[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());
----------------
hanhanW wrote:

naming nit: the other similar function is using `oneElemVecType`. I think we should use the same name to reduce variants. I'm +1 on `oneElemVecType`.

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