|
@@ -0,0 +1,553 @@
|
|
|
+//===-- DxilConstantFolding.cpp - Fold dxil intrinsics into constants -----===//
|
|
|
+//
|
|
|
+// The LLVM Compiler Infrastructure
|
|
|
+//
|
|
|
+// This file is distributed under the University of Illinois Open Source
|
|
|
+// License. See LICENSE.TXT for details.
|
|
|
+//
|
|
|
+// Copyright (C) Microsoft Corporation. All rights reserved.
|
|
|
+//
|
|
|
+//===----------------------------------------------------------------------===//
|
|
|
+//
|
|
|
+//
|
|
|
+//===----------------------------------------------------------------------===//
|
|
|
+#include "llvm/Analysis/DxilConstantFolding.h"
|
|
|
+#include "llvm/Analysis/ConstantFolding.h"
|
|
|
+#include "llvm/ADT/SmallPtrSet.h"
|
|
|
+#include "llvm/ADT/SmallVector.h"
|
|
|
+#include "llvm/ADT/StringMap.h"
|
|
|
+#include "llvm/Analysis/TargetLibraryInfo.h"
|
|
|
+#include "llvm/Analysis/ValueTracking.h"
|
|
|
+#include "llvm/Config/config.h"
|
|
|
+#include "llvm/IR/Constants.h"
|
|
|
+#include "llvm/IR/DataLayout.h"
|
|
|
+#include "llvm/IR/DerivedTypes.h"
|
|
|
+#include "llvm/IR/Function.h"
|
|
|
+#include "llvm/IR/GetElementPtrTypeIterator.h"
|
|
|
+#include "llvm/IR/GlobalVariable.h"
|
|
|
+#include "llvm/IR/Instructions.h"
|
|
|
+#include "llvm/IR/Intrinsics.h"
|
|
|
+#include "llvm/IR/Operator.h"
|
|
|
+#include "llvm/Support/ErrorHandling.h"
|
|
|
+#include "llvm/Support/MathExtras.h"
|
|
|
+#include <cerrno>
|
|
|
+#include <cmath>
|
|
|
+#include <algorithm>
|
|
|
+#include <functional>
|
|
|
+
|
|
|
+#include "dxc/HLSL/Dxil.h"
|
|
|
+
|
|
|
+using namespace llvm;
|
|
|
+using namespace hlsl;
|
|
|
+
|
|
|
+// Check if the given function is a dxil intrinsic and if so extract the
|
|
|
+// opcode for the instrinsic being called.
|
|
|
+static bool GetDxilOpcode(StringRef Name, ArrayRef<Constant *> Operands, OP::OpCode &out) {
|
|
|
+ if (!OP::IsDxilOpFuncName(Name))
|
|
|
+ return false;
|
|
|
+ if (!Operands.size())
|
|
|
+ return false;
|
|
|
+ if (ConstantInt *ci = dyn_cast<ConstantInt>(Operands[0])) {
|
|
|
+ uint64_t opcode = ci->getLimitedValue();
|
|
|
+ if (opcode < static_cast<uint64_t>(OP::OpCode::NumOpCodes)) {
|
|
|
+ out = static_cast<OP::OpCode>(opcode);
|
|
|
+ return true;
|
|
|
+ }
|
|
|
+ }
|
|
|
+
|
|
|
+ return false;
|
|
|
+}
|
|
|
+
|
|
|
+// Typedefs for passing function pointers to evaluate float constants.
|
|
|
+typedef double(__cdecl *NativeFPUnaryOp)(double);
|
|
|
+typedef std::function<APFloat::opStatus(APFloat&)> APFloatUnaryOp;
|
|
|
+
|
|
|
+/// Currently APFloat versions of these functions do not exist, so we use
|
|
|
+/// the host native double versions. Float versions are not called
|
|
|
+/// directly but for all these it is true (float)(f((double)arg)) ==
|
|
|
+/// f(arg). Long double not supported yet.
|
|
|
+///
|
|
|
+/// Calls out to the llvm constant folding function to do the real work.
|
|
|
+static Constant *DxilConstantFoldFP(NativeFPUnaryOp NativeFP, ConstantFP *C, Type *Ty) {
|
|
|
+ double V = llvm::getValueAsDouble(C);
|
|
|
+ return llvm::ConstantFoldFP(NativeFP, V, Ty);
|
|
|
+}
|
|
|
+
|
|
|
+// Constant fold using the provided function on APFloats.
