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"
135 #include "tree-pass.h"
137 #include "coverage.h"
139 #include "tree-flow.h"
141 #include "ipa-prop.h"
144 #define MAX_TIME 1000000000
146 /* Mode incremental inliner operate on:
148 In ALWAYS_INLINE only functions marked
149 always_inline are inlined. This mode is used after detecting cycle during
152 In SIZE mode, only functions that reduce function body size after inlining
153 are inlined, this is used during early inlining.
155 in ALL mode, everything is inlined. This is used during flattening. */
158 INLINE_ALWAYS_INLINE,
159 INLINE_SIZE_NORECURSIVE,
165 cgraph_decide_inlining_incrementally (struct cgraph_node *, enum inlining_mode);
166 static void cgraph_flatten (struct cgraph_node *node);
169 /* Statistics we collect about inlining algorithm. */
170 static int ncalls_inlined;
171 static int nfunctions_inlined;
172 static int overall_size;
173 static gcov_type max_count, max_benefit;
175 /* Holders of ipa cgraph hooks: */
176 static struct cgraph_node_hook_list *function_insertion_hook_holder;
178 static inline struct inline_summary *
179 inline_summary (struct cgraph_node *node)
181 return &node->local.inline_summary;
184 /* Estimate self time of the function after inlining WHAT into TO. */
187 cgraph_estimate_time_after_inlining (int frequency, struct cgraph_node *to,
188 struct cgraph_node *what)
190 gcov_type time = (((gcov_type)what->global.time
191 - inline_summary (what)->time_inlining_benefit)
192 * frequency + CGRAPH_FREQ_BASE / 2) / CGRAPH_FREQ_BASE
201 /* Estimate self time of the function after inlining WHAT into TO. */
204 cgraph_estimate_size_after_inlining (int times, struct cgraph_node *to,
205 struct cgraph_node *what)
207 int size = (what->global.size - inline_summary (what)->size_inlining_benefit) * times + to->global.size;
208 gcc_assert (size >= 0);
212 /* Scale frequency of NODE edges by FREQ_SCALE and increase loop nest
216 update_noncloned_frequencies (struct cgraph_node *node,
217 int freq_scale, int nest)
219 struct cgraph_edge *e;
221 /* We do not want to ignore high loop nest after freq drops to 0. */
224 for (e = node->callees; e; e = e->next_callee)
226 e->loop_nest += nest;
227 e->frequency = e->frequency * (gcov_type) freq_scale / CGRAPH_FREQ_BASE;
228 if (e->frequency > CGRAPH_FREQ_MAX)
229 e->frequency = CGRAPH_FREQ_MAX;
230 if (!e->inline_failed)
231 update_noncloned_frequencies (e->callee, freq_scale, nest);
235 /* E is expected to be an edge being inlined. Clone destination node of
236 the edge and redirect it to the new clone.
237 DUPLICATE is used for bookkeeping on whether we are actually creating new
238 clones or re-using node originally representing out-of-line function call.
241 cgraph_clone_inlined_nodes (struct cgraph_edge *e, bool duplicate,
242 bool update_original)
248 /* We may eliminate the need for out-of-line copy to be output.
249 In that case just go ahead and re-use it. */
250 if (!e->callee->callers->next_caller
251 && cgraph_can_remove_if_no_direct_calls_p (e->callee)
252 /* Don't reuse if more than one function shares a comdat group.
253 If the other function(s) are needed, we need to emit even
254 this function out of line. */
255 && !e->callee->same_comdat_group
256 && !cgraph_new_nodes)
258 gcc_assert (!e->callee->global.inlined_to);
259 if (e->callee->analyzed)
261 overall_size -= e->callee->global.size;
262 nfunctions_inlined++;
265 e->callee->local.externally_visible = false;
266 update_noncloned_frequencies (e->callee, e->frequency, e->loop_nest);
270 struct cgraph_node *n;
271 n = cgraph_clone_node (e->callee, e->callee->decl,
272 e->count, e->frequency, e->loop_nest,
273 update_original, NULL);
274 cgraph_redirect_edge_callee (e, n);
278 if (e->caller->global.inlined_to)
279 e->callee->global.inlined_to = e->caller->global.inlined_to;
281 e->callee->global.inlined_to = e->caller;
282 e->callee->global.stack_frame_offset
283 = e->caller->global.stack_frame_offset
284 + inline_summary (e->caller)->estimated_self_stack_size;
285 peak = e->callee->global.stack_frame_offset
286 + inline_summary (e->callee)->estimated_self_stack_size;
287 if (e->callee->global.inlined_to->global.estimated_stack_size < peak)
288 e->callee->global.inlined_to->global.estimated_stack_size = peak;
289 cgraph_propagate_frequency (e->callee);
291 /* Recursively clone all bodies. */
292 for (e = e->callee->callees; e; e = e->next_callee)
293 if (!e->inline_failed)
294 cgraph_clone_inlined_nodes (e, duplicate, update_original);
297 /* Mark edge E as inlined and update callgraph accordingly. UPDATE_ORIGINAL
298 specify whether profile of original function should be updated. If any new
299 indirect edges are discovered in the process, add them to NEW_EDGES, unless
300 it is NULL. Return true iff any new callgraph edges were discovered as a
301 result of inlining. */
304 cgraph_mark_inline_edge (struct cgraph_edge *e, bool update_original,
305 VEC (cgraph_edge_p, heap) **new_edges)
307 int old_size = 0, new_size = 0;
308 struct cgraph_node *to = NULL, *what;
309 struct cgraph_edge *curr = e;
312 gcc_assert (e->inline_failed);
313 e->inline_failed = CIF_OK;
314 DECL_POSSIBLY_INLINED (e->callee->decl) = true;
316 if (!e->callee->global.inlined)
317 DECL_POSSIBLY_INLINED (e->callee->decl) = true;
318 e->callee->global.inlined = true;
320 cgraph_clone_inlined_nodes (e, true, update_original);
325 /* Now update size of caller and all functions caller is inlined into. */
326 for (;e && !e->inline_failed; e = e->caller->callers)
329 old_size = e->caller->global.size;
330 new_size = cgraph_estimate_size_after_inlining (1, to, what);
331 to->global.size = new_size;
332 to->global.time = cgraph_estimate_time_after_inlining (freq, to, what);
334 gcc_assert (what->global.inlined_to == to);
335 if (new_size > old_size)
336 overall_size += new_size - old_size;
339 if (flag_indirect_inlining)
340 return ipa_propagate_indirect_call_infos (curr, new_edges);
345 /* Mark all calls of EDGE->CALLEE inlined into EDGE->CALLER. */
348 cgraph_mark_inline (struct cgraph_edge *edge)
350 struct cgraph_node *to = edge->caller;
351 struct cgraph_node *what = edge->callee;
352 struct cgraph_edge *e, *next;
354 gcc_assert (!edge->call_stmt_cannot_inline_p);
355 /* Look for all calls, mark them inline and clone recursively
356 all inlined functions. */
357 for (e = what->callers; e; e = next)
359 next = e->next_caller;
360 if (e->caller == to && e->inline_failed)
362 cgraph_mark_inline_edge (e, true, NULL);
369 /* Estimate the growth caused by inlining NODE into all callees. */
372 cgraph_estimate_growth (struct cgraph_node *node)
375 struct cgraph_edge *e;
376 bool self_recursive = false;
378 if (node->global.estimated_growth != INT_MIN)
379 return node->global.estimated_growth;
381 for (e = node->callers; e; e = e->next_caller)
383 if (e->caller == node)
384 self_recursive = true;
385 if (e->inline_failed)
386 growth += (cgraph_estimate_size_after_inlining (1, e->caller, node)
387 - e->caller->global.size);
390 /* ??? Wrong for non-trivially self recursive functions or cases where
391 we decide to not inline for different reasons, but it is not big deal
392 as in that case we will keep the body around, but we will also avoid
394 if (cgraph_only_called_directly_p (node)
395 && !DECL_EXTERNAL (node->decl) && !self_recursive)
396 growth -= node->global.size;
398 node->global.estimated_growth = growth;
402 /* Return false when inlining WHAT into TO is not good idea
403 as it would cause too large growth of function bodies.
