Here is a small addendum which fixes some problems that I ran into when building avr-libc with avr-gcc v3.4.6:

1) The compiler might not recognize -Wno-sign-compare or -Wno-maybe-uninitialized which are used to silence some build warnings.

2) Old GCC does not built-in define __SIZEOF_DOUBLE__ and __SIZEOF_LONG_DOUBLE__, hence do it by hand.



(file #48240)

Here is an updated version of the patch with the following changes:

• It also defines alias symbols for long double provided long double = float.
• It provides long double prototypes in math.h.
• Adds a new file common/alias.h which defines layout-dependent alias macros (C + Asm).
• Turns double into float in the C-part of the lib (dtostr[ef], vfprintf, atof, ...).
• vfprintf depends on sizeof ([long] double) and distinguishes between %lf and %f if their layouts differ.  However, we are running out of flags (which are currently 8 at most).
• It fixes abs / labs implementation so they do no more conflict with stdlib.h.
• Fixes a bunch of unrelated build warnings so it can be built with CFLAGS=-Werror

(file #47938)

FYI, there is the according GCC issue + patch

http://gcc.gnu.org/PR92055

From the doc of https://gcc.gnu.org/bugzilla/attachment.cgi?id=47114&action=diff

+@item -mdouble=@var{bits}
+@opindex mdouble
+Set the size (in bits) of the @code{double} type.  Valid values for
+@var{bits} are 32 and 64. This option is only available if the compiler
+is configured with @code{--with-double=32} or with @code{--with-double=64},
+and the argument of @code{--with-double=} specifies the default layout of
+@code{double}.
+
+@item -mlong-double=@var{bits}
+@opindex mlong-double
+Set the size (in bits) of the @code{long double} type.  Valid values for
+@var{bits} are 32 and 64. This option is only available if the compiler
+is configured with @code{--with-long-double=32} or with
+@code{--with-long-double=64}, and the argument of @code{--with-long-double=}
+specifies the default layout of @code{long double}.

comment #5:

> Thanks for the explanation.
>
> I think I'll apply the patch then as is by now, and we might
> add some words about the usage later on, when the configuration
> is settled.

Great.  The current patch is supposed to be a no-op, so that no observable effect should occur.

comment #3:

> Curious, what's the compiler switch to select 32- vs. 64-bit double? I'd like to add a bit of documentation about the entire topic as well.

I didn't settle on an interface yet, my preferred choice would be configure options

--with-double=[32|64]
--with-long-double=[32|64]

where --with-double= triggers availability of, say, -mdouble64 and also sets the default: --with-double=32 would make 64-bit double available, but the default would still be double=float.

But the problem is not with the (configure) option(s), it's the extension of the multilib space:

We currently have float/double/long-double = 32/32/32, and with the new multilib variants we'd also have 32/32/64 and 32/64/64.  This would mean blowing up the number of multilibs to 57 or beyond...

For libgcc I am using a hack to copy the [long-]double multilibs from the traditional one so that the build-times don't go up.  This works because the multilibs are the same: double=64 vs. double=32 is accomplished by referencing __adddf3 instead of __addsf3, not by a different implementation of __adddf3.

For libc, there are only changes in a few places:

• How and if symbol aliases are provided (i.e. the topic of this issue).

• How printf et al. interprets %f, %lf, %g etc.

• For any functions that use open-coded [long-]double like dtostre (which is ftostre actually).  We can implement this as ftostre and provide an alias or a prototype with asm name depending on double layout.

• And of course the multilib structure.  I still have no clue how to integrate `avr-gcc --print-multi-lib` into the configure process and how to make use of `avr-gcc --print-multi-directory --mmcu=MMCU ...`.  This would be very much appreciated in order to make avr-libc more independent of avr-gcc's version, multilib-structure and multilib-options.  Ideas?

Here is a version without __asm(symbol) that defines a symbol for double only if sizeof(double) == sizeof(float).

        * include/math.h
        (cosf, sinf, tanf, fmodf, modff, sqrtf, cbrtf, hypotf, squaref)
        (floorf, ceilf, frexpf, ldexpf, expf, coshf, sinhf, tanhf, asinf)
        (acosf, atanf, atan2f, logf, log10f, powf, isnanf, isinff)
        (signbitf, fdimf, fmaf, fmaxf, fmin, truncf, roundf, lroundf)
        (lrintf): Turn from macro to proper prototype.
        (fabs, fabsf, isfinite, isfinitef, copysign, copysignf): Implement.

        * libm/fplib/asmdef.h (ALIAS_ENTRY): Remove.
        (ENTRY_DOUBLE32): New .macro that only defines the symbol
        if __SIZEOF_DOUBLE__ == __SIZEOF_FLOAT__.
        * libm/fplib/acos.S: Use correct names for functions that deal
        with float, e.g. acosf instead of acos.  Provide double symbol
        only if __SIZEOF_DOUBLE__ == __SIZEOF_FLOAT__.
        * libm/fplib/asin.S: Same.
        * libm/fplib/atan.S: Same.
        * libm/fplib/atan2.S: Same.
        * libm/fplib/cbrt.S: Same.
        * libm/fplib/ceil.S: Same.
        * libm/fplib/copysign.S: Same.
        * libm/fplib/cos.S: Same.
        * libm/fplib/cosh.S: Same.
        * libm/fplib/exp.S: Same.
        * libm/fplib/fdim.S: Same.
        * libm/fplib/floor.S: Same.
        * libm/fplib/fma.S: Same.
        * libm/fplib/fmax.S: Same.
        * libm/fplib/fmin.S: Same.
        * libm/fplib/fmod.S: Same.
        * libm/fplib/fp_arccos.S: Same.
        * libm/fplib/frexp.S: Same.
        * libm/fplib/hypot.S: Same.
        * libm/fplib/isfinite.S: Same.
        * libm/fplib/isinf.S: Same.
        * libm/fplib/isnan.S: Same.
        * libm/fplib/ldexp.S: Same.
        * libm/fplib/log.S: Same.
        * libm/fplib/log10.S: Same.
        * libm/fplib/lrint.S: Same.
        * libm/fplib/lround.S: Same.
        * libm/fplib/modf.S: Same.
        * libm/fplib/pow.S: Same.
        * libm/fplib/round.S: Same.
        * libm/fplib/signbit.S: Same.
        * libm/fplib/sin.S: Same.
        * libm/fplib/sinh.S: Same.
        * libm/fplib/sqrt.S: Same.
        * libm/fplib/square.S: Same.
        * libm/fplib/tan.S: Same.
        * libm/fplib/tanh.S: Same.
        * libm/fplib/trunc.S: Same.

(file #47711)

Implementation of math.h / libm currently blocks 64-bit double, namely:

1) Function names are wrong, e.g. 32-bit implementation of sine is named "sin", not "sinf".

2) There are "aliases" in math.h (macros actually) that define the float version (like "sinf") to the double version (like "sin").  Problem is that this:

2a) This is the wrong way round ("sin" should be alias of "sinf").

2b) There are no proper prototypes.

2c) Macros are bad, e.g. in C++ when some class implements a method for which math.h defines a macro, C++ functions / methods are silently renamed.

3) math.h is unconditional, i.e. treats double != float just like double = float.