1 /* Inlining decision heuristics.
2 Copyright (C) 2003, 2004, 2007, 2008, 2009, 2010
3 Free Software Foundation, Inc.
4 Contributed by Jan Hubicka
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 /* Inlining decision heuristics
24 We separate inlining decisions from the inliner itself and store it
25 inside callgraph as so called inline plan. Refer to cgraph.c
26 documentation about particular representation of inline plans in the
29 There are three major parts of this file:
31 cgraph_mark_inline implementation
33 This function allows to mark given call inline and performs necessary
34 modifications of cgraph (production of the clones and updating overall
37 inlining heuristics limits
39 These functions allow to check that particular inlining is allowed
40 by the limits specified by user (allowed function growth, overall unit
45 This is implementation of IPA pass aiming to get as much of benefit
46 from inlining obeying the limits checked above.
48 The implementation of particular heuristics is separated from
49 the rest of code to make it easier to replace it with more complicated
50 implementation in the future. The rest of inlining code acts as a
51 library aimed to modify the callgraph and verify that the parameters
52 on code size growth fits.
54 To mark given call inline, use cgraph_mark_inline function, the
55 verification is performed by cgraph_default_inline_p and
56 cgraph_check_inline_limits.
58 The heuristics implements simple knapsack style algorithm ordering
59 all functions by their "profitability" (estimated by code size growth)
60 and inlining them in priority order.
62 cgraph_decide_inlining implements heuristics taking whole callgraph
63 into account, while cgraph_decide_inlining_incrementally considers
64 only one function at a time and is used by early inliner.
66 The inliner itself is split into several passes:
68 pass_inline_parameters
70 This pass computes local properties of functions that are used by inliner:
71 estimated function body size, whether function is inlinable at all and
72 stack frame consumption.
74 Before executing any of inliner passes, this local pass has to be applied
75 to each function in the callgraph (ie run as subpass of some earlier
76 IPA pass). The results are made out of date by any optimization applied
81 Simple local inlining pass inlining callees into current function. This
82 pass makes no global whole compilation unit analysis and this when allowed
83 to do inlining expanding code size it might result in unbounded growth of
86 The pass is run during conversion into SSA form. Only functions already
87 converted into SSA form are inlined, so the conversion must happen in
88 topological order on the callgraph (that is maintained by pass manager).
89 The functions after inlining are early optimized so the early inliner sees
90 unoptimized function itself, but all considered callees are already
91 optimized allowing it to unfold abstraction penalty on C++ effectively and
94 pass_ipa_early_inlining
96 With profiling, the early inlining is also necessary to reduce
97 instrumentation costs on program with high abstraction penalty (doing
98 many redundant calls). This can't happen in parallel with early
99 optimization and profile instrumentation, because we would end up
100 re-instrumenting already instrumented function bodies we brought in via
103 To avoid this, this pass is executed as IPA pass before profiling. It is
104 simple wrapper to pass_early_inlining and ensures first inlining.
108 This is the main pass implementing simple greedy algorithm to do inlining
109 of small functions that results in overall growth of compilation unit and
110 inlining of functions called once. The pass compute just so called inline
111 plan (representation of inlining to be done in callgraph) and unlike early
112 inlining it is not performing the inlining itself.
116 This pass performs actual inlining according to pass_ipa_inline on given
117 function. Possible the function body before inlining is saved when it is
118 needed for further inlining later.
123 #include "coretypes.h"
126 #include "tree-inline.h"
127 #include "langhooks.h"
130 #include "diagnostic.h"
131 #include "gimple-pretty-print.h"
136 #include "tree-pass.h"
138 #include "coverage.h"
140 #include "tree-flow.h"
142 #include "ipa-prop.h"
145 #define MAX_TIME 1000000000
147 /* Mode incremental inliner operate on:
149 In ALWAYS_INLINE only functions marked
150 always_inline are inlined. This mode is used after detecting cycle during
153 In SIZE mode, only functions that reduce function body size after inlining
154 are inlined, this is used during early inlining.
156 in ALL mode, everything is inlined. This is used during flattening. */
159 INLINE_ALWAYS_INLINE,
160 INLINE_SIZE_NORECURSIVE,
166 cgraph_decide_inlining_incrementally (struct cgraph_node *, enum inlining_mode);
167 static void cgraph_flatten (struct cgraph_node *node);
170 /* Statistics we collect about inlining algorithm. */
171 static int ncalls_inlined;
172 static int nfunctions_inlined;
173 static int overall_size;
174 static gcov_type max_count, max_benefit;
176 /* Holders of ipa cgraph hooks: */
177 static struct cgraph_node_hook_list *function_insertion_hook_holder;
179 static inline struct inline_summary *
180 inline_summary (struct cgraph_node *node)
182 return &node->local.inline_summary;
185 /* Estimate self time of the function after inlining WHAT into TO. */
188 cgraph_estimate_time_after_inlining (int frequency, struct cgraph_node *to,
189 struct cgraph_node *what)
191 gcov_type time = (((gcov_type)what->global.time
192 - inline_summary (what)->time_inlining_benefit)
193 * frequency + CGRAPH_FREQ_BASE / 2) / CGRAPH_FREQ_BASE
202 /* Estimate self time of the function after inlining WHAT into TO. */
205 cgraph_estimate_size_after_inlining (int times, struct cgraph_node *to,
206 struct cgraph_node *what)
208 int size = ((what->global.size - inline_summary (what)->size_inlining_benefit)
209 * times + to->global.size);
210 gcc_assert (size >= 0);
214 /* Scale frequency of NODE edges by FREQ_SCALE and increase loop nest
218 update_noncloned_frequencies (struct cgraph_node *node,
219 int freq_scale, int nest)
221 struct cgraph_edge *e;
223 /* We do not want to ignore high loop nest after freq drops to 0. */
226 for (e = node->callees; e; e = e->next_callee)
228 e->loop_nest += nest;
229 e->frequency = e->frequency * (gcov_type) freq_scale / CGRAPH_FREQ_BASE;
230 if (e->frequency > CGRAPH_FREQ_MAX)
231 e->frequency = CGRAPH_FREQ_MAX;
232 if (!e->inline_failed)
233 update_noncloned_frequencies (e->callee, freq_scale, nest);
237 /* E is expected to be an edge being inlined. Clone destination node of
238 the edge and redirect it to the new clone.
239 DUPLICATE is used for bookkeeping on whether we are actually creating new
240 clones or re-using node originally representing out-of-line function call.
243 cgraph_clone_inlined_nodes (struct cgraph_edge *e, bool duplicate,
244 bool update_original)
250 /* We may eliminate the need for out-of-line copy to be output.
251 In that case just go ahead and re-use it. */
252 if (!e->callee->callers->next_caller
253 && cgraph_can_remove_if_no_direct_calls_p (e->callee)
254 /* Don't reuse if more than one function shares a comdat group.