404 When ONE_ONLY is true, assume that only one call site is going
405 to be inlined, otherwise figure out how many call sites in
406 TO calls WHAT and verify that all can be inlined.
410 cgraph_check_inline_limits (struct cgraph_node *to, struct cgraph_node *what,
411 cgraph_inline_failed_t *reason, bool one_only)
414 struct cgraph_edge *e;
417 HOST_WIDE_INT stack_size_limit, inlined_stack;
422 for (e = to->callees; e; e = e->next_callee)
423 if (e->callee == what)
426 if (to->global.inlined_to)
427 to = to->global.inlined_to;
429 /* When inlining large function body called once into small function,
430 take the inlined function as base for limiting the growth. */
431 if (inline_summary (to)->self_size > inline_summary(what)->self_size)
432 limit = inline_summary (to)->self_size;
434 limit = inline_summary (what)->self_size;
436 limit += limit * PARAM_VALUE (PARAM_LARGE_FUNCTION_GROWTH) / 100;
438 /* Check the size after inlining against the function limits. But allow
439 the function to shrink if it went over the limits by forced inlining. */
440 newsize = cgraph_estimate_size_after_inlining (times, to, what);
441 if (newsize >= to->global.size
442 && newsize > PARAM_VALUE (PARAM_LARGE_FUNCTION_INSNS)
446 *reason = CIF_LARGE_FUNCTION_GROWTH_LIMIT;
450 stack_size_limit = inline_summary (to)->estimated_self_stack_size;
452 stack_size_limit += stack_size_limit * PARAM_VALUE (PARAM_STACK_FRAME_GROWTH) / 100;
454 inlined_stack = (to->global.stack_frame_offset
455 + inline_summary (to)->estimated_self_stack_size
456 + what->global.estimated_stack_size);
457 if (inlined_stack > stack_size_limit
458 && inlined_stack > PARAM_VALUE (PARAM_LARGE_STACK_FRAME))
461 *reason = CIF_LARGE_STACK_FRAME_GROWTH_LIMIT;
467 /* Return true when function N is small enough to be inlined. */
470 cgraph_default_inline_p (struct cgraph_node *n, cgraph_inline_failed_t *reason)
474 if (n->local.disregard_inline_limits)
477 if (!flag_inline_small_functions && !DECL_DECLARED_INLINE_P (decl))
480 *reason = CIF_FUNCTION_NOT_INLINE_CANDIDATE;
487 *reason = CIF_BODY_NOT_AVAILABLE;
491 if (DECL_DECLARED_INLINE_P (decl))
493 if (n->global.size >= MAX_INLINE_INSNS_SINGLE)
496 *reason = CIF_MAX_INLINE_INSNS_SINGLE_LIMIT;
502 if (n->global.size >= MAX_INLINE_INSNS_AUTO)
505 *reason = CIF_MAX_INLINE_INSNS_AUTO_LIMIT;
513 /* Return true when inlining WHAT would create recursive inlining.
514 We call recursive inlining all cases where same function appears more than
515 once in the single recursion nest path in the inline graph. */
518 cgraph_recursive_inlining_p (struct cgraph_node *to,
519 struct cgraph_node *what,
520 cgraph_inline_failed_t *reason)
523 if (to->global.inlined_to)
524 recursive = what->decl == to->global.inlined_to->decl;
526 recursive = what->decl == to->decl;
527 /* Marking recursive function inline has sane semantic and thus we should
529 if (recursive && reason)
530 *reason = (what->local.disregard_inline_limits
531 ? CIF_RECURSIVE_INLINING : CIF_UNSPECIFIED);
535 /* A cost model driving the inlining heuristics in a way so the edges with
536 smallest badness are inlined first. After each inlining is performed
537 the costs of all caller edges of nodes affected are recomputed so the
538 metrics may accurately depend on values such as number of inlinable callers
539 of the function or function body size. */
542 cgraph_edge_badness (struct cgraph_edge *edge, bool dump)
546 (cgraph_estimate_size_after_inlining (1, edge->caller, edge->callee)
547 - edge->caller->global.size);
551 fprintf (dump_file, " Badness calculcation for %s -> %s\n",
552 cgraph_node_name (edge->caller),
553 cgraph_node_name (edge->callee));
554 fprintf (dump_file, " growth %i, time %i-%i, size %i-%i\n",
556 edge->callee->global.time,
557 inline_summary (edge->callee)->time_inlining_benefit,
558 edge->callee->global.size,
559 inline_summary (edge->callee)->size_inlining_benefit);
562 /* Always prefer inlining saving code size. */
565 badness = INT_MIN - growth;
567 fprintf (dump_file, " %i: Growth %i < 0\n", (int) badness,
571 /* When profiling is available, base priorities -(#calls / growth).
572 So we optimize for overall number of "executed" inlined calls. */
577 ((double) edge->count * INT_MIN / max_count / (max_benefit + 1)) *
578 (inline_summary (edge->callee)->time_inlining_benefit + 1)) / growth;
582 " %i (relative %f): profile info. Relative count %f"
583 " * Relative benefit %f\n",
584 (int) badness, (double) badness / INT_MIN,
585 (double) edge->count / max_count,
586 (double) (inline_summary (edge->callee)->
587 time_inlining_benefit + 1) / (max_benefit + 1));
591 /* When function local profile is available, base priorities on
592 growth / frequency, so we optimize for overall frequency of inlined
593 calls. This is not too accurate since while the call might be frequent
594 within function, the function itself is infrequent.
