This is the mail archive of the gcc-patches@gcc.gnu.org mailing list for the GCC project.


Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]
Other format: [Raw text]

Re: Possible fix for 30% performance regression in PR 33928


On Wed, May 6, 2009 at 10:09 PM, Bradley Lucier <lucier@math.purdue.edu> wrote:
>
> On May 6, 2009, at 1:59 AM, Steven Bosscher wrote:
>
>> And in any case, just adding back those functions doesn't explain why
>> your code gets better. Blindly adding stuff back without understanding
>> why that helps
>
> I don't hesitate to admit a deep lack of understanding about the internals
> of gcc. ?On the other hand, I do understand roughly the goals of
> "canonicalize address computations" and "choose the best addressing mode
> based on some cost", and that was why after some thought I added back
> precisely these two functions.
>
> So, are there other places in the compiler that "canonicalize address
> computations" and/or "choose the best addressing mode based on some cost"?
>
>> isn't going to get us anywhere closer to simplifying
>> CSE.
>
>
>> There was a reason to remove them: fwprop should do the same
>> transformations already.
>
>
> Good, there's a clear statement: ?Are you saying that fwprop should perform
> what canon_for_address and find_best_addr were doing in the old CSE?

It should basically the same perform the same transformation, yes.

I would like to understand what we're missing. The problem with your
PR has always been (and still is) that the test case is just huge, so
it is very difficult to debug. Maybe, with this patch of yours, we can
investigate this more easily (I plan to do that).

Ciao!
Steven


Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]