|
|
@@ -307,767 +307,6 @@ FunctionPass *llvm::createScalarReplAggregatesHLSLPass(bool UseDomTree, bool Pro
|
|
|
return new SROA_SSAUp_HLSL(Promote);
|
|
|
}
|
|
|
|
|
|
-//===----------------------------------------------------------------------===//
|
|
|
-// Convert To Scalar Optimization.
|
|
|
-//===----------------------------------------------------------------------===//
|
|
|
-
|
|
|
-namespace {
|
|
|
-/// ConvertToScalarInfo - This class implements the "Convert To Scalar"
|
|
|
-/// optimization, which scans the uses of an alloca and determines if it can
|
|
|
-/// rewrite it in terms of a single new alloca that can be mem2reg'd.
|
|
|
-class ConvertToScalarInfo {
|
|
|
- /// AllocaSize - The size of the alloca being considered in bytes.
|
|
|
- unsigned AllocaSize;
|
|
|
- const DataLayout &DL;
|
|
|
- unsigned ScalarLoadThreshold;
|
|
|
-
|
|
|
- /// IsNotTrivial - This is set to true if there is some access to the object
|
|
|
- /// which means that mem2reg can't promote it.
|
|
|
- bool IsNotTrivial;
|
|
|
-
|
|
|
- /// ScalarKind - Tracks the kind of alloca being considered for promotion,
|
|
|
- /// computed based on the uses of the alloca rather than the LLVM type system.
|
|
|
- enum {
|
|
|
- Unknown,
|
|
|
-
|
|
|
- // Accesses via GEPs that are consistent with element access of a vector
|
|
|
- // type. This will not be converted into a vector unless there is a later
|
|
|
- // access using an actual vector type.
|
|
|
- ImplicitVector,
|
|
|
-
|
|
|
- // Accesses via vector operations and GEPs that are consistent with the
|
|
|
- // layout of a vector type.
|
|
|
- Vector,
|
|
|
-
|
|
|
- // An integer bag-of-bits with bitwise operations for insertion and
|
|
|
- // extraction. Any combination of types can be converted into this kind
|
|
|
- // of scalar.
|
|
|
- Integer
|
|
|
- } ScalarKind;
|
|
|
-
|
|
|
- /// VectorTy - This tracks the type that we should promote the vector to if
|
|
|
- /// it is possible to turn it into a vector. This starts out null, and if it
|
|
|
- /// isn't possible to turn into a vector type, it gets set to VoidTy.
|
|
|
- VectorType *VectorTy;
|
|
|
-
|
|
|
- /// HadNonMemTransferAccess - True if there is at least one access to the
|
|
|
- /// alloca that is not a MemTransferInst. We don't want to turn structs into
|
|
|
- /// large integers unless there is some potential for optimization.
|
|
|
- bool HadNonMemTransferAccess;
|
|
|
-
|
|
|
- /// HadDynamicAccess - True if some element of this alloca was dynamic.
|
|
|
- /// We don't yet have support for turning a dynamic access into a large
|
|
|
- /// integer.
|
|
|
- bool HadDynamicAccess;
|
|
|
-
|
|
|
-public:
|
|
|
- explicit ConvertToScalarInfo(unsigned Size, const DataLayout &DL,
|
|
|
- unsigned SLT)
|
|
|
- : AllocaSize(Size), DL(DL), ScalarLoadThreshold(SLT), IsNotTrivial(false),
|
|
|
- ScalarKind(Unknown), VectorTy(nullptr), HadNonMemTransferAccess(false),
|
|
|
- HadDynamicAccess(false) {}
|
|
|
-
|
|
|
- AllocaInst *TryConvert(AllocaInst *AI);
|
|
|
-
|
|
|
-private:
|
|
|
- bool CanConvertToScalar(Value *V, uint64_t Offset, Value *NonConstantIdx);
|
|
|
- void MergeInTypeForLoadOrStore(Type *In, uint64_t Offset);
|
|
|
- bool MergeInVectorType(VectorType *VInTy, uint64_t Offset);
|
|
|
- void ConvertUsesToScalar(Value *Ptr, AllocaInst *NewAI, uint64_t Offset,
|
|
|
- Value *NonConstantIdx);
|
|
|
-
|
|
|
- Value *ConvertScalar_ExtractValue(Value *NV, Type *ToType, uint64_t Offset,
|
|
|
- Value *NonConstantIdx,
|
|
|
- IRBuilder<> &Builder);
|
|
|
- Value *ConvertScalar_InsertValue(Value *StoredVal, Value *ExistingVal,
|
|
|
- uint64_t Offset, Value *NonConstantIdx,
|
|
|
- IRBuilder<> &Builder);
|
|
|
-};
|
|
|
-} // end anonymous namespace.
|
|
|
-
|
|
|
-/// TryConvert - Analyze the specified alloca, and if it is safe to do so,
|
|
|
-/// rewrite it to be a new alloca which is mem2reg'able. This returns the new
|
|
|
-/// alloca if possible or null if not.
|
|
|
-AllocaInst *ConvertToScalarInfo::TryConvert(AllocaInst *AI) {
|
|
|
- // If we can't convert this scalar, or if mem2reg can trivially do it, bail
|
|
|
- // out.
|
|
|
- if (!CanConvertToScalar(AI, 0, nullptr) || !IsNotTrivial)
|
|
|
- return nullptr;
|
|
|
-
|
|
|
- // If an alloca has only memset / memcpy uses, it may still have an Unknown
|
|
|
- // ScalarKind. Treat it as an Integer below.
|
|
|
- if (ScalarKind == Unknown)
|
|
|
- ScalarKind = Integer;
|
|
|
-
|
|
|
- if (ScalarKind == Vector && VectorTy->getBitWidth() != AllocaSize * 8)
|
|
|
- ScalarKind = Integer;
|
|
|
-
|
|
|
- // If we were able to find a vector type that can handle this with
|
|
|
- // insert/extract elements, and if there was at least one use that had
|
|
|
- // a vector type, promote this to a vector. We don't want to promote
|
|
|
- // random stuff that doesn't use vectors (e.g. <9 x double>) because then
|
|
|
- // we just get a lot of insert/extracts. If at least one vector is
|
|
|
- // involved, then we probably really do have a union of vector/array.
|
|
|
- Type *NewTy;
|
|
|
- if (ScalarKind == Vector) {
|
|
|
- assert(VectorTy && "Missing type for vector scalar.");
|
|
|
- DEBUG(dbgs() << "CONVERT TO VECTOR: " << *AI << "\n TYPE = " << *VectorTy
|
|
|
- << '\n');
|
|
|
- NewTy = VectorTy; // Use the vector type.
|
|
|
- } else {
|
|
|
- unsigned BitWidth = AllocaSize * 8;
|
|
|
-
|
|
|
- // Do not convert to scalar integer if the alloca size exceeds the
|
|
|
- // scalar load threshold.
|
|
|
- if (BitWidth > ScalarLoadThreshold)
|
|
|
- return nullptr;
|
|
|
-
|
|
|
- if ((ScalarKind == ImplicitVector || ScalarKind == Integer) &&
|
|
|
- !HadNonMemTransferAccess && !DL.fitsInLegalInteger(BitWidth))
|
|
|
- return nullptr;
|
|
|
- // Dynamic accesses on integers aren't yet supported. They need us to shift
|
|
|
- // by a dynamic amount which could be difficult to work out as we might not
|
|
|
- // know whether to use a left or right shift.
|
|
|
- if (ScalarKind == Integer && HadDynamicAccess)
|
|
|
- return nullptr;
|
|
|
-
|
|
|
- DEBUG(dbgs() << "CONVERT TO SCALAR INTEGER: " << *AI << "\n");
|
|
|
- // Create and insert the integer alloca.
|
|
|
- NewTy = IntegerType::get(AI->getContext(), BitWidth);
|
|
|
- }
|
|
|
- AllocaInst *NewAI =
|
|
|
- new AllocaInst(NewTy, nullptr, "", AI->getParent()->begin());
|
|
|
- ConvertUsesToScalar(AI, NewAI, 0, nullptr);
|
|
|
- return NewAI;
|
|
|
-}
|
|
|
-
|
|
|
-/// MergeInTypeForLoadOrStore - Add the 'In' type to the accumulated vector type
|
|
|
-/// (VectorTy) so far at the offset specified by Offset (which is specified in
|
|
|
-/// bytes).
|
|
|
-///
|
|
|
-/// There are two cases we handle here:
|
|
|
-/// 1) A union of vector types of the same size and potentially its elements.
|
|
|
-/// Here we turn element accesses into insert/extract element operations.
|
|
|
-/// This promotes a <4 x float> with a store of float to the third element
|
|
|
-/// into a <4 x float> that uses insert element.
|
|
|
-/// 2) A fully general blob of memory, which we turn into some (potentially
|
|
|
-/// large) integer type with extract and insert operations where the loads
|
|
|
-/// and stores would mutate the memory. We mark this by setting VectorTy
|
|
|
-/// to VoidTy.
|
|
|
-void ConvertToScalarInfo::MergeInTypeForLoadOrStore(Type *In, uint64_t Offset) {
|
|
|
- // If we already decided to turn this into a blob of integer memory, there is
|
|
|
- // nothing to be done.
|
|
|
- if (ScalarKind == Integer)
|
|
|
- return;
|
|
|
-
|
|
|
- // If this could be contributing to a vector, analyze it.
|
|
|
-
|
|
|
- // If the In type is a vector that is the same size as the alloca, see if it
|
|
|
- // matches the existing VecTy.
|
|
|
- if (VectorType *VInTy = dyn_cast<VectorType>(In)) {
|
|
|
- if (MergeInVectorType(VInTy, Offset))
|
|
|
- return;
|
|
|
- } else if (In->isFloatTy() || In->isDoubleTy() ||
|
|
|
- (In->isIntegerTy() && In->getPrimitiveSizeInBits() >= 8 &&
|
|
|
- isPowerOf2_32(In->getPrimitiveSizeInBits()))) {
|
|
|
- // Full width accesses can be ignored, because they can always be turned
|
|
|
- // into bitcasts.
|
|
|
- unsigned EltSize = In->getPrimitiveSizeInBits() / 8;
|
|
|
- if (EltSize == AllocaSize)
|
|
|
- return;
|
|
|
-
|
|
|
- // If we're accessing something that could be an element of a vector, see
|
|
|
- // if the implied vector agrees with what we already have and if Offset is
|
|
|
- // compatible with it.
|
|
|
- if (Offset % EltSize == 0 && AllocaSize % EltSize == 0 &&
|
|
|
- (!VectorTy ||
|
|
|
- EltSize == VectorTy->getElementType()->getPrimitiveSizeInBits() / 8)) {
|
|
|
- if (!VectorTy) {
|
|
|
- ScalarKind = ImplicitVector;
|
|
|
- VectorTy = VectorType::get(In, AllocaSize / EltSize);
|
|
|
- }
|
|
|
- return;
|
|
|
- }
|
|
|
- }
|
|
|
-
|
|
|
- // Otherwise, we have a case that we can't handle with an optimized vector
|
|
|
- // form. We can still turn this into a large integer.