596 Other objective to optimize for is number of different calls inlined.
597 We add the estimated growth after inlining all functions to bias the
598 priorities slightly in this direction (so fewer times called functions
599 of the same size gets priority). */
600 else if (flag_guess_branch_prob)
602 int div = edge->frequency * 100 / CGRAPH_FREQ_BASE + 1;
605 badness = growth * 10000;
607 MIN (100 * inline_summary (edge->callee)->time_inlining_benefit /
608 (edge->callee->global.time + 1) +1, 100);
612 /* Decrease badness if call is nested. */
613 /* Compress the range so we don't overflow. */
615 div = 10000 + ceil_log2 (div) - 8;
620 growth_for_all = cgraph_estimate_growth (edge->callee);
621 badness += growth_for_all;
622 if (badness > INT_MAX)
627 " %i: guessed profile. frequency %i, overall growth %i,"
628 " benefit %i%%, divisor %i\n",
629 (int) badness, edge->frequency, growth_for_all, benefitperc, div);
632 /* When function local profile is not available or it does not give
633 useful information (ie frequency is zero), base the cost on
634 loop nest and overall size growth, so we optimize for overall number
635 of functions fully inlined in program. */
638 int nest = MIN (edge->loop_nest, 8);
639 badness = cgraph_estimate_growth (edge->callee) * 256;
641 /* Decrease badness if call is nested. */
649 fprintf (dump_file, " %i: no profile. nest %i\n", (int) badness,
653 /* Ensure that we did not overflow in all the fixed point math above. */
654 gcc_assert (badness >= INT_MIN);
655 gcc_assert (badness <= INT_MAX - 1);
656 /* Make recursive inlining happen always after other inlining is done. */
657 if (cgraph_recursive_inlining_p (edge->caller, edge->callee, NULL))
663 /* Recompute heap nodes for each of caller edge. */
666 update_caller_keys (fibheap_t heap, struct cgraph_node *node,
667 bitmap updated_nodes)
669 struct cgraph_edge *edge;
670 cgraph_inline_failed_t failed_reason;
672 if (!node->local.inlinable
673 || node->global.inlined_to)
675 if (bitmap_bit_p (updated_nodes, node->uid))
677 bitmap_set_bit (updated_nodes, node->uid);
678 node->global.estimated_growth = INT_MIN;
680 if (!node->local.inlinable)
682 /* Prune out edges we won't inline into anymore. */
683 if (!cgraph_default_inline_p (node, &failed_reason))
685 for (edge = node->callers; edge; edge = edge->next_caller)
688 fibheap_delete_node (heap, (fibnode_t) edge->aux);
690 if (edge->inline_failed)
691 edge->inline_failed = failed_reason;
696 for (edge = node->callers; edge; edge = edge->next_caller)
697 if (edge->inline_failed)
699 int badness = cgraph_edge_badness (edge, false);
702 fibnode_t n = (fibnode_t) edge->aux;
703 gcc_assert (n->data == edge);
704 if (n->key == badness)
707 /* fibheap_replace_key only increase the keys. */
708 if (badness < n->key)
710 fibheap_replace_key (heap, n, badness);
711 gcc_assert (n->key == badness);
714 fibheap_delete_node (heap, (fibnode_t) edge->aux);
716 edge->aux = fibheap_insert (heap, badness, edge);
720 /* Recompute heap nodes for each of caller edges of each of callees. */
723 update_callee_keys (fibheap_t heap, struct cgraph_node *node,
724 bitmap updated_nodes)
726 struct cgraph_edge *e;
727 node->global.estimated_growth = INT_MIN;
729 for (e = node->callees; e; e = e->next_callee)
730 if (e->inline_failed)
731 update_caller_keys (heap, e->callee, updated_nodes);
732 else if (!e->inline_failed)
733 update_callee_keys (heap, e->callee, updated_nodes);
736 /* Enqueue all recursive calls from NODE into priority queue depending on
737 how likely we want to recursively inline the call. */
740 lookup_recursive_calls (struct cgraph_node *node, struct cgraph_node *where,
744 struct cgraph_edge *e;
745 for (e = where->callees; e; e = e->next_callee)
746 if (e->callee == node)
748 /* When profile feedback is available, prioritize by expected number
749 of calls. Without profile feedback we maintain simple queue
750 to order candidates via recursive depths. */
751 fibheap_insert (heap,
752 !max_count ? priority++
753 : -(e->count / ((max_count + (1<<24) - 1) / (1<<24))),
756 for (e = where->callees; e; e = e->next_callee)
757 if (!e->inline_failed)
758 lookup_recursive_calls (node, e->callee, heap);
761 /* Decide on recursive inlining: in the case function has recursive calls,
762 inline until body size reaches given argument. If any new indirect edges
763 are discovered in the process, add them to *NEW_EDGES, unless NEW_EDGES
767 cgraph_decide_recursive_inlining (struct cgraph_node *node,
768 VEC (cgraph_edge_p, heap) **new_edges)
770 int limit = PARAM_VALUE (PARAM_MAX_INLINE_INSNS_RECURSIVE_AUTO);
771 int max_depth = PARAM_VALUE (PARAM_MAX_INLINE_RECURSIVE_DEPTH_AUTO);
772 int probability = PARAM_VALUE (PARAM_MIN_INLINE_RECURSIVE_PROBABILITY);
774 struct cgraph_edge *e;
775 struct cgraph_node *master_clone, *next;
779 /* It does not make sense to recursively inline always-inline functions
780 as we are going to sorry() on the remaining calls anyway. */
781 if (node->local.disregard_inline_limits
782 && lookup_attribute ("always_inline", DECL_ATTRIBUTES (node->decl)))
785 if (optimize_function_for_size_p (DECL_STRUCT_FUNCTION (node->decl))
786 || (!flag_inline_functions && !DECL_DECLARED_INLINE_P (node->decl)))
789 if (DECL_DECLARED_INLINE_P (node->decl))
791 limit = PARAM_VALUE (PARAM_MAX_INLINE_INSNS_RECURSIVE);
792 max_depth = PARAM_VALUE (PARAM_MAX_INLINE_RECURSIVE_DEPTH);
795 /* Make sure that function is small enough to be considered for inlining. */
797 || cgraph_estimate_size_after_inlining (1, node, node) >= limit)
799 heap = fibheap_new ();
800 lookup_recursive_calls (node, node, heap);
801 if (fibheap_empty (heap))
803 fibheap_delete (heap);
809 " Performing recursive inlining on %s\n",
810 cgraph_node_name (node));
812 /* We need original clone to copy around. */
813 master_clone = cgraph_clone_node (node, node->decl,
814 node->count, CGRAPH_FREQ_BASE, 1,
816 master_clone->needed = true;
817 for (e = master_clone->callees; e; e = e->next_callee)
818 if (!e->inline_failed)
819 cgraph_clone_inlined_nodes (e, true, false);
821 /* Do the inlining and update list of recursive call during process. */
822 while (!fibheap_empty (heap)
823 && (cgraph_estimate_size_after_inlining (1, node, master_clone)
826 struct cgraph_edge *curr
827 = (struct cgraph_edge *) fibheap_extract_min (heap);
828 struct cgraph_node *cnode;
831 for (cnode = curr->caller;
832 cnode->global.inlined_to; cnode = cnode->callers->caller)
833 if (node->decl == curr->callee->decl)
835 if (depth > max_depth)
839 " maximal depth reached\n");
845 if (!cgraph_maybe_hot_edge_p (curr))
848 fprintf (dump_file, " Not inlining cold call\n");
851 if (curr->count * 100 / node->count < probability)
855 " Probability of edge is too small\n");
863 " Inlining call of depth %i", depth);
866 fprintf (dump_file, " called approx. %.2f times per call",
867 (double)curr->count / node->count);
869 fprintf (dump_file, "\n");
871 cgraph_redirect_edge_callee (curr, master_clone);
872 cgraph_mark_inline_edge (curr, false, new_edges);
873 lookup_recursive_calls (node, curr->callee, heap);
876 if (!fibheap_empty (heap) && dump_file)
877 fprintf (dump_file, " Recursive inlining growth limit met.\n");
879 fibheap_delete (heap);
882 "\n Inlined %i times, body grown from size %i to %i, time %i to %i\n", n,
883 master_clone->global.size, node->global.size,
884 master_clone->global.time, node->global.time);
886 /* Remove master clone we used for inlining. We rely that clones inlined
887 into master clone gets queued just before master clone so we don't
889 for (node = cgraph_nodes; node != master_clone;
893 if (node->global.inlined_to == master_clone)
894 cgraph_remove_node (node);
896 cgraph_remove_node (master_clone);
897 /* FIXME: Recursive inlining actually reduces number of calls of the
898 function. At this place we should probably walk the function and
899 inline clones and compensate the counts accordingly. This probably
900 doesn't matter much in practice. */
904 /* Set inline_failed for all callers of given function to REASON. */
907 cgraph_set_inline_failed (struct cgraph_node *node,
908 cgraph_inline_failed_t reason)
910 struct cgraph_edge *e;
913 fprintf (dump_file, "Inlining failed: %s\n",
914 cgraph_inline_failed_string (reason));
915 for (e = node->callers; e; e = e->next_caller)
916 if (e->inline_failed)
917 e->inline_failed = reason;
920 /* Given whole compilation unit estimate of INSNS, compute how large we can
921 allow the unit to grow. */
923 compute_max_insns (int insns)
925 int max_insns = insns;
926 if (max_insns < PARAM_VALUE (PARAM_LARGE_UNIT_INSNS))
927 max_insns = PARAM_VALUE (PARAM_LARGE_UNIT_INSNS);
929 return ((HOST_WIDEST_INT) max_insns
930 * (100 + PARAM_VALUE (PARAM_INLINE_UNIT_GROWTH)) / 100);
933 /* Compute badness of all edges in NEW_EDGES and add them to the HEAP. */
935 add_new_edges_to_heap (fibheap_t heap, VEC (cgraph_edge_p, heap) *new_edges)
937 while (VEC_length (cgraph_edge_p, new_edges) > 0)
939 struct cgraph_edge *edge = VEC_pop (cgraph_edge_p, new_edges);
941 gcc_assert (!edge->aux);
942 edge->aux = fibheap_insert (heap, cgraph_edge_badness (edge, false), edge);
947 /* We use greedy algorithm for inlining of small functions:
948 All inline candidates are put into prioritized heap based on estimated
949 growth of the overall number of instructions and then update the estimates.
