Richard Earnshaw wrote:
> I've just seen a sudden large increase in the number of gcc testsuite
> failures for ARM after updating my newlib sources. It looks like libc
> has quietly (probably erroneously) grown a link-time dependency on libm:
>
> libc.a:lib_a-strtod.o: U __fpclassifyd
> libc.a:lib_a-svfprintf.o: U __fpclassifyd
> libc.a:lib_a-svfscanf.o: U __fpclassifyd
> libc.a:lib_a-svfwprintf.o: U __fpclassifyd
> libc.a:lib_a-svfwscanf.o: U __fpclassifyd
> libc.a:lib_a-vfprintf.o: U __fpclassifyd
> libc.a:lib_a-vfscanf.o: U __fpclassifyd
> libc.a:lib_a-vfwprintf.o: U __fpclassifyd
> libc.a:lib_a-vfwscanf.o: U __fpclassifyd
> libc.a:lib_a-wcstod.o: U __fpclassifyd
>
> libm.a:lib_a-s_fpclassify.o:00000000 T __fpclassifyd
>
>
> My guess is that this was caused by the following change:
>
> 2009-06-16 Craig Howland <howland@LGSInnovations.com>
>
> * libc/include/math.h: Simplify fpclassify, isinf, isnan, and signbit
> macros to remove un-necessary extension use. isinf and isnan also
> changed to use fpclassify. isfinite macro modified to run faster by
> only calling fpclassify once instead of possibly twice.
>
>
>
>
Good catch. There are a set of math objs in libc that we copy over in
the top-level Makefile.am.
Changing infinity testing to use fpclassify caused the problem. I have
added the fpclassify routines to
the MATHOBJS_IN_LIBC list and regenerated Makefile.in.
2009-06-23 Jeff Johnston <jjohnstn@redhat.com>
* Makefile.am (MATHOBJS_IN_LIBC): Add s_fpclassify and
sf_fpclassify as these are now used for infinity testing.
* Makefile.in: Regenerated.
-- Jeff J.