1 //===- AMDGPUPerfHintAnalysis.cpp - analysis of functions memory traffic --===//
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 /// \file
10 /// \brief Analyzes if a function potentially memory bound and if a kernel
11 /// kernel may benefit from limiting number of waves to reduce cache thrashing.
12 ///
13 //===----------------------------------------------------------------------===//
14 
15 #include "AMDGPU.h"
16 #include "AMDGPUPerfHintAnalysis.h"
17 #include "Utils/AMDGPUBaseInfo.h"
18 #include "llvm/ADT/SmallSet.h"
19 #include "llvm/ADT/Statistic.h"
20 #include "llvm/Analysis/CallGraph.h"
21 #include "llvm/Analysis/ValueTracking.h"
22 #include "llvm/CodeGen/TargetLowering.h"
23 #include "llvm/CodeGen/TargetPassConfig.h"
24 #include "llvm/CodeGen/TargetSubtargetInfo.h"
25 #include "llvm/IR/Constants.h"
26 #include "llvm/IR/Instructions.h"
27 #include "llvm/IR/IntrinsicInst.h"
28 #include "llvm/IR/Module.h"
29 #include "llvm/IR/ValueMap.h"
30 #include "llvm/Support/CommandLine.h"
31 
32 using namespace llvm;
33 
34 #define DEBUG_TYPE "amdgpu-perf-hint"
35 
36 static cl::opt<unsigned>
37     MemBoundThresh("amdgpu-membound-threshold", cl::init(50), cl::Hidden,
38                    cl::desc("Function mem bound threshold in %"));
39 
40 static cl::opt<unsigned>
41     LimitWaveThresh("amdgpu-limit-wave-threshold", cl::init(50), cl::Hidden,
42                     cl::desc("Kernel limit wave threshold in %"));
43 
44 static cl::opt<unsigned>
45     IAWeight("amdgpu-indirect-access-weight", cl::init(1000), cl::Hidden,
46              cl::desc("Indirect access memory instruction weight"));
47 
48 static cl::opt<unsigned>
49     LSWeight("amdgpu-large-stride-weight", cl::init(1000), cl::Hidden,
50              cl::desc("Large stride memory access weight"));
51 
52 static cl::opt<unsigned>
53     LargeStrideThresh("amdgpu-large-stride-threshold", cl::init(64), cl::Hidden,
54                       cl::desc("Large stride memory access threshold"));
55 
56 STATISTIC(NumMemBound, "Number of functions marked as memory bound");
57 STATISTIC(NumLimitWave, "Number of functions marked as needing limit wave");
58 
59 char llvm::AMDGPUPerfHintAnalysis::ID = 0;
60 char &llvm::AMDGPUPerfHintAnalysisID = AMDGPUPerfHintAnalysis::ID;
61 
62 INITIALIZE_PASS(AMDGPUPerfHintAnalysis, DEBUG_TYPE,
63                 "Analysis if a function is memory bound", true, true)
64 
65 namespace {
66 
67 struct AMDGPUPerfHint {
68   friend AMDGPUPerfHintAnalysis;
69 
70 public:
AMDGPUPerfHint__anone97d0fd90111::AMDGPUPerfHint71   AMDGPUPerfHint(AMDGPUPerfHintAnalysis::FuncInfoMap &FIM_,
72                  const TargetLowering *TLI_)
73       : FIM(FIM_), DL(nullptr), TLI(TLI_) {}
74 
75   bool runOnFunction(Function &F);
76 
77 private:
78   struct MemAccessInfo {
79     const Value *V;
80     const Value *Base;
81     int64_t Offset;
MemAccessInfo__anone97d0fd90111::AMDGPUPerfHint::MemAccessInfo82     MemAccessInfo() : V(nullptr), Base(nullptr), Offset(0) {}
83     bool isLargeStride(MemAccessInfo &Reference) const;
84 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
print__anone97d0fd90111::AMDGPUPerfHint::MemAccessInfo85     Printable print() const {
86       return Printable([this](raw_ostream &OS) {
87         OS << "Value: " << *V << '\n'
88            << "Base: " << *Base << " Offset: " << Offset << '\n';
89       });
90     }
91 #endif
92   };
93 
94   MemAccessInfo makeMemAccessInfo(Instruction *) const;
95 
96   MemAccessInfo LastAccess; // Last memory access info
97 
98   AMDGPUPerfHintAnalysis::FuncInfoMap &FIM;
99 
100   const DataLayout *DL;
101 
102   const TargetLowering *TLI;
103 
104   AMDGPUPerfHintAnalysis::FuncInfo *visit(const Function &F);
105   static bool isMemBound(const AMDGPUPerfHintAnalysis::FuncInfo &F);
106   static bool needLimitWave(const AMDGPUPerfHintAnalysis::FuncInfo &F);
107 
108   bool isIndirectAccess(const Instruction *Inst) const;
109 
110   /// Check if the instruction is large stride.
