[Mlir-commits] [mlir] 6067129 - Revert "[mlir][vector] Add a pattern to fuse extract(constant_mask) (#81057)" (#83275)
llvmlistbot at llvm.org
llvmlistbot at llvm.org
Wed Feb 28 07:53:01 PST 2024
Author: Quinn Dawkins
Date: 2024-02-28T10:52:57-05:00
New Revision: 6067129fbe8944d5bd39dc1bc2331ce5ed6ecee8
URL: https://github.com/llvm/llvm-project/commit/6067129fbe8944d5bd39dc1bc2331ce5ed6ecee8
DIFF: https://github.com/llvm/llvm-project/commit/6067129fbe8944d5bd39dc1bc2331ce5ed6ecee8.diff
LOG: Revert "[mlir][vector] Add a pattern to fuse extract(constant_mask) (#81057)" (#83275)
This reverts commit 5cdb8c0c8854d08ac7ca131ce3e8d78a32eb6716.
This pattern is producing incorrect IR. For example,
```mlir
func.func @extract_subvector_from_constant_mask() -> vector<16xi1> {
%mask = vector.constant_mask [2, 3] : vector<16x16xi1>
%extract = vector.extract %mask[8] : vector<16xi1> from vector<16x16xi1>
return %extract : vector<16xi1>
}
```
Canonicalizes to
```mlir
func.func @extract_subvector_from_constant_mask() -> vector<16xi1> {
%0 = vector.constant_mask [3] : vector<16xi1>
return %0 : vector<16xi1>
}
```
Where it should be a zero mask because the extraction index (8) is
greater than the constant mask size along that dim (2).
Added:
Modified:
mlir/lib/Dialect/Vector/IR/VectorOps.cpp
mlir/test/Dialect/Vector/canonicalize.mlir
Removed:
################################################################################
diff --git a/mlir/lib/Dialect/Vector/IR/VectorOps.cpp b/mlir/lib/Dialect/Vector/IR/VectorOps.cpp
index 8c341a347ff662..5be6a628904cdf 100644
--- a/mlir/lib/Dialect/Vector/IR/VectorOps.cpp
+++ b/mlir/lib/Dialect/Vector/IR/VectorOps.cpp
@@ -2040,77 +2040,6 @@ class ExtractOpFromCreateMask final : public OpRewritePattern<ExtractOp> {
}
};
-// Patterns to rewrite ExtractOp(ConstantMaskOp)
-//
-// When the result of ExtractOp is a subvector of input, we can rewrite it as
-// a ConstantMaskOp with subvector ranks.
-//
-// ExtractOp(ConstantMaskOp) -> ConstantMaskOp
-//
-// When the result of ExtractOp is a scalar, we can get the scalar value
-// directly.
-//
-// ExtractOp(ConstantMaskOp) -> ConstantOp
-class ExtractOpFromConstantMask final : public OpRewritePattern<ExtractOp> {
-public:
- using OpRewritePattern::OpRewritePattern;
-
- LogicalResult matchAndRewrite(ExtractOp extractOp,
- PatternRewriter &rewriter) const override {
- auto constantMaskOp =
- extractOp.getVector().getDefiningOp<vector::ConstantMaskOp>();
- if (!constantMaskOp)
- return failure();
-
- // All indices must be static.
- ArrayRef<int64_t> extractOpPos = extractOp.getStaticPosition();
- unsigned dynamicPosCount =
- llvm::count_if(extractOpPos, ShapedType::isDynamic);
- // If there is any dynamic position in ExtractOp, we cannot determine the
- // scalar value.
- if (dynamicPosCount)
- return failure();
-
- ArrayRef<Attribute> maskDimSizes =
- constantMaskOp.getMaskDimSizes().getValue();
- Type resultTy = extractOp.getResult().getType();
- if (resultTy.isa<mlir::VectorType>()) {
- auto resultVectorTy = resultTy.cast<mlir::VectorType>();
- int64_t resultRank = resultVectorTy.getRank();
- int64_t n = maskDimSizes.size();
- std::vector<int64_t> indices;
- for (auto i = n - resultRank; i < n; ++i)
- indices.push_back(cast<IntegerAttr>(maskDimSizes[i]).getInt());
-
- rewriter.replaceOpWithNewOp<vector::ConstantMaskOp>(
- extractOp, resultVectorTy,
- vector::getVectorSubscriptAttr(rewriter, indices));
-
- return success();
- } else if (resultTy.isa<mlir::IntegerType>()) {
- // ConstantMaskOp creates and returns a vector mask where elements of the
- // result vector are set to ‘0’ or ‘1’, based on whether the element
- // indices are contained within a hyper-rectangular region.
- // We go through ExtractOp static positions to determine the position is
- // within the hyper-rectangular region or not.
- Type boolType = rewriter.getI1Type();
- IntegerAttr setAttr = IntegerAttr::get(boolType, 1);
- for (size_t i = 0, end = extractOpPos.size(); i < end; ++i) {
- if (cast<IntegerAttr>(maskDimSizes[i]).getInt() <= extractOpPos[i]) {
- setAttr = IntegerAttr::get(boolType, 0);
- break;
- }
- }
-
- rewriter.replaceOpWithNewOp<arith::ConstantOp>(extractOp, boolType,
- setAttr);
- return success();
- }
-
- return failure();
- }
-};
-
// Folds extract(shape_cast(..)) into shape_cast when the total element count
// does not change.
