Re: gas 2.16 and assembly macros -- entry_64.S build failure

[Jan Beulich]

>>> On 01.10.10 at 02:26, "H. Peter Anvin" <hpa@xxxxxxxxx> wrote:
> On 09/16/2010 04:21 AM, Jan Beulich wrote:
>>>
>>> Yes, to only generate CFI on binutils that allows us to write sane 
>>> looking code.
>>>
>>> I.e. to disable CONFIG_AS_CFI on binutils that are broken for this.
>> 
>> Again - that won't help, as it's the macro invocation that gas
>> fails one, not one of the actual .cfi_* directives.
>> 
> 
> Looking again at this issue with the binutils version from hell
> (sigh)... I'm running quickly out of ideas.
> 
> The problem is that cpp inserts spaces around expansions, so:
> 
> 	pushq_cfi $(USER_DS)
> 
> ... turns into something like ...
> 
> 	pushq_cfi $( ( 5 * 8 + 3 ) )
> 
> ... which these old versions of gas considers multiple arguments to the
> macro, even though there is no comma anywhere.  We can defang *some* of
> these problems by using cpp macros to kill them off:
> 
> #define pushq_cfi pushq
> 
> ... but that doesn't work with the macros like movq_cfi.  On those, we
> could argue that at least people won't put $ on them, but cpp will still
> split them apart with spaces; this apparently causes problems at least
> as soon as there is an expression more complicated than addition
> involved (apparently plus signs are okay, but minus signs aren't!)
> 
> I'm completely lost about how to deal with this.   We can't simply
> defang the macros -- at least not in a way that is likely to *stay*
> working -- and dropping the macros is seriously going to impact the
> debuggability of the kernel.  One way, of course, is to simply declare
> binutils 2.16 and 2.15.9x (which is apparently included in
> RHEL/CentOS 4) to be broken beyond repair unless distros backport a fix,
> and in many ways I think that is the preferred option, but I don't know
> if that makes sense to others...

Would excessive parenthesisation (in the header files) be an
acceptable workaround? cpp inserts the spaces only when
a preprocessing identifier that expands to a token sequence
ending in a number is followed by ., +, or -. (The asymmetry
in behavior between + and - then results from gas considering
- a symbol char, but not +, which is the case even in current
mainline, and which I think ought to be fixed - I'll bring this
up on their mailing list -, but that fix would break the kernel
build in its current shape afaict.)

Basically, all constants potentially used in assembly expressions
passed to macros would need to change from e.g.

#define RIP		128

to

#define RIP		(128)

I'll put together a patch if this is considered acceptable, and if
it turns out to work across all cpp/gas version combinations
that I can reasonably try.

Jan

--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/