[patch/hppa-linux] MD_FALLBACK_FRAME_STATE_FOR definition

[Andrew Haley]

Randolph Chung writes:
 > > You'll need a PC adjustment for boundary cases.  It's quite possible
 > > that you haven't yet seen a test case that fails.  (Unless hppa leaves
 > > PC pointing at the insn after the one that traps, in which case you
 > > may ignore this advice.)
 > 
 > do you have a specific test case for me? :) 

No, and it's impossible certainly to write one unless you resort to
assembly language.  You need a faulting instruction that's the first
instruction in a EH region.

 > or point me to the bit of
 > code that uses this "adjusted PC"?

Yes.  See below.

 > there are several things:
 > - on palinux, the "PC" (iaoq[0]) has the 2 lower bits set for userspace
 >   addresses, so if the fault occured in 0x10780, the iaoq[0] in the
 >   sigcontext will have 0x10783. (instructions are all 4-bytes wide, so
 >   the next insn will usually be at 0x10784)

This should be fine.  However, relying on this is so very weird that
it might deserve a comment...

 > - pa has a two-level deep "PC", so there's an iaoq[0] and iaoq[1].
 >   iaoq[1] points to the next insn to be executed. in some cases this is
 >   not iaoq[0]+4 (for some branching cases).

Well, I presume your MD_FALLBACK_FRAME_STATE_FOR will do the Right
Thing.  SEGV on most processors leaves PC pointing at the faulting
insn, presumably so that it can be restarted; I dunno about hppa.

 > so... i need to know how this is used to know if/how we need to adjust
 > the PC.

It's here.  Note that, unlesss PC is adjusted, this "context->ra - 1"
may move the address into a previous region.  That's the crux of the
matter: "context->ra - 1" *must* point into an insn that's in your EH
region.

static _Unwind_Reason_Code
uw_frame_state_for (struct _Unwind_Context *context, _Unwind_FrameState *fs)
{
  const struct dwarf_fde *fde;
  const struct dwarf_cie *cie;
  const unsigned char *aug, *insn, *end;

  memset (fs, 0, sizeof (*fs));
  context->args_size = 0;
  context->lsda = 0;

  if (context->ra == 0)
    return _URC_END_OF_STACK;

  fde = _Unwind_Find_FDE (context->ra - 1, &context->bases);
 
Andrew.