LogicalResult foldExtractFromShapeCastToShapeCast(ExtractOp extractOp,
@@ -2137,8 +2066,7 @@ LogicalResult foldExtractFromShapeCastToShapeCast(ExtractOp extractOp,
void ExtractOp::getCanonicalizationPatterns(RewritePatternSet &results,
MLIRContext *context) {
results.add<ExtractOpSplatConstantFolder, ExtractOpNonSplatConstantFolder,
- ExtractOpFromBroadcast, ExtractOpFromCreateMask,
- ExtractOpFromConstantMask>(context);
+ ExtractOpFromBroadcast, ExtractOpFromCreateMask>(context);
results.add(foldExtractFromShapeCastToShapeCast);
}
diff --git a/mlir/test/Dialect/Vector/canonicalize.mlir b/mlir/test/Dialect/Vector/canonicalize.mlir
index dc486181ebe960..e6f045e12e5197 100644
--- a/mlir/test/Dialect/Vector/canonicalize.mlir
+++ b/mlir/test/Dialect/Vector/canonicalize.mlir
@@ -2567,76 +2567,3 @@ func.func @load_store_forwarding_rank_mismatch(%v0: vector<4x1x1xf32>, %arg0: te
tensor<4x4x4xf32>, vector<1x100x4x5xf32>
return %r : vector<1x100x4x5xf32>
}
-
-// -----
-
-// CHECK-LABEL: func.func @extract_true_from_constant_mask() -> i1 {
-func.func @extract_true_from_constant_mask() -> i1 {
-// CHECK: %[[TRUE:.*]] = arith.constant true
-// CHECK-NEXT: return %[[TRUE]] : i1
- %mask = vector.constant_mask [2, 2, 3] : vector<4x4x4xi1>
- %extract = vector.extract %mask[1, 1, 2] : i1 from vector<4x4x4xi1>
- return %extract : i1
-}
-
-// -----
-
-// CHECK-LABEL: func.func @extract_false_from_constant_mask() -> i1 {
-func.func @extract_false_from_constant_mask() -> i1 {
-// CHECK: %[[FALSE:.*]] = arith.constant false
-// CHECK-NEXT: return %[[FALSE]] : i1
- %mask = vector.constant_mask [2, 2, 3] : vector<4x4x4xi1>
- %extract = vector.extract %mask[1, 2, 2] : i1 from vector<4x4x4xi1>
- return %extract : i1
-}
-
-// -----
-
-// CHECK-LABEL: func.func @extract_from_create_mask() -> i1 {
-func.func @extract_from_create_mask() -> i1 {
-// CHECK: %[[TRUE:.*]] = arith.constant true
-// CHECK-NEXT: return %[[TRUE]] : i1
- %c2 = arith.constant 2 : index
- %c3 = arith.constant 3 : index
- %mask = vector.create_mask %c2, %c2, %c3 : vector<4x4x4xi1>
- %extract = vector.extract %mask[1, 1, 2] : i1 from vector<4x4x4xi1>
- return %extract : i1
-}
-
-// -----
-
-// CHECK-LABEL: func.func @extract_subvector_from_constant_mask() ->
-// CHECK-SAME: vector<6xi1> {
-func.func @extract_subvector_from_constant_mask() -> vector<6xi1> {
-// CHECK: %[[S0:.*]] = vector.constant_mask [4] : vector<6xi1>
-// CHECK-NEXT: return %[[S0]] : vector<6xi1>
- %mask = vector.constant_mask [2, 3, 4] : vector<4x5x6xi1>
- %extract = vector.extract %mask[1, 2] : vector<6xi1> from vector<4x5x6xi1>
- return %extract : vector<6xi1>
-}
-
-// -----
-
-// CHECK-LABEL: func.func @extract_scalar_with_dynamic_positions(
-// CHECK-SAME: %[[INDEX:.*]]: index) -> i1 {
-func.func @extract_scalar_with_dynamic_positions(%index: index) -> i1 {
-// CHECK: %[[S0:.*]] = vector.constant_mask [2, 2, 3] : vector<4x4x4xi1>
-// CHECK-NEXT: %[[S1:.*]] = vector.extract %[[S0]][1, 1, %[[INDEX]]] : i1 from vector<4x4x4xi1>
-// CHECK-NEXT: return %[[S1]] : i1
- %mask = vector.constant_mask [2, 2, 3] : vector<4x4x4xi1>
- %extract = vector.extract %mask[1, 1, %index] : i1 from vector<4x4x4xi1>
- return %extract : i1
-}
-
-// -----
-
-// CHECK-LABEL: func.func @extract_subvector_with_dynamic_positions
-// CHECK-SAME: %[[INDEX:.*]]: index) -> vector<6xi1> {
-func.func @extract_subvector_with_dynamic_positions(%index: index) -> vector<6xi1> {
-// CHECK: %[[S0:.*]] = vector.constant_mask [2, 3, 4] : vector<4x5x6xi1>
-// CHECK-NEXT: %[[S1:.*]] = vector.extract %[[S0]][1, %[[INDEX]]] : vector<6xi1> from vector<4x5x6xi1>
-// CHECK-NEXT: return %[[S1]] : vector<6xi1>
- %mask = vector.constant_mask [2, 3, 4] : vector<4x5x6xi1>
- %extract = vector.extract %mask[1, %index] : vector<6xi1> from vector<4x5x6xi1>
- return %extract : vector<6xi1>
-}
More information about the Mlir-commits
mailing list