Re: svn commit: r1001413 - in /subversion/branches/performance/subversion: include/svn_string.h libsvn_subr/svn_string.c
One correction (in the comments -- which are great, BTW):
s/descression/discretion
-Ed
On Mon, Sep 27, 2010 at 5:05 PM, Hyrum K. Wright
hyrum_wri...@mail.utexas.edu wrote:
+1 to merge to trunk.
On Sun, Sep 26, 2010 at 6:01 AM, stef...@apache.org wrote:
Author: stefan2
Date: Sun Sep 26 11:01:03 2010
New Revision: 1001413
URL: http://svn.apache.org/viewvc?rev=1001413view=rev
Log:
Extensively document the benefits of svn_stringbuf_appendbyte and
the rationals behind its implementation. To simplify the explanations,
one statement had to be moved.
* subversion/include/svn_string.h
(svn_stringbuf_appendbyte): extend docstring to indicate that this method
is cheaper to call and execute
* subversion/libsvn_subr/svn_string.c
(svn_stringbuf_appendbyte): reorder statements for easier description;
add extensive description about the optimizations done
Modified:
subversion/branches/performance/subversion/include/svn_string.h
subversion/branches/performance/subversion/libsvn_subr/svn_string.c
Modified: subversion/branches/performance/subversion/include/svn_string.h
URL:
http://svn.apache.org/viewvc/subversion/branches/performance/subversion/include/svn_string.h?rev=1001413r1=1001412r2=1001413view=diff
==
--- subversion/branches/performance/subversion/include/svn_string.h
(original)
+++ subversion/branches/performance/subversion/include/svn_string.h Sun Sep
26 11:01:03 2010
@@ -259,6 +259,10 @@ void
svn_stringbuf_fillchar(svn_stringbuf_t *str, unsigned char c);
/** Append a single character @a byte onto @a targetstr.
+ * This is an optimized version of @ref svn_stringbuf_appendbytes
+ * that is much faster to call and execute. Gains vary with the ABI.
+ * The advantages extend beyond the actual call because the reduced
+ * register pressure allows for more optimization within the caller.
*
* reallocs if necessary. @a targetstr is affected, nothing else is.
* @since New in 1.7.
Modified: subversion/branches/performance/subversion/libsvn_subr/svn_string.c
URL:
http://svn.apache.org/viewvc/subversion/branches/performance/subversion/libsvn_subr/svn_string.c?rev=1001413r1=1001412r2=1001413view=diff
==
--- subversion/branches/performance/subversion/libsvn_subr/svn_string.c
(original)
+++ subversion/branches/performance/subversion/libsvn_subr/svn_string.c Sun
Sep 26 11:01:03 2010
@@ -393,28 +393,60 @@ svn_stringbuf_ensure(svn_stringbuf_t *st
}
+/* WARNING - Optimized code ahead!
+ * This function has been hand-tuned for performance. Please read
+ * the comments below before modifying the code.
+ */
void
svn_stringbuf_appendbyte(svn_stringbuf_t *str, char byte)
{
+ char *dest;
+ apr_size_t old_len = str-len;
+
/* In most cases, there will be pre-allocated memory left
* to just write the new byte at the end of the used section
* and terminate the string properly.
*/
- apr_size_t old_len = str-len;
- if (str-blocksize old_len + 1)
+ if (str-blocksize old_len + 1)
{
- char *dest = str-data;
+ /* The following read does not depend this write, so we
+ * can issue the write first to minimize register pressure:
+ * The value of old_len+1 is no longer needed; on most processors,
+ * dest[old_len+1] will be calculated implicitly as part of
+ * the addressing scheme.
+ */
+ str-len = old_len+1;
+
+ /* Since the compiler cannot be sure that *src-data and *src
+ * don't overlap, we read src-data *once* before writing
+ * to *src-data. Replacing dest with str-data would force
+ * the compiler to read it again after the first byte.
+ */
+ dest = str-data;
+ /* If not already available in a register as per ABI, load
+ * byte into the register (e.g. the one freed from old_len+1),
+ * then write it to the string buffer and terminate it properly.
+ *
+ * Including the byte fetch, all operations so far could be
+ * issued at once and be scheduled at the CPU's descression.
+ * Most likely, no-one will soon depend on the data that will be
+ * written in this function. So, no stalls there, either.
+ */
dest[old_len] = byte;
dest[old_len+1] = '\0';
-
- str-len = old_len+1;
}
else
{
/* we need to re-allocate the string buffer
* - let the more generic implementation take care of that part
*/
+
+ /* Depending on the ABI, byte is a register value. If we were
+ * to take its address directly, the compiler might decide to
+ * put in on the stack *unconditionally*, even if that would
+ * only be necessary for this block.
