1 | //===-- CSPreInliner.h - Profile guided preinliner ---------------- C++ -*-===// |
2 | // |
3 | // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. |
4 | // See https://llvm.org/LICENSE.txt for license information. |
5 | // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception |
6 | // |
7 | //===----------------------------------------------------------------------===// |
8 | |
9 | #ifndef LLVM_TOOLS_LLVM_PROFGEN_PGOINLINEADVISOR_H |
10 | #define LLVM_TOOLS_LLVM_PROFGEN_PGOINLINEADVISOR_H |
11 | |
12 | #include "ProfiledBinary.h" |
13 | #include "llvm/ADT/PriorityQueue.h" |
14 | #include "llvm/ProfileData/ProfileCommon.h" |
15 | #include "llvm/ProfileData/SampleProf.h" |
16 | #include "llvm/Transforms/IPO/ProfiledCallGraph.h" |
17 | #include "llvm/Transforms/IPO/SampleContextTracker.h" |
18 | |
19 | using namespace llvm; |
20 | using namespace sampleprof; |
21 | |
22 | namespace llvm { |
23 | namespace sampleprof { |
24 | |
25 | // Inline candidate seen from profile |
26 | struct ProfiledInlineCandidate { |
27 | ProfiledInlineCandidate(const FunctionSamples *Samples, uint64_t Count, |
28 | uint32_t Size) |
29 | : CalleeSamples(Samples), CallsiteCount(Count), SizeCost(Size) {} |
30 | // Context-sensitive function profile for inline candidate |
31 | const FunctionSamples *CalleeSamples; |
32 | // Call site count for an inline candidate |
33 | // TODO: make sure entry count for context profile and call site |
34 | // target count for corresponding call are consistent. |
35 | uint64_t CallsiteCount; |
36 | // Size proxy for function under particular call context. |
37 | uint64_t SizeCost; |
38 | }; |
39 | |
40 | // Inline candidate comparer using call site weight |
41 | struct ProfiledCandidateComparer { |
42 | bool operator()(const ProfiledInlineCandidate &LHS, |
43 | const ProfiledInlineCandidate &RHS) { |
44 | // Always prioritize inlining zero-sized functions as they do not affect the |
45 | // size budget. This could happen when all of the callee's code is gone and |
46 | // only pseudo probes are left. |
47 | if ((LHS.SizeCost == 0 || RHS.SizeCost == 0) && |
48 | (LHS.SizeCost != RHS.SizeCost)) |
49 | return RHS.SizeCost == 0; |
50 | |
51 | if (LHS.CallsiteCount != RHS.CallsiteCount) |
52 | return LHS.CallsiteCount < RHS.CallsiteCount; |
53 | |
54 | if (LHS.SizeCost != RHS.SizeCost) |
55 | return LHS.SizeCost > RHS.SizeCost; |
56 | |
57 | // Tie breaker using GUID so we have stable/deterministic inlining order |
58 | assert(LHS.CalleeSamples && RHS.CalleeSamples && |
59 | "Expect non-null FunctionSamples" ); |
60 | return LHS.CalleeSamples->getGUID() < RHS.CalleeSamples->getGUID(); |
61 | } |
62 | }; |
63 | |
64 | using ProfiledCandidateQueue = |
65 | PriorityQueue<ProfiledInlineCandidate, std::vector<ProfiledInlineCandidate>, |
66 | ProfiledCandidateComparer>; |
67 | |
68 | // Pre-compilation inliner based on context-sensitive profile. |
69 | // The PreInliner estimates inline decision using hotness from profile |
70 | // and cost estimation from machine code size. It helps merges context |
71 | // profile globally and achieves better post-inine profile quality, which |
72 | // otherwise won't be possible for ThinLTO. It also reduce context profile |
73 | // size by only keep context that is estimated to be inlined. |
74 | class CSPreInliner { |
75 | public: |
76 | CSPreInliner(SampleContextTracker &Tracker, ProfiledBinary &Binary, |
77 | ProfileSummary *Summary); |
78 | void run(); |
79 | |
80 | private: |
81 | bool getInlineCandidates(ProfiledCandidateQueue &CQueue, |
82 | const FunctionSamples *FCallerContextSamples); |
83 | std::vector<FunctionId> buildTopDownOrder(); |
84 | void processFunction(FunctionId Name); |
85 | bool shouldInline(ProfiledInlineCandidate &Candidate); |
86 | uint32_t getFuncSize(const ContextTrieNode *ContextNode); |
87 | bool UseContextCost; |
88 | SampleContextTracker &ContextTracker; |
89 | ProfiledBinary &Binary; |
90 | ProfileSummary *Summary; |
91 | }; |
92 | |
93 | } // end namespace sampleprof |
94 | } // end namespace llvm |
95 | |
96 | #endif |
97 | |