for the PHI argument. If so, then we replace
265 Equivalences are looked up based on their value (think of it as
266 the RHS of an assignment). A value may be an SSA_NAME or an
267 invariant. We may have several SSA_NAMEs with the same value,
268 so with each value we have a list of SSA_NAMEs that have the
271 /* As we enter each block we record the value for any edge equivalency
272 leading to this block. If no such edge equivalency exists, then we
273 record NULL. These equivalences are live until we leave the dominator
274 subtree rooted at the block where we record the equivalency. */
275 static VEC(tree
,heap
) *equiv_stack
;
277 /* Global hash table implementing a mapping from invariant values
278 to a list of SSA_NAMEs which have the same value. We might be
279 able to reuse tree-vn for this code. */
282 /* Main structure for recording equivalences into our hash table. */
283 struct equiv_hash_elt
285 /* The value/key of this entry. */
288 /* List of SSA_NAMEs which have the same value/key. */
289 VEC(tree
,heap
) *equivalences
;
292 static void uncprop_initialize_block (struct dom_walk_data
*, basic_block
);
293 static void uncprop_finalize_block (struct dom_walk_data
*, basic_block
);
294 static void uncprop_into_successor_phis (struct dom_walk_data
*, basic_block
);
296 /* Hashing and equality routines for the hash table. */
299 equiv_hash (const void *p
)
301 tree
const value
= ((const struct equiv_hash_elt
*)p
)->value
;
302 return iterative_hash_expr (value
, 0);
306 equiv_eq (const void *p1
, const void *p2
)
308 tree value1
= ((const struct equiv_hash_elt
*)p1
)->value
;
309 tree value2
= ((const struct equiv_hash_elt
*)p2
)->value
;
311 return operand_equal_p (value1
, value2
, 0);
314 /* Free an instance of equiv_hash_elt. */
319 struct equiv_hash_elt
*elt
= (struct equiv_hash_elt
*) p
;
320 VEC_free (tree
, heap
, elt
->equivalences
);
324 /* Remove the most recently recorded equivalency for VALUE. */
327 remove_equivalence (tree value
)
329 struct equiv_hash_elt equiv_hash_elt
, *equiv_hash_elt_p
;
332 equiv_hash_elt
.value
= value
;
333 equiv_hash_elt
.equivalences
= NULL
;
335 slot
= htab_find_slot (equiv
, &equiv_hash_elt
, NO_INSERT
);
337 equiv_hash_elt_p
= (struct equiv_hash_elt
*) *slot
;
338 VEC_pop (tree
, equiv_hash_elt_p
->equivalences
);
341 /* Record EQUIVALENCE = VALUE into our hash table. */
344 record_equiv (tree value
, tree equivalence
)
346 struct equiv_hash_elt
*equiv_hash_elt
;
349 equiv_hash_elt
= XNEW (struct equiv_hash_elt
);
350 equiv_hash_elt
->value
= value
;
351 equiv_hash_elt
->equivalences
= NULL
;
353 slot
= htab_find_slot (equiv
, equiv_hash_elt
, INSERT
);
356 *slot
= (void *) equiv_hash_elt
;
358 free (equiv_hash_elt
);
360 equiv_hash_elt
= (struct equiv_hash_elt
*) *slot
;
362 VEC_safe_push (tree
, heap
, equiv_hash_elt
->equivalences
, equivalence
);
365 /* Main driver for un-cprop. */
368 tree_ssa_uncprop (void)
370 struct dom_walk_data walk_data
;
373 associate_equivalences_with_edges ();
375 /* Create our global data structures. */
376 equiv
= htab_create (1024, equiv_hash
, equiv_eq
, equiv_free
);
377 equiv_stack
= VEC_alloc (tree
, heap
, 2);
379 /* We're going to do a dominator walk, so ensure that we have
380 dominance information. */
381 calculate_dominance_info (CDI_DOMINATORS
);
383 /* Setup callbacks for the generic dominator tree walker. */
384 walk_data
.walk_stmts_backward
= false;
385 walk_data
.dom_direction
= CDI_DOMINATORS
;
386 walk_data
.initialize_block_local_data
= NULL
;
387 walk_data
.before_dom_children_before_stmts
= uncprop_initialize_block
;
388 walk_data
.before_dom_children_walk_stmts
= NULL
;
389 walk_data
.before_dom_children_after_stmts
= uncprop_into_successor_phis
;
390 walk_data
.after_dom_children_before_stmts
= NULL
;
391 walk_data
.after_dom_children_walk_stmts
= NULL
;
392 walk_data
.after_dom_children_after_stmts
= uncprop_finalize_block
;
393 walk_data
.global_data
= NULL
;
394 walk_data
.block_local_data_size
= 0;
395 walk_data
.interesting_blocks
= NULL
;
397 /* Now initialize the dominator walker. */
398 init_walk_dominator_tree (&walk_data
);
400 /* Recursively walk the dominator tree undoing unprofitable
401 constant/copy propagations. */
402 walk_dominator_tree (&walk_data
, ENTRY_BLOCK_PTR
);
404 /* Finalize and clean up. */
405 fini_walk_dominator_tree (&walk_data
);
407 /* EQUIV_STACK should already be empty at this point, so we just
408 need to empty elements out of the hash table, free EQUIV_STACK,
409 and cleanup the AUX field on the edges. */
411 VEC_free (tree
, heap
, equiv_stack
);
417 FOR_EACH_EDGE (e
, ei
, bb
->succs
)