|
|
|
- ScalarKind = Integer;
|
|
|
-}
|
|
|
-
|
|
|
-/// MergeInVectorType - Handles the vector case of MergeInTypeForLoadOrStore,
|
|
|
-/// returning true if the type was successfully merged and false otherwise.
|
|
|
-bool ConvertToScalarInfo::MergeInVectorType(VectorType *VInTy,
|
|
|
- uint64_t Offset) {
|
|
|
- if (VInTy->getBitWidth() / 8 == AllocaSize && Offset == 0) {
|
|
|
- // If we're storing/loading a vector of the right size, allow it as a
|
|
|
- // vector. If this the first vector we see, remember the type so that
|
|
|
- // we know the element size. If this is a subsequent access, ignore it
|
|
|
- // even if it is a differing type but the same size. Worst case we can
|
|
|
- // bitcast the resultant vectors.
|
|
|
- if (!VectorTy)
|
|
|
- VectorTy = VInTy;
|
|
|
- ScalarKind = Vector;
|
|
|
- return true;
|
|
|
- }
|
|
|
-
|
|
|
- return false;
|
|
|
-}
|
|
|
-
|
|
|
-/// CanConvertToScalar - V is a pointer. If we can convert the pointee and all
|
|
|
-/// its accesses to a single vector type, return true and set VecTy to
|
|
|
-/// the new type. If we could convert the alloca into a single promotable
|
|
|
-/// integer, return true but set VecTy to VoidTy. Further, if the use is not a
|
|
|
-/// completely trivial use that mem2reg could promote, set IsNotTrivial. Offset
|
|
|
-/// is the current offset from the base of the alloca being analyzed.
|
|
|
-///
|
|
|
-/// If we see at least one access to the value that is as a vector type, set the
|
|
|
-/// SawVec flag.
|
|
|
-bool ConvertToScalarInfo::CanConvertToScalar(Value *V, uint64_t Offset,
|
|
|
- Value *NonConstantIdx) {
|
|
|
- for (User *U : V->users()) {
|
|
|
- Instruction *UI = cast<Instruction>(U);
|
|
|
-
|
|
|
- if (LoadInst *LI = dyn_cast<LoadInst>(UI)) {
|
|
|
- // Don't break volatile loads.
|
|
|
- if (!LI->isSimple())
|
|
|
- return false;
|
|
|
-
|
|
|
- HadNonMemTransferAccess = true;
|
|
|
- MergeInTypeForLoadOrStore(LI->getType(), Offset);
|
|
|
- continue;
|
|
|
- }
|
|
|
-
|
|
|
- if (StoreInst *SI = dyn_cast<StoreInst>(UI)) {
|
|
|
- // Storing the pointer, not into the value?
|
|
|
- if (SI->getOperand(0) == V || !SI->isSimple())
|
|
|
- return false;
|
|
|
-
|
|
|
- HadNonMemTransferAccess = true;
|
|
|
- MergeInTypeForLoadOrStore(SI->getOperand(0)->getType(), Offset);
|
|
|
- continue;
|
|
|
- }
|
|
|
-
|
|
|
- if (BitCastInst *BCI = dyn_cast<BitCastInst>(UI)) {
|
|
|
- if (!onlyUsedByLifetimeMarkers(BCI))
|
|
|
- IsNotTrivial = true; // Can't be mem2reg'd.
|
|
|
- if (!CanConvertToScalar(BCI, Offset, NonConstantIdx))
|
|
|
- return false;
|
|
|
- continue;
|
|
|
- }
|
|
|
-
|
|
|
- if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(UI)) {
|
|
|
- // If this is a GEP with a variable indices, we can't handle it.
|
|
|
- PointerType *PtrTy = dyn_cast<PointerType>(GEP->getPointerOperandType());
|
|
|
- if (!PtrTy)
|
|
|
- return false;
|
|
|
-
|
|
|
- // Compute the offset that this GEP adds to the pointer.
|
|
|
- SmallVector<Value *, 8> Indices(GEP->op_begin() + 1, GEP->op_end());
|
|
|
- Value *GEPNonConstantIdx = nullptr;
|
|
|
- if (!GEP->hasAllConstantIndices()) {
|
|
|
- if (!isa<VectorType>(PtrTy->getElementType()))
|
|
|
- return false;
|
|
|
- if (NonConstantIdx)
|
|
|
- return false;
|
|
|
- GEPNonConstantIdx = Indices.pop_back_val();
|
|
|
- if (!GEPNonConstantIdx->getType()->isIntegerTy(32))
|
|
|
- return false;
|
|
|
- HadDynamicAccess = true;
|
|
|
- } else
|
|
|
- GEPNonConstantIdx = NonConstantIdx;
|
|
|
- uint64_t GEPOffset = DL.getIndexedOffset(PtrTy, Indices);
|
|
|
- // See if all uses can be converted.
|
|
|
- if (!CanConvertToScalar(GEP, Offset + GEPOffset, GEPNonConstantIdx))
|
|
|
- return false;
|
|
|
- IsNotTrivial = true; // Can't be mem2reg'd.
|
|
|
- HadNonMemTransferAccess = true;
|
|
|
- continue;
|
|
|
- }
|
|
|
-
|
|
|
- // If this is a constant sized memset of a constant value (e.g. 0) we can
|
|
|
- // handle it.
|
|
|
- if (MemSetInst *MSI = dyn_cast<MemSetInst>(UI)) {
|
|
|
- // Store to dynamic index.
|
|
|
- if (NonConstantIdx)
|
|
|
- return false;
|
|
|
- // Store of constant value.
|
|
|
- if (!isa<ConstantInt>(MSI->getValue()))
|
|
|
- return false;
|
|
|
-
|
|
|
- // Store of constant size.
|
|
|
- ConstantInt *Len = dyn_cast<ConstantInt>(MSI->getLength());
|
|
|
- if (!Len)
|
|
|
- return false;
|
|
|
-
|
|
|
- // If the size differs from the alloca, we can only convert the alloca to
|
|
|
- // an integer bag-of-bits.
|
|
|
- // FIXME: This should handle all of the cases that are currently accepted
|
|
|
- // as vector element insertions.
|
|
|
- if (Len->getZExtValue() != AllocaSize || Offset != 0)
|
|
|
- ScalarKind = Integer;
|
|
|
-
|
|
|
- IsNotTrivial = true; // Can't be mem2reg'd.
|
|
|
- HadNonMemTransferAccess = true;
|
|
|
- continue;
|
|
|
- }
|
|
|
-
|
|
|
- // If this is a memcpy or memmove into or out of the whole allocation, we
|
|
|
- // can handle it like a load or store of the scalar type.
|
|
|
- if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(UI)) {
|
|
|
- // Store to dynamic index.
|
|
|
- if (NonConstantIdx)
|
|
|
- return false;
|
|
|
- ConstantInt *Len = dyn_cast<ConstantInt>(MTI->getLength());
|
|
|
- if (!Len || Len->getZExtValue() != AllocaSize || Offset != 0)
|
|
|
- return false;
|
|
|
-
|
|
|
- IsNotTrivial = true; // Can't be mem2reg'd.
|
|
|
- continue;
|
|
|
- }
|
|
|
-
|
|
|
- // If this is a lifetime intrinsic, we can handle it.
|
|
|
- if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(UI)) {
|
|
|
- if (II->getIntrinsicID() == Intrinsic::lifetime_start ||
|
|
|
- II->getIntrinsicID() == Intrinsic::lifetime_end) {
|
|
|
- continue;
|
|
|
- }
|
|
|
- }
|
|
|
-
|
|
|
- // Otherwise, we cannot handle this!
|
|
|
- return false;
|
|
|
- }
|
|
|
-
|
|
|
- return true;
|
|
|
-}
|
|
|
-
|
|
|
-/// ConvertUsesToScalar - Convert all of the users of Ptr to use the new alloca
|
|
|
-/// directly. This happens when we are converting an "integer union" to a
|
|
|
-/// single integer scalar, or when we are converting a "vector union" to a
|
|
|
-/// vector with insert/extractelement instructions.
|
|
|
-///
|
|
|
-/// Offset is an offset from the original alloca, in bits that need to be
|
|
|
-/// shifted to the right. By the end of this, there should be no uses of Ptr.
|
|
|
-void ConvertToScalarInfo::ConvertUsesToScalar(Value *Ptr, AllocaInst *NewAI,
|
|
|
- uint64_t Offset,
|
|
|
- Value *NonConstantIdx) {
|
|
|
- while (!Ptr->use_empty()) {
|
|
|
- Instruction *User = cast<Instruction>(Ptr->user_back());
|
|
|
-
|
|
|
- if (BitCastInst *CI = dyn_cast<BitCastInst>(User)) {
|
|
|
- ConvertUsesToScalar(CI, NewAI, Offset, NonConstantIdx);
|
|
|
- CI->eraseFromParent();
|
|
|
- continue;
|
|
|
- }
|
|
|
-
|
|
|
- if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(User)) {
|
|
|
- // Compute the offset that this GEP adds to the pointer.
|
|
|
- SmallVector<Value *, 8> Indices(GEP->op_begin() + 1, GEP->op_end());
|
|
|
- Value *GEPNonConstantIdx = nullptr;
|
|
|
- if (!GEP->hasAllConstantIndices()) {
|
|
|
- assert(!NonConstantIdx &&
|
|
|
- "Dynamic GEP reading from dynamic GEP unsupported");
|
|
|
- GEPNonConstantIdx = Indices.pop_back_val();
|
|
|
- } else
|
|
|
- GEPNonConstantIdx = NonConstantIdx;
|
|
|
- uint64_t GEPOffset =
|
|
|
- DL.getIndexedOffset(GEP->getPointerOperandType(), Indices);
|
|
|
- ConvertUsesToScalar(GEP, NewAI, Offset + GEPOffset * 8,
|
|
|
- GEPNonConstantIdx);
|
|
|
- GEP->eraseFromParent();
|
|
|
- continue;
|
|
|
- }
|
|
|
-
|
|
|
- IRBuilder<> Builder(User);
|
|
|
-
|
|
|
- if (LoadInst *LI = dyn_cast<LoadInst>(User)) {
|
|
|
- // The load is a bit extract from NewAI shifted right by Offset bits.
|
|
|
- Value *LoadedVal = Builder.CreateLoad(NewAI);
|
|
|
- Value *NewLoadVal = ConvertScalar_ExtractValue(
|
|
|
- LoadedVal, LI->getType(), Offset, NonConstantIdx, Builder);
|
|
|
- LI->replaceAllUsesWith(NewLoadVal);
|
|
|
- LI->eraseFromParent();
|
|
|
- continue;
|
|
|
- }
|
|
|
-
|
|
|
- if (StoreInst *SI = dyn_cast<StoreInst>(User)) {
|
|
|
- assert(SI->getOperand(0) != Ptr && "Consistency error!");
|
|
|
- Instruction *Old = Builder.CreateLoad(NewAI, NewAI->getName() + ".in");
|
|
|
- Value *New = ConvertScalar_InsertValue(SI->getOperand(0), Old, Offset,
|
|
|
- NonConstantIdx, Builder);
|
|
|
- Builder.CreateStore(New, NewAI);
|
|
|
- SI->eraseFromParent();
|
|
|
-
|
|
|
- // If the load we just inserted is now dead, then the inserted store
|
|
|
- // overwrote the entire thing.
