Front page | perl.perl5.porters |
Postings from May 2016
On Tue, May 03, 2016 at 07:40:22PM +0100, Nicholas Clark wrote:
> On Tue, May 03, 2016 at 04:56:31PM +0100, Dave Mitchell wrote:
> > Rik,
> > I hadn't realised until recently that the Coro breakage caused by
> > consting the MGF vtable array in commit v5.21.7-98-gc910fea, is still
> > unresolved.
> >
> > I propose that we revert this commit for 5.24 (we need another RC anyway
> > for the SvGROW) issue.
>
> What does this gain?
>
> Coro still won't compile against 5.24.x without other changes (due to your
> context stack reworking), and still won't compile against 5.22.x (because this
> wouldn't be backported).
>
> > This should have relatively minimal effect (apart from 4K extra memory
> > usage per process), since all it would be doing is removing a constraint
> > on XS code (the vtable array can now be modified by XS code again).
>
> So as best I can tell, all this achieves is a 4K increase in memory usage
> per process. It doesn't fix Coro. What's the upside?
Apparently it is hard to make Coro work without the vtable being
modifiable. The vtable consting, although desirable, is a relatively
trivial gain for perl, and its loss is relatively small. It just means
that an extra 4K from a 2Mb executable can be shared across processes when
2 or more perls are running on the same system. And allows us to stop
XS code from doing weird stuff with some of our internals.
So it's a partial fix for Coro. I have no idea what use Coro makes of the
context stack, or whether its doing something subtle which is
fundamentally unfixable under the recent context changes in blead; but it
may happen that its easily fixable within Coro. In that case, that makes
it theoretically possible for a Coro to be released that runs on perl
5.24.0 and later. Without the const revert, we may remove this possibility
altogether.
I had a very quick look at the Coro src; the diff below allows it to
compile under blead (with the revert), and it then passes 18 of its 20
test scripts. I have no idea whether the two failing scripts represent a
"nearly there" state or a "fundamentally unfixable" state.
diff --git a/Coro-6.49/Coro/State.xs b/Coro-6.49/Coro/State.xs
index 6d631bd..c2cbe7b 100644
--- a/Coro-6.49/Coro/State.xs
+++ b/Coro-6.49/Coro/State.xs
@@ -1395,7 +1395,18 @@ runops_trace (pTHX)
PUSHMARK (SP);
PUSHs (&PL_sv_yes);
PUSHs (fullname);
- PUSHs (CxHASARGS (cx) ? sv_2mortal (newRV_inc ((SV *)cx->blk_sub.argarray)) : &PL_sv_undef);
+ {
+ AV *aa;
+# if PERL_VERSION_ATLEAST(5,23,8)
+ aa = ((AV*)(AvARRAY(MUTABLE_AV(
+ PadlistARRAY(CvPADLIST(cx->blk_sub.cv))[
+ CvDEPTH(cx->blk_sub.cv)]))[0]));
+#else
+ aa = cx->blk_sub.argarray;
+#endif
+
+ PUSHs (CxHASARGS (cx) ? sv_2mortal (newRV_inc ((SV *)aa)) : &PL_sv_undef);
+ }
PUTBACK;
cb = hv_fetch ((HV *)SvRV (coro_current), "_trace_sub_cb", sizeof ("_trace_sub_cb") - 1, 0);
if (cb) call_sv (*cb, G_KEEPERR | G_EVAL | G_VOID | G_DISCARD);
--
Indomitable in retreat, invincible in advance, insufferable in victory
-- Churchill on Montgomery