1 /* Function splitting pass
3 Free Software Foundation, Inc.
4 Contributed by Jan Hubicka <jh@suse.cz>
6 This file is part of GCC.
8 GCC is free software; you can redistribute it and/or modify it under
9 the terms of the GNU General Public License as published by the Free
10 Software Foundation; either version 3, or (at your option) any later
13 GCC is distributed in the hope that it will be useful, but WITHOUT ANY
14 WARRANTY; without even the implied warranty of MERCHANTABILITY or
15 FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
18 You should have received a copy of the GNU General Public License
19 along with GCC; see the file COPYING3. If not see
20 <http://www.gnu.org/licenses/>. */
22 /* The purpose of this pass is to split function bodies to improve
23 inlining. I.e. for function of the form:
48 When func becomes inlinable and when cheap_test is often true, inlining func,
49 but not fund.part leads to performance imrovement similar as inlining
50 original func while the code size growth is smaller.
52 The pass is organized in three stages:
53 1) Collect local info about basic block into BB_INFO structure and
54 compute function body estimated size and time.
55 2) Via DFS walk find all possible basic blocks where we can split
57 3) If split point is found, split at the specified BB by creating a clone
58 and updating function to call it.
60 The decisions what functions to split are in execute_split_functions
63 There are several possible future improvements for this pass including:
65 1) Splitting to break up large functions
66 2) Splitting to reduce stack frame usage
67 3) Allow split part of function to use values computed in the header part.
68 The values needs to be passed to split function, perhaps via same
69 interface as for nested functions or as argument.
70 4) Support for simple rematerialization. I.e. when split part use
71 value computed in header from function parameter in very cheap way, we
72 can just recompute it.
73 5) Support splitting of nested functions.
74 6) Support non-SSA arguments.
75 7) There is nothing preventing us from producing multiple parts of single function
76 when needed or splitting also the parts. */
80 #include "coretypes.h"
85 #include "tree-flow.h"
86 #include "tree-pass.h"
89 #include "diagnostic.h"
90 #include "tree-dump.h"
91 #include "tree-inline.h"
94 #include "gimple-pretty-print.h"
96 /* Per basic block info. */
104 DEF_VEC_ALLOC_O(bb_info,heap);
106 static VEC(bb_info, heap) *bb_info_vec;
108 /* Description of split point. */
112 /* Size of the partitions. */
113 unsigned int header_time, header_size, split_time, split_size;
115 /* SSA names that need to be passed into spit funciton. */
116 bitmap ssa_names_to_pass;
118 /* Basic block where we split (that will become entry point of new function. */
119 basic_block entry_bb;
121 /* Basic blocks we are splitting away. */
124 /* True when return value is computed on split part and thus it needs
126 bool split_part_set_retval;
129 /* Best split point found. */
131 struct split_point best_split_point;
133 static tree find_retval (basic_block return_bb);
135 /* Callback for walk_stmt_load_store_addr_ops. If T is non-ssa automatic
136 variable, check it if it is present in bitmap passed via DATA. */
139 test_nonssa_use (gimple stmt ATTRIBUTE_UNUSED, tree t,
140 void *data ATTRIBUTE_UNUSED)
142 t = get_base_address (t);
144 if (t && !is_gimple_reg (t)
145 && ((TREE_CODE (t) == VAR_DECL
146 && auto_var_in_fn_p (t, current_function_decl))
147 || (TREE_CODE (t) == RESULT_DECL)
148 || (TREE_CODE (t) == PARM_DECL)))
149 return bitmap_bit_p ((bitmap)data, DECL_UID (t));
151 /* For DECL_BY_REFERENCE, the return value is actually pointer. We want to pretend
152 that the value pointed to is actual result decl. */
153 if (t && (TREE_CODE (t) == MEM_REF || INDIRECT_REF_P (t))
154 && TREE_CODE (TREE_OPERAND (t, 0)) == SSA_NAME
155 && TREE_CODE (SSA_NAME_VAR (TREE_OPERAND (t, 0))) == RESULT_DECL
156 && DECL_BY_REFERENCE (DECL_RESULT (current_function_decl)))
157 return bitmap_bit_p ((bitmap)data, DECL_UID (DECL_RESULT (current_function_decl)));
161 /* Dump split point CURRENT. */
164 dump_split_point (FILE * file, struct split_point *current)
167 "Split point at BB %i header time:%i header size: %i"
168 " split time: %i split size: %i\n bbs: ",
169 current->entry_bb->index, current->header_time,
170 current->header_size, current->split_time, current->split_size);
171 dump_bitmap (file, current->split_bbs);
172 fprintf (file, " SSA names to pass: ");
173 dump_bitmap (file, current->ssa_names_to_pass);
176 /* Look for all BBs in header that might lead to split part and verify that
177 they are not defining any of SSA vars used by split part.