|
|
|
- if (Old->use_empty())
|
|
|
- Old->eraseFromParent();
|
|
|
- continue;
|
|
|
- }
|
|
|
-
|
|
|
- // If this is a constant sized memset of a constant value (e.g. 0) we can
|
|
|
- // transform it into a store of the expanded constant value.
|
|
|
- if (MemSetInst *MSI = dyn_cast<MemSetInst>(User)) {
|
|
|
- assert(MSI->getRawDest() == Ptr && "Consistency error!");
|
|
|
- assert(!NonConstantIdx && "Cannot replace dynamic memset with insert");
|
|
|
- int64_t SNumBytes = cast<ConstantInt>(MSI->getLength())->getSExtValue();
|
|
|
- if (SNumBytes > 0 && (SNumBytes >> 32) == 0) {
|
|
|
- unsigned NumBytes = static_cast<unsigned>(SNumBytes);
|
|
|
- unsigned Val = cast<ConstantInt>(MSI->getValue())->getZExtValue();
|
|
|
-
|
|
|
- // Compute the value replicated the right number of times.
|
|
|
- APInt APVal(NumBytes * 8, Val);
|
|
|
-
|
|
|
- // Splat the value if non-zero.
|
|
|
- if (Val)
|
|
|
- for (unsigned i = 1; i != NumBytes; ++i)
|
|
|
- APVal |= APVal << 8;
|
|
|
-
|
|
|
- Instruction *Old = Builder.CreateLoad(NewAI, NewAI->getName() + ".in");
|
|
|
- Value *New = ConvertScalar_InsertValue(
|
|
|
- ConstantInt::get(User->getContext(), APVal), Old, Offset, nullptr,
|
|
|
- Builder);
|
|
|
- Builder.CreateStore(New, NewAI);
|
|
|
-
|
|
|
- // If the load we just inserted is now dead, then the memset overwrote
|
|
|
- // the entire thing.
|
|
|
- if (Old->use_empty())
|
|
|
- Old->eraseFromParent();
|
|
|
- }
|
|
|
- MSI->eraseFromParent();
|
|
|
- continue;
|
|
|
- }
|
|
|
-
|
|
|
- // If this is a memcpy or memmove into or out of the whole allocation, we
|
|
|
- // can handle it like a load or store of the scalar type.
|
|
|
- if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(User)) {
|
|
|
- assert(Offset == 0 && "must be store to start of alloca");
|
|
|
- assert(!NonConstantIdx && "Cannot replace dynamic transfer with insert");
|
|
|
-
|
|
|
- // If the source and destination are both to the same alloca, then this is
|
|
|
- // a noop copy-to-self, just delete it. Otherwise, emit a load and store
|
|
|
- // as appropriate.
|
|
|
- AllocaInst *OrigAI = cast<AllocaInst>(GetUnderlyingObject(Ptr, DL, 0));
|
|
|
-
|
|
|
- if (GetUnderlyingObject(MTI->getSource(), DL, 0) != OrigAI) {
|
|
|
- // Dest must be OrigAI, change this to be a load from the original
|
|
|
- // pointer (bitcasted), then a store to our new alloca.
|
|
|
- assert(MTI->getRawDest() == Ptr && "Neither use is of pointer?");
|
|
|
- Value *SrcPtr = MTI->getSource();
|
|
|
- PointerType *SPTy = cast<PointerType>(SrcPtr->getType());
|
|
|
- PointerType *AIPTy = cast<PointerType>(NewAI->getType());
|
|
|
- if (SPTy->getAddressSpace() != AIPTy->getAddressSpace()) {
|
|
|
- AIPTy = PointerType::get(AIPTy->getElementType(),
|
|
|
- SPTy->getAddressSpace());
|
|
|
- }
|
|
|
- SrcPtr = Builder.CreateBitCast(SrcPtr, AIPTy);
|
|
|
-
|
|
|
- LoadInst *SrcVal = Builder.CreateLoad(SrcPtr, "srcval");
|
|
|
- SrcVal->setAlignment(MTI->getAlignment());
|
|
|
- Builder.CreateStore(SrcVal, NewAI);
|
|
|
- } else if (GetUnderlyingObject(MTI->getDest(), DL, 0) != OrigAI) {
|
|
|
- // Src must be OrigAI, change this to be a load from NewAI then a store
|
|
|
- // through the original dest pointer (bitcasted).
|
|
|
- assert(MTI->getRawSource() == Ptr && "Neither use is of pointer?");
|
|
|
- LoadInst *SrcVal = Builder.CreateLoad(NewAI, "srcval");
|
|
|
-
|
|
|
- PointerType *DPTy = cast<PointerType>(MTI->getDest()->getType());
|
|
|
- PointerType *AIPTy = cast<PointerType>(NewAI->getType());
|
|
|
- if (DPTy->getAddressSpace() != AIPTy->getAddressSpace()) {
|
|
|
- AIPTy = PointerType::get(AIPTy->getElementType(),
|
|
|
- DPTy->getAddressSpace());
|
|
|
- }
|
|
|
- Value *DstPtr = Builder.CreateBitCast(MTI->getDest(), AIPTy);
|
|
|
-
|
|
|
- StoreInst *NewStore = Builder.CreateStore(SrcVal, DstPtr);
|
|
|
- NewStore->setAlignment(MTI->getAlignment());
|
|
|
- } else {
|
|
|
- // Noop transfer. Src == Dst
|
|
|
- }
|
|
|
-
|
|
|
- MTI->eraseFromParent();
|
|
|
- continue;
|
|
|
- }
|
|
|
-
|
|
|
- if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(User)) {
|
|
|
- if (II->getIntrinsicID() == Intrinsic::lifetime_start ||
|
|
|
- II->getIntrinsicID() == Intrinsic::lifetime_end) {
|
|
|
- // There's no need to preserve these, as the resulting alloca will be
|
|
|
- // converted to a register anyways.
|
|
|
- II->eraseFromParent();
|
|
|
- continue;
|
|
|
- }
|
|
|
- }
|
|
|
-
|
|
|
- llvm_unreachable("Unsupported operation!");
|
|
|
- }
|
|
|
-}
|
|
|
-
|
|
|
-/// ConvertScalar_ExtractValue - Extract a value of type ToType from an integer
|
|
|
-/// or vector value FromVal, extracting the bits from the offset specified by
|
|
|
-/// Offset. This returns the value, which is of type ToType.
|
|
|
-///
|
|
|
-/// This happens when we are converting an "integer union" to a single
|
|
|
-/// integer scalar, or when we are converting a "vector union" to a vector with
|
|
|
-/// insert/extractelement instructions.
|
|
|
-///
|
|
|
-/// Offset is an offset from the original alloca, in bits that need to be
|
|
|
-/// shifted to the right.
|
|
|
-Value *ConvertToScalarInfo::ConvertScalar_ExtractValue(Value *FromVal,
|
|
|
- Type *ToType,
|
|
|
- uint64_t Offset,
|
|
|
- Value *NonConstantIdx,
|
|
|
- IRBuilder<> &Builder) {
|
|
|
- // If the load is of the whole new alloca, no conversion is needed.
|
|
|
- Type *FromType = FromVal->getType();
|
|
|
- if (FromType == ToType && Offset == 0)
|
|
|
- return FromVal;
|
|
|
-
|
|
|
- // If the result alloca is a vector type, this is either an element
|
|
|
- // access or a bitcast to another vector type of the same size.
|
|
|
- if (VectorType *VTy = dyn_cast<VectorType>(FromType)) {
|
|
|
- unsigned FromTypeSize = DL.getTypeAllocSize(FromType);
|
|
|
- unsigned ToTypeSize = DL.getTypeAllocSize(ToType);
|
|
|
- if (FromTypeSize == ToTypeSize)
|
|
|
- return Builder.CreateBitCast(FromVal, ToType);
|
|
|
-
|
|
|
- // Otherwise it must be an element access.
|
|
|
- unsigned Elt = 0;
|
|
|
- if (Offset) {
|
|
|
- unsigned EltSize = DL.getTypeAllocSizeInBits(VTy->getElementType());
|
|
|
- Elt = Offset / EltSize;
|
|
|
- assert(EltSize * Elt == Offset && "Invalid modulus in validity checking");
|
|
|
- }
|
|
|
- // Return the element extracted out of it.
|
|
|
- Value *Idx;
|
|
|
- if (NonConstantIdx) {
|
|
|
- if (Elt)
|
|
|
- Idx = Builder.CreateAdd(NonConstantIdx, Builder.getInt32(Elt),
|
|
|
- "dyn.offset");
|
|
|
- else
|
|
|
- Idx = NonConstantIdx;
|
|
|
- } else
|
|
|
- Idx = Builder.getInt32(Elt);
|
|
|
- Value *V = Builder.CreateExtractElement(FromVal, Idx);
|
|
|
- if (V->getType() != ToType)
|
|
|
- V = Builder.CreateBitCast(V, ToType);
|
|
|
- return V;
|
|
|
- }
|
|
|
-
|
|
|
- // If ToType is a first class aggregate, extract out each of the pieces and
|
|
|
- // use insertvalue's to form the FCA.
|
|
|
- if (StructType *ST = dyn_cast<StructType>(ToType)) {
|
|
|
- assert(!NonConstantIdx &&
|
|
|
- "Dynamic indexing into struct types not supported");
|
|
|
- const StructLayout &Layout = *DL.getStructLayout(ST);
|
|
|
- Value *Res = UndefValue::get(ST);
|
|
|
- for (unsigned i = 0, e = ST->getNumElements(); i != e; ++i) {
|
|
|
- Value *Elt = ConvertScalar_ExtractValue(
|
|
|
- FromVal, ST->getElementType(i),
|
|
|
- Offset + Layout.getElementOffsetInBits(i), nullptr, Builder);
|
|
|
- Res = Builder.CreateInsertValue(Res, Elt, i);
|
|
|
- }
|
|
|
- return Res;
|
|
|
- }
|
|
|
-
|
|
|
- if (ArrayType *AT = dyn_cast<ArrayType>(ToType)) {
|
|
|
- assert(!NonConstantIdx &&
|
|
|
- "Dynamic indexing into array types not supported");
|
|
|
- uint64_t EltSize = DL.getTypeAllocSizeInBits(AT->getElementType());
|
|
|
- Value *Res = UndefValue::get(AT);
|
|
|
- for (unsigned i = 0, e = AT->getNumElements(); i != e; ++i) {
|
|
|
- Value *Elt =
|
|
|
- ConvertScalar_ExtractValue(FromVal, AT->getElementType(),
|
|
|
- Offset + i * EltSize, nullptr, Builder);
|
|
|
- Res = Builder.CreateInsertValue(Res, Elt, i);
|
|
|
- }
|
|
|
- return Res;
|
|
|
- }
|
|
|
-
|
|
|
- // Otherwise, this must be a union that was converted to an integer value.
