1 /* Function splitting pass
2 Copyright (C) 2010, 2011
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 improvement 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 function. */
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, void *data)
141 t = get_base_address (t);
143 if (!t || is_gimple_reg (t))
146 if (TREE_CODE (t) == PARM_DECL
147 || (TREE_CODE (t) == VAR_DECL
148 && auto_var_in_fn_p (t, current_function_decl))
149 || TREE_CODE (t) == RESULT_DECL
150 || TREE_CODE (t) == LABEL_DECL)
151 return bitmap_bit_p ((bitmap)data, DECL_UID (t));
153 /* For DECL_BY_REFERENCE, the return value is actually a pointer. We want
154 to pretend that the value pointed to is actual result decl. */
155 if ((TREE_CODE (t) == MEM_REF || INDIRECT_REF_P (t))
156 && TREE_CODE (TREE_OPERAND (t, 0)) == SSA_NAME
157 && TREE_CODE (SSA_NAME_VAR (TREE_OPERAND (t, 0))) == RESULT_DECL
158 && DECL_BY_REFERENCE (DECL_RESULT (current_function_decl)))
160 bitmap_bit_p ((bitmap)data,
161 DECL_UID (DECL_RESULT (current_function_decl)));
166 /* Dump split point CURRENT. */
169 dump_split_point (FILE * file, struct split_point *current)
172 "Split point at BB %i header time:%i header size: %i"
173 " split time: %i split size: %i\n bbs: ",
174 current->entry_bb->index, current->header_time,
175 current->header_size, current->split_time, current->split_size);
176 dump_bitmap (file, current->split_bbs);
177 fprintf (file, " SSA names to pass: ");
178 dump_bitmap (file, current->ssa_names_to_pass);
181 /* Look for all BBs in header that might lead to the split part and verify
182 that they are not defining any non-SSA var used by the split part.
183 Parameters are the same as for consider_split. */
186 verify_non_ssa_vars (struct split_point *current, bitmap non_ssa_vars,
187 basic_block return_bb)
189 bitmap seen = BITMAP_ALLOC (NULL);
190 VEC (basic_block,heap) *worklist = NULL;
195 FOR_EACH_EDGE (e, ei, current->entry_bb->preds)
196 if (e->src != ENTRY_BLOCK_PTR
197 && !bitmap_bit_p (current->split_bbs, e->src->index))
199 VEC_safe_push (basic_block, heap, worklist, e->src);
200 bitmap_set_bit (seen, e->src->index);
203 while (!VEC_empty (basic_block, worklist))
205 gimple_stmt_iterator bsi;
206 basic_block bb = VEC_pop (basic_block, worklist);
208 FOR_EACH_EDGE (e, ei, bb->preds)
209 if (e->src != ENTRY_BLOCK_PTR
210 && bitmap_set_bit (seen, e->src->index))
212 gcc_checking_assert (!bitmap_bit_p (current->split_bbs,
214 VEC_safe_push (basic_block, heap, worklist, e->src);
216 for (bsi = gsi_start_bb (bb); !gsi_end_p (bsi); gsi_next (&bsi))
218 gimple stmt = gsi_stmt (bsi);
219 if (is_gimple_debug (stmt))
221 if (walk_stmt_load_store_addr_ops
222 (stmt, non_ssa_vars, test_nonssa_use, test_nonssa_use,
228 if (gimple_code (stmt) == GIMPLE_LABEL
229 && test_nonssa_use (stmt, gimple_label_label (stmt),
236 for (bsi = gsi_start_phis (bb); !gsi_end_p (bsi); gsi_next (&bsi))
238 if (walk_stmt_load_store_addr_ops
239 (gsi_stmt (bsi), non_ssa_vars, test_nonssa_use, test_nonssa_use,
246 FOR_EACH_EDGE (e, ei, bb->succs)
248 if (e->dest != return_bb)
250 for (bsi = gsi_start_phis (return_bb); !gsi_end_p (bsi);
253 gimple stmt = gsi_stmt (bsi);
254 tree op = gimple_phi_arg_def (stmt, e->dest_idx);
256 if (!is_gimple_reg (gimple_phi_result (stmt)))
258 if (TREE_CODE (op) != SSA_NAME
259 && test_nonssa_use (stmt, op, non_ssa_vars))
269 VEC_free (basic_block, heap, worklist);
273 /* We found an split_point CURRENT. NON_SSA_VARS is bitmap of all non ssa
274 variables used and RETURN_BB is return basic block.
