Re: [PATCH] MIPS gas: Reallow branches to global file-internal labels

[Thiemo Seufer <ica2_ts at csv dot ica dot uni-stuttgart dot de>]

Alexandre Oliva wrote:
> On Feb 12, 2003, Thiemo Seufer <ica2_ts@csv.ica.uni-stuttgart.de> wrote:
> 
> > This seems to be intentional, as the testcase was adjusted accordingly.
> 
> Yes, it is intentional.  Global symbols are supposed to be overridable
> according to ELF library semantics.  Resolving them locally, like the
> IRIX assembler does, looks like a bug to me.  If we can't represent
> it, we can't, live it with.

But not all objects have to be (conforming) ELF libraries.

> > However, this breaks much code, especially performance-critical
> > hand-coded one.
> 
> Hand-coded assembly can always be adjusted so as to use local labels,
> such that only local labels are used as branch targets.

It loses performance and enlarges the code, because conditional branches
have to be converted to branch-and-jump sequences.

> > To my knowledge, it affects at least the Linux
> > Kernel, glibc and dietlibc.
> 
> Kernel is fixed.  Glibc fixes are still awaiting approval.  I haven't
> looked at dietlibc.

IMHO the assembler should reallow branches to globals instead, at least
for non-ELF-libraries.


Thiemo