|
|
|
- IntegerType *NTy = cast<IntegerType>(FromVal->getType());
|
|
|
-
|
|
|
- // If this is a big-endian system and the load is narrower than the
|
|
|
- // full alloca type, we need to do a shift to get the right bits.
|
|
|
- int ShAmt = 0;
|
|
|
- if (DL.isBigEndian()) {
|
|
|
- // On big-endian machines, the lowest bit is stored at the bit offset
|
|
|
- // from the pointer given by getTypeStoreSizeInBits. This matters for
|
|
|
- // integers with a bitwidth that is not a multiple of 8.
|
|
|
- ShAmt = DL.getTypeStoreSizeInBits(NTy) - DL.getTypeStoreSizeInBits(ToType) -
|
|
|
- Offset;
|
|
|
- } else {
|
|
|
- ShAmt = Offset;
|
|
|
- }
|
|
|
-
|
|
|
- // Note: we support negative bitwidths (with shl) which are not defined.
|
|
|
- // We do this to support (f.e.) loads off the end of a structure where
|
|
|
- // only some bits are used.
|
|
|
- if (ShAmt > 0 && (unsigned)ShAmt < NTy->getBitWidth())
|
|
|
- FromVal = Builder.CreateLShr(FromVal,
|
|
|
- ConstantInt::get(FromVal->getType(), ShAmt));
|
|
|
- else if (ShAmt < 0 && (unsigned)-ShAmt < NTy->getBitWidth())
|
|
|
- FromVal = Builder.CreateShl(FromVal,
|
|
|
- ConstantInt::get(FromVal->getType(), -ShAmt));
|
|
|
-
|
|
|
- // Finally, unconditionally truncate the integer to the right width.
|
|
|
- unsigned LIBitWidth = DL.getTypeSizeInBits(ToType);
|
|
|
- if (LIBitWidth < NTy->getBitWidth())
|
|
|
- FromVal = Builder.CreateTrunc(
|
|
|
- FromVal, IntegerType::get(FromVal->getContext(), LIBitWidth));
|
|
|
- else if (LIBitWidth > NTy->getBitWidth())
|
|
|
- FromVal = Builder.CreateZExt(
|
|
|
- FromVal, IntegerType::get(FromVal->getContext(), LIBitWidth));
|
|
|
-
|
|
|
- // If the result is an integer, this is a trunc or bitcast.
|
|
|
- if (ToType->isIntegerTy()) {
|
|
|
- // Should be done.
|
|
|
- } else if (ToType->isFloatingPointTy() || ToType->isVectorTy()) {
|
|
|
- // Just do a bitcast, we know the sizes match up.
|
|
|
- FromVal = Builder.CreateBitCast(FromVal, ToType);
|
|
|
- } else {
|
|
|
- // Otherwise must be a pointer.
|
|
|
- FromVal = Builder.CreateIntToPtr(FromVal, ToType);
|
|
|
- }
|
|
|
- assert(FromVal->getType() == ToType && "Didn't convert right?");
|
|
|
- return FromVal;
|
|
|
-}
|
|
|
-
|
|
|
-/// ConvertScalar_InsertValue - Insert the value "SV" into the existing integer
|
|
|
-/// or vector value "Old" at the offset specified by Offset.
|
|
|
-///
|
|
|
-/// This happens when we are converting an "integer union" to a
|
|
|
-/// single integer scalar, or when we are converting a "vector union" to a
|
|
|
-/// vector with insert/extractelement instructions.
|
|
|
-///
|
|
|
-/// Offset is an offset from the original alloca, in bits that need to be
|
|
|
-/// shifted to the right.
|
|
|
-///
|
|
|
-/// NonConstantIdx is an index value if there was a GEP with a non-constant
|
|
|
-/// index value. If this is 0 then all GEPs used to find this insert address
|
|
|
-/// are constant.
|
|
|
-Value *ConvertToScalarInfo::ConvertScalar_InsertValue(Value *SV, Value *Old,
|
|
|
- uint64_t Offset,
|
|
|
- Value *NonConstantIdx,
|
|
|
- IRBuilder<> &Builder) {
|
|
|
- // Convert the stored type to the actual type, shift it left to insert
|
|
|
- // then 'or' into place.
|
|
|
- Type *AllocaType = Old->getType();
|
|
|
- LLVMContext &Context = Old->getContext();
|
|
|
-
|
|
|
- if (VectorType *VTy = dyn_cast<VectorType>(AllocaType)) {
|
|
|
- uint64_t VecSize = DL.getTypeAllocSizeInBits(VTy);
|
|
|
- uint64_t ValSize = DL.getTypeAllocSizeInBits(SV->getType());
|
|
|
-
|
|
|
- // Changing the whole vector with memset or with an access of a different
|
|
|
- // vector type?
|
|
|
- if (ValSize == VecSize)
|
|
|
- return Builder.CreateBitCast(SV, AllocaType);
|
|
|
-
|
|
|
- // Must be an element insertion.
|
|
|
- Type *EltTy = VTy->getElementType();
|
|
|
- if (SV->getType() != EltTy)
|
|
|
- SV = Builder.CreateBitCast(SV, EltTy);
|
|
|
- uint64_t EltSize = DL.getTypeAllocSizeInBits(EltTy);
|
|
|
- unsigned Elt = Offset / EltSize;
|
|
|
- Value *Idx;
|
|
|
- if (NonConstantIdx) {
|
|
|
- if (Elt)
|
|
|
- Idx = Builder.CreateAdd(NonConstantIdx, Builder.getInt32(Elt),
|
|
|
- "dyn.offset");
|
|
|
- else
|
|
|
- Idx = NonConstantIdx;
|
|
|
- } else
|
|
|
- Idx = Builder.getInt32(Elt);
|
|
|
- return Builder.CreateInsertElement(Old, SV, Idx);
|
|
|
- }
|
|
|
-
|
|
|
- // If SV is a first-class aggregate value, insert each value recursively.
|
|
|
- if (StructType *ST = dyn_cast<StructType>(SV->getType())) {
|
|
|
- assert(!NonConstantIdx &&
|
|
|
- "Dynamic indexing into struct types not supported");
|
|
|
- const StructLayout &Layout = *DL.getStructLayout(ST);
|
|
|
- for (unsigned i = 0, e = ST->getNumElements(); i != e; ++i) {
|
|
|
- Value *Elt = Builder.CreateExtractValue(SV, i);
|
|
|
- Old = ConvertScalar_InsertValue(Elt, Old,
|
|
|
- Offset + Layout.getElementOffsetInBits(i),
|
|
|
- nullptr, Builder);
|
|
|
- }
|
|
|
- return Old;
|
|
|
- }
|
|
|
-
|
|
|
- if (ArrayType *AT = dyn_cast<ArrayType>(SV->getType())) {
|
|
|
- assert(!NonConstantIdx &&
|
|
|
- "Dynamic indexing into array types not supported");
|
|
|
- uint64_t EltSize = DL.getTypeAllocSizeInBits(AT->getElementType());
|
|
|
- for (unsigned i = 0, e = AT->getNumElements(); i != e; ++i) {
|
|
|
- Value *Elt = Builder.CreateExtractValue(SV, i);
|
|
|
- Old = ConvertScalar_InsertValue(Elt, Old, Offset + i * EltSize, nullptr,
|
|
|
- Builder);
|
|
|
- }
|
|
|
- return Old;
|
|
|
- }
|
|
|
-
|
|
|
- // If SV is a float, convert it to the appropriate integer type.
|
|
|
- // If it is a pointer, do the same.
|
|
|
- unsigned SrcWidth = DL.getTypeSizeInBits(SV->getType());
|
|
|
- unsigned DestWidth = DL.getTypeSizeInBits(AllocaType);
|
|
|
- unsigned SrcStoreWidth = DL.getTypeStoreSizeInBits(SV->getType());
|
|
|
- unsigned DestStoreWidth = DL.getTypeStoreSizeInBits(AllocaType);
|
|
|
- if (SV->getType()->isFloatingPointTy() || SV->getType()->isVectorTy())
|
|
|
- SV =
|
|
|
- Builder.CreateBitCast(SV, IntegerType::get(SV->getContext(), SrcWidth));
|
|
|
- else if (SV->getType()->isPointerTy())
|
|
|
- SV = Builder.CreatePtrToInt(SV, DL.getIntPtrType(SV->getType()));
|
|
|
-
|
|
|
- // Zero extend or truncate the value if needed.
|
|
|
- if (SV->getType() != AllocaType) {
|
|
|
- if (SV->getType()->getPrimitiveSizeInBits() <
|
|
|
- AllocaType->getPrimitiveSizeInBits())
|
|
|
- SV = Builder.CreateZExt(SV, AllocaType);
|
|
|
- else {
|
|
|
- // Truncation may be needed if storing more than the alloca can hold
|
|
|
- // (undefined behavior).
|
|
|
- SV = Builder.CreateTrunc(SV, AllocaType);
|
|
|
- SrcWidth = DestWidth;
|
|
|
- SrcStoreWidth = DestStoreWidth;
|
|
|
- }
|
|
|
- }
|
|
|
-
|
|
|
- // If this is a big-endian system and the store is narrower than the
|
|
|
- // full alloca type, we need to do a shift to get the right bits.
|
|
|
- int ShAmt = 0;
|
|
|
- if (DL.isBigEndian()) {
|
|
|
- // On big-endian machines, the lowest bit is stored at the bit offset
|
|
|
- // from the pointer given by getTypeStoreSizeInBits. This matters for
|
|
|
- // integers with a bitwidth that is not a multiple of 8.
|
|
|
- ShAmt = DestStoreWidth - SrcStoreWidth - Offset;
|
|
|
- } else {
|
|
|
- ShAmt = Offset;
|
|
|
- }
|
|
|
-
|
|
|
- // Note: we support negative bitwidths (with shr) which are not defined.
|
|
|
- // We do this to support (f.e.) stores off the end of a structure where
|
|
|
- // only some bits in the structure are set.
|
|
|
- APInt Mask(APInt::getLowBitsSet(DestWidth, SrcWidth));
|
|
|
- if (ShAmt > 0 && (unsigned)ShAmt < DestWidth) {
|
|
|
- SV = Builder.CreateShl(SV, ConstantInt::get(SV->getType(), ShAmt));
|
|
|
- Mask <<= ShAmt;
|
|
|
- } else if (ShAmt < 0 && (unsigned)-ShAmt < DestWidth) {
|
|
|
- SV = Builder.CreateLShr(SV, ConstantInt::get(SV->getType(), -ShAmt));
|
|
|
- Mask = Mask.lshr(-ShAmt);
|
|
|
- }
|
|
|
-
|
|
|
- // Mask out the bits we are about to insert from the old value, and or
|
|
|
- // in the new bits.
|
|
|
- if (SrcWidth != DestWidth) {
|
|
|
- assert(DestWidth > SrcWidth);
|
|
|
- Old = Builder.CreateAnd(Old, ConstantInt::get(Context, ~Mask), "mask");
|
|
|
- SV = Builder.CreateOr(Old, SV, "ins");
|
|
|
- }
|
|
|
- return SV;
|
|
|
-}
|
|
|
-
|
|
|
//===----------------------------------------------------------------------===//
|
|
|
// SRoA Driver
|
|
|
//===----------------------------------------------------------------------===//
|
|
|
@@ -6571,778 +5810,3 @@ INITIALIZE_PASS(LowerStaticGlobalIntoAlloca, "static-global-to-alloca",
|
|
|
ModulePass *llvm::createLowerStaticGlobalIntoAlloca() {
|
|
|
return new LowerStaticGlobalIntoAlloca();
|
|
|
}
|
|
|
-
|
|
|
-//===----------------------------------------------------------------------===//
|
|
|
-// Lower one type to another type.
