[Mlir-commits] [mlir] [mlir] Compose expand of collapse to cast (PR #172864)

[Maya Amrami]

================
@@ -748,14 +748,18 @@ bool CastOp::areCastCompatible(TypeRange inputs, TypeRange outputs) {
       // source memref is static and the value in the target memref is the
       // same. They are also compatible if either one is dynamic (see
       // description of MemRefCastOp for details).
+      // Note that for dimensions of size 1, the stride can differ.
       auto checkCompatible = [](int64_t a, int64_t b) {
         return (ShapedType::isDynamic(a) || ShapedType::isDynamic(b) || a == b);
       };
       if (!checkCompatible(aOffset, bOffset))
         return false;
-      for (const auto &aStride : enumerate(aStrides))
-        if (!checkCompatible(aStride.value(), bStrides[aStride.index()]))
+      for (const auto &[index, aStride] : enumerate(aStrides)) {
+        if (aT.getDimSize(index) == 1)
----------------
amrami wrote:

This is valid for all dimensions.
Adding another lit test that shows the mismatch in strides when the dimension is internal.

https://github.com/llvm/llvm-project/pull/172864