[MLIR] Allow constFoldBinaryOp to fold (T1, T1) -> T2
#151410
Conversation
The `constFoldBinaryOp` helper function had limited support for different input and output types, but the static type of the underlying value (e.g. `APInt`) had to match between the inputs and the output. This worked fine for int comparisons of the form `(intN, intN) -> int1`, as the static type signature was `(APInt, APInt) -> APInt`. However, float comparisons map `(floatN, floatN) -> int1`, with a static type signature of `(APFloat, APFloat) -> APInt`. This use case wasn't supported by `constFoldBinaryOp`. `constFoldBinaryOp` now accepts an optional template argument overriding the return type in case it differs from the input type. If the new template argument isn't provided, the default behavior is unchanged (i.e. the return type will be assumed to match the input type).
|
Thank you for submitting a Pull Request (PR) to the LLVM Project! This PR will be automatically labeled and the relevant teams will be notified. If you wish to, you can add reviewers by using the "Reviewers" section on this page. If this is not working for you, it is probably because you do not have write permissions for the repository. In which case you can instead tag reviewers by name in a comment by using If you have received no comments on your PR for a week, you can request a review by "ping"ing the PR by adding a comment “Ping”. The common courtesy "ping" rate is once a week. Please remember that you are asking for valuable time from other developers. If you have further questions, they may be answered by the LLVM GitHub User Guide. You can also ask questions in a comment on this PR, on the LLVM Discord or on the forums. |
|
@llvm/pr-subscribers-mlir Author: Matthias Guenther (mrguenther) ChangesThe This worked fine for int comparisons of the form
Full diff: https://github.com/llvm/llvm-project/pull/151410.diff 3 Files Affected:
diff --git a/mlir/include/mlir/Dialect/CommonFolders.h b/mlir/include/mlir/Dialect/CommonFolders.h
index b5a12426aff80..79832139b8610 100644
--- a/mlir/include/mlir/Dialect/CommonFolders.h
+++ b/mlir/include/mlir/Dialect/CommonFolders.h
@@ -15,10 +15,16 @@
#ifndef MLIR_DIALECT_COMMONFOLDERS_H
#define MLIR_DIALECT_COMMONFOLDERS_H
+#include "mlir/IR/Attributes.h"
+#include "mlir/IR/BuiltinAttributeInterfaces.h"
#include "mlir/IR/BuiltinAttributes.h"
-#include "mlir/IR/BuiltinTypes.h"
+#include "mlir/IR/BuiltinTypeInterfaces.h"
+#include "mlir/IR/Types.h"
#include "llvm/ADT/ArrayRef.h"
#include "llvm/ADT/STLExtras.h"
+
+#include <cassert>
+#include <cstddef>
#include <optional>
namespace mlir {
@@ -30,11 +36,13 @@ class PoisonAttr;
/// Uses `resultType` for the type of the returned attribute.
/// Optional PoisonAttr template argument allows to specify 'poison' attribute
/// which will be directly propagated to result.
-template <class AttrElementT,
+template <class AttrElementT, //
class ElementValueT = typename AttrElementT::ValueType,
class PoisonAttr = ub::PoisonAttr,
+ class ResultAttrElementT = AttrElementT,
+ class ResultElementValueT = typename ResultAttrElementT::ValueType,
class CalculationT = function_ref<
- std::optional<ElementValueT>(ElementValueT, ElementValueT)>>
+ std::optional<ResultElementValueT>(ElementValueT, ElementValueT)>>
Attribute constFoldBinaryOpConditional(ArrayRef<Attribute> operands,
Type resultType,
CalculationT &&calculate) {
@@ -65,7 +73,7 @@ Attribute constFoldBinaryOpConditional(ArrayRef<Attribute> operands,
if (!calRes)
return {};
- return AttrElementT::get(resultType, *calRes);
+ return ResultAttrElementT::get(resultType, *calRes);
}
if (isa<SplatElementsAttr>(operands[0]) &&
@@ -99,7 +107,7 @@ Attribute constFoldBinaryOpConditional(ArrayRef<Attribute> operands,
return {};
auto lhsIt = *maybeLhsIt;
auto rhsIt = *maybeRhsIt;
- SmallVector<ElementValueT, 4> elementResults;
+ SmallVector<ResultElementValueT, 4> elementResults;
elementResults.reserve(lhs.getNumElements());
for (size_t i = 0, e = lhs.getNumElements(); i < e; ++i, ++lhsIt, ++rhsIt) {
auto elementResult = calculate(*lhsIt, *rhsIt);
@@ -119,11 +127,13 @@ Attribute constFoldBinaryOpConditional(ArrayRef<Attribute> operands,
/// attribute.