255 If the other function(s) are needed, we need to emit even
256 this function out of line. */
257 && !e->callee->same_comdat_group
258 && !cgraph_new_nodes)
260 gcc_assert (!e->callee->global.inlined_to);
261 if (e->callee->analyzed)
263 overall_size -= e->callee->global.size;
264 nfunctions_inlined++;
267 e->callee->local.externally_visible = false;
268 update_noncloned_frequencies (e->callee, e->frequency, e->loop_nest);
272 struct cgraph_node *n;
273 n = cgraph_clone_node (e->callee, e->callee->decl,
274 e->count, e->frequency, e->loop_nest,
275 update_original, NULL);
276 cgraph_redirect_edge_callee (e, n);
280 if (e->caller->global.inlined_to)
281 e->callee->global.inlined_to = e->caller->global.inlined_to;
283 e->callee->global.inlined_to = e->caller;
284 e->callee->global.stack_frame_offset
285 = e->caller->global.stack_frame_offset
286 + inline_summary (e->caller)->estimated_self_stack_size;
287 peak = e->callee->global.stack_frame_offset
288 + inline_summary (e->callee)->estimated_self_stack_size;
289 if (e->callee->global.inlined_to->global.estimated_stack_size < peak)
290 e->callee->global.inlined_to->global.estimated_stack_size = peak;
291 cgraph_propagate_frequency (e->callee);
293 /* Recursively clone all bodies. */
294 for (e = e->callee->callees; e; e = e->next_callee)
295 if (!e->inline_failed)
296 cgraph_clone_inlined_nodes (e, duplicate, update_original);
299 /* Mark edge E as inlined and update callgraph accordingly. UPDATE_ORIGINAL
300 specify whether profile of original function should be updated. If any new
301 indirect edges are discovered in the process, add them to NEW_EDGES, unless
302 it is NULL. Return true iff any new callgraph edges were discovered as a
303 result of inlining. */
306 cgraph_mark_inline_edge (struct cgraph_edge *e, bool update_original,
307 VEC (cgraph_edge_p, heap) **new_edges)
309 int old_size = 0, new_size = 0;
310 struct cgraph_node *to = NULL, *what;
311 struct cgraph_edge *curr = e;
314 gcc_assert (e->inline_failed);
315 e->inline_failed = CIF_OK;
316 DECL_POSSIBLY_INLINED (e->callee->decl) = true;
318 cgraph_clone_inlined_nodes (e, true, update_original);
323 /* Now update size of caller and all functions caller is inlined into. */
324 for (;e && !e->inline_failed; e = e->caller->callers)
327 old_size = e->caller->global.size;
328 new_size = cgraph_estimate_size_after_inlining (1, to, what);
329 to->global.size = new_size;
330 to->global.time = cgraph_estimate_time_after_inlining (freq, to, what);
332 gcc_assert (what->global.inlined_to == to);
333 if (new_size > old_size)
334 overall_size += new_size - old_size;
337 if (flag_indirect_inlining)
338 return ipa_propagate_indirect_call_infos (curr, new_edges);
343 /* Mark all calls of EDGE->CALLEE inlined into EDGE->CALLER. */
346 cgraph_mark_inline (struct cgraph_edge *edge)
348 struct cgraph_node *to = edge->caller;
349 struct cgraph_node *what = edge->callee;
350 struct cgraph_edge *e, *next;
352 gcc_assert (!edge->call_stmt_cannot_inline_p);
353 /* Look for all calls, mark them inline and clone recursively
354 all inlined functions. */
355 for (e = what->callers; e; e = next)
357 next = e->next_caller;
358 if (e->caller == to && e->inline_failed)
360 cgraph_mark_inline_edge (e, true, NULL);
367 /* Estimate the growth caused by inlining NODE into all callees. */
370 cgraph_estimate_growth (struct cgraph_node *node)
373 struct cgraph_edge *e;
374 bool self_recursive = false;
376 if (node->global.estimated_growth != INT_MIN)
377 return node->global.estimated_growth;
379 for (e = node->callers; e; e = e->next_caller)
381 if (e->caller == node)
382 self_recursive = true;
383 if (e->inline_failed)
384 growth += (cgraph_estimate_size_after_inlining (1, e->caller, node)
385 - e->caller->global.size);
388 /* ??? Wrong for non-trivially self recursive functions or cases where
389 we decide to not inline for different reasons, but it is not big deal
390 as in that case we will keep the body around, but we will also avoid
392 if (cgraph_will_be_removed_from_program_if_no_direct_calls (node)
393 && !DECL_EXTERNAL (node->decl) && !self_recursive)
394 growth -= node->global.size;
396 node->global.estimated_growth = growth;
400 /* Return false when inlining WHAT into TO is not good idea
401 as it would cause too large growth of function bodies.
402 When ONE_ONLY is true, assume that only one call site is going
403 to be inlined, otherwise figure out how many call sites in
404 TO calls WHAT and verify that all can be inlined.
408 cgraph_check_inline_limits (struct cgraph_node *to, struct cgraph_node *what,
409 cgraph_inline_failed_t *reason, bool one_only)
412 struct cgraph_edge *e;
415 HOST_WIDE_INT stack_size_limit, inlined_stack;
420 for (e = to->callees; e; e = e->next_callee)
421 if (e->callee == what)
424 if (to->global.inlined_to)
425 to = to->global.inlined_to;
427 /* When inlining large function body called once into small function,
428 take the inlined function as base for limiting the growth. */
429 if (inline_summary (to)->self_size > inline_summary(what)->self_size)
430 limit = inline_summary (to)->self_size;
432 limit = inline_summary (what)->self_size;
434 limit += limit * PARAM_VALUE (PARAM_LARGE_FUNCTION_GROWTH) / 100;
436 /* Check the size after inlining against the function limits. But allow
437 the function to shrink if it went over the limits by forced inlining. */
438 newsize = cgraph_estimate_size_after_inlining (times, to, what);
439 if (newsize >= to->global.size
440 && newsize > PARAM_VALUE (PARAM_LARGE_FUNCTION_INSNS)
444 *reason = CIF_LARGE_FUNCTION_GROWTH_LIMIT;
448 stack_size_limit = inline_summary (to)->estimated_self_stack_size;
450 stack_size_limit += stack_size_limit * PARAM_VALUE (PARAM_STACK_FRAME_GROWTH) / 100;
452 inlined_stack = (to->global.stack_frame_offset
453 + inline_summary (to)->estimated_self_stack_size
454 + what->global.estimated_stack_size);
455 if (inlined_stack > stack_size_limit
456 && inlined_stack > PARAM_VALUE (PARAM_LARGE_STACK_FRAME))
459 *reason = CIF_LARGE_STACK_FRAME_GROWTH_LIMIT;
465 /* Return true when function N is small enough to be inlined. */
468 cgraph_default_inline_p (struct cgraph_node *n, cgraph_inline_failed_t *reason)
472 if (n->local.disregard_inline_limits)
475 if (!flag_inline_small_functions && !DECL_DECLARED_INLINE_P (decl))
478 *reason = CIF_FUNCTION_NOT_INLINE_CANDIDATE;
485 *reason = CIF_BODY_NOT_AVAILABLE;
489 if (DECL_DECLARED_INLINE_P (decl))
491 if (n->global.size >= MAX_INLINE_INSNS_SINGLE)
494 *reason = CIF_MAX_INLINE_INSNS_SINGLE_LIMIT;
500 if (n->global.size >= MAX_INLINE_INSNS_AUTO)
503 *reason = CIF_MAX_INLINE_INSNS_AUTO_LIMIT;
511 /* Return true when inlining WHAT would create recursive inlining.
512 We call recursive inlining all cases where same function appears more than
513 once in the single recursion nest path in the inline graph. */
516 cgraph_recursive_inlining_p (struct cgraph_node *to,
517 struct cgraph_node *what,
518 cgraph_inline_failed_t *reason)
521 if (to->global.inlined_to)
522 recursive = what->decl == to->global.inlined_to->decl;
524 recursive = what->decl == to->decl;
525 /* Marking recursive function inline has sane semantic and thus we should
527 if (recursive && reason)
528 *reason = (what->local.disregard_inline_limits
529 ? CIF_RECURSIVE_INLINING : CIF_UNSPECIFIED);
533 /* A cost model driving the inlining heuristics in a way so the edges with
534 smallest badness are inlined first. After each inlining is performed
535 the costs of all caller edges of nodes affected are recomputed so the
536 metrics may accurately depend on values such as number of inlinable callers
537 of the function or function body size. */
540 cgraph_edge_badness (struct cgraph_edge *edge, bool dump)
544 (cgraph_estimate_size_after_inlining (1, edge->caller, edge->callee)
545 - edge->caller->global.size);
547 if (edge->callee->local.disregard_inline_limits)
552 fprintf (dump_file, " Badness calculcation for %s -> %s\n",
553 cgraph_node_name (edge->caller),
554 cgraph_node_name (edge->callee));
555 fprintf (dump_file, " growth %i, time %i-%i, size %i-%i\n",
557 edge->callee->global.time,
558 inline_summary (edge->callee)->time_inlining_benefit,
559 edge->callee->global.size,
560 inline_summary (edge->callee)->size_inlining_benefit);
563 /* Always prefer inlining saving code size. */
566 badness = INT_MIN - growth;
568 fprintf (dump_file, " %i: Growth %i < 0\n", (int) badness,
572 /* When profiling is available, base priorities -(#calls / growth).
573 So we optimize for overall number of "executed" inlined calls. */
578 ((double) edge->count * INT_MIN / max_count / (max_benefit + 1)) *
579 (inline_summary (edge->callee)->time_inlining_benefit + 1)) / growth;
583 " %i (relative %f): profile info. Relative count %f"
584 " * Relative benefit %f\n",
585 (int) badness, (double) badness / INT_MIN,
586 (double) edge->count / max_count,
587 (double) (inline_summary (edge->callee)->
588 time_inlining_benefit + 1) / (max_benefit + 1));
592 /* When function local profile is available, base priorities on
593 growth / frequency, so we optimize for overall frequency of inlined
594 calls. This is not too accurate since while the call might be frequent
595 within function, the function itself is infrequent.
