[mlir][Transforms] Add 1:N matchAndRewrite overload
#116470
Open
Conversation
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
matthias-springer
force-pushed
the
users/matthias-springer/1n_pattern
branch
3 times, most recently
from
de3b79f
to
fe38d4b
Compare
Apply suggestions from code review Co-authored-by: Markus Böck <[email protected]> address comments [WIP] 1:N conversion pattern update test cases
matthias-springer
force-pushed
the
users/matthias-springer/1n_pattern
branch
from
fe38d4b
to
153310a
Compare
matthias-springer
requested review from
aartbik,
PeimingLiu and
yinying-lisa-li
as code owners
llvmbot
added
mlir:core
|
@llvm/pr-subscribers-mlir-scf @llvm/pr-subscribers-mlir-core Author: Matthias Springer (matthias-springer) ChangesThis commit adds a new The existing LogicalResult ConversionPattern::matchAndRewrite(
Operation *op, ArrayRef<Value> operands /*adaptor values*/,
ConversionPatternRewriter &rewriter) const;This commit adds a LogicalResult ConversionPattern::matchAndRewrite(
Operation *op, ArrayRef<ValueRange> operands /*adaptor values*/,
ConversionPatternRewriter &rewriter) const {
// Note: getOneToOneAdaptorOperands produces a fatal error if at least one
// ValueRange has 0 or more than 1 value.
return matchAndRewrite(op, getOneToOneAdaptorOperands(operands), rewriter);
}The Before handing the adaptor values to a This commit also updates the the "decompose call graphs" transformation and the "sparse tensor codegen" transformation to use the new 1:N Note for LLVM conversion: If you are using a type converter with 1:N type conversion rules or if your patterns are performing 1:N replacements (via Patch is 70.77 KiB, truncated to 20.00 KiB below, full version: https://github.com/llvm/llvm-project/pull/116470.diff 9 Files Affected:
diff --git a/mlir/include/mlir/Conversion/LLVMCommon/Pattern.h b/mlir/include/mlir/Conversion/LLVMCommon/Pattern.h
index f3bf5b66398e09..86ea87b55af1cd 100644
--- a/mlir/include/mlir/Conversion/LLVMCommon/Pattern.h
+++ b/mlir/include/mlir/Conversion/LLVMCommon/Pattern.h
@@ -143,6 +143,8 @@ template <typename SourceOp>
class ConvertOpToLLVMPattern : public ConvertToLLVMPattern {
public:
using OpAdaptor = typename SourceOp::Adaptor;
+ using OneToNOpAdaptor =
+ typename SourceOp::template GenericAdaptor<ArrayRef<ValueRange>>;
explicit ConvertOpToLLVMPattern(const LLVMTypeConverter &typeConverter,
PatternBenefit benefit = 1)
@@ -153,8 +155,13 @@ class ConvertOpToLLVMPattern : public ConvertToLLVMPattern {
/// Wrappers around the RewritePattern methods that pass the derived op type.
void rewrite(Operation *op, ArrayRef<Value> operands,
ConversionPatternRewriter &rewriter) const final {
- rewrite(cast<SourceOp>(op), OpAdaptor(operands, cast<SourceOp>(op)),
- rewriter);
+ auto sourceOp = cast<SourceOp>(op);
+ rewrite(sourceOp, OpAdaptor(operands, sourceOp), rewriter);
+ }
+ void rewrite(Operation *op, ArrayRef<ValueRange> operands,
+ ConversionPatternRewriter &rewriter) const final {
+ auto sourceOp = cast<SourceOp>(op);
+ rewrite(sourceOp, OneToNOpAdaptor(operands, sourceOp), rewriter);
}
LogicalResult match(Operation *op) const final {
return match(cast<SourceOp>(op));
@@ -162,8 +169,15 @@ class ConvertOpToLLVMPattern : public ConvertToLLVMPattern {
LogicalResult
matchAndRewrite(Operation *op, ArrayRef<Value> operands,
ConversionPatternRewriter &rewriter) const final {
- return matchAndRewrite(cast<SourceOp>(op),
- OpAdaptor(operands, cast<SourceOp>(op)), rewriter);
+ auto sourceOp = cast<SourceOp>(op);
+ return matchAndRewrite(sourceOp, OpAdaptor(operands, sourceOp), rewriter);
+ }
+ LogicalResult
+ matchAndRewrite(Operation *op, ArrayRef<ValueRange> operands,
+ ConversionPatternRewriter &rewriter) const final {
+ auto sourceOp = cast<SourceOp>(op);
+ return matchAndRewrite(sourceOp, OneToNOpAdaptor(operands, sourceOp),
+ rewriter);
}
/// Rewrite and Match methods that operate on the SourceOp type. These must be
@@ -175,6 +189,12 @@ class ConvertOpToLLVMPattern : public ConvertToLLVMPattern {
ConversionPatternRewriter &rewriter) const {
llvm_unreachable("must override rewrite or matchAndRewrite");
}
+ virtual void rewrite(SourceOp op, OneToNOpAdaptor adaptor,
+ ConversionPatternRewriter &rewriter) const {
+ SmallVector<Value> oneToOneOperands =
+ getOneToOneAdaptorOperands(adaptor.getOperands());
+ rewrite(op, OpAdaptor(oneToOneOperands, adaptor), rewriter);
+ }
virtual LogicalResult
matchAndRewrite(SourceOp op, OpAdaptor adaptor,
ConversionPatternRewriter &rewriter) const {
@@ -183,6 +203,13 @@ class ConvertOpToLLVMPattern : public ConvertToLLVMPattern {
rewrite(op, adaptor, rewriter);
return success();
}
+ virtual LogicalResult
+ matchAndRewrite(SourceOp op, OneToNOpAdaptor adaptor,
+ ConversionPatternRewriter &rewriter) const {
+ SmallVector<Value> oneToOneOperands =
+ getOneToOneAdaptorOperands(adaptor.getOperands());
+ return matchAndRewrite(op, OpAdaptor(oneToOneOperands, adaptor), rewriter);
+ }
private:
using ConvertToLLVMPattern::match;
diff --git a/mlir/include/mlir/Transforms/DialectConversion.h b/mlir/include/mlir/Transforms/DialectConversion.h
index de47765006f81e..e4eeb39b9c0741 100644
--- a/mlir/include/mlir/Transforms/DialectConversion.h
+++ b/mlir/include/mlir/Transforms/DialectConversion.h
@@ -537,6 +537,10 @@ class ConversionPattern : public RewritePattern {
ConversionPatternRewriter &rewriter) const {
llvm_unreachable("unimplemented rewrite");
}
+ virtual void rewrite(Operation *op, ArrayRef<ValueRange> operands,
+ ConversionPatternRewriter &rewriter) const {
+ rewrite(op, getOneToOneAdaptorOperands(operands), rewriter);
+ }
/// Hook for derived classes to implement combined matching and rewriting.