951 INLINED and INLINED_CALEES are just pointers to arrays large enough
952 to be passed to cgraph_inlined_into and cgraph_inlined_callees. */
955 cgraph_decide_inlining_of_small_functions (void)
957 struct cgraph_node *node;
958 struct cgraph_edge *edge;
959 cgraph_inline_failed_t failed_reason;
960 fibheap_t heap = fibheap_new ();
961 bitmap updated_nodes = BITMAP_ALLOC (NULL);
962 int min_size, max_size;
963 VEC (cgraph_edge_p, heap) *new_indirect_edges = NULL;
965 if (flag_indirect_inlining)
966 new_indirect_edges = VEC_alloc (cgraph_edge_p, heap, 8);
969 fprintf (dump_file, "\nDeciding on smaller functions:\n");
971 /* Put all inline candidates into the heap. */
973 for (node = cgraph_nodes; node; node = node->next)
975 if (!node->local.inlinable || !node->callers)
978 fprintf (dump_file, "Considering inline candidate %s.\n", cgraph_node_name (node));
980 node->global.estimated_growth = INT_MIN;
981 if (!cgraph_default_inline_p (node, &failed_reason))
983 cgraph_set_inline_failed (node, failed_reason);
987 for (edge = node->callers; edge; edge = edge->next_caller)
988 if (edge->inline_failed)
990 gcc_assert (!edge->aux);
991 edge->aux = fibheap_insert (heap, cgraph_edge_badness (edge, false), edge);
995 max_size = compute_max_insns (overall_size);
996 min_size = overall_size;
998 while (overall_size <= max_size
999 && !fibheap_empty (heap))
1001 int old_size = overall_size;
1002 struct cgraph_node *where, *callee;
1003 int badness = fibheap_min_key (heap);
1005 cgraph_inline_failed_t not_good = CIF_OK;
1007 edge = (struct cgraph_edge *) fibheap_extract_min (heap);
1008 gcc_assert (edge->aux);
1010 if (!edge->inline_failed)
1012 #ifdef ENABLE_CHECKING
1013 gcc_assert (cgraph_edge_badness (edge, false) == badness);
1015 callee = edge->callee;
1017 growth = (cgraph_estimate_size_after_inlining (1, edge->caller, edge->callee)
1018 - edge->caller->global.size);
1023 "\nConsidering %s with %i size\n",
1024 cgraph_node_name (edge->callee),
1025 edge->callee->global.size);
1027 " to be inlined into %s in %s:%i\n"
1028 " Estimated growth after inlined into all callees is %+i insns.\n"
1029 " Estimated badness is %i, frequency %.2f.\n",
1030 cgraph_node_name (edge->caller),
1031 flag_wpa ? "unknown"
1032 : gimple_filename ((const_gimple) edge->call_stmt),
1033 flag_wpa ? -1 : gimple_lineno ((const_gimple) edge->call_stmt),
1034 cgraph_estimate_growth (edge->callee),
1036 edge->frequency / (double)CGRAPH_FREQ_BASE);
1038 fprintf (dump_file," Called "HOST_WIDEST_INT_PRINT_DEC"x\n", edge->count);
1039 if (dump_flags & TDF_DETAILS)
1040 cgraph_edge_badness (edge, true);
1043 /* When not having profile info ready we don't weight by any way the
1044 position of call in procedure itself. This means if call of
1045 function A from function B seems profitable to inline, the recursive
1046 call of function A in inline copy of A in B will look profitable too
1047 and we end up inlining until reaching maximal function growth. This
1048 is not good idea so prohibit the recursive inlining.
1050 ??? When the frequencies are taken into account we might not need this
1053 We need to be cureful here, in some testcases, e.g. directivec.c in
1054 libcpp, we can estimate self recursive function to have negative growth
1055 for inlining completely.
1059 where = edge->caller;
1060 while (where->global.inlined_to)
1062 if (where->decl == edge->callee->decl)
1064 where = where->callers->caller;
1066 if (where->global.inlined_to)
1069 = (edge->callee->local.disregard_inline_limits
1070 ? CIF_RECURSIVE_INLINING : CIF_UNSPECIFIED);
1072 fprintf (dump_file, " inline_failed:Recursive inlining performed only for function itself.\n");
1077 if (edge->callee->local.disregard_inline_limits)
1079 else if (!cgraph_maybe_hot_edge_p (edge))
1080 not_good = CIF_UNLIKELY_CALL;
1081 else if (!flag_inline_functions
1082 && !DECL_DECLARED_INLINE_P (edge->callee->decl))
1083 not_good = CIF_NOT_DECLARED_INLINED;
1084 else if (optimize_function_for_size_p (DECL_STRUCT_FUNCTION(edge->caller->decl)))
1085 not_good = CIF_OPTIMIZING_FOR_SIZE;
1086 if (not_good && growth > 0 && cgraph_estimate_growth (edge->callee) > 0)
1088 if (!cgraph_recursive_inlining_p (edge->caller, edge->callee,
1089 &edge->inline_failed))
1091 edge->inline_failed = not_good;
1093 fprintf (dump_file, " inline_failed:%s.\n",
1094 cgraph_inline_failed_string (edge->inline_failed));
1098 if (!cgraph_default_inline_p (edge->callee, &edge->inline_failed))
1100 if (!cgraph_recursive_inlining_p (edge->caller, edge->callee,
1101 &edge->inline_failed))
1104 fprintf (dump_file, " inline_failed:%s.\n",
1105 cgraph_inline_failed_string (edge->inline_failed));
1109 if (!tree_can_inline_p (edge))
1112 fprintf (dump_file, " inline_failed:%s.\n",
1113 cgraph_inline_failed_string (edge->inline_failed));
1116 if (cgraph_recursive_inlining_p (edge->caller, edge->callee,
1117 &edge->inline_failed))
1119 where = edge->caller;
1120 if (where->global.inlined_to)
1121 where = where->global.inlined_to;
1122 if (!cgraph_decide_recursive_inlining (where,
1123 flag_indirect_inlining
1124 ? &new_indirect_edges : NULL))
1126 if (flag_indirect_inlining)
1127 add_new_edges_to_heap (heap, new_indirect_edges);
1128 update_callee_keys (heap, where, updated_nodes);
1132 struct cgraph_node *callee;
1133 if (edge->call_stmt_cannot_inline_p
1134 || !cgraph_check_inline_limits (edge->caller, edge->callee,
1135 &edge->inline_failed, true))
1138 fprintf (dump_file, " Not inlining into %s:%s.\n",
1139 cgraph_node_name (edge->caller),
1140 cgraph_inline_failed_string (edge->inline_failed));
1143 callee = edge->callee;
1144 cgraph_mark_inline_edge (edge, true, &new_indirect_edges);
1145 if (flag_indirect_inlining)
1146 add_new_edges_to_heap (heap, new_indirect_edges);
1148 update_callee_keys (heap, callee, updated_nodes);
1150 where = edge->caller;
1151 if (where->global.inlined_to)
1152 where = where->global.inlined_to;
1154 /* Our profitability metric can depend on local properties
1155 such as number of inlinable calls and size of the function body.