597 Other objective to optimize for is number of different calls inlined.
598 We add the estimated growth after inlining all functions to bias the
599 priorities slightly in this direction (so fewer times called functions
600 of the same size gets priority). */
601 else if (flag_guess_branch_prob)
603 int div = edge->frequency * 100 / CGRAPH_FREQ_BASE + 1;
606 badness = growth * 10000;
608 MIN (100 * inline_summary (edge->callee)->time_inlining_benefit /
609 (edge->callee->global.time + 1) +1, 100);
613 /* Decrease badness if call is nested. */
614 /* Compress the range so we don't overflow. */
616 div = 10000 + ceil_log2 (div) - 8;
621 growth_for_all = cgraph_estimate_growth (edge->callee);
622 badness += growth_for_all;
623 if (badness > INT_MAX)
628 " %i: guessed profile. frequency %i, overall growth %i,"
629 " benefit %i%%, divisor %i\n",
630 (int) badness, edge->frequency, growth_for_all, benefitperc, div);
633 /* When function local profile is not available or it does not give
634 useful information (ie frequency is zero), base the cost on
635 loop nest and overall size growth, so we optimize for overall number
636 of functions fully inlined in program. */
639 int nest = MIN (edge->loop_nest, 8);
640 badness = cgraph_estimate_growth (edge->callee) * 256;
642 /* Decrease badness if call is nested. */
650 fprintf (dump_file, " %i: no profile. nest %i\n", (int) badness,
654 /* Ensure that we did not overflow in all the fixed point math above. */
655 gcc_assert (badness >= INT_MIN);
656 gcc_assert (badness <= INT_MAX - 1);
657 /* Make recursive inlining happen always after other inlining is done. */
658 if (cgraph_recursive_inlining_p (edge->caller, edge->callee, NULL))
664 /* Recompute badness of EDGE and update its key in HEAP if needed. */
666 update_edge_key (fibheap_t heap, struct cgraph_edge *edge)
668 int badness = cgraph_edge_badness (edge, false);
671 fibnode_t n = (fibnode_t) edge->aux;
672 gcc_checking_assert (n->data == edge);
674 /* fibheap_replace_key only decrease the keys.
675 When we increase the key we do not update heap
676 and instead re-insert the element once it becomes
678 if (badness < n->key)
680 fibheap_replace_key (heap, n, badness);
681 gcc_checking_assert (n->key == badness);
685 edge->aux = fibheap_insert (heap, badness, edge);
688 /* Recompute heap nodes for each of caller edge. */
691 update_caller_keys (fibheap_t heap, struct cgraph_node *node,
692 bitmap updated_nodes)
694 struct cgraph_edge *edge;
695 cgraph_inline_failed_t failed_reason;
697 if (!node->local.inlinable
698 || node->global.inlined_to)
700 if (bitmap_bit_p (updated_nodes, node->uid))
702 bitmap_set_bit (updated_nodes, node->uid);
703 node->global.estimated_growth = INT_MIN;
705 /* See if there is something to do. */
706 for (edge = node->callers; edge; edge = edge->next_caller)
707 if (edge->inline_failed)
711 /* Prune out edges we won't inline into anymore. */
712 if (!cgraph_default_inline_p (node, &failed_reason))
714 for (; edge; edge = edge->next_caller)
717 fibheap_delete_node (heap, (fibnode_t) edge->aux);
719 if (edge->inline_failed)
720 edge->inline_failed = failed_reason;
725 for (; edge; edge = edge->next_caller)
726 if (edge->inline_failed)
727 update_edge_key (heap, edge);
730 /* Recompute heap nodes for each uninlined call.
731 This is used when we know that edge badnesses are going only to increase
732 (we introduced new call site) and thus all we need is to insert newly
733 created edges into heap. */
736 update_callee_keys (fibheap_t heap, struct cgraph_node *node,
737 bitmap updated_nodes)
739 struct cgraph_edge *e = node->callees;
740 node->global.estimated_growth = INT_MIN;
745 if (!e->inline_failed && e->callee->callees)
746 e = e->callee->callees;
750 && e->callee->local.inlinable
751 && !bitmap_bit_p (updated_nodes, e->callee->uid))
753 node->global.estimated_growth = INT_MIN;
754 /* If function becomes uninlinable, we need to remove it from the heap. */
755 if (!cgraph_default_inline_p (e->callee, &e->inline_failed))
756 update_caller_keys (heap, e->callee, updated_nodes);
758 /* Otherwise update just edge E. */
759 update_edge_key (heap, e);
767 if (e->caller == node)
769 e = e->caller->callers;
771 while (!e->next_callee);
777 /* Recompute heap nodes for each of caller edges of each of callees.
778 Walk recursively into all inline clones. */
781 update_all_callee_keys (fibheap_t heap, struct cgraph_node *node,
782 bitmap updated_nodes)
784 struct cgraph_edge *e = node->callees;
785 node->global.estimated_growth = INT_MIN;
790 if (!e->inline_failed && e->callee->callees)
791 e = e->callee->callees;
794 if (e->inline_failed)
795 update_caller_keys (heap, e->callee, updated_nodes);
802 if (e->caller == node)
804 e = e->caller->callers;
806 while (!e->next_callee);
812 /* Enqueue all recursive calls from NODE into priority queue depending on
813 how likely we want to recursively inline the call. */
816 lookup_recursive_calls (struct cgraph_node *node, struct cgraph_node *where,
820 struct cgraph_edge *e;
821 for (e = where->callees; e; e = e->next_callee)
822 if (e->callee == node)
824 /* When profile feedback is available, prioritize by expected number
825 of calls. Without profile feedback we maintain simple queue
826 to order candidates via recursive depths. */
827 fibheap_insert (heap,
828 !max_count ? priority++
829 : -(e->count / ((max_count + (1<<24) - 1) / (1<<24))),
832 for (e = where->callees; e; e = e->next_callee)
833 if (!e->inline_failed)
834 lookup_recursive_calls (node, e->callee, heap);
837 /* Decide on recursive inlining: in the case function has recursive calls,
838 inline until body size reaches given argument. If any new indirect edges
839 are discovered in the process, add them to *NEW_EDGES, unless NEW_EDGES
843 cgraph_decide_recursive_inlining (struct cgraph_node *node,
844 VEC (cgraph_edge_p, heap) **new_edges)
846 int limit = PARAM_VALUE (PARAM_MAX_INLINE_INSNS_RECURSIVE_AUTO);
847 int max_depth = PARAM_VALUE (PARAM_MAX_INLINE_RECURSIVE_DEPTH_AUTO);
848 int probability = PARAM_VALUE (PARAM_MIN_INLINE_RECURSIVE_PROBABILITY);
850 struct cgraph_edge *e;
851 struct cgraph_node *master_clone, *next;
855 /* It does not make sense to recursively inline always-inline functions
856 as we are going to sorry() on the remaining calls anyway. */
857 if (node->local.disregard_inline_limits
858 && lookup_attribute ("always_inline", DECL_ATTRIBUTES (node->decl)))
861 if (optimize_function_for_size_p (DECL_STRUCT_FUNCTION (node->decl))
862 || (!flag_inline_functions && !DECL_DECLARED_INLINE_P (node->decl)))
865 if (DECL_DECLARED_INLINE_P (node->decl))
867 limit = PARAM_VALUE (PARAM_MAX_INLINE_INSNS_RECURSIVE);
868 max_depth = PARAM_VALUE (PARAM_MAX_INLINE_RECURSIVE_DEPTH);
871 /* Make sure that function is small enough to be considered for inlining. */
873 || cgraph_estimate_size_after_inlining (1, node, node) >= limit)
875 heap = fibheap_new ();
876 lookup_recursive_calls (node, node, heap);
877 if (fibheap_empty (heap))
879 fibheap_delete (heap);
885 " Performing recursive inlining on %s\n",
886 cgraph_node_name (node));
888 /* We need original clone to copy around. */
889 master_clone = cgraph_clone_node (node, node->decl,