178 Parameters are the same as for consider_split. */
181 verify_non_ssa_vars (struct split_point *current, bitmap non_ssa_vars,
182 basic_block return_bb)
184 bitmap seen = BITMAP_ALLOC (NULL);
185 VEC (basic_block,heap) *worklist = NULL;
190 FOR_EACH_EDGE (e, ei, current->entry_bb->preds)
191 if (e->src != ENTRY_BLOCK_PTR
192 && !bitmap_bit_p (current->split_bbs, e->src->index))
194 VEC_safe_push (basic_block, heap, worklist, e->src);
195 bitmap_set_bit (seen, e->src->index);
198 while (!VEC_empty (basic_block, worklist))
200 gimple_stmt_iterator bsi;
201 basic_block bb = VEC_pop (basic_block, worklist);
203 FOR_EACH_EDGE (e, ei, bb->preds)
204 if (e->src != ENTRY_BLOCK_PTR
205 && !bitmap_bit_p (seen, e->src->index))
207 gcc_checking_assert (!bitmap_bit_p (current->split_bbs,
209 VEC_safe_push (basic_block, heap, worklist, e->src);
210 bitmap_set_bit (seen, e->src->index);
212 for (bsi = gsi_start_bb (bb); !gsi_end_p (bsi); gsi_next (&bsi))
214 if (is_gimple_debug (gsi_stmt (bsi)))
216 if (walk_stmt_load_store_addr_ops
217 (gsi_stmt (bsi), non_ssa_vars, test_nonssa_use,
218 test_nonssa_use, test_nonssa_use))
224 for (bsi = gsi_start_phis (bb); !gsi_end_p (bsi); gsi_next (&bsi))
226 if (walk_stmt_load_store_addr_ops
227 (gsi_stmt (bsi), non_ssa_vars, test_nonssa_use,
228 test_nonssa_use, test_nonssa_use))
234 FOR_EACH_EDGE (e, ei, bb->succs)
236 if (e->dest != return_bb)
238 for (bsi = gsi_start_phis (return_bb); !gsi_end_p (bsi);
241 gimple stmt = gsi_stmt (bsi);
242 tree op = gimple_phi_arg_def (stmt, e->dest_idx);
244 if (!is_gimple_reg (gimple_phi_result (stmt)))
246 if (TREE_CODE (op) != SSA_NAME
247 && test_nonssa_use (stmt, op, non_ssa_vars))
257 VEC_free (basic_block, heap, worklist);
261 /* We found an split_point CURRENT. NON_SSA_VARS is bitmap of all non ssa
262 variables used and RETURN_BB is return basic block.
263 See if we can split function here. */
266 consider_split (struct split_point *current, bitmap non_ssa_vars,
267 basic_block return_bb)
270 unsigned int num_args = 0;
271 unsigned int call_overhead;
274 gimple_stmt_iterator bsi;
276 int incomming_freq = 0;
279 if (dump_file && (dump_flags & TDF_DETAILS))
280 dump_split_point (dump_file, current);
282 FOR_EACH_EDGE (e, ei, current->entry_bb->preds)
283 if (!bitmap_bit_p (current->split_bbs, e->src->index))
284 incomming_freq += EDGE_FREQUENCY (e);
286 /* Do not split when we would end up calling function anyway. */
288 >= (ENTRY_BLOCK_PTR->frequency
289 * PARAM_VALUE (PARAM_PARTIAL_INLINING_ENTRY_PROBABILITY) / 100))
291 if (dump_file && (dump_flags & TDF_DETAILS))
293 " Refused: incomming frequency is too large.\n");
297 if (!current->header_size)
299 if (dump_file && (dump_flags & TDF_DETAILS))
300 fprintf (dump_file, " Refused: header empty\n");
305 /* Verify that PHI args on entry are either virutal or all their operands
306 incomming from header are the same. */
307 for (bsi = gsi_start_phis (current->entry_bb); !gsi_end_p (bsi); gsi_next (&bsi))
309 gimple stmt = gsi_stmt (bsi);
312 if (!is_gimple_reg (gimple_phi_result (stmt)))
314 for (i = 0; i < gimple_phi_num_args (stmt); i++)
316 edge e = gimple_phi_arg_edge (stmt, i);
317 if (!bitmap_bit_p (current->split_bbs, e->src->index))
319 tree edge_val = gimple_phi_arg_def (stmt, i);
320 if (val && edge_val != val)
322 if (dump_file && (dump_flags & TDF_DETAILS))
324 " Refused: entry BB has PHI with multiple variants\n");
333 /* See what argument we will pass to the split function and compute
335 call_overhead = eni_size_weights.call_cost;
336 for (parm = DECL_ARGUMENTS (current_function_decl); parm;
337 parm = DECL_CHAIN (parm))
339 if (!is_gimple_reg (parm))
341 if (bitmap_bit_p (non_ssa_vars, DECL_UID (parm)))
343 if (dump_file && (dump_flags & TDF_DETAILS))
345 " Refused: need to pass non-ssa param values\n");
349 else if (gimple_default_def (cfun, parm)
350 && bitmap_bit_p (current->ssa_names_to_pass,
351 SSA_NAME_VERSION (gimple_default_def
354 if (!VOID_TYPE_P (TREE_TYPE (parm)))
355 call_overhead += estimate_move_cost (TREE_TYPE (parm));
359 if (!VOID_TYPE_P (TREE_TYPE (current_function_decl)))
360 call_overhead += estimate_move_cost (TREE_TYPE (current_function_decl));
362 if (current->split_size <= call_overhead)
364 if (dump_file && (dump_flags & TDF_DETAILS))
366 " Refused: split size is smaller than call overhead\n");
369 if (current->header_size + call_overhead
370 >= (unsigned int)(DECL_DECLARED_INLINE_P (current_function_decl)
371 ? MAX_INLINE_INSNS_SINGLE
372 : MAX_INLINE_INSNS_AUTO))
374 if (dump_file && (dump_flags & TDF_DETAILS))
376 " Refused: header size is too large for inline candidate\n");
380 /* FIXME: we currently can pass only SSA function parameters to the split
381 arguments. Once parm_adjustment infrastructure is supported by cloning,
382 we can pass more than that. */
383 if (num_args != bitmap_count_bits (current->ssa_names_to_pass))
386 if (dump_file && (dump_flags & TDF_DETAILS))
388 " Refused: need to pass non-param values\n");
392 /* When there are non-ssa vars used in the split region, see if they
393 are used in the header region. If so, reject the split.