|
|
|
+static Constant *HLSLConstantFoldAPFloat(APFloatUnaryOp NativeFP, ConstantFP *C, Type *Ty) {
|
|
|
+ APFloat APF = C->getValueAPF();
|
|
|
+
|
|
|
+ if (NativeFP(APF) != APFloat::opStatus::opOK)
|
|
|
+ return nullptr;
|
|
|
+
|
|
|
+ return ConstantFP::get(Ty->getContext(), APF);
|
|
|
+}
|
|
|
+
|
|
|
+// Constant fold a round dxil intrinsic.
|
|
|
+static Constant *HLSLConstantFoldRound(APFloat::roundingMode roundingMode, ConstantFP *C, Type *Ty) {
|
|
|
+ APFloatUnaryOp f = [roundingMode](APFloat &x) { return x.roundToIntegral(roundingMode); };
|
|
|
+ return HLSLConstantFoldAPFloat(f, C, Ty);
|
|
|
+}
|
|
|
+
|
|
|
+namespace {
|
|
|
+// Wrapper for call operands that "shifts past" the hlsl intrinsic opcode.
|
|
|
+// Also provides accessors that dyn_cast the operand to a constant type.
|
|
|
+class DxilIntrinsicOperands {
|
|
|
+public:
|
|
|
+ DxilIntrinsicOperands(ArrayRef<Constant *> RawCallOperands) : m_RawCallOperands(RawCallOperands) {}
|
|
|
+ Constant * const &operator[](size_t index) const {
|
|
|
+ return m_RawCallOperands[index + 1];
|
|
|
+ }
|
|
|
+
|
|
|
+ ConstantInt *GetConstantInt(size_t index) const {
|
|
|
+ return dyn_cast<ConstantInt>(this->operator[](index));
|
|
|
+ }
|
|
|
+
|
|
|
+ ConstantFP *GetConstantFloat(size_t index) const {
|
|
|
+ return dyn_cast<ConstantFP>(this->operator[](index));
|
|
|
+ }
|
|
|
+
|
|
|
+ size_t Size() const {
|
|
|
+ return m_RawCallOperands.size() - 1;
|
|
|
+ }
|
|
|
+private:
|
|
|
+ ArrayRef<Constant *> m_RawCallOperands;
|
|
|
+};
|
|
|
+}
|
|
|
+
|
|
|
+/// We only fold functions with finite arguments. Folding NaN and inf is
|
|
|
+/// likely to be aborted with an exception anyway, and some host libms
|
|
|
+/// have known errors raising exceptions.
|
|
|
+static bool IsFinite(ConstantFP *C) {
|
|
|
+ if (C->getValueAPF().isNaN() || C->getValueAPF().isInfinity())
|
|
|
+ return false;
|
|
|
+
|
|
|
+ return true;
|
|
|
+}
|
|
|
+
|
|
|
+// Check that the op is non-null and finite.
|
|
|
+static bool IsValidOp(ConstantFP *C) {
|
|
|
+ if (!C || !IsFinite(C))
|
|
|
+ return false;
|
|
|
+
|
|
|
+ return true;
|
|
|
+}
|
|
|
+
|
|
|
+// Check that all ops are valid.
|
|
|
+static bool AllValidOps(ArrayRef<ConstantFP *> Ops) {
|
|
|
+ return std::all_of(Ops.begin(), Ops.end(), IsValidOp);
|
|
|
+}
|
|
|
+
|
|
|
+// Constant fold unary floating point intrinsics.
