Imo the best solution would be special "fast" floating-point types, that have less strict requirements.
I personal almost always use -ffast-math by default in my C programs that care about performance, because I almost never care enough about the loss in accuracy. The only case I remember needing it was when doing some random number distribution tests where I cared about subnormals, and got confused for a second because they didn't seem to exist (-ffast-math disables them on x86).
That or a scoped optimization directive. GCC does allow `__attribute__((optimize("-ffast-math")))` as a function-wide attribute, but Clang doesn't seem to have an equivalent and the standard syntax `[[gcc::optimize("-ffast-math")]]` doesn't seem to work as well. In any case, such optimization should be visible from the code in my opinion.
GP seems to be saying that you can flag individual functions in GCC, thereby avoiding this issue: only flagged functions would be compiled with fast math semantics.
This only work if it's a leaf function that will throw away the result. If you feed the result of your --fast-math function into other working code you risk breaking it.
`-ffast-math` is fully local, asides from GCC's unexpected `crtfastmath.o` linkage which is global.
Functions with `-ffast-math` enabled still return fp values via usual registers and in usual formats. If some function `f` is expected to return -1.0 to 1.0 for particluar inputs, `-ffast-math` can only make it to return 1.001 or NaN instead. If another function without `-ffast-math` expects and doesn't verify f's return value, it will surely misbehave, but only because the original analysis of f no longer holds.
`-ffast-math` the compiler option is bad because this effect is not evident from the code. Anything visible in the code should be okay.
I personal almost always use -ffast-math by default in my C programs that care about performance, because I almost never care enough about the loss in accuracy. The only case I remember needing it was when doing some random number distribution tests where I cared about subnormals, and got confused for a second because they didn't seem to exist (-ffast-math disables them on x86).