diff --git a/simd-checksum-x86_64.cpp b/simd-checksum-x86_64.cpp
index 33f26e92..438ce369 100644
--- a/simd-checksum-x86_64.cpp
+++ b/simd-checksum-x86_64.cpp
@@ -85,9 +85,7 @@ typedef long long __m256i_u __attribute__((__vector_size__(32), __may_alias__, _
 #define SSE2_HADDS_EPI16(a, b) _mm_adds_epi16(SSE2_INTERLEAVE_EVEN_EPI16(a, b), SSE2_INTERLEAVE_ODD_EPI16(a, b))
 #define SSE2_MADDUBS_EPI16(a, b) _mm_adds_epi16(SSE2_MULU_EVEN_EPI8(a, b), SSE2_MULU_ODD_EPI8(a, b))
 
-#ifndef USE_ROLL_ASM
-__attribute__ ((target("default"))) MVSTATIC int32 get_checksum1_avx2_64(schar* buf, int32 len, int32 i, uint32* ps1, uint32* ps2) { return i; }
-#endif
+__attribute__ ((target("default"))) MVSTATIC int32 get_checksum1_avx2(schar* buf, int32 len, int32 i, uint32* ps1, uint32* ps2) { return i; }
 __attribute__ ((target("default"))) MVSTATIC int32 get_checksum1_ssse3_32(schar* buf, int32 len, int32 i, uint32* ps1, uint32* ps2) { return i; }
 __attribute__ ((target("default"))) MVSTATIC int32 get_checksum1_sse2_32(schar* buf, int32 len, int32 i, uint32* ps1, uint32* ps2) { return i; }
 
@@ -316,6 +314,10 @@ __attribute__ ((target("sse2"))) MVSTATIC int32 get_checksum1_sse2_32(schar* buf
 #ifdef USE_ROLL_ASM /* { */
 
 extern "C" __attribute__ ((target("avx2"))) int32 get_checksum1_avx2_asm(schar* buf, int32 len, int32 i, uint32* ps1, uint32* ps2);
+__attribute__ ((target("avx2"))) MVSTATIC int32 get_checksum1_avx2(schar* buf, int32 len, int32 i, uint32* ps1, uint32* ps2)
+{
+	return get_checksum1_avx2_asm(buf, len, i, ps1, ps2);
+}
 
 #else /* } { */
 
@@ -432,6 +434,10 @@ __attribute__ ((target("avx2"))) MVSTATIC int32 get_checksum1_avx2_64(schar* buf
     }
     return i;
 }
+__attribute__ ((target("avx2"))) MVSTATIC int32 get_checksum1_avx2(schar* buf, int32 len, int32 i, uint32* ps1, uint32* ps2)
+{
+	return get_checksum1_avx2_64(buf, len, i, ps1, ps2);
+}
 
 #endif /* } !USE_ROLL_ASM */
 
@@ -461,11 +467,7 @@ static inline uint32 get_checksum1_cpp(char *buf1, int32 len)
     uint32 s2 = 0;
 
     // multiples of 64 bytes using AVX2 (if available)
-#ifdef USE_ROLL_ASM
-    i = get_checksum1_avx2_asm((schar*)buf1, len, i, &s1, &s2);
-#else
-    i = get_checksum1_avx2_64((schar*)buf1, len, i, &s1, &s2);
-#endif
+    i = get_checksum1_avx2((schar*)buf1, len, i, &s1, &s2);
 
     // multiples of 32 bytes using SSSE3 (if available)
     i = get_checksum1_ssse3_32((schar*)buf1, len, i, &s1, &s2);
@@ -534,11 +536,7 @@ int main() {
     benchmark("Raw-C", get_checksum1_default_1, (schar*)buf, BLOCK_LEN);
     benchmark("SSE2", get_checksum1_sse2_32, (schar*)buf, BLOCK_LEN);
     benchmark("SSSE3", get_checksum1_ssse3_32, (schar*)buf, BLOCK_LEN);
-#ifdef USE_ROLL_ASM
-    benchmark("AVX2-ASM", get_checksum1_avx2_asm, (schar*)buf, BLOCK_LEN);
-#else
-    benchmark("AVX2", get_checksum1_avx2_64, (schar*)buf, BLOCK_LEN);
-#endif
+    benchmark("AVX2", get_checksum1_avx2, (schar*)buf, BLOCK_LEN);
 
     free(buf);
     return 0;