virtual LogicalResult
@@ -547,6 +551,11 @@ class ConversionPattern : public RewritePattern {
rewrite(op, operands, rewriter);
return success();
}
+ virtual LogicalResult
+ matchAndRewrite(Operation *op, ArrayRef<ValueRange> operands,
+ ConversionPatternRewriter &rewriter) const {
+ return matchAndRewrite(op, getOneToOneAdaptorOperands(operands), rewriter);
+ }
/// Attempt to match and rewrite the IR root at the specified operation.
LogicalResult matchAndRewrite(Operation *op,
@@ -574,6 +583,15 @@ class ConversionPattern : public RewritePattern {
: RewritePattern(std::forward<Args>(args)...),
typeConverter(&typeConverter) {}
+ /// Given an array of value ranges, which are the inputs to a 1:N adaptor,
+ /// try to extract the single value of each range to construct a the inputs
+ /// for a 1:1 adaptor.
+ ///
+ /// This function produces a fatal error if at least one range has 0 or
+ /// more than 1 value: "pattern 'name' does not support 1:N conversion"
+ SmallVector<Value>
+ getOneToOneAdaptorOperands(ArrayRef<ValueRange> operands) const;
+
protected:
/// An optional type converter for use by this pattern.
const TypeConverter *typeConverter = nullptr;
@@ -589,6 +607,8 @@ template <typename SourceOp>
class OpConversionPattern : public ConversionPattern {
public:
using OpAdaptor = typename SourceOp::Adaptor;
+ using OneToNOpAdaptor =
+ typename SourceOp::template GenericAdaptor<ArrayRef<ValueRange>>;
OpConversionPattern(MLIRContext *context, PatternBenefit benefit = 1)
: ConversionPattern(SourceOp::getOperationName(), benefit, context) {}
@@ -607,12 +627,24 @@ class OpConversionPattern : public ConversionPattern {
auto sourceOp = cast<SourceOp>(op);
rewrite(sourceOp, OpAdaptor(operands, sourceOp), rewriter);
}
+ void rewrite(Operation *op, ArrayRef<ValueRange> operands,
+ ConversionPatternRewriter &rewriter) const final {
+ auto sourceOp = cast<SourceOp>(op);
+ rewrite(sourceOp, OneToNOpAdaptor(operands, sourceOp), rewriter);
+ }
LogicalResult
matchAndRewrite(Operation *op, ArrayRef<Value> operands,
ConversionPatternRewriter &rewriter) const final {
auto sourceOp = cast<SourceOp>(op);
return matchAndRewrite(sourceOp, OpAdaptor(operands, sourceOp), rewriter);
}
+ LogicalResult
+ matchAndRewrite(Operation *op, ArrayRef<ValueRange> operands,
+ ConversionPatternRewriter &rewriter) const final {
+ auto sourceOp = cast<SourceOp>(op);
+ return matchAndRewrite(sourceOp, OneToNOpAdaptor(operands, sourceOp),
+ rewriter);
+ }
/// Rewrite and Match methods that operate on the SourceOp type. These must be
/// overridden by the derived pattern class.
@@ -623,6 +655,12 @@ class OpConversionPattern : public ConversionPattern {
ConversionPatternRewriter &rewriter) const {
llvm_unreachable("must override matchAndRewrite or a rewrite method");
}
+ virtual void rewrite(SourceOp op, OneToNOpAdaptor adaptor,
+ ConversionPatternRewriter &rewriter) const {
+ SmallVector<Value> oneToOneOperands =
+ getOneToOneAdaptorOperands(adaptor.getOperands());
+ rewrite(op, OpAdaptor(oneToOneOperands, adaptor), rewriter);
+ }
virtual LogicalResult
matchAndRewrite(SourceOp op, OpAdaptor adaptor,
ConversionPatternRewriter &rewriter) const {
@@ -631,6 +669,13 @@ class OpConversionPattern : public ConversionPattern {
rewrite(op, adaptor, rewriter);
return success();
}
+ virtual LogicalResult
+ matchAndRewrite(SourceOp op, OneToNOpAdaptor adaptor,
+ ConversionPatternRewriter &rewriter) const {
+ SmallVector<Value> oneToOneOperands =
+ getOneToOneAdaptorOperands(adaptor.getOperands());
+ return matchAndRewrite(op, OpAdaptor(oneToOneOperands, adaptor), rewriter);
+ }
private:
using ConversionPattern::matchAndRewrite;
@@ -656,11 +701,20 @@ class OpInterfaceConversionPattern : public ConversionPattern {
ConversionPatternRewriter &rewriter) const final {
rewrite(cast<SourceOp>(op), operands, rewriter);
}
+ void rewrite(Operation *op, ArrayRef<ValueRange> operands,
+ ConversionPatternRewriter &rewriter) const final {
+ rewrite(cast<SourceOp>(op), operands, rewriter);
+ }
LogicalResult
matchAndRewrite(Operation *op, ArrayRef<Value> operands,
ConversionPatternRewriter &rewriter) const final {
return matchAndRewrite(cast<SourceOp>(op), operands, rewriter);
}
+ LogicalResult
+ matchAndRewrite(Operation *op, ArrayRef<ValueRange> operands,
+ ConversionPatternRewriter &rewriter) const final {
+ return matchAndRewrite(cast<SourceOp>(op), operands, rewriter);
+ }
/// Rewrite and Match methods that operate on the SourceOp type. These must be
/// overridden by the derived pattern class.
