Prevent some invalid associativity in to_float implementation for avx2 under -ffast-math

Fix  #1264

Author: serge-sans-paille

include/xsimd/arch/xsimd_avx2.hpp

@@ -542,6 +542,9 @@ namespace xsimd
                 __m256i mask = _mm256_setr_epi16(0xFFFF, 0xFFFF, 0xFFFF, 0x0000, 0xFFFF, 0xFFFF, 0xFFFF, 0x0000,
                                                  0xFFFF, 0xFFFF, 0xFFFF, 0x0000, 0xFFFF, 0xFFFF, 0xFFFF, 0x0000);
                 __m256i xL = _mm256_or_si256(_mm256_and_si256(mask, x), _mm256_andnot_si256(mask, _mm256_castpd_si256(_mm256_set1_pd(0x0010000000000000)))); //  2^52
+#ifdef __FAST_MATH__
+                volatile // prevents some association with the add below, when converting from (-1, -1, 0, 0)
+#endif
                 __m256d f = _mm256_sub_pd(_mm256_castsi256_pd(xH), _mm256_set1_pd(442726361368656609280.)); //  3*2^67 + 2^52
                 return _mm256_add_pd(f, _mm256_castsi256_pd(xL));
             }