[llvm-commits] [PATCH] Use an alternate symbol for function-size calc

[Joerg Sonnenberger]

On Fri, Dec 30, 2011 at 10:01:10AM -0600, Hal Finkel wrote:
> On Fri, 2011-12-30 at 15:50 +0100, Joerg Sonnenberger wrote:
> > On Thu, Dec 29, 2011 at 09:09:55PM -0500, Rafael Ávila de Espíndola wrote:
> > > On 29/12/11 03:20 PM, Hal Finkel wrote:
> > > > This small patch fixes a compatibility problem between the ppc linux asm
> > > > printer and recent versions of binutils (as explained in the test case).
> > > > With this patch, llvm's behavior will again match that of gcc. It
> > > > touches CodeGen/AsmPrinter; please review.
> > > 
> > > Do you know when this was changed in binutils? Maybe we should always
> > > use the local symbol instead of making this a ppc only change.
> > 
> > Or more importantly, why it was changed. There seems to be a plethora of
> > questionable changes in binutils lately.
> 
> To be fair, the current behavior produces assembly like:
> test1:
>   .quad .L.test1,.TOC. at tocbase
>   .previous
> .L.test1:
>   ...
> .Ltmp0:
>   .size       test1, .Ltmp0-test1
> 
> so this is asking gas to produce a size constant by subtracting two
> labels from different (sub)sections. Is that generally legal? Once I
> realized what it was doing, I was surprised that it had ever worked.

OK, this is a bit questionable. What's the purpose of the .previous
again? Could we use the more modern .pushsection/.popsection instead?
They are a bit less nasty in the backend...

I don't think there is a good reason for not just always creating the
local label to avoid the difference.

Joerg