On Fri, Nov 6, 2020 at 4:51 PM George Koehler wrote:
> Your ilogb fix is ok gkoehler@
>
It's annoying that C and/or ieee754 and the original hardware
implementation in the x87 instructions diverged in their definitions, but
the former is what matters and libm needs to follow that. ok guenther@
> On Sat, 31 Oct 2020 16:09:07 +0100 (CET)
> Mark Kettenis wrote:
>
> > - Dropping the amd64 and i386 versions. Fixing the corner cases in
> > assembly is hard, and the C implementation should be fast enough for
> > regular floating-point values.
>
> The amd64 and i386 assembly uses the x87 fxtract instruction. I feel
> that x87 instructions are obsolete on amd64. <...>
Umm, no? The amd64 ABI defines "long double" as matching the format used
by the x87 instructions; a function returning that type returns it in the
x87 %st(0) register and a review of libm finds many of the functions are
naturally implemented by the native x87 instructions.
I believe the main issue in this case is that the standard evolved away
from what Kahan originally did as a consultant with Intel. For a note on
how the ieee754 standard originated in (large) part from what Kahan did
with Intel, see
https://people.eecs.berkeley.edu/~wkahan/ieee754status/754story.html
I'm not nearly enough of a numerical analyst to judge the decision of the
standard.
Philip Guenther
