chromatic (via RT) wrote:
Several tests fail with the CGP runcore (parrot -C) when multidispatch
re-enters bytecode -- in specific, anything that calls into src/pic.c from
Parrot_pcc_invoke_sub_from_sig_object causes failures.
The problem appears to be that CGP's PIC tries to poke into the bytecode
directly to find get_params and similar opcodes, while parameter-passing
information is in a context after Parrot_pcc_invoke_sub_from_sig_object.
One workaround is to enforce the use of the normal runcore only for calls back
into Parrot from Parrot_pcc_invoke_sub_from_sig_object, but that's just a
workaround. A better solution is to fix the PIC code to look in the
appropriate place in the context for return values. That seems like a good
task for the calling conventions branch.
I've been debating whether src/pic.c should be removed entirely. It's a
swodge of code that only gives a tiny speedup for 4 opcodes: 'infix',
'get_params', 'set_args', and 'set_returns'. And the 'infix' opcode has
been deprecated and will be removed in the calling conventions branch
(it was a nasty hack for the old MMD system). And it's not even really
the right speedup for 'get_params', 'set_args', and 'set_returns'.
There are also a pile of old, nasty MMD functions that should be deleted
and are only called from src/pic.c.
Allison
