750e90e440
Many machine-learning applications (and most software written at AMD) expect the operation that truncates floats to 8-bit floats to be saturatinng. That is, they expect `truncf 256.0 : f32 to f8E4M3FNUZ` to yield `240.0`, not `NaN`, and similarly for negative numbers. However, the underlying hardware instruction that can be used for this truncation implements overflow-to-NaN semantics. To enable handling this usecase, we add the saturate-fp8-truncf option to ArithToAMDGPU (off by default), which causes the requisite clamping code to be emitted. Said clamping code ensures that Inf and NaN are passed through exactly (and thus trancate to NaN). Per review feedback, this commit efactors createScalarOrSplatConstant() to the Arith dialect utilities and uses it in this code. It also fixes naming of existing patterns and switches from vector.extractelement/insertelement to vector.extract/insert.
1155 lines
46 KiB
C++
1155 lines
46 KiB
C++
//===- EmulateWideInt.cpp - Wide integer operation emulation ----*- C++ -*-===//
|
|
//
|
|
// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
|
|
// See https://llvm.org/LICENSE.txt for license information.
|
|
// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
|
|
//
|
|
//===----------------------------------------------------------------------===//
|
|
|
|
#include "mlir/Dialect/Arith/Transforms/Passes.h"
|
|
|
|
#include "mlir/Dialect/Arith/IR/Arith.h"
|
|
#include "mlir/Dialect/Arith/Transforms/WideIntEmulationConverter.h"
|
|
#include "mlir/Dialect/Arith/Utils/Utils.h"
|
|
#include "mlir/Dialect/Func/IR/FuncOps.h"
|
|
#include "mlir/Dialect/Func/Transforms/FuncConversions.h"
|
|
#include "mlir/Dialect/Vector/IR/VectorOps.h"
|
|
#include "mlir/IR/BuiltinTypes.h"
|
|
#include "mlir/IR/TypeUtilities.h"
|
|
#include "mlir/Support/LogicalResult.h"
|
|
#include "mlir/Transforms/DialectConversion.h"
|
|
#include "llvm/ADT/APInt.h"
|
|
#include "llvm/Support/FormatVariadic.h"
|
|
#include "llvm/Support/MathExtras.h"
|
|
#include <cassert>
|
|
|
|
namespace mlir::arith {
|
|
#define GEN_PASS_DEF_ARITHEMULATEWIDEINT
|
|
#include "mlir/Dialect/Arith/Transforms/Passes.h.inc"
|
|
} // namespace mlir::arith
|
|
|
|
using namespace mlir;
|
|
|
|
//===----------------------------------------------------------------------===//
|
|
// Common Helper Functions
|
|
//===----------------------------------------------------------------------===//
|
|
|
|
/// Returns N bottom and N top bits from `value`, where N = `newBitWidth`.
|
|
/// Treats `value` as a 2*N bits-wide integer.
|
|
/// The bottom bits are returned in the first pair element, while the top bits
|
|
/// in the second one.
|
|
static std::pair<APInt, APInt> getHalves(const APInt &value,
|
|
unsigned newBitWidth) {
|
|
APInt low = value.extractBits(newBitWidth, 0);
|
|
APInt high = value.extractBits(newBitWidth, newBitWidth);
|
|
return {std::move(low), std::move(high)};
|
|
}
|
|
|
|
/// Returns the type with the last (innermost) dimension reduced to x1.
|
|
/// Scalarizes 1D vector inputs to match how we extract/insert vector values,
|
|
/// e.g.:
|
|
/// - vector<3x2xi16> --> vector<3x1xi16>
|
|
/// - vector<2xi16> --> i16
|
|
static Type reduceInnermostDim(VectorType type) {
|
|
if (type.getShape().size() == 1)
|
|
return type.getElementType();
|
|
|
|
auto newShape = to_vector(type.getShape());
|
|
newShape.back() = 1;
|
|
return VectorType::get(newShape, type.getElementType());
|
|
}
|
|
|
|
/// Extracts the `input` vector slice with elements at the last dimension offset
|
|
/// by `lastOffset`. Returns a value of vector type with the last dimension
|
|
/// reduced to x1 or fully scalarized, e.g.:
|
|
/// - vector<3x2xi16> --> vector<3x1xi16>
|
|
/// - vector<2xi16> --> i16
|
|
static Value extractLastDimSlice(ConversionPatternRewriter &rewriter,
|
|
Location loc, Value input,
|
|
int64_t lastOffset) {
|
|
ArrayRef<int64_t> shape = cast<VectorType>(input.getType()).getShape();
|
|
assert(lastOffset < shape.back() && "Offset out of bounds");
|
|
|
|
// Scalarize the result in case of 1D vectors.
|
|
if (shape.size() == 1)
|
|
return rewriter.create<vector::ExtractOp>(loc, input, lastOffset);
|
|
|
|
SmallVector<int64_t> offsets(shape.size(), 0);
|
|
offsets.back() = lastOffset;
|
|
auto sizes = llvm::to_vector(shape);
|
|
sizes.back() = 1;
|
|
SmallVector<int64_t> strides(shape.size(), 1);
|
|
|
|
return rewriter.create<vector::ExtractStridedSliceOp>(loc, input, offsets,
|
|
sizes, strides);
|
|
}
|
|
|
|
/// Extracts two vector slices from the `input` whose type is `vector<...x2T>`,
|
|
/// with the first element at offset 0 and the second element at offset 1.
|
|
static std::pair<Value, Value>
|
|
extractLastDimHalves(ConversionPatternRewriter &rewriter, Location loc,
|
|
Value input) {
|
|
return {extractLastDimSlice(rewriter, loc, input, 0),
|
|
extractLastDimSlice(rewriter, loc, input, 1)};
|
|
}
|
|
|
|
// Performs a vector shape cast to drop the trailing x1 dimension. If the
|
|
// `input` is a scalar, this is a noop.
|
|
static Value dropTrailingX1Dim(ConversionPatternRewriter &rewriter,
|
|
Location loc, Value input) {
|
|
auto vecTy = dyn_cast<VectorType>(input.getType());
|
|
if (!vecTy)
|
|
return input;
|
|
|
|
// Shape cast to drop the last x1 dimension.
|
|
ArrayRef<int64_t> shape = vecTy.getShape();
|
|
assert(shape.size() >= 2 && "Expected vector with at list two dims");
|
|
assert(shape.back() == 1 && "Expected the last vector dim to be x1");
|
|
|
|
auto newVecTy = VectorType::get(shape.drop_back(), vecTy.getElementType());
|
|
return rewriter.create<vector::ShapeCastOp>(loc, newVecTy, input);
|
|
}
|
|
|
|
/// Performs a vector shape cast to append an x1 dimension. If the
|
|
/// `input` is a scalar, this is a noop.
|
|
static Value appendX1Dim(ConversionPatternRewriter &rewriter, Location loc,
|
|
Value input) {
|
|
auto vecTy = dyn_cast<VectorType>(input.getType());
|
|
if (!vecTy)
|
|
return input;
|
|
|
|
// Add a trailing x1 dim.
|
|
auto newShape = llvm::to_vector(vecTy.getShape());
|
|
newShape.push_back(1);
|
|
auto newTy = VectorType::get(newShape, vecTy.getElementType());
|
|
return rewriter.create<vector::ShapeCastOp>(loc, newTy, input);
|
|
}
|
|
|
|
/// Inserts the `source` vector slice into the `dest` vector at offset
|
|
/// `lastOffset` in the last dimension. `source` can be a scalar when `dest` is
|
|
/// a 1D vector.
