Follow @Openwall on Twitter for new release announcements and other news
[<prev] [next>] [<thread-prev] [day] [month] [year] [list]
Message-ID: <20261120201440.GM5716@brightrain.aerifal.cx>
Date: Fri, 20 Nov 2026 15:14:40 -0500
From: Rich Felker <dalias@...c.org>
To: Graeme Harker <graeme.harker@...il.com>
Cc: musl@...ts.openwall.com
Subject: Re: [PATCH] m68k: fix soft-float build of setjmp and longjmp

On Sat, Nov 14, 2026 at 09:14:37PM +0000, Graeme Harker wrote:
> On Fri, 13 Nov 2026 at 18:40, Rich Felker <dalias@...c.org> wrote:
> > Is there an ABI document that covers the softfloat ABI? In particular,
> > when archs have softfloat ABIs it's important to check whether, when
> > running them on a cpu that does have hard float, the floating point
> > registers are all call-clobbered, or whether some are "call-saved if
> > present" like on arm.
> 
> I don't know of an ABI document for it; the SVR4 m68k supplement
> assumes a 68881. What there is, is GCC's behaviour, and there it is
> the former: there is no equivalent of arm's softfp, where code built
> for the soft-float calling convention may still use fpu registers.
> 
> - -msoft-float, -mhard-float and -m68881 all act on one mask bit
>   (HARD_FLOAT in m68k.opt), so a translation unit is one or the other.
> 
> - With hard float off, m68k_conditional_register_usage() in m68k.cc
>   marks every fpu register fixed, so the compiler never uses one.
>   gcc 15.2.0 rejects an explicit 'register double x asm("fp2")' under
>   -msoft-float.
> 
> - The two are already different ABIs, not one ABI with optional
>   registers: with -m68881 a float, double or long double is returned
>   in %fp0, with -msoft-float in %d0/%d1. The comment on
>   m68k_libcall_value() says so: "It also makes -msoft-float code
>   incompatible with hard-float code, so people have to be careful not
>   to mix the two."
> 
> So no soft-float caller can have a live value in fp2-fp7 across
> setjmp, on any cpu, and __HAVE_68881__ is the same condition
> arch/m68k/reloc.h uses to pick ld-musl-m68k-sf.so.1.
> 
> One caveat: nothing in the ELF file records which of the two an
> object uses (e_flags is 0 for both, and ld links them together
> without complaint), so a hard-float object linked against an -sf libc
> is not caught at link time. It is broken in any case, for every
> function that returns a floating-point value, not only for setjmp.

OK, thanks. I'm merging this with the additional factoring of the
.s->.S rename as a separate commit so that the rename can be followed
with exact-hash rather than just heuristics.

Rich

Powered by blists - more mailing lists

Confused about mailing lists and their use? Read about mailing lists on Wikipedia and check out these guidelines on proper formatting of your messages.