275 See if we can split function here. */
278 consider_split (struct split_point *current, bitmap non_ssa_vars,
279 basic_block return_bb)
282 unsigned int num_args = 0;
283 unsigned int call_overhead;
286 gimple_stmt_iterator bsi;
288 int incoming_freq = 0;
291 if (dump_file && (dump_flags & TDF_DETAILS))
292 dump_split_point (dump_file, current);
294 FOR_EACH_EDGE (e, ei, current->entry_bb->preds)
295 if (!bitmap_bit_p (current->split_bbs, e->src->index))
296 incoming_freq += EDGE_FREQUENCY (e);
298 /* Do not split when we would end up calling function anyway. */
300 >= (ENTRY_BLOCK_PTR->frequency
301 * PARAM_VALUE (PARAM_PARTIAL_INLINING_ENTRY_PROBABILITY) / 100))
303 if (dump_file && (dump_flags & TDF_DETAILS))
305 " Refused: incoming frequency is too large.\n");
309 if (!current->header_size)
311 if (dump_file && (dump_flags & TDF_DETAILS))
312 fprintf (dump_file, " Refused: header empty\n");
316 /* Verify that PHI args on entry are either virtual or all their operands
317 incoming from header are the same. */
318 for (bsi = gsi_start_phis (current->entry_bb); !gsi_end_p (bsi); gsi_next (&bsi))
320 gimple stmt = gsi_stmt (bsi);
323 if (!is_gimple_reg (gimple_phi_result (stmt)))
325 for (i = 0; i < gimple_phi_num_args (stmt); i++)
327 edge e = gimple_phi_arg_edge (stmt, i);
328 if (!bitmap_bit_p (current->split_bbs, e->src->index))
330 tree edge_val = gimple_phi_arg_def (stmt, i);
331 if (val && edge_val != val)
333 if (dump_file && (dump_flags & TDF_DETAILS))
335 " Refused: entry BB has PHI with multiple variants\n");
344 /* See what argument we will pass to the split function and compute
346 call_overhead = eni_size_weights.call_cost;
347 for (parm = DECL_ARGUMENTS (current_function_decl); parm;
348 parm = DECL_CHAIN (parm))
350 if (!is_gimple_reg (parm))
352 if (bitmap_bit_p (non_ssa_vars, DECL_UID (parm)))
354 if (dump_file && (dump_flags & TDF_DETAILS))
356 " Refused: need to pass non-ssa param values\n");
360 else if (gimple_default_def (cfun, parm)
361 && bitmap_bit_p (current->ssa_names_to_pass,
362 SSA_NAME_VERSION (gimple_default_def
365 if (!VOID_TYPE_P (TREE_TYPE (parm)))
366 call_overhead += estimate_move_cost (TREE_TYPE (parm));
370 if (!VOID_TYPE_P (TREE_TYPE (current_function_decl)))
371 call_overhead += estimate_move_cost (TREE_TYPE (current_function_decl));
373 if (current->split_size <= call_overhead)
375 if (dump_file && (dump_flags & TDF_DETAILS))
377 " Refused: split size is smaller than call overhead\n");
380 if (current->header_size + call_overhead
381 >= (unsigned int)(DECL_DECLARED_INLINE_P (current_function_decl)
382 ? MAX_INLINE_INSNS_SINGLE
383 : MAX_INLINE_INSNS_AUTO))
385 if (dump_file && (dump_flags & TDF_DETAILS))
387 " Refused: header size is too large for inline candidate\n");
391 /* FIXME: we currently can pass only SSA function parameters to the split
392 arguments. Once parm_adjustment infrastructure is supported by cloning,
393 we can pass more than that. */
394 if (num_args != bitmap_count_bits (current->ssa_names_to_pass))
397 if (dump_file && (dump_flags & TDF_DETAILS))
399 " Refused: need to pass non-param values\n");
403 /* When there are non-ssa vars used in the split region, see if they
404 are used in the header region. If so, reject the split.
405 FIXME: we can use nested function support to access both. */
406 if (!bitmap_empty_p (non_ssa_vars)
407 && !verify_non_ssa_vars (current, non_ssa_vars, return_bb))
409 if (dump_file && (dump_flags & TDF_DETAILS))
411 " Refused: split part has non-ssa uses\n");
414 /* See if retval used by return bb is computed by header or split part.
415 When it is computed by split part, we need to produce return statement
416 in the split part and add code to header to pass it around.
418 This is bit tricky to test:
419 1) When there is no return_bb or no return value, we always pass
421 2) Invariants are always computed by caller.
422 3) For SSA we need to look if defining statement is in header or split part
423 4) For non-SSA we need to look where the var is computed. */
424 retval = find_retval (return_bb);
426 current->split_part_set_retval = true;
427 else if (is_gimple_min_invariant (retval))
428 current->split_part_set_retval = false;
429 /* Special case is value returned by reference we record as if it was non-ssa
430 set to result_decl. */
431 else if (TREE_CODE (retval) == SSA_NAME
432 && TREE_CODE (SSA_NAME_VAR (retval)) == RESULT_DECL
433 && DECL_BY_REFERENCE (DECL_RESULT (current_function_decl)))
434 current->split_part_set_retval
435 = bitmap_bit_p (non_ssa_vars, DECL_UID (SSA_NAME_VAR (retval)));
436 else if (TREE_CODE (retval) == SSA_NAME)
437 current->split_part_set_retval
438 = (!SSA_NAME_IS_DEFAULT_DEF (retval)
439 && (bitmap_bit_p (current->split_bbs,
440 gimple_bb (SSA_NAME_DEF_STMT (retval))->index)
441 || gimple_bb (SSA_NAME_DEF_STMT (retval)) == return_bb));
442 else if (TREE_CODE (retval) == PARM_DECL)
443 current->split_part_set_retval = false;
444 else if (TREE_CODE (retval) == VAR_DECL
445 || TREE_CODE (retval) == RESULT_DECL)
446 current->split_part_set_retval
447 = bitmap_bit_p (non_ssa_vars, DECL_UID (retval));
449 current->split_part_set_retval = true;
451 /* split_function fixes up at most one PHI non-virtual PHI node in return_bb,
452 for the return value. If there are other PHIs, give up. */
453 if (return_bb != EXIT_BLOCK_PTR)
455 gimple_stmt_iterator psi;
457 for (psi = gsi_start_phis (return_bb); !gsi_end_p (psi); gsi_next (&psi))
458 if (is_gimple_reg (gimple_phi_result (gsi_stmt (psi)))
460 && current->split_part_set_retval
461 && TREE_CODE (retval) == SSA_NAME
462 && !DECL_BY_REFERENCE (DECL_RESULT (current_function_decl))
463 && SSA_NAME_DEF_STMT (retval) == gsi_stmt (psi)))
465 if (dump_file && (dump_flags & TDF_DETAILS))
467 " Refused: return bb has extra PHIs\n");
472 if (dump_file && (dump_flags & TDF_DETAILS))
473 fprintf (dump_file, " Accepted!\n");
475 /* At the moment chose split point with lowest frequency and that leaves
476 out smallest size of header.
