https://gcc.gnu.org/bugzilla/show_bug.cgi?id=91645
--- Comment #11 from Jakub Jelinek <jakub at gcc dot gnu.org> ---
(In reply to Aldy Hernandez from comment #9)
> It looks like what we want for this test is actually !isgreaterequal() not
> isless(), since we want to exclude the possibility of a NAN. Like this:
>
> float test (float x)
> {
> if (!__builtin_isgreaterequal (x, 0.0))
> __builtin_unreachable();
> return sqrtf (x);
> }
No, what we emit for sqrtf is effectively:
if (!__builtin_isless (x, 0.0))
asm ("sqrtss %0, %0" : "+x" (x));
else
x = sqrtf (x); // The library version
return x;
So, the
if (__builtin_isless (x, 0.0)
__builtin_unreachable ();
is all we need.
We should either ask in tree-call-cdce.cc before trying to create this whether
x range is [-0.0, +Inf] NAN or some subset thereof, in that case just use
the .SQRT internal call, or if it is [-Inf, nextafterf (-0.0, -Inf)] or subset
thereof,
then use the library call only.
Or try to fold it in vrp2 or when, but in that case we don't know that easily
whether
the comparison can be optimized away completely (it still raises exceptions on
sNaN).