890 node->count, CGRAPH_FREQ_BASE, 1,
892 master_clone->needed = true;
893 for (e = master_clone->callees; e; e = e->next_callee)
894 if (!e->inline_failed)
895 cgraph_clone_inlined_nodes (e, true, false);
897 /* Do the inlining and update list of recursive call during process. */
898 while (!fibheap_empty (heap)
899 && (cgraph_estimate_size_after_inlining (1, node, master_clone)
902 struct cgraph_edge *curr
903 = (struct cgraph_edge *) fibheap_extract_min (heap);
904 struct cgraph_node *cnode;
907 for (cnode = curr->caller;
908 cnode->global.inlined_to; cnode = cnode->callers->caller)
909 if (node->decl == curr->callee->decl)
911 if (depth > max_depth)
915 " maximal depth reached\n");
921 if (!cgraph_maybe_hot_edge_p (curr))
924 fprintf (dump_file, " Not inlining cold call\n");
927 if (curr->count * 100 / node->count < probability)
931 " Probability of edge is too small\n");
939 " Inlining call of depth %i", depth);
942 fprintf (dump_file, " called approx. %.2f times per call",
943 (double)curr->count / node->count);
945 fprintf (dump_file, "\n");
947 cgraph_redirect_edge_callee (curr, master_clone);
948 cgraph_mark_inline_edge (curr, false, new_edges);
949 lookup_recursive_calls (node, curr->callee, heap);
952 if (!fibheap_empty (heap) && dump_file)
953 fprintf (dump_file, " Recursive inlining growth limit met.\n");
955 fibheap_delete (heap);
958 "\n Inlined %i times, body grown from size %i to %i, time %i to %i\n", n,
959 master_clone->global.size, node->global.size,
960 master_clone->global.time, node->global.time);
962 /* Remove master clone we used for inlining. We rely that clones inlined
963 into master clone gets queued just before master clone so we don't
965 for (node = cgraph_nodes; node != master_clone;
969 if (node->global.inlined_to == master_clone)
970 cgraph_remove_node (node);
972 cgraph_remove_node (master_clone);
973 /* FIXME: Recursive inlining actually reduces number of calls of the
974 function. At this place we should probably walk the function and
975 inline clones and compensate the counts accordingly. This probably
976 doesn't matter much in practice. */
980 /* Set inline_failed for all callers of given function to REASON. */
983 cgraph_set_inline_failed (struct cgraph_node *node,
984 cgraph_inline_failed_t reason)
986 struct cgraph_edge *e;
989 fprintf (dump_file, "Inlining failed: %s\n",
990 cgraph_inline_failed_string (reason));
991 for (e = node->callers; e; e = e->next_caller)
992 if (e->inline_failed)
993 e->inline_failed = reason;
996 /* Given whole compilation unit estimate of INSNS, compute how large we can
997 allow the unit to grow. */
999 compute_max_insns (int insns)
1001 int max_insns = insns;
1002 if (max_insns < PARAM_VALUE (PARAM_LARGE_UNIT_INSNS))
1003 max_insns = PARAM_VALUE (PARAM_LARGE_UNIT_INSNS);
1005 return ((HOST_WIDEST_INT) max_insns
1006 * (100 + PARAM_VALUE (PARAM_INLINE_UNIT_GROWTH)) / 100);
1009 /* Compute badness of all edges in NEW_EDGES and add them to the HEAP. */
1011 add_new_edges_to_heap (fibheap_t heap, VEC (cgraph_edge_p, heap) *new_edges)
1013 while (VEC_length (cgraph_edge_p, new_edges) > 0)
1015 struct cgraph_edge *edge = VEC_pop (cgraph_edge_p, new_edges);
1017 gcc_assert (!edge->aux);
1018 if (edge->callee->local.inlinable
1019 && cgraph_default_inline_p (edge->callee, &edge->inline_failed))
1020 edge->aux = fibheap_insert (heap, cgraph_edge_badness (edge, false), edge);
1025 /* We use greedy algorithm for inlining of small functions:
1026 All inline candidates are put into prioritized heap based on estimated
1027 growth of the overall number of instructions and then update the estimates.
1029 INLINED and INLINED_CALEES are just pointers to arrays large enough
1030 to be passed to cgraph_inlined_into and cgraph_inlined_callees. */
1033 cgraph_decide_inlining_of_small_functions (void)
1035 struct cgraph_node *node;
1036 struct cgraph_edge *edge;
1037 cgraph_inline_failed_t failed_reason;
1038 fibheap_t heap = fibheap_new ();
1039 bitmap updated_nodes = BITMAP_ALLOC (NULL);
1040 int min_size, max_size;
1041 VEC (cgraph_edge_p, heap) *new_indirect_edges = NULL;
1043 if (flag_indirect_inlining)
1044 new_indirect_edges = VEC_alloc (cgraph_edge_p, heap, 8);
1047 fprintf (dump_file, "\nDeciding on smaller functions:\n");
1049 /* Put all inline candidates into the heap. */
1051 for (node = cgraph_nodes; node; node = node->next)
1053 if (!node->local.inlinable || !node->callers)
1056 fprintf (dump_file, "Considering inline candidate %s.\n", cgraph_node_name (node));
1058 node->global.estimated_growth = INT_MIN;
1059 if (!cgraph_default_inline_p (node, &failed_reason))
1061 cgraph_set_inline_failed (node, failed_reason);
1065 for (edge = node->callers; edge; edge = edge->next_caller)
1066 if (edge->inline_failed)
1068 gcc_assert (!edge->aux);
1069 edge->aux = fibheap_insert (heap, cgraph_edge_badness (edge, false), edge);
1073 max_size = compute_max_insns (overall_size);
1074 min_size = overall_size;
1076 while (overall_size <= max_size
1077 && !fibheap_empty (heap))
1079 int old_size = overall_size;
1080 struct cgraph_node *where, *callee;
1081 int badness = fibheap_min_key (heap);
1082 int current_badness;
1084 cgraph_inline_failed_t not_good = CIF_OK;
1086 edge = (struct cgraph_edge *) fibheap_extract_min (heap);
1087 gcc_assert (edge->aux);
1089 if (!edge->inline_failed)
1092 /* When updating the edge costs, we only decrease badness in the keys.
1093 When the badness increase, we keep the heap as it is and re-insert
1095 current_badness = cgraph_edge_badness (edge, false);
1096 gcc_assert (current_badness >= badness);
1097 if (current_badness != badness)
1099 edge->aux = fibheap_insert (heap, current_badness, edge);
1103 callee = edge->callee;
1105 growth = (cgraph_estimate_size_after_inlining (1, edge->caller, edge->callee)
1106 - edge->caller->global.size);
1111 "\nConsidering %s with %i size\n",
1112 cgraph_node_name (edge->callee),
1113 edge->callee->global.size);
1115 " to be inlined into %s in %s:%i\n"
1116 " Estimated growth after inlined into all callees is %+i insns.\n"
1117 " Estimated badness is %i, frequency %.2f.\n",
1118 cgraph_node_name (edge->caller),
1119 flag_wpa ? "unknown"
1120 : gimple_filename ((const_gimple) edge->call_stmt),
1121 flag_wpa ? -1 : gimple_lineno ((const_gimple) edge->call_stmt),
1122 cgraph_estimate_growth (edge->callee),
1124 edge->frequency / (double)CGRAPH_FREQ_BASE);
1126 fprintf (dump_file," Called "HOST_WIDEST_INT_PRINT_DEC"x\n", edge->count);
1127 if (dump_flags & TDF_DETAILS)
1128 cgraph_edge_badness (edge, true);
1131 /* When not having profile info ready we don't weight by any way the
1132 position of call in procedure itself. This means if call of
1133 function A from function B seems profitable to inline, the recursive
1134 call of function A in inline copy of A in B will look profitable too
1135 and we end up inlining until reaching maximal function growth. This
1136 is not good idea so prohibit the recursive inlining.
1138 ??? When the frequencies are taken into account we might not need this
1141 We need to be cureful here, in some testcases, e.g. directivec.c in
1142 libcpp, we can estimate self recursive function to have negative growth
1143 for inlining completely.
