[mlir][vector] Add a check to ensure input vector rank equals target shape rank

[CoTinker]

The crash is caused because, during IR transformation, the vector-unrolling pass (using ExtractStridedSliceOp) attempts to slice an
input vector of higher rank using a target vector of lower rank, which is not supported. Fixes #148368.

[llvmbot]

@llvm/pr-subscribers-mlir-vector

@llvm/pr-subscribers-mlir

Author: Longsheng Mou (CoTinker)

Changes

The crash is caused because, during IR transformation, the vector-unrolling pass (using ExtractStridedSliceOp) attempts to slice an
input vector of higher rank using a target vector of lower rank, which is not supported. Fixed #148368.

---

Full diff: https://github.com/llvm/llvm-project/pull/149239.diff

2 Files Affected:

• (modified) mlir/lib/Dialect/Vector/Transforms/VectorUnroll.cpp (+15-1)
• (modified) mlir/test/Dialect/Vector/vector-unroll-options.mlir (+22)

diff --git a/mlir/lib/Dialect/Vector/Transforms/VectorUnroll.cpp b/mlir/lib/Dialect/Vector/Transforms/VectorUnroll.cpp
index 693f4f955994d..be911901c2afc 100644
--- a/mlir/lib/Dialect/Vector/Transforms/VectorUnroll.cpp
+++ b/mlir/lib/Dialect/Vector/Transforms/VectorUnroll.cpp
@@ -169,7 +169,13 @@ struct UnrollTransferReadPattern
     auto sourceVectorType = readOp.getVectorType();
     SmallVector<int64_t> strides(targetShape->size(), 1);
     Location loc = readOp.getLoc();
-    ArrayRef<int64_t> originalSize = readOp.getVectorType().getShape();
+    ArrayRef<int64_t> originalSize = sourceVectorType.getShape();
+    // Bail-out if rank(source) != rank(target). The main limitation here is the
+    // fact that `InsertStridedSliceOp` requires the rank for the input and
+    // output to match. If needed, we can relax this later.
+    if (originalSize.size() != targetShape->size())
+      return rewriter.notifyMatchFailure(
+          readOp, "expected source vector rank to match target shape rank");
 
     // Prepare the result vector;
     Value result = rewriter.create<arith::ConstantOp>(
@@ -224,6 +230,14 @@ struct UnrollTransferWritePattern
     SmallVector<int64_t> strides(targetShape->size(), 1);
     Location loc = writeOp.getLoc();
     ArrayRef<int64_t> originalSize = sourceVectorType.getShape();
+    // Bail-out if rank(source) != rank(target). The main limitation here is the
+    // fact that `ExtractStridedSlice` requires the rank for the input and
+    // output to match. If needed, we can relax this later.
+    if (originalSize.size() != targetShape->size())
+      return rewriter.notifyMatchFailure(
+          writeOp,
+          "expected source input vector rank to match target shape rank");
+
     SmallVector<Value> originalIndices(writeOp.getIndices().begin(),
                                        writeOp.getIndices().end());
     SmallVector<int64_t> loopOrder =
diff --git a/mlir/test/Dialect/Vector/vector-unroll-options.mlir b/mlir/test/Dialect/Vector/vector-unroll-options.mlir
index e129cd5c40b9c..8f6945468feb3 100644
--- a/mlir/test/Dialect/Vector/vector-unroll-options.mlir
+++ b/mlir/test/Dialect/Vector/vector-unroll-options.mlir
@@ -420,3 +420,25 @@ func.func @vector_store_2D(%mem: memref<4x4xf16>, %v: vector<4x4xf16>) {
   // CHECK: vector.store %[[V2]], %[[ARG0]][%[[C2]], %[[C0]]] : memref<4x4xf16>, vector<2x2xf16>
   // CHECK: %[[V3:.*]] = vector.extract_strided_slice %[[ARG1]] {offsets = [2, 2], sizes = [2, 2], strides = [1, 1]} : vector<4x4xf16> to vector<2x2xf16>
   // CHECK: vector.store %[[V3]], %[[ARG0]][%[[C2]], %[[C2]]] : memref<4x4xf16>, vector<2x2xf16>
+
+func.func @vector_transfer_read(%arg0: memref<6x34x62xi8>) -> vector<6x34x62xi8> {
+  %c0_i8 = arith.constant 0 : i8
+  %c0 = arith.constant 0 : index
+  %0 = vector.transfer_read %arg0[%c0, %c0, %c0], %c0_i8 : memref<6x34x62xi8>, vector<6x34x62xi8>
+  return %0 : vector<6x34x62xi8>
+}
+// CHECK-LABEL: func @vector_transfer_read
+//   CHECK-NOT: vector.intert_strided_slice
+//       CHECK: vector.transfer_read
+//       CHECK: return
+
+func.func @vector_transfer_write(%arg0: vector<6x34x62xi8>) {
+  %c0 = arith.constant 0 : index
+  %alloc = memref.alloc() : memref<6x34x62xi8>
+  vector.transfer_write %arg0, %alloc[%c0, %c0, %c0]: vector<6x34x62xi8>, memref<6x34x62xi8>
+  return
+}
+// CHECK-LABEL: func @vector_transfer_write
+//   CHECK-NOT: vector.extract_strided_slice
+//       CHECK: vector.transfer_write
+//       CHECK: return

[newling]

Thank you for providing a fix!

[newling]

There's a file vector-transfer-unroll.mlir which might be a better place for these tests.

Please add comments above the tests describing what is being tested.

Such tests generally are prefixed with negative_

[newling]

insert_strided_slice can have a smaller source vector: https://mlir.llvm.org/docs/Dialects/Vector/#vectorinsert_strided_slice-vectorinsertstridedsliceop

[newling]

Hi @CoTinker thanks for changes. Just to note: in general it is preferable to not squash commits, it is easier for reviewers to follow what's happening if the commits are kept. When PR's are merged, the commits get squashed into 1 by default, but in the review process it's nice to see the incremental changes. Just for future reference. Thanks!

[CoTinker]

Hi @CoTinker thanks for changes. Just to note: in general it is preferable to not squash commits, it is easier for reviewers to follow what's happening if the commits are kept. When PR's are merged, the commits get squashed into 1 by default, but in the review process it's nice to see the incremental changes. Just for future reference. Thanks!

I see, thanks for your advice.

[banach-space]

Thanks!

[dcaballe]

Thanks!

[banach-space]

LG % request for a comment in test. Thanks!

@newling, are you OK with this change?