From 3898cb596a92680d4dcba6988852133bbe957608 Mon Sep 17 00:00:00 2001
From: Andrew Noyes <andrew@weaselab.dev>
Date: Thu, 1 Aug 2024 15:24:09 -0700
Subject: [PATCH] Use _mm512_loadu_epi32 instead of memcpy

The latter somehow generated a vmovdqa64 instruction that segfaulted on
unaligned memory, with gcc and --coverage. Hmm.
---
 ConflictSet.cpp | 7 +++----
 1 file changed, 3 insertions(+), 4 deletions(-)

diff --git a/ConflictSet.cpp b/ConflictSet.cpp
index 1588883..ece8e60 100644
--- a/ConflictSet.cpp
+++ b/ConflictSet.cpp
@@ -1988,12 +1988,11 @@ uint32_t compare16_32bit(const InternalVersionT *vs, InternalVersionT rv) {
 
 __attribute__((target("avx512f"))) uint32_t
 compare16_32bit_avx512(const InternalVersionT *vs, InternalVersionT rv) {
-  __m512i w;
-  memcpy(&w, vs, sizeof(w));
   uint32_t r;
   memcpy(&r, &rv, sizeof(r));
-  return _mm512_cmpgt_epi32_mask(_mm512_sub_epi32(w, _mm512_set1_epi32(r)),
-                                 _mm512_setzero_epi32());
+  return _mm512_cmpgt_epi32_mask(
+      _mm512_sub_epi32(_mm512_loadu_epi32(vs), _mm512_set1_epi32(r)),
+      _mm512_setzero_epi32());
 }
 #endif