111   /// The purpose is to identify memory access pattern like:
112   /// x = a[i];
113   /// y = a[i+1000];
114   /// z = a[i+2000];
115   /// In the above example, the second and third memory access will be marked
116   /// large stride memory access.
117   bool isLargeStride(const Instruction *Inst);
118 
119   bool isGlobalAddr(const Value *V) const;
120   bool isLocalAddr(const Value *V) const;
121   bool isConstantAddr(const Value *V) const;
122 };
123 
getMemoryInstrPtr(const Instruction * Inst)124 static const Value *getMemoryInstrPtr(const Instruction *Inst) {
125   if (auto LI = dyn_cast<LoadInst>(Inst)) {
126     return LI->getPointerOperand();
127   }
128   if (auto SI = dyn_cast<StoreInst>(Inst)) {
129     return SI->getPointerOperand();
130   }
131   if (auto AI = dyn_cast<AtomicCmpXchgInst>(Inst)) {
132     return AI->getPointerOperand();
133   }
134   if (auto AI = dyn_cast<AtomicRMWInst>(Inst)) {
135     return AI->getPointerOperand();
136   }
137   if (auto MI = dyn_cast<AnyMemIntrinsic>(Inst)) {
138     return MI->getRawDest();
139   }
140 
141   return nullptr;
142 }
143 
isIndirectAccess(const Instruction * Inst) const144 bool AMDGPUPerfHint::isIndirectAccess(const Instruction *Inst) const {
145   LLVM_DEBUG(dbgs() << "[isIndirectAccess] " << *Inst << '\n');
146   SmallSet<const Value *, 32> WorkSet;
147   SmallSet<const Value *, 32> Visited;
148   if (const Value *MO = getMemoryInstrPtr(Inst)) {
149     if (isGlobalAddr(MO))
150       WorkSet.insert(MO);
151   }
152 
153   while (!WorkSet.empty()) {
154     const Value *V = *WorkSet.begin();
155     WorkSet.erase(*WorkSet.begin());
156     if (!Visited.insert(V).second)
157       continue;
158     LLVM_DEBUG(dbgs() << "  check: " << *V << '\n');
159 
160     if (auto LD = dyn_cast<LoadInst>(V)) {
161       auto M = LD->getPointerOperand();
162       if (isGlobalAddr(M) || isLocalAddr(M) || isConstantAddr(M)) {
163         LLVM_DEBUG(dbgs() << "    is IA\n");
164         return true;
165       }
166       continue;
167     }
168 
169     if (auto GEP = dyn_cast<GetElementPtrInst>(V)) {
170       auto P = GEP->getPointerOperand();
171       WorkSet.insert(P);
172       for (unsigned I = 1, E = GEP->getNumIndices() + 1; I != E; ++I)
173         WorkSet.insert(GEP->getOperand(I));
174       continue;
175     }
176 
177     if (auto U = dyn_cast<UnaryInstruction>(V)) {
178       WorkSet.insert(U->getOperand(0));
179       continue;
180     }
181 
182     if (auto BO = dyn_cast<BinaryOperator>(V)) {
183       WorkSet.insert(BO->getOperand(0));
184       WorkSet.insert(BO->getOperand(1));
185       continue;
186     }
187 
188     if (auto S = dyn_cast<SelectInst>(V)) {
189       WorkSet.insert(S->getFalseValue());
190       WorkSet.insert(S->getTrueValue());
191       continue;
192     }
193 
194     if (auto E = dyn_cast<ExtractElementInst>(V)) {
195       WorkSet.insert(E->getVectorOperand());
196       continue;
197     }
198 
199     LLVM_DEBUG(dbgs() << "    dropped\n");
200   }
201 
202   LLVM_DEBUG(dbgs() << "  is not IA\n");
203   return false;
204 }
205 
visit(const Function & F)206 AMDGPUPerfHintAnalysis::FuncInfo *AMDGPUPerfHint::visit(const Function &F) {
207   AMDGPUPerfHintAnalysis::FuncInfo &FI = FIM[&F];
208 
209   LLVM_DEBUG(dbgs() << "[AMDGPUPerfHint] process " << F.getName() << '\n');
210 
211   for (auto &B : F) {
212     LastAccess = MemAccessInfo();
213     for (auto &I : B) {
214       if (getMemoryInstrPtr(&I)) {
215         if (isIndirectAccess(&I))
