Re: [PATCH] math: move i386 sqrtf to C

[Rich Felker <dalias@libc.org>]

On Fri, Jan 10, 2020 at 12:18:58AM +0100, Szabolcs Nagy wrote:
> * Rich Felker <dalias@libc.org> [2020-01-09 17:00:14 -0500]:
> > On Thu, Jan 09, 2020 at 10:00:06PM +0100, Szabolcs Nagy wrote:
> > > > Note that eval_as_float only helps if -ffloat-store is used, which is
> > > > a nasty hack and also nonconforming, arguably worse than the behavior
> > > > without it, so we should probably drop use of that as a fallback, and
> > > > use fp_barrier[f] instead if needed.
> > > 
> > > i think -ffloat-store almost always drops excess precision
> > > including returns and assignments, so with that no
> > > annotation is needed. but yes the way the annotation is
> > > defined now is not useful against broken compilers or
> > > non-standard excess precision setting, in glibc the
> > > annotation is defined differently (with inline asm).
> > 
> > I was thinking in the context of wanting to remove from configure the:
> > 
> > || { test "$ARCH" = i386 && tryflag CFLAGS_C99FSE -ffloat-store ; }
> > 
> > which is probably doing more harm than good. Do you know if there are
> > things that'd break if we did that? I think eval_as_float should
> > probably be defined as fp_barrierf to make it safe in your code,
> > conditional on FLT_EVAL_METHOD>0 (and likewise >1 for eval_as_double).
> 
> i think -fexcess-precision=standard was introduced in
> gcc 4.5 and to get reliable behaviour before that we
> needed -ffloat-store.

I don't think the behavior was "reliable" with -ffloat-store; it's
wrong with respect to the defined meaning of FLT_EVAL_METHOD.

> since we had -ffloat-store i turned off the volatile
> hacks in commit 6d3f1a39c14b12026df84f386875b094e3652990
> and later completely removed the annotations in commit
> 9b0fcb441a44456c7b071c7cdaf90403f81ec05a

Thanks for these references.

> on new compilers -fexcess-precision=standard is used,
> but that turned out to do too many stores on the fdlibm
> code (which is why glibc kept using =fast), so in commit
> e216951f509b71da193da2fc63e25b998740d58b i started using
> float_t and double_t to get fast code in standard mode.
> (of course this made things worse for -ffloat-store).

This was the right thing to do, and I think it largely but not
entirely eliminates the need for caring about how the compiler handles
this, except in a few cases. It could probably be eliminated in more.
For example the argument reduction code cited above could use the
right constants for double_t rather than double to avoid the need to
store/load to drop excess precision.

> i think we would need to add back the old annotations
> to make old compilers safe without -ffloat-store.
> (fdlibm often raises fenv exceptions via a final rounding
> before return, those could be often handled more cleanly
> by __math_oflow etc helpers, but since it was not designed
> for inline errno handling some normal return paths can
> raise fp exceptions too and thus need eval_as_* annotation).

I think I asked you before, but from a standpoint of fenv stuff, I'm
confused why the eval_as_* things are useful at all; it looks like you
would need fp_barrier* to ensure they're actually evaluated (e.g. in
the presence of LTO with a compiler that doesn't honor fenv right).

But I think it's also useful to distinguish between possibility of
wrong exceptions being raised, which is a rather minor issue since
some widely-used compilers don't even support fenv reasonably at at
all, and the possibility of wrong values being returned for functions
where the result is required to be correctly rounded. I would deem it
a serious problem for sqrt[f] or fma[f] to return the wrong value when
compiled with gcc3 or pcc. I don't think I would particularly care if
exceptions failed to be raised properly when compiled with gcc3 or
pcc, though. So I probably would like to ensure that, whatever code we
end up with in i386 sqrt[f].c, it it ends up working even if the
compiler does not handle excess precision correctly.

Rich