[Mlir-commits] [mlir] [mlir][memref] Fix computeCollapsedLayoutMap for contiguous dynamic dim (PR #136485)

[Matthias Springer]

================
@@ -2401,11 +2402,19 @@ computeCollapsedLayoutMap(MemRefType srcType,
     if (!ShapedType::isDynamic(srcShape[ref.back()]) || ref.size() == 1) {
       resultStrides.push_back(srcStrides[ref.back()]);
     } else {
-      // Dynamically-sized dims may turn out to be dims of size 1 at runtime, so
-      // the corresponding stride may have to be skipped. (See above comment.)
-      // Therefore, the result stride cannot be statically determined and must
-      // be dynamic.
-      resultStrides.push_back(ShapedType::kDynamic);
+      bool contiguousSrcDim = srcStrides[ref.back()] == 1;
+      bool dynamicSizeIsPreserved =
+          std::all_of(ref.begin(), ref.end() - 1,
----------------
matthias-springer wrote:

Why is it `ref.begin() -> ref.end() - 1`? All dimensions except for one must be `1`, right? In that case, it does not matter where non-unit dimension is?

https://github.com/llvm/llvm-project/pull/136485