The current value used for ulp near zero is wrong,
and this commit fixes it such that ulp(0) is the smallest
subnormal value nearest to zero, which makes the most
sense for testing values near zero. Note that this is not
what Java does; they use the nearest normal value, which
is less accurate than what we want for glibc. Note that
there is no correct implementation of ulp since there
is no strict mathmatical definition that is accepted by
all groups using IEEE 754.
Previously with the large ulp values near zero there
were tests that previously passed, but were in fact
billions of ulp away from the precise answer. With this
commit we now need to disable one of the cpow tests which
is revealed to be inaccurate (bug 14473).
---
2013-05-24 Carlos O'Donell <carlos@redhat.com>
* math/libm-test.inc (MAX_EXP): Define.
(ULPDIFF): Define.
(ulp): New function.
(check_float_internal): Use ULPDIFF.
(cpow_test): Disable failing test.
(check_ulp): Test ulp() implemetnation.
(main): Call check_ulp before starting tests.
Use the most accurate hex literals possible for the answers to the
cos and sincos tests that vary according to the error in the rounding
of PI/2.
---
2013-04-24 Carlos O'Donell <carlos@redhat.com>
* math/libm-test.inc (cos_test): Use accurate hex constants.
(sincost_test): Likewise.
The value of PI is never exactly PI in any floating point representation,
and the value of PI/2 is never PI/2. It is wrong to expect cos(M_PI_2l)
to return 0, instead it will return an answer that is non-zero because
M_PI_2l doesn't round to exactly PI/2 in the type used.
That is to say that the correct answer is to do the following:
* Take PI or PI/2.
* Round to the floating point representation.
* Take the rounded value and compute an infinite precision cos or sin.
* Use the rounded result of the infinite precision cos or sin as the
answer to the test.
I used printf to do the type rounding, and Wolfram's Alpha to do the
infinite precision cos calculations.
The following changes bring x86-64 and x86 to 1/2 ulp for two tests.
It shows that the x86 cos implementation is quite good, and that
our test are flawed.
Unfortunately given that the rounding errors are type dependent we
need to fix this for each type. No regressions on x86-64 or x86.
---
2013-04-11 Carlos O'Donell <carlos@redhat.com>
* math/libm-test.inc (cos_test): Fix PI/2 test.
(sincos_test): Likewise.
* sysdeps/x86_64/fpu/libm-test-ulps: Regenerate.
* sysdeps/i386/fpu/libm-test-ulps: Regenerate.
run-via-rtld-prefix checks whether the program to be run is a static
test and skips if it is. This is fine, except that it assumes that
the program to be run is the second $^, which is true only for tests.
This change creates an rtld-prefix, which is simply the dynamic linker
prefix with the necessary arguments and uses that in the non-test
targets.
Document the use of the convenience testrun.sh script for
running the libm test.
---
2013-04-06 Carlos O'Donell <carlos@redhat.com>
* math/README.libm-test (How can I generate "libm-test-ulps"?):
Use testrun.sh to run libm tests.
The wiki "Regeneration" page has this to say about update ULPs.
"The libm-test-ulps files are semiautomatically updated. To
update an ulps baseline, run each of the failing tests (test-float,
test-double, etc.) with -u; this will generate a file called ULPs;
concatenate each of those files with the existing libm-test-ulps
file, after removing any entries for particularly huge numbers of
ulps that you do not want to mark as expected. Then run
gen-libm-test.pl -n -u FILE where FILE is the concatenated file
produced in the previous step. This generates a file called
NewUlps which is the new sorted version of libm-test-ulps."
The same information is listed in math/README.libm-test, and is a
lot of manual work that you often want to run over-and-over again
while working on a particular test.
The `regen-ulps' convenience target does this automatically for
developers.
We strictly assume the source tree is readonly and add a
new --output-dir option to libm-test.inc to allow for writing
out ULPs to $(objpfx).
When run the new target does the following:
* Starts with the baseline ULPs file.
* Runs each of the libm math tests with -u.
* Adds new changes seen with -u to the baseline.
* Sorts and prepares the test output with gen-libm-test.pl.
* Leaves math/NewUlps in your build tree to copy to your source
tree, cleanup, and checkin.
The math test documentation in math/README.libm-test is updated
document the new Makefile target.
---
2013-04-06 Carlos O'Donell <carlos@redhat.com>
* Makefile.in (regen-ulps): New target.
* math/Makefile [ifneq (no,$(PERL)]: Declare regen-ulps with .PHONY.
[ifneq (no,$(PERL)] (run-regen-ulps): New variable.
[ifneq (no,$(PERL)] (regen-ulps): New target.
[ifeq (no,$(PERL)] (regen-ulps): New target.
* math/libm-test.inc (ulps_file_name): Define.
(output_dir): New variable.
(options): Add "output-dir" option.
(parse_opt): Handle 'o' case.
(main): If output_dir is non-NULL use it as a prefix
otherwise use "".
* math/README.libm-test: Update `How can I generate "libm-test-ulps"?'
The patch increase the high value to check if expl overflows. Current
high mark value is not really correct, the algorithm accepts high values.
It also adds a correct wrapper function to check for overflow and underflow.
* math/Makefile: Recognize gmp-sysdep_routines.
* sysdeps/sparc/sparc64/multiarch/Makefile: Add VIS3 optimized GMP routines
to sysdeps.
* sysdeps/sparc/sparc64/multiarch/add_n-vis3.S: New file.
