1 /* Function splitting pass
2 Copyright (C) 2010-2022 Free Software Foundation, Inc.
3 Contributed by Jan Hubicka <jh@suse.cz>
5 This file is part of GCC.
7 GCC is free software; you can redistribute it and/or modify it under
8 the terms of the GNU General Public License as published by the Free
9 Software Foundation; either version 3, or (at your option) any later
12 GCC is distributed in the hope that it will be useful, but WITHOUT ANY
13 WARRANTY; without even the implied warranty of MERCHANTABILITY or
14 FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
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/>. */
21 /* The purpose of this pass is to split function bodies to improve
22 inlining. I.e. for function of the form:
47 When func becomes inlinable and when cheap_test is often true, inlining func,
48 but not fund.part leads to performance improvement similar as inlining
49 original func while the code size growth is smaller.
51 The pass is organized in three stages:
52 1) Collect local info about basic block into BB_INFO structure and
53 compute function body estimated size and time.
54 2) Via DFS walk find all possible basic blocks where we can split
56 3) If split point is found, split at the specified BB by creating a clone
57 and updating function to call it.
59 The decisions what functions to split are in execute_split_functions
62 There are several possible future improvements for this pass including:
64 1) Splitting to break up large functions
65 2) Splitting to reduce stack frame usage
66 3) Allow split part of function to use values computed in the header part.
67 The values needs to be passed to split function, perhaps via same
68 interface as for nested functions or as argument.
69 4) Support for simple rematerialization. I.e. when split part use
70 value computed in header from function parameter in very cheap way, we
71 can just recompute it.
72 5) Support splitting of nested functions.
73 6) Support non-SSA arguments.
74 7) There is nothing preventing us from producing multiple parts of single function
75 when needed or splitting also the parts. */
79 #include "coretypes.h"
85 #include "alloc-pool.h"
86 #include "tree-pass.h"
89 #include "diagnostic.h"
90 #include "fold-const.h"
94 #include "gimple-iterator.h"
95 #include "gimplify-me.h"
96 #include "gimple-walk.h"
97 #include "symbol-summary.h"
100 #include "tree-into-ssa.h"
101 #include "tree-dfa.h"
102 #include "tree-inline.h"
103 #include "gimple-pretty-print.h"
104 #include "ipa-fnsummary.h"
108 /* Per basic block info. */
117 static vec
<split_bb_info
> bb_info_vec
;
119 /* Description of split point. */
124 /* Size of the partitions. */
125 sreal header_time
, split_time
;
126 unsigned int header_size
, split_size
;
128 /* SSA names that need to be passed into spit function. */
129 bitmap ssa_names_to_pass
;
131 /* Basic block where we split (that will become entry point of new function. */
132 basic_block entry_bb
;
134 /* Count for entering the split part.
135 This is not count of the entry_bb because it may be in loop. */
138 /* Basic blocks we are splitting away. */
141 /* True when return value is computed on split part and thus it needs
143 bool split_part_set_retval
;
146 /* Best split point found. */
148 class split_point best_split_point
;
150 /* Set of basic blocks that are not allowed to dominate a split point. */
152 static bitmap forbidden_dominators
;
154 static tree
find_retval (basic_block return_bb
);
156 /* Callback for walk_stmt_load_store_addr_ops. If T is non-SSA automatic
157 variable, check it if it is present in bitmap passed via DATA. */
160 test_nonssa_use (gimple
*, tree t
, tree
, void *data
)
162 t
= get_base_address (t
);
164 if (!t
|| is_gimple_reg (t
))
167 if (TREE_CODE (t
) == PARM_DECL
169 && auto_var_in_fn_p (t
, current_function_decl
))
170 || TREE_CODE (t
) == RESULT_DECL
171 /* Normal labels are part of CFG and will be handled gratefully.
172 Forced labels however can be used directly by statements and
173 need to stay in one partition along with their uses. */
174 || (TREE_CODE (t
) == LABEL_DECL
175 && FORCED_LABEL (t
)))
176 return bitmap_bit_p ((bitmap
)data
, DECL_UID (t
));
178 /* For DECL_BY_REFERENCE, the return value is actually a pointer. We want
179 to pretend that the value pointed to is actual result decl. */
180 if ((TREE_CODE (t
) == MEM_REF
|| INDIRECT_REF_P (t
))
181 && TREE_CODE (TREE_OPERAND (t
, 0)) == SSA_NAME
182 && SSA_NAME_VAR (TREE_OPERAND (t
, 0))
183 && TREE_CODE (SSA_NAME_VAR (TREE_OPERAND (t
, 0))) == RESULT_DECL
184 && DECL_BY_REFERENCE (DECL_RESULT (current_function_decl
)))
186 bitmap_bit_p ((bitmap
)data
,
187 DECL_UID (DECL_RESULT (current_function_decl
)));
192 /* Dump split point CURRENT. */
195 dump_split_point (FILE * file
, class split_point
*current
)
198 "Split point at BB %i\n"
199 " header time: %f header size: %i\n"
200 " split time: %f split size: %i\n bbs: ",
201 current
->entry_bb
->index
, current
->header_time
.to_double (),
202 current
->header_size
, current
->split_time
.to_double (),
203 current
->split_size
);
204 dump_bitmap (file
, current
->split_bbs
);
205 fprintf (file
, " SSA names to pass: ");
206 dump_bitmap (file
, current
->ssa_names_to_pass
);
209 /* Look for all BBs in header that might lead to the split part and verify
210 that they are not defining any non-SSA var used by the split part.
211 Parameters are the same as for consider_split. */
214 verify_non_ssa_vars (class split_point
*current
, bitmap non_ssa_vars
,
215 basic_block return_bb
)
217 bitmap seen
= BITMAP_ALLOC (NULL
);
218 vec
<basic_block
> worklist
= vNULL
;
224 FOR_EACH_EDGE (e
, ei
, current
->entry_bb
->preds
)
225 if (e
->src
!= ENTRY_BLOCK_PTR_FOR_FN (cfun
)
226 && !bitmap_bit_p (current
->split_bbs
, e
->src
->index
))
228 worklist
.safe_push (e
->src
);
229 bitmap_set_bit (seen
, e
->src
->index
);
232 while (!worklist
.is_empty ())
234 bb
= worklist
.pop ();
235 FOR_EACH_EDGE (e
, ei
, bb
->preds
)
236 if (e
->src
!= ENTRY_BLOCK_PTR_FOR_FN (cfun
)
237 && bitmap_set_bit (seen
, e
->src
->index
))
239 gcc_checking_assert (!bitmap_bit_p (current
->split_bbs
,
241 worklist
.safe_push (e
->src
);
243 for (gimple_stmt_iterator bsi
= gsi_start_bb (bb
); !gsi_end_p (bsi
);
246 gimple
*stmt
= gsi_stmt (bsi
);
247 if (is_gimple_debug (stmt
))
249 if (walk_stmt_load_store_addr_ops
250 (stmt
, non_ssa_vars
, test_nonssa_use
, test_nonssa_use
,
256 if (glabel
*label_stmt
= dyn_cast
<glabel
*> (stmt
))
257 if (test_nonssa_use (stmt
, gimple_label_label (label_stmt
),
258 NULL_TREE
, non_ssa_vars
))
264 for (gphi_iterator bsi
= gsi_start_phis (bb
); !gsi_end_p (bsi
);
267 if (walk_stmt_load_store_addr_ops
268 (gsi_stmt (bsi
), non_ssa_vars
, test_nonssa_use
, test_nonssa_use
,
275 FOR_EACH_EDGE (e
, ei
, bb
->succs
)
277 if (e
->dest
!= return_bb
)
279 for (gphi_iterator bsi
= gsi_start_phis (return_bb
);
283 gphi
*stmt
= bsi
.phi ();
284 tree op
= gimple_phi_arg_def (stmt
, e
->dest_idx
);
286 if (virtual_operand_p (gimple_phi_result (stmt
)))
288 if (TREE_CODE (op
) != SSA_NAME
289 && test_nonssa_use (stmt
, op
, op
, non_ssa_vars
))
298 /* Verify that the rest of function does not define any label
299 used by the split part. */
300 FOR_EACH_BB_FN (bb
, cfun
)
301 if (!bitmap_bit_p (current
->split_bbs
, bb
->index
)
302 && !bitmap_bit_p (seen
, bb
->index
))
304 gimple_stmt_iterator bsi
;
305 for (bsi
= gsi_start_bb (bb
); !gsi_end_p (bsi
); gsi_next (&bsi
))
306 if (glabel
*label_stmt
= dyn_cast
<glabel
*> (gsi_stmt (bsi
)))
308 if (test_nonssa_use (label_stmt
,
309 gimple_label_label (label_stmt
),
310 NULL_TREE
, non_ssa_vars
))
326 /* If STMT is a call, check the callee against a list of forbidden
327 predicate functions. If a match is found, look for uses of the
328 call result in condition statements that compare against zero.
