[mlir][linalg] PadTensorOp vectorization: Avoid redundant FillOps

[springerm]

Summary

Do not generate FillOps when these would be entirely overwritten.

[springerm]

rebase

mlir/lib/Dialect/Linalg/Transforms/Vectorization.cpp

static LogicalResult tryVectorizeCopy(PatternRewriter &rewriter,
 
     // Generate TransferReadOp.
     SmallVector<Value> readIndices(
         vecType.getRank(), rewriter.create<ConstantIndexOp>(padOp.getLoc(), 0));
     auto read = rewriter.create<vector::TransferReadOp>(
         padOp.getLoc(), vecType, padOp.source(), readIndices, padValue,
         readInBounds);
 
+    // If `dest` is a FillOp and the TransferWriteOp would overwrite the entire
+    // tensor, write directly to the FillOp's operand.
+    if (llvm::equal(vecShape, resultType.getShape()) &&
+        llvm::all_of(writeInBounds, [](bool b) { return b; }))
+      if (auto fill = dest.getDefiningOp<FillOp>())
+        dest = fill.output();
+
     // Generate TransferWriteOp.
     auto writeIndices = ofrToIndexValues(
         rewriter, padOp.getLoc(), padOp.getMixedLowPad());
     rewriter.replaceOpWithNewOp<vector::TransferWriteOp>(
         padOp, read, dest, writeIndices, writeInBounds);
 
     return success();
   }