enterxssub: add CvXSCV bit

Stored in the stash GV for all registered XS methods, to be picked up
by METHOD_CHECK_CACHE, to change the next op to XS.
clear it on redefinition as PP.

testsuite benchmark:
7% faster than without the bit.
overall ~3.5% faster then without enterxssub compile-time optimization.
only faster with some XS calls. Minor slowdown (<1%) without any XS call.

dump.c

@@ -1711,6 +1711,7 @@ const struct flag_to_name gv_flags_names[] = {
     {GVf_MULTI, "MULTI,"},
     {GVf_ASSUMECV, "ASSUMECV,"},
     {GVf_STATIC, "STATIC,"},
+    {GVf_XSCV, "XSCV,"},
 };
 
 const struct flag_to_name gv_flags_imported_names[] = {

gv.h

@@ -176,6 +176,7 @@ Return the CV from the GV.
 #define GVf_IMPORTED_AV	  0x20
 #define GVf_IMPORTED_HV	  0x40
 #define GVf_IMPORTED_CV	  0x80
+#define GVf_XSCV	  0x100
 
 #define GvINTRO(gv)		(GvFLAGS(gv) & GVf_INTRO)
 #define GvINTRO_on(gv)		(GvFLAGS(gv) |= GVf_INTRO)
@@ -211,6 +212,11 @@ Return the CV from the GV.
 #define GvIMPORTED_CV_on(gv)	(GvFLAGS(gv) |= GVf_IMPORTED_CV)
 #define GvIMPORTED_CV_off(gv)	(GvFLAGS(gv) &= ~GVf_IMPORTED_CV)
 
+/* for the cache only */
+#define GvXSCV(gv)		(GvFLAGS(gv) & GVf_XSCV)
+#define GvXSCV_on(gv)		(GvFLAGS(gv) |= GVf_XSCV)
+#define GvXSCV_off(gv)		(GvFLAGS(gv) &= ~GVf_XSCV)
+
 #ifndef PERL_CORE
 #  define GvIN_PAD(gv)		0
 #  define GvIN_PAD_on(gv)	NOOP

op.c

@@ -10220,6 +10220,7 @@ Perl_newXS_len_flags(pTHX_ const char *name, STRLEN len,
             }
         }
 
+        GvXSCV_on(gv);
         CvGV_set(cv, gv);
         if (filename) {
             /* XSUBs can't be perl lang/perl5db.pl debugged
@@ -18620,11 +18621,14 @@ Perl_coresub_op(pTHX_ SV * const coreargssv, const int code,
 
 /*
 =for apidoc report_redefined_cv
+
+If a CV is overwritten, warn by whom
+when use warnings 'redefine' is in effect.
 =cut
 */
 void
 Perl_report_redefined_cv(pTHX_ const SV *name, const CV *old_cv,
-			       SV * const *new_const_svp)
+                         SV * const *new_const_svp)
 {
     const char *hvname;
     bool is_const = !!CvCONST(old_cv);
@@ -18677,10 +18681,10 @@ Perl_report_redefined_cv(pTHX_ const SV *name, const CV *old_cv,
         } else {
         no_caller:
             Perl_warner(aTHX_ packWARN(WARN_REDEFINE),
-			  is_const
-			    ? "Constant subroutine %" SVf " redefined"
-			    : "Subroutine %" SVf " redefined",
-                          SVfARG(name));
+                        is_const
+                          ? "Constant subroutine %" SVf " redefined"
+                          : "Subroutine %" SVf " redefined",
+                        SVfARG(name));
         }
     }
 }

pod/perlcdelta.pod

@@ -110,16 +110,52 @@ as an updated module in the L</Modules and Pragmata> section.
 
 =head1 Performance Enhancements
 
-XXX Changes which enhance performance without changing behaviour go here.
-There may well be none in a stable release.
-
-[ List each enhancement as a =item entry ]
-
 =over 4
 
 =item *
 
-XXX
+XS calls were made faster again. Previously cperl (with 5.22.1c) fell
+from pure perl calls to XS calls and changed the call type to
+ENTERXSSUB for faster subsequent invocations. This made bigger
+examples ~10% faster.  Now the compiler detects calls to XS code and
+changes the type to ENTERXSSUB.  Also the method cache code checks now
+for an XS call and changes the type of the next OP. Overall the
+improvement is ~3.5%, but it depends on the frequency of XS
+calls. It's 0.5-1% slower without any XS call.  Note that p5p decided
+to delete this feature request from it's TODO list.
+
+The optree is a bit slimmer.
+Before:
+
+    $ cperl5.27.0d-nt -Dt -e'use strict'
+    (-e:1)	nextstate
+    (-e:1)	const(PV("strict.pm"))
+    (-e:1)	require
+    (-e:1)	nextstate
+    (-e:1)	pushmark
+    (-e:1)	const(PV("strict"))
+    (-e:1)	method_named(->import)
+    (-e:1)	entersub(strict::import)
+    (-e:1)	enterxssub(strict::import)
+    (-e:1)	leavesub
+
+Now:
+
+    $ cperl5.27.1d-nt -Dt -e'use strict'
+    (-e:1)	nextstate
+    (-e:1)	const(PV("strict.pm"))
+    (-e:1)	require
+    (-e:1)	nextstate
+    (-e:1)	pushmark
+    (-e:1)	const(PV("strict"))
+    (-e:1)	method_named(->import)
+    (-e:1)	enterxssub(strict::import)
+    (-e:1)	leavesub
+
+Note that this is still not optimal, as strict::import can be
+compile-time optimized from a dynamic method call to a static function
+call. A regression from 5.22.1c.
+See L<http://blogs.perl.org/users/rurban/2011/06/how-perl-calls-subs-and-methods.html>
 
 =back
 

pp_hot.c

@@ -5429,6 +5429,7 @@ PP(pp_method)
     RETURN;
 }
 
+/* collapsed the XS check into one bit in the cached gv: GvXSCV */
 #define METHOD_CHECK_CACHE(stash,cache,meth) 				\
     if (SvREADONLY(stash) && !hv_exists_ent(stash, meth, 0)) { ;        \
     } else {                                                            \
@@ -5439,7 +5440,7 @@ PP(pp_method)
                 == (PL_sub_generation + HvMROMETA(stash)->cache_gen)))  \
             {                                                           \
                 CV *cv = GvCV(gv);                                      \
-                if (CvISXSUB(cv) && !CvCONST(cv) && CvROOT(cv) && !PL_perldb \
+                if (GvXSCV(gv) && !PL_perldb                            \
                     && OpTYPE(PL_op->op_next) == OP_ENTERSUB)           \
                 {                                                       \
                     DEBUG_k(Perl_deb(aTHX_ "method_named -> xs %" SVf "\n",\

sv.c

@@ -4129,6 +4129,8 @@ Perl_gv_setref(pTHX_ SV *const dstr, SV *const sstr)
                         ? cv_const_sv((const CV *)sref)
                         : NULL;
                     HV * const stash = GvSTASH((const GV *)dstr);
+                    if (GvXSCV(dstr) && (const_sv || !CvISXSUB(cv)))
+                        GvXSCV_off(dstr);
                     report_redefined_cv(
                             sv_2mortal(
                                    stash