477 In future we might re-consider this heuristics. */
478 if (!best_split_point.split_bbs
479 || best_split_point.entry_bb->frequency > current->entry_bb->frequency
480 || (best_split_point.entry_bb->frequency == current->entry_bb->frequency
481 && best_split_point.split_size < current->split_size))
484 if (dump_file && (dump_flags & TDF_DETAILS))
485 fprintf (dump_file, " New best split point!\n");
486 if (best_split_point.ssa_names_to_pass)
488 BITMAP_FREE (best_split_point.ssa_names_to_pass);
489 BITMAP_FREE (best_split_point.split_bbs);
491 best_split_point = *current;
492 best_split_point.ssa_names_to_pass = BITMAP_ALLOC (NULL);
493 bitmap_copy (best_split_point.ssa_names_to_pass,
494 current->ssa_names_to_pass);
495 best_split_point.split_bbs = BITMAP_ALLOC (NULL);
496 bitmap_copy (best_split_point.split_bbs, current->split_bbs);
500 /* Return basic block containing RETURN statement. We allow basic blocks
504 but return_bb can not be more complex than this.
505 If nothing is found, return EXIT_BLOCK_PTR.
507 When there are multiple RETURN statement, chose one with return value,
508 since that one is more likely shared by multiple code paths.
510 Return BB is special, because for function splitting it is the only
511 basic block that is duplicated in between header and split part of the
514 TODO: We might support multiple return blocks. */
517 find_return_bb (void)
520 basic_block return_bb = EXIT_BLOCK_PTR;
521 gimple_stmt_iterator bsi;
522 bool found_return = false;
523 tree retval = NULL_TREE;
525 if (!single_pred_p (EXIT_BLOCK_PTR))
528 e = single_pred_edge (EXIT_BLOCK_PTR);
529 for (bsi = gsi_last_bb (e->src); !gsi_end_p (bsi); gsi_prev (&bsi))
531 gimple stmt = gsi_stmt (bsi);
532 if (gimple_code (stmt) == GIMPLE_LABEL || is_gimple_debug (stmt))
534 else if (gimple_code (stmt) == GIMPLE_ASSIGN
536 && gimple_assign_single_p (stmt)
537 && (auto_var_in_fn_p (gimple_assign_rhs1 (stmt),
538 current_function_decl)
539 || is_gimple_min_invariant (gimple_assign_rhs1 (stmt)))
540 && retval == gimple_assign_lhs (stmt))
542 else if (gimple_code (stmt) == GIMPLE_RETURN)
545 retval = gimple_return_retval (stmt);
550 if (gsi_end_p (bsi) && found_return)
556 /* Given return basic block RETURN_BB, see where return value is really
559 find_retval (basic_block return_bb)
561 gimple_stmt_iterator bsi;
562 for (bsi = gsi_start_bb (return_bb); !gsi_end_p (bsi); gsi_next (&bsi))
563 if (gimple_code (gsi_stmt (bsi)) == GIMPLE_RETURN)
564 return gimple_return_retval (gsi_stmt (bsi));
565 else if (gimple_code (gsi_stmt (bsi)) == GIMPLE_ASSIGN)
566 return gimple_assign_rhs1 (gsi_stmt (bsi));
570 /* Callback for walk_stmt_load_store_addr_ops. If T is non-SSA automatic
571 variable, mark it as used in bitmap passed via DATA.
572 Return true when access to T prevents splitting the function. */
575 mark_nonssa_use (gimple stmt ATTRIBUTE_UNUSED, tree t, void *data)
577 t = get_base_address (t);
579 if (!t || is_gimple_reg (t))
582 /* At present we can't pass non-SSA arguments to split function.
583 FIXME: this can be relaxed by passing references to arguments. */
584 if (TREE_CODE (t) == PARM_DECL)
586 if (dump_file && (dump_flags & TDF_DETAILS))
588 "Cannot split: use of non-ssa function parameter.\n");
592 if ((TREE_CODE (t) == VAR_DECL
593 && auto_var_in_fn_p (t, current_function_decl))
594 || TREE_CODE (t) == RESULT_DECL
595 || TREE_CODE (t) == LABEL_DECL)
596 bitmap_set_bit ((bitmap)data, DECL_UID (t));
598 /* For DECL_BY_REFERENCE, the return value is actually a pointer. We want
599 to pretend that the value pointed to is actual result decl. */
600 if ((TREE_CODE (t) == MEM_REF || INDIRECT_REF_P (t))
601 && TREE_CODE (TREE_OPERAND (t, 0)) == SSA_NAME
602 && TREE_CODE (SSA_NAME_VAR (TREE_OPERAND (t, 0))) == RESULT_DECL
603 && DECL_BY_REFERENCE (DECL_RESULT (current_function_decl)))
605 bitmap_bit_p ((bitmap)data,
606 DECL_UID (DECL_RESULT (current_function_decl)));
611 /* Compute local properties of basic block BB we collect when looking for
612 split points. We look for ssa defs and store them in SET_SSA_NAMES,
613 for ssa uses and store them in USED_SSA_NAMES and for any non-SSA automatic
614 vars stored in NON_SSA_VARS.