216           ++FI.IAMInstCount;
217         if (isLargeStride(&I))
218           ++FI.LSMInstCount;
219         ++FI.MemInstCount;
220         ++FI.InstCount;
221         continue;
222       }
223       CallSite CS(const_cast<Instruction *>(&I));
224       if (CS) {
225         Function *Callee = CS.getCalledFunction();
226         if (!Callee || Callee->isDeclaration()) {
227           ++FI.InstCount;
228           continue;
229         }
230         if (&F == Callee) // Handle immediate recursion
231           continue;
232 
233         auto Loc = FIM.find(Callee);
234         if (Loc == FIM.end())
235           continue;
236 
237         FI.MemInstCount += Loc->second.MemInstCount;
238         FI.InstCount += Loc->second.InstCount;
239         FI.IAMInstCount += Loc->second.IAMInstCount;
240         FI.LSMInstCount += Loc->second.LSMInstCount;
241       } else if (auto *GEP = dyn_cast<GetElementPtrInst>(&I)) {
242         TargetLoweringBase::AddrMode AM;
243         auto *Ptr = GetPointerBaseWithConstantOffset(GEP, AM.BaseOffs, *DL);
244         AM.BaseGV = dyn_cast_or_null<GlobalValue>(const_cast<Value *>(Ptr));
245         AM.HasBaseReg = !AM.BaseGV;
246         if (TLI->isLegalAddressingMode(*DL, AM, GEP->getResultElementType(),
247                                        GEP->getPointerAddressSpace()))
248           // Offset will likely be folded into load or store
249           continue;
250         ++FI.InstCount;
251       } else {
252         ++FI.InstCount;
253       }
254     }
255   }
256 
257   return &FI;
258 }
259 
runOnFunction(Function & F)260 bool AMDGPUPerfHint::runOnFunction(Function &F) {
261   const Module &M = *F.getParent();
262   DL = &M.getDataLayout();
263 
264   if (F.hasFnAttribute("amdgpu-wave-limiter") &&
265       F.hasFnAttribute("amdgpu-memory-bound"))
266     return false;
267 
268   const AMDGPUPerfHintAnalysis::FuncInfo *Info = visit(F);
269 
270   LLVM_DEBUG(dbgs() << F.getName() << " MemInst: " << Info->MemInstCount
271                     << '\n'
272                     << " IAMInst: " << Info->IAMInstCount << '\n'
273                     << " LSMInst: " << Info->LSMInstCount << '\n'
274                     << " TotalInst: " << Info->InstCount << '\n');
275 
276   if (isMemBound(*Info)) {
277     LLVM_DEBUG(dbgs() << F.getName() << " is memory bound\n");
278     NumMemBound++;
279     F.addFnAttr("amdgpu-memory-bound", "true");
280   }
281 
282   if (AMDGPU::isEntryFunctionCC(F.getCallingConv()) && needLimitWave(*Info)) {
283     LLVM_DEBUG(dbgs() << F.getName() << " needs limit wave\n");
284     NumLimitWave++;
285     F.addFnAttr("amdgpu-wave-limiter", "true");
286   }
287 
288   return true;
289 }
290 
isMemBound(const AMDGPUPerfHintAnalysis::FuncInfo & FI)291 bool AMDGPUPerfHint::isMemBound(const AMDGPUPerfHintAnalysis::FuncInfo &FI) {
292   return FI.MemInstCount * 100 / FI.InstCount > MemBoundThresh;
293 }
294 
needLimitWave(const AMDGPUPerfHintAnalysis::FuncInfo & FI)295 bool AMDGPUPerfHint::needLimitWave(const AMDGPUPerfHintAnalysis::FuncInfo &FI) {
296   return ((FI.MemInstCount + FI.IAMInstCount * IAWeight +
297            FI.LSMInstCount * LSWeight) *
298           100 / FI.InstCount) > LimitWaveThresh;
299 }
300 
isGlobalAddr(const Value * V) const301 bool AMDGPUPerfHint::isGlobalAddr(const Value *V) const {
302   if (auto PT = dyn_cast<PointerType>(V->getType())) {
303     unsigned As = PT->getAddressSpace();
304     // Flat likely points to global too.
