COWed stack bug (was: IMCC - PerlArray getting trounced)

[Leopold Toetsch]

Will Coleda wrote:

> Well, that was festive. "I can reproduce that bug in 22 lines!"

Great.

The bug (and other reported curruptions) are definitely coming from the
COW logic in register.c. This is what happens:

Setting up the exception handler (which is a continuation) triggers COW
setting of the register stacks. Then on first subroutine return from
__inner, the register stack is unCOWed (the chunk is copied). When then
returning from __outer stack->top->used is still 2 end the same chunk
gets popped off the stack, because the interpreter's stack is still
unchanged.

The register restoring memcpy is the same *twice*:

memcpy (dstpp=0x824ca50, srcpp=0x4017d9a0, len=64)

So returning from __outer places P17 (the RetContinuation of __inner) in
mains registers and ruins the PerlUndef that you wanted to print.

> :get_string() not implemented in class 'RetContinuation'

... which then causes this error.

Just returning from mark_register_stack_cow() makes the bug vanish

(but doesn't fix anything)

As mentioned several times, a COWed buffer needs distinct buffer headers
and shared buffer memory. The current implementation in register *and*
other stacks is broken. *Fixes welcome*.

You can currently avoid the bug by not using Continuations and Exception
Handlers.

Thanks for your test program,
leo