/// processMemCpy - perform simplication of memcpy's. If we have memcpy A which /// copies X to Y, and memcpy B which copies Y to Z, then we can rewrite B to be /// a memcpy from X to Z (or potentially a memmove, depending on circumstances). /// This allows later passes to remove the first memcpy altogether. bool MemCpyOpt::processMemCpy(MemCpyInst *M) { MemoryDependenceAnalysis &MD = getAnalysis<MemoryDependenceAnalysis>(); // The are two possible optimizations we can do for memcpy: // a) memcpy-memcpy xform which exposes redundance for DSE. // b) call-memcpy xform for return slot optimization. MemDepResult dep = MD.getDependency(M); if (!dep.isClobber()) return false; if (!isa<MemCpyInst>(dep.getInst())) { if (CallInst *C = dyn_cast<CallInst>(dep.getInst())) return performCallSlotOptzn(M, C); return false; } MemCpyInst *MDep = cast<MemCpyInst>(dep.getInst()); // We can only transforms memcpy's where the dest of one is the source of the // other if (M->getSource() != MDep->getDest()) return false; // Second, the length of the memcpy's must be the same, or the preceeding one // must be larger than the following one. ConstantInt *C1 = dyn_cast<ConstantInt>(MDep->getLength()); ConstantInt *C2 = dyn_cast<ConstantInt>(M->getLength()); if (!C1 || !C2) return false; uint64_t DepSize = C1->getValue().getZExtValue(); uint64_t CpySize = C2->getValue().getZExtValue(); if (DepSize < CpySize) return false; // Finally, we have to make sure that the dest of the second does not // alias the source of the first AliasAnalysis &AA = getAnalysis<AliasAnalysis>(); if (AA.alias(M->getRawDest(), CpySize, MDep->getRawSource(), DepSize) != AliasAnalysis::NoAlias) return false; else if (AA.alias(M->getRawDest(), CpySize, M->getRawSource(), CpySize) != AliasAnalysis::NoAlias) return false; else if (AA.alias(MDep->getRawDest(), DepSize, MDep->getRawSource(), DepSize) != AliasAnalysis::NoAlias) return false; // If all checks passed, then we can transform these memcpy's const Type *ArgTys[3] = { M->getRawDest()->getType(), MDep->getRawSource()->getType(), M->getLength()->getType() }; Function *MemCpyFun = Intrinsic::getDeclaration( M->getParent()->getParent()->getParent(), M->getIntrinsicID(), ArgTys, 3); Value *Args[5] = { M->getRawDest(), MDep->getRawSource(), M->getLength(), M->getAlignmentCst(), M->getVolatileCst() }; CallInst *C = CallInst::Create(MemCpyFun, Args, Args+5, "", M); // If C and M don't interfere, then this is a valid transformation. If they // did, this would mean that the two sources overlap, which would be bad. if (MD.getDependency(C) == dep) { MD.removeInstruction(M); M->eraseFromParent(); ++NumMemCpyInstr; return true; } // Otherwise, there was no point in doing this, so we remove the call we // inserted and act like nothing happened. MD.removeInstruction(C); C->eraseFromParent(); return false; }
