xdiff: load full words in the inner loop of xdl_hash_record
authorThomas Rast <trast@student.ethz.ch>
Fri, 6 Apr 2012 21:01:23 +0000 (23:01 +0200)
committerJunio C Hamano <gitster@pobox.com>
Tue, 10 Apr 2012 00:03:25 +0000 (17:03 -0700)
Redo the hashing loop in xdl_hash_record in a way that loads an entire
'long' at a time, using masking tricks to see when and where we found
the terminating '\n'.

I stole inspiration and code from the posts by Linus Torvalds around

https://lkml.org/lkml/2012/3/2/452
https://lkml.org/lkml/2012/3/5/6

His method reads the buffers in sizeof(long) increments, and may thus
overrun it by at most sizeof(long)-1 bytes before it sees the final
newline (or hits the buffer length check). I considered padding out
all buffers by a suitable amount to "catch" the overrun, but

* this does not work for mmap()'d buffers: if you map 4096+8 bytes
from a 4096 byte file, accessing the last 8 bytes results in a
SIGBUS on my machine; and

* it would also be extremely ugly because it intrudes deep into the
unpacking machinery.

So I adapted it to not read beyond the buffer at all. Instead, it
reads the final partial word byte-by-byte and strings it together.
Then it can use the same logic as before to finish the hashing.

So far we enable this only on x86_64, where it provides nice speedup
for diff-related work:

Test origin/next tr/xdiff-fast-hash
-----------------------------------------------------------------------------
4000.1: log -3000 (baseline) 0.07(0.05+0.02) 0.08(0.06+0.02) +14.3%
4000.2: log --raw -3000 (tree-only) 0.37(0.33+0.04) 0.37(0.32+0.04) +0.0%
4000.3: log -p -3000 (Myers) 1.75(1.65+0.09) 1.60(1.49+0.10) -8.6%
4000.4: log -p -3000 --histogram 1.73(1.62+0.09) 1.58(1.49+0.08) -8.7%
4000.5: log -p -3000 --patience 2.11(2.00+0.10) 1.94(1.80+0.11) -8.1%

Perhaps other platforms could also benefit. However it does NOT work
on big-endian systems!

[jc: minimum style and compilation fixes]

Signed-off-by: Thomas Rast <trast@student.ethz.ch>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Makefile
xdiff/xutils.c
index be1957a5e986d2e0581123dfe4e0eeb6702c13dc..13ed1b1d29299cf69bbd8cec69f40d313a86a9f5 100644 (file)
--- a/Makefile
+++ b/Makefile
@@ -288,6 +288,11 @@ all::
 # dependency rules.
 #
 # Define NATIVE_CRLF if your platform uses CRLF for line endings.
+#
+# Define XDL_FAST_HASH to use an alternative line-hashing method in
+# the diff algorithm.  It gives a nice speedup if your processor has
+# fast unaligned word loads.  Does NOT work on big-endian systems!
+# Enabled by default on x86_64.
 
 GIT-VERSION-FILE: FORCE
        @$(SHELL_PATH) ./GIT-VERSION-GEN
@@ -864,6 +869,9 @@ EXTLIBS =
 # because maintaining the nesting to match is a pain.  If
 # we had "elif" things would have been much nicer...
 
+ifeq ($(uname_M),x86_64)
+       XDL_FAST_HASH = YesPlease
+endif
 ifeq ($(uname_S),OSF1)
        # Need this for u_short definitions et al
        BASIC_CFLAGS += -D_OSF_SOURCE
@@ -1737,6 +1745,10 @@ ifndef NO_MSGFMT_EXTENDED_OPTIONS
        MSGFMT += --check --statistics
 endif
 
+ifneq (,$(XDL_FAST_HASH))
+       BASIC_CFLAGS += -DXDL_FAST_HASH
+endif
+
 ifeq ($(TCLTK_PATH),)
 NO_TCLTK=NoThanks
 endif
index 0de084e53f5144153373cc66cae7b523b4ae2812..e05b5c96aaa274f79f7d512459b35885be9e8d4d 100644 (file)
@@ -20,6 +20,8 @@
  *
  */
 
