Re: FE Exception triggered by comparison

[Szabolcs Nagy <nsz@port70.net>]

* Rich Felker <dalias@libc.org> [2019-02-27 12:26:41 -0500]:
> On Wed, Feb 27, 2019 at 08:14:07PM +0300, Alexander Monakov wrote:
> > On Wed, 27 Feb 2019, Rich Felker wrote:
> > 
> > > Are there reasons we should perhaps use the __builtin versions of
> > > these when __GNUC__ indicates they're available? I like our bit test
> > > versions we have now, and I think they're sufficiently efficient, but
> > > I'm open to changes if there's a good reason.
> > 
> > Well, it really depends on what one considers 'sufficiently efficient'.
> > Instead of comparing a register with itself and testing flags (2 instructions)
> > you get (for 'int f(double x){return isnan(x);}'):
> > 
> > f:
> >         movabsq $9223372036854775807, %rdx
> >         movq    %xmm0, %rax
> >         andq    %rdx, %rax
> >         movabsq $9218868437227405312, %rdx
> >         cmpq    %rdx, %rax
> >         seta    %al
> >         movzbl  %al, %eax
> >         ret
> > 
> > (note that movq %xmm0, %rax is going to be more costly than a normal
> > move as it crosses from fp to integer domain in the cpu)
> > 
> > I think musl bit test can be implemented more efficiently via right-shifting
> > the representation in %rax first, avoiding 64-bit immediates,
> 
> Or left-shifting rather than masking to get rid of the sign bit?
> That's all it's doing. I don't think right-shift is okay since losing
> any low bits would break the comparison.
> 
> > but even then
> > I'd say the "native" version is preferable.
> 
> I suspect this is probably true, though I also worry a bit whether
> there are archs where it does something inefficient or broken.

e.g. isnan is broken with -fsignaling-nan since it should
not signal but the ucomisd gcc generates does as discussed.

(although it's unlikely to matter much: we dont support
snan in all apis)

but gcc used to generate horrible code for fpclassify things,
nowadays it should be mostly fixed, i don't remember if
there were actual correctness bugs or just inefficient code.

> 
> Ideally the compiler would be able to recognize portable (within IEEE)
> patterns for floating point representation examination and optimize
> them if there's a more efficient way to be able to do it for a
> particular machine.
> 
> Rich