1 //===------- X86InsertPrefetch.cpp - Insert cache prefetch hints ----------===//
2 //
3 //                     The LLVM Compiler Infrastructure
4 //
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
7 //
8 //===----------------------------------------------------------------------===//
9 //
10 // This pass applies cache prefetch instructions based on a profile. The pass
11 // assumes DiscriminateMemOps ran immediately before, to ensure debug info
12 // matches the one used at profile generation time. The profile is encoded in
13 // afdo format (text or binary). It contains prefetch hints recommendations.
14 // Each recommendation is made in terms of debug info locations, a type (i.e.
15 // nta, t{0|1|2}) and a delta. The debug info identifies an instruction with a
16 // memory operand (see X86DiscriminateMemOps). The prefetch will be made for
17 // a location at that memory operand + the delta specified in the
18 // recommendation.
19 //
20 //===----------------------------------------------------------------------===//
21 
22 #include "X86.h"
23 #include "X86InstrBuilder.h"
24 #include "X86InstrInfo.h"
25 #include "X86MachineFunctionInfo.h"
26 #include "X86Subtarget.h"
27 #include "llvm/CodeGen/MachineModuleInfo.h"
28 #include "llvm/IR/DebugInfoMetadata.h"
29 #include "llvm/ProfileData/SampleProf.h"
30 #include "llvm/ProfileData/SampleProfReader.h"
31 #include "llvm/Transforms/IPO/SampleProfile.h"
32 using namespace llvm;
33 using namespace sampleprof;
34 
35 static cl::opt<std::string>
36     PrefetchHintsFile("prefetch-hints-file",
37                       cl::desc("Path to the prefetch hints profile."),
38                       cl::Hidden);
39 namespace {
40 
41 class X86InsertPrefetch : public MachineFunctionPass {
42   void getAnalysisUsage(AnalysisUsage &AU) const override;
43   bool doInitialization(Module &) override;
44 
45   bool runOnMachineFunction(MachineFunction &MF) override;
46   struct PrefetchInfo {
47     unsigned InstructionID;
48     int64_t Delta;
49   };
50   typedef SmallVectorImpl<PrefetchInfo> Prefetches;
51   bool findPrefetchInfo(const FunctionSamples *Samples, const MachineInstr &MI,
52                         Prefetches &prefetches) const;
53 
54 public:
55   static char ID;
56   X86InsertPrefetch(const std::string &PrefetchHintsFilename);
57   StringRef getPassName() const override {
58     return "X86 Insert Cache Prefetches";
59   }
60 
61 private:
62   std::string Filename;
63   std::unique_ptr<SampleProfileReader> Reader;
64 };
65 
66 using PrefetchHints = SampleRecord::CallTargetMap;
67 
68 // Return any prefetching hints for the specified MachineInstruction. The hints
69 // are returned as pairs (name, delta).
70 ErrorOr<PrefetchHints> getPrefetchHints(const FunctionSamples *TopSamples,
71                                         const MachineInstr &MI) {
72   if (const auto &Loc = MI.getDebugLoc())
73     if (const auto *Samples = TopSamples->findFunctionSamples(Loc))
74       return Samples->findCallTargetMapAt(FunctionSamples::getOffset(Loc),
75                                           Loc->getBaseDiscriminator());
76   return std::error_code();
77 }
78 
79 } // end anonymous namespace
80 
81 //===----------------------------------------------------------------------===//
82 //            Implementation
83 //===----------------------------------------------------------------------===//
84 
85 char X86InsertPrefetch::ID = 0;
86 
87 X86InsertPrefetch::X86InsertPrefetch(const std::string &PrefetchHintsFilename)
88     : MachineFunctionPass(ID), Filename(PrefetchHintsFilename) {}
89 
90 /// Return true if the provided MachineInstruction has cache prefetch hints. In
91 /// that case, the prefetch hints are stored, in order, in the Prefetches
92 /// vector.
93 bool X86InsertPrefetch::findPrefetchInfo(const FunctionSamples *TopSamples,
94                                          const MachineInstr &MI,
95                                          Prefetches &Prefetches) const {
96   assert(Prefetches.empty() &&
97          "Expected caller passed empty PrefetchInfo vector.");
98   static const std::pair<const StringRef, unsigned> HintTypes[] = {
99       {"_nta_", X86::PREFETCHNTA},
100       {"_t0_", X86::PREFETCHT0},
101       {"_t1_", X86::PREFETCHT1},
102       {"_t2_", X86::PREFETCHT2},
103   };
104   static const char *SerializedPrefetchPrefix = "__prefetch";
105 
106   const ErrorOr<PrefetchHints> T = getPrefetchHints(TopSamples, MI);
107   if (!T)
108     return false;
109   int16_t max_index = -1;
110   // Convert serialized prefetch hints into PrefetchInfo objects, and populate
111   // the Prefetches vector.
