Ofast compiler flag in PowerPC

[imciner2]

We have been carrying a patch in Julia for quite a few years that removes all the occurrences of -Ofast from Makefile.power (the only place it is ever used in OpenBLAS), because we forbid its use in Julia.

Looking at how it is actually used in the command line, it is used with the fast math optimizations disabled:

CCOMMON_OPT += -Ofast -mcpu=power10 -mtune=power10 -mvsx -fno-fast-math

Those optimizations make up the core part of it, and other than enabling the O3 optimizations, it tells the compiler to also allow possible data races when optimizing, so I am curious why -Ofast used here?

Ideally I would like to work on removing our Julia-specific patch for this, and I can see that done in one of two ways:

1. Introduce a NO_OFAST build option that we can set that guards this.
2. Just remove it from those PowerPC build options.

[martin-frbg]

As far as I can tell, the -Ofast -fno-fast-math was introduced 9 years ago in #846 and has since been copied for POWER9 and POWER10 without questioning (including by contributors working for IBM). I'm not aware of any faults actually being observed from its use, although the implied "allow store data races" does not sound particularly healthy, and neither does stack-arrays at larger problem sizes... Given its history, it might well have been a kludge necessary for the specific version of GCC in use at the time (which would probably have been some 4.x).
I do notice that this option was not carried over to the CMake build system (while the no-fast-math was), so perhaps the
safest option now would be to replace it with -O3 - we certainly want to be fast but not furious ?

[martin-frbg]

closing as fixed by your PR #5323 - thank you