ConvL2INode::Ideal has an optimization to convert convL2I(addL(x,y)) ==> addI(convL2I(x),convL2I(y)). Unfortunately it doesn't check to see if there are other uses of the ConvL2I which would cause it to be computed anyway. The test case from 4850191 has this problem. You end up computing the add twice which is clearly not useful. The fix is to check if there is more than one use of the add.
###@###.### 2005-06-15 01:11:35 GMT