|
|
|
-//===----------------------------------------------------------------------===//
|
|
|
-namespace {
|
|
|
-class LowerTypePass : public ModulePass {
|
|
|
-public:
|
|
|
- explicit LowerTypePass(char &ID)
|
|
|
- : ModulePass(ID) {}
|
|
|
-
|
|
|
- bool runOnModule(Module &M) override;
|
|
|
-private:
|
|
|
- bool runOnFunction(Function &F, bool HasDbgInfo);
|
|
|
- AllocaInst *lowerAlloca(AllocaInst *A);
|
|
|
- GlobalVariable *lowerInternalGlobal(GlobalVariable *GV);
|
|
|
-protected:
|
|
|
- virtual bool needToLower(Value *V) = 0;
|
|
|
- virtual void lowerUseWithNewValue(Value *V, Value *NewV) = 0;
|
|
|
- virtual Type *lowerType(Type *Ty) = 0;
|
|
|
- virtual Constant *lowerInitVal(Constant *InitVal, Type *NewTy) = 0;
|
|
|
- virtual StringRef getGlobalPrefix() = 0;
|
|
|
- virtual void initialize(Module &M) {};
|
|
|
-};
|
|
|
-
|
|
|
-AllocaInst *LowerTypePass::lowerAlloca(AllocaInst *A) {
|
|
|
- IRBuilder<> AllocaBuilder(A);
|
|
|
- Type *NewTy = lowerType(A->getAllocatedType());
|
|
|
- return AllocaBuilder.CreateAlloca(NewTy);
|
|
|
-}
|
|
|
-
|
|
|
-GlobalVariable *LowerTypePass::lowerInternalGlobal(GlobalVariable *GV) {
|
|
|
- Type *NewTy = lowerType(GV->getType()->getPointerElementType());
|
|
|
- // So set init val to undef.
|
|
|
- Constant *InitVal = UndefValue::get(NewTy);
|
|
|
- if (GV->hasInitializer()) {
|
|
|
- Constant *OldInitVal = GV->getInitializer();
|
|
|
- if (isa<ConstantAggregateZero>(OldInitVal))
|
|
|
- InitVal = ConstantAggregateZero::get(NewTy);
|
|
|
- else if (!isa<UndefValue>(OldInitVal)) {
|
|
|
- InitVal = lowerInitVal(OldInitVal, NewTy);
|
|
|
- }
|
|
|
- }
|
|
|
-
|
|
|
- bool isConst = GV->isConstant();
|
|
|
- GlobalVariable::ThreadLocalMode TLMode = GV->getThreadLocalMode();
|
|
|
- unsigned AddressSpace = GV->getType()->getAddressSpace();
|
|
|
- GlobalValue::LinkageTypes linkage = GV->getLinkage();
|
|
|
-
|
|
|
- Module *M = GV->getParent();
|
|
|
- GlobalVariable *NewGV = new llvm::GlobalVariable(
|
|
|
- *M, NewTy, /*IsConstant*/ isConst, linkage,
|
|
|
- /*InitVal*/ InitVal, GV->getName() + getGlobalPrefix(),
|
|
|
- /*InsertBefore*/ nullptr, TLMode, AddressSpace);
|
|
|
- return NewGV;
|
|
|
-}
|
|
|
-
|
|
|
-bool LowerTypePass::runOnFunction(Function &F, bool HasDbgInfo) {
|
|
|
- std::vector<AllocaInst *> workList;
|
|
|
- // Scan the entry basic block, adding allocas to the worklist.
|
|
|
- BasicBlock &BB = F.getEntryBlock();
|
|
|
- for (BasicBlock::iterator I = BB.begin(), E = BB.end(); I != E; ++I) {
|
|
|
- if (!isa<AllocaInst>(I))
|
|
|
- continue;
|
|
|
- AllocaInst *A = cast<AllocaInst>(I);
|
|
|
- if (needToLower(A))
|
|
|
- workList.emplace_back(A);
|
|
|
- }
|
|
|
- LLVMContext &Context = F.getContext();
|
|
|
- for (AllocaInst *A : workList) {
|
|
|
- AllocaInst *NewA = lowerAlloca(A);
|
|
|
- if (HasDbgInfo) {
|
|
|
- // Add debug info.
|
|
|
- DbgDeclareInst *DDI = llvm::FindAllocaDbgDeclare(A);
|
|
|
- if (DDI) {
|
|
|
- Value *DDIVar = MetadataAsValue::get(Context, DDI->getRawVariable());
|
|
|
- Value *DDIExp = MetadataAsValue::get(Context, DDI->getRawExpression());
|
|
|
- Value *VMD = MetadataAsValue::get(Context, ValueAsMetadata::get(NewA));
|
|
|
- IRBuilder<> debugBuilder(DDI);
|
|
|
- debugBuilder.CreateCall(DDI->getCalledFunction(),
|
|
|
- {VMD, DDIVar, DDIExp});
|
|
|
- }
|
|
|
- }
|
|
|
- // Replace users.
|
|
|
- lowerUseWithNewValue(A, NewA);
|
|
|
- // Remove alloca.
|
|
|
- A->eraseFromParent();
|
|
|
- }
|
|
|
- return true;
|
|
|
-}
|
|
|
-
|
|
|
-bool LowerTypePass::runOnModule(Module &M) {
|
|
|
- initialize(M);
|
|
|
- // Load up debug information, to cross-reference values and the instructions
|
|
|
- // used to load them.
|
|
|
- bool HasDbgInfo = getDebugMetadataVersionFromModule(M) != 0;
|
|
|
- llvm::DebugInfoFinder Finder;
|
|
|
- if (HasDbgInfo) {
|
|
|
- Finder.processModule(M);
|
|
|
- }
|
|
|
-
|
|
|
- std::vector<AllocaInst*> multiDimAllocas;
|
|
|
- for (Function &F : M.functions()) {
|
|
|
- if (F.isDeclaration())
|
|
|
- continue;
|
|
|
- runOnFunction(F, HasDbgInfo);
|
|
|
- }
|
|
|
-
|
|
|
- // Work on internal global.
|
|
|
- std::vector<GlobalVariable *> vecGVs;
|
|
|
- for (GlobalVariable &GV : M.globals()) {
|
|
|
- if (dxilutil::IsStaticGlobal(&GV) || dxilutil::IsSharedMemoryGlobal(&GV)) {
|
|
|
- if (needToLower(&GV) && !GV.user_empty())
|
|
|
- vecGVs.emplace_back(&GV);
|
|
|
- }
|
|
|
- }
|
|
|
-
|
|
|
- for (GlobalVariable *GV : vecGVs) {
|
|
|
- GlobalVariable *NewGV = lowerInternalGlobal(GV);
|
|
|
- // Add debug info.
|
|
|
- if (HasDbgInfo) {
|
|
|
- HLModule::UpdateGlobalVariableDebugInfo(GV, Finder, NewGV);
|
|
|
- }
|
|
|
- // Replace users.
|
|
|
- lowerUseWithNewValue(GV, NewGV);
|
|
|
- // Remove GV.
|
|
|
- GV->removeDeadConstantUsers();
|
|
|
- GV->eraseFromParent();
|
|
|
- }
|
|
|
-
|
|
|
- return true;
|
|
|
-}
|
|
|
-
|
|
|
-}
|
|
|
-
|
|
|
-
|
|
|
-//===----------------------------------------------------------------------===//
|
|
|
-// DynamicIndexingVector to Array.
|
|
|
-//===----------------------------------------------------------------------===//
|
|
|
-
|
|
|
-namespace {
|
|
|
-class DynamicIndexingVectorToArray : public LowerTypePass {
|
|
|
- bool ReplaceAllVectors;
|
|
|
-public:
|
|
|
- explicit DynamicIndexingVectorToArray(bool ReplaceAll = false)
|
|
|
- : LowerTypePass(ID), ReplaceAllVectors(ReplaceAll) {}
|
|
|
- static char ID; // Pass identification, replacement for typeid
|
|
|
- void applyOptions(PassOptions O) override;
|
|
|
- void dumpConfig(raw_ostream &OS) override;
|
|
|
-protected:
|
|
|
- bool needToLower(Value *V) override;
|
|
|
- void lowerUseWithNewValue(Value *V, Value *NewV) override;
|
|
|
- Type *lowerType(Type *Ty) override;
|
|
|
- Constant *lowerInitVal(Constant *InitVal, Type *NewTy) override;
|
|
|
- StringRef getGlobalPrefix() override { return ".v"; }
|
|
|
-
|
|
|
-private:
|
|
|
- bool HasVectorDynamicIndexing(Value *V);
|
|
|
- void ReplaceVecGEP(Value *GEP, ArrayRef<Value *> idxList, Value *A,
|
|
|
- IRBuilder<> &Builder);
|
|
|
- void ReplaceVecArrayGEP(Value *GEP, ArrayRef<Value *> idxList, Value *A,
|
|
|
- IRBuilder<> &Builder);
|
|
|
- void ReplaceVectorWithArray(Value *Vec, Value *Array);
|
|
|
- void ReplaceVectorArrayWithArray(Value *VecArray, Value *Array);
|
|
|
- void ReplaceStaticIndexingOnVector(Value *V);
|
|
|
- void ReplaceAddrSpaceCast(ConstantExpr *CE,
|
|
|
- Value *A, IRBuilder<> &Builder);
|
|
|
-};
|
|
|
-
|
|
|
-void DynamicIndexingVectorToArray::applyOptions(PassOptions O) {
|
|
|
- GetPassOptionBool(O, "ReplaceAllVectors", &ReplaceAllVectors,
|
|
|
- ReplaceAllVectors);
|
|
|
-}
|
|
|
-void DynamicIndexingVectorToArray::dumpConfig(raw_ostream &OS) {
|
|
|
- ModulePass::dumpConfig(OS);
|
|
|
- OS << ",ReplaceAllVectors=" << ReplaceAllVectors;
|
|
|
-}
|
|
|
-
|
|
|
-void DynamicIndexingVectorToArray::ReplaceStaticIndexingOnVector(Value *V) {
|
|
|
- for (auto U = V->user_begin(), E = V->user_end(); U != E;) {
|
|
|
- Value *User = *(U++);
|
|
|
- if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(User)) {
|
|
|
- // Only work on element access for vector.
|
|
|
- if (GEP->getNumOperands() == 3) {
|
|
|
- auto Idx = GEP->idx_begin();
|
|
|
- // Skip the pointer idx.
