From: Konstantinos Margaritis <konstantinos@vectorcamp.gr>
Date: Thu, 11 Feb 2021 12:21:57 +0000 (+0200)
Subject: use unaligned loads for short scans
X-Git-Tag: v5.4.1+vectorscan~6^2
X-Git-Url: http://git.ipfire.org/cgi-bin/gitweb.cgi?a=commitdiff_plain;h=9fd94e0062159e49939aa6be7fffdc82039d176f;p=thirdparty%2Fvectorscan.git

use unaligned loads for short scans
---

diff --git a/src/hwlm/noodle_engine_sse.c b/src/hwlm/noodle_engine_sse.c
index fcd753fc..0f14852d 100644
--- a/src/hwlm/noodle_engine_sse.c
+++ b/src/hwlm/noodle_engine_sse.c
@@ -50,7 +50,7 @@ hwlm_error_t scanSingleShort(const struct noodTable *n, const u8 *buf,
         return HWLM_SUCCESS;
     }
     m128 mask128 = noCase ? caseMask : ones128();
-    m128 v = and128(load128(d), mask128);
+    m128 v = and128(loadu128(d), mask128);
 
     // mask out where we can't match
     u32 mask = (0xFFFF >> (16 - l));
@@ -103,7 +103,7 @@ hwlm_error_t scanDoubleShort(const struct noodTable *n, const u8 *buf,
 
     DEBUG_PRINTF("d %zu\n", d - buf);
     m128 mask128 = noCase ? caseMask : ones128();
-    m128 v = and128(load128(d), mask128);
+    m128 v = and128(loadu128(d), mask128);
 
     u32 z = movemask128(and128(lshiftbyte_m128(eq128(mask1, v), 1),
                                eq128(mask2, v)));