------- Comment #4 from rob1weld at aol dot com 2007-06-03 16:05 ------- I copied gcc-4_3-build/i686-pc-linux-gnu/libjava/classpath/native/fdlibm/.libs/libfdlibm.a to my current directory and instead of using "-lm" I used "./libfdlibm.a" ...
Guess what. I propose this simple and useful fix: We need to re-arrange gcc to build a directory like "gcc-4_3-build/libdecnumber" and "gcc-4_3-build/prev-libdecnumber" called "gcc-4_3-build/libfdlibm" "gcc-4_3-build/prev-libfdlibm". The first pass can build it using "-O0" so it is certain to work (or we must fix that) and the next pass can build it with profiling and "-O3" enabled. Profiling works on 4.2.0 (but not on 4.3.0). A message can be printed that if profiling the math library fails to rebuild by typing "make nomathprofile". The third pass can use the profiled information to build the math library in a fully optimized state. The resulting super-optimized math library can then be used for the libjava library in the fourth "library build pass" and any other language that wants to take advantage of this. No effort would be wasted and we would have a fast / accurate library that works on all platforms - OR is the libfdlibm not a good plan and should be removed from libjava? If you have this page on your computer look at /usr/share/doc/glibc-2.5/manual/libc/Errors-in-Math-Functions.html - it is created when you compile libc6. That page is a report of the libc6 tests that are ran when the code is built and shows the resulting errors for every function (as is suggested in IEEE754). It looks like this (edited for brevity): The table lists the ULP values for different architectures. Different architectures have different results since their hardware support for floating-point operations varies and also the existing hardware support is different. Function Alpha Generic ix86 IA64 PowerPC acosf - - - - - acos - - - - - cosf 1 - 1 1 1 cos 2 - 2 2 2 cosl - - 1 1 1 logl - - - - 1 log10f 2 - 1 1 2 log10 1 - - - 1 log10l - - 1 1 1 powl - - - - 1 ... Since we link some of gcc with "-lm" and some with libfdlibm we have an inconsistant math library upon which gcc is based. We can standardize and test our math library to provide the same (correct) result on all platforms (as IEEE 754 requires). Lets remove "-lm" from gcc and use libfdlibm on as many platforms as is possible, where that doesn't work we simply let them use "-lm", like they did before, and they will be no worse off. The rest of us will be much better off. -- http://gcc.gnu.org/bugzilla/show_bug.cgi?id=32180
