[Mlir-commits] [mlir] [mlir][ArmSME] Lower multi-tile stores to a single loop (PR #96187)

[Cullen Rhodes]

================
@@ -373,6 +374,139 @@ struct LegalizeTransferWriteOpsByDecomposition
   }
 };
 
+/// Legalize a multi-tile transfer_write as a single store loop. This is done as
+/// part of type decomposition as at this level we know each tile write is
+/// disjoint, but that information is lost after decomposition (without analysis
+/// to reconstruct it).
+///
+/// Example:
+///
+/// ```
+/// vector.transfer_write %vector, %dest[%y, %x], %mask
+///   : vector<[16]x[8]xi16>, memref<?x?xi16>
+/// ```
+/// Is rewritten to:
+/// ```
+/// scf.for %slice_idx = %c0 to %c8_vscale step %c1 {
+///   %upper_slice_y = arith.addi %slice_idx, %y : index
+///   %upper_slice_mask = vector.extract %mask[%slice_idx]
+///     : vector<[8]xi1> from vector<[16]x[8]xi1>
+///   %upper_slice = vector.extract %upper_tile[%slice_idx]
+///     : vector<[8]xi16> from vector<[8]x[8]xi16>
+///   vector.transfer_write %upper_slice,
+///     %dest[%upper_slice_y, %x], %upper_slice_mask
+///     : vector<[8]xi16>, memref<?x?xi16>
+///   // Same again for the lower tile:
+///   %lower_slice_idx = arith.addi %c8_vscale, %slice_idx : index
+///   %lower_slice_y = arith.addi %lower_slice_idx, %y : index
+///   %lower_slice_mask = vector.extract %mask[%lower_slice_idx]
+///     : vector<[8]xi1> from vector<[16]x[8]xi1>
+///   %lower_slice = vector.extract %lower_tile[%slice_idx]
+///     : vector<[8]xi16> from vector<[8]x[8]xi16>
+///   vector.transfer_write %lower_slice,
+///     %dest[%lower_slice_y, %x], %lower_slice_mask
+///     : vector<[8]xi16>, memref<?x?xi16>
+/// }
----------------
c-rhodes wrote:

this is dense, I think a simpler example like the unmasked `transfer_write_f16_scalable_16x8` would be fine.

https://github.com/llvm/llvm-project/pull/96187