AMDGPURewriteOutArguments.cpp

Go to the documentation of this file.

//===- AMDGPURewriteOutArgumentsPass.cpp - Create struct returns ----------===//
//
// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
// See https://llvm.org/LICENSE.txt for license information.
// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
//
//===----------------------------------------------------------------------===//
//
/// \file This pass attempts to replace out argument usage with a return of a
/// struct.
///
/// We can support returning a lot of values directly in registers, but
/// idiomatic C code frequently uses a pointer argument to return a second value
/// rather than returning a struct by value. GPU stack access is also quite
/// painful, so we want to avoid that if possible. Passing a stack object
/// pointer to a function also requires an additional address expansion code
/// sequence to convert the pointer to be relative to the kernel's scratch wave
/// offset register since the callee doesn't know what stack frame the incoming
/// pointer is relative to.
///
/// The goal is to try rewriting code that looks like this:
///
///  int foo(int a, int b, int* out) {
///     *out = bar();
///     return a + b;
/// }
///
/// into something like this:
///
///  std::pair<int, int> foo(int a, int b) {
///     return std::pair(a + b, bar());
/// }
///
/// Typically the incoming pointer is a simple alloca for a temporary variable
/// to use the API, which if replaced with a struct return will be easily SROA'd
/// out when the stub function we create is inlined
///
/// This pass introduces the struct return, but leaves the unused pointer
/// arguments and introduces a new stub function calling the struct returning
/// body. DeadArgumentElimination should be run after this to clean these up.
//
//===----------------------------------------------------------------------===//
 
#include "AMDGPU.h"
#include "Utils/AMDGPUBaseInfo.h"
#include "llvm/ADT/Statistic.h"
#include "llvm/Analysis/MemorySSA.h"
#include "llvm/Analysis/MemorySSAUpdater.h"
#include "llvm/IR/AttributeMask.h"
#include "llvm/IR/Attributes.h"
#include "llvm/IR/IRBuilder.h"
#include "llvm/IR/Instructions.h"
#include "llvm/InitializePasses.h"
#include "llvm/Pass.h"
#include "llvm/Support/Debug.h"
#include "llvm/Support/raw_ostream.h"
 
#define DEBUG_TYPE "amdgpu-rewrite-out-arguments"
 
using namespace llvm;
 
STATISTIC(NumOutArgumentsReplaced,
          "Number out arguments moved to struct return values");
STATISTIC(NumOutArgumentFunctionsReplaced,
          "Number of functions with out arguments moved to struct return values");
 
namespace {
 
class AMDGPURewriteOutArguments : public FunctionPass {
private:
  const AMDGPUOptions *CLOpts = nullptr;
  const DataLayout *DL = nullptr;
  MemorySSA *MSSA = nullptr;
  MemorySSAUpdater *MSSAU = nullptr;
  AAResults *AA = nullptr;
 
  Type *getStoredType(Value &Arg) const;
  Type *getOutArgumentType(Argument &Arg) const;
 
public:
  static char ID;
 
  AMDGPURewriteOutArguments() : FunctionPass(ID) {}
 
  void getAnalysisUsage(AnalysisUsage &AU) const override {
    AU.addRequired<MemorySSAWrapperPass>();
    AU.addRequired<AAResultsWrapperPass>();
    FunctionPass::getAnalysisUsage(AU);
  }
 
  bool doInitialization(Module &M) override;
  bool runOnFunction(Function &F) override;
};
 
} // end anonymous namespace
 
INITIALIZE_PASS_BEGIN(AMDGPURewriteOutArguments, DEBUG_TYPE,
                      "AMDGPU Rewrite Out Arguments", false, false)
INITIALIZE_PASS_DEPENDENCY(MemorySSAWrapperPass)
INITIALIZE_PASS_END(AMDGPURewriteOutArguments, DEBUG_TYPE,
                    "AMDGPU Rewrite Out Arguments", false, false)
 
char AMDGPURewriteOutArguments::ID = 0;
 
Type *AMDGPURewriteOutArguments::getStoredType(Value &Arg) const {
  const int MaxUses = 10;
  int UseCount = 0;
 