|
|
static Value insertLastDimSlice(ConversionPatternRewriter &rewriter,
|
|
Location loc, Value source, Value dest,
|
|
int64_t lastOffset) {
|
|
ArrayRef<int64_t> shape = cast<VectorType>(dest.getType()).getShape();
|
|
assert(lastOffset < shape.back() && "Offset out of bounds");
|
|
|
|
// Handle scalar source.
|
|
if (isa<IntegerType>(source.getType()))
|
|
return rewriter.create<vector::InsertOp>(loc, source, dest, lastOffset);
|
|
|
|
SmallVector<int64_t> offsets(shape.size(), 0);
|
|
offsets.back() = lastOffset;
|
|
SmallVector<int64_t> strides(shape.size(), 1);
|
|
return rewriter.create<vector::InsertStridedSliceOp>(loc, source, dest,
|
|
offsets, strides);
|
|
}
|
|
|
|
/// Constructs a new vector of type `resultType` by creating a series of
|
|
/// insertions of `resultComponents`, each at the next offset of the last vector
|
|
/// dimension.
|
|
/// When all `resultComponents` are scalars, the result type is `vector<NxT>`;
|
|
/// when `resultComponents` are `vector<...x1xT>`s, the result type is
|
|
/// `vector<...xNxT>`, where `N` is the number of `resultComponents`.
|
|
static Value constructResultVector(ConversionPatternRewriter &rewriter,
|
|
Location loc, VectorType resultType,
|
|
ValueRange resultComponents) {
|
|
llvm::ArrayRef<int64_t> resultShape = resultType.getShape();
|
|
(void)resultShape;
|
|
assert(!resultShape.empty() && "Result expected to have dimensions");
|
|
assert(resultShape.back() == static_cast<int64_t>(resultComponents.size()) &&
|
|
"Wrong number of result components");
|
|
|
|
Value resultVec = createScalarOrSplatConstant(rewriter, loc, resultType, 0);
|
|
for (auto [i, component] : llvm::enumerate(resultComponents))
|
|
resultVec = insertLastDimSlice(rewriter, loc, component, resultVec, i);
|
|
|
|
return resultVec;
|
|
}
|
|
|
|
namespace {
|
|
//===----------------------------------------------------------------------===//
|
|
// ConvertConstant
|
|
//===----------------------------------------------------------------------===//
|
|
|
|
struct ConvertConstant final : OpConversionPattern<arith::ConstantOp> {
|
|
using OpConversionPattern::OpConversionPattern;
|
|
|
|
LogicalResult
|
|
matchAndRewrite(arith::ConstantOp op, OpAdaptor,
|
|
ConversionPatternRewriter &rewriter) const override {
|
|
Type oldType = op.getType();
|
|
auto newType = getTypeConverter()->convertType<VectorType>(oldType);
|
|
if (!newType)
|
|
return rewriter.notifyMatchFailure(
|
|
op, llvm::formatv("unsupported type: {0}", op.getType()));
|
|
|
|
unsigned newBitWidth = newType.getElementTypeBitWidth();
|
|
Attribute oldValue = op.getValueAttr();
|
|
|
|
if (auto intAttr = dyn_cast<IntegerAttr>(oldValue)) {
|
|
auto [low, high] = getHalves(intAttr.getValue(), newBitWidth);
|
|
auto newAttr = DenseElementsAttr::get(newType, {low, high});
|
|
rewriter.replaceOpWithNewOp<arith::ConstantOp>(op, newAttr);
|
|
return success();
|
|
}
|
|
|
|
if (auto splatAttr = dyn_cast<SplatElementsAttr>(oldValue)) {
|
|
auto [low, high] =
|
|
getHalves(splatAttr.getSplatValue<APInt>(), newBitWidth);
|
|
int64_t numSplatElems = splatAttr.getNumElements();
|
|
SmallVector<APInt> values;
|
|
values.reserve(numSplatElems * 2);
|
|
for (int64_t i = 0; i < numSplatElems; ++i) {
|
|
values.push_back(low);
|
|
values.push_back(high);
|
|
}
|
|
|
|
auto attr = DenseElementsAttr::get(newType, values);
|
|
rewriter.replaceOpWithNewOp<arith::ConstantOp>(op, attr);
|
|
return success();
|
|
}
|
|
|
|
if (auto elemsAttr = dyn_cast<DenseElementsAttr>(oldValue)) {
|
|
int64_t numElems = elemsAttr.getNumElements();
|
|
SmallVector<APInt> values;
|
|
values.reserve(numElems * 2);
|
|
for (const APInt &origVal : elemsAttr.getValues<APInt>()) {
|
|
auto [low, high] = getHalves(origVal, newBitWidth);
|
|
values.push_back(std::move(low));
|
|
values.push_back(std::move(high));
|
|
}
|
|
|
|
auto attr = DenseElementsAttr::get(newType, values);
|
|
rewriter.replaceOpWithNewOp<arith::ConstantOp>(op, attr);
|
|
return success();
|
|
}
|
|
|
|
return rewriter.notifyMatchFailure(op.getLoc(),
|
|
"unhandled constant attribute");
|
|
}
|
|
};
|
|
|
|
//===----------------------------------------------------------------------===//
|
|
// ConvertAddI
|
|
//===----------------------------------------------------------------------===//
|
|
|
|
struct ConvertAddI final : OpConversionPattern<arith::AddIOp> {
|
|
using OpConversionPattern::OpConversionPattern;
|
|
|
|
LogicalResult
|
|
matchAndRewrite(arith::AddIOp op, OpAdaptor adaptor,
|
|
ConversionPatternRewriter &rewriter) const override {
|
|
Location loc = op->getLoc();
|
|
auto newTy = getTypeConverter()->convertType<VectorType>(op.getType());
|
|
if (!newTy)
|
|
return rewriter.notifyMatchFailure(
|
|
loc, llvm::formatv("unsupported type: {0}", op.getType()));
|
|
|
|
Type newElemTy = reduceInnermostDim(newTy);
|
|
|
|
auto [lhsElem0, lhsElem1] =
|
|
extractLastDimHalves(rewriter, loc, adaptor.getLhs());
|
|
auto [rhsElem0, rhsElem1] =
|
|
extractLastDimHalves(rewriter, loc, adaptor.getRhs());
|
|
|
|
auto lowSum =
|
|
rewriter.create<arith::AddUIExtendedOp>(loc, lhsElem0, rhsElem0);
|
|
Value overflowVal =
|
|
rewriter.create<arith::ExtUIOp>(loc, newElemTy, lowSum.getOverflow());
|
|
|
|
Value high0 = rewriter.create<arith::AddIOp>(loc, overflowVal, lhsElem1);
|
|
Value high = rewriter.create<arith::AddIOp>(loc, high0, rhsElem1);
|
|
|
|
Value resultVec =
|
|
constructResultVector(rewriter, loc, newTy, {lowSum.getSum(), high});
|
|
rewriter.replaceOp(op, resultVec);
|
|
return success();
|
|
}
|
|
};
|
|
|
|
//===----------------------------------------------------------------------===//
|
|
// ConvertBitwiseBinary
|
|
//===----------------------------------------------------------------------===//
|
|
|
|
/// Conversion pattern template for bitwise binary ops, e.g., `arith.andi`.