329 For each such use, find the block targeted by the condition
330 statement for the nonzero result, and set the bit for this block
331 in the forbidden dominators bitmap. The purpose of this is to avoid
332 selecting a split point where we are likely to lose the chance
333 to optimize away an unused function call. */
336 check_forbidden_calls (gimple
*stmt
)
338 imm_use_iterator use_iter
;
342 /* At the moment, __builtin_constant_p is the only forbidden
343 predicate function call (see PR49642). */
344 if (!gimple_call_builtin_p (stmt
, BUILT_IN_CONSTANT_P
))
347 lhs
= gimple_call_lhs (stmt
);
349 if (!lhs
|| TREE_CODE (lhs
) != SSA_NAME
)
352 FOR_EACH_IMM_USE_FAST (use_p
, use_iter
, lhs
)
355 basic_block use_bb
, forbidden_bb
;
357 edge true_edge
, false_edge
;
360 use_stmt
= dyn_cast
<gcond
*> (USE_STMT (use_p
));
364 /* Assuming canonical form for GIMPLE_COND here, with constant
365 in second position. */
366 op1
= gimple_cond_rhs (use_stmt
);
367 code
= gimple_cond_code (use_stmt
);
368 use_bb
= gimple_bb (use_stmt
);
370 extract_true_false_edges_from_block (use_bb
, &true_edge
, &false_edge
);
372 /* We're only interested in comparisons that distinguish
373 unambiguously from zero. */
374 if (!integer_zerop (op1
) || code
== LE_EXPR
|| code
== GE_EXPR
)
378 forbidden_bb
= false_edge
->dest
;
380 forbidden_bb
= true_edge
->dest
;
382 bitmap_set_bit (forbidden_dominators
, forbidden_bb
->index
);
386 /* If BB is dominated by any block in the forbidden dominators set,
387 return TRUE; else FALSE. */
390 dominated_by_forbidden (basic_block bb
)
395 EXECUTE_IF_SET_IN_BITMAP (forbidden_dominators
, 1, dom_bb
, bi
)
397 if (dominated_by_p (CDI_DOMINATORS
, bb
,
398 BASIC_BLOCK_FOR_FN (cfun
, dom_bb
)))
405 /* For give split point CURRENT and return block RETURN_BB return 1
406 if ssa name VAL is set by split part and 0 otherwise. */
408 split_part_set_ssa_name_p (tree val
, class split_point
*current
,
409 basic_block return_bb
)
411 if (TREE_CODE (val
) != SSA_NAME
)
414 return (!SSA_NAME_IS_DEFAULT_DEF (val
)
415 && (bitmap_bit_p (current
->split_bbs
,
416 gimple_bb (SSA_NAME_DEF_STMT (val
))->index
)
417 || gimple_bb (SSA_NAME_DEF_STMT (val
)) == return_bb
));
420 /* We found an split_point CURRENT. NON_SSA_VARS is bitmap of all non ssa
421 variables used and RETURN_BB is return basic block.
422 See if we can split function here. */
425 consider_split (class split_point
*current
, bitmap non_ssa_vars
,
426 basic_block return_bb
)
429 unsigned int num_args
= 0;
430 unsigned int call_overhead
;
436 bool back_edge
= false;
438 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
439 dump_split_point (dump_file
, current
);
441 current
->count
= profile_count::zero ();
442 FOR_EACH_EDGE (e
, ei
, current
->entry_bb
->preds
)
444 if (e
->flags
& EDGE_DFS_BACK
)
446 if (!bitmap_bit_p (current
->split_bbs
, e
->src
->index
))
447 current
->count
+= e
->count ();
450 /* Do not split when we would end up calling function anyway.
451 Compares are three state, use !(...<...) to also give up when outcome
454 < (ENTRY_BLOCK_PTR_FOR_FN (cfun
)->count
.apply_scale
455 (param_partial_inlining_entry_probability
, 100))))
457 /* When profile is guessed, we cannot expect it to give us
458 realistic estimate on likeliness of function taking the
459 complex path. As a special case, when tail of the function is
460 a loop, enable splitting since inlining code skipping the loop
461 is likely noticeable win. */
463 && profile_status_for_fn (cfun
) != PROFILE_READ
465 < ENTRY_BLOCK_PTR_FOR_FN (cfun
)->count
)
467 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
470 " Split before loop, accepting despite low counts");
471 current
->count
.dump (dump_file
);
472 fprintf (dump_file
, " ");
473 ENTRY_BLOCK_PTR_FOR_FN (cfun
)->count
.dump (dump_file
);
478 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
480 " Refused: incoming frequency is too large.\n");
485 if (!current
->header_size
)
487 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
488 fprintf (dump_file
, " Refused: header empty\n");
492 /* Verify that PHI args on entry are either virtual or all their operands
493 incoming from header are the same. */
494 for (bsi
= gsi_start_phis (current
->entry_bb
); !gsi_end_p (bsi
); gsi_next (&bsi
))
496 gphi
*stmt
= bsi
.phi ();
499 if (virtual_operand_p (gimple_phi_result (stmt
)))
501 for (i
= 0; i
< gimple_phi_num_args (stmt
); i
++)
503 edge e
= gimple_phi_arg_edge (stmt
, i
);
504 if (!bitmap_bit_p (current
->split_bbs
, e
->src
->index
))
506 tree edge_val
= gimple_phi_arg_def (stmt
, i
);
507 if (val
&& edge_val
!= val
)
509 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
511 " Refused: entry BB has PHI with multiple variants\n");
520 /* See what argument we will pass to the split function and compute
522 call_overhead
= eni_size_weights
.call_cost
;
523 for (parm
= DECL_ARGUMENTS (current_function_decl
); parm
;
524 parm
= DECL_CHAIN (parm
))
526 if (!is_gimple_reg (parm
))
528 if (bitmap_bit_p (non_ssa_vars
, DECL_UID (parm
)))
530 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
532 " Refused: need to pass non-ssa param values\n");
538 tree ddef
= ssa_default_def (cfun
, parm
);
540 && bitmap_bit_p (current
->ssa_names_to_pass
,
541 SSA_NAME_VERSION (ddef
)))
543 if (!VOID_TYPE_P (TREE_TYPE (parm
)))
544 call_overhead
+= estimate_move_cost (TREE_TYPE (parm
), false);
549 if (!VOID_TYPE_P (TREE_TYPE (TREE_TYPE (current_function_decl
))))
550 call_overhead
+= estimate_move_cost (TREE_TYPE (TREE_TYPE
551 (current_function_decl
)),
554 if (current
->split_size
<= call_overhead
)
556 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
558 " Refused: split size is smaller than call overhead\n");
561 /* FIXME: The logic here is not very precise, because inliner does use
562 inline predicates to reduce function body size. We add 10 to anticipate
563 that. Next stage1 we should try to be more meaningful here. */
564 if (current
->header_size
+ call_overhead
565 >= (unsigned int)(DECL_DECLARED_INLINE_P (current_function_decl
)
566 ? param_max_inline_insns_single
567 : param_max_inline_insns_auto
) + 10)
569 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
571 " Refused: header size is too large for inline candidate\n");
575 /* Splitting functions brings the target out of comdat group; this will
576 lead to code duplication if the function is reused by other unit.
577 Limit this duplication. This is consistent with limit in tree-sra.cc
578 FIXME: with LTO we ought to be able to do better! */
579 if (DECL_ONE_ONLY (current_function_decl
)
580 && current
->split_size
>= (unsigned int) param_max_inline_insns_auto
+ 10)
582 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
584 " Refused: function is COMDAT and tail is too large\n");
587 /* For comdat functions also reject very small tails; those will likely get
588 inlined back and we do not want to risk the duplication overhead.
589 FIXME: with LTO we ought to be able to do better! */
590 if (DECL_ONE_ONLY (current_function_decl
)
591 && current
->split_size
592 <= (unsigned int) param_early_inlining_insns
/ 2)
594 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
596 " Refused: function is COMDAT and tail is too small\n");
600 /* FIXME: we currently can pass only SSA function parameters to the split
601 arguments. Once parm_adjustment infrastructure is supported by cloning,
602 we can pass more than that. */
603 if (num_args
!= bitmap_count_bits (current
->ssa_names_to_pass
))
606 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
608 " Refused: need to pass non-param values\n");
612 /* When there are non-ssa vars used in the split region, see if they
613 are used in the header region. If so, reject the split.