|
|
|
+static Constant *ConstantFoldUnaryFPIntrinsic(OP::OpCode opcode, Type *Ty, ConstantFP *Op) {
|
|
|
+ switch (opcode) {
|
|
|
+ default: break;
|
|
|
+ case OP::OpCode::FAbs: return DxilConstantFoldFP(fabs, Op, Ty);
|
|
|
+ case OP::OpCode::Saturate: {
|
|
|
+ NativeFPUnaryOp f = [](double x) { return std::max(std::min(x, 1.0), 0.0); };
|
|
|
+ return DxilConstantFoldFP(f, Op, Ty);
|
|
|
+ }
|
|
|
+ case OP::OpCode::Cos: return DxilConstantFoldFP(cos, Op, Ty);
|
|
|
+ case OP::OpCode::Sin: return DxilConstantFoldFP(sin, Op, Ty);
|
|
|
+ case OP::OpCode::Tan: return DxilConstantFoldFP(tan, Op, Ty);
|
|
|
+ case OP::OpCode::Acos: return DxilConstantFoldFP(acos, Op, Ty);
|
|
|
+ case OP::OpCode::Asin: return DxilConstantFoldFP(asin, Op, Ty);
|
|
|
+ case OP::OpCode::Atan: return DxilConstantFoldFP(atan, Op, Ty);
|
|
|
+ case OP::OpCode::Hcos: return DxilConstantFoldFP(cosh, Op, Ty);
|
|
|
+ case OP::OpCode::Hsin: return DxilConstantFoldFP(sinh, Op, Ty);
|
|
|
+ case OP::OpCode::Htan: return DxilConstantFoldFP(tanh, Op, Ty);
|
|
|
+ case OP::OpCode::Exp: return DxilConstantFoldFP(exp2, Op, Ty);
|
|
|
+ case OP::OpCode::Frc: {
|
|
|
+ NativeFPUnaryOp f = [](double x) { double unused; return fabs(modf(x, &unused)); };
|
|
|
+ return DxilConstantFoldFP(f, Op, Ty);
|
|
|
+ }
|
|
|
+ case OP::OpCode::Log: return DxilConstantFoldFP(log2, Op, Ty);
|
|
|
+ case OP::OpCode::Sqrt: return DxilConstantFoldFP(sqrt, Op, Ty);
|
|
|
+ case OP::OpCode::Rsqrt: {
|
|
|
+ NativeFPUnaryOp f = [](double x) { return 1.0 / sqrt(x); };
|
|
|
+ return DxilConstantFoldFP(f, Op, Ty);
|
|
|
+ }
|
|
|
+ case OP::OpCode::Round_ne: return HLSLConstantFoldRound(APFloat::roundingMode::rmNearestTiesToEven, Op, Ty);
|
|
|
+ case OP::OpCode::Round_ni: return HLSLConstantFoldRound(APFloat::roundingMode::rmTowardNegative, Op, Ty);
|
|
|
+ case OP::OpCode::Round_pi: return HLSLConstantFoldRound(APFloat::roundingMode::rmTowardPositive, Op, Ty);
|
|
|
+ case OP::OpCode::Round_z: return HLSLConstantFoldRound(APFloat::roundingMode::rmTowardZero, Op, Ty);
|
|
|
+ }
|
|
|
+
|
|
|
+ return nullptr;
|
|
|
+}
|
|
|
+
|
|
|
+// Constant fold binary floating point intrinsics.
|
|
|
+static Constant *ConstantFoldBinaryFPIntrinsic(OP::OpCode opcode, Type *Ty, ConstantFP *Op1, ConstantFP *Op2) {
|
|
|
+ const APFloat &C1 = Op1->getValueAPF();
|
|
|
+ const APFloat &C2 = Op2->getValueAPF();
|
|
|
+ switch (opcode) {
|
|
|
+ default: break;
|
|
|
+ case OP::OpCode::FMax: return ConstantFP::get(Ty->getContext(), maxnum(C1, C2));
|
|
|
+ case OP::OpCode::FMin: return ConstantFP::get(Ty->getContext(), minnum(C1, C2));
|
|
|
+ }
|
|
|
+
|
|
|
+ return nullptr;
|
|
|
+}
|
|
|
+
|
|
|
+// Constant fold ternary floating point intrinsics.
|
|
|
+static Constant *ConstantFoldTernaryFPIntrinsic(OP::OpCode opcode, Type *Ty, ConstantFP *Op1, ConstantFP *Op2, ConstantFP *Op3) {
|
|
|
+ const APFloat &C1 = Op1->getValueAPF();
|
|
|
+ const APFloat &C2 = Op2->getValueAPF();
|
|
|
+ const APFloat &C3 = Op3->getValueAPF();
|
|
|
+ APFloat::roundingMode roundingMode = APFloat::rmNearestTiesToEven;
|
|
|
+ switch (opcode) {
|
|
|
+ default: break;
|
|
|
+ case OP::OpCode::FMad: {
|
|
|
+ APFloat result(C1);
|
|
|
+ result.multiply(C2, roundingMode);
|
|
|
+ result.add(C3, roundingMode);
|
|
|
+ return ConstantFP::get(Ty->getContext(), result);
|
|
|
+ }
|
|
|
+ case OP::OpCode::Fma: {
|
|
|
+ APFloat result(C1);
|
|
|
+ result.fusedMultiplyAdd(C2, C3, roundingMode);
|
|
|
+ return ConstantFP::get(Ty->getContext(), result);
|
|
|
+ }
|
|
|
+ }
|
|
|
+ return nullptr;
|
|
|
+}
|
|
|
+
|
|
|
+// Compute dot product for arbitrary sized vectors.