394 FIXME: we can use nested function support to access both. */
395 if (!bitmap_empty_p (non_ssa_vars)
396 && !verify_non_ssa_vars (current, non_ssa_vars, return_bb))
398 if (dump_file && (dump_flags & TDF_DETAILS))
400 " Refused: split part has non-ssa uses\n");
403 if (dump_file && (dump_flags & TDF_DETAILS))
404 fprintf (dump_file, " Accepted!\n");
406 /* See if retval used by return bb is computed by header or split part.
407 When it is computed by split part, we need to produce return statement
408 in the split part and add code to header to pass it around.
410 This is bit tricky to test:
411 1) When there is no return_bb or no return value, we always pass
413 2) Invariants are always computed by caller.
414 3) For SSA we need to look if defining statement is in header or split part
415 4) For non-SSA we need to look where the var is computed. */
416 retval = find_retval (return_bb);
418 current->split_part_set_retval = true;
419 else if (is_gimple_min_invariant (retval))
420 current->split_part_set_retval = false;
421 /* Special case is value returned by reference we record as if it was non-ssa
422 set to result_decl. */
423 else if (TREE_CODE (retval) == SSA_NAME
424 && TREE_CODE (SSA_NAME_VAR (retval)) == RESULT_DECL
425 && DECL_BY_REFERENCE (DECL_RESULT (current_function_decl)))
426 current->split_part_set_retval
427 = bitmap_bit_p (non_ssa_vars, DECL_UID (SSA_NAME_VAR (retval)));
428 else if (TREE_CODE (retval) == SSA_NAME)
429 current->split_part_set_retval
430 = (!SSA_NAME_IS_DEFAULT_DEF (retval)
431 && (bitmap_bit_p (current->split_bbs,
432 gimple_bb (SSA_NAME_DEF_STMT (retval))->index)
433 || gimple_bb (SSA_NAME_DEF_STMT (retval)) == return_bb));
434 else if (TREE_CODE (retval) == PARM_DECL)
435 current->split_part_set_retval = false;
436 else if (TREE_CODE (retval) == VAR_DECL
437 || TREE_CODE (retval) == RESULT_DECL)
438 current->split_part_set_retval
439 = bitmap_bit_p (non_ssa_vars, DECL_UID (retval));
441 current->split_part_set_retval = true;
443 /* At the moment chose split point with lowest frequency and that leaves
444 out smallest size of header.
445 In future we might re-consider this heuristics. */
446 if (!best_split_point.split_bbs
447 || best_split_point.entry_bb->frequency > current->entry_bb->frequency
448 || (best_split_point.entry_bb->frequency == current->entry_bb->frequency
449 && best_split_point.split_size < current->split_size))
452 if (dump_file && (dump_flags & TDF_DETAILS))
453 fprintf (dump_file, " New best split point!\n");
454 if (best_split_point.ssa_names_to_pass)
456 BITMAP_FREE (best_split_point.ssa_names_to_pass);
457 BITMAP_FREE (best_split_point.split_bbs);
459 best_split_point = *current;
460 best_split_point.ssa_names_to_pass = BITMAP_ALLOC (NULL);
461 bitmap_copy (best_split_point.ssa_names_to_pass,
462 current->ssa_names_to_pass);
463 best_split_point.split_bbs = BITMAP_ALLOC (NULL);
464 bitmap_copy (best_split_point.split_bbs, current->split_bbs);
468 /* Return basic block containing RETURN statement. We allow basic blocks
472 but return_bb can not be more complex than this.
473 If nothing is found, return EXIT_BLOCK_PTR.
475 When there are multiple RETURN statement, chose one with return value,
476 since that one is more likely shared by multiple code paths.
478 Return BB is special, because for function splitting it is the only
479 basic block that is duplicated in between header and split part of the
482 TODO: We might support multiple return blocks. */
485 find_return_bb (void)
489 basic_block return_bb = EXIT_BLOCK_PTR;
491 if (EDGE_COUNT (EXIT_BLOCK_PTR->preds) == 1)
492 FOR_EACH_EDGE (e, ei, EXIT_BLOCK_PTR->preds)
494 gimple_stmt_iterator bsi;
495 bool found_return = false;
496 tree retval = NULL_TREE;
498 for (bsi = gsi_last_bb (e->src); !gsi_end_p (bsi); gsi_prev (&bsi))
500 gimple stmt = gsi_stmt (bsi);
501 if (gimple_code (stmt) == GIMPLE_LABEL
502 || is_gimple_debug (stmt))
504 else if (gimple_code (stmt) == GIMPLE_ASSIGN
506 && gimple_assign_single_p (stmt)
507 && (auto_var_in_fn_p (gimple_assign_rhs1 (stmt),
508 current_function_decl)
509 || is_gimple_min_invariant
510 (gimple_assign_rhs1 (stmt)))
511 && retval == gimple_assign_lhs (stmt))
513 else if (gimple_code (stmt) == GIMPLE_RETURN)
516 retval = gimple_return_retval (stmt);
521 if (gsi_end_p (bsi) && found_return)
532 /* Given return basicblock RETURN_BB, see where return value is really
535 find_retval (basic_block return_bb)
537 gimple_stmt_iterator bsi;
538 for (bsi = gsi_start_bb (return_bb); !gsi_end_p (bsi); gsi_next (&bsi))
539 if (gimple_code (gsi_stmt (bsi)) == GIMPLE_RETURN)
540 return gimple_return_retval (gsi_stmt (bsi));
541 else if (gimple_code (gsi_stmt (bsi)) == GIMPLE_ASSIGN)
542 return gimple_assign_rhs1 (gsi_stmt (bsi));
546 /* Callback for walk_stmt_load_store_addr_ops. If T is non-ssa automatic
547 variable, mark it as used in bitmap passed via DATA.
