Re: Patch: MIPS support for libjava.

[Andrew Haley <aph at redhat dot com>]

David Daney writes:
 > Andrew Haley wrote:
 > 
 > >Do we have copyright assignment papers from you?
 > >
 > I finally mailed them in last week.  If you don't already have them, you 
 > should have them soon.

Okay.  It can sometimes take a while for the FSF to turn these around.

 > >From what I can see, you jump to the magic code that restores some
 > >registers form the signal frame and then that code jumps to
 > >_Jv_MipselThrowNPE.
 > >
 > >The unwinder sees that the return address points to the middle of
 > >__sig2exdummy and unwinds from there.  The return address of
 > >__sig2exdummy points just after the faulting insn.
 > >
 > That was the general idea.
 > 
 > >
 > >What I can't quite see is how scratch registers that are destroyed by
 > >the signal handler will be restored.  As far as I can see there's
 > >nothing to restore them.  $16..$23 are restored, but not $12..$15 or
 > >$31, and the compiler won't expect these to be trashed by a faulting
 > >memory access.
 > >
 > For some reason I thought that -fnon-call-exceptions might take care of 
 > those.

I don't think that happens on other arches, but I admit that it varies
quite a bit.  The thing to do is copy everything that is saved in the
sigcontext into the unwinder context and the unwinder worry about what
to do with it all.

 > >The official way to do this is to define MD_FALLBACK_FRAME_STATE_FOR
 > >in arch/linux.h.  See gcc/config/rs6000/linux.h for an example.
 > >  
 > >
 > I knew there must be a better way, but did not want to start hacking in 
 > the "core" of gcc.  Perhaps it is time to revisit that decision.

MD_FALLBACK_FRAME_STATE_FOR is a bit tricky, but there's now quite a
few examples for other Linuxes to refer to.  Although the way it works
by recognizing a signal handler frame is rather weird, this is the
best way to do it.  I think you'll find the result is less hairy than
your patch.

Thanks,
Andrew.