[PATCH] D28508: [NVPTX] Lower to sqrt.approx and rsqrt.approx under more circumstances.

[Hal Finkel via Phabricator via llvm-commits]

hfinkel added a comment.

In https://reviews.llvm.org/D28508#643613, @jlebar wrote:

>


...

> But specifically for sqrt (AFAIK) I need to rely on some behavior that's not quite kosher according to the langref.  Specifically, I need to transform llvm.nvvm.sqrt.f into
> 
>   s = llvm.sqrt(arg)
>   select(arg >= -0.0, s, NaN)
>    
> 
> The langref says that llvm.sqrt is undefined behavior if you call it with a negative value, but @arsenm, @mehdi_amini, and I think it should say that it *returns undef* if passed a negative value.  arsenm looked at the blame and the language has been unchanged since llvm.sqrt was added.  His and @mehdi_amini's GPU backends already assume the "returns undef" semantics, and I have code in flight for XLA to do the same.  One of them also did a quick audit and concluded none of our current optimizations use the "undefined behavior" behavior.
> 
> Mehdi pointed me at a larger patch to fix up this and other math intrinsics which I now can't find (sorry), but maybe we can make this smaller change in parallel?
> 
> WDYT?

Sounds reasonable to me.


https://reviews.llvm.org/D28508