[LLVMdev] [RFC] How to fix sqrt vs llvm.sqrt optimization asymmetry

[Hal Finkel]

----- Original Message -----
> 
> On Nov 11, 2013, at 9:30 PM, Hal Finkel <hfinkel at anl.gov> wrote:
> 
> > ----- Original Message -----
> >> Hi Hal, all.
> >> 
> >> I'm not sure why llvm.sqrt is 'special'.  Maybe because there is a
> >> SSE
> >> packed sqrt instruction (SQRTPS) but not e.g. a packed sin
> >> instruction
> >> AFAIK.
> > 
> > This seems relevant:
> > http://lists.cs.uiuc.edu/pipermail/llvmdev/2007-August/010248.html
> > 
> > Chris, et al., does the decision on how to treat sqrt predate our
> > current way of handling errno?
> 
> The intention of llvm.sqrt is that it is the same as sqrt(), but
> doesn't affect errno.  If it helps, I think it would be completely
> reasonable to change this in langref.html:
> 
> "Unlike sqrt in libm, however, llvm.sqrt has undefined behavior for
> ..."
> 
> to "... produces an undefined value", with a link back to
> ##undefined-values.

I'm not sure that helps, because it will prevents sqrt + -fno-math-errno (a readnone sqrt) -> llvm.sqrt -- and thus still prevents the use of llvm.sqrt to vectorize sqrt. I think what will help is just saying something like this:

  "Unlike sqrt in libm, after calling llvm.sqrt with a negative argument (except for -0.0), the state of errno, and any other portions of the floating-point environment that are used to record errors, is undefined."

FWIW, we may want to make a similar change for the other llvm.<libm function> intrinsics, at least for the vector versions (as vector ops on many ISAs don't affect the FP state bits like the scalar ops do).

> 
> Transformations that apply to sqrt() should generally apply to
> llvm.sqrt as well.

And the intent is for this to be true for all of the other llvm.<libm function> intrinsics as well, right?

Thanks again,
Hal

> 
> -Chris
> 

-- 
Hal Finkel
Assistant Computational Scientist
Leadership Computing Facility
Argonne National Laboratory