1 /* Code sinking for trees
2 Copyright (C) 2001-2021 Free Software Foundation, Inc.
3 Contributed by Daniel Berlin <dan@dberlin.org>
5 This file is part of GCC.
7 GCC is free software; you can redistribute it and/or modify
8 it under the terms of the GNU General Public License as published by
9 the Free Software Foundation; either version 3, or (at your option)
12 GCC is distributed in the hope that it will be useful,
13 but WITHOUT ANY WARRANTY; without even the implied warranty of
14 MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
15 GNU General Public License for more details.
17 You should have received a copy of the GNU General Public License
18 along with GCC; see the file COPYING3. If not see
19 <http://www.gnu.org/licenses/>. */
23 #include "coretypes.h"
28 #include "tree-pass.h"
30 #include "gimple-pretty-print.h"
31 #include "fold-const.h"
32 #include "stor-layout.h"
34 #include "gimple-iterator.h"
40 1. Sinking store only using scalar promotion (IE without moving the RHS):
60 Store copy propagation will take care of the store elimination above.
63 2. Sinking using Partial Dead Code Elimination. */
68 /* The number of statements sunk down the flowgraph by code sinking. */
71 /* The number of stores commoned and sunk down by store commoning. */
76 /* Given a PHI, and one of its arguments (DEF), find the edge for
77 that argument and return it. If the argument occurs twice in the PHI node,
81 find_bb_for_arg (gphi
*phi
, tree def
)
84 bool foundone
= false;
85 basic_block result
= NULL
;
86 for (i
= 0; i
< gimple_phi_num_args (phi
); i
++)
87 if (PHI_ARG_DEF (phi
, i
) == def
)
92 result
= gimple_phi_arg_edge (phi
, i
)->src
;
97 /* When the first immediate use is in a statement, then return true if all
98 immediate uses in IMM are in the same statement.
99 We could also do the case where the first immediate use is in a phi node,
100 and all the other uses are in phis in the same basic block, but this
101 requires some expensive checking later (you have to make sure no def/vdef
102 in the statement occurs for multiple edges in the various phi nodes it's
103 used in, so that you only have one place you can sink it to. */
106 all_immediate_uses_same_place (def_operand_p def_p
)
108 tree var
= DEF_FROM_PTR (def_p
);
109 imm_use_iterator imm_iter
;
112 gimple
*firstuse
= NULL
;
113 FOR_EACH_IMM_USE_FAST (use_p
, imm_iter
, var
)
115 if (is_gimple_debug (USE_STMT (use_p
)))
117 if (firstuse
== NULL
)
118 firstuse
= USE_STMT (use_p
);
120 if (firstuse
!= USE_STMT (use_p
))
127 /* Find the nearest common dominator of all of the immediate uses in IMM. */
130 nearest_common_dominator_of_uses (def_operand_p def_p
, bool *debug_stmts
)
132 tree var
= DEF_FROM_PTR (def_p
);
134 basic_block commondom
;
137 imm_use_iterator imm_iter
;
140 FOR_EACH_IMM_USE_FAST (use_p
, imm_iter
, var
)
142 gimple
*usestmt
= USE_STMT (use_p
);
143 basic_block useblock
;
145 if (gphi
*phi
= dyn_cast
<gphi
*> (usestmt
))
147 int idx
= PHI_ARG_INDEX_FROM_USE (use_p
);
149 useblock
= gimple_phi_arg_edge (phi
, idx
)->src
;
151 else if (is_gimple_debug (usestmt
))
158 useblock
= gimple_bb (usestmt
);
161 /* Short circuit. Nothing dominates the entry block. */
162 if (useblock
== ENTRY_BLOCK_PTR_FOR_FN (cfun
))
165 bitmap_set_bit (blocks
, useblock
->index
);
167 commondom
= BASIC_BLOCK_FOR_FN (cfun
, bitmap_first_set_bit (blocks
));
168 EXECUTE_IF_SET_IN_BITMAP (blocks
, 0, j
, bi
)
169 commondom
= nearest_common_dominator (CDI_DOMINATORS
, commondom
,
170 BASIC_BLOCK_FOR_FN (cfun
, j
));
174 /* Given EARLY_BB and LATE_BB, two blocks in a path through the dominator
175 tree, return the best basic block between them (inclusive) to place
178 We want the most control dependent block in the shallowest loop nest.
180 If the resulting block is in a shallower loop nest, then use it. Else
181 only use the resulting block if it has significantly lower execution
182 frequency than EARLY_BB to avoid gratuitous statement movement. We
183 consider statements with VOPS more desirable to move.