548 Return true when access to T prevents splitting the function. */
551 mark_nonssa_use (gimple stmt ATTRIBUTE_UNUSED, tree t,
552 void *data ATTRIBUTE_UNUSED)
554 t = get_base_address (t);
556 if (!t || is_gimple_reg (t))
559 /* At present we can't pass non-SSA arguments to split function.
560 FIXME: this can be relaxed by passing references to arguments. */
561 if (TREE_CODE (t) == PARM_DECL)
563 if (dump_file && (dump_flags & TDF_DETAILS))
564 fprintf (dump_file, "Can not split use of non-ssa function parameter.\n");
568 if ((TREE_CODE (t) == VAR_DECL && auto_var_in_fn_p (t, current_function_decl))
569 || (TREE_CODE (t) == RESULT_DECL))
570 bitmap_set_bit ((bitmap)data, DECL_UID (t));
572 /* For DECL_BY_REFERENCE, the return value is actually pointer. We want to pretend
573 that the value pointed to is actual result decl. */
574 if (t && (TREE_CODE (t) == MEM_REF || INDIRECT_REF_P (t))
575 && TREE_CODE (TREE_OPERAND (t, 0)) == SSA_NAME
576 && TREE_CODE (SSA_NAME_VAR (TREE_OPERAND (t, 0))) == RESULT_DECL
577 && DECL_BY_REFERENCE (DECL_RESULT (current_function_decl)))
578 return bitmap_bit_p ((bitmap)data, DECL_UID (DECL_RESULT (current_function_decl)));
582 /* Compute local properties of basic block BB we collect when looking for
583 split points. We look for ssa defs and store them in SET_SSA_NAMES,
584 for ssa uses and store them in USED_SSA_NAMES and for any non-SSA automatic
585 vars stored in NON_SSA_VARS.
587 When BB has edge to RETURN_BB, collect uses in RETURN_BB too.
589 Return false when BB contains something that prevents it from being put into
593 visit_bb (basic_block bb, basic_block return_bb,
594 bitmap set_ssa_names, bitmap used_ssa_names,
597 gimple_stmt_iterator bsi;
600 bool can_split = true;
602 for (bsi = gsi_start_bb (bb); !gsi_end_p (bsi); gsi_next (&bsi))
604 gimple stmt = gsi_stmt (bsi);
609 if (is_gimple_debug (stmt))
612 /* FIXME: We can split regions containing EH. We can not however
613 split RESX, EH_DISPATCH and EH_POINTER referring to same region
614 into different partitions. This would require tracking of
615 EH regions and checking in consider_split_point if they
616 are not used elsewhere. */
617 if (gimple_code (stmt) == GIMPLE_RESX
618 && stmt_can_throw_external (stmt))
620 if (dump_file && (dump_flags & TDF_DETAILS))
621 fprintf (dump_file, "Can not split external resx.\n");
624 if (gimple_code (stmt) == GIMPLE_EH_DISPATCH)
626 if (dump_file && (dump_flags & TDF_DETAILS))
627 fprintf (dump_file, "Can not split eh dispatch.\n");
631 /* Check builtins that prevent splitting. */
632 if (gimple_code (stmt) == GIMPLE_CALL
633 && (decl = gimple_call_fndecl (stmt)) != NULL_TREE
634 && DECL_BUILT_IN (decl)
635 && DECL_BUILT_IN_CLASS (decl) == BUILT_IN_NORMAL)
636 switch (DECL_FUNCTION_CODE (decl))
638 /* FIXME: once we will allow passing non-parm values to split part,
639 we need to be sure to handle correct builtin_stack_save and
640 builtin_stack_restore. At the moment we are safe; there is no
641 way to store builtin_stack_save result in non-SSA variable
642 since all calls to those are compiler generated. */
644 case BUILT_IN_VA_START:
645 if (dump_file && (dump_flags & TDF_DETAILS))
646 fprintf (dump_file, "Can not split builtin_apply and va_start.\n");
649 case BUILT_IN_EH_POINTER:
650 if (dump_file && (dump_flags & TDF_DETAILS))
651 fprintf (dump_file, "Can not split builtin_eh_pointer.\n");
658 FOR_EACH_SSA_TREE_OPERAND (op, stmt, iter, SSA_OP_DEF)
659 bitmap_set_bit (set_ssa_names, SSA_NAME_VERSION (op));
660 FOR_EACH_SSA_TREE_OPERAND (op, stmt, iter, SSA_OP_USE)
661 bitmap_set_bit (used_ssa_names, SSA_NAME_VERSION (op));
662 can_split &= !walk_stmt_load_store_addr_ops (stmt, non_ssa_vars,
667 for (bsi = gsi_start_phis (bb); !gsi_end_p (bsi); gsi_next (&bsi))
669 gimple stmt = gsi_stmt (bsi);
672 if (is_gimple_debug (stmt))
674 if (!is_gimple_reg (gimple_phi_result (stmt)))
676 bitmap_set_bit (set_ssa_names,
677 SSA_NAME_VERSION (gimple_phi_result (stmt)));
678 for (i = 0; i < gimple_phi_num_args (stmt); i++)
680 tree op = gimple_phi_arg_def (stmt, i);
681 if (TREE_CODE (op) == SSA_NAME)
682 bitmap_set_bit (used_ssa_names, SSA_NAME_VERSION (op));
684 can_split &= !walk_stmt_load_store_addr_ops (stmt, non_ssa_vars,
689 /* Record also uses comming from PHI operand in return BB. */
690 FOR_EACH_EDGE (e, ei, bb->succs)
691 if (e->dest == return_bb)
693 for (bsi = gsi_start_phis (return_bb); !gsi_end_p (bsi); gsi_next (&bsi))
695 gimple stmt = gsi_stmt (bsi);
696 tree op = gimple_phi_arg_def (stmt, e->dest_idx);
698 if (is_gimple_debug (stmt))
700 if (!is_gimple_reg (gimple_phi_result (stmt)))
702 if (TREE_CODE (op) == SSA_NAME)
703 bitmap_set_bit (used_ssa_names, SSA_NAME_VERSION (op));
705 can_split &= !mark_nonssa_use (stmt, op, non_ssa_vars);
711 /* Stack entry for recursive DFS walk in find_split_point. */
715 /* Basic block we are examining. */
718 /* SSA names set and used by the BB and all BBs reachable
719 from it via DFS walk. */
720 bitmap set_ssa_names, used_ssa_names;
723 /* All BBS visited from this BB via DFS walk. */
726 /* Last examined edge in DFS walk. Since we walk unoriented graph,
727 the value is up to sum of incomming and outgoing edges of BB. */
728 unsigned int edge_num;
730 /* Stack entry index of earliest BB reachable from current BB
731 or any BB visited later in DFS valk. */
734 /* Overall time and size of all BBs reached from this BB in DFS walk. */
735 int overall_time, overall_size;
737 /* When false we can not split on this BB. */
740 DEF_VEC_O(stack_entry);
741 DEF_VEC_ALLOC_O(stack_entry,heap);
744 /* Find all articulations and call consider_split on them.