|
|
|
- Idx++;
|
|
|
- ConstantInt *constIdx = cast<ConstantInt>(Idx);
|
|
|
-
|
|
|
- for (auto GEPU = GEP->user_begin(), GEPE = GEP->user_end();
|
|
|
- GEPU != GEPE;) {
|
|
|
- Instruction *GEPUser = cast<Instruction>(*(GEPU++));
|
|
|
-
|
|
|
- IRBuilder<> Builder(GEPUser);
|
|
|
-
|
|
|
- if (LoadInst *ldInst = dyn_cast<LoadInst>(GEPUser)) {
|
|
|
- // Change
|
|
|
- // ld a->x
|
|
|
- // into
|
|
|
- // b = ld a
|
|
|
- // b.x
|
|
|
- Value *ldVal = Builder.CreateLoad(V);
|
|
|
- Value *Elt = Builder.CreateExtractElement(ldVal, constIdx);
|
|
|
- ldInst->replaceAllUsesWith(Elt);
|
|
|
- ldInst->eraseFromParent();
|
|
|
- } else {
|
|
|
- // Change
|
|
|
- // st val, a->x
|
|
|
- // into
|
|
|
- // tmp = ld a
|
|
|
- // tmp.x = val
|
|
|
- // st tmp, a
|
|
|
- // Must be store inst here.
|
|
|
- StoreInst *stInst = cast<StoreInst>(GEPUser);
|
|
|
- Value *val = stInst->getValueOperand();
|
|
|
- Value *ldVal = Builder.CreateLoad(V);
|
|
|
- ldVal = Builder.CreateInsertElement(ldVal, val, constIdx);
|
|
|
- Builder.CreateStore(ldVal, V);
|
|
|
- stInst->eraseFromParent();
|
|
|
- }
|
|
|
- }
|
|
|
- GEP->eraseFromParent();
|
|
|
- } else if (GEP->getNumIndices() == 1) {
|
|
|
- Value *Idx = *GEP->idx_begin();
|
|
|
- if (ConstantInt *C = dyn_cast<ConstantInt>(Idx)) {
|
|
|
- if (C->getLimitedValue() == 0) {
|
|
|
- GEP->replaceAllUsesWith(V);
|
|
|
- GEP->eraseFromParent();
|
|
|
- }
|
|
|
- }
|
|
|
- }
|
|
|
- }
|
|
|
- }
|
|
|
-}
|
|
|
-
|
|
|
-bool DynamicIndexingVectorToArray::needToLower(Value *V) {
|
|
|
- Type *Ty = V->getType()->getPointerElementType();
|
|
|
- if (dyn_cast<VectorType>(Ty)) {
|
|
|
- if (isa<GlobalVariable>(V) || ReplaceAllVectors) {
|
|
|
- return true;
|
|
|
- }
|
|
|
- // Don't lower local vector which only static indexing.
|
|
|
- if (HasVectorDynamicIndexing(V)) {
|
|
|
- return true;
|
|
|
- } else {
|
|
|
- // Change vector indexing with ld st.
|
|
|
- ReplaceStaticIndexingOnVector(V);
|
|
|
- return false;
|
|
|
- }
|
|
|
- } else if (ArrayType *AT = dyn_cast<ArrayType>(Ty)) {
|
|
|
- // Array must be replaced even without dynamic indexing to remove vector
|
|
|
- // type in dxil.
|
|
|
- // TODO: optimize static array index in later pass.
|
|
|
- Type *EltTy = dxilutil::GetArrayEltTy(AT);
|
|
|
- return isa<VectorType>(EltTy);
|
|
|
- }
|
|
|
- return false;
|
|
|
-}
|
|
|
-
|
|
|
-void DynamicIndexingVectorToArray::ReplaceVecGEP(Value *GEP, ArrayRef<Value *> idxList,
|
|
|
- Value *A, IRBuilder<> &Builder) {
|
|
|
- Value *newGEP = Builder.CreateGEP(A, idxList);
|
|
|
- if (GEP->getType()->getPointerElementType()->isVectorTy()) {
|
|
|
- ReplaceVectorWithArray(GEP, newGEP);
|
|
|
- } else {
|
|
|
- GEP->replaceAllUsesWith(newGEP);
|
|
|
- }
|
|
|
-}
|
|
|
-
|
|
|
-void DynamicIndexingVectorToArray::ReplaceAddrSpaceCast(ConstantExpr *CE,
|
|
|
- Value *A, IRBuilder<> &Builder) {
|
|
|
- // create new AddrSpaceCast.
|
|
|
- Value *NewAddrSpaceCast = Builder.CreateAddrSpaceCast(
|
|
|
- A,
|
|
|
- PointerType::get(A->getType()->getPointerElementType(),
|
|
|
- CE->getType()->getPointerAddressSpace()));
|
|
|
- ReplaceVectorWithArray(CE, NewAddrSpaceCast);
|
|
|
-}
|
|
|
-
|
|
|
-void DynamicIndexingVectorToArray::ReplaceVectorWithArray(Value *Vec, Value *A) {
|
|
|
- unsigned size = Vec->getType()->getPointerElementType()->getVectorNumElements();
|
|
|
- for (auto U = Vec->user_begin(); U != Vec->user_end();) {
|
|
|
- User *User = (*U++);
|
|
|
-
|
|
|
- // GlobalVariable user.
|
|
|
- if (ConstantExpr * CE = dyn_cast<ConstantExpr>(User)) {
|
|
|
- if (User->user_empty())
|
|
|
- continue;
|
|
|
- if (GEPOperator *GEP = dyn_cast<GEPOperator>(User)) {
|
|
|
- IRBuilder<> Builder(Vec->getContext());
|
|
|
- SmallVector<Value *, 4> idxList(GEP->idx_begin(), GEP->idx_end());
|
|
|
- ReplaceVecGEP(GEP, idxList, A, Builder);
|
|
|
- continue;
|
|
|
- } else if (CE->getOpcode() == Instruction::AddrSpaceCast) {
|
|
|
- IRBuilder<> Builder(Vec->getContext());
|
|
|
- ReplaceAddrSpaceCast(CE, A, Builder);
|
|
|
- continue;
|
|
|
- }
|
|
|
- DXASSERT(0, "not implemented yet");
|
|
|
- }
|
|
|
- // Instrution user.
|
|
|
- Instruction *UserInst = cast<Instruction>(User);
|
|
|
- IRBuilder<> Builder(UserInst);
|
|
|
- if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(User)) {
|
|
|
- SmallVector<Value *, 4> idxList(GEP->idx_begin(), GEP->idx_end());
|
|
|
- ReplaceVecGEP(cast<GEPOperator>(GEP), idxList, A, Builder);
|
|
|
- GEP->eraseFromParent();
|
|
|
- } else if (LoadInst *ldInst = dyn_cast<LoadInst>(User)) {
|
|
|
- // If ld whole struct, need to split the load.
|
|
|
- Value *newLd = UndefValue::get(ldInst->getType());
|
|
|
- Value *zero = Builder.getInt32(0);
|
|
|
- for (unsigned i = 0; i < size; i++) {
|
|
|
- Value *idx = Builder.getInt32(i);
|
|
|
- Value *GEP = Builder.CreateInBoundsGEP(A, {zero, idx});
|
|
|
- Value *Elt = Builder.CreateLoad(GEP);
|
|
|
- newLd = Builder.CreateInsertElement(newLd, Elt, i);
|
|
|
- }
|
|
|
- ldInst->replaceAllUsesWith(newLd);
|
|
|
- ldInst->eraseFromParent();
|
|
|
- } else if (StoreInst *stInst = dyn_cast<StoreInst>(User)) {
|
|
|
- Value *val = stInst->getValueOperand();
|
|
|
- Value *zero = Builder.getInt32(0);
|
|
|
- for (unsigned i = 0; i < size; i++) {
|
|
|
- Value *Elt = Builder.CreateExtractElement(val, i);
|
|
|
- Value *idx = Builder.getInt32(i);
|
|
|
- Value *GEP = Builder.CreateInBoundsGEP(A, {zero, idx});
|
|
|
- Builder.CreateStore(Elt, GEP);
|
|
|
- }
|
|
|
- stInst->eraseFromParent();
|
|
|
- } else {
|
|
|
- // Vector parameter should be lowered.
|
|
|
- // No function call should use vector.
|
|
|
- DXASSERT(0, "not implement yet");
|
|
|
- }
|
|
|
- }
|
|
|
-}
|
|
|
-
|
|
|
-void DynamicIndexingVectorToArray::ReplaceVecArrayGEP(Value *GEP,
|
|
|
- ArrayRef<Value *> idxList, Value *A,
|
|
|
- IRBuilder<> &Builder) {
|
|
|
- Value *newGEP = Builder.CreateGEP(A, idxList);
|
|
|
- Type *Ty = GEP->getType()->getPointerElementType();
|
|
|
- if (Ty->isVectorTy()) {
|
|
|
- ReplaceVectorWithArray(GEP, newGEP);
|
|
|
- } else if (Ty->isArrayTy()) {
|
|
|
- ReplaceVectorArrayWithArray(GEP, newGEP);
|
|
|
- } else {
|
|
|
- DXASSERT(Ty->isSingleValueType(), "must be vector subscript here");
|
|
|
- GEP->replaceAllUsesWith(newGEP);
|
|
|
- }
|
|
|
-}
|
|
|
-
|
|
|
-void DynamicIndexingVectorToArray::ReplaceVectorArrayWithArray(Value *VA, Value *A) {
|
|
|
- for (auto U = VA->user_begin(); U != VA->user_end();) {
|
|
|
- User *User = *(U++);
|
|
|
- if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(User)) {
|
|
|
- IRBuilder<> Builder(GEP);
|
|
|
- SmallVector<Value *, 4> idxList(GEP->idx_begin(), GEP->idx_end());
|
|
|
- ReplaceVecArrayGEP(GEP, idxList, A, Builder);
|
|
|
- GEP->eraseFromParent();
|
|
|
- } else if (GEPOperator *GEPOp = dyn_cast<GEPOperator>(User)) {
|
|
|
- IRBuilder<> Builder(GEPOp->getContext());
|
|
|
- SmallVector<Value *, 4> idxList(GEPOp->idx_begin(), GEPOp->idx_end());
|
|
|
- ReplaceVecArrayGEP(GEPOp, idxList, A, Builder);
|
|
|
- } else if (BitCastInst *BCI = dyn_cast<BitCastInst>(User)) {
|
|
|
- BCI->setOperand(0, A);
|
|
|
- } else {
|
|
|
- DXASSERT(0, "Array pointer should only used by GEP");
|
|
|
- }
|
|
|
- }
|
|
|
-}
|
|
|
-
|
|
|
-void DynamicIndexingVectorToArray::lowerUseWithNewValue(Value *V, Value *NewV) {
|
|
|
- Type *Ty = V->getType()->getPointerElementType();
|
|
|
- // Replace V with NewV.
