olm_ prefixes for openlibm functions

[ViralBShah]

@vtjnash, @loladiro, Do you think we should have olm_ prefixes for openlibm functions, so that they don't clash with the system libm?

This could be done through a macro, which allows use of libm names, or with a prefix, depending on what the user wants.

[Keno]

That is certainly a possibility, but since we are explicitly resolving symbols in Julia anyway, it shouldn't really matter. But then again it doesn't hurt either.

[ViralBShah]

Agree. At least for julia usage, not critical.

[ViralBShah]

I would like to keep this open, since this would be a generally useful thing to have, even though it is not immediately necessary.

[ViralBShah]

@staticfloat If you are in the guts of openlibm, this would be a good one to tackle. Having olm_ prefixes would make it so much easier to not confuse openlibm functions with system libm, and also silence compiler warnings, which are currently suppressed in the julia compiler.

[staticfloat]

By this do you mean openlibm-specific functions, or just all functions? I kind of like the ability to use openlibm as a drop-in replacement for libm to see how it fares supporting other applications.

[ViralBShah]

The way to do this would be through a Makefile option, which allows you to build a drop-in replacement, or one with the prefixes.

[jiahao]

Perhaps we can close this by using objcopy to rename the exported symbols on demand. (This approach was suggested by @stevengj for renaming OpenBLAS symbols here).

[ViralBShah]

Yes, I was wondering if such a thing was possible. Good to know. Is there an equivalent thing on all OSes?

[stevengj]

objcopy is available on OSX, and on Windows via mingw I think.

[EdSchouten]

The code in #77 would actually make this rather easy to support. All source files will include <openlibm.h> and obtain the prototypes through there. We could simply put a macro in that header and use that in all of the sources:

#define OLM_F(name) olm_##

float OLM_F(cosf)(float x) {
...
}

It would also make the code slightly cleaner, because we could get rid of the _ieee754*() garbage at the same time.

[ViralBShah]

@EdSchouten This is too much to ask - but if you have tried this, could you please submit a PR? It will be so much cleaner to use code with the prefixes, when one needs to be absolutely sure about calling the right versions.

[ViralBShah]

I wondered why the _ieee754* stuff was there.

[ViralBShah]

I am thinking of doing the name mangling, and also simultaneously having the option of using host's fenv.h and various fe* functions from the system libm. That way, we don't have to manage the messy floating point exception stuff, where we are better off using the system provided one.