@@ -668,6 +722,10 @@ class OpInterfaceConversionPattern : public ConversionPattern {
ConversionPatternRewriter &rewriter) const {
llvm_unreachable("must override matchAndRewrite or a rewrite method");
}
+ virtual void rewrite(SourceOp op, ArrayRef<ValueRange> operands,
+ ConversionPatternRewriter &rewriter) const {
+ rewrite(op, getOneToOneAdaptorOperands(operands), rewriter);
+ }
virtual LogicalResult
matchAndRewrite(SourceOp op, ArrayRef<Value> operands,
ConversionPatternRewriter &rewriter) const {
@@ -676,6 +734,11 @@ class OpInterfaceConversionPattern : public ConversionPattern {
rewrite(op, operands, rewriter);
return success();
}
+ virtual LogicalResult
+ matchAndRewrite(SourceOp op, ArrayRef<ValueRange> operands,
+ ConversionPatternRewriter &rewriter) const {
+ return matchAndRewrite(op, getOneToOneAdaptorOperands(operands), rewriter);
+ }
private:
using ConversionPattern::matchAndRewrite;
diff --git a/mlir/lib/Dialect/Func/Transforms/DecomposeCallGraphTypes.cpp b/mlir/lib/Dialect/Func/Transforms/DecomposeCallGraphTypes.cpp
index a08764326a80b6..03be00328bda33 100644
--- a/mlir/lib/Dialect/Func/Transforms/DecomposeCallGraphTypes.cpp
+++ b/mlir/lib/Dialect/Func/Transforms/DecomposeCallGraphTypes.cpp
@@ -13,40 +13,6 @@
using namespace mlir;
using namespace mlir::func;
-//===----------------------------------------------------------------------===//
-// Helper functions
-//===----------------------------------------------------------------------===//
-
-/// If the given value can be decomposed with the type converter, decompose it.
-/// Otherwise, return the given value.
-// TODO: Value decomposition should happen automatically through a 1:N adaptor.
-// This function will disappear when the 1:1 and 1:N drivers are merged.
-static SmallVector<Value> decomposeValue(OpBuilder &builder, Location loc,
- Value value,
- const TypeConverter *converter) {
- // Try to convert the given value's type. If that fails, just return the
- // given value.
- SmallVector<Type> convertedTypes;
- if (failed(converter->convertType(value.getType(), convertedTypes)))
- return {value};
- if (convertedTypes.empty())
- return {};
-
- // If the given value's type is already legal, just return the given value.
- TypeRange convertedTypeRange(convertedTypes);
- if (convertedTypeRange == TypeRange(value.getType()))
- return {value};
-
- // Try to materialize a target conversion. If the materialization did not
- // produce values of the requested type, the materialization failed. Just
- // return the given value in that case.
- SmallVector<Value> result = converter->materializeTargetConversion(
- builder, loc, convertedTypeRange, value);
- if (result.empty())
- return {value};
- return result;
-}
-
//===----------------------------------------------------------------------===//
// DecomposeCallGraphTypesForFuncArgs
//===----------------------------------------------------------------------===//
@@ -102,16 +68,11 @@ struct DecomposeCallGraphTypesForReturnOp
using OpConversionPattern::OpConversionPattern;
LogicalResult
- matchAndRewrite(ReturnOp op, OpAdaptor adaptor,
+ matchAndRewrite(ReturnOp op, OneToNOpAdaptor adaptor,
ConversionPatternRewriter &rewriter) const final {
SmallVector<Value, 2> newOperands;
- for (Value operand : adaptor.getOperands()) {
- // TODO: We can directly take the values from the adaptor once this is a
- // 1:N conversion pattern.
- llvm::append_range(newOperands,
- decomposeValue(rewriter, operand.getLoc(), operand,
- getTypeConverter()));
- }
+ for (ValueRange operand : adaptor.getOperands())
+ llvm::append_range(newOperands, operand);
rewriter.replaceOpWithNewOp<ReturnOp>(op, newOperands);
return success();
}
@@ -128,18 +89,13 @@ struct DecomposeCallGraphTypesForCallOp : public OpConversionPattern<CallOp> {
using OpConversionPattern::OpConversionPattern;
LogicalResult
- matchAndRewrite(CallOp op, OpAdaptor adaptor,
+ matchAndRewrite(CallOp op, OneToNOpAdaptor adaptor,
ConversionPatternRewriter &rewriter) const final {
// Create the operands list of the new `CallOp`.
SmallVector<Value, 2> newOperands;
- for (Value operand : adaptor.getOperands()) {
- // TODO: We can directly take the values from the adaptor once this is a
- // 1:N conversion pattern.
- llvm::append_range(newOperands,
- decomposeValue(rewriter, operand.getLoc(), operand,
- getTypeConverter()));
- }
+ for (ValueRange operand : adaptor.getOperands())
+ llvm::append_range(newOperands, operand);
// Create the new result types for the new `CallOp` and track the number of
// replacement types for each original op result.
diff --git a/mlir/lib/Dialect/Func/Transforms/FuncConversions.cpp b/mlir/lib/Dialect/Func/Transforms/FuncConversions.cpp
index eb444d665ff260..d81f822f7d4b51 100644
--- a/mlir/lib/Dialect/Func/Transforms/FuncConversions.cpp
+++ b/mlir/lib/Dialect/Func/Transforms/FuncConversions.cpp
@@ -21,7 +21,7 @@ struct CallOpSignatureConversion : public OpConversionPattern<CallOp> {
/// Hook for derived classes to implement combined matching and rewriting.
LogicalResult
- matchAndRewrite(CallOp callOp, OpAdaptor adaptor,
+ matchAndRewrite(CallOp callOp, OneToNOpAdaptor adaptor,
ConversionPatternRewriter &rewriter) const override {
// Convert the original function results.
SmallVector<Type, 1> convertedResults;
@@ -37,7 +37,8 @@ struct CallOpSignatureConversion : public OpConversionPattern<CallOp> {
// Substitute with the new result types from the corresponding FuncType
// conversion.
rewriter.replaceOpWithNewOp<CallOp>(
- callOp, callOp.getCallee(), convertedResults, adaptor.getOperands());
+ callOp, callOp.getCallee(), convertedResults,
+ getOneToOneAdaptorOperands(adaptor.getOperands()));
return success();
}
};
diff --git a/mlir/lib/Dialect/SCF/Transforms/StructuralTypeConversions.cpp b/mlir/lib/Dialect/SCF/Transforms/StructuralTypeConversions.cpp
index 93a78056db1944..c0589044c26ecb 100644
--- a/mlir/lib/Dialect/SCF/Transforms/StructuralTypeConversions.cpp
+++ b/mlir/lib/Dialect/SCF/Transforms/StructuralTypeConversions.cpp
@@ -16,20 +16,18 @@ using namespace mlir::scf;
namespace {
-// Unpacks the single unrealized_conversion_cast using the list of inputs
-// e.g., return [%b, %c, %d] for %a = unrealized_conversion_cast(%b, %c, %d)
-static void unpackUnrealizedConversionCast(Value v,
- SmallVectorImpl<Value> &unpacked) {
- if (auto cast =
- dyn_cast_or_null<UnrealizedConversionCastOp>(v.getDefiningOp())) {
- if (cast.getInputs().size() != 1) {
- // 1 : N type conversion.