112   for (const auto &S_V : *T) {
113     StringRef Name = S_V.getKey();
114     if (Name.consume_front(SerializedPrefetchPrefix)) {
115       int64_t D = static_cast<int64_t>(S_V.second);
116       unsigned IID = 0;
117       for (const auto &HintType : HintTypes) {
118         if (Name.startswith(HintType.first)) {
119           Name = Name.drop_front(HintType.first.size());
120           IID = HintType.second;
121           break;
122         }
123       }
124       if (IID == 0)
125         return false;
126       uint8_t index = 0;
127       Name.consumeInteger(10, index);
128 
129       if (index >= Prefetches.size())
130         Prefetches.resize(index + 1);
131       Prefetches[index] = {IID, D};
132       max_index = std::max(max_index, static_cast<int16_t>(index));
133     }
134   }
135   assert(max_index + 1 >= 0 &&
136          "Possible overflow: max_index + 1 should be positive.");
137   assert(static_cast<size_t>(max_index + 1) == Prefetches.size() &&
138          "The number of prefetch hints received should match the number of "
139          "PrefetchInfo objects returned");
140   return !Prefetches.empty();
141 }
142 
143 bool X86InsertPrefetch::doInitialization(Module &M) {
144   if (Filename.empty())
145     return false;
146 
147   LLVMContext &Ctx = M.getContext();
148   ErrorOr<std::unique_ptr<SampleProfileReader>> ReaderOrErr =
149       SampleProfileReader::create(Filename, Ctx);
150   if (std::error_code EC = ReaderOrErr.getError()) {
151     std::string Msg = "Could not open profile: " + EC.message();
152     Ctx.diagnose(DiagnosticInfoSampleProfile(Filename, Msg,
153                                              DiagnosticSeverity::DS_Warning));
154     return false;
155   }
156   Reader = std::move(ReaderOrErr.get());
157   Reader->read();
158   return true;
159 }
160 
161 void X86InsertPrefetch::getAnalysisUsage(AnalysisUsage &AU) const {
162   AU.setPreservesAll();
163   AU.addRequired<MachineModuleInfo>();
164 }
165 
166 bool X86InsertPrefetch::runOnMachineFunction(MachineFunction &MF) {
167   if (!Reader)
168     return false;
169   const FunctionSamples *Samples = Reader->getSamplesFor(MF.getFunction());
170   if (!Samples)
171     return false;
172 
173   bool Changed = false;
174 
175   const TargetInstrInfo *TII = MF.getSubtarget().getInstrInfo();
176   SmallVector<PrefetchInfo, 4> Prefetches;
177   for (auto &MBB : MF) {
178     for (auto MI = MBB.instr_begin(); MI != MBB.instr_end();) {
179       auto Current = MI;
180       ++MI;
181 
182       int Offset = X86II::getMemoryOperandNo(Current->getDesc().TSFlags);
183       if (Offset < 0)
184         continue;
185       Prefetches.clear();
186       if (!findPrefetchInfo(Samples, *Current, Prefetches))
187         continue;
188       assert(!Prefetches.empty() &&
189              "The Prefetches vector should contain at least a value if "
190              "findPrefetchInfo returned true.");
191       for (auto &PrefInfo : Prefetches) {
192         unsigned PFetchInstrID = PrefInfo.InstructionID;
193         int64_t Delta = PrefInfo.Delta;
194         const MCInstrDesc &Desc = TII->get(PFetchInstrID);
195         MachineInstr *PFetch =
196             MF.CreateMachineInstr(Desc, Current->getDebugLoc(), true);
197         MachineInstrBuilder MIB(MF, PFetch);
198         unsigned Bias = X86II::getOperandBias(Current->getDesc());
199         int MemOpOffset = Offset + Bias;
200 
201         assert(X86::AddrBaseReg == 0 && X86::AddrScaleAmt == 1 &&
202                X86::AddrIndexReg == 2 && X86::AddrDisp == 3 &&
203                X86::AddrSegmentReg == 4 &&
204                "Unexpected change in X86 operand offset order.");
205 
206         // This assumes X86::AddBaseReg = 0, {...}ScaleAmt = 1, etc.
207         // FIXME(mtrofin): consider adding a:
208         //     MachineInstrBuilder::set(unsigned offset, op).
209         MIB.addReg(Current->getOperand(MemOpOffset + X86::AddrBaseReg).getReg())
210             .addImm(
211                 Current->getOperand(MemOpOffset + X86::AddrScaleAmt).getImm())
212             .addReg(
213                 Current->getOperand(MemOpOffset + X86::AddrIndexReg).getReg())
214             .addImm(Current->getOperand(MemOpOffset + X86::AddrDisp).getImm() +
215                     Delta)
216             .addReg(Current->getOperand(MemOpOffset + X86::AddrSegmentReg)
217                         .getReg());
218 
219         if (!Current->memoperands_empty()) {
220           MachineMemOperand *CurrentOp = *(Current->memoperands_begin());
221           MIB.addMemOperand(MF.getMachineMemOperand(
222               CurrentOp, CurrentOp->getOffset() + Delta, CurrentOp->getSize()));
223         }
224 
225         // Insert before Current. This is because Current may clobber some of
226         // the registers used to describe the input memory operand.
227         MBB.insert(Current, PFetch);
228         Changed = true;
229       }
230     }
231   }
232   return Changed;
233 }
234 
235 FunctionPass *llvm::createX86InsertPrefetchPass() {
236   return new X86InsertPrefetch(PrefetchHintsFile);
237 }
238