614 FIXME: we can use nested function support to access both. */
615 if (!bitmap_empty_p (non_ssa_vars
)
616 && !verify_non_ssa_vars (current
, non_ssa_vars
, return_bb
))
618 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
620 " Refused: split part has non-ssa uses\n");
624 /* If the split point is dominated by a forbidden block, reject
626 if (!bitmap_empty_p (forbidden_dominators
)
627 && dominated_by_forbidden (current
->entry_bb
))
629 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
631 " Refused: split point dominated by forbidden block\n");
635 /* See if retval used by return bb is computed by header or split part.
636 When it is computed by split part, we need to produce return statement
637 in the split part and add code to header to pass it around.
639 This is bit tricky to test:
640 1) When there is no return_bb or no return value, we always pass
642 2) Invariants are always computed by caller.
643 3) For SSA we need to look if defining statement is in header or split part
644 4) For non-SSA we need to look where the var is computed. */
645 retval
= find_retval (return_bb
);
648 /* If there is a return_bb with no return value in function returning
649 value by reference, also make the split part return void, otherwise
650 we expansion would try to create a non-POD temporary, which is
652 if (return_bb
!= EXIT_BLOCK_PTR_FOR_FN (cfun
)
653 && DECL_RESULT (current_function_decl
)
654 && DECL_BY_REFERENCE (DECL_RESULT (current_function_decl
)))
655 current
->split_part_set_retval
= false;
657 current
->split_part_set_retval
= true;
659 else if (is_gimple_min_invariant (retval
))
660 current
->split_part_set_retval
= false;
661 /* Special case is value returned by reference we record as if it was non-ssa
662 set to result_decl. */
663 else if (TREE_CODE (retval
) == SSA_NAME
664 && SSA_NAME_VAR (retval
)
665 && TREE_CODE (SSA_NAME_VAR (retval
)) == RESULT_DECL
666 && DECL_BY_REFERENCE (DECL_RESULT (current_function_decl
)))
667 current
->split_part_set_retval
668 = bitmap_bit_p (non_ssa_vars
, DECL_UID (SSA_NAME_VAR (retval
)));
669 else if (TREE_CODE (retval
) == SSA_NAME
)
670 current
->split_part_set_retval
671 = split_part_set_ssa_name_p (retval
, current
, return_bb
);
672 else if (TREE_CODE (retval
) == PARM_DECL
)
673 current
->split_part_set_retval
= false;
674 else if (VAR_P (retval
)
675 || TREE_CODE (retval
) == RESULT_DECL
)
676 current
->split_part_set_retval
677 = bitmap_bit_p (non_ssa_vars
, DECL_UID (retval
));
679 current
->split_part_set_retval
= true;
681 /* split_function fixes up at most one PHI non-virtual PHI node in return_bb,
682 for the return value. If there are other PHIs, give up. */
683 if (return_bb
!= EXIT_BLOCK_PTR_FOR_FN (cfun
))
687 for (psi
= gsi_start_phis (return_bb
); !gsi_end_p (psi
); gsi_next (&psi
))
688 if (!virtual_operand_p (gimple_phi_result (psi
.phi ()))
690 && current
->split_part_set_retval
691 && TREE_CODE (retval
) == SSA_NAME
692 && !DECL_BY_REFERENCE (DECL_RESULT (current_function_decl
))
693 && SSA_NAME_DEF_STMT (retval
) == psi
.phi ()))
695 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
697 " Refused: return bb has extra PHIs\n");
702 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
703 fprintf (dump_file
, " Accepted!\n");
705 /* At the moment chose split point with lowest count and that leaves
706 out smallest size of header.
707 In future we might re-consider this heuristics. */
708 if (!best_split_point
.split_bbs
709 || best_split_point
.count
711 || (best_split_point
.count
== current
->count
712 && best_split_point
.split_size
< current
->split_size
))
715 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
716 fprintf (dump_file
, " New best split point!\n");
717 if (best_split_point
.ssa_names_to_pass
)
719 BITMAP_FREE (best_split_point
.ssa_names_to_pass
);
720 BITMAP_FREE (best_split_point
.split_bbs
);
722 best_split_point
= *current
;
723 best_split_point
.ssa_names_to_pass
= BITMAP_ALLOC (NULL
);
724 bitmap_copy (best_split_point
.ssa_names_to_pass
,
725 current
->ssa_names_to_pass
);
726 best_split_point
.split_bbs
= BITMAP_ALLOC (NULL
);
727 bitmap_copy (best_split_point
.split_bbs
, current
->split_bbs
);
731 /* Return basic block containing RETURN statement. We allow basic blocks
735 but return_bb cannot be more complex than this (except for
736 -fsanitize=thread we allow TSAN_FUNC_EXIT () internal call in there).
737 If nothing is found, return the exit block.
739 When there are multiple RETURN statement, chose one with return value,
740 since that one is more likely shared by multiple code paths.
742 Return BB is special, because for function splitting it is the only
743 basic block that is duplicated in between header and split part of the
746 TODO: We might support multiple return blocks. */
749 find_return_bb (void)
752 basic_block return_bb
= EXIT_BLOCK_PTR_FOR_FN (cfun
);
753 gimple_stmt_iterator bsi
;
754 bool found_return
= false;
755 tree retval
= NULL_TREE
;
757 if (!single_pred_p (EXIT_BLOCK_PTR_FOR_FN (cfun
)))
760 e
= single_pred_edge (EXIT_BLOCK_PTR_FOR_FN (cfun
));
761 for (bsi
= gsi_last_bb (e
->src
); !gsi_end_p (bsi
); gsi_prev (&bsi
))
763 gimple
*stmt
= gsi_stmt (bsi
);
764 if (gimple_code (stmt
) == GIMPLE_LABEL
765 || is_gimple_debug (stmt
)
766 || gimple_clobber_p (stmt
))
768 else if (gimple_code (stmt
) == GIMPLE_ASSIGN
770 && gimple_assign_single_p (stmt
)
771 && (auto_var_in_fn_p (gimple_assign_rhs1 (stmt
),
772 current_function_decl
)
773 || is_gimple_min_invariant (gimple_assign_rhs1 (stmt
)))
774 && retval
== gimple_assign_lhs (stmt
))
776 else if (greturn
*return_stmt
= dyn_cast
<greturn
*> (stmt
))
779 retval
= gimple_return_retval (return_stmt
);
781 /* For -fsanitize=thread, allow also TSAN_FUNC_EXIT () in the return
783 else if ((flag_sanitize
& SANITIZE_THREAD
)
784 && gimple_call_internal_p (stmt
, IFN_TSAN_FUNC_EXIT
))
789 if (gsi_end_p (bsi
) && found_return
)
795 /* Given return basic block RETURN_BB, see where return value is really
798 find_retval (basic_block return_bb
)
800 gimple_stmt_iterator bsi
;
801 for (bsi
= gsi_start_bb (return_bb
); !gsi_end_p (bsi
); gsi_next (&bsi
))
802 if (greturn
*return_stmt
= dyn_cast
<greturn
*> (gsi_stmt (bsi
)))
803 return gimple_return_retval (return_stmt
);
804 else if (gimple_code (gsi_stmt (bsi
)) == GIMPLE_ASSIGN
805 && !gimple_clobber_p (gsi_stmt (bsi
)))
806 return gimple_assign_rhs1 (gsi_stmt (bsi
));
810 /* Callback for walk_stmt_load_store_addr_ops. If T is non-SSA automatic
811 variable, mark it as used in bitmap passed via DATA.
812 Return true when access to T prevents splitting the function. */
815 mark_nonssa_use (gimple
*, tree t
, tree
, void *data
)
817 t
= get_base_address (t
);
819 if (!t
|| is_gimple_reg (t
))
822 /* At present we can't pass non-SSA arguments to split function.
