1 //===- AMDGPURewriteOutArgumentsPass.cpp - Create struct returns ----------===//
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 This pass attempts to replace out argument usage with a return of a
10 /// struct.
11 ///
12 /// We can support returning a lot of values directly in registers, but
13 /// idiomatic C code frequently uses a pointer argument to return a second value
14 /// rather than returning a struct by value. GPU stack access is also quite
15 /// painful, so we want to avoid that if possible. Passing a stack object
16 /// pointer to a function also requires an additional address expansion code
17 /// sequence to convert the pointer to be relative to the kernel's scratch wave
18 /// offset register since the callee doesn't know what stack frame the incoming
19 /// pointer is relative to.
20 ///
21 /// The goal is to try rewriting code that looks like this:
22 ///
23 /// int foo(int a, int b, int* out) {
24 /// *out = bar();
25 /// return a + b;
26 /// }
27 ///
28 /// into something like this:
29 ///
30 /// std::pair<int, int> foo(int a, int b) {
31 /// return std::pair(a + b, bar());
32 /// }
33 ///
34 /// Typically the incoming pointer is a simple alloca for a temporary variable
35 /// to use the API, which if replaced with a struct return will be easily SROA'd
36 /// out when the stub function we create is inlined
37 ///
38 /// This pass introduces the struct return, but leaves the unused pointer
39 /// arguments and introduces a new stub function calling the struct returning
40 /// body. DeadArgumentElimination should be run after this to clean these up.
41 //
42 //===----------------------------------------------------------------------===//
43
44 #include "AMDGPU.h"
45 #include "Utils/AMDGPUBaseInfo.h"
46 #include "llvm/ADT/SmallSet.h"
47 #include "llvm/ADT/Statistic.h"
48 #include "llvm/Analysis/MemoryDependenceAnalysis.h"
49 #include "llvm/IR/IRBuilder.h"
50 #include "llvm/IR/Instructions.h"
51 #include "llvm/InitializePasses.h"
52 #include "llvm/Pass.h"
53 #include "llvm/Support/CommandLine.h"
54 #include "llvm/Support/Debug.h"
55 #include "llvm/Support/raw_ostream.h"
56
57 #define DEBUG_TYPE "amdgpu-rewrite-out-arguments"
58
59 using namespace llvm;
60
61 static cl::opt<bool> AnyAddressSpace(
62 "amdgpu-any-address-space-out-arguments",
63 cl::desc("Replace pointer out arguments with "
64 "struct returns for non-private address space"),
65 cl::Hidden,
66 cl::init(false));
67
68 static cl::opt<unsigned> MaxNumRetRegs(
69 "amdgpu-max-return-arg-num-regs",
70 cl::desc("Approximately limit number of return registers for replacing out arguments"),
71 cl::Hidden,
72 cl::init(16));
73
74 STATISTIC(NumOutArgumentsReplaced,
75 "Number out arguments moved to struct return values");
76 STATISTIC(NumOutArgumentFunctionsReplaced,
77 "Number of functions with out arguments moved to struct return values");
78
79 namespace {
80
81 class AMDGPURewriteOutArguments : public FunctionPass {
82 private:
83 const DataLayout *DL = nullptr;
84 MemoryDependenceResults *MDA = nullptr;
85
86 Type *getStoredType(Value &Arg) const;
87 Type *getOutArgumentType(Argument &Arg) const;
88
89 public:
90 static char ID;
91
AMDGPURewriteOutArguments()92 AMDGPURewriteOutArguments() : FunctionPass(ID) {}
93
getAnalysisUsage(AnalysisUsage & AU) const94 void getAnalysisUsage(AnalysisUsage &AU) const override {
95 AU.addRequired<MemoryDependenceWrapperPass>();
96 FunctionPass::getAnalysisUsage(AU);
97 }
98
99 bool doInitialization(Module &M) override;
100 bool runOnFunction(Function &F) override;
101 };
102
103 } // end anonymous namespace
104
105 INITIALIZE_PASS_BEGIN(AMDGPURewriteOutArguments, DEBUG_TYPE,
106 "AMDGPU Rewrite Out Arguments", false, false)
107 INITIALIZE_PASS_DEPENDENCY(MemoryDependenceWrapperPass)
108 INITIALIZE_PASS_END(AMDGPURewriteOutArguments, DEBUG_TYPE,
109 "AMDGPU Rewrite Out Arguments", false, false)