1156 After inlining these properties might change for the function we
1157 inlined into (since it's body size changed) and for the functions
1158 called by function we inlined (since number of it inlinable callers
1160 update_caller_keys (heap, where, updated_nodes);
1162 /* We removed one call of the function we just inlined. If offline
1163 copy is still needed, be sure to update the keys. */
1164 if (callee != where && !callee->global.inlined_to)
1165 update_caller_keys (heap, callee, updated_nodes);
1166 bitmap_clear (updated_nodes);
1171 " Inlined into %s which now has size %i and self time %i,"
1172 "net change of %+i.\n",
1173 cgraph_node_name (edge->caller),
1174 edge->caller->global.time,
1175 edge->caller->global.size,
1176 overall_size - old_size);
1178 if (min_size > overall_size)
1180 min_size = overall_size;
1181 max_size = compute_max_insns (min_size);
1184 fprintf (dump_file, "New minimal size reached: %i\n", min_size);
1187 while (!fibheap_empty (heap))
1189 int badness = fibheap_min_key (heap);
1191 edge = (struct cgraph_edge *) fibheap_extract_min (heap);
1192 gcc_assert (edge->aux);
1194 if (!edge->inline_failed)
1196 #ifdef ENABLE_CHECKING
1197 gcc_assert (cgraph_edge_badness (edge, false) == badness);
1202 "\nSkipping %s with %i size\n",
1203 cgraph_node_name (edge->callee),
1204 edge->callee->global.size);
1206 " called by %s in %s:%i\n"
1207 " Estimated growth after inlined into all callees is %+i insns.\n"
1208 " Estimated badness is %i, frequency %.2f.\n",
1209 cgraph_node_name (edge->caller),
1210 gimple_filename ((const_gimple) edge->call_stmt),
1211 gimple_lineno ((const_gimple) edge->call_stmt),
1212 cgraph_estimate_growth (edge->callee),
1214 edge->frequency / (double)CGRAPH_FREQ_BASE);
1216 fprintf (dump_file," Called "HOST_WIDEST_INT_PRINT_DEC"x\n", edge->count);
1217 if (dump_flags & TDF_DETAILS)
1218 cgraph_edge_badness (edge, true);
1220 if (!edge->callee->local.disregard_inline_limits && edge->inline_failed
1221 && !cgraph_recursive_inlining_p (edge->caller, edge->callee,
1222 &edge->inline_failed))
1223 edge->inline_failed = CIF_INLINE_UNIT_GROWTH_LIMIT;
1226 if (new_indirect_edges)
1227 VEC_free (cgraph_edge_p, heap, new_indirect_edges);
1228 fibheap_delete (heap);
1229 BITMAP_FREE (updated_nodes);
1232 /* Flatten NODE from the IPA inliner. */
1235 cgraph_flatten (struct cgraph_node *node)
1237 struct cgraph_edge *e;
1239 /* We shouldn't be called recursively when we are being processed. */
1240 gcc_assert (node->aux == NULL);
1242 node->aux = (void *)(size_t) INLINE_ALL;
1244 for (e = node->callees; e; e = e->next_callee)
1246 struct cgraph_node *orig_callee;
1248 if (e->call_stmt_cannot_inline_p)
1251 if (!e->callee->analyzed)
1255 "Not inlining: Function body not available.\n");
1259 /* We've hit cycle? It is time to give up. */
1264 "Not inlining %s into %s to avoid cycle.\n",
1265 cgraph_node_name (e->callee),
1266 cgraph_node_name (e->caller));
1267 e->inline_failed = CIF_RECURSIVE_INLINING;
1271 /* When the edge is already inlined, we just need to recurse into
1272 it in order to fully flatten the leaves. */
1273 if (!e->inline_failed)
1275 cgraph_flatten (e->callee);
1279 if (cgraph_recursive_inlining_p (node, e->callee, &e->inline_failed))
1282 fprintf (dump_file, "Not inlining: recursive call.\n");
1286 if (!tree_can_inline_p (e))
1289 fprintf (dump_file, "Not inlining: %s",
1290 cgraph_inline_failed_string (e->inline_failed));
1294 /* Inline the edge and flatten the inline clone. Avoid
1295 recursing through the original node if the node was cloned. */
1297 fprintf (dump_file, " Inlining %s into %s.\n",
1298 cgraph_node_name (e->callee),
1299 cgraph_node_name (e->caller));
1300 orig_callee = e->callee;
1301 cgraph_mark_inline_edge (e, true, NULL);
1302 if (e->callee != orig_callee)
1303 orig_callee->aux = (void *)(size_t) INLINE_ALL;
1304 cgraph_flatten (e->callee);
1305 if (e->callee != orig_callee)
1306 orig_callee->aux = NULL;
1312 /* Decide on the inlining. We do so in the topological order to avoid
1313 expenses on updating data structures. */
1316 cgraph_decide_inlining (void)
1318 struct cgraph_node *node;
1320 struct cgraph_node **order =
1321 XCNEWVEC (struct cgraph_node *, cgraph_n_nodes);
1324 int initial_size = 0;
1326 cgraph_remove_function_insertion_hook (function_insertion_hook_holder);
1327 if (in_lto_p && flag_indirect_inlining)
1328 ipa_update_after_lto_read ();
1329 if (flag_indirect_inlining)
1330 ipa_create_all_structures_for_iinln ();
1334 for (node = cgraph_nodes; node; node = node->next)
1337 struct cgraph_edge *e;
1339 gcc_assert (inline_summary (node)->self_size == node->global.size);
1340 initial_size += node->global.size;
1341 for (e = node->callees; e; e = e->next_callee)
1342 if (max_count < e->count)
1343 max_count = e->count;
1344 if (max_benefit < inline_summary (node)->time_inlining_benefit)
1345 max_benefit = inline_summary (node)->time_inlining_benefit;
1347 gcc_assert (in_lto_p
1349 || (profile_info && flag_branch_probabilities));
1350 overall_size = initial_size;
1352 nnodes = cgraph_postorder (order);
1356 "\nDeciding on inlining. Starting with size %i.\n",
1359 for (node = cgraph_nodes; node; node = node->next)
1363 fprintf (dump_file, "\nFlattening functions:\n");
1365 /* In the first pass handle functions to be flattened. Do this with
1366 a priority so none of our later choices will make this impossible. */
1367 for (i = nnodes - 1; i >= 0; i--)
1371 /* Handle nodes to be flattened, but don't update overall unit
1372 size. Calling the incremental inliner here is lame,
1373 a simple worklist should be enough. What should be left
1374 here from the early inliner (if it runs) is cyclic cases.
