[LLVMdev] PR 12207: sign extension of parameters and return values on x86-64

[Duncan Sands]

Hi Dirkjan,

 > I recently spent a bit of time looking into
> http://llvm.org/bugs/show_bug.cgi?id=12207. After talking to a bunch
> of different projects (including libffi and python, which is where I
> actually ran into issues), I found out that it seems that LLVM is in
> the wrong here after all. I noted such on the issue, but haven't
> gotten a response.
>
> There's a recent relevant thread in
> http://gcc.gnu.org/ml/gcc/2013-01/msg00447.html which discusses the
> issues here, which seems to come to the conclusion that LLVM's
> behavior is incorrect. Does the current lack of response on the bug
> signal disagreement with that diagnosis, or just disinterest?
>
> Since this affects stuff like Python (or really lots of software using
> libffi to make calls), it would be great to get this fixed.

I think the reason for the lack of action is that there is a plan to get
rid of zext/sext parameter- and return-value-attributes altogether.  See
http://nondot.org/sabre/LLVMNotes/ExtendedIntegerResults.txt
However in order to implement this scheme, parameter attributes need to
be made more powerful (so you can say things like "sexted from iN" where
N may be 1, 8, 16, ... or even 33).  Maybe they now are powerful enough thanks
to Bill's recent work (hopefully Bill will chime in here).

Ciao, Duncan.