[Mlir-commits] [mlir] [mlir][ArmSME] Add rudimentary support for tile spills to the stack (PR #76086)

[Cullen Rhodes]

================
@@ -40,6 +132,209 @@ IntegerAttr getTileIdOrError(arm_sme::ArmSMETileOpInterface op) {
   return tileId;
 }
 
+/// Creates a alloca matching the size of tile used by `tileOp`. The alloca is
+/// placed in the first block of the function.
+static memref::AllocaOp
+createAllocaForTile(RewriterBase &rewriter, Location loc,
+                    FunctionOpInterface func,
+                    arm_sme::ArmSMETileOpInterface tileOp) {
+  RewriterBase::InsertionGuard g(rewriter);
+  // Move to the first operation in the function.
+  rewriter.setInsertionPoint(&func.getBlocks().front().front());
+  // Create an alloca matching the tile size of the `tileOp`.
+  auto vscale = rewriter.create<vector::VectorScaleOp>(loc);
+  auto tileElementType =
+      llvm::cast<VectorType>(tileOp.getTileType()).getElementType();
+  auto memrefType = MemRefType::get(
+      {ShapedType::kDynamic, ShapedType::kDynamic}, tileElementType);
+  auto minElements = arm_sme::getSMETileSliceMinNumElts(tileElementType);
+  auto minElementsOp =
+      rewriter.create<arith::ConstantIndexOp>(loc, minElements);
+  auto vectorLen = rewriter.create<arith::MulIOp>(loc, vscale, minElementsOp);
+  auto alloca = rewriter.create<memref::AllocaOp>(
+      loc, memrefType, ValueRange{vectorLen, vectorLen});
+  return alloca;
+}
+
+/// Finds or creates an alloca for a spill of a tile.
+static memref::AllocaOp
+getOrCreateTileMemory(RewriterBase &rewriter, Location loc,
+                      FunctionOpInterface func,
+                      arm_sme::ArmSMETileOpInterface tileOp, unsigned tileId) {
+  // Find an alloca at the top of the function tagged with a
+  // 'arm_sme.in_memory_tile_id' that matches `tileId`.
+  for (auto &op : func.getBlocks().front()) {
+    auto alloca = llvm::dyn_cast<memref::AllocaOp>(op);
+    if (!alloca)
+      continue;
+    auto inMemoryTileId = llvm::dyn_cast_or_null<IntegerAttr>(
+        alloca->getDiscardableAttr(kInMemoryTileId));
+    if (!inMemoryTileId)
+      continue;
+    if (inMemoryTileId.getInt() == tileId)
+      return alloca;
+  }
+  // Otherwise, create a new alloca:
+  auto alloca = createAllocaForTile(rewriter, loc, func, tileOp);
+  alloca->setDiscardableAttr(kInMemoryTileId,
+                             rewriter.getI32IntegerAttr(tileId));
+  return alloca;
+}
+
+/// Very naive lowering of in-memory tiles (i.e. tiles that were not assigned a
+/// hardware tile ID) to ArmSME intrinsics. Currently, this works by assigning
+/// the op to tile 0, then emitting a full tile swap between ZA and memory
+/// before + after the tile op.
+///
+/// Example:
+///
+///    arm_sme.tile_op { tile_id = <IN MEMORY TILE> }
+///
+/// is converted to:
+///     // At function entry:
+///     %alloca = memref.alloca ... : memref<?x?xty>
+///
+///     // Around op:
+///     scf.for %slice_idx {
+///       %current_slice = "arm_sme.intr.read.horiz" ... <{tile_id = 0 : i32}>
+///       "arm_sme.intr.ld1h.horiz"(%alloca, %slice_idx)  <{tile_id = 0 : i32}>
+///       vector.store %current_slice, %alloca[%slice_idx, %c0]
+///     }
+///     arm_sme.tile_op { tile_id = 0 }
+///     scf.for %slice_idx {
+///       %current_slice = "arm_sme.intr.read.horiz" ... <{tile_id = 0 : i32}>
+///       "arm_sme.intr.ld1h.horiz"(%alloca, %slice_idx)  <{tile_id = 0 : i32}>
+///       vector.store %current_slice, %alloca[%slice_idx, %c0]
+///     }
+///
+/// Note that these spills/fills are not inserted earlier as concept of a
+/// register, and the need to swap the contents, can't really be represented
+/// correctly at a high level in MLIR.
+///
+/// TODO: Reduce the spills/reloads to single slices where possible.
+struct ConvertArmSMESpillsAndFillsToLLVM : public ConvertToLLVMPattern {
+
+  ConvertArmSMESpillsAndFillsToLLVM(StringRef rootOpName,
+                                    const LLVMTypeConverter &typeConverter,
+                                    PatternBenefit benefit)
+      : ConvertToLLVMPattern(rootOpName, &typeConverter.getContext(),
+                             typeConverter, benefit) {}
+
+  LogicalResult
+  matchAndRewrite(Operation *op, ArrayRef<Value> operands,
+                  ConversionPatternRewriter &rewriter) const override {
+    auto tileOp = cast<arm_sme::ArmSMETileOpInterface>(op);
+    // Tile has a real (hardware) tile. No spills/reloads required.
+    if (!tileOp.isInMemoryTile())
+      return failure();
+
+    // Step 1. Create an alloca for the tile at the top of the function (if one
+    // does not already exist).
+    auto loc = tileOp.getLoc();
+    auto func = tileOp->getParentOfType<FunctionOpInterface>();
+    auto tileAlloca = getOrCreateTileMemory(rewriter, loc, func, tileOp,
+                                            tileOp.getTileId().getInt());
+
+    // Step 2. Assign the op a real tile ID.
+    // For simplicity, we always use tile 0.
+    auto zeroTileId = rewriter.getI32IntegerAttr(0);
+    {
+      rewriter.startRootUpdate(tileOp);
+      tileOp.setTileId(zeroTileId);
+      rewriter.finalizeRootUpdate(tileOp);
+    }
+
+    VectorType tileVectorType = tileOp.getTileType();
+    auto sliceType = VectorType::Builder(tileOp.getTileType()).dropDim(0);
----------------
c-rhodes wrote:

```suggestion
    auto sliceType = VectorType::Builder(tileVectorType).dropDim(0);
```

https://github.com/llvm/llvm-project/pull/76086