616 When BB has edge to RETURN_BB, collect uses in RETURN_BB too.
618 Return false when BB contains something that prevents it from being put into
622 visit_bb (basic_block bb, basic_block return_bb,
623 bitmap set_ssa_names, bitmap used_ssa_names,
626 gimple_stmt_iterator bsi;
629 bool can_split = true;
631 for (bsi = gsi_start_bb (bb); !gsi_end_p (bsi); gsi_next (&bsi))
633 gimple stmt = gsi_stmt (bsi);
638 if (is_gimple_debug (stmt))
641 /* FIXME: We can split regions containing EH. We can not however
642 split RESX, EH_DISPATCH and EH_POINTER referring to same region
643 into different partitions. This would require tracking of
644 EH regions and checking in consider_split_point if they
645 are not used elsewhere. */
646 if (gimple_code (stmt) == GIMPLE_RESX
647 && stmt_can_throw_external (stmt))
649 if (dump_file && (dump_flags & TDF_DETAILS))
650 fprintf (dump_file, "Cannot split: external resx.\n");
653 if (gimple_code (stmt) == GIMPLE_EH_DISPATCH)
655 if (dump_file && (dump_flags & TDF_DETAILS))
656 fprintf (dump_file, "Cannot split: eh dispatch.\n");
660 /* Check builtins that prevent splitting. */
661 if (gimple_code (stmt) == GIMPLE_CALL
662 && (decl = gimple_call_fndecl (stmt)) != NULL_TREE
663 && DECL_BUILT_IN (decl)
664 && DECL_BUILT_IN_CLASS (decl) == BUILT_IN_NORMAL)
665 switch (DECL_FUNCTION_CODE (decl))
667 /* FIXME: once we will allow passing non-parm values to split part,
668 we need to be sure to handle correct builtin_stack_save and
669 builtin_stack_restore. At the moment we are safe; there is no
670 way to store builtin_stack_save result in non-SSA variable
671 since all calls to those are compiler generated. */
673 case BUILT_IN_APPLY_ARGS:
674 case BUILT_IN_VA_START:
675 if (dump_file && (dump_flags & TDF_DETAILS))
677 "Cannot split: builtin_apply and va_start.\n");
680 case BUILT_IN_EH_POINTER:
681 if (dump_file && (dump_flags & TDF_DETAILS))
682 fprintf (dump_file, "Cannot split: builtin_eh_pointer.\n");
689 FOR_EACH_SSA_TREE_OPERAND (op, stmt, iter, SSA_OP_DEF)
690 bitmap_set_bit (set_ssa_names, SSA_NAME_VERSION (op));
691 FOR_EACH_SSA_TREE_OPERAND (op, stmt, iter, SSA_OP_USE)
692 bitmap_set_bit (used_ssa_names, SSA_NAME_VERSION (op));
693 can_split &= !walk_stmt_load_store_addr_ops (stmt, non_ssa_vars,
698 for (bsi = gsi_start_phis (bb); !gsi_end_p (bsi); gsi_next (&bsi))
700 gimple stmt = gsi_stmt (bsi);
703 if (is_gimple_debug (stmt))
705 if (!is_gimple_reg (gimple_phi_result (stmt)))
707 bitmap_set_bit (set_ssa_names,
708 SSA_NAME_VERSION (gimple_phi_result (stmt)));
709 for (i = 0; i < gimple_phi_num_args (stmt); i++)
711 tree op = gimple_phi_arg_def (stmt, i);
712 if (TREE_CODE (op) == SSA_NAME)
713 bitmap_set_bit (used_ssa_names, SSA_NAME_VERSION (op));
715 can_split &= !walk_stmt_load_store_addr_ops (stmt, non_ssa_vars,
720 /* Record also uses coming from PHI operand in return BB. */
721 FOR_EACH_EDGE (e, ei, bb->succs)
722 if (e->dest == return_bb)
724 for (bsi = gsi_start_phis (return_bb); !gsi_end_p (bsi); gsi_next (&bsi))
726 gimple stmt = gsi_stmt (bsi);
727 tree op = gimple_phi_arg_def (stmt, e->dest_idx);
729 if (is_gimple_debug (stmt))
731 if (!is_gimple_reg (gimple_phi_result (stmt)))
733 if (TREE_CODE (op) == SSA_NAME)
734 bitmap_set_bit (used_ssa_names, SSA_NAME_VERSION (op));
736 can_split &= !mark_nonssa_use (stmt, op, non_ssa_vars);
742 /* Stack entry for recursive DFS walk in find_split_point. */
746 /* Basic block we are examining. */
749 /* SSA names set and used by the BB and all BBs reachable
750 from it via DFS walk. */
751 bitmap set_ssa_names, used_ssa_names;
754 /* All BBS visited from this BB via DFS walk. */
757 /* Last examined edge in DFS walk. Since we walk unoriented graph,
758 the value is up to sum of incoming and outgoing edges of BB. */
759 unsigned int edge_num;
761 /* Stack entry index of earliest BB reachable from current BB
762 or any BB visited later in DFS walk. */
765 /* Overall time and size of all BBs reached from this BB in DFS walk. */
766 int overall_time, overall_size;
768 /* When false we can not split on this BB. */
771 DEF_VEC_O(stack_entry);
772 DEF_VEC_ALLOC_O(stack_entry,heap);
775 /* Find all articulations and call consider_split on them.
776 OVERALL_TIME and OVERALL_SIZE is time and size of the function.
778 We perform basic algorithm for finding an articulation in a graph
779 created from CFG by considering it to be an unoriented graph.
781 The articulation is discovered via DFS walk. We collect earliest
782 basic block on stack that is reachable via backward edge. Articulation
783 is any basic block such that there is no backward edge bypassing it.
784 To reduce stack usage we maintain heap allocated stack in STACK vector.
785 AUX pointer of BB is set to index it appears in the stack or -1 once
786 it is visited and popped off the stack.