- unpacked.append(cast.getInputs().begin(), cast.getInputs().end());
- return;
- }
- }
- // 1 : 1 type conversion.
- unpacked.push_back(v);
+/// Flatten the given value ranges into a single vector of values.
+static SmallVector<Value> flattenValues(ArrayRef<ValueRange> values) {
+ SmallVector<Value> result;
+ for (const auto &vals : values)
+ llvm::append_range(result, vals);
+ return result;
+}
+
+/// Assert that the given value range contains a single value and return it.
+static Value getSingleValue(ValueRange values) {
+ assert(values.size() == 1 && "expected single value");
+ return values.front();
}
// CRTP
@@ -40,19 +38,21 @@ class Structural1ToNConversionPattern : public OpConversionPattern<SourceOp> {
public:
using OpConversionPattern<SourceOp>::typeConverter;
using OpConversionPattern<SourceOp>::OpConversionPattern;
- using OpAdaptor = typename OpConversionPattern<SourceOp>::OpAdaptor;
+ using OneToNOpAdaptor =
+ typename OpConversionPattern<SourceOp>::OneToNOpAdaptor;
//
// Derived classes should provide the following method which performs the
// actual conversion. It should return std::nullopt upon conversion failure
// and return the converted operation upon success.
//
- // std::optional<SourceOp> convertSourceOp(SourceOp op, OpAdaptor adaptor,
- // ConversionPatternRewriter &rewriter,
- // TypeRange dstTypes) const;
+ // std::optional<SourceOp> convertSourceOp(
+ // SourceOp op, OneToNOpAdaptor adaptor,
+ // ConversionPatternRewriter &rewriter,
+ // TypeRange dstTypes) const;
LogicalResult
- matchAndRewrite(SourceOp op, OpAdaptor adaptor,
+ matchAndRewrite(SourceOp op, OneToNOpAdaptor adaptor,
ConversionPatternRewriter &rewriter) const override {
SmallVector<Type> dstTypes;
SmallVector<unsigned> offsets;
@@ -73,28 +73,15 @@ class Structural1ToNConversionPattern : public OpConversionPattern<SourceOp> {
return rewriter.notifyMatchFailure(op, "could not convert operation");
// Packs the return value.
- SmallVector<Value> packedRets;
+ SmallVector<ValueRange> packedRets;
for (unsigned i = 1, e = offsets.size(); i < e; i++) {
unsigned start = offsets[i - 1], end = offsets[i];
unsigned len = end - start;
ValueRange mappedValue = newOp->getResults().slice(start, len);
- if (len != 1) {
- // 1 : N type conversion.
- Type origType = op.getResultTypes()[i - 1];
- Value mat = typeConverter->materializeSourceConversion(
- rewriter, op.getLoc(), origType, mappedValue);
- if (!mat) {
- return rewriter.notifyMatchFailure(
- op, "Failed to materialize 1:N type conversion");
- }
- packedRets.push_back(mat);
- } else {
- // 1 : 1 type conversion.
- packedRets.push_back(mappedValue.front());
- }
+ packedRets.push_back(mappedValue);
}
- rewriter.replaceOp(op, packedRets);
+ rewriter.replaceOpWithMultiple(op, packedRets);
return success();
}
};
@@ -105,7 +92,7 @@ class ConvertForOpTypes
using Structural1ToNConversionPattern::Structural1ToNConversionPattern;
// The callback required by CRTP.
- std::optional<ForOp> convertSourceOp(ForOp op, OpAdaptor adaptor,
+ std::optional<ForOp> convertSourceOp(ForOp op, OneToNOpAdaptor adaptor,
ConversionPatternRewriter &rewriter,
TypeRange dstTypes) const {
// Create a empty new op and inline the regions from the old op.
@@ -129,16 +116,13 @@ class ConvertForOpTypes
if (failed(rewriter.convertRegionTypes(&op.getRegion(), *typeConverter)))
return std::nullopt;
- // Unpacked the iteration arguments.
- SmallVector<Value> flatArgs;
- for (Value arg : adaptor.getInitArgs())
- unpackUnrealizedConversionCast(arg, flatArgs);
-
// ...
[truncated]
|
|
@llvm/pr-subscribers-mlir-sparse Author: Matthias Springer (matthias-springer) ChangesThis commit adds a new The existing LogicalResult ConversionPattern::matchAndRewrite(
Operation *op, ArrayRef<Value> operands /*adaptor values*/,
ConversionPatternRewriter &rewriter) const;This commit adds a LogicalResult ConversionPattern::matchAndRewrite(
Operation *op, ArrayRef<ValueRange> operands /*adaptor values*/,
ConversionPatternRewriter &rewriter) const {
// Note: getOneToOneAdaptorOperands produces a fatal error if at least one
// ValueRange has 0 or more than 1 value.
return matchAndRewrite(op, getOneToOneAdaptorOperands(operands), rewriter);
}The Before handing the adaptor values to a This commit also updates the the "decompose call graphs" transformation and the "sparse tensor codegen" transformation to use the new 1:N Note for LLVM conversion: If you are using a type converter with 1:N type conversion rules or if your patterns are performing 1:N replacements (via Patch is 70.77 KiB, truncated to 20.00 KiB below, full version: https://github.com/llvm/llvm-project/pull/116470.diff 9 Files Affected:
diff --git a/mlir/include/mlir/Conversion/LLVMCommon/Pattern.h b/mlir/include/mlir/Conversion/LLVMCommon/Pattern.h
index f3bf5b66398e09..86ea87b55af1cd 100644
--- a/mlir/include/mlir/Conversion/LLVMCommon/Pattern.h
+++ b/mlir/include/mlir/Conversion/LLVMCommon/Pattern.h
@@ -143,6 +143,8 @@ template <typename SourceOp>
class ConvertOpToLLVMPattern : public ConvertToLLVMPattern {
public:
using OpAdaptor = typename SourceOp::Adaptor;
+ using OneToNOpAdaptor =
+ typename SourceOp::template GenericAdaptor<ArrayRef<ValueRange>>;
explicit ConvertOpToLLVMPattern(const LLVMTypeConverter &typeConverter,
PatternBenefit benefit = 1)
@@ -153,8 +155,13 @@ class ConvertOpToLLVMPattern : public ConvertToLLVMPattern {
/// Wrappers around the RewritePattern methods that pass the derived op type.
