[Mlir-commits] [mlir] [MLIR][XeGPU] Add lowering from transfer_read/transfer_write to load_gather/store_scatter (PR #152429)

[Jianhui Li]

================
@@ -155,6 +152,300 @@ createNdDescriptor(PatternRewriter &rewriter, Location loc,
   return ndDesc;
 }
 
+// Adjusts the strides of a memref according to a given permutation map for
+// vector operations.
+//
+// This function updates the last `vecRank` elements of the `strides` array to
+// reflect the permutation specified by `permMap`. The permutation is applied
+// to the innermost dimensions of the memref, corresponding to the vector
+// shape. This is typically used when lowering vector transfer operations with
+// permutation maps to memory accesses, ensuring that the memory strides match
+// the logical permutation of vector dimensions.
+//
+// Example:
+//   Suppose we have a memref of rank 4 with strides `[s0, s1, s2, s3]` and a
+//   vector of rank 2. If the permutation map swaps the last two dimensions
+//   (e.g., [0, 1] -> [1, 0]), then after calling this function, the last two
+//   strides will be swapped:
+//     Original strides: [s0, s1, s2, s3]
+//     After permutation: [s0, s1, s3, s2]
+//
+void adjustStridesForPermutation(Operation *op, PatternRewriter &rewriter,
----------------
Jianhui-Li wrote:

added

https://github.com/llvm/llvm-project/pull/152429