-fpic on ppc with SVR4 ABI

-fpic on ppc with SVR4 ABI

[H.J. Lu]

Hi,

When I tried to compile glibc for powerpc using egcs, I got

vfscanf.c: In function `_IO_vfscanf':
vfscanf.c:1166: internal error--unrecognizable insn:
(insn 26472 516 506 (set (reg:SI 8 r8)
        (unspec[ 
                (symbol_ref:SI ("@__ctype_b"))
                (mem:SI (plus:SI (reg:SI 31 r31)
                        (const_int 128)))
            ]  8)) -1 (nil)
    (nil))

when making stdio-common/vfscanf.os with -fpic. -fPIC is fine.
I was wondering if anyone had tried to use egcs to build shared
libraries on powerpc with -fpic and SVR4 ABI.

BTW, I cannot find a small test case. The one I got is several
hundred lines.

Thanks.


-- 
H.J. Lu (hjl@gnu.ai.mit.edu)

Re: -fpic on ppc with SVR4 ABI

[Geoffrey KEATING]

> From: hjl@lucon.org (H.J. Lu)
> Date: Thu, 16 Oct 1997 13:37:11 -0700 (PDT)

> When I tried to compile glibc for powerpc using egcs, I got
> 
> vfscanf.c: In function `_IO_vfscanf':
> vfscanf.c:1166: internal error--unrecognizable insn:
> (insn 26472 516 506 (set (reg:SI 8 r8)
>         (unspec[ 
>                 (symbol_ref:SI ("@__ctype_b"))
>                 (mem:SI (plus:SI (reg:SI 31 r31)
>                         (const_int 128)))
>             ]  8)) -1 (nil)
>     (nil))
> 
> when making stdio-common/vfscanf.os with -fpic. -fPIC is fine.
> I was wondering if anyone had tried to use egcs to build shared
> libraries on powerpc with -fpic and SVR4 ABI.
> 
> BTW, I cannot find a small test case. The one I got is several
> hundred lines.
> 
> Thanks.

I have a medium-sized test case somewhere; this particular bug is
triggered by gcc putting the GOT register in memory, which the
.md file can't cope with.

I have a patch for egcs at
< http://discus.anu.edu.au/~geoffk/egcs-geoffk.diff.gz >, which fixes
this and many other bugs. I have another more extensive patch at home
which fixes even more bugs, most notably a bug involving constant
folding on switch statements.  The first patch was sent to egcs-bugs
and will eventually end up in the distribution; I'm still testing the
second patch.

I've been told that current binutils don't work either (I use my own
extensively patched version which works fine), and there's a typo in
glibc in sysdeps/powerpc/dl-machine.h ('finaladdr' should be
'finalvalue' or something).

--
Geoff Keating <Geoff.Keating@anu.edu.au>