void rewrite(Operation *op, ArrayRef<Value> operands,
ConversionPatternRewriter &rewriter) const final {
- rewrite(cast<SourceOp>(op), OpAdaptor(operands, cast<SourceOp>(op)),
- rewriter);
+ auto sourceOp = cast<SourceOp>(op);
+ rewrite(sourceOp, OpAdaptor(operands, sourceOp), rewriter);
+ }
+ void rewrite(Operation *op, ArrayRef<ValueRange> operands,
+ ConversionPatternRewriter &rewriter) const final {
+ auto sourceOp = cast<SourceOp>(op);
+ rewrite(sourceOp, OneToNOpAdaptor(operands, sourceOp), rewriter);
}
LogicalResult match(Operation *op) const final {
return match(cast<SourceOp>(op));
@@ -162,8 +169,15 @@ class ConvertOpToLLVMPattern : public ConvertToLLVMPattern {
LogicalResult
matchAndRewrite(Operation *op, ArrayRef<Value> operands,
ConversionPatternRewriter &rewriter) const final {
- return matchAndRewrite(cast<SourceOp>(op),
- OpAdaptor(operands, cast<SourceOp>(op)), rewriter);
+ auto sourceOp = cast<SourceOp>(op);
+ return matchAndRewrite(sourceOp, OpAdaptor(operands, sourceOp), rewriter);
+ }
+ LogicalResult
+ matchAndRewrite(Operation *op, ArrayRef<ValueRange> operands,
+ ConversionPatternRewriter &rewriter) const final {
+ auto sourceOp = cast<SourceOp>(op);
+ return matchAndRewrite(sourceOp, OneToNOpAdaptor(operands, sourceOp),
+ rewriter);
}
/// Rewrite and Match methods that operate on the SourceOp type. These must be
@@ -175,6 +189,12 @@ class ConvertOpToLLVMPattern : public ConvertToLLVMPattern {
ConversionPatternRewriter &rewriter) const {
llvm_unreachable("must override rewrite or matchAndRewrite");
}
+ virtual void rewrite(SourceOp op, OneToNOpAdaptor adaptor,
+ ConversionPatternRewriter &rewriter) const {
+ SmallVector<Value> oneToOneOperands =
+ getOneToOneAdaptorOperands(adaptor.getOperands());
+ rewrite(op, OpAdaptor(oneToOneOperands, adaptor), rewriter);
+ }
virtual LogicalResult
matchAndRewrite(SourceOp op, OpAdaptor adaptor,
ConversionPatternRewriter &rewriter) const {
@@ -183,6 +203,13 @@ class ConvertOpToLLVMPattern : public ConvertToLLVMPattern {
rewrite(op, adaptor, rewriter);
return success();
}
+ virtual LogicalResult
+ matchAndRewrite(SourceOp op, OneToNOpAdaptor adaptor,
+ ConversionPatternRewriter &rewriter) const {
+ SmallVector<Value> oneToOneOperands =
+ getOneToOneAdaptorOperands(adaptor.getOperands());
+ return matchAndRewrite(op, OpAdaptor(oneToOneOperands, adaptor), rewriter);
+ }
private:
using ConvertToLLVMPattern::match;
diff --git a/mlir/include/mlir/Transforms/DialectConversion.h b/mlir/include/mlir/Transforms/DialectConversion.h
index de47765006f81e..e4eeb39b9c0741 100644
--- a/mlir/include/mlir/Transforms/DialectConversion.h
+++ b/mlir/include/mlir/Transforms/DialectConversion.h
@@ -537,6 +537,10 @@ class ConversionPattern : public RewritePattern {
ConversionPatternRewriter &rewriter) const {
llvm_unreachable("unimplemented rewrite");
}
+ virtual void rewrite(Operation *op, ArrayRef<ValueRange> operands,
+ ConversionPatternRewriter &rewriter) const {
+ rewrite(op, getOneToOneAdaptorOperands(operands), rewriter);
+ }
/// Hook for derived classes to implement combined matching and rewriting.
virtual LogicalResult
@@ -547,6 +551,11 @@ class ConversionPattern : public RewritePattern {
rewrite(op, operands, rewriter);
return success();
}
+ virtual LogicalResult
+ matchAndRewrite(Operation *op, ArrayRef<ValueRange> operands,
+ ConversionPatternRewriter &rewriter) const {
+ return matchAndRewrite(op, getOneToOneAdaptorOperands(operands), rewriter);
+ }
/// Attempt to match and rewrite the IR root at the specified operation.
LogicalResult matchAndRewrite(Operation *op,
@@ -574,6 +583,15 @@ class ConversionPattern : public RewritePattern {
: RewritePattern(std::forward<Args>(args)...),
typeConverter(&typeConverter) {}
+ /// Given an array of value ranges, which are the inputs to a 1:N adaptor,
+ /// try to extract the single value of each range to construct a the inputs
+ /// for a 1:1 adaptor.
+ ///
+ /// This function produces a fatal error if at least one range has 0 or
+ /// more than 1 value: "pattern 'name' does not support 1:N conversion"
+ SmallVector<Value>
+ getOneToOneAdaptorOperands(ArrayRef<ValueRange> operands) const;
+
protected:
/// An optional type converter for use by this pattern.