  SmallVector<Use *> Worklist(llvm::make_pointer_range(Arg.uses()));
 
  Type *StoredType = nullptr;
  while (!Worklist.empty()) {
    Use *U = Worklist.pop_back_val();
 
    if (auto *BCI = dyn_cast<BitCastInst>(U->getUser())) {
      for (Use &U : BCI->uses())
        Worklist.push_back(&U);
      continue;
    }
 
    if (auto *SI = dyn_cast<StoreInst>(U->getUser())) {
      if (UseCount++ > MaxUses)
        return nullptr;
 
      if (!SI->isSimple() ||
          U->getOperandNo() != StoreInst::getPointerOperandIndex())
        return nullptr;
 
      if (StoredType && StoredType != SI->getValueOperand()->getType())
        return nullptr; // More than one type.
      StoredType = SI->getValueOperand()->getType();
      continue;
    }
 
    // Unsupported user.
    return nullptr;
  }
 
  return StoredType;
}
 
Type *AMDGPURewriteOutArguments::getOutArgumentType(Argument &Arg) const {
  const unsigned MaxOutArgSizeBytes = 4 * CLOpts->max_return_arg_num_regs;
  PointerType *ArgTy = dyn_cast<PointerType>(Arg.getType());
 
  // TODO: It might be useful for any out arguments, not just privates.
  if (!ArgTy ||
      (ArgTy->getAddressSpace() != DL->getAllocaAddrSpace() &&
       !CLOpts->any_address_space_out_arguments) ||
      Arg.hasByValAttr() || Arg.hasStructRetAttr()) {
    return nullptr;
  }
 
  Type *StoredType = getStoredType(Arg);
  if (!StoredType || DL->getTypeStoreSize(StoredType) > MaxOutArgSizeBytes)
    return nullptr;
 
  return StoredType;
}
 
bool AMDGPURewriteOutArguments::doInitialization(Module &M) {
  DL = &M.getDataLayout();
  return false;
}
 
static StoreInst *findStoreForOutArgument(BasicBlock *BB, Argument *OutArg,
                                          MemorySSA &MSSA,
                                          BatchAAResults &BAA) {
  MemoryLocation ArgLoc = MemoryLocation::getBeforeOrAfter(OutArg);
  const auto *Accesses = MSSA.getBlockAccesses(BB);
  if (!Accesses)
    return nullptr;
 
  for (const MemoryAccess &Access : reverse(*Accesses)) {
    const auto *UseOrDef = dyn_cast<MemoryUseOrDef>(&Access);
    if (!UseOrDef)
      continue;
 
    Instruction *I = UseOrDef->getMemoryInst();
 
    // Return the must-alias store to the out argument.
    if (auto *Store = dyn_cast<StoreInst>(I))
      if (Store->getPointerOperand() == OutArg)
        return Store;
 
    if (auto *FI = dyn_cast<FenceInst>(I))
      if (FI->getOrdering() == AtomicOrdering::Release)
        continue;
 
    if (auto *LI = dyn_cast<LoadInst>(I)) {
      if (LI->isAtomic()) {
        // May-alias reads with monotonic ordering are ignored.
        if (isStrongerThan(LI->getOrdering(), AtomicOrdering::Monotonic))
          return nullptr;
        continue;
      }
    }
 
    // Any other memory access that may read or write the location prevents the
    // rewrite.
    if (isModOrRefSet(BAA.getModRefInfo(I, ArgLoc)))
      return nullptr;
  }
 
  return nullptr;
}
 
bool AMDGPURewriteOutArguments::runOnFunction(Function &F) {
  if (skipFunction(F))
    return false;
 
  CLOpts = &F.getContext().getLibraryOptions().get<AMDGPUOptions>();
  const unsigned MaxNumRetRegs = CLOpts->max_return_arg_num_regs;
 