+ */
char b = byte;
svn_stringbuf_appendbytes(str, b, 1);
}
Re: svn commit: r1001413 - in /subversion/branches/performance/subversion: include/svn_string.h libsvn_subr/svn_string.c
+1 to merge to trunk.
On Sun, Sep 26, 2010 at 6:01 AM, stef...@apache.org wrote:
Author: stefan2
Date: Sun Sep 26 11:01:03 2010
New Revision: 1001413
URL: http://svn.apache.org/viewvc?rev=1001413view=rev
Log:
Extensively document the benefits of svn_stringbuf_appendbyte and
the rationals behind its implementation. To simplify the explanations,
one statement had to be moved.
* subversion/include/svn_string.h
(svn_stringbuf_appendbyte): extend docstring to indicate that this method
is cheaper to call and execute
* subversion/libsvn_subr/svn_string.c
(svn_stringbuf_appendbyte): reorder statements for easier description;
add extensive description about the optimizations done
Modified:
subversion/branches/performance/subversion/include/svn_string.h
subversion/branches/performance/subversion/libsvn_subr/svn_string.c
Modified: subversion/branches/performance/subversion/include/svn_string.h
URL:
http://svn.apache.org/viewvc/subversion/branches/performance/subversion/include/svn_string.h?rev=1001413r1=1001412r2=1001413view=diff
==
--- subversion/branches/performance/subversion/include/svn_string.h (original)
+++ subversion/branches/performance/subversion/include/svn_string.h Sun Sep
26 11:01:03 2010
@@ -259,6 +259,10 @@ void
svn_stringbuf_fillchar(svn_stringbuf_t *str, unsigned char c);
/** Append a single character @a byte onto @a targetstr.
+ * This is an optimized version of @ref svn_stringbuf_appendbytes
+ * that is much faster to call and execute. Gains vary with the ABI.
+ * The advantages extend beyond the actual call because the reduced
+ * register pressure allows for more optimization within the caller.
*
* reallocs if necessary. @a targetstr is affected, nothing else is.
* @since New in 1.7.
Modified: subversion/branches/performance/subversion/libsvn_subr/svn_string.c
URL:
http://svn.apache.org/viewvc/subversion/branches/performance/subversion/libsvn_subr/svn_string.c?rev=1001413r1=1001412r2=1001413view=diff
==
--- subversion/branches/performance/subversion/libsvn_subr/svn_string.c
(original)
+++ subversion/branches/performance/subversion/libsvn_subr/svn_string.c Sun
Sep 26 11:01:03 2010
@@ -393,28 +393,60 @@ svn_stringbuf_ensure(svn_stringbuf_t *st
}
+/* WARNING - Optimized code ahead!
+ * This function has been hand-tuned for performance. Please read
+ * the comments below before modifying the code.
+ */
void
svn_stringbuf_appendbyte(svn_stringbuf_t *str, char byte)
{
+ char *dest;
+ apr_size_t old_len = str-len;
+
/* In most cases, there will be pre-allocated memory left
* to just write the new byte at the end of the used section
* and terminate the string properly.
*/
- apr_size_t old_len = str-len;
- if (str-blocksize old_len + 1)
+ if (str-blocksize old_len + 1)
{
- char *dest = str-data;
+ /* The following read does not depend this write, so we
+ * can issue the write first to minimize register pressure:
+ * The value of old_len+1 is no longer needed; on most processors,
+ * dest[old_len+1] will be calculated implicitly as part of
+ * the addressing scheme.
+ */
+ str-len = old_len+1;
+
+ /* Since the compiler cannot be sure that *src-data and *src
+ * don't overlap, we read src-data *once* before writing
+ * to *src-data. Replacing dest with str-data would force
+ * the compiler to read it again after the first byte.
+ */
+ dest = str-data;
+ /* If not already available in a register as per ABI, load
+ * byte into the register (e.g. the one freed from old_len+1),
+ * then write it to the string buffer and terminate it properly.
+ *
+ * Including the byte fetch, all operations so far could be
+ * issued at once and be scheduled at the CPU's descression.
+ * Most likely, no-one will soon depend on the data that will be
+ * written in this function. So, no stalls there, either.
+ */
dest[old_len] = byte;
dest[old_len+1] = '\0';
-
- str-len = old_len+1;
}
else
{
/* we need to re-allocate the string buffer
* - let the more generic implementation take care of that part
*/
+
+ /* Depending on the ABI, byte is a register value. If we were
+ * to take its address directly, the compiler might decide to
+ * put in on the stack *unconditionally*, even if that would
+ * only be necessary for this block.
+ */
char b = byte;
svn_stringbuf_appendbytes(str, b, 1);
}
