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

[Rich Felker <dalias@libc.org>]

On Thu, Jan 09, 2020 at 10:00:06PM +0100, Szabolcs Nagy wrote:
> > If we want to ensure correct rounding (important for sqrt[f]) even on
> > broken compilers (some ppl use gcc 3.x, and pcc may be broken too?)
> > perhaps we should just do the store from asm too?
> 
> that would be a bit safer, but then correct compiler would
> store twice i think.

If you did something like:

	float y = expr_with_excess_precision;
	__asm__( "" : "+m"(y));
	return y;

then I think you'd get just one store and one load, as intended. It
seems to work as intended here. Oddly though my local gcc (7.3) is
gratuitously pushing/popping a single gpr to align stack to 8 (but not
16) despite being a leaf function. No idea why.

> it's hard to get excited about this
> issue: it only matters on m68k and i386 which are not the
> main targets for new float code (and old code had to deal
> with this and bigger brokenness already).

Indeed, but context of present thread is getting rid of the i386 asm
files so it's relevant here.

> > 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).

Rich