110
111 char AMDGPURewriteOutArguments::ID = 0;
112
getStoredType(Value & Arg) const113 Type *AMDGPURewriteOutArguments::getStoredType(Value &Arg) const {
114 const int MaxUses = 10;
115 int UseCount = 0;
116
117 SmallVector<Use *> Worklist;
118 for (Use &U : Arg.uses())
119 Worklist.push_back(&U);
120
121 Type *StoredType = nullptr;
122 while (!Worklist.empty()) {
123 Use *U = Worklist.pop_back_val();
124
125 if (auto *BCI = dyn_cast<BitCastInst>(U->getUser())) {
126 for (Use &U : BCI->uses())
127 Worklist.push_back(&U);
128 continue;
129 }
130
131 if (auto *SI = dyn_cast<StoreInst>(U->getUser())) {
132 if (UseCount++ > MaxUses)
133 return nullptr;
134
135 if (!SI->isSimple() ||
136 U->getOperandNo() != StoreInst::getPointerOperandIndex())
137 return nullptr;
138
139 if (StoredType && StoredType != SI->getValueOperand()->getType())
140 return nullptr; // More than one type.
141 StoredType = SI->getValueOperand()->getType();
142 continue;
143 }
144
145 // Unsupported user.
146 return nullptr;
147 }
148
149 return StoredType;
150 }
151
getOutArgumentType(Argument & Arg) const152 Type *AMDGPURewriteOutArguments::getOutArgumentType(Argument &Arg) const {
153 const unsigned MaxOutArgSizeBytes = 4 * MaxNumRetRegs;
154 PointerType *ArgTy = dyn_cast<PointerType>(Arg.getType());
155
156 // TODO: It might be useful for any out arguments, not just privates.
157 if (!ArgTy || (ArgTy->getAddressSpace() != DL->getAllocaAddrSpace() &&
158 !AnyAddressSpace) ||
159 Arg.hasByValAttr() || Arg.hasStructRetAttr()) {
160 return nullptr;
161 }
162
163 Type *StoredType = getStoredType(Arg);
164 if (!StoredType || DL->getTypeStoreSize(StoredType) > MaxOutArgSizeBytes)
165 return nullptr;
166
167 return StoredType;
168 }
169
doInitialization(Module & M)170 bool AMDGPURewriteOutArguments::doInitialization(Module &M) {
171 DL = &M.getDataLayout();
172 return false;
173 }
174
runOnFunction(Function & F)175 bool AMDGPURewriteOutArguments::runOnFunction(Function &F) {
176 if (skipFunction(F))
177 return false;
178
179 // TODO: Could probably handle variadic functions.
180 if (F.isVarArg() || F.hasStructRetAttr() ||
181 AMDGPU::isEntryFunctionCC(F.getCallingConv()))
182 return false;
183
184 MDA = &getAnalysis<MemoryDependenceWrapperPass>().getMemDep();
185
186 unsigned ReturnNumRegs = 0;
187 SmallDenseMap<int, Type *, 4> OutArgIndexes;
188 SmallVector<Type *, 4> ReturnTypes;
189 Type *RetTy = F.getReturnType();
190 if (!RetTy->isVoidTy()) {
191 ReturnNumRegs = DL->getTypeStoreSize(RetTy) / 4;
192
193 if (ReturnNumRegs >= MaxNumRetRegs)
194 return false;
195
196 ReturnTypes.push_back(RetTy);
197 }
198
199 SmallVector<std::pair<Argument *, Type *>, 4> OutArgs;
200 for (Argument &Arg : F.args()) {
201 if (Type *Ty = getOutArgumentType(Arg)) {
202 LLVM_DEBUG(dbgs() << "Found possible out argument " << Arg
203 << " in function " << F.getName() << '\n');
204 OutArgs.push_back({&Arg, Ty});
205 }
206 }
207
208 if (OutArgs.empty())
209 return false;
210
211 using ReplacementVec = SmallVector<std::pair<Argument *, Value *>, 4>;
212
213 DenseMap<ReturnInst *, ReplacementVec> Replacements;
214
215 SmallVector<ReturnInst *, 4> Returns;
216 for (BasicBlock &BB : F) {
217 if (ReturnInst *RI = dyn_cast<ReturnInst>(&BB.back()))
218 Returns.push_back(RI);
219 }
220
221 if (Returns.empty())
222 return false;
223
224 bool Changing;
225
226 do {
227 Changing = false;
228
229 // Keep retrying if we are able to successfully eliminate an argument. This
230 // helps with cases with multiple arguments which may alias, such as in a
231 // sincos implementation. If we have 2 stores to arguments, on the first
232 // attempt the MDA query will succeed for the second store but not the
233 // first. On the second iteration we've removed that out clobbering argument
234 // (by effectively moving it into another function) and will find the second
235 // argument is OK to move.
