This is the mail archive of the
libc-alpha@sourceware.org
mailing list for the glibc project.
expm1 ulps
- From: "Joseph S. Myers" <joseph at codesourcery dot com>
- To: libc-alpha at sourceware dot org
- Date: Tue, 15 Jul 2008 14:43:04 +0000 (UTC)
- Subject: expm1 ulps
If you test glibc on i686 with GCC 4.3, you get a test-ildoubl failure:
Failure: Test: expm1 (1) == M_El - 1.0
Result:
is: 1.71828182845904523532e+00 0xd.bf0a8b14576953500000p-3
should be: 1.71828182845904523543e+00 0xd.bf0a8b14576953600000p-3
difference: 1.08420217248550443401e-19 0x8.00000000000000000000p-66
ulp : 1.0000
max.ulp : 0.0000
Maximal error of `expm1'
is : 1 ulp
accepted: 0 ulp
What happens is that the inline expansion of expm1l uses __builtin_expm1l,
and GCC 4.3 optimizes calls to __builtin_expm1l with constant argument to
a correctly rounded result using MPFR. The result returned is thus the
value of e-1 rounded once to long double precision. However, the test
expects M_El - 1.0, and the result of rounding e to long double precision,
then subtracting 1, differs in the last place from the result of rounding
e-1 to long double precision (the latter has smaller exponent, and the
last bit is 1).
There are two obvious approaches possible to fixing this. The first patch
below changes the expectation to a decimal expansion for e-1 (taken from
that of M_El) rather than doing arithmetic in the expected value. This in
turn requires ulps to be set for the out-of-line version of expm1. It
might also need ulps to be set for the inline version for older compilers
if they should continue to pass the test, and possibly for other targets.
The second patch below takes the alternative approach of keeping the
existing expectation (which has the wrong bit in the last place) and
setting ulps for the inline expansion of expm1, which avoids the risk of
breaking the test for other targets.
2008-07-15 Joseph Myers <joseph@codesourcery.com>
* math/libm-test.inc (expm1_test): Give decimal expansion for
expected value of expm1(1).
* sysdeps/i386/fpu/libm-test-ulps: Add long double ulps for expm1.
Index: math/libm-test.inc
===================================================================
RCS file: /cvs/glibc/libc/math/libm-test.inc,v
retrieving revision 1.81
diff -u -r1.81 libm-test.inc
--- math/libm-test.inc 22 May 2008 19:59:10 -0000 1.81
+++ math/libm-test.inc 15 Jul 2008 14:27:09 -0000
@@ -2594,7 +2594,7 @@
#endif
TEST_f_f (expm1, nan_value, nan_value);
- TEST_f_f (expm1, 1, M_El - 1.0);
+ TEST_f_f (expm1, 1, 1.7182818284590452353602874713526625L);
TEST_f_f (expm1, 0.75L, 1.11700001661267466854536981983709561L);
END (expm1);
Index: sysdeps/i386/fpu/libm-test-ulps
===================================================================
RCS file: /cvs/glibc/libc/sysdeps/i386/fpu/libm-test-ulps,v
retrieving revision 1.44
diff -u -r1.44 libm-test-ulps
--- sysdeps/i386/fpu/libm-test-ulps 15 Jan 2006 17:59:37 -0000 1.44
+++ sysdeps/i386/fpu/libm-test-ulps 15 Jul 2008 14:27:09 -0000
@@ -453,6 +453,10 @@
ildouble: 8
ldouble: 8
+# expm1
+Test "expm1 (1) == 1.7182818284590452353602874713526625":
+ldouble: 1
+
# gamma
Test "gamma (-0.5) == log(2*sqrt(pi))":
double: 1
@@ -1134,6 +1138,9 @@
ildouble: 8
ldouble: 8
+Function: "expm1":
+ldouble: 1
+
Function: "gamma":
double: 1
idouble: 1
2008-07-15 Joseph Myers <joseph@codesourcery.com>
* sysdeps/i386/fpu/libm-test-ulps: Add inline long double ulps for
expm1.
Index: sysdeps/i386/fpu/libm-test-ulps
===================================================================
RCS file: /cvs/glibc/libc/sysdeps/i386/fpu/libm-test-ulps,v
retrieving revision 1.44
diff -u -r1.44 libm-test-ulps
--- sysdeps/i386/fpu/libm-test-ulps 15 Jan 2006 17:59:37 -0000 1.44
+++ sysdeps/i386/fpu/libm-test-ulps 15 Jul 2008 14:28:47 -0000
@@ -453,6 +453,10 @@
ildouble: 8
ldouble: 8
+# expm1
+Test "expm1 (1) == M_El - 1.0":
+ildouble: 1
+
# gamma
Test "gamma (-0.5) == log(2*sqrt(pi))":
double: 1
@@ -1134,6 +1138,9 @@
ildouble: 8
ldouble: 8
+Function: "expm1":
+ildouble: 1
+
Function: "gamma":
double: 1
idouble: 1
--
Joseph S. Myers
joseph@codesourcery.com