proxy: speedup stats proxy

do some sad snprintf unrolling to reduce CPU usage when dumping high
numbers of user stats.

proto_proxy.c

@@ -214,11 +214,29 @@ void process_proxy_stats(void *arg, ADD_STAT add_stats, void *c) {
 
     // return all of the user generated stats
     if (ctx->user_stats_namebuf) {
+        char vbuf[INCR_MAX_STORAGE_LEN];
+        char *e = NULL; // ptr into vbuf
+        const char *pfx = "user_";
+        const size_t pfxlen = strlen(pfx);
         for (int x = 0; x < stats_num; x++) {
             if (us[x].cname) {
                 char *name = ctx->user_stats_namebuf + us[x].cname;
-                snprintf(key_str, STAT_KEY_LEN-1, "user_%s", name);
-                APPEND_STAT(key_str, "%llu", (unsigned long long)counters[x]);
+                size_t nlen = strlen(name);
+                if (nlen > STAT_KEY_LEN-6) {
+                    // impossible, but for paranoia.
+                    nlen = STAT_KEY_LEN;
+                }
+                // avoiding an snprintf call for some performance ("user_%s")
+                memcpy(key_str, pfx, pfxlen);
+                memcpy(key_str+pfxlen, name, nlen);
+                key_str[pfxlen+nlen] = '\0';
+
+                // APPEND_STAT() calls another snprintf, which calls our
+                // add_stats argument. Lets skip yet another snprintf with
+                // some unrolling.
+                e = itoa_u64(counters[x], vbuf);
+                *(e+1) = '\0';
+                add_stats(key_str, pfxlen+nlen, vbuf, e-vbuf, c);
             }
         }
     }