Sure, but I still prefer reverting commit 9d7ab0e at the same time.

Also, this seems to leave macOS/AppleClang as the only other (major/supported) environment w/ these pragmas on math functions... Are we sure they're needed there at all? Can't we just simply remove them? @zisoft

Sure, but I still prefer reverting

Latest squashed and force-pushed commit includes the revert.

Are we sure they're needed there at all?

I would even say they are very likely not wanted and possibly harmful.

The whole story reminds me of

DT_OMP_DECLARE_SIMD()
static inline float sqf(const float x)

in rcd demosaicing we compile with some extra flags

  #pragma GCC optimize ("fast-math", "fp-contract=fast", "finite-math-only", "no-math-errno")

definitely improving performance but using above sqf() definition lead to artifacts.

Performance-wise, the biggest impact comes from -ffast-math, and that's almost certainly the cause of artifacts. With that option enabled, the compiler is free to replace divisions by multiplications with the reciprocal, for which it uses a fast-but-approximate reciprocal instruction with the result that the division only has 12 bts of precision instead of 24....

Yes, so avoided divs there.

Yet, the specific problem there was related to the simd sqf() function.