1375 Ideally when processing callees we stop inlining at the
1376 entry of cycles, possibly cloning that entry point and
1377 try to flatten itself turning it into a self-recursive
1379 if (lookup_attribute ("flatten",
1380 DECL_ATTRIBUTES (node->decl)) != NULL)
1384 "Flattening %s\n", cgraph_node_name (node));
1385 cgraph_flatten (node);
1389 cgraph_decide_inlining_of_small_functions ();
1391 if (flag_inline_functions_called_once)
1394 fprintf (dump_file, "\nDeciding on functions called once:\n");
1396 /* And finally decide what functions are called once. */
1397 for (i = nnodes - 1; i >= 0; i--)
1402 && !node->callers->next_caller
1403 && cgraph_only_called_directly_p (node)
1404 && node->local.inlinable
1405 && node->callers->inline_failed
1406 && node->callers->caller != node
1407 && node->callers->caller->global.inlined_to != node
1408 && !node->callers->call_stmt_cannot_inline_p
1409 && !DECL_EXTERNAL (node->decl)
1410 && !DECL_COMDAT (node->decl))
1412 cgraph_inline_failed_t reason;
1413 old_size = overall_size;
1417 "\nConsidering %s size %i.\n",
1418 cgraph_node_name (node), node->global.size);
1420 " Called once from %s %i insns.\n",
1421 cgraph_node_name (node->callers->caller),
1422 node->callers->caller->global.size);
1425 if (cgraph_check_inline_limits (node->callers->caller, node,
1428 struct cgraph_node *caller = node->callers->caller;
1429 cgraph_mark_inline (node->callers);
1432 " Inlined into %s which now has %i size"
1433 " for a net change of %+i size.\n",
1434 cgraph_node_name (caller),
1435 caller->global.size,
1436 overall_size - old_size);
1442 " Not inlining: %s.\n",
1443 cgraph_inline_failed_string (reason));
1449 /* Free ipa-prop structures if they are no longer needed. */
1450 if (flag_indirect_inlining)
1451 ipa_free_all_structures_after_iinln ();
1455 "\nInlined %i calls, eliminated %i functions, "
1456 "size %i turned to %i size.\n\n",
1457 ncalls_inlined, nfunctions_inlined, initial_size,
1463 /* Return true when N is leaf function. Accept cheap (pure&const) builtins
1464 in leaf functions. */
1466 leaf_node_p (struct cgraph_node *n)
1468 struct cgraph_edge *e;
1469 for (e = n->callees; e; e = e->next_callee)
1470 if (!DECL_BUILT_IN (e->callee->decl)
1471 || (!TREE_READONLY (e->callee->decl)
1472 || DECL_PURE_P (e->callee->decl)))
1477 /* Decide on the inlining. We do so in the topological order to avoid
1478 expenses on updating data structures. */
1481 cgraph_decide_inlining_incrementally (struct cgraph_node *node,
1482 enum inlining_mode mode)
1484 struct cgraph_edge *e;
1485 bool inlined = false;
1486 cgraph_inline_failed_t failed_reason;
1488 #ifdef ENABLE_CHECKING
1489 verify_cgraph_node (node);
1492 if (mode != INLINE_ALWAYS_INLINE && mode != INLINE_SIZE_NORECURSIVE
1493 && lookup_attribute ("flatten", DECL_ATTRIBUTES (node->decl)) != NULL)
1496 fprintf (dump_file, "Incrementally flattening %s\n",
1497 cgraph_node_name (node));
1501 /* First of all look for always inline functions. */
1502 if (mode != INLINE_SIZE_NORECURSIVE)
1503 for (e = node->callees; e; e = e->next_callee)
1505 if (!e->callee->local.disregard_inline_limits
1506 && (mode != INLINE_ALL || !e->callee->local.inlinable))
1508 if (e->call_stmt_cannot_inline_p)
1512 "Considering to always inline inline candidate %s.\n",
1513 cgraph_node_name (e->callee));
1514 if (cgraph_recursive_inlining_p (node, e->callee, &e->inline_failed))
1517 fprintf (dump_file, "Not inlining: recursive call.\n");
1520 if (!tree_can_inline_p (e))
1525 cgraph_inline_failed_string (e->inline_failed));
1528 if (gimple_in_ssa_p (DECL_STRUCT_FUNCTION (node->decl))
1529 != gimple_in_ssa_p (DECL_STRUCT_FUNCTION (e->callee->decl)))
1532 fprintf (dump_file, "Not inlining: SSA form does not match.\n");
1535 if (!e->callee->analyzed)
1539 "Not inlining: Function body no longer available.\n");
1544 fprintf (dump_file, " Inlining %s into %s.\n",
1545 cgraph_node_name (e->callee),
1546 cgraph_node_name (e->caller));
1547 cgraph_mark_inline (e);
1551 /* Now do the automatic inlining. */
1552 if (mode != INLINE_ALL && mode != INLINE_ALWAYS_INLINE
1553 /* Never inline regular functions into always-inline functions
1554 during incremental inlining. */
1555 && !node->local.disregard_inline_limits)
1557 bitmap visited = BITMAP_ALLOC (NULL);
1558 for (e = node->callees; e; e = e->next_callee)
1560 int allowed_growth = 0;
1561 if (!e->callee->local.inlinable
1562 || !e->inline_failed
1563 || e->callee->local.disregard_inline_limits)
1565 /* We are inlining a function to all call-sites in node
1566 or to none. So visit each candidate only once. */
1567 if (!bitmap_set_bit (visited, e->callee->uid))
1570 fprintf (dump_file, "Considering inline candidate %s.\n",
1571 cgraph_node_name (e->callee));
1572 if (cgraph_recursive_inlining_p (node, e->callee, &e->inline_failed))
1575 fprintf (dump_file, "Not inlining: recursive call.\n");
1578 if (gimple_in_ssa_p (DECL_STRUCT_FUNCTION (node->decl))
1579 != gimple_in_ssa_p (DECL_STRUCT_FUNCTION (e->callee->decl)))
1583 "Not inlining: SSA form does not match.\n");
1587 if (cgraph_maybe_hot_edge_p (e) && leaf_node_p (e->callee)
1588 && optimize_function_for_speed_p (cfun))
1589 allowed_growth = PARAM_VALUE (PARAM_EARLY_INLINING_INSNS);
1591 /* When the function body would grow and inlining the function
1592 won't eliminate the need for offline copy of the function,
1594 if (((mode == INLINE_SIZE || mode == INLINE_SIZE_NORECURSIVE)
1595 || (!flag_inline_functions
1596 && !DECL_DECLARED_INLINE_P (e->callee->decl)))
1597 && (cgraph_estimate_size_after_inlining (1, e->caller, e->callee)
1598 > e->caller->global.size + allowed_growth)
1599 && cgraph_estimate_growth (e->callee) > allowed_growth)
1603 "Not inlining: code size would grow by %i.\n",
1604 cgraph_estimate_size_after_inlining (1, e->caller,
1606 - e->caller->global.size);
1609 if (!cgraph_check_inline_limits (node, e->callee, &e->inline_failed,
1611 || e->call_stmt_cannot_inline_p)
1614 fprintf (dump_file, "Not inlining: %s.\n",
1615 cgraph_inline_failed_string (e->inline_failed));
1618 if (!e->callee->analyzed)
1622 "Not inlining: Function body no longer available.\n");
1625 if (!tree_can_inline_p (e))
1629 "Not inlining: %s.",
1630 cgraph_inline_failed_string (e->inline_failed));
1633 if (cgraph_default_inline_p (e->callee, &failed_reason))
1636 fprintf (dump_file, " Inlining %s into %s.\n",
1637 cgraph_node_name (e->callee),
1638 cgraph_node_name (e->caller));
1639 cgraph_mark_inline (e);
1643 BITMAP_FREE (visited);
1648 /* Because inlining might remove no-longer reachable nodes, we need to
1649 keep the array visible to garbage collector to avoid reading collected
