Re: [avr-libc-dev] libm for doubles -- how much effort is needed?

[Joerg Wunsch]

As Tobias Frost wrote:

> currently I have to do an estimation how much time it will take to
> implement double maths to the avr-libm. Based on the outcoming
> estimation, it might happen that the company will sponsor this.

Perhaps Dmitry can answer that.

Note that parts of our libm.a functionality rather belongs into
libgcc.a because they are implementing GCC auxiliary functions.

> Well, I have no idea, how much work the adaption of the GCC will be,
> and -- beside the implementation of the math itself --.

Implementing them into GCC is the first step needed so.  You could
have 64-bit doubles in GCC without an accompanying libm.a, but you
could not (usefully) starting to implement a 64-bit FP math library
without compiler support for the number format.

It is probably not too hard to implement, but as usual: Der Teufel
steckt im Detail.  I'd also like to have it as an option similar to
-mint8, say -mdouble32, defaulting to 64-bit doubles, so the compiler
is by default compliant to the C standard, yet we expect most AVR
users to simply put -mdouble32 into their Makefiles then.  Just as
an idea:

sizeof()    float     double   long double
-mdouble32  4         4        8
default     4         8        8

That would give people some flexibility to have 64-bit FP numbers
even when using -mdouble32 (by using long double).

Of course, the library mess has to be sorted out, too: not only the
correct libm.a needs to be selected depending on -mdouble32, but
functions from the standard library dealing with FP numbers (strtod()
comes to mind) also have to work correctly.  That could perhaps be
handled with some preprocessor tricks that divert the actual strtod()
to some internal __strtod32(), or __strtod64(), respectively.

-- 
cheers, J"org               .-.-.   --... ...--   -.. .  DL8DTL

http://www.sax.de/~joerg/                        NIC: JW11-RIPE
Never trust an operating system you don't have sources for. ;-)