Author: stefan2
Date: Sun Feb 8 16:47:33 2015
New Revision: 1658205
URL: http://svn.apache.org/r1658205
Log:
Squeeze a few more percent efficiency from ra_svn protocol parser.
As it turns out, some reports like 'svn diff --summary' have a high
density of command words. Parse them efficiently.
* subversion/libsvn_ra_svn/marshal.c
(MAX_WORD_LENGTH): Pick a slightly smaller, carefully selected value
that will fit with the updated logic in ...
(read_item): ... here. Add a fast parser path for "word" tokens.
Modified:
subversion/branches/ra-svn-tuning/subversion/libsvn_ra_svn/marshal.c
Modified: subversion/branches/ra-svn-tuning/subversion/libsvn_ra_svn/marshal.c
URL:
http://svn.apache.org/viewvc/subversion/branches/ra-svn-tuning/subversion/libsvn_ra_svn/marshal.c?rev=1658205&r1=1658204&r2=1658205&view=diff
==============================================================================
--- subversion/branches/ra-svn-tuning/subversion/libsvn_ra_svn/marshal.c
(original)
+++ subversion/branches/ra-svn-tuning/subversion/libsvn_ra_svn/marshal.c Sun
Feb 8 16:47:33 2015
@@ -60,9 +60,10 @@
/* We don't use "words" longer than this in our protocol. The longest word
* we are currently using is only about 16 chars long but we leave room for
- * longer future capability and command names.
+ * longer future capability and command names. See read_item() to understand
+ * why MAX_WORD_LENGTH - 1 should be a multiple of 8.
*/
-#define MAX_WORD_LENGTH 31
+#define MAX_WORD_LENGTH 25
/* The generic parsers will use the following value to limit the recursion
* depth to some reasonable value. The current protocol implementation
@@ -1060,17 +1061,40 @@ static svn_error_t *read_item(svn_ra_svn
char *p = buffer + 1;
buffer[0] = c;
- while (1)
+ if (conn->read_ptr + MAX_WORD_LENGTH <= conn->read_end)
{
- SVN_ERR(readbuf_getchar(conn, pool, p));
- if (!svn_ctype_isalnum(*p) && *p != '-')
- break;
+ /* Fast path: we can simply take a chunk from the read
+ * buffer and inspect it with no overflow checks etc.
+ *
+ * Copying these 24 bytes unconditionally is also faster
+ * than a variable-sized memcpy. Note that P is at BUFFER[1].
+ */
+ memcpy(p, conn->read_ptr, MAX_WORD_LENGTH - 1);
+ *end = 0;
+
+ /* This will terminate at P == END because of *END == NUL. */
+ while (svn_ctype_isalnum(*p) || *p == '-')
+ ++p;
- if (++p == end)
- return svn_error_create(SVN_ERR_RA_SVN_MALFORMED_DATA, NULL,
- _("Word is too long"));
+ /* Only now do we mark data as actually read. */
+ conn->read_ptr += p - buffer;
+ }
+ else
+ {
+ /* Slow path. Byte-by-byte copying and checking for
+ * input and output buffer boundaries. */
+ for (p = buffer + 1; p != end; ++p)
+ {
+ SVN_ERR(readbuf_getchar(conn, pool, p));
+ if (!svn_ctype_isalnum(*p) && *p != '-')
+ break;
+ }
}
+ if (p == end)
+ return svn_error_create(SVN_ERR_RA_SVN_MALFORMED_DATA, NULL,
+ _("Word is too long"));
+
c = *p;
*p = '\0';
