Skip to content

[mlir][Tensor] Check for out-of-bounds slice in insert/extract_slice verifier #130487

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
merged 2 commits into from
Mar 12, 2025
Merged

[mlir][Tensor] Check for out-of-bounds slice in insert/extract_slice verifier #130487

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
654a40d
[mlir][Tensor] Check for out-of-bounds extraction in `extract_slice` …
matthias-springer Mar 9, 2025
8690690
address comments
matthias-springer Mar 11, 2025
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
61 changes: 57 additions & 4 deletions mlir/lib/Dialect/Tensor/IR/TensorOps.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -2352,13 +2352,52 @@ static LogicalResult produceSliceErrorMsg(SliceVerificationResult result,
}
}

/// Verify that the offsets/sizes/strides-style access into the given tensor
/// is in-bounds. Only static information is verified.
static LogicalResult verifyInBoundsSlice(Operation *op,
RankedTensorType tensorType,
ArrayRef<int64_t> staticOffsets,
ArrayRef<int64_t> staticSizes,
ArrayRef<int64_t> staticStrides) {
for (int64_t i = 0, e = tensorType.getRank(); i < e; ++i) {
// Nothing to verify for dynamic source dims.
if (tensorType.isDynamicDim(i))
continue;
// Nothing to verify if the offset is dynamic.
if (ShapedType::isDynamic(staticOffsets[i]))
continue;
if (staticOffsets[i] >= tensorType.getDimSize(i))
return op->emitOpError("offset ")
<< i << " is out-of-bounds: " << staticOffsets[i]
<< " >= " << tensorType.getDimSize(i);
if (ShapedType::isDynamic(staticSizes[i]) ||
ShapedType::isDynamic(staticStrides[i]))
continue;
int64_t lastPos =
staticOffsets[i] + (staticSizes[i] - 1) * staticStrides[i];
if (lastPos >= tensorType.getDimSize(i))
return op->emitOpError("slice along dimension ")
<< i << " runs out-of-bounds: " << lastPos
<< " >= " << tensorType.getDimSize(i);
}
return success();
}

/// Verifier for ExtractSliceOp.
LogicalResult ExtractSliceOp::verify() {
RankedTensorType sourceType = getSourceType();

// Verify result type against inferred type.
RankedTensorType expectedType = ExtractSliceOp::inferResultType(
getSourceType(), getMixedOffsets(), getMixedSizes(), getMixedStrides());
sourceType, getMixedOffsets(), getMixedSizes(), getMixedStrides());
SliceVerificationResult result = isRankReducedType(expectedType, getType());
return produceSliceErrorMsg(result, *this, expectedType);
if (result != SliceVerificationResult::Success)
return produceSliceErrorMsg(result, *this, expectedType);

// Verify that offsets, sizes, strides do not run out-of-bounds with respect
// to the source tensor.
return verifyInBoundsSlice(getOperation(), sourceType, getStaticOffsets(),
getStaticSizes(), getStaticStrides());
}

llvm::SmallBitVector ExtractSliceOp::getDroppedDims() {
Expand Down Expand Up @@ -2729,11 +2768,18 @@ static SliceVerificationResult verifyInsertSliceOp(

/// Verifier for InsertSliceOp.
LogicalResult InsertSliceOp::verify() {
// Verify result type against inferred type.
RankedTensorType expectedType;
SliceVerificationResult result =
verifyInsertSliceOp(getSourceType(), getType(), getStaticOffsets(),
getStaticSizes(), getStaticStrides(), &expectedType);
return produceSliceErrorMsg(result, *this, expectedType);
if (result != SliceVerificationResult::Success)
return produceSliceErrorMsg(result, *this, expectedType);

// Verify that offsets, sizes, strides do not run out-of-bounds with respect
// to the source tensor.
return verifyInBoundsSlice(getOperation(), getDestType(), getStaticOffsets(),
getStaticSizes(), getStaticStrides());
}