1147 where = edge->caller;
1148 while (where->global.inlined_to)
1150 if (where->decl == edge->callee->decl)
1152 where = where->callers->caller;
1154 if (where->global.inlined_to)
1157 = (edge->callee->local.disregard_inline_limits
1158 ? CIF_RECURSIVE_INLINING : CIF_UNSPECIFIED);
1160 fprintf (dump_file, " inline_failed:Recursive inlining performed only for function itself.\n");
1165 if (edge->callee->local.disregard_inline_limits)
1167 else if (!cgraph_maybe_hot_edge_p (edge))
1168 not_good = CIF_UNLIKELY_CALL;
1169 else if (!flag_inline_functions
1170 && !DECL_DECLARED_INLINE_P (edge->callee->decl))
1171 not_good = CIF_NOT_DECLARED_INLINED;
1172 else if (optimize_function_for_size_p (DECL_STRUCT_FUNCTION(edge->caller->decl)))
1173 not_good = CIF_OPTIMIZING_FOR_SIZE;
1174 if (not_good && growth > 0 && cgraph_estimate_growth (edge->callee) > 0)
1176 if (!cgraph_recursive_inlining_p (edge->caller, edge->callee,
1177 &edge->inline_failed))
1179 edge->inline_failed = not_good;
1181 fprintf (dump_file, " inline_failed:%s.\n",
1182 cgraph_inline_failed_string (edge->inline_failed));
1186 if (!cgraph_default_inline_p (edge->callee, &edge->inline_failed))
1188 if (!cgraph_recursive_inlining_p (edge->caller, edge->callee,
1189 &edge->inline_failed))
1192 fprintf (dump_file, " inline_failed:%s.\n",
1193 cgraph_inline_failed_string (edge->inline_failed));
1197 if (!tree_can_inline_p (edge))
1200 fprintf (dump_file, " inline_failed:%s.\n",
1201 cgraph_inline_failed_string (edge->inline_failed));
1204 if (cgraph_recursive_inlining_p (edge->caller, edge->callee,
1205 &edge->inline_failed))
1207 where = edge->caller;
1208 if (where->global.inlined_to)
1209 where = where->global.inlined_to;
1210 if (!cgraph_decide_recursive_inlining (where,
1211 flag_indirect_inlining
1212 ? &new_indirect_edges : NULL))
1214 if (flag_indirect_inlining)
1215 add_new_edges_to_heap (heap, new_indirect_edges);
1216 update_all_callee_keys (heap, where, updated_nodes);
1220 struct cgraph_node *callee;
1221 if (edge->call_stmt_cannot_inline_p
1222 || !cgraph_check_inline_limits (edge->caller, edge->callee,
1223 &edge->inline_failed, true))
1226 fprintf (dump_file, " Not inlining into %s:%s.\n",
1227 cgraph_node_name (edge->caller),
1228 cgraph_inline_failed_string (edge->inline_failed));
1231 callee = edge->callee;
1232 gcc_checking_assert (!callee->global.inlined_to);
1233 cgraph_mark_inline_edge (edge, true, &new_indirect_edges);
1234 if (flag_indirect_inlining)
1235 add_new_edges_to_heap (heap, new_indirect_edges);
1237 /* We inlined last offline copy to the body. This might lead
1238 to callees of function having fewer call sites and thus they
1239 may need updating. */
1240 if (callee->global.inlined_to)
1241 update_all_callee_keys (heap, callee, updated_nodes);
1243 update_callee_keys (heap, edge->callee, updated_nodes);
1245 where = edge->caller;
1246 if (where->global.inlined_to)
1247 where = where->global.inlined_to;
1249 /* Our profitability metric can depend on local properties
1250 such as number of inlinable calls and size of the function body.
1251 After inlining these properties might change for the function we
1252 inlined into (since it's body size changed) and for the functions
1253 called by function we inlined (since number of it inlinable callers
1255 update_caller_keys (heap, where, updated_nodes);
1257 /* We removed one call of the function we just inlined. If offline
1258 copy is still needed, be sure to update the keys. */
1259 if (callee != where && !callee->global.inlined_to)
1260 update_caller_keys (heap, callee, updated_nodes);
1261 bitmap_clear (updated_nodes);
1266 " Inlined into %s which now has size %i and self time %i,"
1267 "net change of %+i.\n",
1268 cgraph_node_name (edge->caller),
1269 edge->caller->global.time,
1270 edge->caller->global.size,
1271 overall_size - old_size);
1273 if (min_size > overall_size)
1275 min_size = overall_size;
1276 max_size = compute_max_insns (min_size);
1279 fprintf (dump_file, "New minimal size reached: %i\n", min_size);
1282 while (!fibheap_empty (heap))
1284 int badness = fibheap_min_key (heap);
1286 edge = (struct cgraph_edge *) fibheap_extract_min (heap);
1287 gcc_assert (edge->aux);
1289 if (!edge->inline_failed)
1291 #ifdef ENABLE_CHECKING
1292 gcc_assert (cgraph_edge_badness (edge, false) >= badness);
1297 "\nSkipping %s with %i size\n",
1298 cgraph_node_name (edge->callee),
1299 edge->callee->global.size);
1301 " called by %s in %s:%i\n"
1302 " Estimated growth after inlined into all callees is %+i insns.\n"
1303 " Estimated badness is %i, frequency %.2f.\n",
1304 cgraph_node_name (edge->caller),
1305 flag_wpa ? "unknown"
1306 : gimple_filename ((const_gimple) edge->call_stmt),
1307 flag_wpa ? -1 : gimple_lineno ((const_gimple) edge->call_stmt),
1308 cgraph_estimate_growth (edge->callee),
1310 edge->frequency / (double)CGRAPH_FREQ_BASE);
1312 fprintf (dump_file," Called "HOST_WIDEST_INT_PRINT_DEC"x\n", edge->count);
1313 if (dump_flags & TDF_DETAILS)
1314 cgraph_edge_badness (edge, true);
1316 if (!edge->callee->local.disregard_inline_limits && edge->inline_failed
1317 && !cgraph_recursive_inlining_p (edge->caller, edge->callee,
1318 &edge->inline_failed))
1319 edge->inline_failed = CIF_INLINE_UNIT_GROWTH_LIMIT;
1322 if (new_indirect_edges)
1323 VEC_free (cgraph_edge_p, heap, new_indirect_edges);
1324 fibheap_delete (heap);
1325 BITMAP_FREE (updated_nodes);
1328 /* Flatten NODE from the IPA inliner. */
1331 cgraph_flatten (struct cgraph_node *node)
1333 struct cgraph_edge *e;
1335 /* We shouldn't be called recursively when we are being processed. */
1336 gcc_assert (node->aux == NULL);
1338 node->aux = (void *)(size_t) INLINE_ALL;
1340 for (e = node->callees; e; e = e->next_callee)
1342 struct cgraph_node *orig_callee;
1344 if (e->call_stmt_cannot_inline_p)
1347 if (!e->callee->analyzed)
1351 "Not inlining: Function body not available.\n");
1355 /* We've hit cycle? It is time to give up. */
1360 "Not inlining %s into %s to avoid cycle.\n",
1361 cgraph_node_name (e->callee),
1362 cgraph_node_name (e->caller));
1363 e->inline_failed = CIF_RECURSIVE_INLINING;
1367 /* When the edge is already inlined, we just need to recurse into
1368 it in order to fully flatten the leaves. */
1369 if (!e->inline_failed)
1371 cgraph_flatten (e->callee);
1375 if (cgraph_recursive_inlining_p (node, e->callee, &e->inline_failed))
1378 fprintf (dump_file, "Not inlining: recursive call.\n");
1382 if (!tree_can_inline_p (e))
1385 fprintf (dump_file, "Not inlining: %s",
1386 cgraph_inline_failed_string (e->inline_failed));
1390 /* Inline the edge and flatten the inline clone. Avoid
1391 recursing through the original node if the node was cloned. */
1393 fprintf (dump_file, " Inlining %s into %s.\n",
1394 cgraph_node_name (e->callee),
1395 cgraph_node_name (e->caller));
1396 orig_callee = e->callee;
1397 cgraph_mark_inline_edge (e, true, NULL);
1398 if (e->callee != orig_callee)
1399 orig_callee->aux = (void *)(size_t) INLINE_ALL;
1400 cgraph_flatten (e->callee);
1401 if (e->callee != orig_callee)
1402 orig_callee->aux = NULL;
1408 /* Decide on the inlining. We do so in the topological order to avoid
1409 expenses on updating data structures. */
1412 cgraph_decide_inlining (void)
1414 struct cgraph_node *node;
1416 struct cgraph_node **order =
1417 XCNEWVEC (struct cgraph_node *, cgraph_n_nodes);
1420 int initial_size = 0;
1422 cgraph_remove_function_insertion_hook (function_insertion_hook_holder);
1423 if (in_lto_p && flag_indirect_inlining)
1424 ipa_update_after_lto_read ();
1425 if (flag_indirect_inlining)
1426 ipa_create_all_structures_for_iinln ();
1430 for (node = cgraph_nodes; node; node = node->next)
1433 struct cgraph_edge *e;
1435 gcc_assert (inline_summary (node)->self_size == node->global.size);
1436 initial_size += node->global.size;
1437 for (e = node->callees; e; e = e->next_callee)
1438 if (max_count < e->count)
1439 max_count = e->count;
1440 if (max_benefit < inline_summary (node)->time_inlining_benefit)
1441 max_benefit = inline_summary (node)->time_inlining_benefit;
1443 gcc_assert (in_lto_p
1445 || (profile_info && flag_branch_probabilities));
1446 overall_size = initial_size;
1448 nnodes = cgraph_postorder (order);
1452 "\nDeciding on inlining. Starting with size %i.\n",
1455 for (node = cgraph_nodes; node; node = node->next)
1459 fprintf (dump_file, "\nFlattening functions:\n");
1461 /* In the first pass handle functions to be flattened. Do this with
1462 a priority so none of our later choices will make this impossible. */
1463 for (i = nnodes - 1; i >= 0; i--)
1467 /* Handle nodes to be flattened, but don't update overall unit
1468 size. Calling the incremental inliner here is lame,
1469 a simple worklist should be enough. What should be left
1470 here from the early inliner (if it runs) is cyclic cases.
