[Mlir-commits] [mlir] [mlir][ArmSME] More precisely model dataflow in ArmSME to SCF lowerings (PR #73922)

[Benjamin Maxwell]

================
@@ -202,32 +209,36 @@ struct TileLoadOpWithMaskAndPadZeroConversion
     // Initialize tile with zero to satisfy padding. Inactive cols will be
     // zeroed anyway since the loads use zeroing predication. For inactive rows
     // however, no load will occur so these need to be zeroed.
-    auto tile = tileLoadOp.createOpAndForwardTileId<arm_sme::ZeroOp>(
+    auto initTile = tileLoadOp.createOpAndForwardTileId<arm_sme::ZeroOp>(
         rewriter, loc, tileType);
 
     // Create a loop to load the active tile slices from memory.
     auto step = rewriter.create<arith::ConstantIndexOp>(loc, 1);
     auto lowerBound = rewriter.create<arith::ConstantIndexOp>(loc, 0);
     auto upperBound = numRows;
-    auto forOp = rewriter.create<scf::ForOp>(loc, lowerBound, upperBound, step);
+    auto forOp = rewriter.create<scf::ForOp>(loc, lowerBound, upperBound, step,
+                                             ValueRange{initTile});
 
     rewriter.setInsertionPointToStart(forOp.getBody());
 
     // Create 'arm_sme.load_tile_slice' to load tile slice from memory into
     // tile.
     SmallVector<Value> memrefIndices;
     auto tileSliceIndex = forOp.getInductionVar();
+    auto currentTile = forOp.getRegionIterArg(0);
----------------
MacDue wrote:

Naming is tricky, but it's the tile at the start of the current iteration of the loop, which is taken from the `iter_args`. The `updatedTile` would be the tile you get from the operation in the loop (i.e. the result of `move_vector_to_tile_slice`), and `outputTile` the result of the `scf.for` (at least that's how I see it :))

https://github.com/llvm/llvm-project/pull/73922