305     return As == AMDGPUAS::GLOBAL_ADDRESS || As == AMDGPUAS::FLAT_ADDRESS;
306   }
307   return false;
308 }
309 
isLocalAddr(const Value * V) const310 bool AMDGPUPerfHint::isLocalAddr(const Value *V) const {
311   if (auto PT = dyn_cast<PointerType>(V->getType()))
312     return PT->getAddressSpace() == AMDGPUAS::LOCAL_ADDRESS;
313   return false;
314 }
315 
isLargeStride(const Instruction * Inst)316 bool AMDGPUPerfHint::isLargeStride(const Instruction *Inst) {
317   LLVM_DEBUG(dbgs() << "[isLargeStride] " << *Inst << '\n');
318 
319   MemAccessInfo MAI = makeMemAccessInfo(const_cast<Instruction *>(Inst));
320   bool IsLargeStride = MAI.isLargeStride(LastAccess);
321   if (MAI.Base)
322     LastAccess = std::move(MAI);
323 
324   return IsLargeStride;
325 }
326 
327 AMDGPUPerfHint::MemAccessInfo
makeMemAccessInfo(Instruction * Inst) const328 AMDGPUPerfHint::makeMemAccessInfo(Instruction *Inst) const {
329   MemAccessInfo MAI;
330   const Value *MO = getMemoryInstrPtr(Inst);
331 
332   LLVM_DEBUG(dbgs() << "[isLargeStride] MO: " << *MO << '\n');
333   // Do not treat local-addr memory access as large stride.
334   if (isLocalAddr(MO))
335     return MAI;
336 
337   MAI.V = MO;
338   MAI.Base = GetPointerBaseWithConstantOffset(MO, MAI.Offset, *DL);
339   return MAI;
340 }
341 
isConstantAddr(const Value * V) const342 bool AMDGPUPerfHint::isConstantAddr(const Value *V) const {
343   if (auto PT = dyn_cast<PointerType>(V->getType())) {
344     unsigned As = PT->getAddressSpace();
345     return As == AMDGPUAS::CONSTANT_ADDRESS ||
346            As == AMDGPUAS::CONSTANT_ADDRESS_32BIT;
347   }
348   return false;
349 }
350 
isLargeStride(MemAccessInfo & Reference) const351 bool AMDGPUPerfHint::MemAccessInfo::isLargeStride(
352     MemAccessInfo &Reference) const {
353 
354   if (!Base || !Reference.Base || Base != Reference.Base)
355     return false;
356 
357   uint64_t Diff = Offset > Reference.Offset ? Offset - Reference.Offset
358                                             : Reference.Offset - Offset;
359   bool Result = Diff > LargeStrideThresh;
360   LLVM_DEBUG(dbgs() << "[isLargeStride compare]\n"
361                << print() << "<=>\n"
362                << Reference.print() << "Result:" << Result << '\n');
363   return Result;
364 }
365 } // namespace
366 
runOnSCC(CallGraphSCC & SCC)367 bool AMDGPUPerfHintAnalysis::runOnSCC(CallGraphSCC &SCC) {
368   auto *TPC = getAnalysisIfAvailable<TargetPassConfig>();
369   if (!TPC)
370     return false;
371 
372   const TargetMachine &TM = TPC->getTM<TargetMachine>();
373 
374   bool Changed = false;
375   for (CallGraphNode *I : SCC) {
376     Function *F = I->getFunction();
377     if (!F || F->isDeclaration())
378       continue;
379 
380     const TargetSubtargetInfo *ST = TM.getSubtargetImpl(*F);
381     AMDGPUPerfHint Analyzer(FIM, ST->getTargetLowering());
382 
383     if (Analyzer.runOnFunction(*F))
384       Changed = true;
385   }
386 
387   return Changed;
388 }
389 
isMemoryBound(const Function * F) const390 bool AMDGPUPerfHintAnalysis::isMemoryBound(const Function *F) const {
391   auto FI = FIM.find(F);
392   if (FI == FIM.end())
393     return false;
394 
395   return AMDGPUPerfHint::isMemBound(FI->second);
396 }
397 
needsWaveLimiter(const Function * F) const398 bool AMDGPUPerfHintAnalysis::needsWaveLimiter(const Function *F) const {
399   auto FI = FIM.find(F);
400   if (FI == FIM.end())
401     return false;
402 
403   return AMDGPUPerfHint::needLimitWave(FI->second);
404 }
405