745 OVERALL_TIME and OVERALL_SIZE is time and size of the function.
747 We perform basic algorithm for finding an articulation in a graph
748 created from CFG by considering it to be an unoriented graph.
750 The articulation is discovered via DFS walk. We collect earliest
751 basic block on stack that is reachable via backward edge. Articulation
752 is any basic block such that there is no backward edge bypassing it.
753 To reduce stack usage we maintain heap allocated stack in STACK vector.
754 AUX pointer of BB is set to index it appears in the stack or -1 once
755 it is visited and popped off the stack.
757 The algorithm finds articulation after visiting the whole component
758 reachable by it. This makes it convenient to collect information about
759 the component used by consider_split. */
762 find_split_points (int overall_time, int overall_size)
765 VEC(stack_entry, heap) *stack = NULL;
767 basic_block return_bb = find_return_bb ();
768 struct split_point current;
770 current.header_time = overall_time;
771 current.header_size = overall_size;
772 current.split_time = 0;
773 current.split_size = 0;
774 current.ssa_names_to_pass = BITMAP_ALLOC (NULL);
776 first.bb = ENTRY_BLOCK_PTR;
778 first.overall_time = 0;
779 first.overall_size = 0;
780 first.earliest = INT_MAX;
781 first.set_ssa_names = 0;
782 first.used_ssa_names = 0;
783 first.bbs_visited = 0;
784 VEC_safe_push (stack_entry, heap, stack, &first);
785 ENTRY_BLOCK_PTR->aux = (void *)(intptr_t)-1;
787 while (!VEC_empty (stack_entry, stack))
789 stack_entry *entry = VEC_last (stack_entry, stack);
791 /* We are walking an acyclic graph, so edge_num counts
792 succ and pred edges together. However when considering
793 articulation, we want to have processed everything reachable
794 from articulation but nothing that reaches into it. */
795 if (entry->edge_num == EDGE_COUNT (entry->bb->succs)
796 && entry->bb != ENTRY_BLOCK_PTR)
798 int pos = VEC_length (stack_entry, stack);
799 entry->can_split &= visit_bb (entry->bb, return_bb,
800 entry->set_ssa_names,
801 entry->used_ssa_names,
802 entry->non_ssa_vars);
803 if (pos <= entry->earliest && !entry->can_split
804 && dump_file && (dump_flags & TDF_DETAILS))
806 "found articulation at bb %i but can not split\n",
808 if (pos <= entry->earliest && entry->can_split)
810 if (dump_file && (dump_flags & TDF_DETAILS))
811 fprintf (dump_file, "found articulation at bb %i\n",
813 current.entry_bb = entry->bb;
814 current.ssa_names_to_pass = BITMAP_ALLOC (NULL);
815 bitmap_and_compl (current.ssa_names_to_pass,
816 entry->used_ssa_names, entry->set_ssa_names);
817 current.header_time = overall_time - entry->overall_time;
818 current.header_size = overall_size - entry->overall_size;
819 current.split_time = entry->overall_time;
820 current.split_size = entry->overall_size;
821 current.split_bbs = entry->bbs_visited;
822 consider_split (¤t, entry->non_ssa_vars, return_bb);
823 BITMAP_FREE (current.ssa_names_to_pass);
826 /* Do actual DFS walk. */
828 < (EDGE_COUNT (entry->bb->succs)
829 + EDGE_COUNT (entry->bb->preds)))
833 if (entry->edge_num < EDGE_COUNT (entry->bb->succs))
835 e = EDGE_SUCC (entry->bb, entry->edge_num);
840 e = EDGE_PRED (entry->bb, entry->edge_num
841 - EDGE_COUNT (entry->bb->succs));
847 /* New BB to visit, push it to the stack. */
848 if (dest != return_bb && dest != EXIT_BLOCK_PTR
851 stack_entry new_entry;
854 new_entry.edge_num = 0;
855 new_entry.overall_time
