[Mlir-commits] [mlir] [MLIR] Implement emulation of static indexing subbyte type vector stores (PR #115922)

[Han-Chung Wang]

================
@@ -292,6 +297,104 @@ emulatedVectorLoad(OpBuilder &rewriter, Location loc, Value base,
       newLoad);
 }
 
+/// Emits `memref.generic_atomic_rmw` op to store a subbyte-sized value to a
+/// byte in memory, with a mask. The `valueToStore` is a vector of subbyte-sized
+/// elements, with size of 8 bits, and the mask is used to select which elements
+/// to store.
+///
+/// Inputs:
+///   linearizedMemref = |2|2|2|2| : <4xi2> (<1xi8>)
+///   linearizedIndex = 2
+///   valueToStore = |3|3|3|3| : vector<4xi2>
+///   mask = |0|0|1|1| : vector<4xi1>
+///
+/// Result:
+///   linearizedMemref = |2|2|3|3| : <4xi2> (<1xi8>)
+static void atomicStore(OpBuilder &builder, Location loc,
+                        MemRefValue linearizedMemref, Value linearizedIndex,
+                        VectorValue valueToStore, Value mask) {
+  assert(valueToStore.getType().getRank() == 1 && "expected 1-D vector");
+
+  // Create an atomic load-modify-write region using
+  // `memref.generic_atomic_rmw`.
+  auto atomicOp = builder.create<memref::GenericAtomicRMWOp>(
+      loc, linearizedMemref, ValueRange{linearizedIndex});
+  Value origValue = atomicOp.getCurrentValue();
+
+  OpBuilder::InsertionGuard guard(builder);
+  builder.setInsertionPointToStart(atomicOp.getBody());
+
+  // Construct the vector value from the scalar.
+  auto oneVectorType = VectorType::get({1}, origValue.getType());
+  Value origVecValue = builder.create<vector::FromElementsOp>(
+      loc, oneVectorType, ValueRange{origValue});
+  origVecValue = builder.create<vector::BitCastOp>(loc, valueToStore.getType(),
+                                                   origVecValue);
+
+  // Construct the final masked value and yield it.
+  Value maskedValue =
+      builder.create<arith::SelectOp>(loc, mask, valueToStore, origVecValue);
+  maskedValue =
+      builder.create<vector::BitCastOp>(loc, oneVectorType, maskedValue);
+  auto scalarMaskedValue =
+      builder.create<vector::ExtractOp>(loc, maskedValue, 0);
+  builder.create<memref::AtomicYieldOp>(loc, scalarMaskedValue);
+}
+
+/// Generate a non-atomic read-modify-write sequence for subbyte storing.
+/// It has similar logic to `atomicStore`, but without the atomicity.
+static void rmwStore(OpBuilder &builder, Location loc,
+                     MemRefValue linearizedMemref, Value linearizedIndex,
+                     VectorValue valueToStore, Value mask) {
+  assert(valueToStore.getType().getRank() == 1 && "expected 1-D vector");
+
+  // Construct the vector value from the scalar.
----------------
hanhanW wrote:

This and the below comment does not match the code. It is loading a value from memref, but not construct the value from a scalar.

https://github.com/llvm/llvm-project/pull/115922