430 /* We have finished processing the dominator children of BB, perform
431 any finalization actions in preparation for leaving this node in
432 the dominator tree. */
435 uncprop_finalize_block (struct dom_walk_data
*walk_data ATTRIBUTE_UNUSED
,
436 basic_block bb ATTRIBUTE_UNUSED
)
438 /* Pop the topmost value off the equiv stack. */
439 tree value
= VEC_pop (tree
, equiv_stack
);
441 /* If that value was non-null, then pop the topmost equivalency off
442 its equivalency stack. */
444 remove_equivalence (value
);
447 /* Unpropagate values from PHI nodes in successor blocks of BB. */
450 uncprop_into_successor_phis (struct dom_walk_data
*walk_data ATTRIBUTE_UNUSED
,
456 /* For each successor edge, first temporarily record any equivalence
457 on that edge. Then unpropagate values in any PHI nodes at the
458 destination of the edge. Then remove the temporary equivalence. */
459 FOR_EACH_EDGE (e
, ei
, bb
->succs
)
461 gimple_seq phis
= phi_nodes (e
->dest
);
462 gimple_stmt_iterator gsi
;
464 /* If there are no PHI nodes in this destination, then there is
465 no sense in recording any equivalences. */
469 /* Record any equivalency associated with E. */
472 struct edge_equivalency
*equiv
= (struct edge_equivalency
*) e
->aux
;
473 record_equiv (equiv
->rhs
, equiv
->lhs
);
476 /* Walk over the PHI nodes, unpropagating values. */
477 for (gsi
= gsi_start (phis
) ; !gsi_end_p (gsi
); gsi_next (&gsi
))
479 /* Sigh. We'll have more efficient access to this one day. */
480 gimple phi
= gsi_stmt (gsi
);
481 tree arg
= PHI_ARG_DEF (phi
, e
->dest_idx
);
482 struct equiv_hash_elt equiv_hash_elt
;
485 /* If the argument is not an invariant, or refers to the same
486 underlying variable as the PHI result, then there's no
487 point in un-propagating the argument. */
488 if (!is_gimple_min_invariant (arg
)
489 && SSA_NAME_VAR (arg
) != SSA_NAME_VAR (PHI_RESULT (phi
)))
492 /* Lookup this argument's value in the hash table. */
493 equiv_hash_elt
.value
= arg
;
494 equiv_hash_elt
.equivalences
= NULL
;
495 slot
= htab_find_slot (equiv
, &equiv_hash_elt
, NO_INSERT
);
499 struct equiv_hash_elt
*elt
= (struct equiv_hash_elt
*) *slot
;
502 /* Walk every equivalence with the same value. If we find
503 one with the same underlying variable as the PHI result,
504 then replace the value in the argument with its equivalent
505 SSA_NAME. Use the most recent equivalence as hopefully
506 that results in shortest lifetimes. */
507 for (j
= VEC_length (tree
, elt
->equivalences
) - 1; j
>= 0; j
--)
509 tree equiv
= VEC_index (tree
, elt
->equivalences
, j
);
511 if (SSA_NAME_VAR (equiv
) == SSA_NAME_VAR (PHI_RESULT (phi
)))
513 SET_PHI_ARG_DEF (phi
, e
->dest_idx
, equiv
);
520 /* If we had an equivalence associated with this edge, remove it. */
523 struct edge_equivalency
*equiv
= (struct edge_equivalency
*) e
->aux
;
524 remove_equivalence (equiv
->rhs
);
529 /* Ignoring loop backedges, if BB has precisely one incoming edge then
530 return that edge. Otherwise return NULL. */
532 single_incoming_edge_ignoring_loop_edges (basic_block bb
)
538 FOR_EACH_EDGE (e
, ei
, bb
->preds
)
540 /* A loop back edge can be identified by the destination of
541 the edge dominating the source of the edge. */
542 if (dominated_by_p (CDI_DOMINATORS
, e
->src
, e
->dest
))
545 /* If we have already seen a non-loop edge, then we must have
546 multiple incoming non-loop edges and thus we return NULL. */
550 /* This is the first non-loop incoming edge we have found. Record
559 uncprop_initialize_block (struct dom_walk_data
*walk_data ATTRIBUTE_UNUSED
,
564 bool recorded
= false;
566 /* If this block is dominated by a single incoming edge and that edge
567 has an equivalency, then record the equivalency and push the
568 VALUE onto EQUIV_STACK. Else push a NULL entry on EQUIV_STACK. */
569 parent
= get_immediate_dominator (CDI_DOMINATORS
, bb
);
572 e
= single_incoming_edge_ignoring_loop_edges (bb
);
574 if (e
&& e
->src
== parent
&& e
->aux
)
576 struct edge_equivalency
*equiv
= (struct edge_equivalency
*) e
->aux
;
578 record_equiv (equiv
->rhs
, equiv
->lhs
);
579 VEC_safe_push (tree
, heap
, equiv_stack
, equiv
->rhs
);
585 VEC_safe_push (tree
, heap
, equiv_stack
, NULL_TREE
);
591 return flag_tree_dom
!= 0;
594 struct gimple_opt_pass pass_uncprop
=
598 "uncprop", /* name */
599 gate_uncprop
, /* gate */
600 tree_ssa_uncprop
, /* execute */
603 0, /* static_pass_number */
604 TV_TREE_SSA_UNCPROP
, /* tv_id */
605 PROP_cfg
| PROP_ssa
, /* properties_required */
606 0, /* properties_provided */
607 0, /* properties_destroyed */
608 0, /* todo_flags_start */
609 TODO_dump_func
| TODO_verify_ssa
/* todo_flags_finish */