788 The algorithm finds articulation after visiting the whole component
789 reachable by it. This makes it convenient to collect information about
790 the component used by consider_split. */
793 find_split_points (int overall_time, int overall_size)
796 VEC(stack_entry, heap) *stack = NULL;
798 basic_block return_bb = find_return_bb ();
799 struct split_point current;
801 current.header_time = overall_time;
802 current.header_size = overall_size;
803 current.split_time = 0;
804 current.split_size = 0;
805 current.ssa_names_to_pass = BITMAP_ALLOC (NULL);
807 first.bb = ENTRY_BLOCK_PTR;
809 first.overall_time = 0;
810 first.overall_size = 0;
811 first.earliest = INT_MAX;
812 first.set_ssa_names = 0;
813 first.used_ssa_names = 0;
814 first.bbs_visited = 0;
815 VEC_safe_push (stack_entry, heap, stack, &first);
816 ENTRY_BLOCK_PTR->aux = (void *)(intptr_t)-1;
818 while (!VEC_empty (stack_entry, stack))
820 stack_entry *entry = VEC_last (stack_entry, stack);
822 /* We are walking an acyclic graph, so edge_num counts
823 succ and pred edges together. However when considering
824 articulation, we want to have processed everything reachable
825 from articulation but nothing that reaches into it. */
826 if (entry->edge_num == EDGE_COUNT (entry->bb->succs)
827 && entry->bb != ENTRY_BLOCK_PTR)
829 int pos = VEC_length (stack_entry, stack);
830 entry->can_split &= visit_bb (entry->bb, return_bb,
831 entry->set_ssa_names,
832 entry->used_ssa_names,
833 entry->non_ssa_vars);
834 if (pos <= entry->earliest && !entry->can_split
835 && dump_file && (dump_flags & TDF_DETAILS))
837 "found articulation at bb %i but can not split\n",
839 if (pos <= entry->earliest && entry->can_split)
841 if (dump_file && (dump_flags & TDF_DETAILS))
842 fprintf (dump_file, "found articulation at bb %i\n",
844 current.entry_bb = entry->bb;
845 current.ssa_names_to_pass = BITMAP_ALLOC (NULL);
846 bitmap_and_compl (current.ssa_names_to_pass,
847 entry->used_ssa_names, entry->set_ssa_names);
848 current.header_time = overall_time - entry->overall_time;
849 current.header_size = overall_size - entry->overall_size;
850 current.split_time = entry->overall_time;
851 current.split_size = entry->overall_size;
852 current.split_bbs = entry->bbs_visited;
853 consider_split (¤t, entry->non_ssa_vars, return_bb);
854 BITMAP_FREE (current.ssa_names_to_pass);
857 /* Do actual DFS walk. */
859 < (EDGE_COUNT (entry->bb->succs)
860 + EDGE_COUNT (entry->bb->preds)))
864 if (entry->edge_num < EDGE_COUNT (entry->bb->succs))
866 e = EDGE_SUCC (entry->bb, entry->edge_num);
871 e = EDGE_PRED (entry->bb, entry->edge_num
872 - EDGE_COUNT (entry->bb->succs));
878 /* New BB to visit, push it to the stack. */
879 if (dest != return_bb && dest != EXIT_BLOCK_PTR
882 stack_entry new_entry;
885 new_entry.edge_num = 0;
886 new_entry.overall_time
887 = VEC_index (bb_info, bb_info_vec, dest->index)->time;
888 new_entry.overall_size
889 = VEC_index (bb_info, bb_info_vec, dest->index)->size;
890 new_entry.earliest = INT_MAX;
891 new_entry.set_ssa_names = BITMAP_ALLOC (NULL);
892 new_entry.used_ssa_names = BITMAP_ALLOC (NULL);
893 new_entry.bbs_visited = BITMAP_ALLOC (NULL);
894 new_entry.non_ssa_vars = BITMAP_ALLOC (NULL);
895 new_entry.can_split = true;
896 bitmap_set_bit (new_entry.bbs_visited, dest->index);
897 VEC_safe_push (stack_entry, heap, stack, &new_entry);
898 dest->aux = (void *)(intptr_t)VEC_length (stack_entry, stack);
900 /* Back edge found, record the earliest point. */
901 else if ((intptr_t)dest->aux > 0
902 && (intptr_t)dest->aux < entry->earliest)
903 entry->earliest = (intptr_t)dest->aux;
905 /* We are done with examining the edges. Pop off the value from stack
906 and merge stuff we accumulate during the walk. */
907 else if (entry->bb != ENTRY_BLOCK_PTR)
909 stack_entry *prev = VEC_index (stack_entry, stack,
910 VEC_length (stack_entry, stack) - 2);
912 entry->bb->aux = (void *)(intptr_t)-1;
913 prev->can_split &= entry->can_split;
914 if (prev->set_ssa_names)
916 bitmap_ior_into (prev->set_ssa_names, entry->set_ssa_names);
917 bitmap_ior_into (prev->used_ssa_names, entry->used_ssa_names);
918 bitmap_ior_into (prev->bbs_visited, entry->bbs_visited);
919 bitmap_ior_into (prev->non_ssa_vars, entry->non_ssa_vars);
921 if (prev->earliest > entry->earliest)
922 prev->earliest = entry->earliest;
923 prev->overall_time += entry->overall_time;
924 prev->overall_size += entry->overall_size;
925 BITMAP_FREE (entry->set_ssa_names);
926 BITMAP_FREE (entry->used_ssa_names);
927 BITMAP_FREE (entry->bbs_visited);
928 BITMAP_FREE (entry->non_ssa_vars);
929 VEC_pop (stack_entry, stack);
932 VEC_pop (stack_entry, stack);
934 ENTRY_BLOCK_PTR->aux = NULL;
937 VEC_free (stack_entry, heap, stack);
938 BITMAP_FREE (current.ssa_names_to_pass);
941 /* Split function at SPLIT_POINT. */
944 split_function (struct split_point *split_point)
946 VEC (tree, heap) *args_to_pass = NULL;
947 bitmap args_to_skip = BITMAP_ALLOC (NULL);
950 struct cgraph_node *node;
951 basic_block return_bb = find_return_bb ();
953 gimple_stmt_iterator gsi;