// FIXME: Should try to pick the most likely to be profitable allocas first. bool AMDGPUPromoteAlloca::handleAlloca(AllocaInst &I, bool SufficientLDS) { // Array allocations are probably not worth handling, since an allocation of // the array type is the canonical form. if (!I.isStaticAlloca() || I.isArrayAllocation()) return false; IRBuilder<> Builder(&I); // First try to replace the alloca with a vector Type *AllocaTy = I.getAllocatedType(); DEBUG(dbgs() << "Trying to promote " << I << '\n'); if (tryPromoteAllocaToVector(&I, AS)) return true; // Promoted to vector. const Function &ContainingFunction = *I.getParent()->getParent(); CallingConv::ID CC = ContainingFunction.getCallingConv(); // Don't promote the alloca to LDS for shader calling conventions as the work // item ID intrinsics are not supported for these calling conventions. // Furthermore not all LDS is available for some of the stages. switch (CC) { case CallingConv::AMDGPU_KERNEL: case CallingConv::SPIR_KERNEL: break; default: DEBUG(dbgs() << " promote alloca to LDS not supported with calling convention.\n"); return false; } // Not likely to have sufficient local memory for promotion. if (!SufficientLDS) return false; const AMDGPUSubtarget &ST = TM->getSubtarget<AMDGPUSubtarget>(ContainingFunction); unsigned WorkGroupSize = ST.getFlatWorkGroupSizes(ContainingFunction).second; const DataLayout &DL = Mod->getDataLayout(); unsigned Align = I.getAlignment(); if (Align == 0) Align = DL.getABITypeAlignment(I.getAllocatedType()); // FIXME: This computed padding is likely wrong since it depends on inverse // usage order. // // FIXME: It is also possible that if we're allowed to use all of the memory // could could end up using more than the maximum due to alignment padding. uint32_t NewSize = alignTo(CurrentLocalMemUsage, Align); uint32_t AllocSize = WorkGroupSize * DL.getTypeAllocSize(AllocaTy); NewSize += AllocSize; if (NewSize > LocalMemLimit) { DEBUG(dbgs() << " " << AllocSize << " bytes of local memory not available to promote\n"); return false; } CurrentLocalMemUsage = NewSize; std::vector<Value*> WorkList; if (!collectUsesWithPtrTypes(&I, &I, WorkList)) { DEBUG(dbgs() << " Do not know how to convert all uses\n"); return false; } DEBUG(dbgs() << "Promoting alloca to local memory\n"); Function *F = I.getParent()->getParent(); Type *GVTy = ArrayType::get(I.getAllocatedType(), WorkGroupSize); GlobalVariable *GV = new GlobalVariable( *Mod, GVTy, false, GlobalValue::InternalLinkage, UndefValue::get(GVTy), Twine(F->getName()) + Twine('.') + I.getName(), nullptr, GlobalVariable::NotThreadLocal, AS.LOCAL_ADDRESS); GV->setUnnamedAddr(GlobalValue::UnnamedAddr::Global); GV->setAlignment(I.getAlignment()); Value *TCntY, *TCntZ; std::tie(TCntY, TCntZ) = getLocalSizeYZ(Builder); Value *TIdX = getWorkitemID(Builder, 0); Value *TIdY = getWorkitemID(Builder, 1); Value *TIdZ = getWorkitemID(Builder, 2); Value *Tmp0 = Builder.CreateMul(TCntY, TCntZ, "", true, true); Tmp0 = Builder.CreateMul(Tmp0, TIdX); Value *Tmp1 = Builder.CreateMul(TIdY, TCntZ, "", true, true); Value *TID = Builder.CreateAdd(Tmp0, Tmp1); TID = Builder.CreateAdd(TID, TIdZ); Value *Indices[] = { Constant::getNullValue(Type::getInt32Ty(Mod->getContext())), TID }; Value *Offset = Builder.CreateInBoundsGEP(GVTy, GV, Indices); I.mutateType(Offset->getType()); I.replaceAllUsesWith(Offset); I.eraseFromParent(); for (Value *V : WorkList) { CallInst *Call = dyn_cast<CallInst>(V); if (!Call) { if (ICmpInst *CI = dyn_cast<ICmpInst>(V)) { Value *Src0 = CI->getOperand(0); Type *EltTy = Src0->getType()->getPointerElementType(); PointerType *NewTy = PointerType::get(EltTy, AS.LOCAL_ADDRESS); if (isa<ConstantPointerNull>(CI->getOperand(0))) CI->setOperand(0, ConstantPointerNull::get(NewTy)); if (isa<ConstantPointerNull>(CI->getOperand(1))) CI->setOperand(1, ConstantPointerNull::get(NewTy)); continue; } // The operand's value should be corrected on its own and we don't want to // touch the users. if (isa<AddrSpaceCastInst>(V)) continue; Type *EltTy = V->getType()->getPointerElementType(); PointerType *NewTy = PointerType::get(EltTy, AS.LOCAL_ADDRESS); // FIXME: It doesn't really make sense to try to do this for all // instructions. V->mutateType(NewTy); // Adjust the types of any constant operands. if (SelectInst *SI = dyn_cast<SelectInst>(V)) { if (isa<ConstantPointerNull>(SI->getOperand(1))) SI->setOperand(1, ConstantPointerNull::get(NewTy)); if (isa<ConstantPointerNull>(SI->getOperand(2))) SI->setOperand(2, ConstantPointerNull::get(NewTy)); } else if (PHINode *Phi = dyn_cast<PHINode>(V)) { for (unsigned I = 0, E = Phi->getNumIncomingValues(); I != E; ++I) { if (isa<ConstantPointerNull>(Phi->getIncomingValue(I))) Phi->setIncomingValue(I, ConstantPointerNull::get(NewTy)); } } continue; } IntrinsicInst *Intr = cast<IntrinsicInst>(Call); Builder.SetInsertPoint(Intr); switch (Intr->getIntrinsicID()) { case Intrinsic::lifetime_start: case Intrinsic::lifetime_end: // These intrinsics are for address space 0 only Intr->eraseFromParent(); continue; case Intrinsic::memcpy: { MemCpyInst *MemCpy = cast<MemCpyInst>(Intr); Builder.CreateMemCpy(MemCpy->getRawDest(), MemCpy->getDestAlignment(), MemCpy->getRawSource(), MemCpy->getSourceAlignment(), MemCpy->getLength(), MemCpy->isVolatile()); Intr->eraseFromParent(); continue; } case Intrinsic::memmove: { MemMoveInst *MemMove = cast<MemMoveInst>(Intr); Builder.CreateMemMove(MemMove->getRawDest(), MemMove->getDestAlignment(), MemMove->getRawSource(), MemMove->getSourceAlignment(), MemMove->getLength(), MemMove->isVolatile()); Intr->eraseFromParent(); continue; } case Intrinsic::memset: { MemSetInst *MemSet = cast<MemSetInst>(Intr); Builder.CreateMemSet(MemSet->getRawDest(), MemSet->getValue(), MemSet->getLength(), MemSet->getDestAlignment(), MemSet->isVolatile()); Intr->eraseFromParent(); continue; } case Intrinsic::invariant_start: case Intrinsic::invariant_end: case Intrinsic::invariant_group_barrier: Intr->eraseFromParent(); // FIXME: I think the invariant marker should still theoretically apply, // but the intrinsics need to be changed to accept pointers with any // address space. continue; case Intrinsic::objectsize: { Value *Src = Intr->getOperand(0); Type *SrcTy = Src->getType()->getPointerElementType(); Function *ObjectSize = Intrinsic::getDeclaration(Mod, Intrinsic::objectsize, { Intr->getType(), PointerType::get(SrcTy, AS.LOCAL_ADDRESS) } ); CallInst *NewCall = Builder.CreateCall( ObjectSize, {Src, Intr->getOperand(1), Intr->getOperand(2)}); Intr->replaceAllUsesWith(NewCall); Intr->eraseFromParent(); continue; } default: Intr->print(errs()); llvm_unreachable("Don't know how to promote alloca intrinsic use."); } } return true; }
