[Mlir-commits] [mlir] [mlir][Vector] Update patterns for flattening vector.xfer Ops (2/N) (PR #73523)

[Andrzej Warzyński]

================
@@ -487,26 +487,75 @@ class TransferWriteDropUnitDimsPattern
 
 } // namespace
 
-/// Return true if the memref type has its inner dimension matching the given
-/// shape. Otherwise return false.
-static int64_t hasMatchingInnerContigousShape(MemRefType memrefType,
-                                              ArrayRef<int64_t> targetShape) {
-  auto shape = memrefType.getShape();
-  SmallVector<int64_t> strides;
+/// Return true if `vectorType` is a contiguous slice of `memrefType`.
+///
+/// Compares `vectorType` against the trailing dimensions (*) of `memrefType`
+/// to check whether `vectorType` is a contiguous slice of `memrefType`.
+///
+/// There are two cases:
+///
+/// 1. The trailing dimensions of `memrefType` match the dimensions of
+/// `vectorType` excluding the front dim (the leading dim of `vectorType` does
+/// not matter in this case):
+///
+///   vector<2x4x3x2xi32> vs memref<5x4x3x2xi32> (contiguous slice)
+///   vector<2x4x2x2xi32> vs memref<5x4x3x2xi32> (non-contiguous slice)
+///
+/// 2. The trailing dimension of `memrefType` match the trailing dimensions of
+/// `vectorType` (i.e. at least 2 leading dims of `vectorType` don't match). The
+/// first dim of `vectorType` that does not match can be arbitrary, but the
+/// remaining leading dims have to be 1:
+///
+///   vector<1x1x2x2xi32> vs memref<5x4x3x2xi32> (contiguous slice)
+///   vector<2x1x2x2xi32> vs memref<5x4x3x2xi32> (non-contiguous slice)
+///
+/// In both cases `memrefType` has to be contiguous (this is checked by looking
+/// at strides).
+///
+/// (*) Only relevant in cases when the rank(vectorType) < rank(memrefType)
+/// TODO: Update
+static bool isContiguousSlice(MemRefType memrefType, VectorType vectorType) {
+
+  ArrayRef<int64_t> targetShape = vectorType.getShape();
+  auto targetShapeTrailingDims = targetShape.drop_front(1);
+
+  // Not used
----------------
banach-space wrote:

Change from https://github.com/llvm/llvm-project/pull/73522

https://github.com/llvm/llvm-project/pull/73523