Re: [tcb] Merge PHI nodes -- 7.4% reduction of PHI nodes (Take 2)

[Jeffrey A Law <law at redhat dot com>]

On Mon, 2004-09-20 at 17:10, Kazu Hirata wrote:
> Hi Jeff,
> 
> > BB3:
> >   x1 = PHI (0 (BB0), 1 (BB1))
> > 
> > BB4:
> >   x2 = PHI (x1 (BB3), x0 (BB2))
> >   if (x2 == 0)
> >     goto BB5;
> >   else
> >     goto BB6;
> 
> Thanks for detailed explanation.  By the way, I'd like to point out
> that I often see a variant of your example like so:
> 
> BB3:
>   x1 = PHI (0 (BB0), 1 (BB1))
>   x2 = (_Bool) x1;              <- Notice this cast!
> 
> BB4:
>   x2 = PHI (x1 (BB3), x0 (BB2))
>   if (x2)
>     goto BB5;
>   else
>     goto BB6;
> 
> Notice that my PHI merge pass cannot get at this king as the two PHI
> nodes are separated by a cast to _Bool, meaning that we miss a jump
> threading opportunities.
Yes.  That's the kind of thing I'd like to fix with a tree combiner
pass.   It's not a 100% perfect fit, but it's pretty close.  That's
the primary reason why I haven't dove into your pass to kill those
casts.

jeff