185 This pass would obviously benefit from PDO as it utilizes block
186 frequencies. It would also benefit from recomputing frequencies
187 if profile data is not available since frequencies often get out
188 of sync with reality. */
191 select_best_block (basic_block early_bb
,
195 basic_block best_bb
= late_bb
;
196 basic_block temp_bb
= late_bb
;
199 while (temp_bb
!= early_bb
)
201 /* If we've moved into a lower loop nest, then that becomes
203 if (bb_loop_depth (temp_bb
) < bb_loop_depth (best_bb
))
206 /* Walk up the dominator tree, hopefully we'll find a shallower
208 temp_bb
= get_immediate_dominator (CDI_DOMINATORS
, temp_bb
);
211 /* If we found a shallower loop nest, then we always consider that
212 a win. This will always give us the most control dependent block
213 within that loop nest. */
214 if (bb_loop_depth (best_bb
) < bb_loop_depth (early_bb
))
217 /* Get the sinking threshold. If the statement to be moved has memory
218 operands, then increase the threshold by 7% as those are even more
219 profitable to avoid, clamping at 100%. */
220 threshold
= param_sink_frequency_threshold
;
221 if (gimple_vuse (stmt
) || gimple_vdef (stmt
))
228 /* If BEST_BB is at the same nesting level, then require it to have
229 significantly lower execution frequency to avoid gratuitous movement. */
230 if (bb_loop_depth (best_bb
) == bb_loop_depth (early_bb
)
231 /* If result of comparsion is unknown, prefer EARLY_BB.
232 Thus use !(...>=..) rather than (...<...) */
233 && !(best_bb
->count
.apply_scale (100, 1)
234 >= early_bb
->count
.apply_scale (threshold
, 1)))
237 /* No better block found, so return EARLY_BB, which happens to be the
238 statement's original block. */
242 /* Given a statement (STMT) and the basic block it is currently in (FROMBB),
243 determine the location to sink the statement to, if any.
244 Returns true if there is such location; in that case, TOGSI points to the
245 statement before that STMT should be moved. */
248 statement_sink_location (gimple
*stmt
, basic_block frombb
,
249 gimple_stmt_iterator
*togsi
, bool *zero_uses_p
)
252 use_operand_p one_use
= NULL_USE_OPERAND_P
;
257 imm_use_iterator imm_iter
;
259 *zero_uses_p
= false;
261 /* We only can sink assignments and non-looping const/pure calls. */
263 if (!is_gimple_assign (stmt
)
264 && (!is_gimple_call (stmt
)
265 || !((cf
= gimple_call_flags (stmt
)) & (ECF_CONST
|ECF_PURE
))
266 || (cf
& ECF_LOOPING_CONST_OR_PURE
)))
269 /* We only can sink stmts with a single definition. */
270 def_p
= single_ssa_def_operand (stmt
, SSA_OP_ALL_DEFS
);
271 if (def_p
== NULL_DEF_OPERAND_P
)
274 /* There are a few classes of things we can't or don't move, some because we
275 don't have code to handle it, some because it's not profitable and some
276 because it's not legal.
278 We can't sink things that may be global stores, at least not without
279 calculating a lot more information, because we may cause it to no longer
280 be seen by an external routine that needs it depending on where it gets
283 We can't sink statements that end basic blocks without splitting the
284 incoming edge for the sink location to place it there.
286 We can't sink statements that have volatile operands.
288 We don't want to sink dead code, so anything with 0 immediate uses is not
291 Don't sink BLKmode assignments if current function has any local explicit
292 register variables, as BLKmode assignments may involve memcpy or memset
293 calls or, on some targets, inline expansion thereof that sometimes need
294 to use specific hard registers.