957 tree retval = NULL, real_retval = NULL;
958 bool split_part_return_p = false;
959 gimple last_stmt = NULL;
960 bool conv_needed = false;
966 fprintf (dump_file, "\n\nSplitting function at:\n");
967 dump_split_point (dump_file, split_point);
970 /* Collect the parameters of new function and args_to_skip bitmap. */
971 for (parm = DECL_ARGUMENTS (current_function_decl);
972 parm; parm = DECL_CHAIN (parm), num++)
973 if (!is_gimple_reg (parm)
974 || !gimple_default_def (cfun, parm)
975 || !bitmap_bit_p (split_point->ssa_names_to_pass,
976 SSA_NAME_VERSION (gimple_default_def (cfun, parm))))
977 bitmap_set_bit (args_to_skip, num);
980 arg = gimple_default_def (cfun, parm);
981 if (TYPE_MAIN_VARIANT (DECL_ARG_TYPE (parm))
982 != TYPE_MAIN_VARIANT (TREE_TYPE (arg)))
985 arg = fold_convert (DECL_ARG_TYPE (parm), arg);
987 VEC_safe_push (tree, heap, args_to_pass, arg);
990 /* See if the split function will return. */
991 FOR_EACH_EDGE (e, ei, return_bb->preds)
992 if (bitmap_bit_p (split_point->split_bbs, e->src->index))
995 split_part_return_p = true;
997 /* Add return block to what will become the split function.
998 We do not return; no return block is needed. */
999 if (!split_part_return_p)
1001 /* We have no return block, so nothing is needed. */
1002 else if (return_bb == EXIT_BLOCK_PTR)
1004 /* When we do not want to return value, we need to construct
1005 new return block with empty return statement.
1006 FIXME: Once we are able to change return type, we should change function
1007 to return void instead of just outputting function with undefined return
1008 value. For structures this affects quality of codegen. */
1009 else if (!split_point->split_part_set_retval
1010 && find_retval (return_bb))
1012 bool redirected = true;
1013 basic_block new_return_bb = create_basic_block (NULL, 0, return_bb);
1014 gimple_stmt_iterator gsi = gsi_start_bb (new_return_bb);
1015 gsi_insert_after (&gsi, gimple_build_return (NULL), GSI_NEW_STMT);
1019 FOR_EACH_EDGE (e, ei, return_bb->preds)
1020 if (bitmap_bit_p (split_point->split_bbs, e->src->index))
1022 new_return_bb->count += e->count;
1023 new_return_bb->frequency += EDGE_FREQUENCY (e);
1024 redirect_edge_and_branch (e, new_return_bb);
1029 e = make_edge (new_return_bb, EXIT_BLOCK_PTR, 0);
1030 e->probability = REG_BR_PROB_BASE;
1031 e->count = new_return_bb->count;
1032 bitmap_set_bit (split_point->split_bbs, new_return_bb->index);
1034 /* When we pass around the value, use existing return block. */
1036 bitmap_set_bit (split_point->split_bbs, return_bb->index);
1038 /* If RETURN_BB has virtual operand PHIs, they must be removed and the
1039 virtual operand marked for renaming as we change the CFG in a way that
1040 tree-inline is not able to compensate for.
1042 Note this can happen whether or not we have a return value. If we have
1043 a return value, then RETURN_BB may have PHIs for real operands too. */
1044 if (return_bb != EXIT_BLOCK_PTR)
1046 for (gsi = gsi_start_phis (return_bb); !gsi_end_p (gsi);)
1048 gimple stmt = gsi_stmt (gsi);
1049 if (is_gimple_reg (gimple_phi_result (stmt)))
1054 mark_virtual_phi_result_for_renaming (stmt);
1055 remove_phi_node (&gsi, true);
1059 /* Now create the actual clone. */
1060 rebuild_cgraph_edges ();
1061 node = cgraph_function_versioning (cgraph_node (current_function_decl),
1064 split_point->split_bbs,
1065 split_point->entry_bb, "part");
1066 /* For usual cloning it is enough to clear builtin only when signature
1067 changes. For partial inlining we however can not expect the part
1068 of builtin implementation to have same semantic as the whole. */
1069 if (DECL_BUILT_IN (node->decl))
1071 DECL_BUILT_IN_CLASS (node->decl) = NOT_BUILT_IN;
1072 DECL_FUNCTION_CODE (node->decl) = (enum built_in_function) 0;
1074 cgraph_node_remove_callees (cgraph_node (current_function_decl));
1075 if (!split_part_return_p)
1076 TREE_THIS_VOLATILE (node->decl) = 1;
1078 dump_function_to_file (node->decl, dump_file, dump_flags);
1080 /* Create the basic block we place call into. It is the entry basic block
1081 split after last label. */
1082 call_bb = split_point->entry_bb;
1083 for (gsi = gsi_start_bb (call_bb); !gsi_end_p (gsi);)
1084 if (gimple_code (gsi_stmt (gsi)) == GIMPLE_LABEL)
1086 last_stmt = gsi_stmt (gsi);
1091 e = split_block (split_point->entry_bb, last_stmt);
1094 /* Produce the call statement. */
1095 gsi = gsi_last_bb (call_bb);
1097 FOR_EACH_VEC_ELT (tree, args_to_pass, i, arg)
1098 if (!is_gimple_val (arg))
1100 arg = force_gimple_operand_gsi (&gsi, arg, true, NULL_TREE,
1101 false, GSI_NEW_STMT);
1102 VEC_replace (tree, args_to_pass, i, arg);
1104 call = gimple_build_call_vec (node->decl, args_to_pass);
1105 gimple_set_block (call, DECL_INITIAL (current_function_decl));
1107 /* We avoid address being taken on any variable used by split part,
1108 so return slot optimization is always possible. Moreover this is
1109 required to make DECL_BY_REFERENCE work. */
1110 if (aggregate_value_p (DECL_RESULT (current_function_decl),
1111 TREE_TYPE (current_function_decl)))
1112 gimple_call_set_return_slot_opt (call, true);