void AMDGPUPromoteAlloca::handleAlloca(AllocaInst &I) { // Array allocations are probably not worth handling, since an allocation of // the array type is the canonical form. if (!I.isStaticAlloca() || I.isArrayAllocation()) return; IRBuilder<> Builder(&I); // First try to replace the alloca with a vector Type *AllocaTy = I.getAllocatedType(); DEBUG(dbgs() << "Trying to promote " << I << '\n'); if (tryPromoteAllocaToVector(&I)) return; DEBUG(dbgs() << " alloca is not a candidate for vectorization.\n"); const Function &ContainingFunction = *I.getParent()->getParent(); // FIXME: We should also try to get this value from the reqd_work_group_size // function attribute if it is available. unsigned WorkGroupSize = AMDGPU::getMaximumWorkGroupSize(ContainingFunction); int AllocaSize = WorkGroupSize * Mod->getDataLayout().getTypeAllocSize(AllocaTy); if (AllocaSize > LocalMemAvailable) { DEBUG(dbgs() << " Not enough local memory to promote alloca.\n"); return; } std::vector<Value*> WorkList; if (!collectUsesWithPtrTypes(&I, WorkList)) { DEBUG(dbgs() << " Do not know how to convert all uses\n"); return; } DEBUG(dbgs() << "Promoting alloca to local memory\n"); LocalMemAvailable -= AllocaSize; Function *F = I.getParent()->getParent(); Type *GVTy = ArrayType::get(I.getAllocatedType(), WorkGroupSize); GlobalVariable *GV = new GlobalVariable( *Mod, GVTy, false, GlobalValue::InternalLinkage, UndefValue::get(GVTy), Twine(F->getName()) + Twine('.') + I.getName(), nullptr, GlobalVariable::NotThreadLocal, AMDGPUAS::LOCAL_ADDRESS); GV->setUnnamedAddr(true); GV->setAlignment(I.getAlignment()); Value *TCntY, *TCntZ; std::tie(TCntY, TCntZ) = getLocalSizeYZ(Builder); Value *TIdX = getWorkitemID(Builder, 0); Value *TIdY = getWorkitemID(Builder, 1); Value *TIdZ = getWorkitemID(Builder, 2); Value *Tmp0 = Builder.CreateMul(TCntY, TCntZ, "", true, true); Tmp0 = Builder.CreateMul(Tmp0, TIdX); Value *Tmp1 = Builder.CreateMul(TIdY, TCntZ, "", true, true); Value *TID = Builder.CreateAdd(Tmp0, Tmp1); TID = Builder.CreateAdd(TID, TIdZ); Value *Indices[] = { Constant::getNullValue(Type::getInt32Ty(Mod->getContext())), TID }; Value *Offset = Builder.CreateInBoundsGEP(GVTy, GV, Indices); I.mutateType(Offset->getType()); I.replaceAllUsesWith(Offset); I.eraseFromParent(); for (Value *V : WorkList) { CallInst *Call = dyn_cast<CallInst>(V); if (!Call) { Type *EltTy = V->getType()->getPointerElementType(); PointerType *NewTy = PointerType::get(EltTy, AMDGPUAS::LOCAL_ADDRESS); // The operand's value should be corrected on its own. if (isa<AddrSpaceCastInst>(V)) continue; // FIXME: It doesn't really make sense to try to do this for all // instructions. V->mutateType(NewTy); continue; } IntrinsicInst *Intr = dyn_cast<IntrinsicInst>(Call); if (!Intr) { // FIXME: What is this for? It doesn't make sense to promote arbitrary // function calls. If the call is to a defined function that can also be // promoted, we should be able to do this once that function is also // rewritten. std::vector<Type*> ArgTypes; for (unsigned ArgIdx = 0, ArgEnd = Call->getNumArgOperands(); ArgIdx != ArgEnd; ++ArgIdx) { ArgTypes.push_back(Call->getArgOperand(ArgIdx)->getType()); } Function *F = Call->getCalledFunction(); FunctionType *NewType = FunctionType::get(Call->getType(), ArgTypes, F->isVarArg()); Constant *C = Mod->getOrInsertFunction((F->getName() + ".local").str(), NewType, F->getAttributes()); Function *NewF = cast<Function>(C); Call->setCalledFunction(NewF); continue; } Builder.SetInsertPoint(Intr); switch (Intr->getIntrinsicID()) { case Intrinsic::lifetime_start: case Intrinsic::lifetime_end: // These intrinsics are for address space 0 only Intr->eraseFromParent(); continue; case Intrinsic::memcpy: { MemCpyInst *MemCpy = cast<MemCpyInst>(Intr); Builder.CreateMemCpy(MemCpy->getRawDest(), MemCpy->getRawSource(), MemCpy->getLength(), MemCpy->getAlignment(), MemCpy->isVolatile()); Intr->eraseFromParent(); continue; } case Intrinsic::memmove: { MemMoveInst *MemMove = cast<MemMoveInst>(Intr); Builder.CreateMemMove(MemMove->getRawDest(), MemMove->getRawSource(), MemMove->getLength(), MemMove->getAlignment(), MemMove->isVolatile()); Intr->eraseFromParent(); continue; } case Intrinsic::memset: { MemSetInst *MemSet = cast<MemSetInst>(Intr); Builder.CreateMemSet(MemSet->getRawDest(), MemSet->getValue(), MemSet->getLength(), MemSet->getAlignment(), MemSet->isVolatile()); Intr->eraseFromParent(); continue; } case Intrinsic::invariant_start: case Intrinsic::invariant_end: case Intrinsic::invariant_group_barrier: Intr->eraseFromParent(); // FIXME: I think the invariant marker should still theoretically apply, // but the intrinsics need to be changed to accept pointers with any // address space. continue; case Intrinsic::objectsize: { Value *Src = Intr->getOperand(0); Type *SrcTy = Src->getType()->getPointerElementType(); Function *ObjectSize = Intrinsic::getDeclaration(Mod, Intrinsic::objectsize, { Intr->getType(), PointerType::get(SrcTy, AMDGPUAS::LOCAL_ADDRESS) } ); CallInst *NewCall = Builder.CreateCall(ObjectSize, { Src, Intr->getOperand(1) }); Intr->replaceAllUsesWith(NewCall); Intr->eraseFromParent(); continue; } default: Intr->dump(); llvm_unreachable("Don't know how to promote alloca intrinsic use."); } } }
void AMDGPUPromoteAlloca::visitAlloca(AllocaInst &I) { IRBuilder<> Builder(&I); // First try to replace the alloca with a vector Type *AllocaTy = I.getAllocatedType(); DEBUG(dbgs() << "Trying to promote " << I << '\n'); if (tryPromoteAllocaToVector(&I)) return; DEBUG(dbgs() << " alloca is not a candidate for vectorization.\n"); // FIXME: This is the maximum work group size. We should try to get // value from the reqd_work_group_size function attribute if it is // available. unsigned WorkGroupSize = 256; int AllocaSize = WorkGroupSize * Mod->getDataLayout()->getTypeAllocSize(AllocaTy); if (AllocaSize > LocalMemAvailable) { DEBUG(dbgs() << " Not enough local memory to promote alloca.