856 = VEC_index (bb_info, bb_info_vec, dest->index)->time;
857 new_entry.overall_size
858 = VEC_index (bb_info, bb_info_vec, dest->index)->size;
859 new_entry.earliest = INT_MAX;
860 new_entry.set_ssa_names = BITMAP_ALLOC (NULL);
861 new_entry.used_ssa_names = BITMAP_ALLOC (NULL);
862 new_entry.bbs_visited = BITMAP_ALLOC (NULL);
863 new_entry.non_ssa_vars = BITMAP_ALLOC (NULL);
864 new_entry.can_split = true;
865 bitmap_set_bit (new_entry.bbs_visited, dest->index);
866 VEC_safe_push (stack_entry, heap, stack, &new_entry);
867 dest->aux = (void *)(intptr_t)VEC_length (stack_entry, stack);
869 /* Back edge found, record the earliest point. */
870 else if ((intptr_t)dest->aux > 0
871 && (intptr_t)dest->aux < entry->earliest)
872 entry->earliest = (intptr_t)dest->aux;
874 /* We are done with examing the edges. pop off the value from stack and
875 merge stuff we cummulate during the walk. */
876 else if (entry->bb != ENTRY_BLOCK_PTR)
878 stack_entry *prev = VEC_index (stack_entry, stack,
879 VEC_length (stack_entry, stack) - 2);
881 entry->bb->aux = (void *)(intptr_t)-1;
882 prev->can_split &= entry->can_split;
883 if (prev->set_ssa_names)
885 bitmap_ior_into (prev->set_ssa_names, entry->set_ssa_names);
886 bitmap_ior_into (prev->used_ssa_names, entry->used_ssa_names);
887 bitmap_ior_into (prev->bbs_visited, entry->bbs_visited);
888 bitmap_ior_into (prev->non_ssa_vars, entry->non_ssa_vars);
890 if (prev->earliest > entry->earliest)
891 prev->earliest = entry->earliest;
892 prev->overall_time += entry->overall_time;
893 prev->overall_size += entry->overall_size;
894 BITMAP_FREE (entry->set_ssa_names);
895 BITMAP_FREE (entry->used_ssa_names);
896 BITMAP_FREE (entry->bbs_visited);
897 BITMAP_FREE (entry->non_ssa_vars);
898 VEC_pop (stack_entry, stack);
901 VEC_pop (stack_entry, stack);
903 ENTRY_BLOCK_PTR->aux = NULL;
906 BITMAP_FREE (current.ssa_names_to_pass);
909 /* Split function at SPLIT_POINT. */
912 split_function (struct split_point *split_point)
914 VEC (tree, heap) *args_to_pass = NULL;
915 bitmap args_to_skip = BITMAP_ALLOC (NULL);
918 struct cgraph_node *node;
919 basic_block return_bb = find_return_bb ();
921 gimple_stmt_iterator gsi;
925 tree retval = NULL, real_retval = NULL;
926 bool split_part_return_p = false;
927 gimple last_stmt = NULL;
931 fprintf (dump_file, "\n\nSplitting function at:\n");
932 dump_split_point (dump_file, split_point);
935 /* Collect the parameters of new function and args_to_skip bitmap. */
936 for (parm = DECL_ARGUMENTS (current_function_decl);
937 parm; parm = DECL_CHAIN (parm), num++)
938 if (!is_gimple_reg (parm)
939 || !gimple_default_def (cfun, parm)
940 || !bitmap_bit_p (split_point->ssa_names_to_pass,
941 SSA_NAME_VERSION (gimple_default_def (cfun, parm))))
942 bitmap_set_bit (args_to_skip, num);
944 VEC_safe_push (tree, heap, args_to_pass, gimple_default_def (cfun, parm));
946 /* See if the split function will return. */
947 FOR_EACH_EDGE (e, ei, return_bb->preds)
948 if (bitmap_bit_p (split_point->split_bbs, e->src->index))
951 split_part_return_p = true;
953 /* Add return block to what will become the split function.
954 We do not return; no return block is needed. */
955 if (!split_part_return_p)
957 /* We have no return block, so nothing is needed. */
958 else if (return_bb == EXIT_BLOCK_PTR)
960 /* When we do not want to return value, we need to construct
961 new return block with empty return statement.
