In Perl_mro_isa_changed_in(), no need to call an RVALUE hv_fetch() then

hv_store() a new HV if fetch drew a blank, as we can detect the new SV
an LVALUE fetch gave us, and replace it with a new HV, which will be
more efficient.

p4raw-id: //depot/perl@32168

1 files changed, 15 insertions, 4 deletions

diff --git a/mro.c b/mro.c
index aca1ee8abe..95e0ef1d04 100644
--- a/mro.c
+++ b/mro.c
@@ -550,12 +550,23 @@ Perl_mro_isa_changed_in(pTHX_ HV* stash)
         SV* const sv = *svp++;
         HV* mroisarev;
 
-        HE *he = hv_fetch_ent(PL_isarev, sv, 0, 0);
-        if(!he) {
-            he = hv_store_ent(PL_isarev, sv, (SV*)newHV(), 0);
-        }
+        HE *he = hv_fetch_ent(PL_isarev, sv, TRUE, 0);
+
+	/* That fetch should not fail.  But if it had to create a new SV for
+	   us, then we can detect it, because it will not be the correct type.
+	   Probably faster and cleaner for us to free that scalar [very little
+	   code actually executed to free it] and create a new HV than to
+	   copy&paste [SIN!] the code from newHV() to allow us to upgrade the
+	   new SV from SVt_NULL.  */
+
         mroisarev = (HV*)HeVAL(he);
 
+	if(SvTYPE(mroisarev) != SVt_PVHV) {
+	    SvREFCNT_dec(mroisarev);
+	    mroisarev = newHV();    
+	    HeVAL(he) = (SV *) mroisarev; 
+        }
+
 	/* This hash only ever contains PL_sv_yes. Storing it over itself is
 	   almost as cheap as calling hv_exists, so on aggregate we expect to
 	   save time by not making two calls to the common HV code for the