Re: [PATCH] PowerPC, add support for printing non-trivial C++ object for the finish command.

[Ulrich Weigand <Ulrich.Weigand@de.ibm.com>]

Hi Carl, hi Tom,

>Carl> The following patch fixes an issue on Powerpc where the function
return
>Carl> value for a C++ program prints garbage when the function return
type is
>Carl> a class object.  The fix is based on an x86-64 patch for the
same
>Carl> issue.
>
>Carl> The patch has been tested on Power 10 with no
regressions.  Please let
>Carl> me know if this patch is acceptable for mainline.   Thanks.
>
>FWIW, here at AdaCore, we carry a very similar patch for
>ppc-linux-tdep.c.  I've appended it.
>
>I haven't really submitted this yet since I wanted to investigate the
>ABI difference that lead to the change.

As to my understanding of the ELFv2 ABI, while r3 contains a pointer
to the struct return buffer *on function entry*, the ABI does *not*
guarantee that this will remain live throughout the whole function,
or be returned in r3.

Of course, in many simple examples the compiler will not bother
allocating anything else to r3, in particular if it stores
the return value shortly before returning.  But it's not guaranteed,
and therefore either of your patches is not really correct IMO.

(A possible always-correct fix would be to use the on-entry value
of r3 if we have the relevant on-entry DWARF info.)

Note that this is actually the same on ppc as on many (most?) other
platform ABIs.  Most test case differences are simply due to the
fact that the test case uses a struct that happens to be returned
in registers on x86_64, but in a buffer on ppc64le.


Bye,
Ulrich