236 for (const auto &Pair : OutArgs) {
237 bool ThisReplaceable = true;
238 SmallVector<std::pair<ReturnInst *, StoreInst *>, 4> ReplaceableStores;
239
240 Argument *OutArg = Pair.first;
241 Type *ArgTy = Pair.second;
242
243 // Skip this argument if converting it will push us over the register
244 // count to return limit.
245
246 // TODO: This is an approximation. When legalized this could be more. We
247 // can ask TLI for exactly how many.
248 unsigned ArgNumRegs = DL->getTypeStoreSize(ArgTy) / 4;
249 if (ArgNumRegs + ReturnNumRegs > MaxNumRetRegs)
250 continue;
251
252 // An argument is convertible only if all exit blocks are able to replace
253 // it.
254 for (ReturnInst *RI : Returns) {
255 BasicBlock *BB = RI->getParent();
256
257 MemDepResult Q = MDA->getPointerDependencyFrom(
258 MemoryLocation::getBeforeOrAfter(OutArg), true, BB->end(), BB, RI);
259 StoreInst *SI = nullptr;
260 if (Q.isDef())
261 SI = dyn_cast<StoreInst>(Q.getInst());
262
263 if (SI) {
264 LLVM_DEBUG(dbgs() << "Found out argument store: " << *SI << '\n');
265 ReplaceableStores.emplace_back(RI, SI);
266 } else {
267 ThisReplaceable = false;
268 break;
269 }
270 }
271
272 if (!ThisReplaceable)
273 continue; // Try the next argument candidate.
274
275 for (std::pair<ReturnInst *, StoreInst *> Store : ReplaceableStores) {
276 Value *ReplVal = Store.second->getValueOperand();
277
278 auto &ValVec = Replacements[Store.first];
279 if (llvm::any_of(ValVec,
280 [OutArg](const std::pair<Argument *, Value *> &Entry) {
281 return Entry.first == OutArg;
282 })) {
283 LLVM_DEBUG(dbgs()
284 << "Saw multiple out arg stores" << *OutArg << '\n');
285 // It is possible to see stores to the same argument multiple times,
286 // but we expect these would have been optimized out already.
287 ThisReplaceable = false;
288 break;
289 }
290
291 ValVec.emplace_back(OutArg, ReplVal);
292 Store.second->eraseFromParent();
293 }
294
295 if (ThisReplaceable) {
296 ReturnTypes.push_back(ArgTy);
297 OutArgIndexes.insert({OutArg->getArgNo(), ArgTy});
298 ++NumOutArgumentsReplaced;
299 Changing = true;
300 }
301 }
302 } while (Changing);
303
304 if (Replacements.empty())
305 return false;
306
307 LLVMContext &Ctx = F.getParent()->getContext();
308 StructType *NewRetTy = StructType::create(Ctx, ReturnTypes, F.getName());
309
310 FunctionType *NewFuncTy = FunctionType::get(NewRetTy,
311 F.getFunctionType()->params(),
312 F.isVarArg());
313
314 LLVM_DEBUG(dbgs() << "Computed new return type: " << *NewRetTy << '\n');
315
316 Function *NewFunc = Function::Create(NewFuncTy, Function::PrivateLinkage,
317 F.getName() + ".body");
318 F.getParent()->getFunctionList().insert(F.getIterator(), NewFunc);
319 NewFunc->copyAttributesFrom(&F);
320 NewFunc->setComdat(F.getComdat());
321
322 // We want to preserve the function and param attributes, but need to strip
323 // off any return attributes, e.g. zeroext doesn't make sense with a struct.
