Re: IA64 EH

[Jason Merrill <jason at cygnus dot com>]

>>>>> Andrew Macleod <amacleod@cygnus.com> writes:

 >>>   We should make a decision as to whether or not we want to use the
 >>> IA64-style exception handling on other platforms or not.  There's no
 >>> backwards compatibility issue here beacuse we're breaking the ABI in
 >>> the next release anyhow -- and removing support for the old ABI.  So,
 >>> the decision is only: a) does that make sense, and b) how hard is it.
 >>> We won't get a second chance; if we do want to do this, we want to do
 >>> it now.

 > from rth:

 >>> On Fri, May 26, 2000 at 09:21:32AM +0200, Martin v. Loewis wrote:
 >>> > > No.  The IA-64 bits that were checked in only affect unwinding.
 >>> > > The format that Intel chose is extremely IA-64 specific, and has
 >>> > > no real advantage over dwarf2 at all.

 > Yeah, its very ia64 specific. Its more compact in the object, but it
 > would require a new set of unwind opcode definitions tailored for each
 > target, and it is also distasteful (IMO) in that it generates the unwind
 > info in the assembler instead of the compiler. Dwarf2 is more generic,
 > so its seems more preferable for porting activity.

Agreed.  But we do need to make some improvements to the dwarf2 format; in
particular, we should use relative relocs whenever possible to avoid having
to fix them up at runtime.

 >>> > I guess a different way to phrase Mark's question is: 
 >>> > 
 >>> > a) Can the new C++ EH ABI
 >>> >    (http://reality.sgi.com/dehnert_engr/cxx/abi-eh.html) be
 >>> >    implemented on top of the dwarf2 unwinding information?
 >>> 
 >>> Yes, for the exact same reason: dwarf2 only affects the actual
 >>> stack unwinding, which the document you mention assumes by reference
 >>> from the Intel IA-64 manuals.  What's being described there is
 >>> the interface between the program and the unwinding runtime.

 > I believe Jason Merrill has a slightly modified version of that which he
 > was hoping/planning to suggest as the EH ABI. Its quite similar, but
 > with a couple of enhancements.  I'll let him comment on it since I
 > haven't looked at it for a while.

I'm not sure what enhancements you're referring to; if I had any
enhancements to suggest, I'd want to get them into the common ABI.

Jason