diff --git a/perl/README b/perl/README
index 2ef686c2..d5fc2693 100644
--- a/perl/README
+++ b/perl/README
@@ -8,10 +8,29 @@ SYNOPSIS
 DESCRIPTION
     This module converts Perl data structures to MessagePack and vice versa.
 
+ABOUT MESSAGEPACK FORMAT
     MessagePack is a binary-based efficient object serialization format. It
     enables to exchange structured objects between many languages like JSON.
     But unlike JSON, it is very fast and small.
 
+  ADVANTAGES
+    PORTABILITY
+        Messagepack is language independent binary serialize format.
+
+    SMALL SIZE
+            say length(JSON::XS::encode_json({a=>1, b=>2}));   # => 13
+            say length(Storable::nfreeze({a=>1, b=>2}));       # => 21
+            say length(Data::MessagePack->pack({a=>1, b=>2})); # => 7
+
+        MessagePack format saves memory than JSON and Storable format.
+
+    STREAMING DESERIALIZER
+        MessagePack supports streaming deserializer. It is useful for
+        networking such as RPC.
+
+    If you want to get more informations about messagepack format, please
+    visit to <http://msgpack.org/>.
+
 METHODS
     my $packed = Data::MessagePack->pack($data);
         pack the $data to messagepack format string.
@@ -23,6 +42,28 @@ Configuration Variables
     $Data::MessagePack::PreferInteger
         Pack the string as int when the value looks like int(EXPERIMENTAL).
 
+SPEED
+    This is result of benchmark/serialize.pl and benchmark/deserialize.pl on
+    my SC440(Linux 2.6.32-23-server #37-Ubuntu SMP).
+
+        -- serialize
+        JSON::XS: 2.3
+        Data::MessagePack: 0.20
+        Storable: 2.21
+        Benchmark: timing 1000000 iterations of json, mp, storable...
+              json:  5 wallclock secs ( 3.95 usr +  0.00 sys =  3.95 CPU) @ 253164.56/s (n=1000000)
+                mp:  3 wallclock secs ( 2.69 usr +  0.00 sys =  2.69 CPU) @ 371747.21/s (n=1000000)
+          storable: 26 wallclock secs (27.21 usr +  0.00 sys = 27.21 CPU) @ 36751.19/s (n=1000000)
+
+        -- deserialize
+        JSON::XS: 2.3
+        Data::MessagePack: 0.20
+        Storable: 2.21
+        Benchmark: timing 1000000 iterations of json, mp, storable...
+              json:  4 wallclock secs ( 4.45 usr +  0.00 sys =  4.45 CPU) @ 224719.10/s (n=1000000)
+                mp:  6 wallclock secs ( 5.45 usr +  0.00 sys =  5.45 CPU) @ 183486.24/s (n=1000000)
+          storable:  7 wallclock secs ( 7.77 usr +  0.00 sys =  7.77 CPU) @ 128700.13/s (n=1000000)
+
 AUTHORS
     Tokuhiro Matsuno
 
@@ -40,5 +81,5 @@ LICENSE
     under the same terms as Perl itself.
 
 SEE ALSO
-    <http://msgpack.org/>
+    <http://msgpack.org/> is official web site for MessagePack format.
 