324 NewFunc->stealArgumentListFrom(F);
325
326 AttributeMask RetAttrs;
327 RetAttrs.addAttribute(Attribute::SExt);
328 RetAttrs.addAttribute(Attribute::ZExt);
329 RetAttrs.addAttribute(Attribute::NoAlias);
330 NewFunc->removeRetAttrs(RetAttrs);
331 // TODO: How to preserve metadata?
332
333 // Move the body of the function into the new rewritten function, and replace
334 // this function with a stub.
335 NewFunc->splice(NewFunc->begin(), &F);
336
337 for (std::pair<ReturnInst *, ReplacementVec> &Replacement : Replacements) {
338 ReturnInst *RI = Replacement.first;
339 IRBuilder<> B(RI);
340 B.SetCurrentDebugLocation(RI->getDebugLoc());
341
342 int RetIdx = 0;
343 Value *NewRetVal = PoisonValue::get(NewRetTy);
344
345 Value *RetVal = RI->getReturnValue();
346 if (RetVal)
347 NewRetVal = B.CreateInsertValue(NewRetVal, RetVal, RetIdx++);
348
349 for (std::pair<Argument *, Value *> ReturnPoint : Replacement.second)
350 NewRetVal = B.CreateInsertValue(NewRetVal, ReturnPoint.second, RetIdx++);
351
352 if (RetVal)
353 RI->setOperand(0, NewRetVal);
354 else {
355 B.CreateRet(NewRetVal);
356 RI->eraseFromParent();
357 }
358 }
359
360 SmallVector<Value *, 16> StubCallArgs;
361 for (Argument &Arg : F.args()) {
362 if (OutArgIndexes.count(Arg.getArgNo())) {
363 // It's easier to preserve the type of the argument list. We rely on
364 // DeadArgumentElimination to take care of these.
365 StubCallArgs.push_back(PoisonValue::get(Arg.getType()));
366 } else {
367 StubCallArgs.push_back(&Arg);
368 }
369 }
370
371 BasicBlock *StubBB = BasicBlock::Create(Ctx, "", &F);
372 IRBuilder<> B(StubBB);
373 CallInst *StubCall = B.CreateCall(NewFunc, StubCallArgs);
374
375 int RetIdx = RetTy->isVoidTy() ? 0 : 1;
376 for (Argument &Arg : F.args()) {
377 if (!OutArgIndexes.count(Arg.getArgNo()))
378 continue;
379
380 PointerType *ArgType = cast<PointerType>(Arg.getType());
381
382 Type *EltTy = OutArgIndexes[Arg.getArgNo()];
383 const auto Align =
384 DL->getValueOrABITypeAlignment(Arg.getParamAlign(), EltTy);
385
386 Value *Val = B.CreateExtractValue(StubCall, RetIdx++);
387 Type *PtrTy = Val->getType()->getPointerTo(ArgType->getAddressSpace());
388
389 // We can peek through bitcasts, so the type may not match.
390 Value *PtrVal = B.CreateBitCast(&Arg, PtrTy);
391
392 B.CreateAlignedStore(Val, PtrVal, Align);
393 }
394
395 if (!RetTy->isVoidTy()) {
396 B.CreateRet(B.CreateExtractValue(StubCall, 0));
397 } else {
398 B.CreateRetVoid();
399 }
400
401 // The function is now a stub we want to inline.
402 F.addFnAttr(Attribute::AlwaysInline);
403
404 ++NumOutArgumentFunctionsReplaced;
405 return true;
406 }
407
createAMDGPURewriteOutArgumentsPass()408 FunctionPass *llvm::createAMDGPURewriteOutArgumentsPass() {
409 return new AMDGPURewriteOutArguments();
410 }
411