/// If we have two consecutive InsertSliceOp writing to the same slice, we
Expand Down Expand Up @@ -3747,11 +3793,18 @@ LogicalResult ParallelInsertSliceOp::verify() {
return this->emitError("expected ParallelCombiningOpInterface parent, got:")
<< *(getOperation()->getParentOp());

// Verify result type against inferred type.
RankedTensorType expectedType;
SliceVerificationResult result =
verifyInsertSliceOp(getSourceType(), getDestType(), getStaticOffsets(),
getStaticSizes(), getStaticStrides(), &expectedType);
return produceSliceErrorMsg(result, *this, expectedType);
if (result != SliceVerificationResult::Success)
return produceSliceErrorMsg(result, *this, expectedType);

// Verify that offsets, sizes, strides do not run out-of-bounds with respect
// to the source tensor.
return verifyInBoundsSlice(getOperation(), getDestType(), getStaticOffsets(),
getStaticSizes(), getStaticStrides());
}

void ParallelInsertSliceOp::getCanonicalizationPatterns(
Expand Down
2 changes: 1 addition & 1 deletion mlir/test/Dialect/Tensor/bubble-up-extract-slice-op.mlir
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -31,7 +31,7 @@ func.func @no_bubble_up_extract_slice_on_non_contiguous(%src: tensor<60xf32>) ->

func.func @no_bubble_up_extract_slice_on_stride(%src: tensor<60xf32>) -> tensor<1x1x5xf32> {
%expand = tensor.expand_shape %src [[0, 1, 2]] output_shape [2, 3, 10] : tensor<60xf32> into tensor<2x3x10xf32>
%extract = tensor.extract_slice %expand[0, 0, 5][1, 1, 5][1, 1, 2] : tensor<2x3x10xf32> to tensor<1x1x5xf32>
%extract = tensor.extract_slice %expand[0, 0, 0][1, 1, 5][1, 1, 2] : tensor<2x3x10xf32> to tensor<1x1x5xf32>
return %extract : tensor<1x1x5xf32>
}

Expand Down
8 changes: 4 additions & 4 deletions mlir/test/Dialect/Tensor/drop-redundant-insert-slice-rank-expansion.mlir
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -51,16 +51,16 @@ func.func @no_fold_more_unit_dims_insert_slice_of_extract_slice(%in : tensor<?x8

// -----

func.func @no_fold_strided_insert_slice_of_extract_slice(%in : tensor<?x8x2x8xf32>, %dest : tensor<1x4x4xf32>) -> tensor<1x4x4xf32> {
func.func @no_fold_strided_insert_slice_of_extract_slice(%in : tensor<?x8x2x8xf32>, %dest : tensor<1x15x15xf32>) -> tensor<1x15x15xf32> {
%extracted_slice = tensor.extract_slice %in[0, 0, 0, 0] [1, 8, 1, 8] [1, 1, 1, 1] : tensor<?x8x2x8xf32> to tensor<8x8xf32>
%inserted_slice = tensor.insert_slice %extracted_slice into %dest[0, 0, 0] [1, 8, 8] [1, 2, 2] : tensor<8x8xf32> into tensor<1x4x4xf32>
return %inserted_slice : tensor<1x4x4xf32>
%inserted_slice = tensor.insert_slice %extracted_slice into %dest[0, 0, 0] [1, 8, 8] [1, 2, 2] : tensor<8x8xf32> into tensor<1x15x15xf32>
return %inserted_slice : tensor<1x15x15xf32>
}
// CHECK-LABEL: func.func @no_fold_strided_insert_slice_of_extract_slice(
// CHECK-SAME: %[[ARG0:.*]]: tensor<?x8x2x8xf32>
// CHECK: %[[EXTRACTED_SLICE:.*]] = tensor.extract_slice %[[ARG0]]
// CHECK: %[[INSERTED_SLICE:.*]] = tensor.insert_slice %[[EXTRACTED_SLICE]]
// CHECK: return %[[INSERTED_SLICE]] : tensor<1x4x4xf32>
// CHECK: return %[[INSERTED_SLICE]] : tensor<1x15x15xf32>

