Exception handling support for a target

From: Tim Northover via llvm-dev 

Emitting directives in the epilogue is hard because the directives
apply to all instructions after in program-counter order. So if you
have an epilogue in the middle of a function and emit CFI directives
saying the registers are back where they were then the unwinder will
think that applies to the rest of the function too.

To fix that you'd have to emit yet more directives in the basic block
immeditately following the epilogue. Most people don't bother with
either because you'd only really need it if you expected to have to
catch an exception in the epilogue (which is impossible on most

Also, the
directives are also for debug purpose (.debug_frame), right? I guess I only
have to make sure directives work for exception handling, then debug works
as well?

Yep. Technically you can stop the debugger in the middle of a prologue
or epilogue, at which point having more fine-grained directives can
help. In practice there's no real reason to do that so people don't
emit directives for it (at least not as a high priority).



OpenVMS’ EH model is for full asynchronous prologue/epilogue EH. We do more
than just program-level EH, but have to deal with OS events (timers going off,
asynch I/O completion, mailboxes filled with data, etc.) which could result in
an unwind occurring.

We’re just about finished with our proposal for extending the x86 compact EH data
format to cover all the cases plus the desire to create the additional .cfi directives.
Feedback and pointers will be most helpful. Look for this as a new thread sometime
early next week (we’re having our final review on Monday before I post it).


Are you planning on extending LLVM to add support for non-call exceptions? Currently, LLVM is not able to produce correct unwind tables for arbitrary exception points because it does not split basic blocks in such a way that values that are created in a basic block will be visible in at the end. For example, if I do

int i = 0;
try {
} catch (…) {
  return i;

The SSA form of this will look something more like this:

int i0 = 0;
try {
  i1 = 1;
  i2 = 2;
} catch (…) {
  return 1;

But if you are allowed to throw exceptions on any instruction, then you end up with something more like:

int i0 = 0;
try {
  i1 = 1;
  i2 = 2;
} catch (…) {
  return phi({a,0},{b, i1},{c,i2});

With the current structure of LLVM’s exception handling, the fact that i may be incremented after the only invoke instruction is not visible.

I’ve never been entirely happy with how LLVM models exceptions, as calls with two possible returns, but extending that model to properly support unwinds through arbitrary instructions will leave us in the degenerate case of having single-instruction basic blocks. I’d much rather that we split blocks on places where values visible in an exception handler change, rather than on specific places where exceptions are known to be thrown, but this will require moving to a region-based representation of exceptions, which is a lot of work. Without this, it’s easy to unwind from arbitrary exceptions, unless you care about having the right values in your handlers.


I think it’s valuable to have precise / asynchronous unwind information without adding support for non-call exceptions. Profilers and debuggers need to be able to unwind the stack from arbitrary instruction boundaries, but they don’t need to run exception handlers. We can still declare that outside the model. Speaking of which, barring bugs, we do support precise unwind info on Win64.

Even if we were to support non-call exceptions, we would enumerate the set of supported potentially trapping operations that we support, i.e. load, store, div, FP ops, etc. The edges would either be explicit, as they are today, or implicit. We already have implicit unwind edges out of the function from non-nounwind calls, so this isn’t as crazy as it sounds.

We would definitely say that ADDs, PHIs, and other things don’t trap, so in your example it wouldn’t matter if the value of ‘i’ was 1 or 0 if the exception gets raised at ‘b’. There would be no control-dependent, possibly-unwinding instruction, either unwinding to the caller or choosing to run the catch handler would be correct. It’s equivalent to the asynchronous exception arriving a little sooner or a little later, which the outside observer has no way to control.

OpenVMS’ non-call exception support on Alpha & Itanium does have some restrictions. We’ll do some study and attempt to address the concerns when I post the doc next week.

However, our primary focus is precise exception modelling in the prologue/epilogue of routines. You need precise info on the register saves, stack pointer save/restore, etc. that the unwinder can use. On OpenVMS, we don’t invoke a routine’s personality routine if we’re in the prologue but the prologue isn’t finished. We go up a frame and start searching there. This level of support is needed in our system code (C and BLISS) as well in customer code in Fortran, Pascal, BASIC, COBOL, etc.