From: Victor Julien <victor@inliniac.net>
Date: Thu, 14 Nov 2013 11:51:06 +0000 (+0100)
Subject: memcmp: don't use SSE intrinsics if less that 16 bytes are available in SSE_4_2 version.
X-Git-Tag: suricata-2.0beta2~155
X-Git-Url: http://git.ipfire.org/cgi-bin/gitweb.cgi?a=commitdiff_plain;h=49a54713daafde0edbecdd5e7e8e81dfb342ad37;p=thirdparty%2Fsuricata.git

memcmp: don't use SSE intrinsics if less that 16 bytes are available in SSE_4_2 version.
---

diff --git a/src/util-memcmp.h b/src/util-memcmp.h
index f6fc3a0c59..8443526318 100644
--- a/src/util-memcmp.h
+++ b/src/util-memcmp.h
@@ -66,6 +66,12 @@ static inline int SCMemcmp(void *s1, void *s2, size_t n)
     size_t m = 0;
 
     do {
+        /* apparently we can't just read 16 bytes even though
+         * it almost always works fine :) */
+        if (likely(n - m < 16)) {
+            return memcmp(s1, s2, n - m) ? 1 : 0;
+        }
+
         /* load the buffers into the 128bit vars */
         b1 = _mm_loadu_si128((const __m128i *) s1);
         b2 = _mm_loadu_si128((const __m128i *) s2);
@@ -102,6 +108,12 @@ static inline int SCMemcmpLowercase(void *s1, void *s2, size_t n)
     __m128i uplow = _mm_set1_epi8(0x20);
 
     do {
+        /* apparently we can't just read 16 bytes even though
+         * it almost always works fine :) */
+        if (likely(n - m < 16)) {
+            return MemcmpLowercase(s1, s2, n - m);
+        }
+
         b1 = _mm_loadu_si128((const __m128i *) s1);
         b2 = _mm_loadu_si128((const __m128i *) s2);
         size_t len = n - m;