Since C99, `FLT_EVAL_METHOD` should be left for the compiler/libc to
define. Its redefinition breaks compilation with picolibc as the
target's libc, since it defines said symbol in math.h before the libm
define is evaluated by the compiler.
In its place, there is a check to make sure floating point type sizes
are what are expected to be, triggering a compilation error if those
assumptions are no longer valid.
Co-authored-by: Angus Gratton <angus@redyak.com.au>
Signed-off-by: Alessandro Gatti <a.gatti@frob.it>
This code is imported from musl, to match existing code in libm_dbl.
The file is also added to the build in stm32/Makefile. It's not needed by
the core code but, similar to c5cc64175be32cb1e4f3f1a249667bc9f5a12613,
allows round() to be used by user C modules or board extensions.
Identical to cd527bb324ade952d11a134859d38bf5272c165e but for doubles.
This gives a -2.754% improvement on bm_float.py, and -35% improvement on
calling sqrt in a loop.
