[Mlir-commits] [mlir] [mlir][tensor] Fix crash in expand_shape fold with dynamic result type (PR #183785)

[llvmlistbot at llvm.org]

llvmbot wrote:


<!--LLVM PR SUMMARY COMMENT-->

@llvm/pr-subscribers-mlir-tensor

Author: Mehdi Amini (joker-eph)

<details>
<summary>Changes</summary>

`foldReshapeOp` (in `ReshapeOpsUtils.h`) and `FoldReshapeWithConstant` (in `TensorOps.cpp`) both tried to create a new `DenseElementsAttr` constant when folding a reshape op whose operand is a constant. Neither checked that the result type was statically shaped before doing so, but `DenseElementsAttr::reshape()` and `DenseElementsAttr::getFromRawBuffer()` both assert `hasStaticShape()`.

Guard both fold paths with a `hasStaticShape()` check so they return early when the result type contains a dynamic dimension.

Fixes #<!-- -->177845

---
Full diff: https://github.com/llvm/llvm-project/pull/183785.diff


3 Files Affected:

- (modified) mlir/include/mlir/Dialect/Utils/ReshapeOpsUtils.h (+8-3) 
- (modified) mlir/lib/Dialect/Tensor/IR/TensorOps.cpp (+4) 
- (modified) mlir/test/Dialect/Tensor/canonicalize.mlir (+16) 


``````````diff
diff --git a/mlir/include/mlir/Dialect/Utils/ReshapeOpsUtils.h b/mlir/include/mlir/Dialect/Utils/ReshapeOpsUtils.h
index 525663e5cd6c5..2e8c0e269995e 100644
--- a/mlir/include/mlir/Dialect/Utils/ReshapeOpsUtils.h
+++ b/mlir/include/mlir/Dialect/Utils/ReshapeOpsUtils.h
@@ -90,9 +90,14 @@ static OpFoldResult foldReshapeOp(ReshapeOpTy reshapeOp,
   if (reshapeOp.getSrcType() == reshapeOp.getType())
     return reshapeOp.getSrc();
 
-  // Reshape of a constant can be replaced with a new constant.
-  if (auto elements = dyn_cast_or_null<DenseElementsAttr>(operands.front()))
-    return elements.reshape(cast<ShapedType>(reshapeOp.getResult().getType()));
+  // Reshape of a constant can be replaced with a new constant, but only when
+  // the result type has a static shape. DenseElementsAttr::reshape requires
+  // a static shape to preserve the element count invariant.
+  if (auto elements = dyn_cast_or_null<DenseElementsAttr>(operands.front())) {
+    auto resultType = cast<ShapedType>(reshapeOp.getResult().getType());
+    if (resultType.hasStaticShape())
+      return elements.reshape(resultType);
+  }
 
   // Fold if the producer reshape source has the same shape with at most 1
   // dynamic dimension.
diff --git a/mlir/lib/Dialect/Tensor/IR/TensorOps.cpp b/mlir/lib/Dialect/Tensor/IR/TensorOps.cpp
index 4c0ab7c9ec8a0..30b56dbcc10f3 100644
--- a/mlir/lib/Dialect/Tensor/IR/TensorOps.cpp
+++ b/mlir/lib/Dialect/Tensor/IR/TensorOps.cpp
@@ -2119,6 +2119,10 @@ struct FoldReshapeWithConstant : OpRewritePattern<TensorReshapeOp> {
       return failure();
     if (!attr || !attr.isSplat())
       return failure();
+    // DenseElementsAttr requires a static shape; skip folding for dynamic
+    // result types.
+    if (!reshapeOp.getResultType().hasStaticShape())
+      return failure();
     DenseElementsAttr newAttr = DenseElementsAttr::getFromRawBuffer(
         reshapeOp.getResultType(), attr.getRawData());
     rewriter.replaceOpWithNewOp<arith::ConstantOp>(reshapeOp, newAttr);
diff --git a/mlir/test/Dialect/Tensor/canonicalize.mlir b/mlir/test/Dialect/Tensor/canonicalize.mlir
index fc499da5422fc..b6ae7b2168774 100644
--- a/mlir/test/Dialect/Tensor/canonicalize.mlir
+++ b/mlir/test/Dialect/Tensor/canonicalize.mlir
@@ -1672,6 +1672,22 @@ func.func @reshape_splat_constant_float64() -> tensor<2x4x2xf64> {
 
 // -----
 
+// Regression test for https://github.com/llvm/llvm-project/issues/177845:
+// tensor.expand_shape of a constant to a dynamic shape must not crash.
+// FoldReshapeWithConstant must not call DenseElementsAttr::getFromRawBuffer
+// when the result type is dynamic (getFromRawBuffer requires static shape).
+
+// CHECK-LABEL: @expand_shape_splat_constant_dynamic_result
+//       CHECK:   arith.constant
+//       CHECK:   tensor.expand_shape
+func.func @expand_shape_splat_constant_dynamic_result(%n: index) -> tensor<?xi32> {
+  %cst = arith.constant dense<1> : tensor<i32>
+  %result = tensor.expand_shape %cst [] output_shape [%n] : tensor<i32> into tensor<?xi32>
+  return %result : tensor<?xi32>
+}
+
+// -----
+
 // CHECK-LABEL: func @fold_rank
 func.func @fold_rank() -> (index) {
   %const_0 = arith.constant dense<[[[1, -2, 1, 36]], [[0, 2, -1, 64]]]>

``````````

</details>


https://github.com/llvm/llvm-project/pull/183785