|
|
template <typename BinaryOp>
|
|
struct ConvertBitwiseBinary final : OpConversionPattern<BinaryOp> {
|
|
using OpConversionPattern<BinaryOp>::OpConversionPattern;
|
|
using OpAdaptor = typename OpConversionPattern<BinaryOp>::OpAdaptor;
|
|
|
|
LogicalResult
|
|
matchAndRewrite(BinaryOp op, OpAdaptor adaptor,
|
|
ConversionPatternRewriter &rewriter) const override {
|
|
Location loc = op->getLoc();
|
|
auto newTy = this->getTypeConverter()->template convertType<VectorType>(
|
|
op.getType());
|
|
if (!newTy)
|
|
return rewriter.notifyMatchFailure(
|
|
loc, llvm::formatv("unsupported type: {0}", op.getType()));
|
|
|
|
auto [lhsElem0, lhsElem1] =
|
|
extractLastDimHalves(rewriter, loc, adaptor.getLhs());
|
|
auto [rhsElem0, rhsElem1] =
|
|
extractLastDimHalves(rewriter, loc, adaptor.getRhs());
|
|
|
|
Value resElem0 = rewriter.create<BinaryOp>(loc, lhsElem0, rhsElem0);
|
|
Value resElem1 = rewriter.create<BinaryOp>(loc, lhsElem1, rhsElem1);
|
|
Value resultVec =
|
|
constructResultVector(rewriter, loc, newTy, {resElem0, resElem1});
|
|
rewriter.replaceOp(op, resultVec);
|
|
return success();
|
|
}
|
|
};
|
|
|
|
//===----------------------------------------------------------------------===//
|
|
// ConvertCmpI
|
|
//===----------------------------------------------------------------------===//
|
|
|
|
/// Returns the matching unsigned version of the given predicate `pred`, or the
|
|
/// same predicate if `pred` is not a signed.
|
|
static arith::CmpIPredicate toUnsignedPredicate(arith::CmpIPredicate pred) {
|
|
using P = arith::CmpIPredicate;
|
|
switch (pred) {
|
|
case P::sge:
|
|
return P::uge;
|
|
case P::sgt:
|
|
return P::ugt;
|
|
case P::sle:
|
|
return P::ule;
|
|
case P::slt:
|
|
return P::ult;
|
|
default:
|
|
return pred;
|
|
}
|
|
}
|
|
|
|
struct ConvertCmpI final : OpConversionPattern<arith::CmpIOp> {
|
|
using OpConversionPattern::OpConversionPattern;
|
|
|
|
LogicalResult
|
|
matchAndRewrite(arith::CmpIOp op, OpAdaptor adaptor,
|
|
ConversionPatternRewriter &rewriter) const override {
|
|
Location loc = op->getLoc();
|
|
auto inputTy =
|
|
getTypeConverter()->convertType<VectorType>(op.getLhs().getType());
|
|
if (!inputTy)
|
|
return rewriter.notifyMatchFailure(
|
|
loc, llvm::formatv("unsupported type: {0}", op.getType()));
|
|
|
|
arith::CmpIPredicate highPred = adaptor.getPredicate();
|
|
arith::CmpIPredicate lowPred = toUnsignedPredicate(highPred);
|
|
|
|
auto [lhsElem0, lhsElem1] =
|
|
extractLastDimHalves(rewriter, loc, adaptor.getLhs());
|
|
auto [rhsElem0, rhsElem1] =
|
|
extractLastDimHalves(rewriter, loc, adaptor.getRhs());
|
|
|
|
Value lowCmp =
|
|
rewriter.create<arith::CmpIOp>(loc, lowPred, lhsElem0, rhsElem0);
|
|
Value highCmp =
|
|
rewriter.create<arith::CmpIOp>(loc, highPred, lhsElem1, rhsElem1);
|
|
|
|
Value cmpResult{};
|
|
switch (highPred) {
|
|
case arith::CmpIPredicate::eq: {
|
|
cmpResult = rewriter.create<arith::AndIOp>(loc, lowCmp, highCmp);
|
|
break;
|
|
}
|
|
case arith::CmpIPredicate::ne: {
|
|
cmpResult = rewriter.create<arith::OrIOp>(loc, lowCmp, highCmp);
|
|
break;
|
|
}
|
|
default: {
|
|
// Handle inequality checks.
|
|
Value highEq = rewriter.create<arith::CmpIOp>(
|
|
loc, arith::CmpIPredicate::eq, lhsElem1, rhsElem1);
|
|
cmpResult =
|
|
rewriter.create<arith::SelectOp>(loc, highEq, lowCmp, highCmp);
|
|
break;
|
|
}
|
|
}
|
|
|
|
assert(cmpResult && "Unhandled case");
|
|
rewriter.replaceOp(op, dropTrailingX1Dim(rewriter, loc, cmpResult));
|
|
return success();
|
|
}
|
|
};
|
|
|
|
//===----------------------------------------------------------------------===//
|
|
// ConvertMulI
|
|
//===----------------------------------------------------------------------===//
|
|
|
|
struct ConvertMulI final : OpConversionPattern<arith::MulIOp> {
|
|
using OpConversionPattern::OpConversionPattern;
|
|
|
|
LogicalResult
|
|
matchAndRewrite(arith::MulIOp op, OpAdaptor adaptor,
|
|
ConversionPatternRewriter &rewriter) const override {
|
|
Location loc = op->getLoc();
|
|
auto newTy = getTypeConverter()->convertType<VectorType>(op.getType());
|
|
if (!newTy)
|
|
return rewriter.notifyMatchFailure(
|
|
loc, llvm::formatv("unsupported type: {0}", op.getType()));
|
|
|
|
auto [lhsElem0, lhsElem1] =
|
|
extractLastDimHalves(rewriter, loc, adaptor.getLhs());
|
|
auto [rhsElem0, rhsElem1] =
|
|
extractLastDimHalves(rewriter, loc, adaptor.getRhs());
|
|
|
|
// The multiplication algorithm used is the standard (long) multiplication.
|
|
// Multiplying two i2N integers produces (at most) an i4N result, but
|
|
// because the calculation of top i2N is not necessary, we omit it.
|
|
auto mulLowLow =
|
|
rewriter.create<arith::MulUIExtendedOp>(loc, lhsElem0, rhsElem0);
|
|
Value mulLowHi = rewriter.create<arith::MulIOp>(loc, lhsElem0, rhsElem1);
|
|
Value mulHiLow = rewriter.create<arith::MulIOp>(loc, lhsElem1, rhsElem0);
|
|
|
|
Value resLow = mulLowLow.getLow();
|
|
Value resHi =
|
|
rewriter.create<arith::AddIOp>(loc, mulLowLow.getHigh(), mulLowHi);
|
|
resHi = rewriter.create<arith::AddIOp>(loc, resHi, mulHiLow);
|
|
|
|
Value resultVec =
|
|
constructResultVector(rewriter, loc, newTy, {resLow, resHi});
|
|
rewriter.replaceOp(op, resultVec);
|
|
return success();
|
|
}
|
|
};
|
|
|
|
//===----------------------------------------------------------------------===//
|
|
// ConvertExtSI
|
|
//===----------------------------------------------------------------------===//
|
|
|
|
struct ConvertExtSI final : OpConversionPattern<arith::ExtSIOp> {
|
|
using OpConversionPattern::OpConversionPattern;
|
|
|
|
LogicalResult
|
|
matchAndRewrite(arith::ExtSIOp op, OpAdaptor adaptor,
|
|
ConversionPatternRewriter &rewriter) const override {
|
|
Location loc = op->getLoc();
|
|
auto newTy = getTypeConverter()->convertType<VectorType>(op.getType());
|
|
if (!newTy)
|
|
return rewriter.notifyMatchFailure(
|
|
loc, llvm::formatv("unsupported type: {0}", op.getType()));
|
|
|
|
Type newResultComponentTy = reduceInnermostDim(newTy);
|
|
|
|
// Sign-extend the input value to determine the low half of the result.
