Re: scheduler speedup

[Jan Hubicka <jh at suse dot cz>]

> Err, doesn't this slow down init_regions when
> defined (ENABLE_CHECKING) && ! defined (HAVE_conditional_execution)?
What I do avoid is the multiple calls to updating liveness information
(I do rebuild update information at once, before checksuming).
It does not solve the problems of removing death notes, but it cuts the
time of life_analyzis, so overall time is better even with enable-checking.

I also wanted to keep the enable-checking/!enable-checking code as close as
possible, so it does exactly same thing modulo the verification code.  The
second "whole chain pass" is very cheap compared to the first "all regions"
pass, so I don't think we need to worry about it.

We do many passes over all insns anyway and this one is particulary cheap.

Honza
> 
> It seems to me in that case you do two passes over all the insns.  One
> to count the notes, another to zap them.  That seems rather dumb to me.
> 
> I think the code should look something like
> 
> if (CHECK_DEAD_NOTES)
>   count & remove death notes
> else
>   just remove death notes
> 
> 
> Or am I missing something here?
> 
> jeff
> 
>