Hi Kurchi, I think we've converged on the code changes. Please prepare and post
another webrev for a final cross-check before pushing.
What follows is I think merely residual disagreement over the philosophy of how
to handle generic casts vs reification. :-)
On 4/6/12 3:06 AM, Rémi Forax wrote:
On 04/05/2012 11:04 PM, Stuart Marks wrote:
I'm somewhat skeptical of making code changes now based on potential future
benefits when/if generics become reified. This was discussed before; see
http://mail.openjdk.java.net/pipermail/jdk8-dev/2011-December/000454.html
In that message, John Rose said "If the best practices have to change, then
we'll have to change that code again. Or maybe the retrofit strategy will
have to take account of the existing code idioms. In any case, we'll cross
that bridge when we get to it. (Coping with reification in this case is a
decision to make tomorrow, not today.)"
I disagree with John. The main issue with generics nowadays is that
most of the people doesn't care about a cast to a type variable because
everybody knows about erasure. So codes are written with an implementation
glitch in mind.
Frankly, I don't know if reification will appear (yes it's a kind of magical)
or not
but I think it's a sloppy path to not consider all casts as equals.
In order to program effectively with generics, I think you have to understand
erasure and its implications. It may have been an unfortunate choice, but
erasure is part of the language and we have to deal with it and in some cases
rely on it. I don't think it's merely an "implementation glitch."
The difficulty I have with reification is that while there are proposals
floating around for how it could be done, nobody really knows how it will
eventually turn out, nor whether it will actually be done. If it is eventually
done, there will legal and illegal constructs, constructs that generate
warnings, and perhaps a style guide for how to use reified generics properly.
Right now, we can *imagine* what these future rules might be, but it seems
untenable to me to try to make today's code conform to those imaginary future
rules, especially in the absence of tools to help support those rules.
If unmaskNull return a V, the code of equals will upcast the value from Object
to V
to just after downcast it from V to Object,
I think it's better that unmask to return Object and upcast it to V when it's
necessary.
Certainly there are cases where there's a redundant downcast and upcast. In a
reified world, will this be a significant expense? Really, I have no idea.
s'marks