/// Optional PoisonAttr template argument allows to specify 'poison' attribute
/// which will be directly propagated to result.
-template <class AttrElementT,
+template <class AttrElementT, //
class ElementValueT = typename AttrElementT::ValueType,
class PoisonAttr = ub::PoisonAttr,
+ class ResultAttrElementT = AttrElementT,
+ class ResultElementValueT = typename ResultAttrElementT::ValueType,
class CalculationT = function_ref<
- std::optional<ElementValueT>(ElementValueT, ElementValueT)>>
+ std::optional<ResultElementValueT>(ElementValueT, ElementValueT)>>
Attribute constFoldBinaryOpConditional(ArrayRef<Attribute> operands,
CalculationT &&calculate) {
assert(operands.size() == 2 && "binary op takes two operands");
@@ -153,36 +163,41 @@ Attribute constFoldBinaryOpConditional(ArrayRef<Attribute> operands,
return {};
return constFoldBinaryOpConditional<AttrElementT, ElementValueT, PoisonAttr,
+ ResultAttrElementT, ResultElementValueT,
CalculationT>(
operands, lhsType, std::forward<CalculationT>(calculate));
}
template <class AttrElementT,
class ElementValueT = typename AttrElementT::ValueType,
- class PoisonAttr = void,
+ class PoisonAttr = void, //
+ class ResultAttrElementT = AttrElementT,
+ class ResultElementValueT = typename ResultAttrElementT::ValueType,
class CalculationT =
- function_ref<ElementValueT(ElementValueT, ElementValueT)>>
+ function_ref<ResultElementValueT(ElementValueT, ElementValueT)>>
Attribute constFoldBinaryOp(ArrayRef<Attribute> operands, Type resultType,
CalculationT &&calculate) {
- return constFoldBinaryOpConditional<AttrElementT, ElementValueT, PoisonAttr>(
+ return constFoldBinaryOpConditional<AttrElementT, ElementValueT, PoisonAttr,
+ ResultAttrElementT>(
operands, resultType,
- [&](ElementValueT a, ElementValueT b) -> std::optional<ElementValueT> {
- return calculate(a, b);
- });
+ [&](ElementValueT a, ElementValueT b)
+ -> std::optional<ResultElementValueT> { return calculate(a, b); });
}
-template <class AttrElementT,
+template <class AttrElementT, //
class ElementValueT = typename AttrElementT::ValueType,
class PoisonAttr = ub::PoisonAttr,
+ class ResultAttrElementT = AttrElementT,
+ class ResultElementValueT = typename ResultAttrElementT::ValueType,
class CalculationT =
- function_ref<ElementValueT(ElementValueT, ElementValueT)>>
+ function_ref<ResultElementValueT(ElementValueT, ElementValueT)>>
Attribute constFoldBinaryOp(ArrayRef<Attribute> operands,
CalculationT &&calculate) {
- return constFoldBinaryOpConditional<AttrElementT, ElementValueT, PoisonAttr>(
+ return constFoldBinaryOpConditional<AttrElementT, ElementValueT, PoisonAttr,
+ ResultAttrElementT>(
operands,
- [&](ElementValueT a, ElementValueT b) -> std::optional<ElementValueT> {
- return calculate(a, b);
- });
+ [&](ElementValueT a, ElementValueT b)
+ -> std::optional<ResultElementValueT> { return calculate(a, b); });
}
/// Performs constant folding `calculate` with element-wise behavior on the one
diff --git a/mlir/unittests/Dialect/CMakeLists.txt b/mlir/unittests/Dialect/CMakeLists.txt
index aea247547473d..bbe1c021ec4bf 100644
--- a/mlir/unittests/Dialect/CMakeLists.txt
+++ b/mlir/unittests/Dialect/CMakeLists.txt
@@ -1,5 +1,6 @@
add_mlir_unittest(MLIRDialectTests
BroadcastShapeTest.cpp
+ CommonFoldersTest.cpp
)
mlir_target_link_libraries(MLIRDialectTests
PRIVATE
diff --git a/mlir/unittests/Dialect/CommonFoldersTest.cpp b/mlir/unittests/Dialect/CommonFoldersTest.cpp
new file mode 100644
index 0000000000000..31d9c592dabc1
--- /dev/null
+++ b/mlir/unittests/Dialect/CommonFoldersTest.cpp
@@ -0,0 +1,61 @@
+//===- CommonFoldersTest.cpp - tests for folder-pattern helper templates --===//
+//
+// 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/CommonFolders.h"
+#include "mlir/IR/BuiltinAttributes.h"
+#include "mlir/IR/BuiltinTypeInterfaces.h"
+#include "mlir/IR/BuiltinTypes.h"