297 if (stmt_ends_bb_p (stmt
)
298 || gimple_has_side_effects (stmt
)
299 || (cfun
->has_local_explicit_reg_vars
300 && TYPE_MODE (TREE_TYPE (gimple_get_lhs (stmt
))) == BLKmode
))
303 /* Return if there are no immediate uses of this stmt. */
304 if (has_zero_uses (DEF_FROM_PTR (def_p
)))
310 if (SSA_NAME_OCCURS_IN_ABNORMAL_PHI (DEF_FROM_PTR (def_p
)))
313 FOR_EACH_SSA_USE_OPERAND (use_p
, stmt
, iter
, SSA_OP_ALL_USES
)
315 tree use
= USE_FROM_PTR (use_p
);
316 if (SSA_NAME_OCCURS_IN_ABNORMAL_PHI (use
))
322 /* If stmt is a store the one and only use needs to be the VOP
324 if (virtual_operand_p (DEF_FROM_PTR (def_p
)))
326 FOR_EACH_IMM_USE_FAST (use_p
, imm_iter
, DEF_FROM_PTR (def_p
))
328 gimple
*use_stmt
= USE_STMT (use_p
);
330 /* A killing definition is not a use. */
331 if ((gimple_has_lhs (use_stmt
)
332 && operand_equal_p (gimple_get_lhs (stmt
),
333 gimple_get_lhs (use_stmt
), 0))
334 || stmt_kills_ref_p (use_stmt
, gimple_get_lhs (stmt
)))
336 /* If use_stmt is or might be a nop assignment then USE_STMT
337 acts as a use as well as definition. */
339 && ref_maybe_used_by_stmt_p (use_stmt
,
340 gimple_get_lhs (stmt
)))
345 if (gimple_code (use_stmt
) != GIMPLE_PHI
)
357 /* If all the immediate uses are not in the same place, find the nearest
358 common dominator of all the immediate uses. For PHI nodes, we have to
359 find the nearest common dominator of all of the predecessor blocks, since
360 that is where insertion would have to take place. */
361 else if (gimple_vuse (stmt
)
362 || !all_immediate_uses_same_place (def_p
))
364 bool debug_stmts
= false;
365 basic_block commondom
= nearest_common_dominator_of_uses (def_p
,
368 if (commondom
== frombb
)
371 /* If this is a load then do not sink past any stores.
372 Look for virtual definitions in the path from frombb to the sink
373 location computed from the real uses and if found, adjust
374 that it a common dominator. */
375 if (gimple_vuse (stmt
))
377 /* Do not sink loads from hard registers. */
378 if (gimple_assign_single_p (stmt
)
379 && TREE_CODE (gimple_assign_rhs1 (stmt
)) == VAR_DECL
380 && DECL_HARD_REGISTER (gimple_assign_rhs1 (stmt
)))
383 imm_use_iterator imm_iter
;
385 FOR_EACH_IMM_USE_FAST (use_p
, imm_iter
, gimple_vuse (stmt
))
387 gimple
*use_stmt
= USE_STMT (use_p
);
388 basic_block bb
= gimple_bb (use_stmt
);
389 /* For PHI nodes the block we know sth about is the incoming block
391 if (gimple_code (use_stmt
) == GIMPLE_PHI
)
393 /* In case the PHI node post-dominates the current insert
394 location we can disregard it. But make sure it is not
395 dominating it as well as can happen in a CFG cycle. */
397 && !dominated_by_p (CDI_DOMINATORS
, commondom
, bb
)
398 && dominated_by_p (CDI_POST_DOMINATORS
, commondom
, bb
)
399 /* If the blocks are possibly within the same irreducible
400 cycle the above check breaks down. */
401 && !(bb
->flags
& commondom
->flags
& BB_IRREDUCIBLE_LOOP
))
403 bb
= EDGE_PRED (bb
, PHI_ARG_INDEX_FROM_USE (use_p
))->src
;
405 else if (!gimple_vdef (use_stmt
))
407 /* If the use is not dominated by the path entry it is not on
409 if (!dominated_by_p (CDI_DOMINATORS
, bb
, frombb
))
411 /* There is no easy way to disregard defs not on the path from
412 frombb to commondom so just consider them all. */
413 commondom
= nearest_common_dominator (CDI_DOMINATORS
,
415 if (commondom
== frombb
)
420 /* Our common dominator has to be dominated by frombb in order to be a
421 trivially safe place to put this statement, since it has multiple
423 if (!dominated_by_p (CDI_DOMINATORS
, commondom
, frombb
))
426 commondom
= select_best_block (frombb
, commondom
, stmt
);
428 if (commondom
== frombb
)
431 *togsi
= gsi_after_labels (commondom
);
437 FOR_EACH_IMM_USE_FAST (one_use
, imm_iter