|
|
|
+static Constant *ComputeDot(Type *Ty, ArrayRef<ConstantFP *> A, ArrayRef<ConstantFP *> B) {
|
|
|
+ if (A.size() != B.size() || !A.size()) {
|
|
|
+ assert(false && "invalid call to compute dot");
|
|
|
+ return nullptr;
|
|
|
+ }
|
|
|
+
|
|
|
+ if (!AllValidOps(A) || !AllValidOps(B))
|
|
|
+ return nullptr;
|
|
|
+
|
|
|
+ APFloat::roundingMode roundingMode = APFloat::roundingMode::rmNearestTiesToEven;
|
|
|
+ APFloat sum = APFloat::getZero(A[0]->getValueAPF().getSemantics());
|
|
|
+ for (int i = 0, e = A.size(); i != e; ++i) {
|
|
|
+ APFloat val(A[i]->getValueAPF());
|
|
|
+ val.multiply(B[i]->getValueAPF(), roundingMode);
|
|
|
+ sum.add(val, roundingMode);
|
|
|
+ }
|
|
|
+
|
|
|
+ return ConstantFP::get(Ty->getContext(), sum);
|
|
|
+
|
|
|
+}
|
|
|
+
|
|
|
+// Constant folding for dot2, dot3, and dot4.
|
|
|
+static Constant *ConstantFoldDot(OP::OpCode opcode, Type *Ty, const DxilIntrinsicOperands &operands) {
|
|
|
+ switch (opcode) {
|
|
|
+ default: break;
|
|
|
+ case OP::OpCode::Dot2: {
|
|
|
+ ConstantFP *Ax = operands.GetConstantFloat(0);
|
|
|
+ ConstantFP *Ay = operands.GetConstantFloat(1);
|
|
|
+ ConstantFP *Bx = operands.GetConstantFloat(2);
|
|
|
+ ConstantFP *By = operands.GetConstantFloat(3);
|
|
|
+ return ComputeDot(Ty, { Ax, Ay }, { Bx, By });
|
|
|
+ }
|
|
|
+ case OP::OpCode::Dot3: {
|
|
|
+ ConstantFP *Ax = operands.GetConstantFloat(0);
|
|
|
+ ConstantFP *Ay = operands.GetConstantFloat(1);
|
|
|
+ ConstantFP *Az = operands.GetConstantFloat(2);
|
|
|
+ ConstantFP *Bx = operands.GetConstantFloat(3);
|
|
|
+ ConstantFP *By = operands.GetConstantFloat(4);
|
|
|
+ ConstantFP *Bz = operands.GetConstantFloat(5);
|
|
|
+ return ComputeDot(Ty, { Ax, Ay, Az }, { Bx, By, Bz });
|
|
|
+ }
|
|
|
+ case OP::OpCode::Dot4: {
|
|
|
+ ConstantFP *Ax = operands.GetConstantFloat(0);
|
|
|
+ ConstantFP *Ay = operands.GetConstantFloat(1);
|
|
|
+ ConstantFP *Az = operands.GetConstantFloat(2);
|
|
|
+ ConstantFP *Aw = operands.GetConstantFloat(3);
|
|
|
+ ConstantFP *Bx = operands.GetConstantFloat(4);
|
|
|
+ ConstantFP *By = operands.GetConstantFloat(5);
|
|
|
+ ConstantFP *Bz = operands.GetConstantFloat(6);
|
|
|
+ ConstantFP *Bw = operands.GetConstantFloat(7);
|
|
|
+ return ComputeDot(Ty, { Ax, Ay, Az, Aw }, { Bx, By, Bz, Bw });
|
|
|
+ }
|
|
|
+ }
|
|
|
+
|
|
|
+ return nullptr;
|
|
|
+}
|
|
|
+
|
|
|
+// Constant fold a Bfrev dxil intrinsic.
