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

[Andrzej Warzyński]

================
@@ -309,6 +314,90 @@ 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.
+///
+/// Before:
+///   memory = |ab|cd|ef|12| : <4xi2> (<1xi8>)
+///   valueToStore = |01|23|45|67| : vector<4xi2>
+///   mask = |0|0|1|1| : vector<4xi1>
+/// After:
+///   memory = |ab|cd|45|67| : <4xi2> (<1xi8>)
+static void atomicStore(OpBuilder &builder, Location loc,
+                        MemRefValue linearizedMemref, Value linearizedIndex,
+                        VectorValue valueToStore, Value mask,
+                        int64_t numSrcElemsPerDest) {
+  auto atomicOp = builder.create<memref::GenericAtomicRMWOp>(
+      loc, linearizedMemref, ValueRange{linearizedIndex});
+  Value origValue = atomicOp.getCurrentValue();
+
+  OpBuilder::InsertionGuard guard(builder);
+  builder.setInsertionPointToStart(atomicOp.getBody());
+
+  // i8 -> <1xi8> -> <numSrcElemsPerDest x i.>:
+  auto oneVectorType = VectorType::get({1}, origValue.getType());
+  auto fromElem = builder.create<vector::FromElementsOp>(loc, oneVectorType,
+                                                         ValueRange{origValue});
+  auto vectorBitCast =
+      builder.create<vector::BitCastOp>(loc, valueToStore.getType(), fromElem);
+
+  auto select =
+      builder.create<arith::SelectOp>(loc, mask, valueToStore, vectorBitCast);
+  auto bitcast2 = builder.create<vector::BitCastOp>(loc, oneVectorType, select);
+  auto extract = builder.create<vector::ExtractOp>(loc, bitcast2, 0);
+  builder.create<memref::AtomicYieldOp>(loc, extract.getResult());
+}
+
+/// 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 &rewriter, Location loc,
+                     MemRefValue linearizedMemref, Value linearizedIndex,
+                     VectorValue value, Value mask,
+                     int64_t numSrcElemsPerDest) {
+  auto emulatedIOType =
+      VectorType::get({1}, linearizedMemref.getType().getElementType());
+  auto elemLoad = rewriter.create<vector::LoadOp>(
+      loc, emulatedIOType, linearizedMemref, ValueRange{linearizedIndex});
+  auto fromBitcast = rewriter.create<vector::BitCastOp>(
+      loc,
+      VectorType::get({numSrcElemsPerDest}, value.getType().getElementType()),
+      elemLoad);
+  auto select = rewriter.create<arith::SelectOp>(loc, mask, fromBitcast, value);
+  auto toBitcast =
+      rewriter.create<vector::BitCastOp>(loc, emulatedIOType, select);
+  rewriter.create<vector::StoreOp>(loc, toBitcast, linearizedMemref,
+                                   linearizedIndex);
+}
+
+/// Extract a slice from vector `vector`, with the size of `sliceNumElements`,
+/// and insert it into a zero, byte vector at offset `byteOffset`. For example:
----------------
banach-space wrote:

```suggestion
/// Extract `sliceNumElements` from `vector`,
/// and insert it into an empty vector at offset `byteOffset`.
```
? What's the difference between `sliceOffset` and `byteOffset`?

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