Re: Structure-return bug on powerpc32

[Wim Lewis <wiml at hhhh dot org>]

On May 22, 2009, at 12:55 PM, Andreas Tobler wrote:
> Wim Lewis wrote:
> > I've been looking into some libffi test failures on netbsd/ppc. To  
> > make a long story short, under some circumstances (returning a  
> > small aggregate that's not a multiple of 4 bytes), ffi_call() will  
> > write past the end of the buffer passed to rvalue.  [....]
> Which ABI does netbsd/ppc use? The original SYSV or the same as  
> Linux? If it is the original SYSV you might try to build like  
> FreeBSD does.

NetBSD uses the same ABI as FreeBSD, as far as I know --- in fact,  
it's building libffi with TARGET=POWERPC_FREEBSD. :) The other tests  
which depend on the difference between the SYSV and "GCC" ABIs (small  
aggregates returned in registers vs. a hidden pointer argument) all  
pass. And the actual returned value is correct --- it's just stomping  
on some nearby memory as well.

The bug is not with the procedure call convention itself, but with  
the way that the return value is copied from registers into the  
buffer passed to ffi_call(). Once I found the relevant code in  
powerpc/sysv.S, it was pretty clear that it will write past the end  
of its buffer in situations like this. I'm surprised that the bug  
doesn't occur on FreeBSD as well, but perhaps there's some subtle  
point I'm missing.

I've written up a test case which should detect the problem even if  
the compiler doesn't happen to put something important immediately  
after res_dbl. It fails in the expected way on my netbsd5/ppc  
machine, and passes on the other systems I have handy (openbsd4.4/ 
i386, darwin9/i386, darwin9/x86_64). It also passes on darwin8/ppc,  
apparently because the darwin calling convention is to return even  
small aggregates using a hidden pointer argument.

Here's the test program:
    http://www.hhhh.org/wiml/tmp/odd_struct.c

It'd be interesting to run it on an architecture with more strict  
alignment requirements as well, like MIPS or SPARC.


regards
   Wim Lewis / wiml@hhhh.org