Avoid reifying GVs in rv2cv

\&foo no longer reifies GVs in the stash:

$ ./miniperl -e 'sub foo{} warn $::{foo}; \&foo; warn $::{foo}'
CODE(0x7fab6282ad98) at -e line 1.
CODE(0x7fab6282ad98) at -e line 1.

Sub calls still reify them though, because of the way ck_subr cur-
rently works.

Constant proxies are still upgraded to full GVs for now, just to mini-
mise the churn per patch.

This makes it possible for OP_GVs to point to things other than GVs,
and these things are stored in the pad under threads.  Hence, pad_tidy
could turn on PADTMP, and then IS_PADGV becomes true when it is
upgraded to a glob, so refgen will fail assertions.  There is actually
no need to turn on PADTMP in pad_tidy, since it will already be on for
op targets.  (We need to get rid of IS_PADGV one of these days.  It
doesn’t actually do anything.)

1 files changed, 1 insertions, 3 deletions

diff --git a/pad.c b/pad.c
index 285c2043c3..eb42ab8622 100644
--- a/pad.c
+++ b/pad.c
@@ -1794,9 +1794,7 @@ Perl_pad_tidy(pTHX_ padtidy_type type)
 	    if (!PL_curpad[ix] || SvIMMORTAL(PL_curpad[ix])
 		 || IS_PADGV(PL_curpad[ix]) || IS_PADCONST(PL_curpad[ix]))
 		continue;
-	    if (!SvPADMY(PL_curpad[ix])) {
-		SvPADTMP_on(PL_curpad[ix]);
-	    } else if (!SvFAKE(namep[ix])) {
+	    if (SvPADMY(PL_curpad[ix]) && !SvFAKE(namep[ix])) {
 		/* This is a work around for how the current implementation of
 		   ?{ } blocks in regexps interacts with lexicals.