On Sat, Sep 10, 2016 at 09:47:55PM -0700, Philip Guenther wrote:
>
> On systems that don't have a native version, we use an implementation of
> sqrtl() (square-root of long double) that -- to do its job -- pokes at the
> floating-point exception state and rounding mode. In particular, at a key
> point it clears any previous "inexact" exception and sets the rounding
> mode to "toward zero" and then does a division. It then tests whether
> that raised an "inexact" exception and fixes the result up based on that,
> and finally restores the rounding mode before returning.
>
> The current version does that using the old, non-standard fp* routines:
> fp{set,get}sticky() and fpsetround(). This diff switches it to the new,
> standardized fe* routines: fe{clear,test}except() and fe{get,set}round().
>
> (Why bother? The fp* routines are defined in libc, while the fe* routines
> are defined inside libm itself...which means that with some symbol
> redirection they can be made to call directly, without going through the
> PLT. This diff is thus a prelude to the larger diff I have sitting in my
> tree to do exactly that, reducing 136 PLT entries to just 22 on amd64, for
> example. Even on a HW-FP-poor arch like mips64 it gets reduced from 204
> to only 56 PLT entries.)
>
> ok?
>
looks correct to me, ok tb
