Faster bucket order compare for global ids.

2 files changed, 18 insertions, 34 deletions

diff --git a/document/src/vespa/document/base/globalid.cpp b/document/src/vespa/document/base/globalid.cpp
index bdf81d1c3c8..4f96ac0e659 100644
--- a/document/src/vespa/document/base/globalid.cpp
+++ b/document/src/vespa/document/base/globalid.cpp
@@ -39,39 +39,6 @@ getHexVal(char c)
 
 namespace document {
 
-bool
-GlobalId::BucketOrderCmp::operator()(const GlobalId &lhs, const GlobalId &rhs) const
-{
-    const unsigned char * __restrict__ a = lhs._gid._buffer;
-    const unsigned char * __restrict__ b = rhs._gid._buffer;
-    int diff;
-    if ((diff = compare(a[0], b[0])) != 0) {
-        return diff < 0;
-    }
-    if ((diff = compare(a[1], b[1])) != 0) {
-        return diff < 0;
-    }
-    if ((diff = compare(a[2], b[2])) != 0) {
-        return diff < 0;
-    }
-    if ((diff = compare(a[3], b[3])) != 0) {
-        return diff < 0;
-    }
-    if ((diff = compare(a[8], b[8])) != 0) {
-        return diff < 0;
-    }
-    if ((diff = compare(a[9], b[9])) != 0) {
-        return diff < 0;
-    }
-    if ((diff = compare(a[10], b[10])) != 0) {
-        return diff < 0;
-    }
-    if ((diff = compare(a[11], b[11])) != 0) {
-        return diff < 0;
-    }
-    return lhs < rhs;
-}
-
 vespalib::string GlobalId::toString() const {
     vespalib::asciistream out;
     out << "gid(0x" << vespalib::hex;
diff --git a/document/src/vespa/document/base/globalid.h b/document/src/vespa/document/base/globalid.h
index 1bb257b7cbf..ac30e189033 100644
--- a/document/src/vespa/document/base/globalid.h
+++ b/document/src/vespa/document/base/globalid.h
@@ -64,7 +64,24 @@ public:
      * given bucket.
      */
     struct BucketOrderCmp {
-        bool operator()(const GlobalId &lhs, const GlobalId &rhs) const;
+        bool operator()(const GlobalId &lhs, const GlobalId &rhs) const {
+            const uint32_t * __restrict__ a = lhs._gid._nums;
+            const uint32_t * __restrict__ b = rhs._gid._nums;
+            if (a[0] != b[0]) {
+                return bitswap(a[0]) < bitswap(b[0]);
+            }
+            if (a[2] != b[2]) {
+                return bitswap(a[2]) < bitswap(b[2]);
+            }
+            return __builtin_bswap32(a[1]) < __builtin_bswap32(b[1]);
+        }
+        static uint32_t bitswap(uint32_t value) {
+            value = ((value & 0x55555555) << 1) | ((value & 0xaaaaaaaa) >> 1);
+            value = ((value & 0x33333333) << 2) | ((value & 0xcccccccc) >> 2);
+            value = ((value & 0x0f0f0f0f) << 4) | ((value & 0xf0f0f0f0) >> 4);
+            return __builtin_bswap32(value);
+        }
+
         //These 2 compare methods are exposed only for testing
         static int compareRaw(unsigned char a, unsigned char b) {
             return a - b;