where necessary.
The log1p() function is provided to compute an accurate value of
log(1 + x), even for tiny values of x. The i387 FPU provides the
fyl2xp1 instruction for this purpose.
However, since the range of the fyl2xp1 function is limited to
-(1 - (sqrt(2) / 2)) <= x <= sqrt(2) - 1
(-0.292893 <= x <= 0.414214)
we need to check if the argument is in the valid range.
In order to reduce the cost for testing the range, we only use
fyl2xp1 if the argument is in the range
-0.25 <= x <= 0.25
which can be checked with just one conditional branch.
Fixes PR lib/22599 by Ray Brownrigg.
cf92bf760d