823 FIXME: this can be relaxed by passing references to arguments. */
824 if (TREE_CODE (t
) == PARM_DECL
)
826 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
828 "Cannot split: use of non-ssa function parameter.\n");
832 if ((VAR_P (t
) && auto_var_in_fn_p (t
, current_function_decl
))
833 || TREE_CODE (t
) == RESULT_DECL
834 || (TREE_CODE (t
) == LABEL_DECL
&& FORCED_LABEL (t
)))
835 bitmap_set_bit ((bitmap
)data
, DECL_UID (t
));
837 /* For DECL_BY_REFERENCE, the return value is actually a pointer. We want
838 to pretend that the value pointed to is actual result decl. */
839 if ((TREE_CODE (t
) == MEM_REF
|| INDIRECT_REF_P (t
))
840 && TREE_CODE (TREE_OPERAND (t
, 0)) == SSA_NAME
841 && SSA_NAME_VAR (TREE_OPERAND (t
, 0))
842 && TREE_CODE (SSA_NAME_VAR (TREE_OPERAND (t
, 0))) == RESULT_DECL
843 && DECL_BY_REFERENCE (DECL_RESULT (current_function_decl
)))
845 bitmap_bit_p ((bitmap
)data
,
846 DECL_UID (DECL_RESULT (current_function_decl
)));
851 /* Compute local properties of basic block BB we collect when looking for
852 split points. We look for ssa defs and store them in SET_SSA_NAMES,
853 for ssa uses and store them in USED_SSA_NAMES and for any non-SSA automatic
854 vars stored in NON_SSA_VARS.
856 When BB has edge to RETURN_BB, collect uses in RETURN_BB too.
858 Return false when BB contains something that prevents it from being put into
862 visit_bb (basic_block bb
, basic_block return_bb
,
863 bitmap set_ssa_names
, bitmap used_ssa_names
,
868 bool can_split
= true;
870 for (gimple_stmt_iterator bsi
= gsi_start_bb (bb
); !gsi_end_p (bsi
);
873 gimple
*stmt
= gsi_stmt (bsi
);
877 if (is_gimple_debug (stmt
))
880 if (gimple_clobber_p (stmt
))
883 /* FIXME: We can split regions containing EH. We cannot however
884 split RESX, EH_DISPATCH and EH_POINTER referring to same region
885 into different partitions. This would require tracking of
886 EH regions and checking in consider_split_point if they
887 are not used elsewhere. */
888 if (gimple_code (stmt
) == GIMPLE_RESX
)
890 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
891 fprintf (dump_file
, "Cannot split: resx.\n");
894 if (gimple_code (stmt
) == GIMPLE_EH_DISPATCH
)
896 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
897 fprintf (dump_file
, "Cannot split: eh dispatch.\n");
901 /* Check calls that would prevent splitting. */
902 if (gimple_code (stmt
) == GIMPLE_CALL
)
904 if (tree decl
= gimple_call_fndecl (stmt
))
906 /* Check builtins that would prevent splitting. */
907 if (fndecl_built_in_p (decl
, BUILT_IN_NORMAL
))
908 switch (DECL_FUNCTION_CODE (decl
))
910 /* FIXME: once we will allow passing non-parm values to
911 split part, we need to be sure to handle correct
912 builtin_stack_save and builtin_stack_restore. At the
913 moment we are safe; there is no way to store
914 builtin_stack_save result in non-SSA variable since all
915 calls to those are compiler generated. */
917 case BUILT_IN_APPLY_ARGS
:
918 case BUILT_IN_VA_START
:
919 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
921 "Cannot split: builtin_apply and va_start.\n");
924 case BUILT_IN_EH_POINTER
:
925 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
927 "Cannot split: builtin_eh_pointer.\n");
934 /* Calls to functions (which have the warning or error
935 attribute on them) should not be split off into another
937 if (lookup_attribute ("warning", DECL_ATTRIBUTES (decl
))
938 || lookup_attribute ("error", DECL_ATTRIBUTES (decl
)))
940 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
942 "Cannot split: warning or error attribute.\n");
948 FOR_EACH_SSA_TREE_OPERAND (op
, stmt
, iter
, SSA_OP_DEF
)
949 bitmap_set_bit (set_ssa_names
, SSA_NAME_VERSION (op
));
950 FOR_EACH_SSA_TREE_OPERAND (op
, stmt
, iter
, SSA_OP_USE
)
951 bitmap_set_bit (used_ssa_names
, SSA_NAME_VERSION (op
));
952 can_split
&= !walk_stmt_load_store_addr_ops (stmt
, non_ssa_vars
,
957 for (gphi_iterator bsi
= gsi_start_phis (bb
); !gsi_end_p (bsi
);
960 gphi
*stmt
= bsi
.phi ();
963 if (virtual_operand_p (gimple_phi_result (stmt
)))
965 bitmap_set_bit (set_ssa_names
,
966 SSA_NAME_VERSION (gimple_phi_result (stmt
)));
967 for (i
= 0; i
< gimple_phi_num_args (stmt
); i
++)
969 tree op
= gimple_phi_arg_def (stmt
, i
);
970 if (TREE_CODE (op
) == SSA_NAME
)
971 bitmap_set_bit (used_ssa_names
, SSA_NAME_VERSION (op
));
973 can_split
&= !walk_stmt_load_store_addr_ops (stmt
, non_ssa_vars
,
978 /* Record also uses coming from PHI operand in return BB. */
979 FOR_EACH_EDGE (e
, ei
, bb
->succs
)
980 if (e
->dest
== return_bb
)
982 for (gphi_iterator bsi
= gsi_start_phis (return_bb
);
986 gphi
*stmt
= bsi
.phi ();
987 tree op
= gimple_phi_arg_def (stmt
, e
->dest_idx
);
989 if (virtual_operand_p (gimple_phi_result (stmt
)))
991 if (TREE_CODE (op
) == SSA_NAME
)
992 bitmap_set_bit (used_ssa_names
, SSA_NAME_VERSION (op
));
994 can_split
&= !mark_nonssa_use (stmt
, op
, op
, non_ssa_vars
);
1000 /* Stack entry for recursive DFS walk in find_split_point. */
1005 /* Basic block we are examining. */
1008 /* SSA names set and used by the BB and all BBs reachable
1009 from it via DFS walk. */
1010 bitmap set_ssa_names
, used_ssa_names
;
1011 bitmap non_ssa_vars
;
1013 /* All BBS visited from this BB via DFS walk. */
1016 /* Last examined edge in DFS walk. Since we walk unoriented graph,
1017 the value is up to sum of incoming and outgoing edges of BB. */
1018 unsigned int edge_num
;
1020 /* Stack entry index of earliest BB reachable from current BB
1021 or any BB visited later in DFS walk. */
1024 /* Overall time and size of all BBs reached from this BB in DFS walk. */
1028 /* When false we cannot split on this BB. */
1033 /* Find all articulations and call consider_split on them.
1034 OVERALL_TIME and OVERALL_SIZE is time and size of the function.
1036 We perform basic algorithm for finding an articulation in a graph
1037 created from CFG by considering it to be an unoriented graph.
1039 The articulation is discovered via DFS walk. We collect earliest
1040 basic block on stack that is reachable via backward edge. Articulation
1041 is any basic block such that there is no backward edge bypassing it.
1042 To reduce stack usage we maintain heap allocated stack in STACK vector.
1043 AUX pointer of BB is set to index it appears in the stack or -1 once
1044 it is visited and popped off the stack.
