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 #include "llvm/Target/TargetMachine.h"
32
33 using namespace llvm;
34
35 #define DEBUG_TYPE "amdgpu-perf-hint"
36
37 static cl::opt<unsigned>
38 MemBoundThresh("amdgpu-membound-threshold", cl::init(50), cl::Hidden,
39 cl::desc("Function mem bound threshold in %"));
40
41 static cl::opt<unsigned>
42 LimitWaveThresh("amdgpu-limit-wave-threshold", cl::init(50), cl::Hidden,
43 cl::desc("Kernel limit wave threshold in %"));
44
45 static cl::opt<unsigned>
46 IAWeight("amdgpu-indirect-access-weight", cl::init(1000), cl::Hidden,
47 cl::desc("Indirect access memory instruction weight"));
48
49 static cl::opt<unsigned>
50 LSWeight("amdgpu-large-stride-weight", cl::init(1000), cl::Hidden,
51 cl::desc("Large stride memory access weight"));
52
53 static cl::opt<unsigned>
54 LargeStrideThresh("amdgpu-large-stride-threshold", cl::init(64), cl::Hidden,
55 cl::desc("Large stride memory access threshold"));
56
57 STATISTIC(NumMemBound, "Number of functions marked as memory bound");
58 STATISTIC(NumLimitWave, "Number of functions marked as needing limit wave");
59
60 char llvm::AMDGPUPerfHintAnalysis::ID = 0;
61 char &llvm::AMDGPUPerfHintAnalysisID = AMDGPUPerfHintAnalysis::ID;
62
63 INITIALIZE_PASS(AMDGPUPerfHintAnalysis, DEBUG_TYPE,
64 "Analysis if a function is memory bound", true, true)
65
66 namespace {
67
68 struct AMDGPUPerfHint {
69 friend AMDGPUPerfHintAnalysis;
70
71 public:
AMDGPUPerfHint__anonf21017e30111::AMDGPUPerfHint72 AMDGPUPerfHint(AMDGPUPerfHintAnalysis::FuncInfoMap &FIM_,
73 const TargetLowering *TLI_)
74 : FIM(FIM_), DL(nullptr), TLI(TLI_) {}
75
76 bool runOnFunction(Function &F);
77
78 private:
79 struct MemAccessInfo {
80 const Value *V;
81 const Value *Base;
82 int64_t Offset;
MemAccessInfo__anonf21017e30111::AMDGPUPerfHint::MemAccessInfo83 MemAccessInfo() : V(nullptr), Base(nullptr), Offset(0) {}
84 bool isLargeStride(MemAccessInfo &Reference) const;
85 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
print__anonf21017e30111::AMDGPUPerfHint::MemAccessInfo86 Printable print() const {
87 return Printable([this](raw_ostream &OS) {
88 OS << "Value: " << *V << '\n'
89 << "Base: " << *Base << " Offset: " << Offset << '\n';
90 });
91 }
92 #endif
93 };
94
95 MemAccessInfo makeMemAccessInfo(Instruction *) const;
96
97 MemAccessInfo LastAccess; // Last memory access info
98
99 AMDGPUPerfHintAnalysis::FuncInfoMap &FIM;
100
101 const DataLayout *DL;
102
103 const TargetLowering *TLI;
104
105 AMDGPUPerfHintAnalysis::FuncInfo *visit(const Function &F);
106 static bool isMemBound(const AMDGPUPerfHintAnalysis::FuncInfo &F);
107 static bool needLimitWave(const AMDGPUPerfHintAnalysis::FuncInfo &F);
108
109 bool isIndirectAccess(const Instruction *Inst) const;
110
111 /// Check if the instruction is large stride.
112 /// The purpose is to identify memory access pattern like:
113 /// x = a[i];
114 /// y = a[i+1000];
115 /// z = a[i+2000];
116 /// In the above example, the second and third memory access will be marked
117 /// large stride memory access.
