Re: [PATCH] x86: fix CFI macro invocations to deal with shortcomings in gas

["Alexander van Heukelum" <heukelum@fastmail.fm>]

Hello Jan and all,

On Wed, 20 Oct 2010 07:43 +0100, "Jan Beulich" <JBeulich@novell.com> wrote:
> >>> On 19.10.10 at 23:03, "Alexander van Heukelum" <heukelum@fastmail.fm> wrote:
> >> -       pushl_cfi (TI_sysenter_return-THREAD_SIZE+8+4*4)(%esp)
> >> +       pushl_cfi TI_sysenter_return-THREAD_SIZE_asm+8+4*4(%esp)
> > 
> > This expands to:
> >     pushl_cfi (60)-(8192)+8+4*4(%esp)
> > 
> > I'm sorry to say that Ubuntu 6.06's gas (2.16.91 20060118) still chokes
> > with "too many positional arguments" on this line.
> 
> It escapes me where it would split the obviously single argument, and
> I know I checked all official versions from 2.15 onwards. Could you
> try whether that specific gas would be okay with
> 
>      pushl_cfi ((60)-(8192)+8+4*4)(%esp)

Yes, this works! Adding the parentheses to the patch makes things go
on Ubuntu 6.06 (i386). I hope we can get rid of the old gas, but feel free
to add an Acked-by: Alexander van Heukelum <heukelum@fastmail.fm>.

> I had intentionally removed the surrounding parentheses since
> those are considered by newer gas when determining arguments,
> and thus I could expose eventual problems with older gas even
> on newer versions.
> 
> Further, could you experiment (or ideally debug) where is splits
> the argument. Something like
> 
> .macro m arg1 arg2=0
> 	.long	\arg1, \arg2
> .endm
> 
> .data
> _start:
> 	m	(60)-(8192)+8+4*4

Contents of section .data:
 0000 3c000000 18e0ffff

So it splits at the minus sign, even though there is no space.

Greetings,
    Alexander

> might be handy - assembling with -alm=<filename> should allow
> you to inspect where the argument got split.
> 
> Thanks, Jan
>