Merge branch 'sverker/master/optimize-atom-enc-dec/ERIERL-150'

3 files changed, 101 insertions, 17 deletions

diff --git a/lib/erl_interface/configure.in b/lib/erl_interface/configure.in
index 0a8fbf513c..696ebf5ca0 100644
--- a/lib/erl_interface/configure.in
+++ b/lib/erl_interface/configure.in
@@ -106,6 +106,19 @@ if test $ac_cv_sizeof_long = 8; then
   CFLAGS="$CFLAGS -DEI_64BIT"
 fi
 
+LM_HARDWARE_ARCH
+
+AC_MSG_CHECKING(for unaligned word access)
+case "$ARCH" in
+    x86|amd64)
+	AC_MSG_RESULT(yes: x86 or amd64)
+	AC_DEFINE(HAVE_UNALIGNED_WORD_ACCESS, 1, [Define if hw supports unaligned word access])
+	;;
+    *)
+	AC_MSG_RESULT(no)
+	;;
+esac
+
 AC_CHECK_TOOL(AR, ar, false)
 if test "$AR" = false; then
   AC_MSG_ERROR([No 'ar' command found in PATH])
diff --git a/lib/erl_interface/src/decode/decode_atom.c b/lib/erl_interface/src/decode/decode_atom.c
index b3bba82434..87cd75b1be 100644
--- a/lib/erl_interface/src/decode/decode_atom.c
+++ b/lib/erl_interface/src/decode/decode_atom.c
@@ -92,6 +92,51 @@ int ei_decode_atom_as(const char *buf, int *index, char* p, int destlen,
 }	
 
 
+
+#ifdef HAVE_UNALIGNED_WORD_ACCESS
+
+#if SIZEOF_VOID_P == SIZEOF_LONG
+typedef unsigned long AsciiWord;
+#elif SIZEOF_VOID_P == SIZEOF_LONG_LONG
+typedef unsigned long long AsciiWord;
+#else
+#  error "Uknown word type"
+#endif
+
+#if SIZEOF_VOID_P == 4
+#  define ASCII_CHECK_MASK ((AsciiWord)0x80808080U)
+#elif SIZEOF_VOID_P == 8
+#  define ASCII_CHECK_MASK ((AsciiWord)0x8080808080808080U)
+#endif
+
+static int ascii_fast_track(char* dst, const char* src, int slen, int destlen)
+{
+    const AsciiWord* src_word = (AsciiWord*) src;
+    const AsciiWord* const src_word_end = src_word + (slen / sizeof(AsciiWord));
+
+    if (destlen < slen)
+        return 0;
+
+    if (dst) {
+        AsciiWord* dst_word = (AsciiWord*)dst;
+
+        while (src_word < src_word_end) {
+            if ((*src_word & ASCII_CHECK_MASK) != 0)
+                break;
+            *dst_word++ = *src_word++;
+        }
+    }
+    else {
+        while (src_word < src_word_end) {
+            if ((*src_word & ASCII_CHECK_MASK) != 0)
+                break;
+            src_word++;
+        }
+    }
+    return (char*)src_word - src;
+}
+#endif /* HAVE_UNALIGNED_WORD_ACCESS */
+
 int utf8_to_latin1(char* dst, const char* src, int slen, int destlen,
 		   erlang_char_encoding* res_encp)
 {
@@ -99,6 +144,15 @@ int utf8_to_latin1(char* dst, const char* src, int slen, int destlen,
     const char* const dst_end = dst + destlen;
     int found_non_ascii = 0;
 
+#ifdef HAVE_UNALIGNED_WORD_ACCESS
+    {
+        int aft = ascii_fast_track(dst, src, slen, destlen);
+        src += aft;
+        slen -= aft;
+        dst += aft;
+    }
+#endif
+
     while (slen > 0) {
 	if (dst >= dst_end) return -1;
 	if ((src[0] & 0x80) == 0) {
@@ -136,6 +190,14 @@ int latin1_to_utf8(char* dst, const char* src, int slen, int destlen,
     const char* const dst_end = dst + destlen;
     int found_non_ascii = 0;
 
+#ifdef HAVE_UNALIGNED_WORD_ACCESS
+    {
+        int aft = ascii_fast_track(dst, src, slen, destlen);
+        dst += aft;
+        src += aft;
+    }
+#endif
+
     while (src < src_end) {
 	if (dst >= dst_end) return -1;
 	if ((src[0] & 0x80) == 0) {
diff --git a/lib/ic/c_src/oe_ei_encode_atom.c b/lib/ic/c_src/oe_ei_encode_atom.c
index 758586d1d4..99a9fe26f0 100644
--- a/lib/ic/c_src/oe_ei_encode_atom.c
+++ b/lib/ic/c_src/oe_ei_encode_atom.c
@@ -20,28 +20,37 @@
  */
 #include <ic.h>
 
+#include <string.h>
+
+
+#define DIRTY_ATOM_ENC_MAX(LATIN1_CHARS) ((LATIN1_CHARS)*2 + 3)
+
 
 int oe_ei_encode_atom(CORBA_Environment *ev, const char *p) { 
   int size = ev->_iout;
+  size_t len = strlen(p);
+
+  if (DIRTY_ATOM_ENC_MAX(len) >= ev->_outbufsz) {
+
+      ei_encode_atom_len(0,&size,p,len);
+
+      if (size >= ev->_outbufsz) {
+        char *buf = ev->_outbuf;
+        int bufsz = ev->_outbufsz + ev->_memchunk;
+
+        while (size >= bufsz)
+          bufsz += ev->_memchunk;
+
+        if ((buf = realloc(buf, bufsz)) == NULL) {
+          CORBA_exc_set(ev, CORBA_SYSTEM_EXCEPTION, NO_MEMORY, "End of heap memory while encoding");
+          return -1;  /* OUT OF MEMORY */
+        }
 
-  ei_encode_atom(0,&size,p);
-
-  if (size >= ev->_outbufsz) {
-    char *buf = ev->_outbuf;
-    int bufsz = ev->_outbufsz + ev->_memchunk;
-    
-    while (size >= bufsz)
-      bufsz += ev->_memchunk;
-    
-    if ((buf = realloc(buf, bufsz)) == NULL) {
-      CORBA_exc_set(ev, CORBA_SYSTEM_EXCEPTION, NO_MEMORY, "End of heap memory while encoding");
-      return -1;  /* OUT OF MEMORY */ 
-    }
-
-    ev->_outbuf = buf;
-    ev->_outbufsz = bufsz;
+        ev->_outbuf = buf;
+        ev->_outbufsz = bufsz;
+      }
   }
 
-  return ei_encode_atom(ev->_outbuf,&ev->_iout,p);
+  return ei_encode_atom_len(ev->_outbuf,&ev->_iout,p,len);
 }