Building for an SH target without FPU

[Oleg Endo]

On Wed, 2019-03-06 at 18:04 -0600, Segher Boessenkool wrote:
> Hi!
> 
> On Wed, Mar 06, 2019 at 07:15:54PM +0100, Sébastien Michelland wrote:
> > I'm having trouble building programs for a SuperH-4 target without
> > FPU.
> > 
> > My current setup uses an sh4eb-nofpu-elf target which usually
> > behaves 
> > very well (much better than sh4eb-elf that uses a lot of fmov for 
> > optimization), but I discovered that using a va_list causes the
> > dreaded, 
> > illegal fmov instruction to be generated.
> > 
> > I've looked up available options and expected -m4-nofpu to be of
> > help, 
> > however GCC does not agree with this estimation:
> > 
> >   sh4eb-nofpu-elf-gcc: error: command line option '-m4-nofpu' is
> > not 
> > supported by this configuration
> > 
> > When building I expected this target to support the option "by
> > default". 
> > I've tried building again and I found no relevant flag, even in
> > the 
> > recursive configure help.
> > 
> > Here is what the compiler looks like in more detail:
> > 
> >   Using built-in specs.
> >   COLLECT_GCC=sh4eb-nofpu-elf-gcc
> > 
> > COLLECT_LTO_WRAPPER=/home/el/opt/sh4eb-nofpu-elf-2.32-
> > 8.3.0/libexec/gcc/sh4eb-nofpu-elf/8.3.0/lto-wrapper
> >   Target: sh4eb-nofpu-elf
> >   Configured with: ../gcc-8.3.0/configure 
> > --prefix=/home/el/opt/sh4eb-nofpu-elf-2.32-8.3.0 
> > --target=sh4eb-nofpu-elf --enable-languages=c,c++ --without-
> > headers 
> > --with-newlib --disable-nls --disable-werror --enable-libssp --
> > enable-lto
> >   Thread model: single
> >   gcc version 8.3.0 (GCC)
> > 
> > How can I fix this setup to generate entirely FPU-free programs?
> 
> There is https://gcc.gnu.org/PR64008 which probably is relevant.
> 
> Configuring with --enable-targets=all will help.
> 
> I use this patch for sh, that may help too:
> 
> diff --git a/gcc/config/sh/sh.h b/gcc/config/sh/sh.h
> index 0204872..7f126ce 100644
> --- a/gcc/config/sh/sh.h
> +++ b/gcc/config/sh/sh.h
> @@ -124,6 +124,9 @@ extern int code_for_indirect_jump_scratch;
>  #define SUPPORT_SH2A_NOFPU 1
>  #endif
>  #if SUPPORT_SH3
> +#define SUPPORT_SH4 1
> +#endif
> +#if SUPPORT_SH4
>  #define SUPPORT_SH4_NOFPU 1
>  #endif
>  #if SUPPORT_SH4_NOFPU
> 

Other than that, is "sh4eb-nofpu-elf" actually a valid recognized
triplet?  Maybe it "works" but does effectively nothing? I don't
remember....

In any case, the problem sounds more like a (default) multilib
configuration.  See also the PR link above and the configuration option
--with-multilib-list


Cheers,
Oleg