// -----

Expand Down
16 changes: 6 additions & 10 deletions mlir/test/Dialect/Tensor/fold-tensor-subset-ops.mlir
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -271,38 +271,34 @@ func.func @insert_slice_of_transfer_write_rank_extending(%t1 : tensor<?x?x12xf32

// -----

// CHECK: #[[$map:.*]] = affine_map<()[s0] -> (s0 + 2)>
// CHECK-LABEL: func @insert_slice_of_insert_slice(
// CHECK-SAME: %[[t:[0-9a-z]*]]: tensor<f32>
// CHECK-SAME: %[[r1:[0-9a-z]*]]: tensor<1x14xf32>
// CHECK-SAME: %[[pos:[0-9a-z]*]]: index
// CHECK: %[[add:.*]] = affine.apply #[[$map]]()[%[[pos]]]
// CHECK: tensor.insert_slice %[[t]] into %[[r1]][4, %[[add]]] [1, 1] [1, 1] : tensor<f32> into tensor<1x14xf32>
// CHECK: tensor.insert_slice %[[t]] into %[[r1]][0, %[[pos]]] [1, 1] [1, 1] : tensor<f32> into tensor<1x14xf32>
func.func @insert_slice_of_insert_slice(%t: tensor<f32>, %r0: tensor<1x1xf32>, %r1: tensor<1x14xf32>, %pos: index)
-> tensor<1x14xf32>
{
%0 = tensor.insert_slice %t into %r0[1, 2] [1, 1] [1, 1]
%0 = tensor.insert_slice %t into %r0[0, 0] [1, 1] [1, 1]
Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Making it all zero might be losing some test coverage. Maybe increase the size of the tensor instead?

Copy link
Member Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

This must be [0, 0], otherwise the two insert_slice cannot be merged.

Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

I dont see why. The existing test is taking an offset of [1, 2] and composing it with [3, %pos] to get [4, %pos + 2] . That seems correct to me. This change is changing the test behavior and potentially dropping coverage.

Copy link
Member Author

@matthias-springer matthias-springer Mar 11, 2025 •
edited
Loading

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

The previous test was incorrect:

// This op is invalid because it inserts into %r0 at out-of-bounds offsets.
%0 = tensor.insert_slice %t into %r0[1, 2] [1, 1] [1, 1]
    : tensor<f32> into tensor<1x1xf32>

To fix it, let's try to make the tensor larger but keep the offsets, as you suggested (actually, we have to make both tensors larger now).

%0 = tensor.insert_slice %t into %r0[1, 2] [1, 1] [1, 1]
    : tensor<f32> into tensor<2x3xf32>
%1 = tensor.insert_slice %0 into %r1[0, %pos] [2, 3] [1, 1] 
    : tensor<2x3xf32> into tensor<2x14xf32>

These two tensor.insert_slice cannot be folded into a single tensor.insert_slice. The reason is because we need to insert both %t and some data from %r0 into %r1.

So the only way to fix the test (so that it still exercises the folding pattern) is to set the offsets to 0.

The existing test is taking an offset of [1, 2] and composing it with [3, %pos] to get [4, %pos + 2] . That seems correct to me. This change is changing the test behavior and potentially dropping coverage.

I think there's no way trigger this composition without the first insert_slice being out-of-bounds. The folding pattern code can probably be simplified.

Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

oh, that makes sense. Thanks for the explanation.