|
|
// Then, check if the low half is negative, and sign-extend the comparison
|
|
// result to get the high half.
|
|
Value newOperand = appendX1Dim(rewriter, loc, adaptor.getIn());
|
|
Value extended = rewriter.createOrFold<arith::ExtSIOp>(
|
|
loc, newResultComponentTy, newOperand);
|
|
Value operandZeroCst =
|
|
createScalarOrSplatConstant(rewriter, loc, newResultComponentTy, 0);
|
|
Value signBit = rewriter.create<arith::CmpIOp>(
|
|
loc, arith::CmpIPredicate::slt, extended, operandZeroCst);
|
|
Value signValue =
|
|
rewriter.create<arith::ExtSIOp>(loc, newResultComponentTy, signBit);
|
|
|
|
Value resultVec =
|
|
constructResultVector(rewriter, loc, newTy, {extended, signValue});
|
|
rewriter.replaceOp(op, resultVec);
|
|
return success();
|
|
}
|
|
};
|
|
|
|
//===----------------------------------------------------------------------===//
|
|
// ConvertExtUI
|
|
//===----------------------------------------------------------------------===//
|
|
|
|
struct ConvertExtUI final : OpConversionPattern<arith::ExtUIOp> {
|
|
using OpConversionPattern::OpConversionPattern;
|
|
|
|
LogicalResult
|
|
matchAndRewrite(arith::ExtUIOp op, OpAdaptor adaptor,
|
|
ConversionPatternRewriter &rewriter) const override {
|
|
Location loc = op->getLoc();
|
|
auto newTy = getTypeConverter()->convertType<VectorType>(op.getType());
|
|
if (!newTy)
|
|
return rewriter.notifyMatchFailure(
|
|
loc, llvm::formatv("unsupported type: {0}", op.getType()));
|
|
|
|
Type newResultComponentTy = reduceInnermostDim(newTy);
|
|
|
|
// Zero-extend the input value to determine the low half of the result.
|
|
// The high half is always zero.
|
|
Value newOperand = appendX1Dim(rewriter, loc, adaptor.getIn());
|
|
Value extended = rewriter.createOrFold<arith::ExtUIOp>(
|
|
loc, newResultComponentTy, newOperand);
|
|
Value zeroCst = createScalarOrSplatConstant(rewriter, loc, newTy, 0);
|
|
Value newRes = insertLastDimSlice(rewriter, loc, extended, zeroCst, 0);
|
|
rewriter.replaceOp(op, newRes);
|
|
return success();
|
|
}
|
|
};
|
|
|
|
//===----------------------------------------------------------------------===//
|
|
// ConvertMaxMin
|
|
//===----------------------------------------------------------------------===//
|
|
|
|
template <typename SourceOp, arith::CmpIPredicate CmpPred>
|
|
struct ConvertMaxMin final : OpConversionPattern<SourceOp> {
|
|
using OpConversionPattern<SourceOp>::OpConversionPattern;
|
|
|
|
LogicalResult
|
|
matchAndRewrite(SourceOp op, typename SourceOp::Adaptor adaptor,
|
|
ConversionPatternRewriter &rewriter) const override {
|
|
Location loc = op->getLoc();
|
|
|
|
Type oldTy = op.getType();
|
|
auto newTy = dyn_cast_or_null<VectorType>(
|
|
this->getTypeConverter()->convertType(oldTy));
|
|
if (!newTy)
|
|
return rewriter.notifyMatchFailure(
|
|
loc, llvm::formatv("unsupported type: {0}", op.getType()));
|
|
|
|
// Rewrite Max*I/Min*I as compare and select over original operands. Let
|
|
// the CmpI and Select emulation patterns handle the final legalization.
|
|
Value cmp =
|
|
rewriter.create<arith::CmpIOp>(loc, CmpPred, op.getLhs(), op.getRhs());
|
|
rewriter.replaceOpWithNewOp<arith::SelectOp>(op, cmp, op.getLhs(),
|
|
op.getRhs());
|
|
return success();
|
|
}
|
|
};
|
|
|
|
// Convert IndexCast ops
|
|
//===----------------------------------------------------------------------===//
|
|
|
|
/// Returns true iff the type is `index` or `vector<...index>`.
|
|
static bool isIndexOrIndexVector(Type type) {
|
|
if (isa<IndexType>(type))
|
|
return true;
|
|
|
|
if (auto vectorTy = dyn_cast<VectorType>(type))
|
|
if (isa<IndexType>(vectorTy.getElementType()))
|
|
return true;
|
|
|
|
return false;
|
|
}
|
|
|
|
template <typename CastOp>
|
|
struct ConvertIndexCastIntToIndex final : OpConversionPattern<CastOp> {
|
|
using OpConversionPattern<CastOp>::OpConversionPattern;
|
|
|
|
LogicalResult
|
|
matchAndRewrite(CastOp op, typename CastOp::Adaptor adaptor,
|
|
ConversionPatternRewriter &rewriter) const override {
|
|
Type resultType = op.getType();
|
|
if (!isIndexOrIndexVector(resultType))
|
|
return failure();
|
|
|
|
Location loc = op.getLoc();
|
|
Type inType = op.getIn().getType();
|
|
auto newInTy =
|
|
this->getTypeConverter()->template convertType<VectorType>(inType);
|
|
if (!newInTy)
|
|
return rewriter.notifyMatchFailure(
|
|
loc, llvm::formatv("unsupported type: {0}", inType));
|
|
|
|
// Discard the high half of the input truncating the original value.
|
|
Value extracted = extractLastDimSlice(rewriter, loc, adaptor.getIn(), 0);
|
|
extracted = dropTrailingX1Dim(rewriter, loc, extracted);
|
|
rewriter.replaceOpWithNewOp<CastOp>(op, resultType, extracted);
|
|
return success();
|
|
}
|
|
};
|
|
|
|
template <typename CastOp, typename ExtensionOp>
|
|
struct ConvertIndexCastIndexToInt final : OpConversionPattern<CastOp> {
|
|
using OpConversionPattern<CastOp>::OpConversionPattern;
|
|
|
|
LogicalResult
|
|
matchAndRewrite(CastOp op, typename CastOp::Adaptor adaptor,
|
|
ConversionPatternRewriter &rewriter) const override {
|
|
Type inType = op.getIn().getType();
|
|
if (!isIndexOrIndexVector(inType))
|
|
return failure();
|
|
|
|
Location loc = op.getLoc();
|
|
auto *typeConverter =
|
|
this->template getTypeConverter<arith::WideIntEmulationConverter>();
|
|
|
|
Type resultType = op.getType();
|
|
auto newTy = typeConverter->template convertType<VectorType>(resultType);
|
|
if (!newTy)
|
|
return rewriter.notifyMatchFailure(
|
|
loc, llvm::formatv("unsupported type: {0}", resultType));
|
|
|
|
// Emit an index cast over the matching narrow type.
|
|
Type narrowTy =
|
|
rewriter.getIntegerType(typeConverter->getMaxTargetIntBitWidth());
|
|
if (auto vecTy = dyn_cast<VectorType>(resultType))
|
|
narrowTy = VectorType::get(vecTy.getShape(), narrowTy);
|
|
|
|
// Sign or zero-extend the result. Let the matching conversion pattern
|
|
// legalize the extension op.