1114 /* Update return value. This is bit tricky. When we do not return,
1115 do nothing. When we return we might need to update return_bb
1116 or produce a new return statement. */
1117 if (!split_part_return_p)
1118 gsi_insert_after (&gsi, call, GSI_NEW_STMT);
1121 e = make_edge (call_bb, return_bb,
1122 return_bb == EXIT_BLOCK_PTR ? 0 : EDGE_FALLTHRU);
1123 e->count = call_bb->count;
1124 e->probability = REG_BR_PROB_BASE;
1126 /* If there is return basic block, see what value we need to store
1127 return value into and put call just before it. */
1128 if (return_bb != EXIT_BLOCK_PTR)
1130 real_retval = retval = find_retval (return_bb);
1132 if (real_retval && split_point->split_part_set_retval)
1134 gimple_stmt_iterator psi;
1136 /* See if we need new SSA_NAME for the result.
1137 When DECL_BY_REFERENCE is true, retval is actually pointer to
1138 return value and it is constant in whole function. */
1139 if (TREE_CODE (retval) == SSA_NAME
1140 && !DECL_BY_REFERENCE (DECL_RESULT (current_function_decl)))
1142 retval = make_ssa_name (SSA_NAME_VAR (retval), call);
1144 /* See if there is PHI defining return value. */
1145 for (psi = gsi_start_phis (return_bb);
1146 !gsi_end_p (psi); gsi_next (&psi))
1147 if (is_gimple_reg (gimple_phi_result (gsi_stmt (psi))))
1150 /* When there is PHI, just update its value. */
1151 if (TREE_CODE (retval) == SSA_NAME
1152 && !gsi_end_p (psi))
1153 add_phi_arg (gsi_stmt (psi), retval, e, UNKNOWN_LOCATION);
1154 /* Otherwise update the return BB itself.
1155 find_return_bb allows at most one assignment to return value,
1156 so update first statement. */
1159 gimple_stmt_iterator bsi;
1160 for (bsi = gsi_start_bb (return_bb); !gsi_end_p (bsi);
1162 if (gimple_code (gsi_stmt (bsi)) == GIMPLE_RETURN)
1164 gimple_return_set_retval (gsi_stmt (bsi), retval);
1167 else if (gimple_code (gsi_stmt (bsi)) == GIMPLE_ASSIGN)
1169 gimple_assign_set_rhs1 (gsi_stmt (bsi), retval);
1172 update_stmt (gsi_stmt (bsi));
1175 if (DECL_BY_REFERENCE (DECL_RESULT (current_function_decl)))
1176 gimple_call_set_lhs (call, build_simple_mem_ref (retval));
1178 gimple_call_set_lhs (call, retval);
1180 gsi_insert_after (&gsi, call, GSI_NEW_STMT);
1182 /* We don't use return block (there is either no return in function or
1183 multiple of them). So create new basic block with return statement.
1188 if (split_point->split_part_set_retval
1189 && !VOID_TYPE_P (TREE_TYPE (TREE_TYPE (current_function_decl))))
1191 retval = DECL_RESULT (current_function_decl);
1193 /* We use temporary register to hold value when aggregate_value_p
1194 is false. Similarly for DECL_BY_REFERENCE we must avoid extra
1196 if (!aggregate_value_p (retval, TREE_TYPE (current_function_decl))
1197 && !DECL_BY_REFERENCE (retval))
1198 retval = create_tmp_reg (TREE_TYPE (retval), NULL);
1199 if (is_gimple_reg (retval))
1201 /* When returning by reference, there is only one SSA name
1202 assigned to RESULT_DECL (that is pointer to return value).
1203 Look it up or create new one if it is missing. */
1204 if (DECL_BY_REFERENCE (retval))
1207 if ((retval_name = gimple_default_def (cfun, retval))
1209 retval = retval_name;
1212 retval_name = make_ssa_name (retval,
1213 gimple_build_nop ());
1214 set_default_def (retval, retval_name);
1215 retval = retval_name;
1218 /* Otherwise produce new SSA name for return value. */
1220 retval = make_ssa_name (retval, call);
1222 if (DECL_BY_REFERENCE (DECL_RESULT (current_function_decl)))
1223 gimple_call_set_lhs (call, build_simple_mem_ref (retval));
1225 gimple_call_set_lhs (call, retval);
1227 gsi_insert_after (&gsi, call, GSI_NEW_STMT);
1228 ret = gimple_build_return (retval);
1229 gsi_insert_after (&gsi, ret, GSI_NEW_STMT);
1232 free_dominance_info (CDI_DOMINATORS);
1233 free_dominance_info (CDI_POST_DOMINATORS);
1234 compute_inline_parameters (node);
1237 /* Execute function splitting pass. */
1240 execute_split_functions (void)
1242 gimple_stmt_iterator bsi;
1244 int overall_time = 0, overall_size = 0;
1246 struct cgraph_node *node = cgraph_node (current_function_decl);
1248 if (flags_from_decl_or_type (current_function_decl) & ECF_NORETURN)
1251 fprintf (dump_file, "Not splitting: noreturn function.\n");
1254 if (MAIN_NAME_P (DECL_NAME (current_function_decl)))
1257 fprintf (dump_file, "Not splitting: main function.\n");
1260 /* This can be relaxed; function might become inlinable after splitting
1261 away the uninlinable part. */
1262 if (!node->local.inlinable)
1265 fprintf (dump_file, "Not splitting: not inlinable.\n");
1268 if (node->local.disregard_inline_limits)
1271 fprintf (dump_file, "Not splitting: disregarding inline limits.\n");
1274 /* This can be relaxed; most of versioning tests actually prevents
1276 if (!tree_versionable_function_p (current_function_decl))
1279 fprintf (dump_file, "Not splitting: not versionable.\n");
1282 /* FIXME: we could support this. */
1283 if (DECL_STRUCT_FUNCTION (current_function_decl)->static_chain_decl)
1286 fprintf (dump_file, "Not splitting: nested function.\n");
1290 /* See if it makes sense to try to split.
