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

[Andrzej Warzyński]

================
@@ -373,6 +374,130 @@ 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
+/// static analysis).
+///
+/// Example (in pseudo-MLIR):
+///
+/// ```
+/// vector.transfer_write vector, dest[x, y], mask
+///   : vector<[16]x[4]xf32>, memref<?x?xf32>
+/// ```
+/// Is rewritten to:
+/// ```
+/// for i in range (0, 4 * vscale) {
+///   let sliceRow = i + tile_n.row * vscale;              ─┐
+///   let sliceCol = tile_n.col * vscale;                   |
+///   slice = vector.extract tile_n[i]                      |
+///     : vector<[4]xf32> from vector<[16]x[4]xf32>         |
+///   slice_mask = vector.extract mask[sliceRow]            |- Repeated 4x for
+///     : vector<[4]xi1> from vector<[16]x[4]xi1>           |  all tiles in
+///   vector.transfer_write                                 |  [16]x[4]
+///     slice, dest[x + sliceRow, y + sliceCol], slice_mask |
+///     : vector<[4]xf32>, memref<?x?xf32>                  ┘
+/// }
+/// ```
+struct LegalizeMultiTileTransferWriteAsStoreLoop
+    : public OneToNOpConversionPattern<vector::TransferWriteOp> {
+  using OneToNOpConversionPattern::OneToNOpConversionPattern;
+
+  LogicalResult
+  matchAndRewrite(vector::TransferWriteOp writeOp, OpAdaptor adaptor,
+                  OneToNPatternRewriter &rewriter) const override {
+    if (writeOp.hasPureTensorSemantics())
+      return rewriter.notifyMatchFailure(
+          writeOp, "TODO: tensor semantics are unsupported");
----------------
banach-space wrote:

Please add a negative test

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