[Mlir-commits] [mlir] [mlir][amdgpu] Sink op creation in scaled conversion intrinsics (NFC) (PR #168542)

[Jakub Kuderski]

================
@@ -1767,15 +1763,21 @@ LogicalResult ScaledExtPacked816OpLowering::matchAndRewrite(
     return rewriter.notifyMatchFailure(op, "type conversion failed");
   }
 
-  Value castedSource =
-      LLVM::BitcastOp::create(rewriter, loc, packedType, source);
-
   std::optional<StringRef> maybeIntrinsic =
       scaledExtPacked816ToIntrinsic(srcElemType, destElemType);
   if (!maybeIntrinsic.has_value())
     return op.emitOpError(
         "no intrinsic matching packed scaled conversion on the given chipset");
 
+  int32_t scaleSel =
+      getScaleSel(blockSize, bitWidth, firstScaleLane, firstScaleByte);
+
+  Value castedScale =
+      LLVM::BitcastOp::create(rewriter, loc, i32, adaptor.getScale());
+
+  Value castedSource =
+      LLVM::BitcastOp::create(rewriter, loc, packedType, source);
----------------
kuhar wrote:

```suggestion
  int32_t scaleSel =
      getScaleSel(blockSize, bitWidth, firstScaleLane, firstScaleByte);
  Value castedScale =
      LLVM::BitcastOp::create(rewriter, loc, i32, adaptor.getScale());
  Value castedSource =
      LLVM::BitcastOp::create(rewriter, loc, packedType, source);
```

https://github.com/llvm/llvm-project/pull/168542