1471 Ideally when processing callees we stop inlining at the
1472 entry of cycles, possibly cloning that entry point and
1473 try to flatten itself turning it into a self-recursive
1475 if (lookup_attribute ("flatten",
1476 DECL_ATTRIBUTES (node->decl)) != NULL)
1480 "Flattening %s\n", cgraph_node_name (node));
1481 cgraph_flatten (node);
1485 cgraph_decide_inlining_of_small_functions ();
1487 if (flag_inline_functions_called_once)
1490 fprintf (dump_file, "\nDeciding on functions called once:\n");
1492 /* And finally decide what functions are called once. */
1493 for (i = nnodes - 1; i >= 0; i--)
1498 && !node->callers->next_caller
1499 && cgraph_will_be_removed_from_program_if_no_direct_calls (node)
1500 && node->local.inlinable
1501 && node->callers->inline_failed
1502 && node->callers->caller != node
1503 && node->callers->caller->global.inlined_to != node
1504 && !node->callers->call_stmt_cannot_inline_p
1505 && !DECL_EXTERNAL (node->decl))
1507 cgraph_inline_failed_t reason;
1508 old_size = overall_size;
1512 "\nConsidering %s size %i.\n",
1513 cgraph_node_name (node), node->global.size);
1515 " Called once from %s %i insns.\n",
1516 cgraph_node_name (node->callers->caller),
1517 node->callers->caller->global.size);
1520 if (cgraph_check_inline_limits (node->callers->caller, node,
1523 struct cgraph_node *caller = node->callers->caller;
1524 cgraph_mark_inline (node->callers);
1527 " Inlined into %s which now has %i size"
1528 " for a net change of %+i size.\n",
1529 cgraph_node_name (caller),
1530 caller->global.size,
1531 overall_size - old_size);
1537 " Not inlining: %s.\n",
1538 cgraph_inline_failed_string (reason));
1544 /* Free ipa-prop structures if they are no longer needed. */
1545 if (flag_indirect_inlining)
1546 ipa_free_all_structures_after_iinln ();
1550 "\nInlined %i calls, eliminated %i functions, "
1551 "size %i turned to %i size.\n\n",
1552 ncalls_inlined, nfunctions_inlined, initial_size,
1558 /* Return true when N is leaf function. Accept cheap (pure&const) builtins
1559 in leaf functions. */
1561 leaf_node_p (struct cgraph_node *n)
1563 struct cgraph_edge *e;
1564 for (e = n->callees; e; e = e->next_callee)
1565 if (!DECL_BUILT_IN (e->callee->decl)
1566 || (!TREE_READONLY (e->callee->decl)
1567 || DECL_PURE_P (e->callee->decl)))
1572 /* Decide on the inlining. We do so in the topological order to avoid
1573 expenses on updating data structures. */
1576 cgraph_decide_inlining_incrementally (struct cgraph_node *node,
1577 enum inlining_mode mode)
1579 struct cgraph_edge *e;
1580 bool inlined = false;
1581 cgraph_inline_failed_t failed_reason;
1583 #ifdef ENABLE_CHECKING
1584 verify_cgraph_node (node);
1587 if (mode != INLINE_ALWAYS_INLINE && mode != INLINE_SIZE_NORECURSIVE
1588 && lookup_attribute ("flatten", DECL_ATTRIBUTES (node->decl)) != NULL)
1591 fprintf (dump_file, "Incrementally flattening %s\n",
1592 cgraph_node_name (node));
1596 /* First of all look for always inline functions. */
1597 if (mode != INLINE_SIZE_NORECURSIVE)
1598 for (e = node->callees; e; e = e->next_callee)
1600 if (!e->callee->local.disregard_inline_limits
1601 && (mode != INLINE_ALL || !e->callee->local.inlinable))
1603 if (e->call_stmt_cannot_inline_p)
1607 "Considering to always inline inline candidate %s.\n",
1608 cgraph_node_name (e->callee));
1609 if (cgraph_recursive_inlining_p (node, e->callee, &e->inline_failed))
1612 fprintf (dump_file, "Not inlining: recursive call.\n");
1615 if (!tree_can_inline_p (e))
1620 cgraph_inline_failed_string (e->inline_failed));
1623 if (gimple_in_ssa_p (DECL_STRUCT_FUNCTION (node->decl))
1624 != gimple_in_ssa_p (DECL_STRUCT_FUNCTION (e->callee->decl)))
1627 fprintf (dump_file, "Not inlining: SSA form does not match.\n");
1630 if (!e->callee->analyzed)
1634 "Not inlining: Function body no longer available.\n");
1639 fprintf (dump_file, " Inlining %s into %s.\n",
1640 cgraph_node_name (e->callee),
1641 cgraph_node_name (e->caller));
1642 cgraph_mark_inline (e);
1646 /* Now do the automatic inlining. */
1647 if (mode != INLINE_ALL && mode != INLINE_ALWAYS_INLINE
1648 /* Never inline regular functions into always-inline functions
1649 during incremental inlining. */
1650 && !node->local.disregard_inline_limits)
1652 bitmap visited = BITMAP_ALLOC (NULL);
1653 for (e = node->callees; e; e = e->next_callee)
1655 int allowed_growth = 0;
1656 if (!e->callee->local.inlinable
1657 || !e->inline_failed
1658 || e->callee->local.disregard_inline_limits)
1660 /* We are inlining a function to all call-sites in node
1661 or to none. So visit each candidate only once. */
1662 if (!bitmap_set_bit (visited, e->callee->uid))
1665 fprintf (dump_file, "Considering inline candidate %s.\n",
1666 cgraph_node_name (e->callee));
1667 if (cgraph_recursive_inlining_p (node, e->callee, &e->inline_failed))
1670 fprintf (dump_file, "Not inlining: recursive call.\n");
1673 if (gimple_in_ssa_p (DECL_STRUCT_FUNCTION (node->decl))
1674 != gimple_in_ssa_p (DECL_STRUCT_FUNCTION (e->callee->decl)))
1678 "Not inlining: SSA form does not match.\n");
1682 if (cgraph_maybe_hot_edge_p (e) && leaf_node_p (e->callee)
1683 && optimize_function_for_speed_p (cfun))
1684 allowed_growth = PARAM_VALUE (PARAM_EARLY_INLINING_INSNS);
1686 /* When the function body would grow and inlining the function
1687 won't eliminate the need for offline copy of the function,
1689 if (((mode == INLINE_SIZE || mode == INLINE_SIZE_NORECURSIVE)
1690 || (!flag_inline_functions
1691 && !DECL_DECLARED_INLINE_P (e->callee->decl)))
1692 && (cgraph_estimate_size_after_inlining (1, e->caller, e->callee)
1693 > e->caller->global.size + allowed_growth)
1694 && cgraph_estimate_growth (e->callee) > allowed_growth)
1698 "Not inlining: code size would grow by %i.\n",
1699 cgraph_estimate_size_after_inlining (1, e->caller,
1701 - e->caller->global.size);
1704 if (!cgraph_check_inline_limits (node, e->callee, &e->inline_failed,
1706 || e->call_stmt_cannot_inline_p)
1709 fprintf (dump_file, "Not inlining: %s.\n",
1710 cgraph_inline_failed_string (e->inline_failed));
1713 if (!e->callee->analyzed)
1717 "Not inlining: Function body no longer available.\n");
1720 if (!tree_can_inline_p (e))
1724 "Not inlining: %s.",
1725 cgraph_inline_failed_string (e->inline_failed));
1728 if (cgraph_default_inline_p (e->callee, &failed_reason))
1731 fprintf (dump_file, " Inlining %s into %s.\n",
1732 cgraph_node_name (e->callee),
1733 cgraph_node_name (e->caller));
1734 cgraph_mark_inline (e);
1738 BITMAP_FREE (visited);
1743 /* Because inlining might remove no-longer reachable nodes, we need to
1744 keep the array visible to garbage collector to avoid reading collected