const TypeConverter *typeConverter = nullptr;
@@ -589,6 +607,8 @@ template <typename SourceOp>
class OpConversionPattern : public ConversionPattern {
public:
using OpAdaptor = typename SourceOp::Adaptor;
+ using OneToNOpAdaptor =
+ typename SourceOp::template GenericAdaptor<ArrayRef<ValueRange>>;
OpConversionPattern(MLIRContext *context, PatternBenefit benefit = 1)
: ConversionPattern(SourceOp::getOperationName(), benefit, context) {}
@@ -607,12 +627,24 @@ class OpConversionPattern : public ConversionPattern {
auto sourceOp = cast<SourceOp>(op);
rewrite(sourceOp, OpAdaptor(operands, sourceOp), rewriter);
}
+ void rewrite(Operation *op, ArrayRef<ValueRange> operands,
+ ConversionPatternRewriter &rewriter) const final {
+ auto sourceOp = cast<SourceOp>(op);
+ rewrite(sourceOp, OneToNOpAdaptor(operands, sourceOp), rewriter);
+ }
LogicalResult
matchAndRewrite(Operation *op, ArrayRef<Value> operands,
ConversionPatternRewriter &rewriter) const final {
auto sourceOp = cast<SourceOp>(op);
return matchAndRewrite(sourceOp, OpAdaptor(operands, sourceOp), rewriter);
}
+ LogicalResult
+ matchAndRewrite(Operation *op, ArrayRef<ValueRange> operands,
+ ConversionPatternRewriter &rewriter) const final {
+ auto sourceOp = cast<SourceOp>(op);
+ return matchAndRewrite(sourceOp, OneToNOpAdaptor(operands, sourceOp),
+ rewriter);
+ }
/// Rewrite and Match methods that operate on the SourceOp type. These must be
/// overridden by the derived pattern class.
@@ -623,6 +655,12 @@ class OpConversionPattern : public ConversionPattern {
ConversionPatternRewriter &rewriter) const {
llvm_unreachable("must override matchAndRewrite or a rewrite method");
}
+ virtual void rewrite(SourceOp op, OneToNOpAdaptor adaptor,
+ ConversionPatternRewriter &rewriter) const {
+ SmallVector<Value> oneToOneOperands =
+ getOneToOneAdaptorOperands(adaptor.getOperands());
+ rewrite(op, OpAdaptor(oneToOneOperands, adaptor), rewriter);
+ }
virtual LogicalResult
matchAndRewrite(SourceOp op, OpAdaptor adaptor,
ConversionPatternRewriter &rewriter) const {
@@ -631,6 +669,13 @@ class OpConversionPattern : public ConversionPattern {
rewrite(op, adaptor, rewriter);
return success();
}
+ virtual LogicalResult
+ matchAndRewrite(SourceOp op, OneToNOpAdaptor adaptor,
+ ConversionPatternRewriter &rewriter) const {
+ SmallVector<Value> oneToOneOperands =
+ getOneToOneAdaptorOperands(adaptor.getOperands());
+ return matchAndRewrite(op, OpAdaptor(oneToOneOperands, adaptor), rewriter);
+ }
private:
using ConversionPattern::matchAndRewrite;
@@ -656,11 +701,20 @@ class OpInterfaceConversionPattern : public ConversionPattern {
ConversionPatternRewriter &rewriter) const final {
rewrite(cast<SourceOp>(op), operands, rewriter);
}
+ void rewrite(Operation *op, ArrayRef<ValueRange> operands,
+ ConversionPatternRewriter &rewriter) const final {
+ rewrite(cast<SourceOp>(op), operands, rewriter);
+ }
LogicalResult
matchAndRewrite(Operation *op, ArrayRef<Value> operands,
ConversionPatternRewriter &rewriter) const final {
return matchAndRewrite(cast<SourceOp>(op), operands, rewriter);
}
+ LogicalResult
+ matchAndRewrite(Operation *op, ArrayRef<ValueRange> operands,
+ ConversionPatternRewriter &rewriter) const final {
+ return matchAndRewrite(cast<SourceOp>(op), operands, rewriter);
+ }
/// Rewrite and Match methods that operate on the SourceOp type. These must be
/// overridden by the derived pattern class.
@@ -668,6 +722,10 @@ class OpInterfaceConversionPattern : public ConversionPattern {
ConversionPatternRewriter &rewriter) const {
llvm_unreachable("must override matchAndRewrite or a rewrite method");
}
+ virtual void rewrite(SourceOp op, ArrayRef<ValueRange> operands,
+ ConversionPatternRewriter &rewriter) const {
+ rewrite(op, getOneToOneAdaptorOperands(operands), rewriter);
+ }
virtual LogicalResult
matchAndRewrite(SourceOp op, ArrayRef<Value> operands,
ConversionPatternRewriter &rewriter) const {
@@ -676,6 +734,11 @@ class OpInterfaceConversionPattern : public ConversionPattern {
rewrite(op, operands, rewriter);
return success();
}
+ virtual LogicalResult
+ matchAndRewrite(SourceOp op, ArrayRef<ValueRange> operands,
+ ConversionPatternRewriter &rewriter) const {
+ return matchAndRewrite(op, getOneToOneAdaptorOperands(operands), rewriter);
+ }
private:
using ConversionPattern::matchAndRewrite;
diff --git a/mlir/lib/Dialect/Func/Transforms/DecomposeCallGraphTypes.cpp b/mlir/lib/Dialect/Func/Transforms/DecomposeCallGraphTypes.cpp
index a08764326a80b6..03be00328bda33 100644
--- a/mlir/lib/Dialect/Func/Transforms/DecomposeCallGraphTypes.cpp
+++ b/mlir/lib/Dialect/Func/Transforms/DecomposeCallGraphTypes.cpp
@@ -13,40 +13,6 @@
using namespace mlir;
using namespace mlir::func;
-//===----------------------------------------------------------------------===//
-// Helper functions
-//===----------------------------------------------------------------------===//
-
-/// If the given value can be decomposed with the type converter, decompose it.
-/// Otherwise, return the given value.
-// TODO: Value decomposition should happen automatically through a 1:N adaptor.
-// This function will disappear when the 1:1 and 1:N drivers are merged.
-static SmallVector<Value> decomposeValue(OpBuilder &builder, Location loc,
- Value value,
- const TypeConverter *converter) {
- // Try to convert the given value's type. If that fails, just return the
- // given value.
- SmallVector<Type> convertedTypes;
- if (failed(converter->convertType(value.getType(), convertedTypes)))
- return {value};
- if (convertedTypes.empty())
- return {};
-
- // If the given value's type is already legal, just return the given value.
- TypeRange convertedTypeRange(convertedTypes);
- if (convertedTypeRange == TypeRange(value.getType()))
- return {value};
-
- // Try to materialize a target conversion. If the materialization did not
- // produce values of the requested type, the materialization failed. Just
- // return the given value in that case.
