From a7b6913ac0dc24c82ed9c77f1d66b599de174d29 Mon Sep 17 00:00:00 2001
From: Weijia Wang <9713184+wegank@users.noreply.github.com>
Date: Sun, 8 Mar 2026 12:02:38 +0100
Subject: [PATCH] Avoid _mm_cvtsi128_si64 on 32-bit builds

---
 src/fglm/linalg-fglm.c | 6 ++++++
 1 file changed, 6 insertions(+)

diff --git a/src/fglm/linalg-fglm.c b/src/fglm/linalg-fglm.c
index e3270c6f..7ced7f02 100644
--- a/src/fglm/linalg-fglm.c
+++ b/src/fglm/linalg-fglm.c
@@ -211,7 +211,13 @@ uint64_t _mm256_hsum(__m256i a)
     __m256i sum_lo = _mm256_add_epi64(a, a_hi);
     __m128i sum_hi = _mm256_extracti128_si256(sum_lo, 1);
     __m128i sum = _mm_add_epi64(_mm256_castsi256_si128(sum_lo), sum_hi);
+#if defined(__x86_64__) || defined(_M_X64)
     return (uint64_t) _mm_cvtsi128_si64(sum);
+#else
+    uint64_t lo = (uint32_t) _mm_cvtsi128_si32(sum);
+    uint64_t hi = (uint32_t) _mm_cvtsi128_si32(_mm_srli_si128(sum, 4));
+    return lo | (hi << 32);
+#endif
 }
 
 uint32_t _nmod32_vec_dot_split_avx2(const uint32_t * vec1, const uint32_t * vec2, int64_t len,