| Commit message (Collapse) | Author | Age | Files | Lines |
|---|
| |
|
|
| |
MFC after: 2 weeks
|
| |
|
|
|
|
|
|
|
|
|
|
| |
be into 12+24 bits of precision for extra-precision multiplication,
but was into 13+24 bits. On i386 with -O1 the bug was hidden by
accidental extra precision, but on amd64, in 2^32 trials the bug
caused about 200000 errors of more than 1 ulp, with a maximum error
of about 80 ulps. Now the maximum error in 2^32 trials on amd64
is 0.8573 ulps. It is still 0.8316 ulps on i386 with -O1.
The nearby decomposition of 1/ln2 and the decomposition of 2/(3ln2) in
the double precision version seem to be sub-optimal but not broken.
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
This uses 2 tricks to improve consistency so that more serious problems
aren't hidden in simple regression tests by noise for the NaNs:
- for a signaling NaN, adding 0.0 generates the invalid exception and
converts to a quiet NaN, and doesn't have too many effects for other
types of args (it converts -0 to +0 in some rounding modes, but that
hopefully doesn't change the result after adding the NaN arg). This
avoids some inconsistencies on i386 and ia64. On these arches, the
result of an operation on 2 NaNs is apparently the largest or the
smallest of the NaNs as bits (consistently largest or smallest for
each arch, but the opposite). I forget which way the comparison
goes and if the sign bit affects it. The quiet bit is is handled
poorly by not always setting it before the comparision or ignoring
it. Thus if one of the args was originally a signaling NaN and the
other was originally a quiet NaN, then the result depends too much
on whether the signaling NaN has been quieted at this point, which
in turn depends on optimizations and promotions. E.g., passing float
signaling NaNs to double functions must quiet them on conversion;
on i387, loading a signaling NaN of type float or double (but not
long double) into a register involves a conversion, so it quiets
signaling NaNs, so if the addition has 2 register operands than it
only sees quiet NaNs, but if the addition has a memory operand then
it sees a signaling NaN iff it is in the memory operand.
- subtraction instead of addition is used to avoid a dubious optimization
in old versions of gcc. For SSE operations, mixing of NaNs apparently
always gives the target operand. This is not as good as the i387
and ia64 behaviour. It doesn't mix NaNs at all, and makes addition
not quite commutative. Old versions of gcc sometimes rewrite x+y
to y+x and thus give different results (in bits) for NaNs. gcc-3.3.3
rewrites x+y to y+x for one of pow() and powf() but not the other,
so starting from float NaN args x and y, powf(x, y) was almost always
different from pow(x, y).
These tricks won't give consistency of 2-arg float and double functions
with long double ones on amd64, since long double ones use the i387
which has different semantics from SSE.
Convert to __FBSDID().
|
| | |
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
[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.
This completes fixing all the bugs in powf() that were routinely found
by ucbtest.
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
(1) The bit for the 1.0 part of bp[k] was right shifted by 4. This seems
to have been caused by a typo in converting e_pow.c to e_powf.c.
(2) The lower 12 bits of ax+bp[k] were not discarded, so t_h was actually
plain ax+bp[k]. This seems to have been caused by a logic error in
the conversion.
These bugs gave wrong results like:
powf(-1.1, 101.0) = -15158.703 (should be -15158.707)
hex values: BF8CCCCD 42CA0000 C66CDAD0 C66CDAD4
Fixing (1) gives a result wrong in the opposite direction (hex C66CDAD8),
and fixing (2) gives the correct result.
ucbtest has been reporting this particular wrong result on i386 systems
with unpatched libraries for 9 years. I finally figured out the extent
of the bugs. On i386's they are normally hidden by extra precision.
We use the trick of representing floats as a sum of 2 floats (one much
smaller) to get extra precision in intermediate calculations without
explicitly using more than float precision. This trick is just a
pessimization when extra precision is available naturally (as it always
is when dealing with IEEE single precision, so the float precision part
of the library is mostly misimplemented). (1) and (2) break the trick
in different ways, except on i386's it turns out that the intermediate
calculations are done in enough precision to mask both the bugs and
the limited precision of the float variables (as far as ucbtest can
check).
ucbtest detects the bugs because it forces float precision, but this
is not a normal mode of operation so the bug normally has little effect
on i386's.
On systems that do float arithmetic in float precision, e.g., amd64's,
there is no accidental extra precision and the bugs just give wrong
results.
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
Fixed pow(x, y) when x is very close to -1.0 and y is a very large odd
integer. E.g., pow(-1.0 - pow(2.0, -52.0), 1.0 + pow(2.0, 52.0)) was
0.0 instead of being very close to -exp(1.0).
PR: 39236
Submitted by: Stephen L Moshier <steve@moshier.net>
e_powf.c:
Apply the same patch although it is just cosmetic because odd integers
large enough to cause the problem are too large to be precisely represented
as floats.
MFC after: 1 week
|
| |
|
|
|
|
|
|
|
|
| |
- float ynf(int n, float x) /* wrapper ynf */
+float
+ynf(int n, float x) /* wrapper ynf */
This is because the __STDC__ stuff was indented.
Reviewed by: md5
|
| |
|
|
| |
Reviewed by: md5
|
| | |
|
| |
|
|
|
|
| |
float versions. Using sqrt() was inefficient.
Obtained from: NetBSD
|
| | |
|
| |
|
|
|
|
|
|
| |
This will make a number of things easier in the future, as well as (finally!)
avoiding the Id-smashing problem which has plagued developers for so long.
Boy, I'm glad we're not using sup anymore. This update would have been
insane otherwise.
|
| | |
|
|
|
-- Begin comments from J.T. Conklin:
The most significant improvement is the addition of "float" versions
of the math functions that take float arguments, return floats, and do
all operations in floating point. This doesn't help (performance)
much on the i386, but they are still nice to have.
The float versions were orginally done by Cygnus' Ian Taylor when
fdlibm was integrated into the libm we support for embedded systems.
I gave Ian a copy of my libm as a starting point since I had already
fixed a lot of bugs & problems in Sun's original code. After he was
done, I cleaned it up a bit and integrated the changes back into my
libm.
-- End comments
Reviewed by: jkh
Submitted by: jtc
|