|
|
Value underlyingVal =
|
|
rewriter.create<CastOp>(loc, narrowTy, adaptor.getIn());
|
|
rewriter.replaceOpWithNewOp<ExtensionOp>(op, resultType, underlyingVal);
|
|
return success();
|
|
}
|
|
};
|
|
|
|
//===----------------------------------------------------------------------===//
|
|
// ConvertSelect
|
|
//===----------------------------------------------------------------------===//
|
|
|
|
struct ConvertSelect final : OpConversionPattern<arith::SelectOp> {
|
|
using OpConversionPattern::OpConversionPattern;
|
|
|
|
LogicalResult
|
|
matchAndRewrite(arith::SelectOp op, OpAdaptor adaptor,
|
|
ConversionPatternRewriter &rewriter) const override {
|
|
Location loc = op->getLoc();
|
|
auto newTy = getTypeConverter()->convertType<VectorType>(op.getType());
|
|
if (!newTy)
|
|
return rewriter.notifyMatchFailure(
|
|
loc, llvm::formatv("unsupported type: {0}", op.getType()));
|
|
|
|
auto [trueElem0, trueElem1] =
|
|
extractLastDimHalves(rewriter, loc, adaptor.getTrueValue());
|
|
auto [falseElem0, falseElem1] =
|
|
extractLastDimHalves(rewriter, loc, adaptor.getFalseValue());
|
|
Value cond = appendX1Dim(rewriter, loc, adaptor.getCondition());
|
|
|
|
Value resElem0 =
|
|
rewriter.create<arith::SelectOp>(loc, cond, trueElem0, falseElem0);
|
|
Value resElem1 =
|
|
rewriter.create<arith::SelectOp>(loc, cond, trueElem1, falseElem1);
|
|
Value resultVec =
|
|
constructResultVector(rewriter, loc, newTy, {resElem0, resElem1});
|
|
rewriter.replaceOp(op, resultVec);
|
|
return success();
|
|
}
|
|
};
|
|
|
|
//===----------------------------------------------------------------------===//
|
|
// ConvertShLI
|
|
//===----------------------------------------------------------------------===//
|
|
|
|
struct ConvertShLI final : OpConversionPattern<arith::ShLIOp> {
|
|
using OpConversionPattern::OpConversionPattern;
|
|
|
|
LogicalResult
|
|
matchAndRewrite(arith::ShLIOp op, OpAdaptor adaptor,
|
|
ConversionPatternRewriter &rewriter) const override {
|
|
Location loc = op->getLoc();
|
|
|
|
Type oldTy = op.getType();
|
|
auto newTy = getTypeConverter()->convertType<VectorType>(oldTy);
|
|
if (!newTy)
|
|
return rewriter.notifyMatchFailure(
|
|
loc, llvm::formatv("unsupported type: {0}", op.getType()));
|
|
|
|
Type newOperandTy = reduceInnermostDim(newTy);
|
|
// `oldBitWidth` == `2 * newBitWidth`
|
|
unsigned newBitWidth = newTy.getElementTypeBitWidth();
|
|
|
|
auto [lhsElem0, lhsElem1] =
|
|
extractLastDimHalves(rewriter, loc, adaptor.getLhs());
|
|
Value rhsElem0 = extractLastDimSlice(rewriter, loc, adaptor.getRhs(), 0);
|
|
|
|
// Assume that the shift amount is < 2 * newBitWidth. Calculate the low and
|
|
// high halves of the results separately:
|
|
// 1. low := LHS.low shli RHS
|
|
//
|
|
// 2. high := a or b or c, where:
|
|
// a) Bits from LHS.high, shifted by the RHS.
|
|
// b) Bits from LHS.low, shifted right. These come into play when
|
|
// RHS < newBitWidth, e.g.:
|
|
// [0000][llll] shli 3 --> [0lll][l000]
|
|
// ^
|
|
// |
|
|
// [llll] shrui (4 - 3)
|
|
// c) Bits from LHS.low, shifted left. These matter when
|
|
// RHS > newBitWidth, e.g.:
|
|
// [0000][llll] shli 7 --> [l000][0000]
|
|
// ^
|
|
// |
|
|
// [llll] shli (7 - 4)
|
|
//
|
|
// Because shifts by values >= newBitWidth are undefined, we ignore the high
|
|
// half of RHS, and introduce 'bounds checks' to account for
|
|
// RHS.low > newBitWidth.
|
|
//
|
|
// TODO: Explore possible optimizations.
|
|
Value zeroCst = createScalarOrSplatConstant(rewriter, loc, newOperandTy, 0);
|
|
Value elemBitWidth =
|
|
createScalarOrSplatConstant(rewriter, loc, newOperandTy, newBitWidth);
|
|
|
|
Value illegalElemShift = rewriter.create<arith::CmpIOp>(
|
|
loc, arith::CmpIPredicate::uge, rhsElem0, elemBitWidth);
|
|
|
|
Value shiftedElem0 =
|
|
rewriter.create<arith::ShLIOp>(loc, lhsElem0, rhsElem0);
|
|
Value resElem0 = rewriter.create<arith::SelectOp>(loc, illegalElemShift,
|
|
zeroCst, shiftedElem0);
|
|
|
|
Value cappedShiftAmount = rewriter.create<arith::SelectOp>(
|
|
loc, illegalElemShift, elemBitWidth, rhsElem0);
|
|
Value rightShiftAmount =
|
|
rewriter.create<arith::SubIOp>(loc, elemBitWidth, cappedShiftAmount);
|
|
Value shiftedRight =
|
|
rewriter.create<arith::ShRUIOp>(loc, lhsElem0, rightShiftAmount);
|
|
Value overshotShiftAmount =
|
|
rewriter.create<arith::SubIOp>(loc, rhsElem0, elemBitWidth);
|
|
Value shiftedLeft =
|
|
rewriter.create<arith::ShLIOp>(loc, lhsElem0, overshotShiftAmount);
|
|
|
|
Value shiftedElem1 =
|
|
rewriter.create<arith::ShLIOp>(loc, lhsElem1, rhsElem0);
|
|
Value resElem1High = rewriter.create<arith::SelectOp>(
|
|
loc, illegalElemShift, zeroCst, shiftedElem1);
|
|
Value resElem1Low = rewriter.create<arith::SelectOp>(
|
|
loc, illegalElemShift, shiftedLeft, shiftedRight);
|
|
Value resElem1 =
|
|
rewriter.create<arith::OrIOp>(loc, resElem1Low, resElem1High);
|
|
|
|
Value resultVec =
|
|
constructResultVector(rewriter, loc, newTy, {resElem0, resElem1});
|
|
rewriter.replaceOp(op, resultVec);
|
|
return success();
|
|
}
|
|
};
|
|
|
|
//===----------------------------------------------------------------------===//
|
|
// ConvertShRUI
|
|
//===----------------------------------------------------------------------===//
|
|
|
|
struct ConvertShRUI final : OpConversionPattern<arith::ShRUIOp> {
|
|
using OpConversionPattern::OpConversionPattern;
|
|
|
|
LogicalResult
|
|
matchAndRewrite(arith::ShRUIOp op, OpAdaptor adaptor,
|
|
ConversionPatternRewriter &rewriter) const override {
|
|
Location loc = op->getLoc();
|
|
|
|
Type oldTy = op.getType();
|
|
auto newTy = getTypeConverter()->convertType<VectorType>(oldTy);
|
|
if (!newTy)
|
|
return rewriter.notifyMatchFailure(
|
|
loc, llvm::formatv("unsupported type: {0}", op.getType()));
|
|
|
|
Type newOperandTy = reduceInnermostDim(newTy);
|
|
// `oldBitWidth` == `2 * newBitWidth`
|
|
unsigned newBitWidth = newTy.getElementTypeBitWidth();
|
|
|
|
auto [lhsElem0, lhsElem1] =
|
|
extractLastDimHalves(rewriter, loc, adaptor.getLhs());
|
|
Value rhsElem0 = extractLastDimSlice(rewriter, loc, adaptor.getRhs(), 0);
|
|
|
|
// Assume that the shift amount is < 2 * newBitWidth. Calculate the low and
|
|
// high halves of the results separately:
|
|
// 1. low := a or b or c, where:
|
|
// a) Bits from LHS.low, shifted by the RHS.