1291 It makes sense to split if we inline, that is if we have direct calls to
1292 handle or direct calls are possibly going to appear as result of indirect
1293 inlining or LTO. Also handle -fprofile-generate as LTO to allow non-LTO
1294 training for LTO -fprofile-use build.
1296 Note that we are not completely conservative about disqualifying functions
1297 called once. It is possible that the caller is called more then once and
1298 then inlining would still benefit. */
1299 if ((!node->callers || !node->callers->next_caller)
1300 && !node->address_taken
1301 && (!flag_lto || !node->local.externally_visible))
1304 fprintf (dump_file, "Not splitting: not called directly "
1305 "or called once.\n");
1309 /* FIXME: We can actually split if splitting reduces call overhead. */
1310 if (!flag_inline_small_functions
1311 && !DECL_DECLARED_INLINE_P (current_function_decl))
1314 fprintf (dump_file, "Not splitting: not autoinlining and function"
1315 " is not inline.\n");
1319 /* Compute local info about basic blocks and determine function size/time. */
1320 VEC_safe_grow_cleared (bb_info, heap, bb_info_vec, last_basic_block + 1);
1321 memset (&best_split_point, 0, sizeof (best_split_point));
1326 int freq = compute_call_stmt_bb_frequency (current_function_decl, bb);
1328 if (dump_file && (dump_flags & TDF_DETAILS))
1329 fprintf (dump_file, "Basic block %i\n", bb->index);
1331 for (bsi = gsi_start_bb (bb); !gsi_end_p (bsi); gsi_next (&bsi))
1333 int this_time, this_size;
1334 gimple stmt = gsi_stmt (bsi);
1336 this_size = estimate_num_insns (stmt, &eni_size_weights);
1337 this_time = estimate_num_insns (stmt, &eni_time_weights) * freq;
1341 if (dump_file && (dump_flags & TDF_DETAILS))
1343 fprintf (dump_file, " freq:%6i size:%3i time:%3i ",
1344 freq, this_size, this_time);
1345 print_gimple_stmt (dump_file, stmt, 0, 0);
1348 overall_time += time;
1349 overall_size += size;
1350 VEC_index (bb_info, bb_info_vec, bb->index)->time = time;
1351 VEC_index (bb_info, bb_info_vec, bb->index)->size = size;
1353 find_split_points (overall_time, overall_size);
1354 if (best_split_point.split_bbs)
1356 split_function (&best_split_point);
1357 BITMAP_FREE (best_split_point.ssa_names_to_pass);
1358 BITMAP_FREE (best_split_point.split_bbs);
1359 todo = TODO_update_ssa | TODO_cleanup_cfg;
1361 VEC_free (bb_info, heap, bb_info_vec);
1366 /* Gate function splitting pass. When doing profile feedback, we want
1367 to execute the pass after profiling is read. So disable one in
1368 early optimization. */
1371 gate_split_functions (void)
1373 return (flag_partial_inlining
1374 && !profile_arc_flag && !flag_branch_probabilities);
1377 struct gimple_opt_pass pass_split_functions =
1381 "fnsplit", /* name */
1382 gate_split_functions, /* gate */
1383 execute_split_functions, /* execute */
1386 0, /* static_pass_number */
1387 TV_IPA_FNSPLIT, /* tv_id */
1388 PROP_cfg, /* properties_required */
1389 0, /* properties_provided */
1390 0, /* properties_destroyed */
1391 0, /* todo_flags_start */
1392 TODO_dump_func /* todo_flags_finish */
1396 /* Gate feedback driven function splitting pass.
1397 We don't need to split when profiling at all, we are producing
1398 lousy code anyway. */
1401 gate_feedback_split_functions (void)
1403 return (flag_partial_inlining
1404 && flag_branch_probabilities);
1407 /* Execute function splitting pass. */
1410 execute_feedback_split_functions (void)
1412 unsigned int retval = execute_split_functions ();
1414 retval |= TODO_rebuild_cgraph_edges;
1418 struct gimple_opt_pass pass_feedback_split_functions =
1422 "feedback_fnsplit", /* name */
1423 gate_feedback_split_functions, /* gate */
1424 execute_feedback_split_functions, /* execute */
1427 0, /* static_pass_number */
1428 TV_IPA_FNSPLIT, /* tv_id */
1429 PROP_cfg, /* properties_required */
1430 0, /* properties_provided */
1431 0, /* properties_destroyed */
1432 0, /* todo_flags_start */
1433 TODO_dump_func /* todo_flags_finish */