Skip to content

Commit

Permalink
powf: Fixed another precision bug in powf() (FreeBSD)
Browse files Browse the repository at this point in the history
Fixed another precision bug in powf(). This one is in the computation
[t=p_l+p_h High]. We multiply t by lg2_h, and want the result to be
exact. For the bogus float case of the high-low decomposition trick, we
normally discard the lowest 12 bits of the fraction for the high part,
keeping 12 bits of precision. That was used for t here, but it doesnt't
work because for some reason we only discard the lowest 9 bits in the
fraction for lg2_h.  Discard another 3 bits of the fraction for t to
compensate.

This bug gave wrong results like:

powf(0.9999999, -2.9999995) = 1.0000002 (should be 1.0000001)
hex values: 3F7FFFFF C03FFFFE 3F800002 3F800001

As explained in the log for the previous commit, the bug is normally
masked by doing float calculations in extra precision on i386's, but is
easily detected by ucbtest on systems that don't have accidental extra
precision.

Reference: freebsd/freebsd-src@5f20e5c
Original Author: Bruce Evans
  • Loading branch information
Fabian Schriever authored and jjohnstn committed Sep 19, 2024
1 parent fb76697 commit 1b7c72f
Showing 1 changed file with 1 addition and 1 deletion.
2 changes: 1 addition & 1 deletion newlib/libm/math/ef_pow.c
Original file line number Diff line number Diff line change
Expand Up @@ -242,7 +242,7 @@ ivln2_l = 7.0526075433e-06; /* 0x36eca570 =1/ln2 tail*/
}
t = p_l+p_h;
GET_FLOAT_WORD(is,t);
SET_FLOAT_WORD(t,is&0xfffff000);
SET_FLOAT_WORD(t,is&0xffff8000);
u = t*lg2_h;
v = (p_l-(t-p_h))*lg2+t*lg2_l;
z = u+v;
Expand Down

0 comments on commit 1b7c72f

Please sign in to comment.