Re: [3.4 patch] Fix PR16314, undefined reference to `__subdf3' forarm-softfloat

[Vladimir Ivanov <vladitx at nucleusys dot com>]

On Tue, 31 Aug 2004, Daniel Jacobowitz wrote:

> On Tue, Aug 31, 2004 at 04:55:15PM +0300, Vladimir Ivanov wrote:
> > Yes, it works, although not exactly in the official gcc 3.4.x (but that's
> > because of other things).
> > Due to using multilibs, when "-mcpu=ep9312" is specified a version of
> > divdf3 is used which returns result in FP register. So, for me the
> > multilib approach solves this problem - I can have arm-linux compiler
> > which produces hard-float, soft-float and crunch (ep9312) code, together
> > with appropriate libgcc.
>
> And where do you get this magical version of divdf3 from?  The version
> normally in ARM libgcc comes from ieee754-df.S, is written in assembly,
> and does not support returning in other registers.

In a patch we did - that's why I wrote "not exactly in the official gcc 
3.4.x". It goes with this modification to the end of lib1funcs.asm:

#ifdef __MAVERICK__
#include "ieee754-df-crunch.S"
#include "ieee754-sf-crunch.S"
#else
#include "ieee754-df.S"
#include "ieee754-sf.S"
#endif

The two new files "ieee754-[s|d]f-crunch.S" take care of returning values
in FP register. Only div[s|d]f3 are present, no need of others for Crunch.

> > What I've been asking was connected to TARGET_LIBGCC2_CFLAGS - is there a
> > way to put specific options per multilib. And also, does someone know how
>
> I don't think there's a mechanism for it.  It seems reasonable to do to
> me, though, so maybe one would be accepted.
>
> > to link against different glibc's based on the multilib options (but
> > that's perhaps slightly [OT]).
>
> MULTI_OSDIRNAMES.

Is this present in 3.4.x? Recursive grep in gcc tree and Google both don't 
show anything.

  -- Vladimir