1652 static GTY ((length ("nnodes"))) struct cgraph_node **order;
1654 /* Do inlining of small functions. Doing so early helps profiling and other
1655 passes to be somewhat more effective and avoids some code duplication in
1656 later real inlining pass for testcases with very many function calls. */
1658 cgraph_early_inlining (void)
1660 struct cgraph_node *node = cgraph_node (current_function_decl);
1661 unsigned int todo = 0;
1664 if (sorrycount || errorcount)
1669 || !flag_early_inlining)
1671 /* When not optimizing or not inlining inline only always-inline
1673 cgraph_decide_inlining_incrementally (node, INLINE_ALWAYS_INLINE);
1674 timevar_push (TV_INTEGRATION);
1675 todo |= optimize_inline_calls (current_function_decl);
1676 timevar_pop (TV_INTEGRATION);
1680 /* We iterate incremental inlining to get trivial cases of indirect
1682 while (iterations < PARAM_VALUE (PARAM_EARLY_INLINER_MAX_ITERATIONS)
1683 && cgraph_decide_inlining_incrementally (node,
1685 ? INLINE_SIZE_NORECURSIVE
1688 timevar_push (TV_INTEGRATION);
1689 todo |= optimize_inline_calls (current_function_decl);
1691 timevar_pop (TV_INTEGRATION);
1694 fprintf (dump_file, "Iterations: %i\n", iterations);
1697 cfun->always_inline_functions_inlined = true;
1702 struct gimple_opt_pass pass_early_inline =
1706 "einline", /* name */
1708 cgraph_early_inlining, /* execute */
1711 0, /* static_pass_number */
1712 TV_INLINE_HEURISTICS, /* tv_id */
1713 0, /* properties_required */
1714 0, /* properties_provided */
1715 0, /* properties_destroyed */
1716 0, /* todo_flags_start */
1717 TODO_dump_func /* todo_flags_finish */
1721 /* When inlining shall be performed. */
1723 cgraph_gate_ipa_early_inlining (void)
1725 return (flag_early_inlining
1727 && (flag_branch_probabilities || flag_test_coverage
1728 || profile_arc_flag));
1731 /* IPA pass wrapper for early inlining pass. We need to run early inlining
1732 before tree profiling so we have stand alone IPA pass for doing so. */
1733 struct simple_ipa_opt_pass pass_ipa_early_inline =
1737 "einline_ipa", /* name */
1738 cgraph_gate_ipa_early_inlining, /* gate */
1742 0, /* static_pass_number */
1743 TV_INLINE_HEURISTICS, /* tv_id */
1744 0, /* properties_required */
1745 0, /* properties_provided */
1746 0, /* properties_destroyed */
1747 0, /* todo_flags_start */
1748 TODO_dump_cgraph /* todo_flags_finish */
1752 /* See if statement might disappear after inlining. We are not terribly
1753 sophisficated, basically looking for simple abstraction penalty wrappers. */
1756 likely_eliminated_by_inlining_p (gimple stmt)
1758 enum gimple_code code = gimple_code (stmt);
1764 if (gimple_num_ops (stmt) != 2)
1767 /* Casts of parameters, loads from parameters passed by reference
1768 and stores to return value or parameters are probably free after
1770 if (gimple_assign_rhs_code (stmt) == CONVERT_EXPR
1771 || gimple_assign_rhs_code (stmt) == NOP_EXPR
1772 || gimple_assign_rhs_code (stmt) == VIEW_CONVERT_EXPR
1773 || gimple_assign_rhs_class (stmt) == GIMPLE_SINGLE_RHS)
1775 tree rhs = gimple_assign_rhs1 (stmt);
1776 tree lhs = gimple_assign_lhs (stmt);
1777 tree inner_rhs = rhs;
1778 tree inner_lhs = lhs;
1779 bool rhs_free = false;
1780 bool lhs_free = false;
1782 while (handled_component_p (inner_lhs) || TREE_CODE (inner_lhs) == INDIRECT_REF)
1783 inner_lhs = TREE_OPERAND (inner_lhs, 0);
1784 while (handled_component_p (inner_rhs)
1785 || TREE_CODE (inner_rhs) == ADDR_EXPR || TREE_CODE (inner_rhs) == INDIRECT_REF)
1786 inner_rhs = TREE_OPERAND (inner_rhs, 0);
1789 if (TREE_CODE (inner_rhs) == PARM_DECL
1790 || (TREE_CODE (inner_rhs) == SSA_NAME
1791 && SSA_NAME_IS_DEFAULT_DEF (inner_rhs)
1792 && TREE_CODE (SSA_NAME_VAR (inner_rhs)) == PARM_DECL))
1794 if (rhs_free && is_gimple_reg (lhs))
1796 if (((TREE_CODE (inner_lhs) == PARM_DECL
1797 || (TREE_CODE (inner_lhs) == SSA_NAME
1798 && SSA_NAME_IS_DEFAULT_DEF (inner_lhs)
1799 && TREE_CODE (SSA_NAME_VAR (inner_lhs)) == PARM_DECL))
1800 && inner_lhs != lhs)
1801 || TREE_CODE (inner_lhs) == RESULT_DECL
1802 || (TREE_CODE (inner_lhs) == SSA_NAME
1803 && TREE_CODE (SSA_NAME_VAR (inner_lhs)) == RESULT_DECL))
1805 if (lhs_free && (is_gimple_reg (rhs) || is_gimple_min_invariant (rhs)))
1807 if (lhs_free && rhs_free)
1816 /* Compute function body size parameters for NODE. */
1819 estimate_function_body_sizes (struct cgraph_node *node)
1822 gcov_type time_inlining_benefit = 0;
1824 int size_inlining_benefit = 0;
1826 gimple_stmt_iterator bsi;
1827 struct function *my_function = DECL_STRUCT_FUNCTION (node->decl);
1830 tree funtype = TREE_TYPE (node->decl);
1832 if (node->local.disregard_inline_limits)
1834 inline_summary (node)->self_time = 0;
1835 inline_summary (node)->self_size = 0;
1836 inline_summary (node)->time_inlining_benefit = 0;
1837 inline_summary (node)->size_inlining_benefit = 0;
1841 fprintf (dump_file, "Analyzing function body size: %s\n",
1842 cgraph_node_name (node));
1844 gcc_assert (my_function && my_function->cfg);
1845 FOR_EACH_BB_FN (bb, my_function)
1847 freq = compute_call_stmt_bb_frequency (node->decl, bb);
1848 for (bsi = gsi_start_bb (bb); !gsi_end_p (bsi); gsi_next (&bsi))
1850 gimple stmt = gsi_stmt (bsi);
1851 int this_size = estimate_num_insns (stmt, &eni_size_weights);
1852 int this_time = estimate_num_insns (stmt, &eni_time_weights);
1854 if (dump_file && (dump_flags & TDF_DETAILS))
1856 fprintf (dump_file, " freq:%6i size:%3i time:%3i ",
1857 freq, this_size, this_time);
1858 print_gimple_stmt (dump_file, stmt, 0, 0);
1863 if (likely_eliminated_by_inlining_p (stmt))
1865 size_inlining_benefit += this_size;
1866 time_inlining_benefit += this_time;
1867 if (dump_file && (dump_flags & TDF_DETAILS))
1868 fprintf (dump_file, " Likely eliminated\n");
1870 gcc_assert (time >= 0);
1871 gcc_assert (size >= 0);
1874 time = (time + CGRAPH_FREQ_BASE / 2) / CGRAPH_FREQ_BASE;
1875 time_inlining_benefit = ((time_inlining_benefit + CGRAPH_FREQ_BASE / 2)
1876 / CGRAPH_FREQ_BASE);
1878 fprintf (dump_file, "Overall function body time: %i-%i size: %i-%i\n",
1879 (int)time, (int)time_inlining_benefit,
1880 size, size_inlining_benefit);
1881 time_inlining_benefit += eni_time_weights.call_cost;
1882 size_inlining_benefit += eni_size_weights.call_cost;
1883 if (!VOID_TYPE_P (TREE_TYPE (funtype)))
1885 int cost = estimate_move_cost (TREE_TYPE (funtype));
1886 time_inlining_benefit += cost;
1887 size_inlining_benefit += cost;
1889 for (arg = DECL_ARGUMENTS (node->decl); arg; arg = TREE_CHAIN (arg))
1890 if (!VOID_TYPE_P (TREE_TYPE (arg)))
1892 int cost = estimate_move_cost (TREE_TYPE (arg));
1893 time_inlining_benefit += cost;
1894 size_inlining_benefit += cost;
