Constexpr fma

[mborland]

Implements: fma, fmaf, and fmal

[jzmaddock]

On the one hand this looks trivial.... on the other hand do we not run the risk of compile-time and runtime fma's returning subtly differing values?

[mborland]

If the compiler doesn't optimize this to an FMA asm call and the builtin is not available then you are correct that it could be off by one rounding. C++20 allows for constexpr asm if you would prefer that I hard code it for major architectures.

[NAThompson]

@jzmaddock : Is an fma for multiprecision infeasible or just a PITA?

[mborland]

@NAThompson This is clean and good for review. The implementation is trivial, but with the documented compiler flag the results are exact. Here is the prototype for ima support in Boost.Multiprecision.

[mborland]

@NAThompson This is clean and good for review. The implementation is trivial, but with the documented compiler flag the results are exact. Here is the prototype for ima support in Boost.Multiprecision.

[NAThompson]

Cool; this looks good. Is the failed job a known issue?

[mborland]

Cool; this looks good. Is the failed job a known issue?

It's a sanitizer issue in autodiff that seems to be popping up consistently.