+#include <limits.h>
+#include <assert.h>
 #include "xinclude.h"
 
 
@@ -276,6 +278,115 @@ static unsigned long xdl_hash_record_with_whitespace(char const **data,
        return ha;
 }
 
+#ifdef XDL_FAST_HASH
+
+#define ONEBYTES       0x0101010101010101ul
+#define NEWLINEBYTES   0x0a0a0a0a0a0a0a0aul
+#define HIGHBITS       0x8080808080808080ul
+
+/* Return the high bit set in the first byte that is a zero */
+static inline unsigned long has_zero(unsigned long a)
+{
+       return ((a - ONEBYTES) & ~a) & HIGHBITS;
+}
+
+#if __WORDSIZE == 64
+
+/*
+ * Jan Achrenius on G+: microoptimized version of
+ * the simpler "(mask & ONEBYTES) * ONEBYTES >> 56"
+ * that works for the bytemasks without having to
+ * mask them first.
+ */
+static inline long count_masked_bytes(unsigned long mask)
+{
+       return mask * 0x0001020304050608 >> 56;
+}
+
+#else  /* 32-bit case */
+
+/* Modified Carl Chatfield G+ version for 32-bit */
+static inline long count_masked_bytes(long mask)
+{
+       /*
+        * (a) gives us
+        *   -1 (0, ff), 0 (ffff) or 1 (ffffff)
+        * (b) gives us
+        *   0 for 0, 1 for (ff ffff ffffff)
+        * (a+b+1) gives us
+        *   correct 0-3 bytemask count result
+        */
+       long a = (mask - 256) >> 23;
+       long b = mask & 1;
+       return a + b + 1;
+}
+
+#endif
+
+unsigned long xdl_hash_record(char const **data, char const *top, long flags)
+{
+       unsigned long hash = 5381;
+       unsigned long a = 0, mask = 0;
+       char const *ptr = *data;
+       char const *end = top - sizeof(unsigned long) + 1;
+
+       if (flags & XDF_WHITESPACE_FLAGS)
+               return xdl_hash_record_with_whitespace(data, top, flags);
+
+       ptr -= sizeof(unsigned long);
+       do {
+               hash += hash << 5;
+               hash ^= a;
+               ptr += sizeof(unsigned long);
+               if (ptr >= end)
+                       break;
+               a = *(unsigned long *)ptr;
+               /* Do we have any '\n' bytes in this word? */
+               mask = has_zero(a ^ NEWLINEBYTES);
+       } while (!mask);
+
+       if (ptr >= end) {
+               /*
+                * There is only a partial word left at the end of the
+                * buffer. Because we may work with a memory mapping,
+                * we have to grab the rest byte by byte instead of
+                * blindly reading it.
+                *
+                * To avoid problems with masking in a signed value,
+                * we use an unsigned char here.
+                */
+               const char *p;
+               for (p = top - 1; p >= ptr; p--)
+                       a = (a << 8) + *((const unsigned char *)p);
+               mask = has_zero(a ^ NEWLINEBYTES);
+               if (!mask)
+                       /*
+                        * No '\n' found in the partial word.  Make a
+                        * mask that matches what we read.
+                        */
+                       mask = 1UL << (8 * (top - ptr) + 7);
+       }
+
+       /* The mask *below* the first high bit set */
+       mask = (mask - 1) & ~mask;
+       mask >>= 7;
+       hash += hash << 5;
+       hash ^= a & mask;
+
+       /* Advance past the last (possibly partial) word */
+       ptr += count_masked_bytes(mask);
+
+       if (ptr < top) {
+               assert(*ptr == '\n');
+               ptr++;
+       }
+
+       *data = ptr;
+
+       return hash;
+}
+
+#else /* XDL_FAST_HASH */
 
 unsigned long xdl_hash_record(char const **data, char const *top, long flags) {
        unsigned long ha = 5381;
@@ -293,6 +404,7 @@ unsigned long xdl_hash_record(char const **data, char const *top, long flags) {
        return ha;
 }
 
+#endif /* XDL_FAST_HASH */
 
 unsigned int xdl_hashbits(unsigned int size) {
        unsigned int val = 1, bits = 0;