yann@1201
|
1 |
Original patch from: gentoo/src/patchsets/glibc/2.9/0090_all_glibc-math-tests.patch
|
yann@1201
|
2 |
|
yann@1201
|
3 |
-= BEGIN original header =-
|
yann@1201
|
4 |
From: "Joseph S. Myers" <joseph at codesourcery dot com>
|
yann@1201
|
5 |
To: libc-alpha at sourceware dot org
|
yann@1201
|
6 |
Date: Tue, 15 Jul 2008 14:43:04 +0000 (UTC)
|
yann@1201
|
7 |
Subject: expm1 ulps
|
yann@1201
|
8 |
|
yann@1201
|
9 |
If you test glibc on i686 with GCC 4.3, you get a test-ildoubl failure:
|
yann@1201
|
10 |
|
yann@1201
|
11 |
Failure: Test: expm1 (1) == M_El - 1.0
|
yann@1201
|
12 |
Result:
|
yann@1201
|
13 |
is: 1.71828182845904523532e+00 0xd.bf0a8b14576953500000p-3
|
yann@1201
|
14 |
should be: 1.71828182845904523543e+00 0xd.bf0a8b14576953600000p-3
|
yann@1201
|
15 |
difference: 1.08420217248550443401e-19 0x8.00000000000000000000p-66
|
yann@1201
|
16 |
ulp : 1.0000
|
yann@1201
|
17 |
max.ulp : 0.0000
|
yann@1201
|
18 |
Maximal error of `expm1'
|
yann@1201
|
19 |
is : 1 ulp
|
yann@1201
|
20 |
accepted: 0 ulp
|
yann@1201
|
21 |
|
yann@1201
|
22 |
What happens is that the inline expansion of expm1l uses __builtin_expm1l,
|
yann@1201
|
23 |
and GCC 4.3 optimizes calls to __builtin_expm1l with constant argument to
|
yann@1201
|
24 |
a correctly rounded result using MPFR. The result returned is thus the
|
yann@1201
|
25 |
value of e-1 rounded once to long double precision. However, the test
|
yann@1201
|
26 |
expects M_El - 1.0, and the result of rounding e to long double precision,
|
yann@1201
|
27 |
then subtracting 1, differs in the last place from the result of rounding
|
yann@1201
|
28 |
e-1 to long double precision (the latter has smaller exponent, and the
|
yann@1201
|
29 |
last bit is 1).
|
yann@1201
|
30 |
|
yann@1201
|
31 |
There are two obvious approaches possible to fixing this. The first patch
|
yann@1201
|
32 |
below changes the expectation to a decimal expansion for e-1 (taken from
|
yann@1201
|
33 |
that of M_El) rather than doing arithmetic in the expected value. This in
|
yann@1201
|
34 |
turn requires ulps to be set for the out-of-line version of expm1. It
|
yann@1201
|
35 |
might also need ulps to be set for the inline version for older compilers
|
yann@1201
|
36 |
if they should continue to pass the test, and possibly for other targets.
|
yann@1201
|
37 |
The second patch below takes the alternative approach of keeping the
|
yann@1201
|
38 |
existing expectation (which has the wrong bit in the last place) and
|
yann@1201
|
39 |
setting ulps for the inline expansion of expm1, which avoids the risk of
|
yann@1201
|
40 |
breaking the test for other targets.
|
yann@1201
|
41 |
|
yann@1201
|
42 |
2008-07-15 Joseph Myers <joseph@codesourcery.com>
|
yann@1201
|
43 |
|
yann@1201
|
44 |
* sysdeps/i386/fpu/libm-test-ulps: Add inline long double ulps for
|
yann@1201
|
45 |
expm1.
|
yann@1201
|
46 |
|
yann@1201
|
47 |
-= END original header =-
|
yann@1201
|
48 |
|
yann@1201
|
49 |
diff -durN glibc-2_9.orig/sysdeps/i386/fpu/libm-test-ulps glibc-2_9/sysdeps/i386/fpu/libm-test-ulps
|
yann@1201
|
50 |
--- glibc-2_9.orig/sysdeps/i386/fpu/libm-test-ulps 2006-01-15 18:59:37.000000000 +0100
|
yann@1201
|
51 |
+++ glibc-2_9/sysdeps/i386/fpu/libm-test-ulps 2009-02-02 22:00:47.000000000 +0100
|
yann@1201
|
52 |
@@ -453,6 +453,10 @@
|
yann@1201
|
53 |
ildouble: 8
|
yann@1201
|
54 |
ldouble: 8
|
yann@1201
|
55 |
|
yann@1201
|
56 |
+# expm1
|
yann@1201
|
57 |
+Test "expm1 (1) == M_El - 1.0":
|
yann@1201
|
58 |
+ildouble: 1
|
yann@1201
|
59 |
+
|
yann@1201
|
60 |
# gamma
|
yann@1201
|
61 |
Test "gamma (-0.5) == log(2*sqrt(pi))":
|
yann@1201
|
62 |
double: 1
|
yann@1201
|
63 |
@@ -1134,6 +1138,9 @@
|
yann@1201
|
64 |
ildouble: 8
|
yann@1201
|
65 |
ldouble: 8
|
yann@1201
|
66 |
|
yann@1201
|
67 |
+Function: "expm1":
|
yann@1201
|
68 |
+ildouble: 1
|
yann@1201
|
69 |
+
|
yann@1201
|
70 |
Function: "gamma":
|
yann@1201
|
71 |
double: 1
|
yann@1201
|
72 |
idouble: 1
|