1896 if (time_inlining_benefit > MAX_TIME)
1897 time_inlining_benefit = MAX_TIME;
1898 if (time > MAX_TIME)
1900 inline_summary (node)->self_time = time;
1901 inline_summary (node)->self_size = size;
1903 fprintf (dump_file, "With function call overhead time: %i-%i size: %i-%i\n",
1904 (int)time, (int)time_inlining_benefit,
1905 size, size_inlining_benefit);
1906 inline_summary (node)->time_inlining_benefit = time_inlining_benefit;
1907 inline_summary (node)->size_inlining_benefit = size_inlining_benefit;
1910 /* Compute parameters of functions used by inliner. */
1912 compute_inline_parameters (struct cgraph_node *node)
1914 HOST_WIDE_INT self_stack_size;
1916 gcc_assert (!node->global.inlined_to);
1918 /* Estimate the stack size for the function. But not at -O0
1919 because estimated_stack_frame_size is a quadratic problem. */
1920 self_stack_size = optimize ? estimated_stack_frame_size () : 0;
1921 inline_summary (node)->estimated_self_stack_size = self_stack_size;
1922 node->global.estimated_stack_size = self_stack_size;
1923 node->global.stack_frame_offset = 0;
1925 /* Can this function be inlined at all? */
1926 node->local.inlinable = tree_inlinable_function_p (node->decl);
1927 if (node->local.inlinable && !node->local.disregard_inline_limits)
1928 node->local.disregard_inline_limits
1929 = DECL_DISREGARD_INLINE_LIMITS (node->decl);
1930 estimate_function_body_sizes (node);
1931 /* Inlining characteristics are maintained by the cgraph_mark_inline. */
1932 node->global.time = inline_summary (node)->self_time;
1933 node->global.size = inline_summary (node)->self_size;
1938 /* Compute parameters of functions used by inliner using
1939 current_function_decl. */
1941 compute_inline_parameters_for_current (void)
1943 compute_inline_parameters (cgraph_node (current_function_decl));
1947 struct gimple_opt_pass pass_inline_parameters =
1951 "inline_param", /* name */
1953 compute_inline_parameters_for_current,/* execute */
1956 0, /* static_pass_number */
1957 TV_INLINE_HEURISTICS, /* tv_id */
1958 0, /* properties_required */
1959 0, /* properties_provided */
1960 0, /* properties_destroyed */
1961 0, /* todo_flags_start */
1962 0 /* todo_flags_finish */
1966 /* This function performs intraprocedural analyzis in NODE that is required to
1967 inline indirect calls. */
1969 inline_indirect_intraprocedural_analysis (struct cgraph_node *node)
1971 struct cgraph_edge *cs;
1975 ipa_initialize_node_params (node);
1976 ipa_detect_param_modifications (node);
1978 ipa_analyze_params_uses (node);
1981 for (cs = node->callees; cs; cs = cs->next_callee)
1983 ipa_count_arguments (cs);
1984 ipa_compute_jump_functions (cs);
1989 ipa_print_node_params (dump_file, node);
1990 ipa_print_node_jump_functions (dump_file, node);
1994 /* Note function body size. */
1996 analyze_function (struct cgraph_node *node)
1998 push_cfun (DECL_STRUCT_FUNCTION (node->decl));
1999 current_function_decl = node->decl;
2001 compute_inline_parameters (node);
2002 if (flag_indirect_inlining)
2003 inline_indirect_intraprocedural_analysis (node);
2005 current_function_decl = NULL;
2009 /* Called when new function is inserted to callgraph late. */
2011 add_new_function (struct cgraph_node *node, void *data ATTRIBUTE_UNUSED)
2013 analyze_function (node);
2016 /* Note function body size. */
2018 inline_generate_summary (void)
2020 struct cgraph_node *node;
2022 function_insertion_hook_holder =
2023 cgraph_add_function_insertion_hook (&add_new_function, NULL);
2025 if (flag_indirect_inlining)
2027 ipa_register_cgraph_hooks ();
2028 ipa_check_create_node_params ();
2029 ipa_check_create_edge_args ();
2032 for (node = cgraph_nodes; node; node = node->next)
2034 analyze_function (node);
2039 /* Apply inline plan to function. */
2041 inline_transform (struct cgraph_node *node)
2043 unsigned int todo = 0;
2044 struct cgraph_edge *e;
2046 /* FIXME: Currently the passmanager is adding inline transform more than once to some
2047 clones. This needs revisiting after WPA cleanups. */
2048 if (cfun->after_inlining)
2051 /* We might need the body of this function so that we can expand
2052 it inline somewhere else. */
2053 if (cgraph_preserve_function_body_p (node->decl))
2054 save_inline_function_body (node);
2056 for (e = node->callees; e; e = e->next_callee)
2057 if (!e->inline_failed || warn_inline)
2062 timevar_push (TV_INTEGRATION);
2063 todo = optimize_inline_calls (current_function_decl);
2064 timevar_pop (TV_INTEGRATION);
2066 cfun->always_inline_functions_inlined = true;
2067 cfun->after_inlining = true;
2068 return todo | execute_fixup_cfg ();
2071 /* Read inline summary. Jump functions are shared among ipa-cp
2072 and inliner, so when ipa-cp is active, we don't need to write them
2076 inline_read_summary (void)
2078 if (flag_indirect_inlining)
2080 ipa_register_cgraph_hooks ();
2082 ipa_prop_read_jump_functions ();
2084 function_insertion_hook_holder =
2085 cgraph_add_function_insertion_hook (&add_new_function, NULL);
2088 /* Write inline summary for node in SET.
2089 Jump functions are shared among ipa-cp and inliner, so when ipa-cp is
2090 active, we don't need to write them twice. */
2093 inline_write_summary (cgraph_node_set set,
2094 varpool_node_set vset ATTRIBUTE_UNUSED)
2096 if (flag_indirect_inlining && !flag_ipa_cp)
2097 ipa_prop_write_jump_functions (set);
2100 /* When to run IPA inlining. Inlining of always-inline functions
2101 happens during early inlining. */
2104 gate_cgraph_decide_inlining (void)
2106 /* ??? We'd like to skip this if not optimizing or not inlining as
2107 all always-inline functions have been processed by early
2108 inlining already. But this at least breaks EH with C++ as
2109 we need to unconditionally run fixup_cfg even at -O0.
2110 So leave it on unconditionally for now. */
2114 struct ipa_opt_pass_d pass_ipa_inline =
2118 "inline", /* name */
2119 gate_cgraph_decide_inlining, /* gate */
2120 cgraph_decide_inlining, /* execute */
2123 0, /* static_pass_number */
2124 TV_INLINE_HEURISTICS, /* tv_id */
2125 0, /* properties_required */
2126 0, /* properties_provided */
2127 0, /* properties_destroyed */
2128 TODO_remove_functions, /* todo_flags_finish */
2129 TODO_dump_cgraph | TODO_dump_func
2130 | TODO_remove_functions | TODO_ggc_collect /* todo_flags_finish */
2132 inline_generate_summary, /* generate_summary */
2133 inline_write_summary, /* write_summary */
2134 inline_read_summary, /* read_summary */
2135 NULL, /* write_optimization_summary */
2136 NULL, /* read_optimization_summary */
2137 NULL, /* stmt_fixup */
2139 inline_transform, /* function_transform */
2140 NULL, /* variable_transform */
2144 #include "gt-ipa-inline.h"