, DEF_FROM_PTR (def_p
))
439 if (is_gimple_debug (USE_STMT (one_use
)))
443 use
= USE_STMT (one_use
);
445 if (gimple_code (use
) != GIMPLE_PHI
)
447 sinkbb
= select_best_block (frombb
, gimple_bb (use
), stmt
);
449 if (sinkbb
== frombb
)
452 if (sinkbb
== gimple_bb (use
))
453 *togsi
= gsi_for_stmt (use
);
455 *togsi
= gsi_after_labels (sinkbb
);
461 sinkbb
= find_bb_for_arg (as_a
<gphi
*> (use
), DEF_FROM_PTR (def_p
));
463 /* This can happen if there are multiple uses in a PHI. */
467 sinkbb
= select_best_block (frombb
, sinkbb
, stmt
);
468 if (!sinkbb
|| sinkbb
== frombb
)
471 /* If the latch block is empty, don't make it non-empty by sinking
472 something into it. */
473 if (sinkbb
== frombb
->loop_father
->latch
474 && empty_block_p (sinkbb
))
477 *togsi
= gsi_after_labels (sinkbb
);
482 /* Very simplistic code to sink common stores from the predecessor through
483 our virtual PHI. We do this before sinking stmts from BB as it might
484 expose sinking opportunities of the merged stores.
485 Once we have partial dead code elimination through sth like SSU-PRE this
486 should be moved there. */
489 sink_common_stores_to_bb (basic_block bb
)
494 if (EDGE_COUNT (bb
->preds
) > 1
495 && (phi
= get_virtual_phi (bb
)))
497 /* Repeat until no more common stores are found. */
500 gimple
*first_store
= NULL
;
501 auto_vec
<tree
, 5> vdefs
;
502 gimple_stmt_iterator gsi
;
504 /* Search for common stores defined by all virtual PHI args.
505 ??? Common stores not present in all predecessors could
506 be handled by inserting a forwarder to sink to. Generally
507 this involves deciding which stores to do this for if
508 multiple common stores are present for different sets of
509 predecessors. See PR11832 for an interesting case. */
510 for (unsigned i
= 0; i
< gimple_phi_num_args (phi
); ++i
)
512 tree arg
= gimple_phi_arg_def (phi
, i
);
513 gimple
*def
= SSA_NAME_DEF_STMT (arg
);
514 if (! is_gimple_assign (def
)
515 || stmt_can_throw_internal (cfun
, def
)
516 || (gimple_phi_arg_edge (phi
, i
)->flags
& EDGE_ABNORMAL
))
518 /* ??? We could handle some cascading with the def being
519 another PHI. We'd have to insert multiple PHIs for
520 the rhs then though (if they are not all equal). */
524 /* ??? Do not try to do anything fancy with aliasing, thus
525 do not sink across non-aliased loads (or even stores,
526 so different store order will make the sinking fail). */
527 bool all_uses_on_phi
= true;
528 imm_use_iterator iter
;
530 FOR_EACH_IMM_USE_FAST (use_p
, iter
, arg
)
531 if (USE_STMT (use_p
) != phi
)
533 all_uses_on_phi
= false;
536 if (! all_uses_on_phi
)
541 /* Check all stores are to the same LHS. */
544 /* ??? We could handle differing SSA uses in the LHS by inserting
546 else if (! operand_equal_p (gimple_assign_lhs (first_store
),
547 gimple_assign_lhs (def
), 0)
548 || (gimple_clobber_p (first_store
)
549 != gimple_clobber_p (def
)))
554 vdefs
.safe_push (arg
);
559 /* Check if we need a PHI node to merge the stored values. */
561 if (!gimple_clobber_p (first_store
))
562 for (unsigned i
= 1; i
< vdefs
.length (); ++i
)
564 gimple
*def
= SSA_NAME_DEF_STMT (vdefs
[i
]);
565 if (! operand_equal_p (gimple_assign_rhs1 (first_store
),
566 gimple_assign_rhs1 (def
), 0))
573 /* We cannot handle aggregate values if we need to merge them. */
574 tree type
= TREE_TYPE (gimple_assign_lhs (first_store
));
576 && ! is_gimple_reg_type (type
))
579 if (dump_enabled_p ())
581 dump_printf_loc (MSG_OPTIMIZED_LOCATIONS
,
583 "sinking common stores %sto ",
584 allsame
? "with same value " : "");
585 dump_generic_expr (MSG_OPTIMIZED_LOCATIONS
, TDF_SLIM
,
586 gimple_assign_lhs (first_store
));
587 dump_printf (MSG_OPTIMIZED_LOCATIONS
, "\n");
590 /* Insert a PHI to merge differing stored values if necessary.
