[Mesa3d-dev] Linux-x86 FAST_MATH macros in mmath.h

[Gareth H. <ga...@va...>]

After discussions with Brian and Josh, I'm adding an fnclex instruction
to the END_FAST_MATH macro to clear any FP exceptions that may have
occurred in the FAST_MATH block.  It seems pointless to have a
GCC-specific END_FAST_MATH, when I can convert the entire block to
inline assembly and get rid of the _FPU_SET* checks etc.

Do people have strong objections to this?  I'm essentially adding GCC as
a requirement to enable the fast math stuff, but with our change in
policy regarding FP exceptions this seems reasonable -- and we can
always add non-GCC code for other Linux-x86 targets as required.

-- Gareth