1747 static GTY ((length ("nnodes"))) struct cgraph_node **order;
1749 /* Do inlining of small functions. Doing so early helps profiling and other
1750 passes to be somewhat more effective and avoids some code duplication in
1751 later real inlining pass for testcases with very many function calls. */
1753 cgraph_early_inlining (void)
1755 struct cgraph_node *node = cgraph_node (current_function_decl);
1756 unsigned int todo = 0;
1764 || !flag_early_inlining)
1766 /* When not optimizing or not inlining inline only always-inline
1768 cgraph_decide_inlining_incrementally (node, INLINE_ALWAYS_INLINE);
1769 timevar_push (TV_INTEGRATION);
1770 todo |= optimize_inline_calls (current_function_decl);
1771 timevar_pop (TV_INTEGRATION);
1775 if (lookup_attribute ("flatten",
1776 DECL_ATTRIBUTES (node->decl)) != NULL)
1780 "Flattening %s\n", cgraph_node_name (node));
1781 cgraph_flatten (node);
1782 timevar_push (TV_INTEGRATION);
1783 todo |= optimize_inline_calls (current_function_decl);
1784 timevar_pop (TV_INTEGRATION);
1786 /* We iterate incremental inlining to get trivial cases of indirect
1788 while (iterations < PARAM_VALUE (PARAM_EARLY_INLINER_MAX_ITERATIONS)
1789 && cgraph_decide_inlining_incrementally (node,
1791 ? INLINE_SIZE_NORECURSIVE
1794 timevar_push (TV_INTEGRATION);
1795 todo |= optimize_inline_calls (current_function_decl);
1797 timevar_pop (TV_INTEGRATION);
1800 fprintf (dump_file, "Iterations: %i\n", iterations);
1803 cfun->always_inline_functions_inlined = true;
1808 struct gimple_opt_pass pass_early_inline =
1812 "einline", /* name */
1814 cgraph_early_inlining, /* execute */
1817 0, /* static_pass_number */
1818 TV_INLINE_HEURISTICS, /* tv_id */
1819 0, /* properties_required */
1820 0, /* properties_provided */
1821 0, /* properties_destroyed */
1822 0, /* todo_flags_start */
1823 TODO_dump_func /* todo_flags_finish */
1827 /* When inlining shall be performed. */
1829 cgraph_gate_ipa_early_inlining (void)
1831 return (flag_early_inlining
1833 && (flag_branch_probabilities || flag_test_coverage
1834 || profile_arc_flag));
1837 /* IPA pass wrapper for early inlining pass. We need to run early inlining
1838 before tree profiling so we have stand alone IPA pass for doing so. */
1839 struct simple_ipa_opt_pass pass_ipa_early_inline =
1843 "einline_ipa", /* name */
1844 cgraph_gate_ipa_early_inlining, /* gate */
1848 0, /* static_pass_number */
1849 TV_INLINE_HEURISTICS, /* tv_id */
1850 0, /* properties_required */
1851 0, /* properties_provided */
1852 0, /* properties_destroyed */
1853 0, /* todo_flags_start */
1854 TODO_dump_cgraph /* todo_flags_finish */
1858 /* See if statement might disappear after inlining. We are not terribly
1859 sophisficated, basically looking for simple abstraction penalty wrappers. */
1862 likely_eliminated_by_inlining_p (gimple stmt)
1864 enum gimple_code code = gimple_code (stmt);
1870 if (gimple_num_ops (stmt) != 2)
1873 /* Casts of parameters, loads from parameters passed by reference
1874 and stores to return value or parameters are probably free after
1876 if (gimple_assign_rhs_code (stmt) == CONVERT_EXPR
1877 || gimple_assign_rhs_code (stmt) == NOP_EXPR
1878 || gimple_assign_rhs_code (stmt) == VIEW_CONVERT_EXPR
1879 || gimple_assign_rhs_class (stmt) == GIMPLE_SINGLE_RHS)
1881 tree rhs = gimple_assign_rhs1 (stmt);
1882 tree lhs = gimple_assign_lhs (stmt);
1883 tree inner_rhs = rhs;
1884 tree inner_lhs = lhs;
1885 bool rhs_free = false;
1886 bool lhs_free = false;
1888 while (handled_component_p (inner_lhs)
1889 || TREE_CODE (inner_lhs) == MEM_REF)
1890 inner_lhs = TREE_OPERAND (inner_lhs, 0);
1891 while (handled_component_p (inner_rhs)
1892 || TREE_CODE (inner_rhs) == ADDR_EXPR
1893 || TREE_CODE (inner_rhs) == MEM_REF)
1894 inner_rhs = TREE_OPERAND (inner_rhs, 0);
1897 if (TREE_CODE (inner_rhs) == PARM_DECL
1898 || (TREE_CODE (inner_rhs) == SSA_NAME
1899 && SSA_NAME_IS_DEFAULT_DEF (inner_rhs)
1900 && TREE_CODE (SSA_NAME_VAR (inner_rhs)) == PARM_DECL))
1902 if (rhs_free && is_gimple_reg (lhs))
1904 if (((TREE_CODE (inner_lhs) == PARM_DECL
1905 || (TREE_CODE (inner_lhs) == SSA_NAME
1906 && SSA_NAME_IS_DEFAULT_DEF (inner_lhs)
1907 && TREE_CODE (SSA_NAME_VAR (inner_lhs)) == PARM_DECL))
1908 && inner_lhs != lhs)
1909 || TREE_CODE (inner_lhs) == RESULT_DECL
1910 || (TREE_CODE (inner_lhs) == SSA_NAME
1911 && TREE_CODE (SSA_NAME_VAR (inner_lhs)) == RESULT_DECL))
1914 && (is_gimple_reg (rhs) || is_gimple_min_invariant (rhs)))
1916 if (lhs_free && rhs_free)
1925 /* Compute function body size parameters for NODE. */
1928 estimate_function_body_sizes (struct cgraph_node *node)
1931 gcov_type time_inlining_benefit = 0;
1933 int size_inlining_benefit = 0;
1935 gimple_stmt_iterator bsi;
1936 struct function *my_function = DECL_STRUCT_FUNCTION (node->decl);
1939 tree funtype = TREE_TYPE (node->decl);
1942 fprintf (dump_file, "Analyzing function body size: %s\n",
1943 cgraph_node_name (node));
1945 gcc_assert (my_function && my_function->cfg);
1946 FOR_EACH_BB_FN (bb, my_function)
1948 freq = compute_call_stmt_bb_frequency (node->decl, bb);
1949 for (bsi = gsi_start_bb (bb); !gsi_end_p (bsi); gsi_next (&bsi))
1951 gimple stmt = gsi_stmt (bsi);
1952 int this_size = estimate_num_insns (stmt, &eni_size_weights);
1953 int this_time = estimate_num_insns (stmt, &eni_time_weights);
1955 if (dump_file && (dump_flags & TDF_DETAILS))
1957 fprintf (dump_file, " freq:%6i size:%3i time:%3i ",
1958 freq, this_size, this_time);
1959 print_gimple_stmt (dump_file, stmt, 0, 0);
1964 if (likely_eliminated_by_inlining_p (stmt))
1966 size_inlining_benefit += this_size;
1967 time_inlining_benefit += this_time;
1968 if (dump_file && (dump_flags & TDF_DETAILS))
1969 fprintf (dump_file, " Likely eliminated\n");
1971 gcc_assert (time >= 0);
1972 gcc_assert (size >= 0);
1975 time = (time + CGRAPH_FREQ_BASE / 2) / CGRAPH_FREQ_BASE;
1976 time_inlining_benefit = ((time_inlining_benefit + CGRAPH_FREQ_BASE / 2)
1977 / CGRAPH_FREQ_BASE);
1979 fprintf (dump_file, "Overall function body time: %i-%i size: %i-%i\n",
1980 (int)time, (int)time_inlining_benefit,
1981 size, size_inlining_benefit);
1982 time_inlining_benefit += eni_time_weights.call_cost;
1983 size_inlining_benefit += eni_size_weights.call_cost;
1984 if (!VOID_TYPE_P (TREE_TYPE (funtype)))
1986 int cost = estimate_move_cost (TREE_TYPE (funtype));
1987 time_inlining_benefit += cost;
1988 size_inlining_benefit += cost;
1990 for (arg = DECL_ARGUMENTS (node->decl); arg; arg = TREE_CHAIN (arg))
1991 if (!VOID_TYPE_P (TREE_TYPE (arg)))
1993 int cost = estimate_move_cost (TREE_TYPE (arg));
