1 | //===- X86DiscriminateMemOps.cpp - Unique IDs for Mem Ops -----------------===// |
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 | /// This pass aids profile-driven cache prefetch insertion by ensuring all |
10 | /// instructions that have a memory operand are distinguishible from each other. |
11 | /// |
12 | //===----------------------------------------------------------------------===// |
13 | |
14 | #include "X86.h" |
15 | #include "X86Subtarget.h" |
16 | #include "llvm/CodeGen/MachineFunctionPass.h" |
17 | #include "llvm/CodeGen/MachineModuleInfo.h" |
18 | #include "llvm/IR/DebugInfoMetadata.h" |
19 | #include "llvm/ProfileData/SampleProf.h" |
20 | #include "llvm/ProfileData/SampleProfReader.h" |
21 | #include "llvm/Support/Debug.h" |
22 | #include <optional> |
23 | using namespace llvm; |
24 | |
25 | #define DEBUG_TYPE "x86-discriminate-memops" |
26 | |
27 | static cl::opt<bool> EnableDiscriminateMemops( |
28 | DEBUG_TYPE, cl::init(Val: false), |
29 | cl::desc("Generate unique debug info for each instruction with a memory " |
30 | "operand. Should be enabled for profile-driven cache prefetching, " |
31 | "both in the build of the binary being profiled, as well as in " |
32 | "the build of the binary consuming the profile." ), |
33 | cl::Hidden); |
34 | |
35 | static cl::opt<bool> BypassPrefetchInstructions( |
36 | "x86-bypass-prefetch-instructions" , cl::init(Val: true), |
37 | cl::desc("When discriminating instructions with memory operands, ignore " |
38 | "prefetch instructions. This ensures the other memory operand " |
39 | "instructions have the same identifiers after inserting " |
40 | "prefetches, allowing for successive insertions." ), |
41 | cl::Hidden); |
42 | |
43 | namespace { |
44 | |
45 | using Location = std::pair<StringRef, unsigned>; |
46 | |
47 | Location diToLocation(const DILocation *Loc) { |
48 | return std::make_pair(x: Loc->getFilename(), y: Loc->getLine()); |
49 | } |
50 | |
51 | /// Ensure each instruction having a memory operand has a distinct <LineNumber, |
52 | /// Discriminator> pair. |
53 | void updateDebugInfo(MachineInstr *MI, const DILocation *Loc) { |
54 | DebugLoc DL(Loc); |
55 | MI->setDebugLoc(DL); |
56 | } |
57 | |
58 | class X86DiscriminateMemOps : public MachineFunctionPass { |
59 | bool runOnMachineFunction(MachineFunction &MF) override; |
60 | StringRef getPassName() const override { |
61 | return "X86 Discriminate Memory Operands" ; |
62 | } |
63 | |
64 | public: |
65 | static char ID; |
66 | |
67 | /// Default construct and initialize the pass. |
68 | X86DiscriminateMemOps(); |
69 | }; |
70 | |
71 | bool IsPrefetchOpcode(unsigned Opcode) { |
72 | return Opcode == X86::PREFETCHNTA || Opcode == X86::PREFETCHT0 || |
73 | Opcode == X86::PREFETCHT1 || Opcode == X86::PREFETCHT2 || |
74 | Opcode == X86::PREFETCHIT0 || Opcode == X86::PREFETCHIT1 || |
75 | Opcode == X86::PREFETCHRST2; |
76 | } |
77 | } // end anonymous namespace |
78 | |
79 | //===----------------------------------------------------------------------===// |
80 | // Implementation |
81 | //===----------------------------------------------------------------------===// |
82 | |
83 | char X86DiscriminateMemOps::ID = 0; |
84 | |
85 | /// Default construct and initialize the pass. |
86 | X86DiscriminateMemOps::X86DiscriminateMemOps() : MachineFunctionPass(ID) {} |
87 | |
88 | bool X86DiscriminateMemOps::runOnMachineFunction(MachineFunction &MF) { |
89 | if (!EnableDiscriminateMemops) |
90 | return false; |
91 | |
92 | DISubprogram *FDI = MF.getFunction().getSubprogram(); |
93 | if (!FDI || !FDI->getUnit()->getDebugInfoForProfiling()) |
94 | return false; |
95 | |
96 | // Have a default DILocation, if we find instructions with memops that don't |
97 | // have any debug info. |
98 | const DILocation *ReferenceDI = |
99 | DILocation::get(Context&: FDI->getContext(), Line: FDI->getLine(), Column: 0, Scope: FDI); |