|
|
|
+static Constant *HLSLConstantFoldBfrev(ConstantInt *C, Type *Ty) {
|
|
|
+ APInt API = C->getValue();
|
|
|
+
|
|
|
+ uint64_t result = 0;
|
|
|
+ if (Ty == Type::getInt32Ty(Ty->getContext())) {
|
|
|
+ uint32_t val = static_cast<uint32_t>(API.getLimitedValue());
|
|
|
+ result = llvm::reverseBits(val);
|
|
|
+ }
|
|
|
+ else if (Ty == Type::getInt16Ty(Ty->getContext())) {
|
|
|
+ uint16_t val = static_cast<uint16_t>(API.getLimitedValue());
|
|
|
+ result = llvm::reverseBits(val);
|
|
|
+ }
|
|
|
+ else if (Ty == Type::getInt64Ty(Ty->getContext())) {
|
|
|
+ uint64_t val = static_cast<uint64_t>(API.getLimitedValue());
|
|
|
+ result = llvm::reverseBits(val);
|
|
|
+ }
|
|
|
+ else {
|
|
|
+ return nullptr;
|
|
|
+ }
|
|
|
+ return ConstantInt::get(Ty, result);
|
|
|
+}
|
|
|
+
|
|
|
+// Handle special case for findfirst* bit functions.
|
|
|
+// When the position is equal to the bitwidth the value was not found
|
|
|
+// and we need to return a result of -1.
|
|
|
+static Constant *HLSLConstantFoldFindBit(Type *Ty, unsigned position, unsigned bitwidth) {
|
|
|
+ if (position == bitwidth)
|
|
|
+ return ConstantInt::get(Ty, APInt::getAllOnesValue(Ty->getScalarSizeInBits()));
|
|
|
+
|
|
|
+ return ConstantInt::get(Ty, position);
|
|
|
+}
|
|
|
+
|
|
|
+// Constant fold unary integer intrinsics.
|
|
|
+static Constant *ConstantFoldUnaryIntIntrinsic(OP::OpCode opcode, Type *Ty, ConstantInt *Op) {
|
|
|
+ APInt API = Op->getValue();
|
|
|
+ switch (opcode) {
|
|
|
+ default: break;
|
|
|
+ case OP::OpCode::Bfrev: return HLSLConstantFoldBfrev(Op, Ty);
|
|
|
+ case OP::OpCode::Countbits: return ConstantInt::get(Ty, API.countPopulation());
|
|
|
+ case OP::OpCode::FirstbitLo: return HLSLConstantFoldFindBit(Ty, API.countTrailingZeros(), API.getBitWidth());
|
|
|
+ case OP::OpCode::FirstbitHi: return HLSLConstantFoldFindBit(Ty, API.countLeadingZeros(), API.getBitWidth());
|
|
|
+ case OP::OpCode::FirstbitSHi: {
|
|
|
+ if (API.isNegative())
|
|
|
+ return HLSLConstantFoldFindBit(Ty, API.countLeadingOnes(), API.getBitWidth());
|
|
|
+ else
|
|
|
+ return HLSLConstantFoldFindBit(Ty, API.countLeadingZeros(), API.getBitWidth());
|
|
|
+ }
|
|
|
+ }
|
|
|
+
|
|
|
+ return nullptr;
|
|
|
+}
|
|
|
+
|
|
|
+// Constant fold binary integer intrinsics.
|
|
|
+static Constant *ConstantFoldBinaryIntIntrinsic(OP::OpCode opcode, Type *Ty, ConstantInt *Op1, ConstantInt *Op2) {
|
|
|
+ APInt C1 = Op1->getValue();
|
|
|
+ APInt C2 = Op2->getValue();
|
|
|
+ switch (opcode) {
|
|
|
+ default: break;
|
|
|
+ case OP::OpCode::IMin: {
|
|
|
+ APInt minVal = C1.slt(C2) ? C1 : C2;
|
|
|
+ return ConstantInt::get(Ty, minVal);
|
|
|
+ }
|
|
|
+ case OP::OpCode::IMax: {
|
|
|
+ APInt maxVal = C1.sgt(C2) ? C1 : C2;
|
|
|
+ return ConstantInt::get(Ty, maxVal);
|
|
|
+ }
|
|
|
+ case OP::OpCode::UMin: {
|
|
|
+ APInt minVal = C1.ult(C2) ? C1 : C2;
|
|
|
+ return ConstantInt::get(Ty, minVal);
|
|
|
+ }
|
|
|
+ case OP::OpCode::UMax: {
|
|
|
+ APInt maxVal = C1.ugt(C2) ? C1 : C2;
|
|
|
+ return ConstantInt::get(Ty, maxVal);
|
|
|
+ }
|
|
|
+ }
|
|
|
+
|
|
|
+ return nullptr;
|
|
|
+}
|
|
|
+
|
|
|
+// Compute bit field extract for ibfe and ubfe.
