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

[Han-Chung Wang]

================
@@ -350,32 +423,154 @@ struct ConvertVectorStore final : OpConversionPattern<vector::StoreOp> {
     // vector.store %bitcast, %alloc[%linear_index] : memref<16xi8>,
     // vector<4xi8>
 
-    auto origElements = op.getValueToStore().getType().getNumElements();
-    if (origElements % scale != 0)
-      return failure();
+    auto origElements = valueToStore.getType().getNumElements();
+    bool isUnalignedEmulation = origElements % numSrcElemsPerDest != 0;
 
     auto stridedMetadata =
         rewriter.create<memref::ExtractStridedMetadataOp>(loc, op.getBase());
 
     OpFoldResult linearizedIndices;
-    std::tie(std::ignore, linearizedIndices) =
+    memref::LinearizedMemRefInfo linearizedInfo;
+    std::tie(linearizedInfo, linearizedIndices) =
         memref::getLinearizedMemRefOffsetAndSize(
             rewriter, loc, srcBits, dstBits,
             stridedMetadata.getConstifiedMixedOffset(),
             stridedMetadata.getConstifiedMixedSizes(),
             stridedMetadata.getConstifiedMixedStrides(),
             getAsOpFoldResult(adaptor.getIndices()));
 
-    auto numElements = origElements / scale;
-    auto bitCast = rewriter.create<vector::BitCastOp>(
-        loc, VectorType::get(numElements, newElementType),
-        op.getValueToStore());
+    auto foldedNumFrontPadElems =
+        isUnalignedEmulation
+            ? getConstantIntValue(linearizedInfo.intraDataOffset)
+            : 0;
+
+    if (!foldedNumFrontPadElems) {
+      // Unimplemented case for dynamic front padding size != 0
+      return failure();
+    }
+
+    TypedValue<MemRefType> emulatedMemref =
+        cast<TypedValue<MemRefType>>(adaptor.getBase());
+
+    // Shortcut: conditions when subbyte store at the front is not needed:
+    // 1. The source vector size is multiple of byte size
+    // 2. The address of the store is aligned to the emulated width boundary
+    if (!isUnalignedEmulation && *foldedNumFrontPadElems == 0) {
+      auto numElements = origElements / numSrcElemsPerDest;
+      auto bitCast = rewriter.create<vector::BitCastOp>(
+          loc, VectorType::get(numElements, newElementType),
+          op.getValueToStore());
+      rewriter.replaceOpWithNewOp<vector::StoreOp>(
+          op, bitCast.getResult(), emulatedMemref,
+          getValueOrCreateConstantIndexOp(rewriter, loc, linearizedIndices));
+      return llvm::success();
----------------
hanhanW wrote:

style nit: drop `llvm::`.

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