Re: mips .bss switching patch

[Aldy Hernandez <aldyh at redhat dot com>]

> > +   else if (! strncmp (NAME, ".bss", 4))				\
> > +     fprintf (F, "\t.section %s,\"aw\",@nobits\n", (NAME));	\
> 
> But you don't know anything about this section except that
> it starts with ".bss".  Bad Juju.

damn!  I actually copied what the arm backend was doing.  perhaps a
strcmp instead of a strncmp? ;-).  actually, i think a couple othe back
ends were doing this strncmp business for the .bss section-- not that that
makes it right.

> 
> int x __attribute__((section(".bssiod"))) = 2;
> 
> You should never make choices like this based on name.  I know
> it happens a lot, but that doesn't make it right.
> 
> Why did this come up?

a customer reported that they were getting the following error message:

	a.s:5: Warning: Ignoring incorrect section type for .bss

for the following C code:

	int a __attribute__ ((aligned(16))) __attribute__ ((section(".bss")));


> I can't think how to solve this properly.  The current documentation
> on section attributes is out of date -- we've supported them on
> uninitialized data for a while.  I'm of a mind that 
> 
>   int x __attribute__((section(".foo")));
>   int y __attribute__((section(".foo"))) = 2;
> 
> should continue to work, which means knowing _not_ to mark the 
> section @nobits before you've seen _all_ of the decls.  Which you
> can't do at the moment.

uggghhh,  what should we do?  nick had suggested that the user compile
with -fno-common to inhibit putting the "a" in .comm.

suggestions?
aldy