|
|
|
+// The comptuation for ibfe and ubfe is the same except for the right shift,
|
|
|
+// which is an arithemetic shift for ibfe and logical shift for ubfe.
|
|
|
+// ubfe: https://msdn.microsoft.com/en-us/library/windows/desktop/hh447243(v=vs.85).aspx
|
|
|
+// ibfe: https://msdn.microsoft.com/en-us/library/windows/desktop/hh447243(v=vs.85).aspx
|
|
|
+static Constant *ComputeBFE(Type *Ty, APInt width, APInt offset, APInt val, std::function<APInt(APInt, APInt)> shr) {
|
|
|
+ const APInt bitwidth(width.getBitWidth(), width.getBitWidth());
|
|
|
+ // Limit width and offset to the bitwidth of the value.
|
|
|
+ width = width.And(bitwidth-1);
|
|
|
+ offset = offset.And(bitwidth-1);
|
|
|
+
|
|
|
+ if (width == 0) {
|
|
|
+ return ConstantInt::get(Ty, 0);
|
|
|
+ }
|
|
|
+ else if ((width + offset).ult(bitwidth)) {
|
|
|
+ APInt dest = val.shl(bitwidth - (width + offset));
|
|
|
+ dest = shr(dest, bitwidth - width);
|
|
|
+ return ConstantInt::get(Ty, dest);
|
|
|
+ }
|
|
|
+ else {
|
|
|
+ APInt dest = shr(val, offset);
|
|
|
+ return ConstantInt::get(Ty, dest);
|
|
|
+ }
|
|
|
+}
|
|
|
+
|
|
|
+// Constant fold ternary integer intrinsic.
|
|
|
+static Constant *ConstantFoldTernaryIntIntrinsic(OP::OpCode opcode, Type *Ty, ConstantInt *Op1, ConstantInt *Op2, ConstantInt *Op3) {
|
|
|
+ APInt C1 = Op1->getValue();
|
|
|
+ APInt C2 = Op2->getValue();
|
|
|
+ APInt C3 = Op3->getValue();
|
|
|
+ switch (opcode) {
|
|
|
+ default: break;
|
|
|
+ case OP::OpCode::IMad:
|
|
|
+ case OP::OpCode::UMad: {
|
|
|
+ // Result is same for signed/unsigned since this is twos complement and we only
|
|
|
+ // keep the lower half of the multiply.
|
|
|
+ APInt result = C1 * C2 + C3;
|
|
|
+ return ConstantInt::get(Ty, result);
|
|
|
+ }
|
|
|
+ case OP::OpCode::Ubfe: return ComputeBFE(Ty, C1, C2, C3, [](APInt val, APInt amt) {return val.lshr(amt); });
|
|
|
+ case OP::OpCode::Ibfe: return ComputeBFE(Ty, C1, C2, C3, [](APInt val, APInt amt) {return val.ashr(amt); });
|
|
|
+ }
|
|
|
+
|
|
|
+ return nullptr;
|
|
|
+}
|
|
|
+
|
|
|
+// Constant fold quaternary integer intrinsic.
|
|
|
+//
|
|
|
+// Currently we only have one quaternary intrinsic: Bfi.
|
|
|
+// The Bfi computaion is described here:
|
|
|
+// https://msdn.microsoft.com/en-us/library/windows/desktop/hh446837(v=vs.85).aspx
|
|
|
+static Constant *ConstantFoldQuaternaryIntInstrinsic(OP::OpCode opcode, Type *Ty, ConstantInt *Op1, ConstantInt *Op2, ConstantInt *Op3, ConstantInt *Op4) {
|
|
|
+ if (opcode != OP::OpCode::Bfi)
|
|
|
+ return nullptr;
|
|
|
+
|
|
|
+ APInt bitwidth(Op1->getValue().getBitWidth(), Op1->getValue().getBitWidth());
|
|
|
+ APInt width = Op1->getValue().And(bitwidth-1);
|
|
|
+ APInt offset = Op2->getValue().And(bitwidth-1);
|
|
|
+ APInt src = Op3->getValue();
|
|
|
+ APInt dst = Op4->getValue();
|
|
|
+ APInt one(bitwidth.getBitWidth(), 1);
|
|
|
+ APInt allOnes = APInt::getAllOnesValue(bitwidth.getBitWidth());
|
|
|
+
|
|
|
+ // bitmask = (((1 << width)-1) << offset) & 0xffffffff
|
|
|
+ // dest = ((src2 << offset) & bitmask) | (src3 & ~bitmask)
|
|
|
+ APInt bitmask = (one.shl(width) - 1).shl(offset).And(allOnes);
|
|
|
+ APInt result = (src.shl(offset).And(bitmask)).Or(dst.And(~bitmask));
|
|
|
+
|
|
|
+ return ConstantInt::get(Ty, result);
|
|
|
+}
|
|
|
+
|
|
|
+// Return true if opcode is for a dot operation.