962 FIXME: Once we are able to change return type, we should change function
963 to return void instead of just outputting function with undefined return
964 value. For structures this affects quality of codegen. */
965 else if (!split_point->split_part_set_retval
966 && find_retval (return_bb))
968 bool redirected = true;
969 basic_block new_return_bb = create_basic_block (NULL, 0, return_bb);
970 gimple_stmt_iterator gsi = gsi_start_bb (new_return_bb);
971 gsi_insert_after (&gsi, gimple_build_return (NULL), GSI_NEW_STMT);
975 FOR_EACH_EDGE (e, ei, return_bb->preds)
976 if (bitmap_bit_p (split_point->split_bbs, e->src->index))
978 new_return_bb->count += e->count;
979 new_return_bb->frequency += EDGE_FREQUENCY (e);
980 redirect_edge_and_branch (e, new_return_bb);
985 e = make_edge (new_return_bb, EXIT_BLOCK_PTR, 0);
986 e->probability = REG_BR_PROB_BASE;
987 e->count = new_return_bb->count;
988 bitmap_set_bit (split_point->split_bbs, new_return_bb->index);
989 /* We change CFG in a way tree-inline is not able to compensate on while
990 updating PHIs. There are only virtuals in return_bb, so recompute
992 for (gsi = gsi_start_phis (return_bb); !gsi_end_p (gsi);)
994 gimple stmt = gsi_stmt (gsi);
995 gcc_assert (!is_gimple_reg (gimple_phi_result (stmt)));
996 mark_sym_for_renaming (SSA_NAME_VAR (PHI_RESULT (stmt)));
997 gsi_remove (&gsi, false);
1000 /* When we pass aorund the value, use existing return block. */
1002 bitmap_set_bit (split_point->split_bbs, return_bb->index);
1004 /* Now create the actual clone. */
1005 rebuild_cgraph_edges ();
1006 node = cgraph_function_versioning (cgraph_node (current_function_decl),
1009 split_point->split_bbs,
1010 split_point->entry_bb, "part");
1011 /* For usual cloning it is enough to clear builtin only when signature
1012 changes. For partial inlining we however can not expect the part
1013 of builtin implementation to have same semantic as the whole. */
1014 if (DECL_BUILT_IN (node->decl))
1016 DECL_BUILT_IN_CLASS (node->decl) = NOT_BUILT_IN;
1017 DECL_FUNCTION_CODE (node->decl) = (enum built_in_function) 0;
1019 cgraph_node_remove_callees (cgraph_node (current_function_decl));
1020 if (!split_part_return_p)
1021 TREE_THIS_VOLATILE (node->decl) = 1;
1023 dump_function_to_file (node->decl, dump_file, dump_flags);
1025 /* Create the basic block we place call into. It is the entry basic block
1026 split after last label. */
1027 call_bb = split_point->entry_bb;
1028 for (gsi = gsi_start_bb (call_bb); !gsi_end_p (gsi);)
1029 if (gimple_code (gsi_stmt (gsi)) == GIMPLE_LABEL)
1031 last_stmt = gsi_stmt (gsi);
1036 e = split_block (split_point->entry_bb, last_stmt);
1039 /* Produce the call statement. */
1040 gsi = gsi_last_bb (call_bb);
1041 call = gimple_build_call_vec (node->decl, args_to_pass);
1042 gimple_set_block (call, DECL_INITIAL (current_function_decl));
1044 /* We avoid address being taken on any variable used by split part,
1045 so return slot optimization is always possible. Moreover this is
1046 required to make DECL_BY_REFERENCE work. */
1047 if (aggregate_value_p (DECL_RESULT (current_function_decl),
1048 TREE_TYPE (current_function_decl)))
1049 gimple_call_set_return_slot_opt (call, true);
1051 /* Update return value. This is bit tricky. When we do not return,
1052 do nothing. When we return we might need to update return_bb
1053 or produce a new return statement. */
1054 if (!split_part_return_p)
1055 gsi_insert_after (&gsi, call, GSI_NEW_STMT);
1058 e = make_edge (call_bb, return_bb,
1059 return_bb == EXIT_BLOCK_PTR ? 0 : EDGE_FALLTHRU);
1060 e->count = call_bb->count;
1061 e->probability = REG_BR_PROB_BASE;
1063 /* If there is return basic block, see what value we need to store
1064 return value into and put call just before it. */
1065 if (return_bb != EXIT_BLOCK_PTR)
1067 real_retval = retval = find_retval (return_bb);
1069 if (real_retval && split_point->split_part_set_retval)
1071 gimple_stmt_iterator psi;
1073 /* See if we need new SSA_NAME for the result.
1074 When DECL_BY_REFERENCE is true, retval is actually pointer to
1075 return value and it is constant in whole function. */
1076 if (TREE_CODE (retval) == SSA_NAME
1077 && !DECL_BY_REFERENCE (DECL_RESULT (current_function_decl)))
1079 retval = make_ssa_name (SSA_NAME_VAR (retval), call);
1081 /* See if there is PHI defining return value. */
1082 for (psi = gsi_start_phis (return_bb);
1083 !gsi_end_p (psi); gsi_next (&psi))
1084 if (is_gimple_reg (gimple_phi_result (gsi_stmt (psi))))
1087 /* When there is PHI, just update its value. */
1088 if (TREE_CODE (retval) == SSA_NAME
1089 && !gsi_end_p (psi))
1090 add_phi_arg (gsi_stmt (psi), retval, e, UNKNOWN_LOCATION);
1091 /* Otherwise update the return BB itself.
1092 find_return_bb allows at most one assignment to return value,
1093 so update first statement. */
1096 gimple_stmt_iterator bsi;
1097 for (bsi = gsi_start_bb (return_bb); !gsi_end_p (bsi);
1099 if (gimple_code (gsi_stmt (bsi)) == GIMPLE_RETURN)
1101 gimple_return_set_retval (gsi_stmt (bsi), retval);
1104 else if (gimple_code (gsi_stmt (bsi)) == GIMPLE_ASSIGN)
1106 gimple_assign_set_rhs1 (gsi_stmt (bsi), retval);
1109 update_stmt (gsi_stmt (bsi));
1112 if (DECL_BY_REFERENCE (DECL_RESULT (current_function_decl)))
1113 gimple_call_set_lhs (call, build_simple_mem_ref (retval));
1115 gimple_call_set_lhs (call, retval);
1117 gsi_insert_after (&gsi, call, GSI_NEW_STMT);
1119 /* We don't use return block (there is either no return in function or
1120 multiple of them). So create new basic block with return statement.
1125 if (split_point->split_part_set_retval
1126 && !VOID_TYPE_P (TREE_TYPE (TREE_TYPE (current_function_decl))))
1128 retval = DECL_RESULT (current_function_decl);
1130 /* We use temporary register to hold value when aggregate_value_p
1131 is false. Similarly for DECL_BY_REFERENCE we must avoid extra
1133 if (!aggregate_value_p (retval, TREE_TYPE (current_function_decl))
1134 && !DECL_BY_REFERENCE (retval))
1135 retval = create_tmp_reg (TREE_TYPE (retval), NULL);
1136 if (is_gimple_reg (retval))
1138 /* When returning by reference, there is only one SSA name
1139 assigned to RESULT_DECL (that is pointer to return value).