591 Note that in general inserting PHIs isn't a very good idea as
592 it makes the job of coalescing and register allocation harder.
593 Even common SSA uses on the rhs/lhs might extend their lifetime
594 across multiple edges by this code motion which makes
595 register allocation harder. */
599 from
= make_ssa_name (type
);
600 gphi
*newphi
= create_phi_node (from
, bb
);
601 for (unsigned i
= 0; i
< vdefs
.length (); ++i
)
603 gimple
*def
= SSA_NAME_DEF_STMT (vdefs
[i
]);
604 add_phi_arg (newphi
, gimple_assign_rhs1 (def
),
605 EDGE_PRED (bb
, i
), UNKNOWN_LOCATION
);
609 from
= gimple_assign_rhs1 (first_store
);
611 /* Remove all stores. */
612 for (unsigned i
= 0; i
< vdefs
.length (); ++i
)
613 TREE_VISITED (vdefs
[i
]) = 1;
614 for (unsigned i
= 0; i
< vdefs
.length (); ++i
)
615 /* If we have more than one use of a VDEF on the PHI make sure
616 we remove the defining stmt only once. */
617 if (TREE_VISITED (vdefs
[i
]))
619 TREE_VISITED (vdefs
[i
]) = 0;
620 gimple
*def
= SSA_NAME_DEF_STMT (vdefs
[i
]);
621 gsi
= gsi_for_stmt (def
);
622 unlink_stmt_vdef (def
);
623 gsi_remove (&gsi
, true);
627 /* Insert the first store at the beginning of the merge BB. */
628 gimple_set_vdef (first_store
, gimple_phi_result (phi
));
629 SSA_NAME_DEF_STMT (gimple_vdef (first_store
)) = first_store
;
630 gimple_phi_set_result (phi
, make_ssa_name (gimple_vop (cfun
)));
631 gimple_set_vuse (first_store
, gimple_phi_result (phi
));
632 gimple_assign_set_rhs1 (first_store
, from
);
633 /* ??? Should we reset first_stores location? */
634 gsi
= gsi_after_labels (bb
);
635 gsi_insert_before (&gsi
, first_store
, GSI_SAME_STMT
);
636 sink_stats
.commoned
++;
638 todo
|= TODO_cleanup_cfg
;
641 /* We could now have empty predecessors that we could remove,
642 forming a proper CFG for further sinking. Note that even
643 CFG cleanup doesn't do this fully at the moment and it
644 doesn't preserve post-dominators in the process either.