|
|
|
- if (Ty->isVectorTy()) {
|
|
|
- ReplaceVectorWithArray(V, NewV);
|
|
|
- } else {
|
|
|
- ReplaceVectorArrayWithArray(V, NewV);
|
|
|
- }
|
|
|
-}
|
|
|
-
|
|
|
-Type *DynamicIndexingVectorToArray::lowerType(Type *Ty) {
|
|
|
- if (VectorType *VT = dyn_cast<VectorType>(Ty)) {
|
|
|
- return ArrayType::get(VT->getElementType(), VT->getNumElements());
|
|
|
- } else if (ArrayType *AT = dyn_cast<ArrayType>(Ty)) {
|
|
|
- SmallVector<ArrayType *, 4> nestArrayTys;
|
|
|
- nestArrayTys.emplace_back(AT);
|
|
|
-
|
|
|
- Type *EltTy = AT->getElementType();
|
|
|
- // support multi level of array
|
|
|
- while (EltTy->isArrayTy()) {
|
|
|
- ArrayType *ElAT = cast<ArrayType>(EltTy);
|
|
|
- nestArrayTys.emplace_back(ElAT);
|
|
|
- EltTy = ElAT->getElementType();
|
|
|
- }
|
|
|
- if (EltTy->isVectorTy()) {
|
|
|
- Type *vecAT = ArrayType::get(EltTy->getVectorElementType(),
|
|
|
- EltTy->getVectorNumElements());
|
|
|
- return CreateNestArrayTy(vecAT, nestArrayTys);
|
|
|
- }
|
|
|
- return nullptr;
|
|
|
- }
|
|
|
- return nullptr;
|
|
|
-}
|
|
|
-
|
|
|
-Constant *DynamicIndexingVectorToArray::lowerInitVal(Constant *InitVal, Type *NewTy) {
|
|
|
- Type *VecTy = InitVal->getType();
|
|
|
- ArrayType *ArrayTy = cast<ArrayType>(NewTy);
|
|
|
- if (VecTy->isVectorTy()) {
|
|
|
- SmallVector<Constant *, 4> Elts;
|
|
|
- for (unsigned i = 0; i < VecTy->getVectorNumElements(); i++) {
|
|
|
- Elts.emplace_back(InitVal->getAggregateElement(i));
|
|
|
- }
|
|
|
- return ConstantArray::get(ArrayTy, Elts);
|
|
|
- } else {
|
|
|
- ArrayType *AT = cast<ArrayType>(VecTy);
|
|
|
- ArrayType *EltArrayTy = cast<ArrayType>(ArrayTy->getElementType());
|
|
|
- SmallVector<Constant *, 4> Elts;
|
|
|
- for (unsigned i = 0; i < AT->getNumElements(); i++) {
|
|
|
- Constant *Elt = lowerInitVal(InitVal->getAggregateElement(i), EltArrayTy);
|
|
|
- Elts.emplace_back(Elt);
|
|
|
- }
|
|
|
- return ConstantArray::get(ArrayTy, Elts);
|
|
|
- }
|
|
|
-}
|
|
|
-
|
|
|
-bool DynamicIndexingVectorToArray::HasVectorDynamicIndexing(Value *V) {
|
|
|
- return dxilutil::HasDynamicIndexing(V);
|
|
|
-}
|
|
|
-
|
|
|
-}
|
|
|
-
|
|
|
-char DynamicIndexingVectorToArray::ID = 0;
|
|
|
-
|
|
|
-INITIALIZE_PASS(DynamicIndexingVectorToArray, "dynamic-vector-to-array",
|
|
|
- "Replace dynamic indexing vector with array", false,
|
|
|
- false)
|
|
|
-
|
|
|
-// Public interface to the DynamicIndexingVectorToArray pass
|
|
|
-ModulePass *llvm::createDynamicIndexingVectorToArrayPass(bool ReplaceAllVector) {
|
|
|
- return new DynamicIndexingVectorToArray(ReplaceAllVector);
|
|
|
-}
|
|
|
-
|
|
|
-//===----------------------------------------------------------------------===//
|
|
|
-// Flatten multi dim array into 1 dim.
|
|
|
-//===----------------------------------------------------------------------===//
|
|
|
-
|
|
|
-namespace {
|
|
|
-
|
|
|
-class MultiDimArrayToOneDimArray : public LowerTypePass {
|
|
|
-public:
|
|
|
- explicit MultiDimArrayToOneDimArray() : LowerTypePass(ID) {}
|
|
|
- static char ID; // Pass identification, replacement for typeid
|
|
|
-protected:
|
|
|
- bool needToLower(Value *V) override;
|
|
|
- void lowerUseWithNewValue(Value *V, Value *NewV) override;
|
|
|
- Type *lowerType(Type *Ty) override;
|
|
|
- Constant *lowerInitVal(Constant *InitVal, Type *NewTy) override;
|
|
|
- StringRef getGlobalPrefix() override { return ".1dim"; }
|
|
|
-};
|
|
|
-
|
|
|
-bool MultiDimArrayToOneDimArray::needToLower(Value *V) {
|
|
|
- Type *Ty = V->getType()->getPointerElementType();
|
|
|
- ArrayType *AT = dyn_cast<ArrayType>(Ty);
|
|
|
- if (!AT)
|
|
|
- return false;
|
|
|
- if (!isa<ArrayType>(AT->getElementType())) {
|
|
|
- return false;
|
|
|
- } else {
|
|
|
- // Merge all GEP.
|
|
|
- HLModule::MergeGepUse(V);
|
|
|
- return true;
|
|
|
- }
|
|
|
-}
|
|
|
-
|
|
|
-void ReplaceMultiDimGEP(User *GEP, Value *OneDim, IRBuilder<> &Builder) {
|
|
|
- gep_type_iterator GEPIt = gep_type_begin(GEP), E = gep_type_end(GEP);
|
|
|
-
|
|
|
- Value *PtrOffset = GEPIt.getOperand();
|
|
|
- ++GEPIt;
|
|
|
- Value *ArrayIdx = GEPIt.getOperand();
|
|
|
- ++GEPIt;
|
|
|
- Value *VecIdx = nullptr;
|
|
|
- for (; GEPIt != E; ++GEPIt) {
|
|
|
- if (GEPIt->isArrayTy()) {
|
|
|
- unsigned arraySize = GEPIt->getArrayNumElements();
|
|
|
- Value *V = GEPIt.getOperand();
|
|
|
- ArrayIdx = Builder.CreateMul(ArrayIdx, Builder.getInt32(arraySize));
|
|
|
- ArrayIdx = Builder.CreateAdd(V, ArrayIdx);
|
|
|
- } else {
|
|
|
- DXASSERT_NOMSG(isa<VectorType>(*GEPIt));
|
|
|
- VecIdx = GEPIt.getOperand();
|
|
|
- }
|
|
|
- }
|
|
|
- Value *NewGEP = nullptr;
|
|
|
- if (!VecIdx)
|
|
|
- NewGEP = Builder.CreateGEP(OneDim, {PtrOffset, ArrayIdx});
|
|
|
- else
|
|
|
- NewGEP = Builder.CreateGEP(OneDim, {PtrOffset, ArrayIdx, VecIdx});
|
|
|
-
|
|
|
- GEP->replaceAllUsesWith(NewGEP);
|
|
|
-}
|
|
|
-
|
|
|
-void MultiDimArrayToOneDimArray::lowerUseWithNewValue(Value *MultiDim, Value *OneDim) {
|
|
|
- LLVMContext &Context = MultiDim->getContext();
|
|
|
- // All users should be element type.
|
|
|
- // Replace users of AI or GV.
|
|
|
- for (auto it = MultiDim->user_begin(); it != MultiDim->user_end();) {
|
|
|
- User *U = *(it++);
|
|
|
- if (U->user_empty())
|
|
|
- continue;
|
|
|
- if (BitCastInst *BCI = dyn_cast<BitCastInst>(U)) {
|
|
|
- BCI->setOperand(0, OneDim);
|
|
|
- continue;
|
|
|
- }
|
|
|
-
|
|
|
- if (ConstantExpr *CE = dyn_cast<ConstantExpr>(U)) {
|
|
|
- IRBuilder<> Builder(Context);
|
|
|
- if (GEPOperator *GEP = dyn_cast<GEPOperator>(U)) {
|
|
|
- // NewGEP must be GEPOperator too.
|
|
|
- // No instruction will be build.
|
|
|
- ReplaceMultiDimGEP(U, OneDim, Builder);
|
|
|
- } else if (CE->getOpcode() == Instruction::AddrSpaceCast) {
|
|
|
- Value *NewAddrSpaceCast = Builder.CreateAddrSpaceCast(
|
|
|
- OneDim,
|
|
|
- PointerType::get(OneDim->getType()->getPointerElementType(),
|
|
|
- CE->getType()->getPointerAddressSpace()));
|
|
|
- lowerUseWithNewValue(CE, NewAddrSpaceCast);
|
|
|
- } else {
|
|
|
- DXASSERT(0, "not implemented");
|
|
|
- }
|
|
|
- } else if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(U)) {
|
|
|
- IRBuilder<> Builder(GEP);
|
|
|
- ReplaceMultiDimGEP(U, OneDim, Builder);
|
|
|
- GEP->eraseFromParent();
|
|
|
- } else {
|
|
|
- DXASSERT(0, "not implemented");
|
|
|
- }
|
|
|
- }
|
|
|
-}
|
|
|
-
|
|
|
-Type *MultiDimArrayToOneDimArray::lowerType(Type *Ty) {
|
|
|
- ArrayType *AT = cast<ArrayType>(Ty);
|
|
|
- unsigned arraySize = AT->getNumElements();
|
|
|
-
|
|
|
- Type *EltTy = AT->getElementType();
|
|
|
- // support multi level of array
|
|
|
- while (EltTy->isArrayTy()) {
|
|
|
- ArrayType *ElAT = cast<ArrayType>(EltTy);
|
|
|
- arraySize *= ElAT->getNumElements();
|
|
|
- EltTy = ElAT->getElementType();
|
|
|
- }
|
|
|
-
|
|
|
- return ArrayType::get(EltTy, arraySize);
|
|
|
-}
|
|
|
-
|
|
|
-void FlattenMultiDimConstArray(Constant *V, std::vector<Constant *> &Elts) {
|
|
|
- if (!V->getType()->isArrayTy()) {
|
|
|
- Elts.emplace_back(V);
|
|
|
- } else {
|
|
|
- ArrayType *AT = cast<ArrayType>(V->getType());
|
|
|
- for (unsigned i = 0; i < AT->getNumElements(); i++) {
|
|
|
- FlattenMultiDimConstArray(V->getAggregateElement(i), Elts);
|
|
|
- }
|
|
|
- }
|
|
|
-}
|
|
|
-
|
|
|
-Constant *MultiDimArrayToOneDimArray::lowerInitVal(Constant *InitVal, Type *NewTy) {
|
|
|
- if (InitVal) {
|
|
|
- // MultiDim array init should be done by store.