diff --git a/perl/perlxs.h b/perl/perlxs.h
new file mode 100644
index 00000000..441682de
--- /dev/null
+++ b/perl/perlxs.h
@@ -0,0 +1,76 @@
+/*
+    perlxs.h - Standard XS header file
+    Copyright (c) Fuji, Goro (gfx)
+*/
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+#define PERL_NO_GET_CONTEXT /* we want efficiency */
+#include <EXTERN.h>
+
+#include <perl.h>
+#define NO_XSLOCKS /* for exceptions */
+#include <XSUB.h>
+
+#ifdef __cplusplus
+} /* extern "C" */
+#endif
+
+#include "ppport.h"
+
+/* portability stuff not supported by ppport.h yet */
+
+#ifndef STATIC_INLINE /* from 5.13.4 */
+#   if defined(__GNUC__) || defined(__cplusplus__) || (defined(__STDC_VERSION__) && (__STDC_VERSION__ >= 199901L))
+#       define STATIC_INLINE static inline
+#   else
+#       define STATIC_INLINE static
+#   endif
+#endif /* STATIC_INLINE */
+
+#ifndef __attribute__format__
+#define __attribute__format__(a,b,c) /* nothing */
+#endif
+
+#ifndef LIKELY /* they are just a compiler's hint */
+#define LIKELY(x)   (x)
+#define UNLIKELY(x) (x)
+#endif
+
+#ifndef newSVpvs_share
+#define newSVpvs_share(s) Perl_newSVpvn_share(aTHX_ STR_WITH_LEN(s), 0U)
+#endif
+
+#ifndef get_cvs
+#define get_cvs(name, flags) get_cv(name, flags)
+#endif
+
+#ifndef GvNAME_get
+#define GvNAME_get GvNAME
+#endif
+#ifndef GvNAMELEN_get
+#define GvNAMELEN_get GvNAMELEN
+#endif
+
+#ifndef CvGV_set
+#define CvGV_set(cv, gv) (CvGV(cv) = (gv))
+#endif
+
+/* general utility */
+
+#if PERL_BCDVERSION >= 0x5008005
+#define LooksLikeNumber(x) looks_like_number(x)
+#else
+#define LooksLikeNumber(x) (SvPOKp(x) ? looks_like_number(x) : (I32)SvNIOKp(x))
+#endif
+
+#define newAV_mortal() (AV*)sv_2mortal((SV*)newAV())
+#define newHV_mortal() (HV*)sv_2mortal((SV*)newHV())
+
+#define DECL_BOOT(name) EXTERN_C XS(CAT2(boot_, name))
+#define CALL_BOOT(name) STMT_START {            \
+        PUSHMARK(SP);                           \
+        CALL_FPTR(CAT2(boot_, name))(aTHX_ cv); \
+    } STMT_END
diff --git a/perl/xs-src/pack.c b/perl/xs-src/pack.c
index 93b2e2f4..62eb0024 100644
--- a/perl/xs-src/pack.c
+++ b/perl/xs-src/pack.c
@@ -2,16 +2,7 @@
  * code is written by tokuhirom.
  * buffer alocation technique is taken from JSON::XS. thanks to mlehmann.
  */
-#ifdef __cplusplus
-extern "C" {
-#endif
-#include "EXTERN.h"
-#include "perl.h"
-#include "XSUB.h"
-#include "ppport.h"
-#ifdef __cplusplus
-};
-#endif
+#include "perlxs.h"
 
 #include "msgpack/pack_define.h"
 
diff --git a/perl/xs-src/unpack.c b/perl/xs-src/unpack.c
index eb6e0ddb..20a07372 100644
--- a/perl/xs-src/unpack.c
+++ b/perl/xs-src/unpack.c
@@ -2,13 +2,9 @@
 extern "C" {
 #endif
 
-#include "EXTERN.h"
-#include "perl.h"
-#include "XSUB.h"
-#include "util.h"
 #define NEED_newRV_noinc
 #define NEED_sv_2pv_flags
-#include "ppport.h"
+#include "perlxs.h"
 
 #ifdef __cplusplus
 };
@@ -38,7 +34,7 @@ typedef struct {
 /* ---------------------------------------------------------------------- */
 /* utility functions                                                      */
 
-static INLINE SV *
+STATIC_INLINE SV *
 get_bool (const char *name) {
     SV * sv = sv_mortalcopy(get_sv( name, 1 ));
 
@@ -60,19 +56,19 @@ static SV* template_data(msgpack_unpack_t* u);
 static int template_execute(msgpack_unpack_t* u,
     const char* data, size_t len, size_t* off);
 
-static INLINE SV* template_callback_root(unpack_user* u)
+STATIC_INLINE SV* template_callback_root(unpack_user* u)
 { return &PL_sv_undef; }
 
-static INLINE int template_callback_uint8(unpack_user* u, uint8_t d, SV** o)
+STATIC_INLINE int template_callback_uint8(unpack_user* u, uint8_t d, SV** o)
 { *o = sv_2mortal(newSVuv(d)); return 0; }
 
-static INLINE int template_callback_uint16(unpack_user* u, uint16_t d, SV** o)
+STATIC_INLINE int template_callback_uint16(unpack_user* u, uint16_t d, SV** o)
 { *o = sv_2mortal(newSVuv(d)); return 0; }
 
-static INLINE int template_callback_uint32(unpack_user* u, uint32_t d, SV** o)
+STATIC_INLINE int template_callback_uint32(unpack_user* u, uint32_t d, SV** o)
 { *o = sv_2mortal(newSVuv(d)); return 0; }
 
