[clang-tools-extra] [clang] [mlir] [llvm] [MLIR][Linalg] Support dynamic sizes in `lower_unpack` (PR #75494)

[lorenzo chelini via cfe-commits]

================
@@ -434,8 +429,21 @@ FailureOr<LowerUnPackOpResult> linalg::lowerUnPack(RewriterBase &rewriter,
       RankedTensorType::Builder(packedTensorType).setShape(stripMinedShape);
   RankedTensorType collapsedType = tensor::CollapseShapeOp::inferCollapsedType(
       stripMinedTensorType, packingMetadata.reassociations);
+
+  // Get dynamic dims from input tensor in order of stripMinedTensor
+  // `tensor.empty` op
+  SmallVector<OpFoldResult, 4> dims =
+      tensor::getMixedSizes(rewriter, loc, unPackOp.getSource());
+  applyPermutationToVector(dims, lastDimsToInsertPositionsPerm);
+  SmallVector<Value, 4> dynDims;
----------------
chelini wrote:

We can avoid the loop using another builder method for emptyOp:

```
SmallVector<OpFoldResult> dims =
      tensor::getMixedSizes(rewriter, loc, unPackOp.getSource());
applyPermutationToVector(dims, lastDimsToInsertPositionsPerm);

auto emptyOp = rewriter.create<tensor::EmptyOp>(
      loc, dims, stripMinedTensorType.getElementType());
```


https://github.com/llvm/llvm-project/pull/75494