1994 time_inlining_benefit += cost;
1995 size_inlining_benefit += cost;
1997 if (time_inlining_benefit > MAX_TIME)
1998 time_inlining_benefit = MAX_TIME;
1999 if (time > MAX_TIME)
2001 inline_summary (node)->self_time = time;
2002 inline_summary (node)->self_size = size;
2004 fprintf (dump_file, "With function call overhead time: %i-%i size: %i-%i\n",
2005 (int)time, (int)time_inlining_benefit,
2006 size, size_inlining_benefit);
2007 inline_summary (node)->time_inlining_benefit = time_inlining_benefit;
2008 inline_summary (node)->size_inlining_benefit = size_inlining_benefit;
2011 /* Compute parameters of functions used by inliner. */
2013 compute_inline_parameters (struct cgraph_node *node)
2015 HOST_WIDE_INT self_stack_size;
2017 gcc_assert (!node->global.inlined_to);
2019 /* Estimate the stack size for the function. But not at -O0
2020 because estimated_stack_frame_size is a quadratic problem. */
2021 self_stack_size = optimize ? estimated_stack_frame_size (node->decl) : 0;
2022 inline_summary (node)->estimated_self_stack_size = self_stack_size;
2023 node->global.estimated_stack_size = self_stack_size;
2024 node->global.stack_frame_offset = 0;
2026 /* Can this function be inlined at all? */
2027 node->local.inlinable = tree_inlinable_function_p (node->decl);
2028 if (node->local.inlinable && !node->local.disregard_inline_limits)
2029 node->local.disregard_inline_limits
2030 = DECL_DISREGARD_INLINE_LIMITS (node->decl);
2031 estimate_function_body_sizes (node);
2032 /* Inlining characteristics are maintained by the cgraph_mark_inline. */
2033 node->global.time = inline_summary (node)->self_time;
2034 node->global.size = inline_summary (node)->self_size;
2039 /* Compute parameters of functions used by inliner using
2040 current_function_decl. */
2042 compute_inline_parameters_for_current (void)
2044 compute_inline_parameters (cgraph_node (current_function_decl));
2048 struct gimple_opt_pass pass_inline_parameters =
2052 "inline_param", /* name */
2054 compute_inline_parameters_for_current,/* execute */
2057 0, /* static_pass_number */
2058 TV_INLINE_HEURISTICS, /* tv_id */
2059 0, /* properties_required */
2060 0, /* properties_provided */
2061 0, /* properties_destroyed */
2062 0, /* todo_flags_start */
2063 0 /* todo_flags_finish */
2067 /* This function performs intraprocedural analyzis in NODE that is required to
2068 inline indirect calls. */
2070 inline_indirect_intraprocedural_analysis (struct cgraph_node *node)
2072 ipa_analyze_node (node);
2073 if (dump_file && (dump_flags & TDF_DETAILS))
2075 ipa_print_node_params (dump_file, node);
2076 ipa_print_node_jump_functions (dump_file, node);
2080 /* Note function body size. */
2082 analyze_function (struct cgraph_node *node)
2084 push_cfun (DECL_STRUCT_FUNCTION (node->decl));
2085 current_function_decl = node->decl;
2087 compute_inline_parameters (node);
2088 if (flag_indirect_inlining)
2089 inline_indirect_intraprocedural_analysis (node);
2091 current_function_decl = NULL;
2095 /* Called when new function is inserted to callgraph late. */
2097 add_new_function (struct cgraph_node *node, void *data ATTRIBUTE_UNUSED)
2099 analyze_function (node);
2102 /* Note function body size. */
2104 inline_generate_summary (void)
2106 struct cgraph_node *node;
2108 function_insertion_hook_holder =
2109 cgraph_add_function_insertion_hook (&add_new_function, NULL);
2111 if (flag_indirect_inlining)
2113 ipa_register_cgraph_hooks ();
2114 ipa_check_create_node_params ();
2115 ipa_check_create_edge_args ();
2118 for (node = cgraph_nodes; node; node = node->next)
2120 analyze_function (node);
2125 /* Apply inline plan to function. */
2127 inline_transform (struct cgraph_node *node)
2129 unsigned int todo = 0;
2130 struct cgraph_edge *e;
2131 bool inline_p = false;
2133 /* FIXME: Currently the passmanager is adding inline transform more than once to some
2134 clones. This needs revisiting after WPA cleanups. */
2135 if (cfun->after_inlining)
2138 /* We might need the body of this function so that we can expand
2139 it inline somewhere else. */
2140 if (cgraph_preserve_function_body_p (node->decl))
2141 save_inline_function_body (node);
2143 for (e = node->callees; e; e = e->next_callee)
2145 cgraph_redirect_edge_call_stmt_to_callee (e);
2146 if (!e->inline_failed || warn_inline)
2152 timevar_push (TV_INTEGRATION);
2153 todo = optimize_inline_calls (current_function_decl);
2154 timevar_pop (TV_INTEGRATION);
2156 cfun->always_inline_functions_inlined = true;
2157 cfun->after_inlining = true;
2158 return todo | execute_fixup_cfg ();
2161 /* Read inline summary. Jump functions are shared among ipa-cp
2162 and inliner, so when ipa-cp is active, we don't need to write them
2166 inline_read_summary (void)
2168 if (flag_indirect_inlining)
2170 ipa_register_cgraph_hooks ();
2172 ipa_prop_read_jump_functions ();
2174 function_insertion_hook_holder =
2175 cgraph_add_function_insertion_hook (&add_new_function, NULL);
2178 /* Write inline summary for node in SET.
2179 Jump functions are shared among ipa-cp and inliner, so when ipa-cp is
2180 active, we don't need to write them twice. */
2183 inline_write_summary (cgraph_node_set set,
2184 varpool_node_set vset ATTRIBUTE_UNUSED)
2186 if (flag_indirect_inlining && !flag_ipa_cp)
2187 ipa_prop_write_jump_functions (set);
2190 /* When to run IPA inlining. Inlining of always-inline functions
2191 happens during early inlining. */
2194 gate_cgraph_decide_inlining (void)
2196 /* ??? We'd like to skip this if not optimizing or not inlining as
2197 all always-inline functions have been processed by early
2198 inlining already. But this at least breaks EH with C++ as
2199 we need to unconditionally run fixup_cfg even at -O0.
2200 So leave it on unconditionally for now. */
2204 struct ipa_opt_pass_d pass_ipa_inline =
2208 "inline", /* name */
2209 gate_cgraph_decide_inlining, /* gate */
2210 cgraph_decide_inlining, /* execute */
2213 0, /* static_pass_number */
2214 TV_INLINE_HEURISTICS, /* tv_id */
2215 0, /* properties_required */
2216 0, /* properties_provided */
2217 0, /* properties_destroyed */
2218 TODO_remove_functions, /* todo_flags_finish */
2219 TODO_dump_cgraph | TODO_dump_func
2220 | TODO_remove_functions | TODO_ggc_collect /* todo_flags_finish */
2222 inline_generate_summary, /* generate_summary */
2223 inline_write_summary, /* write_summary */
2224 inline_read_summary, /* read_summary */
2225 NULL, /* write_optimization_summary */
2226 NULL, /* read_optimization_summary */
2227 NULL, /* stmt_fixup */
2229 inline_transform, /* function_transform */
2230 NULL, /* variable_transform */
2234 #include "gt-ipa-inline.h"