[Mlir-commits] [mlir] [MLIR] Make 1-D memref flattening a prerequisite for vector narrow type emulation (PR #157771)

[Andrzej Warzyński]

================
@@ -931,18 +938,27 @@ struct ConvertVectorMaskedStore final
 // ConvertVectorLoad
 //===----------------------------------------------------------------------===//
 
-// TODO: Document-me
+/// Converts `vector.load` on narrow element types to work with
+/// wider, byte-aligned container types by adjusting load sizes and using
+/// bitcasting.
+///
+/// Example: `vector.load` of `vector<4xi4>` from `memref<3x4xi4>` is emulated
+/// by loading `vector<2xi8>` from the linearized `memref<6xi8>` and bitcasting
+/// back, since each `i8` container holds two `i4` values.
----------------
banach-space wrote:

[nit]
```suggestion
/// Example: `vector.load` of `vector<4xi4>` from `memref<3x4xi4>` is emulated
/// by loading `vector<2xi8>` from the linearized `memref<6xi8>` (each `i8` container holds two `i4` values) and bitcasting  back.
```

https://github.com/llvm/llvm-project/pull/157771