[PATCH] [RFC] loop index promotion pass

Mark Mitchell mark@codesourcery.com
Wed May 20 16:08:00 GMT 2009


stevenb.gcc@gmail.com wrote:

> This is just not true. Once something is in GCC, ripping it out again is
> virtually impossible.

I think this is a general argument, being applied inappropriately.  The
nice thing about this whole TREE-SSA framework is that these passes are
pretty independent; ripping this out isn't going to be that involved.
You'll just remove this file.

> Why can't this pass just be implemented as part of ivcanon? So what if
> it takes extra work. Tough luck. It is the better long-term solution.

If there's a better way to do it, and someone wants to do that, that's
obviously good.  I'd argue, however, that an existing patch, if it is
(a) already done, (b) correct, (c) doesn't impact compile-time
performance, and (d) improves performance is in fact worth
incorporating, if a better option is not actually likely to happen any
time soon.

That may reflect bad taste on my part, but user's care about how the
compiler behaves, not what it looks like.

I realize I'm probably swimming upstream here, though.  If y'all can
trick each other into figuring out a better way to implement this, and
especially into tricking someone into proving that way is better by
implementing it, even better!

-- 
Mark Mitchell
CodeSourcery
mark@codesourcery.com
(650) 331-3385 x713



More information about the Gcc-patches mailing list