645 The mergephi pass might do it though. gcc.dg/tree-ssa/ssa-sink-13.c
646 shows this nicely if you disable tail merging or (same effect)
647 make the stored values unequal. */
653 /* Perform code sinking on BB */
656 sink_code_in_bb (basic_block bb
)
659 gimple_stmt_iterator gsi
;
665 /* Sink common stores from the predecessor through our virtual PHI. */
666 todo
|= sink_common_stores_to_bb (bb
);
668 /* If this block doesn't dominate anything, there can't be any place to sink
669 the statements to. */
670 if (first_dom_son (CDI_DOMINATORS
, bb
) == NULL
)
673 /* We can't move things across abnormal edges, so don't try. */
674 FOR_EACH_EDGE (e
, ei
, bb
->succs
)
675 if (e
->flags
& EDGE_ABNORMAL
)
678 for (gsi
= gsi_last_bb (bb
); !gsi_end_p (gsi
);)
680 gimple
*stmt
= gsi_stmt (gsi
);
681 gimple_stmt_iterator togsi
;
684 if (!statement_sink_location (stmt
, bb
, &togsi
, &zero_uses_p
))
686 gimple_stmt_iterator saved
= gsi
;
687 if (!gsi_end_p (gsi
))
689 /* If we face a dead stmt remove it as it possibly blocks
692 && ! gimple_vdef (stmt
))
694 gsi_remove (&saved
, true);
703 fprintf (dump_file
, "Sinking ");
704 print_gimple_stmt (dump_file
, stmt
, 0, TDF_VOPS
);
705 fprintf (dump_file
, " from bb %d to bb %d\n",
706 bb
->index
, (gsi_bb (togsi
))->index
);
709 /* Update virtual operands of statements in the path we
711 if (gimple_vdef (stmt
))
713 imm_use_iterator iter
;
717 FOR_EACH_IMM_USE_STMT (vuse_stmt
, iter
, gimple_vdef (stmt
))
718 if (gimple_code (vuse_stmt
) != GIMPLE_PHI
)
719 FOR_EACH_IMM_USE_ON_STMT (use_p
, iter
)
720 SET_USE (use_p
, gimple_vuse (stmt
));
723 /* If this is the end of the basic block, we need to insert at the end
724 of the basic block. */
725 if (gsi_end_p (togsi
))
726 gsi_move_to_bb_end (&gsi
, gsi_bb (togsi
));
728 gsi_move_before (&gsi
, &togsi
);
732 /* If we've just removed the last statement of the BB, the
733 gsi_end_p() test below would fail, but gsi_prev() would have
734 succeeded, and we want it to succeed. So we keep track of
735 whether we're at the last statement and pick up the new last
739 gsi
= gsi_last_bb (bb
);
744 if (!gsi_end_p (gsi
))
749 for (son
= first_dom_son (CDI_POST_DOMINATORS
, bb
);
751 son
= next_dom_son (CDI_POST_DOMINATORS
, son
))
753 todo
|= sink_code_in_bb (son
);
759 /* Perform code sinking.
760 This moves code down the flowgraph when we know it would be
761 profitable to do so, or it wouldn't increase the number of
762 executions of the statement.
775 a_6 = PHI (a_5, a_1);
778 we'll transform this into:
789 a_6 = PHI (a_5, a_1);
792 Note that this reduces the number of computations of a = b + c to 1
793 when we take the else edge, instead of 2.
797 const pass_data pass_data_sink_code
=
799 GIMPLE_PASS
, /* type */
801 OPTGROUP_NONE
, /* optinfo_flags */
802 TV_TREE_SINK
, /* tv_id */
803 /* PROP_no_crit_edges is ensured by running split_edges_for_insertion in
804 pass_data_sink_code::execute (). */
805 ( PROP_cfg
| PROP_ssa
), /* properties_required */
806 0, /* properties_provided */
807 0, /* properties_destroyed */
808 0, /* todo_flags_start */
809 TODO_update_ssa
, /* todo_flags_finish */
812 class pass_sink_code
: public gimple_opt_pass
815 pass_sink_code (gcc::context
*ctxt
)
816 : gimple_opt_pass (pass_data_sink_code
, ctxt
)
819 /* opt_pass methods: */
820 virtual bool gate (function
*) { return flag_tree_sink
!= 0; }
821 virtual unsigned int execute (function
*);
823 }; // class pass_sink_code
826 pass_sink_code::execute (function
*fun
)
828 loop_optimizer_init (LOOPS_NORMAL
);
829 split_edges_for_insertion ();
830 connect_infinite_loops_to_exit ();
831 memset (&sink_stats
, 0, sizeof (sink_stats
));
832 calculate_dominance_info (CDI_DOMINATORS
);
833 calculate_dominance_info (CDI_POST_DOMINATORS
);
834 unsigned todo
= sink_code_in_bb (EXIT_BLOCK_PTR_FOR_FN (fun
));
835 statistics_counter_event (fun
, "Sunk statements", sink_stats
.sunk
);
836 statistics_counter_event (fun
, "Commoned stores", sink_stats
.commoned
);
837 free_dominance_info (CDI_POST_DOMINATORS
);
838 remove_fake_exit_edges ();
839 loop_optimizer_finalize ();
847 make_pass_sink_code (gcc::context
*ctxt
)
849 return new pass_sink_code (ctxt
);