* Specific problems in IA64 make check
** test-unwind
Representation of the relevant dynamic context:
non-rewindable
catch frame make cont.
o----o-----a----------b-------------c
\
\ call cont.
o-----o-----------d
A continuation is captured at (c), with a non-rewindable frame in the
dynamic context at (b). If a rewind through that frame was attempted,
Guile would throw to the catch at (a). Then the context unwinds back
past (a), then winds forwards again, and the captured continuation is
called at (d).
We should end up at the catch at (a). On ia64, we get an "illegal
instruction".
The problem is that Guile does not restore the ia64 register backing
store (RBS) stack (which is saved off when the continuation is
captured) until all the unwinding and rewinding is done. Therefore,
when the rewind code (scm_i_dowinds) hits the non-rewindable frame at
(b), the RBS stack hasn't yet been restored. The throw finds the
jmp_buf (for the catch at (a)) correctly from the dynamic context, and
jumps back to (a), but the RBS stack is invalid, hence the illegal
instruction.
This could be fixed by restoring the RBS stack earlier, at the same
point (copy_stack) where the normal stack is restored. But that
causes a problem in the next test...
** continuations.test
The dynamic context diagram for this case is similar:
non-rewindable
catch frame make cont.
a----x-----o----------b-------------c
\
\ call cont.
o-------d
The only significant difference is that the catch point (a) is
upstream of where the dynamic context forks. This means that the RBS
stack at (d) already contains the correct RBS contents for throwing
back to (a), so it doesn't matter whether the RBS stack that was saved
off with the continuation gets restored.
This test passes with the Guile 1.8.4 code, but fails (with an
"illegal instruction") when the code is changed to restore the RBS
stack earlier as described above.
The problem now is that the RBS stack is being restored _too_ early;
specifically when there is still stuff to do that relies on the old
RBS contents. When a continuation is called, the sequence of relevant
events is:
(1) Grow the (normal) stack until it is bigger than the (normal)
stack saved off in the continuation. (scm_dynthrow, grow_stack)
(2) scm_i_dowinds calls itself recursively, such that
(2.1) for each rewind (from (x) to (c)) that will be needed,
another frame is added to the stack (both normal and RBS),
with local variables specifying the required rewind; the
rewinds don't actually happen yet, they will happen when
the stack unwinds again through these frames
(2.2) required unwinds - back from where the continuation was
called (d) to the fork point (x) - are done immediately.
(3) The normal (i.e. non-RBS) stack that was stored in the
continuation is restored (i.e. copied on top of the actual
stack).
Note that this doesn't overwrite the frames that were added in
(2.1), because the growth in (1) ensures that the added frames
are beyond the end of the restored stack.
(4) ? Restore the RBS stack here too ?
(5) Return (from copy_stack) through the (2.1) frames, which means
that the rewinds now happen.
(6) setcontext (or longjmp) to the context (c) where the
continuation was captured.
The trouble is that step (1) does not create space in the RBS stack in
the same kind of way that it does for the normal stack. Therefore, if
the saved (in the continuation) RBS stack is big enough, it can
overwrite the RBS of the (2.1) frames that still need to complete.
This causes an illegal instruction when we return through those frames
and try to perform the rewinds.
* Fix
The key to the fix is that the saved RBS stack only needs to be
restored at some point before the next setcontext call, and that doing
it as close to the setcontext call as possible will avoid bad
interactions with the pre-setcontext stack. Therefore we do the
restoration at the last possible point, immediately before the next
setcontext call.
The situation is complicated by there being two ways that the next
setcontext call can happen.
- If the unwinding and rewinding is all successful, the next
setcontext will be the one from step (6) above. This is the
"normal" continuation invocation case.
- If one of the rewinds throws an error, the next setcontext will
come from the throw implementation code. (And the one in step (6)
will never happen.) This is the rewind error case.
In the rewind error case, the code calling setcontext knows nothing
about the continuation. So to cover both cases, we:
- copy (in step (4) above) the address and length of the
continuation's saved RBS stack to the current thread state
(SCM_I_CURRENT_THREAD)
- modify all setcontext callers so that they check the current
thread state for a saved RBS stack, and restore it if so before
calling setcontext.
* Notes
** I think rewinders cannot rely on using any stack data
Unless it can be guaranteed that the data won't go into a register.
I'm not 100% sure about this, but I think it follows from the fact
that the RBS stack is not restored until after the rewinds have
happened.
Note that this isn't a regression caused by the current fix. In Guile
1.8.4, the RBS stack was restored _after_ the rewinds, and this is
still the case now.
** Most setcontext calls for `throw' don't need to change the RBS stack
In the absence of continuation invocation, the setcontext call in the
throw implementation code always sets context to a place higher up the
same stack (both normal and RBS), hence no stack restoration is
needed.
* Other changes
** Using setcontext for all non-local jumps (for __ia64__)
Along the way, I read a claim somewhere that setcontext was more
reliable than longjmp, in cases where the stack has been manipulated.
I don't now have any reason to believe this, but it seems reasonable
anyway to leave the __ia64__ code using getcontext/setcontext, instead
of setjmp/longjmp.
(I think the only possible argument against this would be performance -
if getcontext was significantly slower than setjmp. It that proves to
be the case, we should revisit this.)
** Capping RBS base for non-main threads
Somewhere else along the way, I hit a problem in GC, involving the RBS
stack of a non-main thread. The problem was, in
SCM_MARK_BACKING_STORE, that scm_ia64_register_backing_store_base was
returning a value that was massively greater than the value of
scm_ia64_ar_bsp, leading to a seg fault. This is because the
implementation of scm_ia64_register_backing_store_base is only valid
for the main thread. I couldn't find a neat way of getting the true
RBS base of a non-main thread, but one idea is simply to call
scm_ia64_ar_bsp when guilifying a thread, and use the value returned
as an upper bound for that thread's RBS base. (Note that the RBS
stack grows upwards.)
(Were it not for scm_init_guile, we could be much more definitive
about this. We could take the value of scm_ia64_ar_bsp as a
definitive base address for the part of the RBS stack that Guile cares
about. We could also then discard
scm_ia64_register_backing_store_base.)
* libguile/struct.c (scm_struct_ref, scm_struct_set_x): "Light" structs
have no hidden words (members of the SCM_STRUCT_DATA(x) array accessed
with negative indices). In that case, determine the number of fields
from the length of the struct layout descriptor. (Most GOOPS instances
are light structs.)
* libguile/goops.c (get_slot_value, set_slot_value): In the struct
allocation case, don't poke the slots array directly -- we should
go through struct-ref/struct-set! code so that we get the
permissions and allocation ('u' versus 'p') correct.