* sysdeps/sparc/sparc64/multiarch/add_n.S: New file.
* sysdeps/sparc/sparc64/multiarch/addmul_1-vis3.S: New file.
* sysdeps/sparc/sparc64/multiarch/addmul_1.S: New file.
* sysdeps/sparc/sparc64/multiarch/mul_1-vis3.S: New file.
* sysdeps/sparc/sparc64/multiarch/mul_1.S: New file.
* sysdeps/sparc/sparc64/multiarch/sub_n-vis3.S: New file.
* sysdeps/sparc/sparc64/multiarch/sub_n.S: New file.
* sysdeps/sparc/sparc64/multiarch/submul_1-vis3.S: New file.
* sysdeps/sparc/sparc64/multiarch/submul_1.S: New file.
With help from Joseph Myers.
* sysdeps/ieee754/ldbl-128/s_atanl.c (__atanl): Handle tiny and
very large arguments properly.
* math/libm-test.inc (atan_test): New tests.
(atan2_test): New tests.
* sysdeps/sparc/fpu/libm-test-ulps: Update.
* sysdeps/x86_64/fpu/libm-test-ulps: Update.
With help from Joseph Myers.
* sysdeps/ieee754/flt-32/e_j0f.c (__ieee754_y0f): Adjust tinyness
cutoff to 2**-13.
* sysdeps/ieee754/flt-32/e_j1f.c (__ieee754_y1f): Adjust tinyness
cutoff to 2**-25.
* sysdeps/ieee754/ldbl-128/e_j0l.c (U0): New constant.
( __ieee754_y0l): Avoid arithmetic underflow when 'x' is very
small.
* sysdeps/ieee754/ldbl-128/e_j1l.c (__ieee754_y1l): Likewise.
* math/libm-test.inc (y0_test): New tests.
(y1_test): New tests.
* sysdeps/i386/fpu/libm-test-ulps: Update.
* sysdeps/x86_64/fpu/libm-test-ulps: Update.
* sysdeps/sparc/fpu/libm-test-ulps: Update.
* sysdeps/i386/i686/fpu/multiarch/Makefile (sysdep_routines):
Add s_sinf-sse2, s_conf-sse2.
* sysdeps/i386/i686/fpu/multiarch/s_sinf-sse2.S: New file.
* sysdeps/i386/i686/fpu/multiarch/s_cosf-sse2.S: New file.
* sysdeps/i386/i686/fpu/multiarch/s_sinf.c: New file.
* sysdeps/i386/i686/fpu/multiarch/s_cosf.c: New file.
* sysdeps/ieee754/flt-32/s_sinf.c (SINF, SINF_FUNC): Add macros
for using routine as __sinf_ia32.
Use macro for function declaration and weak_alias.
* sysdeps/ieee754/flt-32/s_cosf.c (COSF, COSF_FUNC): Add macros
for using routine as __cosf_ia32.
Use macro for function declaration and weak_alias.
* sysdeps/i386/i686/fpu/multiarch/e_expf-sse2.S: Fix Copyright.
* sysdeps/i386/i686/fpu/multiarch/e_expf.c: Fix Copyright.
* sysdeps/x86_64/fpu/s_sinf.S: New file.
* sysdeps/x86_64/fpu/s_cosf.S: New file.
* sysdeps/x86_64/fpu/libm-test-ulps: Update.
* math/libm-test.inc (cos_test): Add more test cases.
(sin_test): Likewise.
(sincos_test): Likewise.
For some tests, just claim that fetestexcept() always returns true,
so the rest of the test can be compiled.
For libm-test, provide known bogus values for unsupported rounding
modes, so fesetround() will return failure.
Elsewhere, just add some #ifdefs to avoid code that uses particular
FP exceptions if the exceptions aren't supported.
POSIX 2008 states that if the input for 'logb[f|l]' is a subnormal number
it should be treated as if it were normalized. This means the
implementation should calculate the log2 of the mantissa and add it to the
subnormal exponent (-126 for float and -1022 for double and IBM long
double). This patch takes care of that.
[BZ #6794]
Following Joseph comments about bug 6794, here is a proposed fix. It turned out
to be a large fix mainly because I had to move some file along to follow libm
files/names conventions.
Basically I have added wrappers (w_ilogb.c, w_ilogbf.c, w_ilogbl.c) that now calls
the symbol '__ieee754_ilogb'. The wrappers checks for '__ieee754_ilogb' output and
set the errno and raise exceptions as expected.
The '__ieee754_ilogb' is implemented in sysdeps. I have moved the 's_ilogb[f|l]' files
to e_ilogb[f|l] and renamed the '__ilogb[f|l]' to '__ieee754_ilogb[f|l]'.
I also found out a bug in i386 and x86-64 assembly coded ilogb implementation where
it raises a FE_DIVBYZERO when argument is '0.0'. I corrected this issue as well.
Finally I added the errno and FE_INVALID tests for 0.0, NaN and +-InF argument. Tested
on i386, x86-64, ppc32 and ppc64.
Joseph Myers
Carlos O'Donell
Richard Henderson
Peter Collingbourne
Adhemerval Zanella
Siddhesh Poyarekar
Thomas Schwinge
David S. Miller
Andreas Schwab
Andreas Schwab
Mike Frysinger
Marcus Shawcroft
Markus Trippelsdorf
Liubov Dmitrieva
Marek Polacek
Chris Metcalf
Roland McGrath
Andreas Jaeger
Ryan S. Arnold