Module Name: src
Committed By: dsl
Date: Sun Mar 16 22:49:27 UTC 2014
Modified Files:
src/tests/lib/libm: t_libm.h
Log Message:
Print the result as a 'long double' - on i386 a return value that
should be infinity might just be too large for 'double' and won't
get converted until it has to be saved to memory.
To generate a diff of this commit:
cvs rdiff -u -r1.4 -r1.5 src/tests/lib/libm/t_libm.h
Please note that diffs are not public domain; they are subject to the
copyright notices on the relevant files.
Modified files:
Index: src/tests/lib/libm/t_libm.h
diff -u src/tests/lib/libm/t_libm.h:1.4 src/tests/lib/libm/t_libm.h:1.5
--- src/tests/lib/libm/t_libm.h:1.4 Sun Mar 16 18:42:21 2014
+++ src/tests/lib/libm/t_libm.h Sun Mar 16 22:49:27 2014
@@ -1,15 +1,18 @@
-/* $NetBSD: t_libm.h,v 1.4 2014/03/16 18:42:21 dsl Exp $ */
+/* $NetBSD: t_libm.h,v 1.5 2014/03/16 22:49:27 dsl Exp $ */
/*
* Check result of fn(arg) is correct within the bounds.
* Should be ok to do the checks using 'double' for 'float' functions.
+ * On i386 float and double values are returned on the x87 stack and might
+ * be out of range for the function - so save and print as 'long double'.
+ * (otherwise you can get 'inf != inf' reported!)
*/
#define T_LIBM_CHECK(subtest, fn, arg, expect, epsilon) do { \
- double r = fn(arg); \
+ long double r = fn(arg); \
double e = fabs(r - expect); \
if (r != expect && e > epsilon) \
atf_tc_fail_nonfatal( \
- "subtest %u: " #fn "(%g) is %g (%.13a) not %g (%.13a), error %g (%.6a) > %g", \
+ "subtest %u: " #fn "(%g) is %Lg (%.14La) not %g (%.13a), error %g (%.6a) > %g", \
subtest, arg, r, r, expect, expect, e, e, epsilon); \
} while (0)
