Hello, I have a couple of questions that maybe someone can help with. Exhibit A is an x86_64 FC3 system:
# objdump -p /usr/lib64/libfftw.so | grep -A20 Version Version References: required from libm.so.6: 0x09691a75 0x00 03 GLIBC_2.2.5 required from libc.so.6: 0x0d696913 0x00 04 GLIBC_2.3 0x09691a75 0x00 02 GLIBC_2.2.5 # rpm -qi fftw | grep ^[VRI] Version : 2.1.5 Vendor: (none) Release : 5 Build Date: Tue 14 Dec 2004 06:56:25 PM CST Install Date: Wed 23 Mar 2005 04:14:50 PM CST Build Host: extras64.linux.duke.edu
Exhibit B is an x86_64 FC5 system:
# objdump -p /usr/lib64/libfftw.so | grep -A20 Version Version References: required from libc.so.6: 0x0d696913 0x00 05 GLIBC_2.3 0x0d696914 0x00 04 GLIBC_2.4 0x09691a75 0x00 03 GLIBC_2.2.5 0x09691974 0x00 02 GLIBC_2.3.4
# rpm -qi fftw2 | grep ^[VRI] Version : 2.1.5 Vendor: (none) Release : 12.fc5 Build Date: Sat 18 Feb 2006 04:08:15 AM CST Install Date: Wed 26 Apr 2006 05:00:30 PM CDT Build Host: hammer2.fedora.redhat.com
The FC3 fftw references libm, the FC5 one doesn't. Is one of the two behaviours considered more proper? Scons' autoconf-like test fails with the FC5 libfftw because it can't resolve sin() and cos(). The FC3 libfftw seems to reference those, too, but the reference to libm takes care of that.
More in general, though, isn't a call to sin() and cos() expensive and shouldn't the compiler just use inline native FP code? In the case of FFTW those two are only used in the setup stages, with the actual FFTs implemented with multiplications and additions (gcc uses SSE as expected on x86_64), but not all code is going to be like that. Are there reasons for them not to be inlined?
Thanks.