1140 Look it up or create new one if it is missing. */
1141 if (DECL_BY_REFERENCE (retval))
1144 if ((retval_name = gimple_default_def (cfun, retval))
1146 retval = retval_name;
1149 retval_name = make_ssa_name (retval,
1150 gimple_build_nop ());
1151 set_default_def (retval, retval_name);
1152 retval = retval_name;
1155 /* Otherwise produce new SSA name for return value. */
1157 retval = make_ssa_name (retval, call);
1159 if (DECL_BY_REFERENCE (DECL_RESULT (current_function_decl)))
1160 gimple_call_set_lhs (call, build_simple_mem_ref (retval));
1162 gimple_call_set_lhs (call, retval);
1164 gsi_insert_after (&gsi, call, GSI_NEW_STMT);
1165 ret = gimple_build_return (retval);
1166 gsi_insert_after (&gsi, ret, GSI_NEW_STMT);
1169 free_dominance_info (CDI_DOMINATORS);
1170 free_dominance_info (CDI_POST_DOMINATORS);
1171 compute_inline_parameters (node);
1174 /* Execute function splitting pass. */
1177 execute_split_functions (void)
1179 gimple_stmt_iterator bsi;
1181 int overall_time = 0, overall_size = 0;
1183 struct cgraph_node *node = cgraph_node (current_function_decl);
1185 if (flags_from_decl_or_type (current_function_decl) & ECF_NORETURN)
1188 fprintf (dump_file, "Not splitting: noreturn function.\n");
1191 if (MAIN_NAME_P (DECL_NAME (current_function_decl)))
1194 fprintf (dump_file, "Not splitting: main function.\n");
1197 /* This can be relaxed; function might become inlinable after splitting
1198 away the uninlinable part. */
1199 if (!node->local.inlinable)
1202 fprintf (dump_file, "Not splitting: not inlinable.\n");
1205 if (node->local.disregard_inline_limits)
1208 fprintf (dump_file, "Not splitting: disregading inline limits.\n");
1211 /* This can be relaxed; most of versioning tests actually prevents
1213 if (!tree_versionable_function_p (current_function_decl))
1216 fprintf (dump_file, "Not splitting: not versionable.\n");
1219 /* FIXME: we could support this. */
1220 if (DECL_STRUCT_FUNCTION (current_function_decl)->static_chain_decl)
1223 fprintf (dump_file, "Not splitting: nested function.\n");
1227 /* See if it makes sense to try to split.
1228 It makes sense to split if we inline, that is if we have direct calls to
1229 handle or direct calls are possibly going to appear as result of indirect
1231 Note that we are not completely conservative about disqualifying functions
1232 called once. It is possible that the caller is called more then once and
1233 then inlining would still benefit. */
1234 if ((!node->callers || !node->callers->next_caller)
1235 && !node->address_taken
1236 && ((!flag_lto && !flag_whopr) || !node->local.externally_visible))
1239 fprintf (dump_file, "Not splitting: not called directly "
1240 "or called once.\n");
1244 /* FIXME: We can actually split if splitting reduces call overhead. */
1245 if (!flag_inline_small_functions
1246 && !DECL_DECLARED_INLINE_P (current_function_decl))
1249 fprintf (dump_file, "Not splitting: not autoinlining and function"
1250 " is not inline.\n");
1254 /* Compute local info about basic blocks and determine function size/time. */
1255 VEC_safe_grow_cleared (bb_info, heap, bb_info_vec, last_basic_block + 1);
1256 memset (&best_split_point, 0, sizeof (best_split_point));
1261 int freq = compute_call_stmt_bb_frequency (current_function_decl, bb);
1263 if (dump_file && (dump_flags & TDF_DETAILS))
1264 fprintf (dump_file, "Basic block %i\n", bb->index);
1266 for (bsi = gsi_start_bb (bb); !gsi_end_p (bsi); gsi_next (&bsi))
1268 int this_time, this_size;
1269 gimple stmt = gsi_stmt (bsi);
1271 this_size = estimate_num_insns (stmt, &eni_size_weights);
1272 this_time = estimate_num_insns (stmt, &eni_time_weights) * freq;
1276 if (dump_file && (dump_flags & TDF_DETAILS))
1278 fprintf (dump_file, " freq:%6i size:%3i time:%3i ",
1279 freq, this_size, this_time);
1280 print_gimple_stmt (dump_file, stmt, 0, 0);
1283 overall_time += time;
1284 overall_size += size;
1285 VEC_index (bb_info, bb_info_vec, bb->index)->time = time;
1286 VEC_index (bb_info, bb_info_vec, bb->index)->size = size;
1288 find_split_points (overall_time, overall_size);
1289 if (best_split_point.split_bbs)
1291 split_function (&best_split_point);
1292 BITMAP_FREE (best_split_point.ssa_names_to_pass);
1293 BITMAP_FREE (best_split_point.split_bbs);
1294 todo = TODO_update_ssa | TODO_cleanup_cfg;
1296 VEC_free (bb_info, heap, bb_info_vec);
1302 gate_split_functions (void)
1304 return flag_partial_inlining;
1307 struct gimple_opt_pass pass_split_functions =
1311 "fnsplit", /* name */
1312 gate_split_functions, /* gate */
1313 execute_split_functions, /* execute */
1316 0, /* static_pass_number */
1317 TV_IPA_FNSPLIT, /* tv_id */
1318 PROP_cfg, /* properties_required */
1319 0, /* properties_provided */
1320 0, /* properties_destroyed */
1321 0, /* todo_flags_start */
1322 TODO_dump_func /* todo_flags_finish */