llvm-project

2a96d573 - [mlir][memref] Enforce consistent reinterpret_cast metadata (#217338)

[mlir][memref] Enforce consistent reinterpret_cast metadata (#217338) This PR implements [[RFC] Clarify `memref.reinterpret_cast` Verification of Dynamic Metadata](https://discourse.llvm.org/t/rfc-clarify-memref-reinterpret-cast-verification-of-dynamic-metadata). ## Context `memref.reinterpret_cast` describes its result through both its mixed offset/size/stride metadata and its result `MemRefType`. Currently, verification allowed a dynamic result-type correspond to static operation metadata, introducing inconsistent descriptors. For example, the verifier permits: ```mlir %r = memref.reinterpret_cast %base to offset: [0], sizes: [%n, 2], strides: [%s, 1] : memref<f32> to memref<?x?xf32, strided<[?, ?], offset: ?>> ``` where the operation constructs a descriptor with a statically known second size and stride, but the result type does not expose those facts. ## Changes This PR consolidates the verifier to require `reinterpret_cast` result type’s sizes, offset, and strides to exactly match the static/dynamic representation and values of the operation metadata. Transforms that create `memref.reinterpret_cast`, including GPU memref decomposition and strided-metadata expansion for `subview` and `collapse_shape`, now derive the result type from the final, potentially folded descriptor instead of inheriting the original operation’s result type. When this descriptor-derived type differs from the original result type, the rewrite inserts a `memref.cast` back to the original type. These rewrites are not type-conversion patterns, so their existing users—including function returns and calls—must retain their original types. Identity stride layouts are canonicalized when possible to avoid unnecessary explicit layouts and casts. Tests are updated to use result types consistent with their `reinterpret_cast` metadata. In the SPIR-V tests, the size `10` is represented by an SSA value because those cases intentionally remain dynamically shaped and lower to element pointers; making the size static would change the SPIR-V pointee type and prevent the existing `reinterpret_cast` lowering from applying. A regression test also verifies that folded `subview` offsets and strides are reflected in the generated `reinterpret_cast` type. --- Assisted-by: Codex (refine implementation + tests). I reviewed all code and tests before submission.