+#include "mlir/IR/MLIRContext.h"
+#include "llvm/ADT/APFloat.h"
+#include "llvm/ADT/APInt.h"
+#include "llvm/Support/Casting.h"
+#include "gmock/gmock.h"
+#include "gtest/gtest.h"
+
+namespace mlir {
+namespace {
+
+using ::llvm::APFloat;
+using ::llvm::APInt;
+using ::mlir::constFoldBinaryOp;
+using ::mlir::DenseElementsAttr;
+using ::mlir::Float32Type;
+using ::mlir::FloatAttr;
+using ::mlir::IntegerAttr;
+using ::mlir::IntegerType;
+using ::mlir::MLIRContext;
+using ::mlir::RankedTensorType;
+using ::testing::ElementsAre;
+
+APInt floatLessThan(APFloat lhs, APFloat rhs) { return APInt(1, lhs < rhs); }
+
+TEST(CommonFoldersTest, FoldFloatComparisonToBoolean) {
+ MLIRContext context;
+ auto vector4xf32 = RankedTensorType::get({4}, Float32Type::get(&context));
+
+ auto lhs = DenseElementsAttr::get(vector4xf32, {-12.9f, 0.0f, 42.5f, -0.01f});
+ auto rhs = DenseElementsAttr::get(vector4xf32, {0.0f, 0.0f, 0.0f, 0.0f});
+
+ auto result = llvm::dyn_cast<DenseElementsAttr>(
+ constFoldBinaryOp<FloatAttr, FloatAttr::ValueType, void, IntegerAttr>(
+ {lhs, rhs}, RankedTensorType::get({4}, IntegerType::get(&context, 1)),
+ floatLessThan));
+ ASSERT_TRUE(result);
+
+ auto resultElementType = result.getElementType();
+ EXPECT_TRUE(resultElementType.isInteger(1));
+
+ const APInt i1True = APInt(1, true);
+ const APInt i1False = APInt(1, false);
+
+ EXPECT_THAT(result.getValues<APInt>(),
+ ElementsAre(i1True, i1False, i1False, i1True));
+}
+
+} // namespace
+} // namespace mlir
|
joker-eph
changed the title
constFoldBinaryOp to fold (T1, T1) -> T2constFoldBinaryOp to fold (T1, T1) -> T2
|
|
||
| EXPECT_THAT(result.getValues<APInt>(), | ||
| ElementsAre(i1True, i1False, i1False, i1True)); | ||
| } |
There was a problem hiding this comment.
Can we get this exercised through one of the canonicalize of the IR tests? (and avoid adding C++ unit-tests)
I suspect the arithmetic dialect could be exercising this?
There was a problem hiding this comment.
Oh, okay. I assumed a C++ unit test was appropriate here since I was testing a C++ helper function, not the patterns making use of it, but I can exercise it by testing a pattern based on it instead if you prefer.
To clarify, should I remove the C++ test entirely once I've exercised it through a tested pattern?
There was a problem hiding this comment.
Hmm, updating the Arith dialect to use this helper function for float comparisons looks non-trivial; the current logic is rather complex, and I'd need to refactor it significantly. I can give it a try if you like, but it might make sense to handle that in a separate PR. Would you be open to submitting this C++ test for now, then deleting it later if and when a LIT test starts covering the same functionality?
There was a problem hiding this comment.
I would rather not see this C++ test landing in the first place.
If arith isn't suitable right now , you can write something in the test dialect to demonstrate the feature.
The
constFoldBinaryOphelper function had limited support for different input and output types, but the static type of the underlying value (e.g.APInt) had to match between the inputs and the output.This worked fine for int comparisons of the form
(intN, intN) -> int1, as the static type signature was(APInt, APInt) -> APInt. However, float comparisons map(floatN, floatN) -> int1, with a static type signature of(APFloat, APFloat) -> APInt. This use case wasn't supported byconstFoldBinaryOp.constFoldBinaryOpnow accepts an optional template argument overriding the return type in case it differs from the input type. If the new template argument isn't provided, the default behavior is unchanged (i.e. the return type will be assumed to match the input type).constFoldUnaryOpreceived similar changes in order to support folding non-cast ops of the form(T1) -> T2(e.g. asignop mapping(floatN) -> sint32).