: tensor<f32> into tensor<1x1xf32>
%1 = tensor.insert_slice %0 into %r1[3, %pos] [1, 1] [1, 1]
%1 = tensor.insert_slice %0 into %r1[0, %pos] [1, 1] [1, 1]
: tensor<1x1xf32> into tensor<1x14xf32>
return %1 : tensor<1x14xf32>
}

// -----

// CHECK-DAG: #[[$map:.*]] = affine_map<()[s0] -> (s0 + 2)>
// CHECK-LABEL: func @insert_slice_of_insert_slice(
// CHECK-SAME: %[[t:[0-9a-z]*]]: tensor<f32>
// CHECK-SAME: %[[r1:[0-9a-z]*]]: tensor<1x14xf32>
// CHECK-SAME: %[[pos:[0-9a-z]*]]: index
// CHECK: %[[composed_pos:.+]] = affine.apply #[[$map]]()[%[[pos]]]
// CHECK: tensor.insert_slice %[[t]] into %[[r1]][3, %[[composed_pos]]] [1, 1] [1, 1] : tensor<f32> into tensor<1x14xf32>
// CHECK: tensor.insert_slice %[[t]] into %[[r1]][0, %[[pos]]] [1, 1] [1, 1] : tensor<f32> into tensor<1x14xf32>
func.func @insert_slice_of_insert_slice(%t: tensor<f32>, %r0: tensor<1xf32>, %r1: tensor<1x14xf32>, %pos: index)
-> tensor<1x14xf32>
{
%0 = tensor.insert_slice %t into %r0[2] [1] [1]
%0 = tensor.insert_slice %t into %r0[0] [1] [1]
: tensor<f32> into tensor<1xf32>
%1 = tensor.insert_slice %0 into %r1[3, %pos] [1, 1] [1, 1]
%1 = tensor.insert_slice %0 into %r1[0, %pos] [1, 1] [1, 1]
: tensor<1xf32> into tensor<1x14xf32>
return %1 : tensor<1x14xf32>
}
Expand Down
32 changes: 32 additions & 0 deletions mlir/test/Dialect/Tensor/invalid.mlir
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -258,6 +258,22 @@ func.func @illegal_num_offsets(%arg0 : tensor<?x?x?xf32>, %arg1 : index, %arg2 :

// -----

func.func @extract_slice_offset_out_of_bounds(%arg0: tensor<10xf32>) {
// expected-error@+1 {{offset 0 is out-of-bounds: 10 >= 10}}
%0 = tensor.extract_slice %arg0 [10][1][1] : tensor<10xf32> to tensor<1xf32>
return
}

// -----

func.func @extract_slice_runs_out_of_bounds(%arg0: tensor<9xf32>) {
// expected-error@+1 {{slice along dimension 0 runs out-of-bounds: 9 >= 9}}
%0 = tensor.extract_slice %arg0 [3][4][2] : tensor<9xf32> to tensor<4xf32>
return
}

// -----

func.func @insert_slice_wrong_result_rank(%t1: tensor<?xf32>, %t2: tensor<?x?xf32>, %idx : index) {
// expected-error @+1 {{expected rank to be smaller or equal to the other rank.}}
%0 = tensor.insert_slice %t2 into %t1[0][4][1] : tensor<?x?xf32> into tensor<?xf32>
Expand Down Expand Up @@ -296,6 +312,22 @@ func.func @insert_slice_wrong_dynamic_type(%t1: tensor<?x4x4xf32>, %t2: tensor<8

// -----

func.func @insert_slice_offset_out_of_bounds(%arg0: tensor<1xf32>, %arg1: tensor<10xf32>) {
// expected-error@+1 {{offset 0 is out-of-bounds: 10 >= 10}}
%0 = tensor.insert_slice %arg0 into %arg1 [10][1][1] : tensor<1xf32> into tensor<10xf32>
return
}

// -----

func.func @insert_slice_runs_out_of_bounds(%arg0: tensor<4xf32>, %arg1: tensor<9xf32>) {
// expected-error@+1 {{slice along dimension 0 runs out-of-bounds: 9 >= 9}}
%0 = tensor.insert_slice %arg0 into %arg1 [3][4][2] : tensor<4xf32> into tensor<9xf32>
return
}

// -----

func.func @illegal_expanding_reshape_static_tensor
(%arg0: tensor<2x3x20xf32>) -> tensor<2x3x2x4x5xf32> {
// expected-error @+1 {{expected dimension 2 of collapsed type to be static value of 40}}
Expand Down