Re: PR26435 (-ftree-loop-linear)

[Daniel Berlin <dberlin at dberlin dot org>]

On Apr 18, 2006, at 5:00 PM, Mark Mitchell wrote:

> As discussed on the gcc list, -ftree-loop-linear consists of several
> chunks of code.  One chunk (to do with perfect loop nests) is  
> apparently
> not 100% algorithmically sound, but has the decided benefit of making
> our SPEC scores substantially better.
>
> My question is: what should we do about this, for 4.1 and for 4.2?
>
> For 4.1, I would suggest separating the problematic chunk of code  
> into a
> separate option, and document the option as experimental, hopefully  
> with
> a note that suggests under what situations the optimization is unsafe.
> That would clarify the situation for users, while not taking away
> performance?

I'm happy to prepare such a patch if that is what we want.

> For 4.2, can we correct the code so that it doesn't do the wrong  
> thing?
Sebastian believes he can.

>  It sounds like we need some additional safety checks to avoid invalid
> transformations; can we add those?  And will that impact SPEC scores?
I will leave these questions to Sebastian, because bersonally, I do  
not believe i will have time to look into it until at least a few  
weeks from now, realistically.