Linux/PPC gcc 2.95.2 - bad access generated when using inline func

[Geoff Keating]

> From: Franz Sirl <Franz.Sirl-kernel@lauterbach.com>
> Date: Wed, 28 Jun 2000 20:57:21 +0200
> Cc: gcc-bugs@gcc.gnu.org
> 
> On Wed, 28 Jun 2000, Geoff Keating wrote:
> > I get the following code for rtl_soft_cli with -O2 and gcc 2.95.2:
> >
> > rtl_soft_cli:
> >         lis 10,rtl_local@ha
> >         lwz 0,rtl_local@l(10)
> >         la 8,rtl_local@l(10)
> >         andi. 9,0,2
> >         bc 12,2,.L1240
> >         lwz 9,-16(1)
> >         lis 11,last_cli@ha
> >         lwz 0,4(9)
> >         stw 0,last_cli@l(11)
> > .L1240:
> >         li 0,2
> >         1:      lwarx   9,0,8
> >         andc    9,9,0
> >         stwcx.  9,0,8
> >         bne     1b
> >         blr
> >
> > I don't see that register 8 has the wrong value.  It looks like it
> > contains the address of the global variable rtl_local.  Could you be
> > more specific about what compiler flags you used, what assembler
> > output you got, and what is wong with it?
> 
> Geoff, it seems the problem is using __builtin_return_address in leaf 
> functions. __builtin_return_address currently returns wrong values if the 
> current function has no stackframe. See this comment in rs6000.h:
> 
> /* Definitions for __builtin_return_address and __builtin_frame_address.
>    __builtin_return_address (0) should give link register (65), enable
>    this. */
> /* This should be uncommented, so that the link register is used, but
>    currently this would result in unmatched insns and spilling fixed
>    registers so we'll leave it for another day.  When these problems are
>    taken care of one additional fetch will be necessary in RETURN_ADDR_RTX.
>    (mrs) */
> /* #define RETURN_ADDR_IN_PREVIOUS_FRAME */
> 
> I uncommented this, but now I get:
> 
> test5.c:10: Internal compiler error:
> test5.c:10: internal error--unrecognizable insn:
> 
> (insn 10 8 0 (set (mem/f:SI (reg:SI 82) 0)
>         (reg:SI 65 lr)) -1 (insn_list 8 (nil))
>     (expr_list:REG_DEAD (reg:SI 65 lr)
>         (expr_list:REG_DEAD (reg:SI 82)
>             (nil))))
> 
> I'm currently playing with rs6000.md, but with my knowledge in there a 
> solution will take a while... I would appreciate some hints :-).

Actually, __builtin_return_address (0) should not give the link
register, because sometimes the link register holds something else,
like the return address of something called by this function.

My suspicion is that the right thing to do is force all functions that
call __builtin_return_address to have a stack frame...  I will think
about it.

-- 
- Geoffrey Keating <geoffk@cygnus.com>