|
|
|
+static bool IsDotOpcode(OP::OpCode opcode) {
|
|
|
+ return opcode == OP::OpCode::Dot2
|
|
|
+ || opcode == OP::OpCode::Dot3
|
|
|
+ || opcode == OP::OpCode::Dot4;
|
|
|
+}
|
|
|
+
|
|
|
+// Top level function to constant fold floating point intrinsics.
|
|
|
+static Constant *ConstantFoldFPIntrinsic(OP::OpCode opcode, Type *Ty, const DxilIntrinsicOperands &IntrinsicOperands) {
|
|
|
+ if (!Ty->isHalfTy() && !Ty->isFloatTy() && !Ty->isDoubleTy())
|
|
|
+ return nullptr;
|
|
|
+
|
|
|
+ if (IntrinsicOperands.Size() == 1) {
|
|
|
+ ConstantFP *Op = IntrinsicOperands.GetConstantFloat(0);
|
|
|
+
|
|
|
+ if (!IsValidOp(Op))
|
|
|
+ return nullptr;
|
|
|
+
|
|
|
+ return ConstantFoldUnaryFPIntrinsic(opcode, Ty, Op);
|
|
|
+ }
|
|
|
+ else if (IntrinsicOperands.Size() == 2) {
|
|
|
+ ConstantFP *Op1 = IntrinsicOperands.GetConstantFloat(0);
|
|
|
+ ConstantFP *Op2 = IntrinsicOperands.GetConstantFloat(1);
|
|
|
+
|
|
|
+ if (!IsValidOp(Op1) || !IsValidOp(Op2))
|
|
|
+ return nullptr;
|
|
|
+
|
|
|
+ return ConstantFoldBinaryFPIntrinsic(opcode, Ty, Op1, Op2);
|
|
|
+ }
|
|
|
+ else if (IntrinsicOperands.Size() == 3) {
|
|
|
+ ConstantFP *Op1 = IntrinsicOperands.GetConstantFloat(0);
|
|
|
+ ConstantFP *Op2 = IntrinsicOperands.GetConstantFloat(1);
|
|
|
+ ConstantFP *Op3 = IntrinsicOperands.GetConstantFloat(2);
|
|
|
+
|
|
|
+ if (!IsValidOp(Op1) || !IsValidOp(Op2) || !IsValidOp(Op3))
|
|
|
+ return nullptr;
|
|
|
+
|
|
|
+ return ConstantFoldTernaryFPIntrinsic(opcode, Ty, Op1, Op2, Op3);
|
|
|
+ }
|
|
|
+ else if (IsDotOpcode(opcode)) {
|
|
|
+ return ConstantFoldDot(opcode, Ty, IntrinsicOperands);
|
|
|
+ }
|
|
|
+
|
|
|
+ return nullptr;
|
|
|
+}
|
|
|
+
|
|
|
+// Top level function to constant fold integer intrinsics.