1046 The algorithm finds articulation after visiting the whole component
1047 reachable by it. This makes it convenient to collect information about
1048 the component used by consider_split. */
1051 find_split_points (basic_block return_bb
, sreal overall_time
, int overall_size
)
1054 vec
<stack_entry
> stack
= vNULL
;
1056 class split_point current
;
1058 current
.header_time
= overall_time
;
1059 current
.header_size
= overall_size
;
1060 current
.split_time
= 0;
1061 current
.split_size
= 0;
1062 current
.ssa_names_to_pass
= BITMAP_ALLOC (NULL
);
1064 first
.bb
= ENTRY_BLOCK_PTR_FOR_FN (cfun
);
1066 first
.overall_time
= 0;
1067 first
.overall_size
= 0;
1068 first
.earliest
= INT_MAX
;
1069 first
.set_ssa_names
= 0;
1070 first
.used_ssa_names
= 0;
1071 first
.non_ssa_vars
= 0;
1072 first
.bbs_visited
= 0;
1073 first
.can_split
= false;
1074 stack
.safe_push (first
);
1075 ENTRY_BLOCK_PTR_FOR_FN (cfun
)->aux
= (void *)(intptr_t)-1;
1077 while (!stack
.is_empty ())
1079 stack_entry
*entry
= &stack
.last ();
1081 /* We are walking an acyclic graph, so edge_num counts
1082 succ and pred edges together. However when considering
1083 articulation, we want to have processed everything reachable
1084 from articulation but nothing that reaches into it. */
1085 if (entry
->edge_num
== EDGE_COUNT (entry
->bb
->succs
)
1086 && entry
->bb
!= ENTRY_BLOCK_PTR_FOR_FN (cfun
))
1088 int pos
= stack
.length ();
1089 entry
->can_split
&= visit_bb (entry
->bb
, return_bb
,
1090 entry
->set_ssa_names
,
1091 entry
->used_ssa_names
,
1092 entry
->non_ssa_vars
);
1093 if (pos
<= entry
->earliest
&& !entry
->can_split
1094 && dump_file
&& (dump_flags
& TDF_DETAILS
))
1096 "found articulation at bb %i but cannot split\n",
1098 if (pos
<= entry
->earliest
&& entry
->can_split
)
1100 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
1101 fprintf (dump_file
, "found articulation at bb %i\n",
1103 current
.entry_bb
= entry
->bb
;
1104 current
.ssa_names_to_pass
= BITMAP_ALLOC (NULL
);
1105 bitmap_and_compl (current
.ssa_names_to_pass
,
1106 entry
->used_ssa_names
, entry
->set_ssa_names
);
1107 current
.header_time
= overall_time
- entry
->overall_time
;
1108 current
.header_size
= overall_size
- entry
->overall_size
;
1109 current
.split_time
= entry
->overall_time
;
1110 current
.split_size
= entry
->overall_size
;
1111 current
.split_bbs
= entry
->bbs_visited
;
1112 consider_split (¤t
, entry
->non_ssa_vars
, return_bb
);
1113 BITMAP_FREE (current
.ssa_names_to_pass
);
1116 /* Do actual DFS walk. */
1118 < (EDGE_COUNT (entry
->bb
->succs
)
1119 + EDGE_COUNT (entry
->bb
->preds
)))
1123 if (entry
->edge_num
< EDGE_COUNT (entry
->bb
->succs
))
1125 e
= EDGE_SUCC (entry
->bb
, entry
->edge_num
);
1130 e
= EDGE_PRED (entry
->bb
, entry
->edge_num
1131 - EDGE_COUNT (entry
->bb
->succs
));
1137 /* New BB to visit, push it to the stack. */
1138 if (dest
!= return_bb
&& dest
!= EXIT_BLOCK_PTR_FOR_FN (cfun
)
1141 stack_entry new_entry
;
1143 new_entry
.bb
= dest
;
1144 new_entry
.edge_num
= 0;
1145 new_entry
.overall_time
1146 = bb_info_vec
[dest
->index
].time
;
1147 new_entry
.overall_size
1148 = bb_info_vec
[dest
->index
].size
;
1149 new_entry
.earliest
= INT_MAX
;
1150 new_entry
.set_ssa_names
= BITMAP_ALLOC (NULL
);
1151 new_entry
.used_ssa_names
= BITMAP_ALLOC (NULL
);
1152 new_entry
.bbs_visited
= BITMAP_ALLOC (NULL
);
1153 new_entry
.non_ssa_vars
= BITMAP_ALLOC (NULL
);
1154 new_entry
.can_split
= true;
1155 bitmap_set_bit (new_entry
.bbs_visited
, dest
->index
);
1156 stack
.safe_push (new_entry
);
1157 dest
->aux
= (void *)(intptr_t)stack
.length ();
1159 /* Back edge found, record the earliest point. */
1160 else if ((intptr_t)dest
->aux
> 0
1161 && (intptr_t)dest
->aux
< entry
->earliest
)
1162 entry
->earliest
= (intptr_t)dest
->aux
;
1164 /* We are done with examining the edges. Pop off the value from stack
1165 and merge stuff we accumulate during the walk. */
1166 else if (entry
->bb
!= ENTRY_BLOCK_PTR_FOR_FN (cfun
))
1168 stack_entry
*prev
= &stack
[stack
.length () - 2];
1170 entry
->bb
->aux
= (void *)(intptr_t)-1;
1171 prev
->can_split
&= entry
->can_split
;
1172 if (prev
->set_ssa_names
)
1174 bitmap_ior_into (prev
->set_ssa_names
, entry
->set_ssa_names
);
1175 bitmap_ior_into (prev
->used_ssa_names
, entry
->used_ssa_names
);
1176 bitmap_ior_into (prev
->bbs_visited
, entry
->bbs_visited
);
1177 bitmap_ior_into (prev
->non_ssa_vars
, entry
->non_ssa_vars
);
1179 if (prev
->earliest
> entry
->earliest
)
1180 prev
->earliest
= entry
->earliest
;
1181 prev
->overall_time
+= entry
->overall_time
;
1182 prev
->overall_size
+= entry
->overall_size
;
1183 BITMAP_FREE (entry
->set_ssa_names
);
1184 BITMAP_FREE (entry
->used_ssa_names
);
1185 BITMAP_FREE (entry
->bbs_visited
);
1186 BITMAP_FREE (entry
->non_ssa_vars
);
1192 ENTRY_BLOCK_PTR_FOR_FN (cfun
)->aux
= NULL
;
1193 FOR_EACH_BB_FN (bb
, cfun
)
1196 BITMAP_FREE (current
.ssa_names_to_pass
);
1199 /* Split function at SPLIT_POINT. */
1202 split_function (basic_block return_bb
, class split_point
*split_point
,
1203 bool add_tsan_func_exit
)
1205 vec
<tree
> args_to_pass
= vNULL
;
1206 bitmap args_to_skip
;
1209 cgraph_node
*node
, *cur_node
= cgraph_node::get (current_function_decl
);
1210 basic_block call_bb
;
1211 gcall
*call
, *tsan_func_exit_call
= NULL
;
1214 tree retval
= NULL
, real_retval
= NULL
;
1215 gimple
*last_stmt
= NULL
;
1221 fprintf (dump_file
, "\n\nSplitting function at:\n");
1222 dump_split_point (dump_file
, split_point
);
1225 if (cur_node
->can_change_signature
)
1226 args_to_skip
= BITMAP_ALLOC (NULL
);
1228 args_to_skip
= NULL
;
1230 /* Collect the parameters of new function and args_to_skip bitmap. */
1231 for (parm
= DECL_ARGUMENTS (current_function_decl
);
1232 parm
; parm
= DECL_CHAIN (parm
), num
++)
1234 && (!is_gimple_reg (parm
)
1235 || (ddef
= ssa_default_def (cfun
, parm
)) == NULL_TREE
1236 || !bitmap_bit_p (split_point
->ssa_names_to_pass
,
1237 SSA_NAME_VERSION (ddef
))))
1238 bitmap_set_bit (args_to_skip
, num
);
1241 /* This parm might not have been used up to now, but is going to be
1242 used, hence register it. */
1243 if (is_gimple_reg (parm
))
1244 arg
= get_or_create_ssa_default_def (cfun
, parm
);
1248 if (!useless_type_conversion_p (DECL_ARG_TYPE (parm
), TREE_TYPE (arg
)))
1249 arg
= fold_convert (DECL_ARG_TYPE (parm
), arg
);
1250 args_to_pass
.safe_push (arg
);
1253 /* See if the split function will return. */
1254 bool split_part_return_p
= false;
1255 FOR_EACH_EDGE (e
, ei
, return_bb
->preds
)
1257 if (bitmap_bit_p (split_point
->split_bbs
, e
->src
->index
))
1258 split_part_return_p
= true;
1261 /* Add return block to what will become the split function.
1262 We do not return; no return block is needed. */
1263 if (!split_part_return_p
)
1265 /* We have no return block, so nothing is needed. */
1266 else if (return_bb
== EXIT_BLOCK_PTR_FOR_FN (cfun
))
1268 /* When we do not want to return value, we need to construct
1269 new return block with empty return statement.
1270 FIXME: Once we are able to change return type, we should change function
1271 to return void instead of just outputting function with undefined return
1272 value. For structures this affects quality of codegen. */
1273 else if ((retval
= find_retval (return_bb
))
1274 && !split_point
->split_part_set_retval
)
1276 bool redirected
= true;
1277 basic_block new_return_bb
= create_basic_block (NULL
, 0, return_bb
);
1278 gimple_stmt_iterator gsi
= gsi_start_bb (new_return_bb
);
1279 gsi_insert_after (&gsi
, gimple_build_return (NULL
), GSI_NEW_STMT
);
1280 new_return_bb
->count
= profile_count::zero ();
1284 FOR_EACH_EDGE (e
, ei
, return_bb
->preds
)
1285 if (bitmap_bit_p (split_point
->split_bbs
, e
->src
->index
))
1287 new_return_bb
->count
+= e
->count ();
1288 redirect_edge_and_branch (e
, new_return_bb
);
1293 e
= make_single_succ_edge (new_return_bb
, EXIT_BLOCK_PTR_FOR_FN (cfun
), 0);
1294 add_bb_to_loop (new_return_bb
, current_loops
->tree_root
);
1295 bitmap_set_bit (split_point
->split_bbs
, new_return_bb
->index
);
1297 /* When we pass around the value, use existing return block. */
1299 bitmap_set_bit (split_point
->split_bbs
, return_bb
->index
);
1301 /* If RETURN_BB has virtual operand PHIs, they must be removed and the
1302 virtual operand marked for renaming as we change the CFG in a way that
1303 tree-inline is not able to compensate for.