118 bool isLargeStride(const Instruction *Inst);
119
120 bool isGlobalAddr(const Value *V) const;
121 bool isLocalAddr(const Value *V) const;
122 bool isConstantAddr(const Value *V) const;
123 };
124
getMemoryInstrPtr(const Instruction * Inst)125 static const Value *getMemoryInstrPtr(const Instruction *Inst) {
126 if (auto LI = dyn_cast<LoadInst>(Inst)) {
127 return LI->getPointerOperand();
128 }
129 if (auto SI = dyn_cast<StoreInst>(Inst)) {
130 return SI->getPointerOperand();
131 }
132 if (auto AI = dyn_cast<AtomicCmpXchgInst>(Inst)) {
133 return AI->getPointerOperand();
134 }
135 if (auto AI = dyn_cast<AtomicRMWInst>(Inst)) {
136 return AI->getPointerOperand();
137 }
138 if (auto MI = dyn_cast<AnyMemIntrinsic>(Inst)) {
139 return MI->getRawDest();
140 }
141
142 return nullptr;
143 }
144
isIndirectAccess(const Instruction * Inst) const145 bool AMDGPUPerfHint::isIndirectAccess(const Instruction *Inst) const {
146 LLVM_DEBUG(dbgs() << "[isIndirectAccess] " << *Inst << '\n');
147 SmallSet<const Value *, 32> WorkSet;
148 SmallSet<const Value *, 32> Visited;
149 if (const Value *MO = getMemoryInstrPtr(Inst)) {
150 if (isGlobalAddr(MO))
151 WorkSet.insert(MO);
152 }
153
154 while (!WorkSet.empty()) {
155 const Value *V = *WorkSet.begin();
156 WorkSet.erase(*WorkSet.begin());
157 if (!Visited.insert(V).second)
158 continue;
159 LLVM_DEBUG(dbgs() << " check: " << *V << '\n');
160
161 if (auto LD = dyn_cast<LoadInst>(V)) {
162 auto M = LD->getPointerOperand();
163 if (isGlobalAddr(M) || isLocalAddr(M) || isConstantAddr(M)) {
164 LLVM_DEBUG(dbgs() << " is IA\n");
165 return true;
166 }
167 continue;
168 }
169
170 if (auto GEP = dyn_cast<GetElementPtrInst>(V)) {
171 auto P = GEP->getPointerOperand();
172 WorkSet.insert(P);
173 for (unsigned I = 1, E = GEP->getNumIndices() + 1; I != E; ++I)
174 WorkSet.insert(GEP->getOperand(I));
175 continue;
176 }
177
178 if (auto U = dyn_cast<UnaryInstruction>(V)) {
179 WorkSet.insert(U->getOperand(0));
180 continue;
181 }
182
183 if (auto BO = dyn_cast<BinaryOperator>(V)) {
184 WorkSet.insert(BO->getOperand(0));
185 WorkSet.insert(BO->getOperand(1));
186 continue;
187 }
188
189 if (auto S = dyn_cast<SelectInst>(V)) {
190 WorkSet.insert(S->getFalseValue());
191 WorkSet.insert(S->getTrueValue());
192 continue;
193 }
194
195 if (auto E = dyn_cast<ExtractElementInst>(V)) {
196 WorkSet.insert(E->getVectorOperand());
197 continue;
198 }
199
200 LLVM_DEBUG(dbgs() << " dropped\n");
201 }
202
203 LLVM_DEBUG(dbgs() << " is not IA\n");
204 return false;
205 }
206
visit(const Function & F)207 AMDGPUPerfHintAnalysis::FuncInfo *AMDGPUPerfHint::visit(const Function &F) {
208 AMDGPUPerfHintAnalysis::FuncInfo &FI = FIM[&F];
209
210 LLVM_DEBUG(dbgs() << "[AMDGPUPerfHint] process " << F.getName() << '\n');
211
212 for (auto &B : F) {
213 LastAccess = MemAccessInfo();
214 for (auto &I : B) {
215 if (getMemoryInstrPtr(&I)) {
216 if (isIndirectAccess(&I))
217 ++FI.IAMInstCount;
218 if (isLargeStride(&I))
219 ++FI.LSMInstCount;
220 ++FI.MemInstCount;
221 ++FI.InstCount;
222 continue;
223 }
224 if (auto *CB = dyn_cast<CallBase>(&I)) {
225 Function *Callee = CB->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