\n"); return; } std::vector<Value*> WorkList; if (!collectUsesWithPtrTypes(&I, WorkList)) { DEBUG(dbgs() << " Do not know how to convert all uses\n"); return; } DEBUG(dbgs() << "Promoting alloca to local memory\n"); LocalMemAvailable -= AllocaSize; GlobalVariable *GV = new GlobalVariable( *Mod, ArrayType::get(I.getAllocatedType(), 256), false, GlobalValue::ExternalLinkage, 0, I.getName(), 0, GlobalVariable::NotThreadLocal, AMDGPUAS::LOCAL_ADDRESS); FunctionType *FTy = FunctionType::get( Type::getInt32Ty(Mod->getContext()), false); AttributeSet AttrSet; AttrSet.addAttribute(Mod->getContext(), 0, Attribute::ReadNone); Value *ReadLocalSizeY = Mod->getOrInsertFunction( "llvm.r600.read.local.size.y", FTy, AttrSet); Value *ReadLocalSizeZ = Mod->getOrInsertFunction( "llvm.r600.read.local.size.z", FTy, AttrSet); Value *ReadTIDIGX = Mod->getOrInsertFunction( "llvm.r600.read.tidig.x", FTy, AttrSet); Value *ReadTIDIGY = Mod->getOrInsertFunction( "llvm.r600.read.tidig.y", FTy, AttrSet); Value *ReadTIDIGZ = Mod->getOrInsertFunction( "llvm.r600.read.tidig.z", FTy, AttrSet); Value *TCntY = Builder.CreateCall(ReadLocalSizeY); Value *TCntZ = Builder.CreateCall(ReadLocalSizeZ); Value *TIdX = Builder.CreateCall(ReadTIDIGX); Value *TIdY = Builder.CreateCall(ReadTIDIGY); Value *TIdZ = Builder.CreateCall(ReadTIDIGZ); Value *Tmp0 = Builder.CreateMul(TCntY, TCntZ); Tmp0 = Builder.CreateMul(Tmp0, TIdX); Value *Tmp1 = Builder.CreateMul(TIdY, TCntZ); Value *TID = Builder.CreateAdd(Tmp0, Tmp1); TID = Builder.CreateAdd(TID, TIdZ); std::vector<Value*> Indices; Indices.push_back(Constant::getNullValue(Type::getInt32Ty(Mod->getContext()))); Indices.push_back(TID); Value *Offset = Builder.CreateGEP(GV, Indices); I.mutateType(Offset->getType()); I.replaceAllUsesWith(Offset); I.eraseFromParent(); for (std::vector<Value*>::iterator i = WorkList.begin(), e = WorkList.end(); i != e; ++i) { Value *V = *i; CallInst *Call = dyn_cast<CallInst>(V); if (!Call) { Type *EltTy = V->getType()->getPointerElementType(); PointerType *NewTy = PointerType::get(EltTy, AMDGPUAS::LOCAL_ADDRESS); // The operand's value should be corrected on its own. if (isa<AddrSpaceCastInst>(V)) continue; // FIXME: It doesn't really make sense to try to do this for all // instructions. V->mutateType(NewTy); continue; } IntrinsicInst *Intr = dyn_cast<IntrinsicInst>(Call); if (!Intr) { std::vector<Type*> ArgTypes; for (unsigned ArgIdx = 0, ArgEnd = Call->getNumArgOperands(); ArgIdx != ArgEnd; ++ArgIdx) { ArgTypes.push_back(Call->getArgOperand(ArgIdx)->getType()); } Function *F = Call->getCalledFunction(); FunctionType *NewType = FunctionType::get(Call->getType(), ArgTypes, F->isVarArg()); Constant *C = Mod->getOrInsertFunction(StringRef(F->getName().str() + ".local"), NewType, F->getAttributes()); Function *NewF = cast<Function>(C); Call->setCalledFunction(NewF); continue; } Builder.SetInsertPoint(Intr); switch (Intr->getIntrinsicID()) { case Intrinsic::lifetime_start: case Intrinsic::lifetime_end: // These intrinsics are for address space 0 only Intr->eraseFromParent(); continue; case Intrinsic::memcpy: { MemCpyInst *MemCpy = cast<MemCpyInst>(Intr); Builder.CreateMemCpy(MemCpy->getRawDest(), MemCpy->getRawSource(), MemCpy->getLength(), MemCpy->getAlignment(), MemCpy->isVolatile()); Intr->eraseFromParent(); continue; } case Intrinsic::memset: { MemSetInst *MemSet = cast<MemSetInst>(Intr); Builder.CreateMemSet(MemSet->getRawDest(), MemSet->getValue(), MemSet->getLength(), MemSet->getAlignment(), MemSet->isVolatile()); Intr->eraseFromParent(); continue; } default: Intr->dump(); llvm_unreachable("Don't know how to promote alloca intrinsic use."); } } }