[PATCH] Enable SLP-vectorization of intrinsics.

[Chandler Carruth]

On Wed, Jan 22, 2014 at 2:13 PM, Hal Finkel <hfinkel at anl.gov> wrote:

> > Well, either removing that or implementing a mechanism to fully model
> > this would resolve the issue.
> > If this is known to be widely broken the former would be the easiest
> > to do, of course. Who should we consult on this, llvm-dev? Should I
> > prepare a patch?
>
> You can try asking on llvmdev. If that does not trigger any useful
> discussion, then just submit a patch.
>
> >
> >
> > This impacts more than just vectorization, I can construct simple
> > testcases where we fail to common scalar sqrt calls.
>
> Yep.


FWIW, I thin kthis would be *really* good to do.

We're right now in the worst of both worlds: we neither support the feature
nor do we take advantage of the fact that we don't support the feature.

If you post to llvmdev, I'll argue strongly for supporting two worlds:

1) math routines can write to errno, but have no interaction with the FENV
business and we don't support modelling that
2) math routines cannot even write to errno

These can be toggled by the existing -f(no-)math-errno flags in Clang that
I think are actually propagated.
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/llvm-commits/attachments/20140124/c8bd097a/attachment.html>