- SmallVector<Value> result = converter->materializeTargetConversion(
- builder, loc, convertedTypeRange, value);
- if (result.empty())
- return {value};
- return result;
-}
-
//===----------------------------------------------------------------------===//
// DecomposeCallGraphTypesForFuncArgs
//===----------------------------------------------------------------------===//
@@ -102,16 +68,11 @@ struct DecomposeCallGraphTypesForReturnOp
using OpConversionPattern::OpConversionPattern;
LogicalResult
- matchAndRewrite(ReturnOp op, OpAdaptor adaptor,
+ matchAndRewrite(ReturnOp op, OneToNOpAdaptor adaptor,
ConversionPatternRewriter &rewriter) const final {
SmallVector<Value, 2> newOperands;
- for (Value operand : adaptor.getOperands()) {
- // TODO: We can directly take the values from the adaptor once this is a
- // 1:N conversion pattern.
- llvm::append_range(newOperands,
- decomposeValue(rewriter, operand.getLoc(), operand,
- getTypeConverter()));
- }
+ for (ValueRange operand : adaptor.getOperands())
+ llvm::append_range(newOperands, operand);
rewriter.replaceOpWithNewOp<ReturnOp>(op, newOperands);
return success();
}
@@ -128,18 +89,13 @@ struct DecomposeCallGraphTypesForCallOp : public OpConversionPattern<CallOp> {
using OpConversionPattern::OpConversionPattern;
LogicalResult
- matchAndRewrite(CallOp op, OpAdaptor adaptor,
+ matchAndRewrite(CallOp op, OneToNOpAdaptor adaptor,
ConversionPatternRewriter &rewriter) const final {
// Create the operands list of the new `CallOp`.
SmallVector<Value, 2> newOperands;
- for (Value operand : adaptor.getOperands()) {
- // TODO: We can directly take the values from the adaptor once this is a
- // 1:N conversion pattern.
- llvm::append_range(newOperands,
- decomposeValue(rewriter, operand.getLoc(), operand,
- getTypeConverter()));
- }
+ for (ValueRange operand : adaptor.getOperands())
+ llvm::append_range(newOperands, operand);
// Create the new result types for the new `CallOp` and track the number of
// replacement types for each original op result.
diff --git a/mlir/lib/Dialect/Func/Transforms/FuncConversions.cpp b/mlir/lib/Dialect/Func/Transforms/FuncConversions.cpp
index eb444d665ff260..d81f822f7d4b51 100644
--- a/mlir/lib/Dialect/Func/Transforms/FuncConversions.cpp
+++ b/mlir/lib/Dialect/Func/Transforms/FuncConversions.cpp
@@ -21,7 +21,7 @@ struct CallOpSignatureConversion : public OpConversionPattern<CallOp> {
/// Hook for derived classes to implement combined matching and rewriting.
LogicalResult
- matchAndRewrite(CallOp callOp, OpAdaptor adaptor,
+ matchAndRewrite(CallOp callOp, OneToNOpAdaptor adaptor,
ConversionPatternRewriter &rewriter) const override {
// Convert the original function results.
SmallVector<Type, 1> convertedResults;
@@ -37,7 +37,8 @@ struct CallOpSignatureConversion : public OpConversionPattern<CallOp> {
// Substitute with the new result types from the corresponding FuncType
// conversion.
rewriter.replaceOpWithNewOp<CallOp>(
- callOp, callOp.getCallee(), convertedResults, adaptor.getOperands());
+ callOp, callOp.getCallee(), convertedResults,
+ getOneToOneAdaptorOperands(adaptor.getOperands()));
return success();
}
};
diff --git a/mlir/lib/Dialect/SCF/Transforms/StructuralTypeConversions.cpp b/mlir/lib/Dialect/SCF/Transforms/StructuralTypeConversions.cpp
index 93a78056db1944..c0589044c26ecb 100644
--- a/mlir/lib/Dialect/SCF/Transforms/StructuralTypeConversions.cpp
+++ b/mlir/lib/Dialect/SCF/Transforms/StructuralTypeConversions.cpp
@@ -16,20 +16,18 @@ using namespace mlir::scf;
namespace {
-// Unpacks the single unrealized_conversion_cast using the list of inputs
-// e.g., return [%b, %c, %d] for %a = unrealized_conversion_cast(%b, %c, %d)
-static void unpackUnrealizedConversionCast(Value v,
- SmallVectorImpl<Value> &unpacked) {
- if (auto cast =
- dyn_cast_or_null<UnrealizedConversionCastOp>(v.getDefiningOp())) {
- if (cast.getInputs().size() != 1) {
- // 1 : N type conversion.
- unpacked.append(cast.getInputs().begin(), cast.getInputs().end());
- return;
- }
- }
- // 1 : 1 type conversion.
- unpacked.push_back(v);
+/// Flatten the given value ranges into a single vector of values.
+static SmallVector<Value> flattenValues(ArrayRef<ValueRange> values) {
+ SmallVector<Value> result;
+ for (const auto &vals : values)
+ llvm::append_range(result, vals);
+ return result;
+}
+
+/// Assert that the given value range contains a single value and return it.
+static Value getSingleValue(ValueRange values) {
+ assert(values.size() == 1 && "expected single value");
+ return values.front();
}
// CRTP
@@ -40,19 +38,21 @@ class Structural1ToNConversionPattern : public OpConversionPattern<SourceOp> {
public:
using OpConversionPattern<SourceOp>::typeConverter;
using OpConversionPattern<SourceOp>::OpConversionPattern;
- using OpAdaptor = typename OpConversionPattern<SourceOp>::OpAdaptor;
+ using OneToNOpAdaptor =
+ typename OpConversionPattern<SourceOp>::OneToNOpAdaptor;
//
// Derived classes should provide the following method which performs the
// actual conversion. It should return std::nullopt upon conversion failure
// and return the converted operation upon success.
//
- // std::optional<SourceOp> convertSourceOp(SourceOp op, OpAdaptor adaptor,
- // ConversionPatternRewriter &rewriter,
- // TypeRange dstTypes) const;
+ // std::optional<SourceOp> convertSourceOp(
+ // SourceOp op, OneToNOpAdaptor adaptor,
+ // ConversionPatternRewriter &rewriter,
+ // TypeRange dstTypes) const;
LogicalResult
- matchAndRewrite(SourceOp op, OpAdaptor adaptor,
+ matchAndRewrite(SourceOp op, OneToNOpAdaptor adaptor,
ConversionPatternRewriter &rewriter) const override {
SmallVector<Type> dstTypes;
SmallVector<unsigned> offsets;
@@ -73,28 +73,15 @@ class Structural1ToNConversionPattern : public OpConversionPattern<SourceOp> {
return rewriter.notifyMatchFailure(op, "could not convert operation");
// Packs the return value.
