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 
92   AMDGPURewriteOutArguments() : FunctionPass(ID) {}
93 
94   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 
113 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 
152 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 
170 bool AMDGPURewriteOutArguments::doInitialization(Module &M) {
171   DL = &M.getDataLayout();
172   return false;
173 }
174 
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 
408 FunctionPass *llvm::createAMDGPURewriteOutArgumentsPass() {
409   return new AMDGPURewriteOutArguments();
410 }
411