1305 Note this can happen whether or not we have a return value. If we have
1306 a return value, then RETURN_BB may have PHIs for real operands too. */
1307 if (return_bb
!= EXIT_BLOCK_PTR_FOR_FN (cfun
))
1310 for (gphi_iterator gsi
= gsi_start_phis (return_bb
);
1313 gphi
*stmt
= gsi
.phi ();
1314 if (!virtual_operand_p (gimple_phi_result (stmt
)))
1319 mark_virtual_phi_result_for_renaming (stmt
);
1320 remove_phi_node (&gsi
, true);
1323 /* In reality we have to rename the reaching definition of the
1324 virtual operand at return_bb as we will eventually release it
1325 when we remove the code region we outlined.
1326 So we have to rename all immediate virtual uses of that region
1327 if we didn't see a PHI definition yet. */
1328 /* ??? In real reality we want to set the reaching vdef of the
1329 entry of the SESE region as the vuse of the call and the reaching
1330 vdef of the exit of the SESE region as the vdef of the call. */
1332 for (gimple_stmt_iterator gsi
= gsi_start_bb (return_bb
);
1336 gimple
*stmt
= gsi_stmt (gsi
);
1337 if (gimple_vuse (stmt
))
1339 gimple_set_vuse (stmt
, NULL_TREE
);
1342 if (gimple_vdef (stmt
))
1347 ipa_param_adjustments
*adjustments
;
1348 bool skip_return
= (!split_part_return_p
1349 || !split_point
->split_part_set_retval
);
1350 /* TODO: Perhaps get rid of args_to_skip entirely, after we make sure the
1351 debug info generation and discrepancy avoiding works well too. */
1352 if ((args_to_skip
&& !bitmap_empty_p (args_to_skip
))
1355 vec
<ipa_adjusted_param
, va_gc
> *new_params
= NULL
;
1357 for (parm
= DECL_ARGUMENTS (current_function_decl
), j
= 0;
1358 parm
; parm
= DECL_CHAIN (parm
), j
++)
1359 if (!args_to_skip
|| !bitmap_bit_p (args_to_skip
, j
))
1361 ipa_adjusted_param adj
;
1362 memset (&adj
, 0, sizeof (adj
));
1363 adj
.op
= IPA_PARAM_OP_COPY
;
1365 adj
.prev_clone_index
= j
;
1366 vec_safe_push (new_params
, adj
);
1368 adjustments
= new ipa_param_adjustments (new_params
, j
, skip_return
);
1373 /* Now create the actual clone. */
1374 cgraph_edge::rebuild_edges ();
1375 node
= cur_node
->create_version_clone_with_body
1376 (vNULL
, NULL
, adjustments
,
1377 split_point
->split_bbs
, split_point
->entry_bb
, "part");
1379 node
->split_part
= true;
1381 if (cur_node
->same_comdat_group
)
1383 /* TODO: call is versionable if we make sure that all
1384 callers are inside of a comdat group. */
1385 cur_node
->calls_comdat_local
= true;
1386 node
->add_to_same_comdat_group (cur_node
);
1390 /* Let's take a time profile for splitted function. */
1391 if (cur_node
->tp_first_run
)
1392 node
->tp_first_run
= cur_node
->tp_first_run
+ 1;
1394 /* For usual cloning it is enough to clear builtin only when signature
1395 changes. For partial inlining we however cannot expect the part
1396 of builtin implementation to have same semantic as the whole. */
1397 if (fndecl_built_in_p (node
->decl
))
1398 set_decl_built_in_function (node
->decl
, NOT_BUILT_IN
, 0);
1400 /* If return_bb contains any clobbers that refer to SSA_NAMEs
1401 set in the split part, remove them. Also reset debug stmts that
1402 refer to SSA_NAMEs set in the split part. */
1403 if (return_bb
!= EXIT_BLOCK_PTR_FOR_FN (cfun
))
1405 gimple_stmt_iterator gsi
= gsi_start_bb (return_bb
);
1406 while (!gsi_end_p (gsi
))
1410 gimple
*stmt
= gsi_stmt (gsi
);
1411 bool remove
= false;
1412 if (gimple_clobber_p (stmt
) || is_gimple_debug (stmt
))
1413 FOR_EACH_SSA_TREE_OPERAND (op
, stmt
, iter
, SSA_OP_USE
)
1415 basic_block bb
= gimple_bb (SSA_NAME_DEF_STMT (op
));
1419 && bitmap_bit_p (split_point
->split_bbs
, bb
->index
))
1421 if (is_gimple_debug (stmt
))
1423 gimple_debug_bind_reset_value (stmt
);
1432 gsi_remove (&gsi
, true);
1438 /* If the original function is declared inline, there is no point in issuing
1439 a warning for the non-inlinable part. */
1440 DECL_NO_INLINE_WARNING_P (node
->decl
) = 1;
1441 cur_node
->remove_callees ();
1442 cur_node
->remove_all_references ();
1443 if (!split_part_return_p
)
1444 TREE_THIS_VOLATILE (node
->decl
) = 1;
1446 dump_function_to_file (node
->decl
, dump_file
, dump_flags
);
1448 /* Create the basic block we place call into. It is the entry basic block
1449 split after last label. */
1450 call_bb
= split_point
->entry_bb
;
1451 for (gimple_stmt_iterator gsi
= gsi_start_bb (call_bb
); !gsi_end_p (gsi
);)
1452 if (gimple_code (gsi_stmt (gsi
)) == GIMPLE_LABEL
)
1454 last_stmt
= gsi_stmt (gsi
);
1459 call_bb
->count
= split_point
->count
;
1460 e
= split_block (split_point
->entry_bb
, last_stmt
);
1463 /* Produce the call statement. */
1464 gimple_stmt_iterator gsi
= gsi_last_bb (call_bb
);
1465 FOR_EACH_VEC_ELT (args_to_pass
, i
, arg
)
1466 if (!is_gimple_val (arg
))
1468 arg
= force_gimple_operand_gsi (&gsi
, arg
, true, NULL_TREE
,
1469 false, GSI_CONTINUE_LINKING
);
1470 args_to_pass
[i
] = arg
;
1472 call
= gimple_build_call_vec (node
->decl
, args_to_pass
);
1473 gimple_set_block (call
, DECL_INITIAL (current_function_decl
));
1474 args_to_pass
.release ();
1476 /* For optimized away parameters, add on the caller side
1478 DEBUG D#X => parm_Y(D)
1479 stmts and associate D#X with parm in decl_debug_args_lookup
1480 vector to say for debug info that if parameter parm had been passed,
1481 it would have value parm_Y(D). */
1484 vec
<tree
, va_gc
> **debug_args
= NULL
;
1485 unsigned i
= 0, len
= 0;
1486 if (MAY_HAVE_DEBUG_BIND_STMTS
)
1488 debug_args
= decl_debug_args_lookup (node
->decl
);
1490 len
= vec_safe_length (*debug_args
);
1492 for (parm
= DECL_ARGUMENTS (current_function_decl
), num
= 0;
1493 parm
; parm
= DECL_CHAIN (parm
), num
++)
1494 if (bitmap_bit_p (args_to_skip
, num
) && is_gimple_reg (parm
))
1499 /* This needs to be done even without
1500 MAY_HAVE_DEBUG_BIND_STMTS, otherwise if it didn't exist
1501 before, we'd end up with different SSA_NAME_VERSIONs
1502 between -g and -g0. */
1503 arg
= get_or_create_ssa_default_def (cfun
, parm
);
1504 if (!MAY_HAVE_DEBUG_BIND_STMTS
|| debug_args
== NULL
)
1507 while (i
< len
&& (**debug_args
)[i
] != DECL_ORIGIN (parm
))
1511 ddecl
= (**debug_args
)[i
+ 1];
1513 = gimple_build_debug_bind (ddecl
, unshare_expr (arg
), call
);
1514 gsi_insert_after (&gsi
, def_temp
, GSI_NEW_STMT
);
1516 BITMAP_FREE (args_to_skip
);
1519 /* We avoid address being taken on any variable used by split part,
1520 so return slot optimization is always possible. Moreover this is
1521 required to make DECL_BY_REFERENCE work. */
1522 if (aggregate_value_p (DECL_RESULT (current_function_decl
),
1523 TREE_TYPE (current_function_decl
))
1524 && (!is_gimple_reg_type (TREE_TYPE (DECL_RESULT (current_function_decl
)))
1525 || DECL_BY_REFERENCE (DECL_RESULT (current_function_decl
))))
1526 gimple_call_set_return_slot_opt (call
, true);
1528 if (add_tsan_func_exit
)
1529 tsan_func_exit_call
= gimple_build_call_internal (IFN_TSAN_FUNC_EXIT
, 0);
1531 /* Update return value. This is bit tricky. When we do not return,
1532 do nothing. When we return we might need to update return_bb
1533 or produce a new return statement. */
1534 if (!split_part_return_p
)
1536 gsi_insert_after (&gsi
, call
, GSI_NEW_STMT
);
1537 if (tsan_func_exit_call
)
1538 gsi_insert_after (&gsi
, tsan_func_exit_call
, GSI_NEW_STMT
);
1542 e
= make_single_succ_edge (call_bb
, return_bb
,
1543 return_bb
== EXIT_BLOCK_PTR_FOR_FN (cfun
)
1544 ? 0 : EDGE_FALLTHRU
);
1546 /* If there is return basic block, see what value we need to store
1547 return value into and put call just before it. */
1548 if (return_bb
!= EXIT_BLOCK_PTR_FOR_FN (cfun
))
1550 real_retval
= retval
;
1551 if (real_retval
&& split_point
->split_part_set_retval
)
1555 /* See if we need new SSA_NAME for the result.
