[PATCH] Infer known bits from dominating conditions

[Philip Reames]

On 03/04/2015 08:29 AM, hfinkel at anl.gov wrote:
> In http://reviews.llvm.org/D7708#134110, @reames wrote:
>
>> Hal - preliminary numbers on the dominance based approach make it look like that might be viable.  I need to cleanup my hacked implementation and fix at least one bug, but my initial impression is that it appears to be fast enough.  I'm a bit surprised by this, but it's what the data says.
>>
>> Worth discussing is whether we should actually incorporate both.  They seem to be good at different cases and maybe we should exploit that.  The use based approach is good for grabbing facts far from the use site (i.e. function entry guards), while the dominance approach is likely more powerful since it's not restricted to direct uses, but is probably best limited to a local scope.  Thoughts?
>
> Yes, perhaps with a FIXME if there is some other "right" way to fix the problem.
Ok, I'll try to get a merged patch up for review tomorrow.  This is 
coming out of my spare time right now, so it might take a day or two.

> I've been thinking of something like LVI that keeps track of known bits. What do you think?
I've had thoughts in similar direction, but I think it would really 
needs to be merged with ValueTracking.  We'd want instcombine to exploit 
it and instcombine uses ValueTracking.  Teaching instcombine about two 
analyzes that don't cross feed seems less than ideal.

I've mentioned the idea of turning VT into an analysis pass to Chandler 
previously.  He had serious reservations about that.  If I remember 
correctly, he was much less concerned about introducing an internal 
cache inside VT.  He was mostly concerned about invalidation bugs.  He 
believes that much of instcombine assumes VT works from first principals 
and that changing that would be *hard*.

I also see this as being orthogonal to the original patch.  That was 
about catching any dominating conditions at all.  This would be about 
making VT better as a whole, not just specific to dominating conditions.

p.s. The other orthogonal work here is propagating conditional constants 
early in the optimizer.  (possibly instcombine?)  If you have thoughts 
here, let's me know.

>
>
> http://reviews.llvm.org/D7708
>
> EMAIL PREFERENCES
>    http://reviews.llvm.org/settings/panel/emailpreferences/
>
>