  // TODO: Could probably handle variadic functions.
  if (F.isVarArg() || F.hasStructRetAttr() ||
      AMDGPU::isEntryFunctionCC(F.getCallingConv()))
    return false;
 
  unsigned ReturnNumRegs = 0;
  // Maps an out-argument number to its field index in the return struct.
  // Fields are in processing order, which the retry loop below can reorder
  // relative to argument order, so the index must be tracked, not recomputed.
  SmallDenseMap<unsigned, unsigned, 4> OutArgIndexes;
  SmallVector<Type *, 4> ReturnTypes;
  Type *RetTy = F.getReturnType();
  if (!RetTy->isVoidTy()) {
    ReturnNumRegs = DL->getTypeStoreSize(RetTy) / 4;
 
    if (ReturnNumRegs >= MaxNumRetRegs)
      return false;
 
    ReturnTypes.push_back(RetTy);
  }
 
  SmallVector<std::pair<Argument *, Type *>, 4> OutArgs;
  for (Argument &Arg : F.args()) {
    if (Type *Ty = getOutArgumentType(Arg)) {
      LLVM_DEBUG(dbgs() << "Found possible out argument " << Arg
                        << " in function " << F.getName() << '\n');
      OutArgs.push_back({&Arg, Ty});
    }
  }
 
  if (OutArgs.empty())
    return false;
 
  using ReplacementVec = SmallVector<std::pair<Argument *, Value *>, 4>;
 
  DenseMap<ReturnInst *, ReplacementVec> Replacements;
 
  SmallVector<ReturnInst *, 4> Returns;
  for (BasicBlock &BB : F) {
    if (ReturnInst *RI = dyn_cast<ReturnInst>(&BB.back()))
      Returns.push_back(RI);
  }
 
  if (Returns.empty())
    return false;
 
  AA = &getAnalysis<AAResultsWrapperPass>().getAAResults();
  MSSA = &getAnalysis<MemorySSAWrapperPass>().getMSSA();
 
  BatchAAResults BatchAA(*AA);
  MemorySSAUpdater MSSAUpdater(MSSA);
  MSSAU = &MSSAUpdater;
 
  bool Changing;
 
  do {
    Changing = false;
 
    // Keep retrying if we are able to successfully eliminate an argument. This
    // helps with cases with multiple arguments which may alias, such as in a
    // sincos implementation. With 2 stores to may-aliasing arguments, MDA
    // returns the second store for the first argument too; the identity guard
    // below rejects it, and a later iteration folds the first argument once the
    // second store has been removed.
    for (const auto &Pair : OutArgs) {
      bool ThisReplaceable = true;
      SmallVector<std::pair<ReturnInst *, StoreInst *>, 4> ReplaceableStores;
 
      Argument *OutArg = Pair.first;
      Type *ArgTy = Pair.second;
 
      // Skip this argument if converting it will push us over the register
      // count to return limit.
 
      // TODO: This is an approximation. When legalized this could be more. We
      // can ask TLI for exactly how many.
      unsigned ArgNumRegs = DL->getTypeStoreSize(ArgTy) / 4;
      if (ArgNumRegs + ReturnNumRegs > MaxNumRetRegs)
        continue;
 
      // An argument is convertible only if all exit blocks are able to replace
      // it.
      for (ReturnInst *RI : Returns) {
        BasicBlock *BB = RI->getParent();
 
        StoreInst *SI = findStoreForOutArgument(BB, OutArg, *MSSA, BatchAA);
        if (SI) {
          LLVM_DEBUG(dbgs() << "Found out argument store: " << *SI << '\n');
          ReplaceableStores.emplace_back(RI, SI);
        } else {
          ThisReplaceable = false;
          break;
        }
      }
 
      if (!ThisReplaceable)
        continue; // Try the next argument candidate.
 
      for (std::pair<ReturnInst *, StoreInst *> Store : ReplaceableStores) {
        Value *ReplVal = Store.second->getValueOperand();
 
        auto &ValVec = Replacements[Store.first];
        if (llvm::is_contained(llvm::make_first_range(ValVec), OutArg)) {
          LLVM_DEBUG(dbgs()
                     << "Saw multiple out arg stores" << *OutArg << '\n');
          // It is possible to see stores to the same argument multiple times,
          // but we expect these would have been optimized out already.
          ThisReplaceable = false;
          break;
        }
 