1556 When DECL_BY_REFERENCE is true, retval is actually pointer to
1557 return value and it is constant in whole function. */
1558 if (TREE_CODE (retval
) == SSA_NAME
1559 && !DECL_BY_REFERENCE (DECL_RESULT (current_function_decl
)))
1561 retval
= copy_ssa_name (retval
, call
);
1563 /* See if there is PHI defining return value. */
1564 for (psi
= gsi_start_phis (return_bb
);
1565 !gsi_end_p (psi
); gsi_next (&psi
))
1566 if (!virtual_operand_p (gimple_phi_result (psi
.phi ())))
1569 /* When there is PHI, just update its value. */
1570 if (TREE_CODE (retval
) == SSA_NAME
1571 && !gsi_end_p (psi
))
1572 add_phi_arg (psi
.phi (), retval
, e
, UNKNOWN_LOCATION
);
1573 /* Otherwise update the return BB itself.
1574 find_return_bb allows at most one assignment to return value,
1575 so update first statement. */
1578 gimple_stmt_iterator bsi
;
1579 for (bsi
= gsi_start_bb (return_bb
); !gsi_end_p (bsi
);
1581 if (greturn
*return_stmt
1582 = dyn_cast
<greturn
*> (gsi_stmt (bsi
)))
1584 gimple_return_set_retval (return_stmt
, retval
);
1587 else if (gimple_code (gsi_stmt (bsi
)) == GIMPLE_ASSIGN
1588 && !gimple_clobber_p (gsi_stmt (bsi
)))
1590 gimple_assign_set_rhs1 (gsi_stmt (bsi
), retval
);
1593 update_stmt (gsi_stmt (bsi
));
1594 /* Also adjust clobbers and debug stmts in return_bb. */
1595 for (bsi
= gsi_start_bb (return_bb
); !gsi_end_p (bsi
);
1598 gimple
*stmt
= gsi_stmt (bsi
);
1599 if (gimple_clobber_p (stmt
)
1600 || is_gimple_debug (stmt
))
1603 use_operand_p use_p
;
1604 bool update
= false;
1605 FOR_EACH_SSA_USE_OPERAND (use_p
, stmt
, iter
,
1607 if (USE_FROM_PTR (use_p
) == real_retval
)
1609 SET_USE (use_p
, retval
);
1618 if (DECL_BY_REFERENCE (DECL_RESULT (current_function_decl
)))
1620 gimple_call_set_lhs (call
, build_simple_mem_ref (retval
));
1621 gsi_insert_after (&gsi
, call
, GSI_NEW_STMT
);
1626 restype
= TREE_TYPE (DECL_RESULT (current_function_decl
));
1627 gsi_insert_after (&gsi
, call
, GSI_NEW_STMT
);
1628 if (!useless_type_conversion_p (TREE_TYPE (retval
), restype
))
1631 tree tem
= create_tmp_reg (restype
);
1632 tem
= make_ssa_name (tem
, call
);
1633 cpy
= gimple_build_assign (retval
, NOP_EXPR
, tem
);
1634 gsi_insert_after (&gsi
, cpy
, GSI_NEW_STMT
);
1637 gimple_call_set_lhs (call
, retval
);
1642 gsi_insert_after (&gsi
, call
, GSI_NEW_STMT
);
1643 if (tsan_func_exit_call
)
1644 gsi_insert_after (&gsi
, tsan_func_exit_call
, GSI_NEW_STMT
);
1646 /* We don't use return block (there is either no return in function or
1647 multiple of them). So create new basic block with return statement.
1652 if (split_point
->split_part_set_retval
1653 && !VOID_TYPE_P (TREE_TYPE (TREE_TYPE (current_function_decl
))))
1655 retval
= DECL_RESULT (current_function_decl
);
1657 /* We use temporary register to hold value when aggregate_value_p
1658 is false. Similarly for DECL_BY_REFERENCE we must avoid extra
1660 if (!aggregate_value_p (retval
, TREE_TYPE (current_function_decl
))
1661 && !DECL_BY_REFERENCE (retval
))
1662 retval
= create_tmp_reg (TREE_TYPE (retval
));
1663 if (is_gimple_reg (retval
))
1665 /* When returning by reference, there is only one SSA name
1666 assigned to RESULT_DECL (that is pointer to return value).
1667 Look it up or create new one if it is missing. */
1668 if (DECL_BY_REFERENCE (retval
))
1669 retval
= get_or_create_ssa_default_def (cfun
, retval
);
1670 /* Otherwise produce new SSA name for return value. */
1672 retval
= make_ssa_name (retval
, call
);
1674 if (DECL_BY_REFERENCE (DECL_RESULT (current_function_decl
)))
1675 gimple_call_set_lhs (call
, build_simple_mem_ref (retval
));
1677 gimple_call_set_lhs (call
, retval
);
1678 gsi_insert_after (&gsi
, call
, GSI_NEW_STMT
);
1682 gsi_insert_after (&gsi
, call
, GSI_NEW_STMT
);
1684 && is_gimple_reg_type (TREE_TYPE (retval
))
1685 && !is_gimple_val (retval
))
1688 = gimple_build_assign (make_ssa_name (TREE_TYPE (retval
)),
1690 retval
= gimple_assign_lhs (g
);
1691 gsi_insert_after (&gsi
, g
, GSI_NEW_STMT
);
1694 if (tsan_func_exit_call
)
1695 gsi_insert_after (&gsi
, tsan_func_exit_call
, GSI_NEW_STMT
);
1696 ret
= gimple_build_return (retval
);
1697 gsi_insert_after (&gsi
, ret
, GSI_NEW_STMT
);
1700 free_dominance_info (CDI_DOMINATORS
);
1701 free_dominance_info (CDI_POST_DOMINATORS
);
1702 compute_fn_summary (node
, true);
1705 /* Execute function splitting pass. */
1708 execute_split_functions (void)
1710 gimple_stmt_iterator bsi
;
1712 sreal overall_time
= 0;
1713 int overall_size
= 0;
1715 struct cgraph_node
*node
= cgraph_node::get (current_function_decl
);
1717 if (flags_from_decl_or_type (current_function_decl
)
1718 & (ECF_NORETURN
|ECF_MALLOC
))
1721 fprintf (dump_file
, "Not splitting: noreturn/malloc function.\n");
1724 if (MAIN_NAME_P (DECL_NAME (current_function_decl
)))
1727 fprintf (dump_file
, "Not splitting: main function.\n");
1730 if (node
->frequency
== NODE_FREQUENCY_UNLIKELY_EXECUTED
)
1733 fprintf (dump_file
, "Not splitting: function is unlikely executed.\n");
1736 /* This can be relaxed; function might become inlinable after splitting
1737 away the uninlinable part. */
1738 if (ipa_fn_summaries
1739 && ipa_fn_summaries
->get (node
)
1740 && !ipa_fn_summaries
->get (node
)->inlinable
)
1743 fprintf (dump_file
, "Not splitting: not inlinable.\n");
1746 if (DECL_DISREGARD_INLINE_LIMITS (node
->decl
))
1749 fprintf (dump_file
, "Not splitting: disregarding inline limits.\n");
1752 /* This can be relaxed; most of versioning tests actually prevents
1754 if (!tree_versionable_function_p (current_function_decl
))
1757 fprintf (dump_file
, "Not splitting: not versionable.\n");
1760 /* FIXME: we could support this. */
1761 if (DECL_STRUCT_FUNCTION (current_function_decl
)->static_chain_decl
)
1764 fprintf (dump_file
, "Not splitting: nested function.\n");
1768 /* See if it makes sense to try to split.