|
|
// b) Bits from LHS.high, shifted left. These matter when
|
|
// RHS < newBitWidth, e.g.:
|
|
// [hhhh][0000] shrui 3 --> [000h][hhh0]
|
|
// ^
|
|
// |
|
|
// [hhhh] shli (4 - 1)
|
|
// c) Bits from LHS.high, shifted right. These come into play when
|
|
// RHS > newBitWidth, e.g.:
|
|
// [hhhh][0000] shrui 7 --> [0000][000h]
|
|
// ^
|
|
// |
|
|
// [hhhh] shrui (7 - 4)
|
|
//
|
|
// 2. high := LHS.high shrui RHS
|
|
//
|
|
// Because shifts by values >= newBitWidth are undefined, we ignore the high
|
|
// half of RHS, and introduce 'bounds checks' to account for
|
|
// RHS.low > newBitWidth.
|
|
//
|
|
// TODO: Explore possible optimizations.
|
|
Value zeroCst = createScalarOrSplatConstant(rewriter, loc, newOperandTy, 0);
|
|
Value elemBitWidth =
|
|
createScalarOrSplatConstant(rewriter, loc, newOperandTy, newBitWidth);
|
|
|
|
Value illegalElemShift = rewriter.create<arith::CmpIOp>(
|
|
loc, arith::CmpIPredicate::uge, rhsElem0, elemBitWidth);
|
|
|
|
Value shiftedElem0 =
|
|
rewriter.create<arith::ShRUIOp>(loc, lhsElem0, rhsElem0);
|
|
Value resElem0Low = rewriter.create<arith::SelectOp>(loc, illegalElemShift,
|
|
zeroCst, shiftedElem0);
|
|
Value shiftedElem1 =
|
|
rewriter.create<arith::ShRUIOp>(loc, lhsElem1, rhsElem0);
|
|
Value resElem1 = rewriter.create<arith::SelectOp>(loc, illegalElemShift,
|
|
zeroCst, shiftedElem1);
|
|
|
|
Value cappedShiftAmount = rewriter.create<arith::SelectOp>(
|
|
loc, illegalElemShift, elemBitWidth, rhsElem0);
|
|
Value leftShiftAmount =
|
|
rewriter.create<arith::SubIOp>(loc, elemBitWidth, cappedShiftAmount);
|
|
Value shiftedLeft =
|
|
rewriter.create<arith::ShLIOp>(loc, lhsElem1, leftShiftAmount);
|
|
Value overshotShiftAmount =
|
|
rewriter.create<arith::SubIOp>(loc, rhsElem0, elemBitWidth);
|
|
Value shiftedRight =
|
|
rewriter.create<arith::ShRUIOp>(loc, lhsElem1, overshotShiftAmount);
|
|
|
|
Value resElem0High = rewriter.create<arith::SelectOp>(
|
|
loc, illegalElemShift, shiftedRight, shiftedLeft);
|
|
Value resElem0 =
|
|
rewriter.create<arith::OrIOp>(loc, resElem0Low, resElem0High);
|
|
|
|
Value resultVec =
|
|
constructResultVector(rewriter, loc, newTy, {resElem0, resElem1});
|
|
rewriter.replaceOp(op, resultVec);
|
|
return success();
|
|
}
|
|
};
|
|
|
|
//===----------------------------------------------------------------------===//
|
|
// ConvertShRSI
|
|
//===----------------------------------------------------------------------===//
|
|
|
|
struct ConvertShRSI final : OpConversionPattern<arith::ShRSIOp> {
|
|
using OpConversionPattern::OpConversionPattern;
|
|
|
|
LogicalResult
|
|
matchAndRewrite(arith::ShRSIOp op, OpAdaptor adaptor,
|
|
ConversionPatternRewriter &rewriter) const override {
|
|
Location loc = op->getLoc();
|
|
|
|
Type oldTy = op.getType();
|
|
auto newTy = getTypeConverter()->convertType<VectorType>(oldTy);
|
|
if (!newTy)
|
|
return rewriter.notifyMatchFailure(
|
|
loc, llvm::formatv("unsupported type: {0}", op.getType()));
|
|
|
|
Value lhsElem1 = extractLastDimSlice(rewriter, loc, adaptor.getLhs(), 1);
|
|
Value rhsElem0 = extractLastDimSlice(rewriter, loc, adaptor.getRhs(), 0);
|
|
|
|
Type narrowTy = rhsElem0.getType();
|
|
int64_t origBitwidth = newTy.getElementTypeBitWidth() * 2;
|
|
|
|
// Rewrite this as an bitwise or of `arith.shrui` and sign extension bits.
|
|
// Perform as many ops over the narrow integer type as possible and let the
|
|
// other emulation patterns convert the rest.
|
|
Value elemZero = createScalarOrSplatConstant(rewriter, loc, narrowTy, 0);
|
|
Value signBit = rewriter.create<arith::CmpIOp>(
|
|
loc, arith::CmpIPredicate::slt, lhsElem1, elemZero);
|
|
signBit = dropTrailingX1Dim(rewriter, loc, signBit);
|
|
|
|
// Create a bit pattern of either all ones or all zeros. Then shift it left
|
|
// to calculate the sign extension bits created by shifting the original
|
|
// sign bit right.
|
|
Value allSign = rewriter.create<arith::ExtSIOp>(loc, oldTy, signBit);
|
|
Value maxShift =
|
|
createScalarOrSplatConstant(rewriter, loc, narrowTy, origBitwidth);
|
|
Value numNonSignExtBits =
|
|
rewriter.create<arith::SubIOp>(loc, maxShift, rhsElem0);
|
|
numNonSignExtBits = dropTrailingX1Dim(rewriter, loc, numNonSignExtBits);
|
|
numNonSignExtBits =
|
|
rewriter.create<arith::ExtUIOp>(loc, oldTy, numNonSignExtBits);
|
|
Value signBits =
|
|
rewriter.create<arith::ShLIOp>(loc, allSign, numNonSignExtBits);
|
|
|
|
// Use original arguments to create the right shift.
|
|
Value shrui =
|
|
rewriter.create<arith::ShRUIOp>(loc, op.getLhs(), op.getRhs());
|
|
Value shrsi = rewriter.create<arith::OrIOp>(loc, shrui, signBits);
|
|
|
|
// Handle shifting by zero. This is necessary when the `signBits` shift is
|
|
// invalid.