        ValVec.emplace_back(OutArg, ReplVal);
        MSSAU->removeMemoryAccess(Store.second);
        Store.second->eraseFromParent();
      }
 
      if (ThisReplaceable) {
        OutArgIndexes.insert({OutArg->getArgNo(), ReturnTypes.size()});
        ReturnTypes.push_back(ArgTy);
        ++NumOutArgumentsReplaced;
        Changing = true;
      }
    }
  } while (Changing);
 
  if (Replacements.empty())
    return false;
 
  LLVMContext &Ctx = F.getContext();
  StructType *NewRetTy = StructType::create(Ctx, ReturnTypes, F.getName());
 
  FunctionType *NewFuncTy = FunctionType::get(NewRetTy,
                                              F.getFunctionType()->params(),
                                              F.isVarArg());
 
  LLVM_DEBUG(dbgs() << "Computed new return type: " << *NewRetTy << '\n');
 
  Function *NewFunc = Function::Create(NewFuncTy, Function::PrivateLinkage,
                                       F.getName() + ".body");
  F.getParent()->getFunctionList().insert(F.getIterator(), NewFunc);
  NewFunc->copyAttributesFrom(&F);
  NewFunc->setComdat(F.getComdat());
 
  // We want to preserve the function and param attributes, but need to strip
  // off any return attributes, e.g. zeroext doesn't make sense with a struct.
  NewFunc->stealArgumentListFrom(F);
 
  NewFunc->removeRetAttrs(AttributeFuncs::typeIncompatible(
      NewRetTy, NewFunc->getAttributes().getRetAttrs()));
  // TODO: How to preserve metadata?
 
  // Move the body of the function into the new rewritten function, and replace
  // this function with a stub.
  NewFunc->splice(NewFunc->begin(), &F);
 
  for (auto &Replacement : Replacements) {
    ReturnInst *RI = Replacement.first;
    IRBuilder<> B(RI);
    B.SetCurrentDebugLocation(RI->getDebugLoc());
 
    Value *NewRetVal = PoisonValue::get(NewRetTy);
 
    Value *RetVal = RI->getReturnValue();
    if (RetVal)
      NewRetVal = B.CreateInsertValue(NewRetVal, RetVal, 0);
 
    // Use OutArgIndexes so body and stub agree on the field for each argument.
    for (std::pair<Argument *, Value *> ReturnPoint : Replacement.second) {
      unsigned FieldIdx = OutArgIndexes.lookup(ReturnPoint.first->getArgNo());
      NewRetVal = B.CreateInsertValue(NewRetVal, ReturnPoint.second, FieldIdx);
    }
 
    if (RetVal)
      RI->setOperand(0, NewRetVal);
    else {
      B.CreateRet(NewRetVal);
      RI->eraseFromParent();
    }
  }
 
  SmallVector<Value *, 16> StubCallArgs;
  for (Argument &Arg : F.args()) {
    if (OutArgIndexes.count(Arg.getArgNo())) {
      // It's easier to preserve the type of the argument list. We rely on
      // DeadArgumentElimination to take care of these.
      StubCallArgs.push_back(PoisonValue::get(Arg.getType()));
    } else {
      StubCallArgs.push_back(&Arg);
    }
  }
 
  BasicBlock *StubBB = BasicBlock::Create(Ctx, "", &F);
  IRBuilder<> B(StubBB);
  CallInst *StubCall = B.CreateCall(NewFunc, StubCallArgs);
 
  for (Argument &Arg : F.args()) {
    auto It = OutArgIndexes.find(Arg.getArgNo());
    if (It == OutArgIndexes.end())
      continue;
 
    unsigned FieldIdx = It->second;
    Type *EltTy = NewRetTy->getElementType(FieldIdx);
    const auto Align =
        DL->getValueOrABITypeAlignment(Arg.getParamAlign(), EltTy);
 
    Value *Val = B.CreateExtractValue(StubCall, FieldIdx);
    B.CreateAlignedStore(Val, &Arg, Align);
  }
 
  if (!RetTy->isVoidTy()) {
    B.CreateRet(B.CreateExtractValue(StubCall, 0));
  } else {
    B.CreateRetVoid();
  }
 
  // The function is now a stub we want to inline.
  F.addFnAttr(Attribute::AlwaysInline);
 
  ++NumOutArgumentFunctionsReplaced;
  return true;
}