1769 It makes sense to split if we inline, that is if we have direct calls to
1770 handle or direct calls are possibly going to appear as result of indirect
1771 inlining or LTO. Also handle -fprofile-generate as LTO to allow non-LTO
1772 training for LTO -fprofile-use build.
1774 Note that we are not completely conservative about disqualifying functions
1775 called once. It is possible that the caller is called more then once and
1776 then inlining would still benefit. */
1778 /* Local functions called once will be completely inlined most of time. */
1779 || (!node
->callers
->next_caller
&& node
->local
))
1780 && !node
->address_taken
1781 && !node
->has_aliases_p ()
1782 && (!flag_lto
|| !node
->externally_visible
))
1785 fprintf (dump_file
, "Not splitting: not called directly "
1786 "or called once.\n");
1790 /* FIXME: We can actually split if splitting reduces call overhead. */
1791 if (!flag_inline_small_functions
1792 && !DECL_DECLARED_INLINE_P (current_function_decl
))
1795 fprintf (dump_file
, "Not splitting: not autoinlining and function"
1796 " is not inline.\n");
1800 if (lookup_attribute ("noinline", DECL_ATTRIBUTES (current_function_decl
)))
1803 fprintf (dump_file
, "Not splitting: function is noinline.\n");
1806 if (lookup_attribute ("section", DECL_ATTRIBUTES (current_function_decl
)))
1809 fprintf (dump_file
, "Not splitting: function is in user defined "
1814 /* We enforce splitting after loop headers when profile info is not
1816 if (profile_status_for_fn (cfun
) != PROFILE_READ
)
1817 mark_dfs_back_edges ();
1819 /* Initialize bitmap to track forbidden calls. */
1820 forbidden_dominators
= BITMAP_ALLOC (NULL
);
1821 calculate_dominance_info (CDI_DOMINATORS
);
1823 /* Compute local info about basic blocks and determine function size/time. */
1824 bb_info_vec
.safe_grow_cleared (last_basic_block_for_fn (cfun
) + 1, true);
1825 best_split_point
.split_bbs
= NULL
;
1826 basic_block return_bb
= find_return_bb ();
1827 int tsan_exit_found
= -1;
1828 FOR_EACH_BB_FN (bb
, cfun
)
1832 sreal freq
= bb
->count
.to_sreal_scale
1833 (ENTRY_BLOCK_PTR_FOR_FN (cfun
)->count
);
1835 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
1836 fprintf (dump_file
, "Basic block %i\n", bb
->index
);
1838 for (bsi
= gsi_start_bb (bb
); !gsi_end_p (bsi
); gsi_next (&bsi
))
1842 gimple
*stmt
= gsi_stmt (bsi
);
1844 this_size
= estimate_num_insns (stmt
, &eni_size_weights
);
1845 this_time
= (sreal
)estimate_num_insns (stmt
, &eni_time_weights
)
1849 check_forbidden_calls (stmt
);
1851 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
1853 fprintf (dump_file
, " freq:%4.2f size:%3i time:%4.2f ",
1854 freq
.to_double (), this_size
, this_time
.to_double ());
1855 print_gimple_stmt (dump_file
, stmt
, 0);
1858 if ((flag_sanitize
& SANITIZE_THREAD
)
1859 && gimple_call_internal_p (stmt
, IFN_TSAN_FUNC_EXIT
))
1861 /* We handle TSAN_FUNC_EXIT for splitting either in the
1862 return_bb, or in its immediate predecessors. */
1863 if ((bb
!= return_bb
&& !find_edge (bb
, return_bb
))
1864 || (tsan_exit_found
!= -1
1865 && tsan_exit_found
!= (bb
!= return_bb
)))
1868 fprintf (dump_file
, "Not splitting: TSAN_FUNC_EXIT"
1869 " in unexpected basic block.\n");
1870 BITMAP_FREE (forbidden_dominators
);
1871 bb_info_vec
.release ();
1874 tsan_exit_found
= bb
!= return_bb
;
1877 overall_time
+= time
;
1878 overall_size
+= size
;
1879 bb_info_vec
[bb
->index
].time
= time
;
1880 bb_info_vec
[bb
->index
].size
= size
;
1882 find_split_points (return_bb
, overall_time
, overall_size
);
1883 if (best_split_point
.split_bbs
)
1885 split_function (return_bb
, &best_split_point
, tsan_exit_found
== 1);
1886 BITMAP_FREE (best_split_point
.ssa_names_to_pass
);
1887 BITMAP_FREE (best_split_point
.split_bbs
);
1888 todo
= TODO_update_ssa
| TODO_cleanup_cfg
;
1890 BITMAP_FREE (forbidden_dominators
);
1891 bb_info_vec
.release ();
1897 const pass_data pass_data_split_functions
=
1899 GIMPLE_PASS
, /* type */
1900 "fnsplit", /* name */
1901 OPTGROUP_NONE
, /* optinfo_flags */
1902 TV_IPA_FNSPLIT
, /* tv_id */
1903 PROP_cfg
, /* properties_required */
1904 0, /* properties_provided */
1905 0, /* properties_destroyed */
1906 0, /* todo_flags_start */
1907 0, /* todo_flags_finish */
1910 class pass_split_functions
: public gimple_opt_pass
1913 pass_split_functions (gcc::context
*ctxt
)
1914 : gimple_opt_pass (pass_data_split_functions
, ctxt
)
1917 /* opt_pass methods: */
1918 virtual bool gate (function
*);
1919 virtual unsigned int execute (function
*)
1921 return execute_split_functions ();
1924 }; // class pass_split_functions
1927 pass_split_functions::gate (function
*)
1929 /* When doing profile feedback, we want to execute the pass after profiling
1930 is read. So disable one in early optimization. */
1931 return (flag_partial_inlining
1932 && !profile_arc_flag
&& !flag_branch_probabilities
);
1938 make_pass_split_functions (gcc::context
*ctxt
)
1940 return new pass_split_functions (ctxt
);
1943 /* Execute function splitting pass. */
1946 execute_feedback_split_functions (void)
1948 unsigned int retval
= execute_split_functions ();
1950 retval
|= TODO_rebuild_cgraph_edges
;
1956 const pass_data pass_data_feedback_split_functions
=
1958 GIMPLE_PASS
, /* type */
1959 "feedback_fnsplit", /* name */
1960 OPTGROUP_NONE
, /* optinfo_flags */
1961 TV_IPA_FNSPLIT
, /* tv_id */
1962 PROP_cfg
, /* properties_required */
1963 0, /* properties_provided */
1964 0, /* properties_destroyed */
1965 0, /* todo_flags_start */
1966 0, /* todo_flags_finish */
1969 class pass_feedback_split_functions
: public gimple_opt_pass
1972 pass_feedback_split_functions (gcc::context
*ctxt
)
1973 : gimple_opt_pass (pass_data_feedback_split_functions
, ctxt
)
1976 /* opt_pass methods: */
1977 virtual bool gate (function
*);
1978 virtual unsigned int execute (function
*)
1980 return execute_feedback_split_functions ();
1983 }; // class pass_feedback_split_functions
1986 pass_feedback_split_functions::gate (function
*)
1988 /* We don't need to split when profiling at all, we are producing
1989 lousy code anyway. */
1990 return (flag_partial_inlining
1991 && flag_branch_probabilities
);
1997 make_pass_feedback_split_functions (gcc::context
*ctxt
)
1999 return new pass_feedback_split_functions (ctxt
);