|
|
Value isNoop = rewriter.create<arith::CmpIOp>(loc, arith::CmpIPredicate::eq,
|
|
rhsElem0, elemZero);
|
|
isNoop = dropTrailingX1Dim(rewriter, loc, isNoop);
|
|
rewriter.replaceOpWithNewOp<arith::SelectOp>(op, isNoop, op.getLhs(),
|
|
shrsi);
|
|
|
|
return success();
|
|
}
|
|
};
|
|
|
|
//===----------------------------------------------------------------------===//
|
|
// ConvertSIToFP
|
|
//===----------------------------------------------------------------------===//
|
|
|
|
struct ConvertSIToFP final : OpConversionPattern<arith::SIToFPOp> {
|
|
using OpConversionPattern::OpConversionPattern;
|
|
|
|
LogicalResult
|
|
matchAndRewrite(arith::SIToFPOp op, OpAdaptor adaptor,
|
|
ConversionPatternRewriter &rewriter) const override {
|
|
Location loc = op.getLoc();
|
|
|
|
Value in = op.getIn();
|
|
Type oldTy = in.getType();
|
|
auto newTy = getTypeConverter()->convertType<VectorType>(oldTy);
|
|
if (!newTy)
|
|
return rewriter.notifyMatchFailure(
|
|
loc, llvm::formatv("unsupported type: {0}", oldTy));
|
|
|
|
unsigned oldBitWidth = getElementTypeOrSelf(oldTy).getIntOrFloatBitWidth();
|
|
Value zeroCst = createScalarOrSplatConstant(rewriter, loc, oldTy, 0);
|
|
Value oneCst = createScalarOrSplatConstant(rewriter, loc, oldTy, 1);
|
|
Value allOnesCst = createScalarOrSplatConstant(
|
|
rewriter, loc, oldTy, APInt::getAllOnes(oldBitWidth));
|
|
|
|
// To avoid operating on very large unsigned numbers, perform the
|
|
// conversion on the absolute value. Then, decide whether to negate the
|
|
// result or not based on that sign bit. We assume two's complement and
|
|
// implement negation by flipping all bits and adding 1.
|
|
// Note that this relies on the the other conversion patterns to legalize
|
|
// created ops and narrow the bit widths.
|
|
Value isNeg = rewriter.create<arith::CmpIOp>(loc, arith::CmpIPredicate::slt,
|
|
in, zeroCst);
|
|
Value bitwiseNeg = rewriter.create<arith::XOrIOp>(loc, in, allOnesCst);
|
|
Value neg = rewriter.create<arith::AddIOp>(loc, bitwiseNeg, oneCst);
|
|
Value abs = rewriter.create<arith::SelectOp>(loc, isNeg, neg, in);
|
|
|
|
Value absResult = rewriter.create<arith::UIToFPOp>(loc, op.getType(), abs);
|
|
Value negResult = rewriter.create<arith::NegFOp>(loc, absResult);
|
|
rewriter.replaceOpWithNewOp<arith::SelectOp>(op, isNeg, negResult,
|
|
absResult);
|
|
return success();
|
|
}
|
|
};
|
|
|
|
//===----------------------------------------------------------------------===//
|
|
// ConvertUIToFP
|
|
//===----------------------------------------------------------------------===//
|
|
|
|
struct ConvertUIToFP final : OpConversionPattern<arith::UIToFPOp> {
|
|
using OpConversionPattern::OpConversionPattern;
|
|
|
|
LogicalResult
|
|
matchAndRewrite(arith::UIToFPOp op, OpAdaptor adaptor,
|
|
ConversionPatternRewriter &rewriter) const override {
|
|
Location loc = op.getLoc();
|
|
|
|
Type oldTy = op.getIn().getType();
|
|
auto newTy = getTypeConverter()->convertType<VectorType>(oldTy);
|
|
if (!newTy)
|
|
return rewriter.notifyMatchFailure(
|
|
loc, llvm::formatv("unsupported type: {0}", oldTy));
|
|
unsigned newBitWidth = newTy.getElementTypeBitWidth();
|
|
|
|
auto [low, hi] = extractLastDimHalves(rewriter, loc, adaptor.getIn());
|
|
Value lowInt = dropTrailingX1Dim(rewriter, loc, low);
|
|
Value hiInt = dropTrailingX1Dim(rewriter, loc, hi);
|
|
Value zeroCst =
|
|
createScalarOrSplatConstant(rewriter, loc, hiInt.getType(), 0);
|
|
|
|
// The final result has the following form:
|
|
// if (hi == 0) return uitofp(low)
|
|
// else return uitofp(low) + uitofp(hi) * 2^BW
|
|
//
|
|
// where `BW` is the bitwidth of the narrowed integer type. We emit a
|
|
// select to make it easier to fold-away the `hi` part calculation when it
|
|
// is known to be zero.
|
|
//
|
|
// Note 1: The emulation is precise only for input values that have exact
|
|
// integer representation in the result floating point type, and may lead
|
|
// loss of precision otherwise.
|
|
//
|
|
// Note 2: We do not strictly need the `hi == 0`, case, but it makes
|
|
// constant folding easier.
|
|
Value hiEqZero = rewriter.create<arith::CmpIOp>(
|
|
loc, arith::CmpIPredicate::eq, hiInt, zeroCst);
|
|
|
|
Type resultTy = op.getType();
|
|
Type resultElemTy = getElementTypeOrSelf(resultTy);
|
|
Value lowFp = rewriter.create<arith::UIToFPOp>(loc, resultTy, lowInt);
|
|
Value hiFp = rewriter.create<arith::UIToFPOp>(loc, resultTy, hiInt);
|
|
|
|
int64_t pow2Int = int64_t(1) << newBitWidth;
|
|
TypedAttr pow2Attr =
|
|
rewriter.getFloatAttr(resultElemTy, static_cast<double>(pow2Int));
|
|
if (auto vecTy = dyn_cast<VectorType>(resultTy))
|
|
pow2Attr = SplatElementsAttr::get(vecTy, pow2Attr);
|
|
|
|
Value pow2Val = rewriter.create<arith::ConstantOp>(loc, resultTy, pow2Attr);
|
|
|
|
Value hiVal = rewriter.create<arith::MulFOp>(loc, hiFp, pow2Val);
|
|
Value result = rewriter.create<arith::AddFOp>(loc, lowFp, hiVal);
|
|
|
|
rewriter.replaceOpWithNewOp<arith::SelectOp>(op, hiEqZero, lowFp, result);
|
|
return success();
|
|
}
|
|
};
|
|
|
|
//===----------------------------------------------------------------------===//
|
|
// ConvertTruncI
|
|
//===----------------------------------------------------------------------===//
|
|
|
|
struct ConvertTruncI final : OpConversionPattern<arith::TruncIOp> {
|
|
using OpConversionPattern::OpConversionPattern;
|
|
|
|
LogicalResult
|
|
matchAndRewrite(arith::TruncIOp op, OpAdaptor adaptor,
|
|
ConversionPatternRewriter &rewriter) const override {
|
|
Location loc = op.getLoc();
|
|
// Check if the result type is legal for this target. Currently, we do not
|
|
// support truncation to types wider than supported by the target.
|
|
if (!getTypeConverter()->isLegal(op.getType()))
|
|
return rewriter.notifyMatchFailure(
|
|
loc, llvm::formatv("unsupported truncation result type: {0}",
|
|
op.getType()));
|
|
|
|
// Discard the high half of the input. Truncate the low half, if
|
|
// necessary.