|
|
|
+static Constant *ConstantFoldIntIntrinsic(OP::OpCode opcode, Type *Ty, const DxilIntrinsicOperands &IntrinsicOperands) {
|
|
|
+ if (Ty->getScalarSizeInBits() > (sizeof(int64_t) * CHAR_BIT))
|
|
|
+ return nullptr;
|
|
|
+
|
|
|
+ if (IntrinsicOperands.Size() == 1) {
|
|
|
+ ConstantInt *Op = IntrinsicOperands.GetConstantInt(0);
|
|
|
+ if (!Op)
|
|
|
+ return nullptr;
|
|
|
+
|
|
|
+ return ConstantFoldUnaryIntIntrinsic(opcode, Ty, Op);
|
|
|
+ }
|
|
|
+ else if (IntrinsicOperands.Size() == 2) {
|
|
|
+ ConstantInt *Op1 = IntrinsicOperands.GetConstantInt(0);
|
|
|
+ ConstantInt *Op2 = IntrinsicOperands.GetConstantInt(1);
|
|
|
+ if (!Op1 || !Op2)
|
|
|
+ return nullptr;
|
|
|
+
|
|
|
+ return ConstantFoldBinaryIntIntrinsic(opcode, Ty, Op1, Op2);
|
|
|
+ }
|
|
|
+ else if (IntrinsicOperands.Size() == 3) {
|
|
|
+ ConstantInt *Op1 = IntrinsicOperands.GetConstantInt(0);
|
|
|
+ ConstantInt *Op2 = IntrinsicOperands.GetConstantInt(1);
|
|
|
+ ConstantInt *Op3 = IntrinsicOperands.GetConstantInt(2);
|
|
|
+ if (!Op1 || !Op2 || !Op3)
|
|
|
+ return nullptr;
|
|
|
+
|
|
|
+ return ConstantFoldTernaryIntIntrinsic(opcode, Ty, Op1, Op2, Op3);
|
|
|
+ }
|
|
|
+ else if (IntrinsicOperands.Size() == 4) {
|
|
|
+ ConstantInt *Op1 = IntrinsicOperands.GetConstantInt(0);
|
|
|
+ ConstantInt *Op2 = IntrinsicOperands.GetConstantInt(1);
|
|
|
+ ConstantInt *Op3 = IntrinsicOperands.GetConstantInt(2);
|
|
|
+ ConstantInt *Op4 = IntrinsicOperands.GetConstantInt(3);
|
|
|
+ if (!Op1 || !Op2 || !Op3 || !Op4)
|
|
|
+ return nullptr;
|
|
|
+
|
|
|
+ return ConstantFoldQuaternaryIntInstrinsic(opcode, Ty, Op1, Op2, Op3, Op4);
|
|
|
+ }
|
|
|
+ return nullptr;
|
|
|
+}
|
|
|
+
|
|
|
+// External entry point to constant fold dxil intrinsics.
|
|
|
+// Called from the llvm constant folding routine.
|
|
|
+Constant *hlsl::ConstantFoldScalarCall(StringRef Name, Type *Ty, ArrayRef<Constant *> RawOperands) {
|
|
|
+ OP::OpCode opcode;
|
|
|
+ if (GetDxilOpcode(Name, RawOperands, opcode)) {
|
|
|
+ DxilIntrinsicOperands IntrinsicOperands(RawOperands);
|
|
|
+
|
|
|
+ if (Ty->isFloatingPointTy()) {
|
|
|
+ return ConstantFoldFPIntrinsic(opcode, Ty, IntrinsicOperands);
|
|
|
+ }
|
|
|
+ else if (Ty->isIntegerTy()) {
|
|
|
+ return ConstantFoldIntIntrinsic(opcode, Ty, IntrinsicOperands);
|
|
|
+ }
|
|
|
+ }
|
|
|
+ return nullptr;
|
|
|
+}
|
|
|
+
|
|
|
+// External entry point to determine if we can constant fold calls to
|
|
|
+// the given function. We have to overestimate the set of functions because
|
|
|
+// we only have the function value here instead of the call. We need the
|
|
|
+// actual call to get the opcode for the intrinsic.
|
|
|
+bool hlsl::CanConstantFoldCallTo(const Function *F) {
|
|
|
+ if (!OP::IsDxilOpFunc(F))
|
|
|
+ return false;
|
|
|
+
|
|
|
+ // Check match using startswith to get all overloads.
|
|
|
+ StringRef Name = F->getName();
|
|
|
+ if (Name.startswith("dx.op.unary"))
|
|
|
+ return true;
|
|
|
+ else if (Name.startswith("dx.op.unaryBits"))
|
|
|
+ return true;
|
|
|
+ else if (Name.startswith("dx.op.binary"))
|
|
|
+ return true;
|
|
|
+ else if (Name.startswith("dx.op.tertiary"))
|
|
|
+ return true;
|
|
|
+ else if (Name.startswith("dx.op.quaternary"))
|
|
|
+ return true;
|
|
|
+ else if (Name.startswith("dx.op.dot"))
|
|
|
+ return true;
|
|
|
+
|
|
|
+ return false;
|
|
|
+}
|