-static INLINE int template_callback_uint64(unpack_user* u, uint64_t d, SV** o)
+STATIC_INLINE int template_callback_uint64(unpack_user* u, uint64_t d, SV** o)
 {
 #if IVSIZE==4
     *o = sv_2mortal(newSVnv(d));
@@ -82,47 +78,47 @@ static INLINE int template_callback_uint64(unpack_user* u, uint64_t d, SV** o)
     return 0;
 }
 
-static INLINE int template_callback_int8(unpack_user* u, int8_t d, SV** o)
+STATIC_INLINE int template_callback_int8(unpack_user* u, int8_t d, SV** o)
 { *o = sv_2mortal(newSViv((long)d)); return 0; }
 
-static INLINE int template_callback_int16(unpack_user* u, int16_t d, SV** o)
+STATIC_INLINE int template_callback_int16(unpack_user* u, int16_t d, SV** o)
 { *o = sv_2mortal(newSViv((long)d)); return 0; }
 
-static INLINE int template_callback_int32(unpack_user* u, int32_t d, SV** o)
+STATIC_INLINE int template_callback_int32(unpack_user* u, int32_t d, SV** o)
 { *o = sv_2mortal(newSViv((long)d)); return 0; }
 
-static INLINE int template_callback_int64(unpack_user* u, int64_t d, SV** o)
+STATIC_INLINE int template_callback_int64(unpack_user* u, int64_t d, SV** o)
 { *o = sv_2mortal(newSViv(d)); return 0; }
 
-static INLINE int template_callback_float(unpack_user* u, float d, SV** o)
+STATIC_INLINE int template_callback_float(unpack_user* u, float d, SV** o)
 { *o = sv_2mortal(newSVnv(d)); return 0; }
 
-static INLINE int template_callback_double(unpack_user* u, double d, SV** o)
+STATIC_INLINE int template_callback_double(unpack_user* u, double d, SV** o)
 { *o = sv_2mortal(newSVnv(d)); return 0; }
 
 /* &PL_sv_undef is not so good. see http://gist.github.com/387743 */
-static INLINE int template_callback_nil(unpack_user* u, SV** o)
+STATIC_INLINE int template_callback_nil(unpack_user* u, SV** o)
 { *o = sv_newmortal(); return 0; }
 
-static INLINE int template_callback_true(unpack_user* u, SV** o)
+STATIC_INLINE int template_callback_true(unpack_user* u, SV** o)
 { *o = get_bool("Data::MessagePack::true") ; return 0; }
 
-static INLINE int template_callback_false(unpack_user* u, SV** o)
+STATIC_INLINE int template_callback_false(unpack_user* u, SV** o)
 { *o = get_bool("Data::MessagePack::false") ; return 0; }
 
-static INLINE int template_callback_array(unpack_user* u, unsigned int n, SV** o)
+STATIC_INLINE int template_callback_array(unpack_user* u, unsigned int n, SV** o)
 { AV* a = (AV*)sv_2mortal((SV*)newAV()); *o = sv_2mortal((SV*)newRV_inc((SV*)a)); av_extend(a, n); return 0; }
 
-static INLINE int template_callback_array_item(unpack_user* u, SV** c, SV* o)
+STATIC_INLINE int template_callback_array_item(unpack_user* u, SV** c, SV* o)
 { av_push((AV*)SvRV(*c), o); SvREFCNT_inc(o); return 0; }  /* FIXME set value directry RARRAY_PTR(obj)[RARRAY_LEN(obj)++] */
 
-static INLINE int template_callback_map(unpack_user* u, unsigned int n, SV** o)
+STATIC_INLINE int template_callback_map(unpack_user* u, unsigned int n, SV** o)
 { HV * h = (HV*)sv_2mortal((SV*)newHV()); *o = sv_2mortal(newRV_inc((SV*)h)); return 0; }
 
-static INLINE int template_callback_map_item(unpack_user* u, SV** c, SV* k, SV* v)
+STATIC_INLINE int template_callback_map_item(unpack_user* u, SV** c, SV* k, SV* v)
 { hv_store_ent((HV*)SvRV(*c), k, v, 0); SvREFCNT_inc(v); return 0; }
 
-static INLINE int template_callback_raw(unpack_user* u, const char* b, const char* p, unsigned int l, SV** o)
+STATIC_INLINE int template_callback_raw(unpack_user* u, const char* b, const char* p, unsigned int l, SV** o)
 { *o = sv_2mortal((l==0) ? newSVpv("", 0) : newSVpv(p, l)); return 0; }
 /* { *o = newSVpvn_flags(p, l, SVs_TEMP); return 0; }  <= this does not works. */