|
|
Value extracted = extractLastDimSlice(rewriter, loc, adaptor.getIn(), 0);
|
|
extracted = dropTrailingX1Dim(rewriter, loc, extracted);
|
|
Value truncated =
|
|
rewriter.createOrFold<arith::TruncIOp>(loc, op.getType(), extracted);
|
|
rewriter.replaceOp(op, truncated);
|
|
return success();
|
|
}
|
|
};
|
|
|
|
//===----------------------------------------------------------------------===//
|
|
// ConvertVectorPrint
|
|
//===----------------------------------------------------------------------===//
|
|
|
|
struct ConvertVectorPrint final : OpConversionPattern<vector::PrintOp> {
|
|
using OpConversionPattern::OpConversionPattern;
|
|
|
|
LogicalResult
|
|
matchAndRewrite(vector::PrintOp op, OpAdaptor adaptor,
|
|
ConversionPatternRewriter &rewriter) const override {
|
|
rewriter.replaceOpWithNewOp<vector::PrintOp>(op, adaptor.getSource());
|
|
return success();
|
|
}
|
|
};
|
|
|
|
//===----------------------------------------------------------------------===//
|
|
// Pass Definition
|
|
//===----------------------------------------------------------------------===//
|
|
|
|
struct EmulateWideIntPass final
|
|
: arith::impl::ArithEmulateWideIntBase<EmulateWideIntPass> {
|
|
using ArithEmulateWideIntBase::ArithEmulateWideIntBase;
|
|
|
|
void runOnOperation() override {
|
|
if (!llvm::isPowerOf2_32(widestIntSupported) || widestIntSupported < 2) {
|
|
signalPassFailure();
|
|
return;
|
|
}
|
|
|
|
Operation *op = getOperation();
|
|
MLIRContext *ctx = op->getContext();
|
|
|
|
arith::WideIntEmulationConverter typeConverter(widestIntSupported);
|
|
ConversionTarget target(*ctx);
|
|
target.addDynamicallyLegalOp<func::FuncOp>([&typeConverter](Operation *op) {
|
|
return typeConverter.isLegal(cast<func::FuncOp>(op).getFunctionType());
|
|
});
|
|
auto opLegalCallback = [&typeConverter](Operation *op) {
|
|
return typeConverter.isLegal(op);
|
|
};
|
|
target.addDynamicallyLegalOp<func::CallOp, func::ReturnOp>(opLegalCallback);
|
|
target
|
|
.addDynamicallyLegalDialect<arith::ArithDialect, vector::VectorDialect>(
|
|
opLegalCallback);
|
|
|
|
RewritePatternSet patterns(ctx);
|
|
arith::populateArithWideIntEmulationPatterns(typeConverter, patterns);
|
|
|
|
if (failed(applyPartialConversion(op, target, std::move(patterns))))
|
|
signalPassFailure();
|
|
}
|
|
};
|
|
} // end anonymous namespace
|
|
|
|
//===----------------------------------------------------------------------===//
|
|
// Public Interface Definition
|
|
//===----------------------------------------------------------------------===//
|
|
|
|
arith::WideIntEmulationConverter::WideIntEmulationConverter(
|
|
unsigned widestIntSupportedByTarget)
|
|
: maxIntWidth(widestIntSupportedByTarget) {
|
|
assert(llvm::isPowerOf2_32(widestIntSupportedByTarget) &&
|
|
"Only power-of-two integers with are supported");
|
|
assert(widestIntSupportedByTarget >= 2 && "Integer type too narrow");
|
|
|
|
// Allow unknown types.
|
|
addConversion([](Type ty) -> std::optional<Type> { return ty; });
|
|
|
|
// Scalar case.
|
|
addConversion([this](IntegerType ty) -> std::optional<Type> {
|
|
unsigned width = ty.getWidth();
|
|
if (width <= maxIntWidth)
|
|
return ty;
|
|
|
|
// i2N --> vector<2xiN>
|
|
if (width == 2 * maxIntWidth)
|
|
return VectorType::get(2, IntegerType::get(ty.getContext(), maxIntWidth));
|
|
|
|
return std::nullopt;
|
|
});
|
|
|
|
// Vector case.
|
|
addConversion([this](VectorType ty) -> std::optional<Type> {
|
|
auto intTy = dyn_cast<IntegerType>(ty.getElementType());
|
|
if (!intTy)
|
|
return ty;
|
|
|
|
unsigned width = intTy.getWidth();
|
|
if (width <= maxIntWidth)
|
|
return ty;
|
|
|
|
// vector<...xi2N> --> vector<...x2xiN>
|
|
if (width == 2 * maxIntWidth) {
|
|
auto newShape = to_vector(ty.getShape());
|
|
newShape.push_back(2);
|
|
return VectorType::get(newShape,
|
|
IntegerType::get(ty.getContext(), maxIntWidth));
|
|
}
|
|
|
|
return std::nullopt;
|
|
});
|
|
|
|
// Function case.
|
|
addConversion([this](FunctionType ty) -> std::optional<Type> {
|
|
// Convert inputs and results, e.g.:
|
|
// (i2N, i2N) -> i2N --> (vector<2xiN>, vector<2xiN>) -> vector<2xiN>
|
|
SmallVector<Type> inputs;
|
|
if (failed(convertTypes(ty.getInputs(), inputs)))
|
|
return std::nullopt;
|
|
|
|
SmallVector<Type> results;
|
|
if (failed(convertTypes(ty.getResults(), results)))
|
|
return std::nullopt;
|
|
|
|
return FunctionType::get(ty.getContext(), inputs, results);
|
|
});
|
|
}
|
|
|
|
void arith::populateArithWideIntEmulationPatterns(
|
|
WideIntEmulationConverter &typeConverter, RewritePatternSet &patterns) {
|
|
// Populate `func.*` conversion patterns.
|
|
populateFunctionOpInterfaceTypeConversionPattern<func::FuncOp>(patterns,
|
|
typeConverter);
|
|
populateCallOpTypeConversionPattern(patterns, typeConverter);
|
|
populateReturnOpTypeConversionPattern(patterns, typeConverter);
|
|
|
|
// Populate `arith.*` conversion patterns.
|
|
patterns.add<
|
|
// Misc ops.
|
|
ConvertConstant, ConvertCmpI, ConvertSelect, ConvertVectorPrint,
|
|
// Binary ops.
|
|
ConvertAddI, ConvertMulI, ConvertShLI, ConvertShRSI, ConvertShRUI,
|
|
ConvertMaxMin<arith::MaxUIOp, arith::CmpIPredicate::ugt>,
|
|
ConvertMaxMin<arith::MaxSIOp, arith::CmpIPredicate::sgt>,
|
|
ConvertMaxMin<arith::MinUIOp, arith::CmpIPredicate::ult>,
|
|
ConvertMaxMin<arith::MinSIOp, arith::CmpIPredicate::slt>,
|
|
// Bitwise binary ops.
|
|
ConvertBitwiseBinary<arith::AndIOp>, ConvertBitwiseBinary<arith::OrIOp>,
|
|
ConvertBitwiseBinary<arith::XOrIOp>,
|
|
// Extension and truncation ops.
|
|
ConvertExtSI, ConvertExtUI, ConvertTruncI,
|
|
// Cast ops.
|
|
ConvertIndexCastIntToIndex<arith::IndexCastOp>,
|
|
ConvertIndexCastIntToIndex<arith::IndexCastUIOp>,
|
|
ConvertIndexCastIndexToInt<arith::IndexCastOp, arith::ExtSIOp>,
|
|
ConvertIndexCastIndexToInt<arith::IndexCastUIOp, arith::ExtUIOp>,
|
|
ConvertSIToFP, ConvertUIToFP>(typeConverter, patterns.getContext());
|
|
}
|