|
|
|
- if (isa<ConstantAggregateZero>(InitVal))
|
|
|
- InitVal = ConstantAggregateZero::get(NewTy);
|
|
|
- else if (isa<UndefValue>(InitVal))
|
|
|
- InitVal = UndefValue::get(NewTy);
|
|
|
- else {
|
|
|
- std::vector<Constant *> Elts;
|
|
|
- FlattenMultiDimConstArray(InitVal, Elts);
|
|
|
- InitVal = ConstantArray::get(cast<ArrayType>(NewTy), Elts);
|
|
|
- }
|
|
|
- } else {
|
|
|
- InitVal = UndefValue::get(NewTy);
|
|
|
- }
|
|
|
- return InitVal;
|
|
|
-}
|
|
|
-
|
|
|
-}
|
|
|
-
|
|
|
-char MultiDimArrayToOneDimArray::ID = 0;
|
|
|
-
|
|
|
-INITIALIZE_PASS(MultiDimArrayToOneDimArray, "multi-dim-one-dim",
|
|
|
- "Flatten multi-dim array into one-dim array", false,
|
|
|
- false)
|
|
|
-
|
|
|
-// Public interface to the SROA_Parameter_HLSL pass
|
|
|
-ModulePass *llvm::createMultiDimArrayToOneDimArrayPass() {
|
|
|
- return new MultiDimArrayToOneDimArray();
|
|
|
-}
|
|
|
-
|
|
|
-//===----------------------------------------------------------------------===//
|
|
|
-// Lower resource into handle.
|
|
|
-//===----------------------------------------------------------------------===//
|
|
|
-
|
|
|
-namespace {
|
|
|
-
|
|
|
-class ResourceToHandle : public LowerTypePass {
|
|
|
-public:
|
|
|
- explicit ResourceToHandle() : LowerTypePass(ID) {}
|
|
|
- static char ID; // Pass identification, replacement for typeid
|
|
|
-protected:
|
|
|
- bool needToLower(Value *V) override;
|
|
|
- void lowerUseWithNewValue(Value *V, Value *NewV) override;
|
|
|
- Type *lowerType(Type *Ty) override;
|
|
|
- Constant *lowerInitVal(Constant *InitVal, Type *NewTy) override;
|
|
|
- StringRef getGlobalPrefix() override { return ".res"; }
|
|
|
- void initialize(Module &M) override;
|
|
|
-private:
|
|
|
- void ReplaceResourceWithHandle(Value *ResPtr, Value *HandlePtr);
|
|
|
- void ReplaceResourceGEPWithHandleGEP(Value *GEP, ArrayRef<Value *> idxList,
|
|
|
- Value *A, IRBuilder<> &Builder);
|
|
|
- void ReplaceResourceArrayWithHandleArray(Value *VA, Value *A);
|
|
|
-
|
|
|
- Type *m_HandleTy;
|
|
|
- HLModule *m_pHLM;
|
|
|
- bool m_bIsLib;
|
|
|
-};
|
|
|
-
|
|
|
-void ResourceToHandle::initialize(Module &M) {
|
|
|
- DXASSERT(M.HasHLModule(), "require HLModule");
|
|
|
- m_pHLM = &M.GetHLModule();
|
|
|
- m_HandleTy = m_pHLM->GetOP()->GetHandleType();
|
|
|
- m_bIsLib = m_pHLM->GetShaderModel()->IsLib();
|
|
|
-}
|
|
|
-
|
|
|
-bool ResourceToHandle::needToLower(Value *V) {
|
|
|
- Type *Ty = V->getType()->getPointerElementType();
|
|
|
- Ty = dxilutil::GetArrayEltTy(Ty);
|
|
|
- return (dxilutil::IsHLSLObjectType(Ty) &&
|
|
|
- !HLModule::IsStreamOutputType(Ty)) &&
|
|
|
- // Skip lib profile.
|
|
|
- !m_bIsLib;
|
|
|
-}
|
|
|
-
|
|
|
-Type *ResourceToHandle::lowerType(Type *Ty) {
|
|
|
- if ((dxilutil::IsHLSLObjectType(Ty) && !HLModule::IsStreamOutputType(Ty))) {
|
|
|
- return m_HandleTy;
|
|
|
- }
|
|
|
-
|
|
|
- ArrayType *AT = cast<ArrayType>(Ty);
|
|
|
-
|
|
|
- SmallVector<ArrayType *, 4> nestArrayTys;
|
|
|
- nestArrayTys.emplace_back(AT);
|
|
|
-
|
|
|
- Type *EltTy = AT->getElementType();
|
|
|
- // support multi level of array
|
|
|
- while (EltTy->isArrayTy()) {
|
|
|
- ArrayType *ElAT = cast<ArrayType>(EltTy);
|
|
|
- nestArrayTys.emplace_back(ElAT);
|
|
|
- EltTy = ElAT->getElementType();
|
|
|
- }
|
|
|
-
|
|
|
- return CreateNestArrayTy(m_HandleTy, nestArrayTys);
|
|
|
-}
|
|
|
-
|
|
|
-Constant *ResourceToHandle::lowerInitVal(Constant *InitVal, Type *NewTy) {
|
|
|
- DXASSERT(isa<UndefValue>(InitVal), "resource cannot have real init val");
|
|
|
- return UndefValue::get(NewTy);
|
|
|
-}
|
|
|
-
|
|
|
-void ResourceToHandle::ReplaceResourceWithHandle(Value *ResPtr,
|
|
|
- Value *HandlePtr) {
|
|
|
- for (auto it = ResPtr->user_begin(); it != ResPtr->user_end();) {
|
|
|
- User *U = *(it++);
|
|
|
- if (LoadInst *LI = dyn_cast<LoadInst>(U)) {
|
|
|
- IRBuilder<> Builder(LI);
|
|
|
- Value *Handle = Builder.CreateLoad(HandlePtr);
|
|
|
- Type *ResTy = LI->getType();
|
|
|
- // Used by createHandle or Store.
|
|
|
- for (auto ldIt = LI->user_begin(); ldIt != LI->user_end();) {
|
|
|
- User *ldU = *(ldIt++);
|
|
|
- if (StoreInst *SI = dyn_cast<StoreInst>(ldU)) {
|
|
|
- Value *TmpRes = HLModule::EmitHLOperationCall(
|
|
|
- Builder, HLOpcodeGroup::HLCast,
|
|
|
- (unsigned)HLCastOpcode::HandleToResCast, ResTy, {Handle},
|
|
|
- *m_pHLM->GetModule());
|
|
|
- SI->replaceUsesOfWith(LI, TmpRes);
|
|
|
- } else {
|
|
|
- CallInst *CI = cast<CallInst>(ldU);
|
|
|
- DXASSERT(hlsl::GetHLOpcodeGroupByName(CI->getCalledFunction()) == HLOpcodeGroup::HLCreateHandle,
|
|
|
- "must be createHandle");
|
|
|
- CI->replaceAllUsesWith(Handle);
|
|
|
- CI->eraseFromParent();
|
|
|
- }
|
|
|
- }
|
|
|
- LI->eraseFromParent();
|
|
|
- } else if (StoreInst *SI = dyn_cast<StoreInst>(U)) {
|
|
|
- Value *Res = SI->getValueOperand();
|
|
|
- IRBuilder<> Builder(SI);
|
|
|
- // CreateHandle from Res.
|
|
|
- Value *Handle = HLModule::EmitHLOperationCall(
|
|
|
- Builder, HLOpcodeGroup::HLCreateHandle,
|
|
|
- /*opcode*/ 0, m_HandleTy, {Res}, *m_pHLM->GetModule());
|
|
|
- // Store Handle to HandlePtr.
|
|
|
- Builder.CreateStore(Handle, HandlePtr);
|
|
|
- // Remove resource Store.
|
|
|
- SI->eraseFromParent();
|
|
|
- } else if (U->user_empty() && isa<GEPOperator>(U)) {
|
|
|
- continue;
|
|
|
- } else {
|
|
|
- CallInst *CI = cast<CallInst>(U);
|
|
|
- IRBuilder<> Builder(CI);
|
|
|
- HLOpcodeGroup group = GetHLOpcodeGroupByName(CI->getCalledFunction());
|
|
|
- // Allow user function to use res ptr as argument.
|
|
|
- if (group == HLOpcodeGroup::NotHL) {
|
|
|
- Value *TmpResPtr = Builder.CreateBitCast(HandlePtr, ResPtr->getType());
|
|
|
- CI->replaceUsesOfWith(ResPtr, TmpResPtr);
|
|
|
- } else {
|
|
|
- DXASSERT(0, "invalid operation on resource");
|
|
|
- }
|
|
|
- }
|
|
|
- }
|
|
|
-}
|
|
|
-
|
|
|
-void ResourceToHandle::ReplaceResourceGEPWithHandleGEP(
|
|
|
- Value *GEP, ArrayRef<Value *> idxList, Value *A, IRBuilder<> &Builder) {
|
|
|
- Value *newGEP = Builder.CreateGEP(A, idxList);
|
|
|
- Type *Ty = GEP->getType()->getPointerElementType();
|
|
|
- if (Ty->isArrayTy()) {
|
|
|
- ReplaceResourceArrayWithHandleArray(GEP, newGEP);
|
|
|
- } else {
|
|
|
- DXASSERT(dxilutil::IsHLSLObjectType(Ty), "must be resource type here");
|
|
|
- ReplaceResourceWithHandle(GEP, newGEP);
|
|
|
- }
|
|
|
-}
|
|
|
-
|
|
|
-void ResourceToHandle::ReplaceResourceArrayWithHandleArray(Value *VA,
|
|
|
- Value *A) {
|
|
|
- for (auto U = VA->user_begin(); U != VA->user_end();) {
|
|
|
- User *User = *(U++);
|
|
|
- if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(User)) {
|
|
|
- IRBuilder<> Builder(GEP);
|
|
|
- SmallVector<Value *, 4> idxList(GEP->idx_begin(), GEP->idx_end());
|
|
|
- ReplaceResourceGEPWithHandleGEP(GEP, idxList, A, Builder);
|
|
|
- GEP->eraseFromParent();
|
|
|
- } else if (GEPOperator *GEPOp = dyn_cast<GEPOperator>(User)) {
|
|
|
- IRBuilder<> Builder(GEPOp->getContext());
|
|
|
- SmallVector<Value *, 4> idxList(GEPOp->idx_begin(), GEPOp->idx_end());
|
|
|
- ReplaceResourceGEPWithHandleGEP(GEPOp, idxList, A, Builder);
|
|
|
- } else {
|
|
|
- DXASSERT(0, "Array pointer should only used by GEP");
|
|
|
- }
|
|
|
- }
|
|
|
-}
|
|
|
-
|
|
|
-void ResourceToHandle::lowerUseWithNewValue(Value *V, Value *NewV) {
|
|
|
- Type *Ty = V->getType()->getPointerElementType();
|
|
|
- // Replace V with NewV.
|
|
|
- if (Ty->isArrayTy()) {
|
|
|
- ReplaceResourceArrayWithHandleArray(V, NewV);
|
|
|
- } else {
|
|
|
- ReplaceResourceWithHandle(V, NewV);
|
|
|
- }
|
|
|
-}
|
|
|
-
|
|
|
-}
|
|
|
-
|
|
|
-char ResourceToHandle::ID = 0;
|
|
|
-
|
|
|
-INITIALIZE_PASS(ResourceToHandle, "resource-handle",
|
|
|
- "Lower resource into handle", false,
|
|
|
- false)
|
|
|
-
|
|
|
-// Public interface to the ResourceToHandle pass
|
|
|
-ModulePass *llvm::createResourceToHandlePass() {
|
|
|
- return new ResourceToHandle();
|
|
|
-}
|
Tristan Labelle