From: Alex Coyte <a.coyte@intel.com>
Date: Mon, 29 Aug 2016 05:07:22 +0000 (+1000)
Subject: rework load_m128_from_u64a()
X-Git-Tag: v4.4.0^2~94
X-Git-Url: http://git.ipfire.org/cgi-bin/gitweb.cgi?a=commitdiff_plain;h=e74b141e958816be48693e7258f2a5da1d8df04f;p=thirdparty%2Fvectorscan.git

rework load_m128_from_u64a()
---

diff --git a/src/util/simd_utils.h b/src/util/simd_utils.h
index dc8922fd..b7cb1c0f 100644
--- a/src/util/simd_utils.h
+++ b/src/util/simd_utils.h
@@ -176,7 +176,15 @@ static really_inline u64a movq(const m128 in) {
 /* another form of movq */
 static really_inline
 m128 load_m128_from_u64a(const u64a *p) {
+#if defined(__GNUC__) && !defined(__INTEL_COMPILER)
+    /* unfortunately _mm_loadl_epi64() is best avoided as it seems to cause
+     * trouble on some older compilers, possibly because it is misdefined to
+     * take an m128 as its parameter */
+    return _mm_set_epi64((__m64)0ULL, (__m64)*p);
+#else
+    /* ICC doesn't like casting to __m64 */
     return _mm_loadl_epi64((const m128 *)p);
+#endif
 }
 
 #define rshiftbyte_m128(a, count_immed) _mm_srli_si128(a, count_immed)