100 | assert(ReferenceDI && "ReferenceDI should not be nullptr" ); |
101 | DenseMap<Location, unsigned> MemOpDiscriminators; |
102 | MemOpDiscriminators[diToLocation(Loc: ReferenceDI)] = 0; |
103 | |
104 | // Figure out the largest discriminator issued for each Location. When we |
105 | // issue new discriminators, we can thus avoid issuing discriminators |
106 | // belonging to instructions that don't have memops. This isn't a requirement |
107 | // for the goals of this pass, however, it avoids unnecessary ambiguity. |
108 | for (auto &MBB : MF) { |
109 | for (auto &MI : MBB) { |
110 | const auto &DI = MI.getDebugLoc(); |
111 | if (!DI) |
112 | continue; |
113 | if (BypassPrefetchInstructions && IsPrefetchOpcode(Opcode: MI.getDesc().Opcode)) |
114 | continue; |
115 | Location Loc = diToLocation(Loc: DI); |
116 | unsigned &Disc = MemOpDiscriminators[Loc]; |
117 | Disc = std::max(a: Disc, b: DI->getBaseDiscriminator()); |
118 | } |
119 | } |
120 | |
121 | // Keep track of the discriminators seen at each Location. If an instruction's |
122 | // DebugInfo has a Location and discriminator we've already seen, replace its |
123 | // discriminator with a new one, to guarantee uniqueness. |
124 | DenseMap<Location, DenseSet<unsigned>> Seen; |
125 | |
126 | bool Changed = false; |
127 | for (auto &MBB : MF) { |
128 | for (auto &MI : MBB) { |
129 | if (X86II::getMemoryOperandNo(TSFlags: MI.getDesc().TSFlags) < 0) |
130 | continue; |
131 | if (BypassPrefetchInstructions && IsPrefetchOpcode(Opcode: MI.getDesc().Opcode)) |
132 | continue; |
133 | const DILocation *DI = MI.getDebugLoc(); |
134 | bool HasDebug = DI; |
135 | if (!HasDebug) { |
136 | DI = ReferenceDI; |
137 | } |
138 | Location L = diToLocation(Loc: DI); |
139 | DenseSet<unsigned> &Set = Seen[L]; |
140 | const std::pair<DenseSet<unsigned>::iterator, bool> TryInsert = |
141 | Set.insert(V: DI->getBaseDiscriminator()); |
142 | if (!TryInsert.second || !HasDebug) { |
143 | unsigned BF, DF, CI = 0; |
144 | DILocation::decodeDiscriminator(D: DI->getDiscriminator(), BD&: BF, DF, CI); |
145 | std::optional<unsigned> EncodedDiscriminator = |
146 | DILocation::encodeDiscriminator(BD: MemOpDiscriminators[L] + 1, DF, CI); |
147 | |
148 | if (!EncodedDiscriminator) { |
149 | // FIXME(mtrofin): The assumption is that this scenario is infrequent/OK |
150 | // not to support. If evidence points otherwise, we can explore synthesizeing |
151 | // unique DIs by adding fake line numbers, or by constructing 64 bit |
152 | // discriminators. |
153 | LLVM_DEBUG(dbgs() << "Unable to create a unique discriminator " |
154 | "for instruction with memory operand in: " |
155 | << DI->getFilename() << " Line: " << DI->getLine() |
156 | << " Column: " << DI->getColumn() |
157 | << ". This is likely due to a large macro expansion. \n" ); |
158 | continue; |
159 | } |
160 | // Since we were able to encode, bump the MemOpDiscriminators. |
161 | ++MemOpDiscriminators[L]; |
162 | DI = DI->cloneWithDiscriminator(Discriminator: *EncodedDiscriminator); |
163 | assert(DI && "DI should not be nullptr" ); |
164 | updateDebugInfo(MI: &MI, Loc: DI); |
165 | Changed = true; |
166 | std::pair<DenseSet<unsigned>::iterator, bool> MustInsert = |
167 | Set.insert(V: DI->getBaseDiscriminator()); |
168 | (void)MustInsert; // Silence warning in release build. |
169 | assert(MustInsert.second && "New discriminator shouldn't be present in set" ); |
170 | } |
171 | |
172 | // Bump the reference DI to avoid cramming discriminators on line 0. |
173 | // FIXME(mtrofin): pin ReferenceDI on blocks or first instruction with DI |
174 | // in a block. It's more consistent than just relying on the last memop |
175 | // instruction we happened to see. |
176 | ReferenceDI = DI; |
177 | } |
178 | } |
179 | return Changed; |
180 | } |
181 | |
182 | FunctionPass *llvm::createX86DiscriminateMemOpsPass() { |
183 | return new X86DiscriminateMemOps(); |
184 | } |
185 | |