- SmallVector<Value> packedRets;
+ SmallVector<ValueRange> packedRets;
for (unsigned i = 1, e = offsets.size(); i < e; i++) {
unsigned start = offsets[i - 1], end = offsets[i];
unsigned len = end - start;
ValueRange mappedValue = newOp->getResults().slice(start, len);
- if (len != 1) {
- // 1 : N type conversion.
- Type origType = op.getResultTypes()[i - 1];
- Value mat = typeConverter->materializeSourceConversion(
- rewriter, op.getLoc(), origType, mappedValue);
- if (!mat) {
- return rewriter.notifyMatchFailure(
- op, "Failed to materialize 1:N type conversion");
- }
- packedRets.push_back(mat);
- } else {
- // 1 : 1 type conversion.
- packedRets.push_back(mappedValue.front());
- }
+ packedRets.push_back(mappedValue);
}
- rewriter.replaceOp(op, packedRets);
+ rewriter.replaceOpWithMultiple(op, packedRets);
return success();
}
};
@@ -105,7 +92,7 @@ class ConvertForOpTypes
using Structural1ToNConversionPattern::Structural1ToNConversionPattern;
// The callback required by CRTP.
- std::optional<ForOp> convertSourceOp(ForOp op, OpAdaptor adaptor,
+ std::optional<ForOp> convertSourceOp(ForOp op, OneToNOpAdaptor adaptor,
ConversionPatternRewriter &rewriter,
TypeRange dstTypes) const {
// Create a empty new op and inline the regions from the old op.
@@ -129,16 +116,13 @@ class ConvertForOpTypes
if (failed(rewriter.convertRegionTypes(&op.getRegion(), *typeConverter)))
return std::nullopt;
- // Unpacked the iteration arguments.
- SmallVector<Value> flatArgs;
- for (Value arg : adaptor.getInitArgs())
- unpackUnrealizedConversionCast(arg, flatArgs);
-
// ...
[truncated]
|
zero9178
reviewed
Nov 16, 2024
Comment on lines
+1611
to
+1612
| for (int i = 0; i < newValues.size(); ++i) | ||
| newVals.push_back(newValues.slice(i, 1)); |
There was a problem hiding this comment.
Suggested change
| for (int i = 0; i < newValues.size(); ++i) | |
| newVals.push_back(newValues.slice(i, 1)); | |
| for (Value value : newValues) | |
| newVals.push_back(value); |
Comment on lines
+1827
to
+1828
| SmallVector<ValueRange> remappedAsRange = llvm::map_to_vector( | ||
| remapped, [](const auto &v) -> ValueRange { return v; }); |
There was a problem hiding this comment.
Suggested change
| SmallVector<ValueRange> remappedAsRange = llvm::map_to_vector( | |
| remapped, [](const auto &v) -> ValueRange { return v; }); | |
| SmallVector<ValueRange> remappedAsRange = llvm::to_vector_of<ValueRange>( | |
| remapped); |
| @@ -1965,19 +2053,19 @@ OperationLegalizer::legalizeWithFold(Operation *op, | |||
| }); | |||
|
|
|||
| // Try to fold the operation. | |||
| SmallVector<Value, 2> replacementValues; | |||
| SmallVector<Value, 2> ValueRange; | |||
There was a problem hiding this comment.
Suggested change
| SmallVector<Value, 2> ValueRange; | |
| SmallVector<Value, 2> valueRange; |
| @@ -537,6 +537,10 @@ class ConversionPattern : public RewritePattern { | |||
| ConversionPatternRewriter &rewriter) const { | |||
| llvm_unreachable("unimplemented rewrite"); | |||
| } | |||
| virtual void rewrite(Operation *op, ArrayRef<ValueRange> operands, | |||
There was a problem hiding this comment.
Think it'd be useful here to add doc comments how this differs from the other rewrite overload so that readers know when they'd want to (or are required to) use this overload rather than the other.
| Value repl = mapping.lookupOrDefault(operand); | ||
| SmallVector<Value> unpacked = unpackNTo1Materialization(repl); | ||
| if (TypeRange(unpacked) == legalTypes) { | ||
| remapped.push_back(unpacked); |
There was a problem hiding this comment.
Suggested change
| remapped.push_back(unpacked); | |
| remapped.push_back(std::move(unpacked)); |
| return {value}; | ||
| assert(castOp->getNumResults() == 1 && "expected single result"); | ||
|
|
||
| SmallVector<Value> result; |
There was a problem hiding this comment.
Do I understand correctly that this recursion for scenarios such as tuple<i32, tuple<i64>> to flatten them properly? I think it might be worth leaving a comment here just in case. This also causes some of the major simplifications in the decompose-call-graph tests right?
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
This commit adds a new
matchAndRewriteoverload toConversionPatternto support 1:N replacements. This is the first of two main PRs that merge the 1:1 and 1:N dialect conversion drivers.The existing
matchAndRewritefunction supports only 1:1 replacements, as can be seen from theArrayRef<Value>parameter.This commit adds a
matchAndRewriteoverload that is called by the dialect conversion driver. By default, this new overload dispatches to the original 1:1matchAndRewriteimplementation. ExistingConversionPatterns do not need to be changed as long as there are no 1:N type conversions or value replacements.The
ConversionValueMapping, which keeps track of value replacements and materializations, still does not support 1:N replacements. We still rely on argument materializations to convert N replacement values back into a single value. TheConversionValueMappingwill be generalized to 1:N mappings in the second main PR.Before handing the adaptor values to a
ConversionPattern, all argument materializations are "unpacked". TheConversionPatternreceives N replacement values and does not see any argument materializations. This implementation strategy allows us to use the 1:N infrastructure/API inConversionPatterns even though some functionality is still missing in the driver. This strategy was chosen to keep the sizes of the PRs smaller and to make it easier for downstream users to adapt to API changes.This commit also updates the the "decompose call graphs" transformation and the "sparse tensor codegen" transformation to use the new 1:N
ConversionPatternAPI.Note for LLVM conversion: If you are using a type converter with 1:N type conversion rules or if your patterns are performing 1:N replacements (via
replaceOpWithMultipleorapplySignatureConversion), conversion pattern applications will start failing (fatal LLVM error) with this error message:pattern 'name' does not support 1:N conversion. The name of the failing pattern is shown in the error message. These patterns must be updated to the new 1:NmatchAndRewriteAPI.