diff --git a/cutypes.h b/cutypes.h
index a8e7277..6c23bb6 100644
--- a/cutypes.h
+++ b/cutypes.h
@@ -4,29 +4,27 @@
 #include <stddef.h>
 #include <stdint.h>
 
-enum cu_arch_e {
-    ARCH_UNKNOWN = 0,
-    x86_64,
-    x86_32,
-    ARM2,
-    ARM3,
-    ARM4T,
-    ARM5,
-    ARM6T2,
-    ARM6,
-    ARM7,
-    ARM7A,
-    ARM7R,
-    ARM7M,
-    ARM7S,
-    AARCH64,
-    MIPS,
-    SUPERH,
-    POWERPC,
-    POWERPC64,
-    SPARC,
-    M68K,
-};
+#define ARCH_UNKNOWN    0
+#define x86_64          1
+#define x86_32          2
+#define ARM2            3
+#define ARM3            4
+#define ARM4T           5
+#define ARM5            6
+#define ARM6T2          7
+#define ARM6            8
+#define ARM7            9
+#define ARM7A           10
+#define ARM7R           11
+#define ARM7M           12
+#define ARM7S           13
+#define AARCH64         14
+#define MIPS            15
+#define SUPERH          16
+#define POWERPC         17
+#define POWERPC64       18
+#define SPARC           19
+#define M68K            20
 
 #if defined(__x86_64__) || defined(_M_X64)
 #   define CU_ARCH x86_64
@@ -72,6 +70,25 @@ enum cu_arch_e {
 #   define CU_ARCH ARCH_UNKNOWN;
 #endif
 
+#if (CU_ARCH == x86_64 || CU_ARCH == x86_32)
+#   define CU_ARCH_X86
+#elif (CU_ARCH >= ARM2 && CU_ARCH <= AARCH64)
+#   define CU_ARCH_ARM
+#   define CU_ARCH_ARM_ARCH __ARM_ARCH
+#elif (CU_ARCH == MIPS)
+#   define CU_ARCH_MIPS
+#elif (CU_ARCH == SUPERH)
+#   define CU_ARCH_SUPERH
+#elif (CU_ARCH == POWERPC || CU_ARCH == POWERPC64)
+#   define CU_ARCH_POWERPC
+#elif (CU_ARCH == SPARC)
+#   define CU_ARCH_SPARC
+#elif (CU_ARCH == M68K)
+#   define CU_ARCH_M68K
+#else
+#   warning "unknown architecture"
+#endif
+
 #if (CU_ARCH == x86_64 || CU_ARCH == AARCH64 || CU_ARCH == POWERPC64)
 #   define CU_64BIT
 #else
diff --git a/endian.c b/endian.c
index ef83c48..b9b011f 100644
--- a/endian.c
+++ b/endian.c
@@ -46,11 +46,39 @@ static uint64_t sw_bswap64(uint64_t val)
         |   (val & 0x00000000000000FFUL) << 56;
 }
 
-#include <stdio.h>
+#if (CU_ARCH == x86_32)
+static uint32_t hw_bswap32(uint32_t val)
+{
+    __asm__ ("bswap %0" : "=r" (val) : "r" (val));
+    return val;
+}
+
+bswap16_t cu_bswap16 = sw_bswap16;
+bswap32_t cu_bswap32 = hw_bswap32;
+bswap64_t cu_bswap64 = sw_bswap64;
+#elif (CU_ARCH == x86_64)
+static uint32_t hw_bswap32(uint32_t val)
+{
+    __asm__ ("bswap %0" : "=r" (val) : "r" (val));
+    return val;
+}
+
+static uint64_t hw_bswap64(uint64_t val)
+{
+    __asm__ ("bswap %0" : "=r" (val) : "r" (val));
+    return val;
+}
+
+bswap16_t cu_bswap16 = sw_bswap16;
+bswap32_t cu_bswap32 = hw_bswap32;
+bswap64_t cu_bswap64 = hw_bswap64;
+#else
+bswap16_t cu_bswap16 = sw_bswap16;
+bswap32_t cu_bswap32 = sw_bswap32;
+bswap64_t cu_bswap64 = sw_bswap64;
+#endif
 
 void cu_endian_init()
 {
     cu_endian = endian_detect();
-    printf("%x\t%x\n", 0x12345678, sw_bswap32(0x12345678));
-    printf("%lx\t%lx\n", 0x12345678abcd8765, sw_bswap64(0x12345678abcd8765));
 }
diff --git a/endian.h b/endian.h
index 6d32640..9c5132b 100644
--- a/endian.h
+++ b/endian.h
@@ -11,6 +11,14 @@ enum cu_endian_e {
 
 extern enum cu_endian_e cu_endian;
 
+typedef uint16_t (*bswap16_t)(uint16_t);
+typedef uint32_t (*bswap32_t)(uint32_t);
+typedef uint64_t (*bswap64_t)(uint64_t);
+
+extern bswap16_t cu_bswap16;
+extern bswap32_t cu_bswap32;
+extern bswap64_t cu_bswap64;
+
 void cu_endian_init();
 
 #endif
\ No newline at end of file
diff --git a/test.c b/test.c
index 9a28daf..4047433 100644
--- a/test.c
+++ b/test.c
@@ -77,7 +77,11 @@ int main()
 
     printf("[cutil]\n");
     printf("cutil_arch\t%u\n", CU_ARCH);
+
+    printf("[endian]\n");
     printf("cutil_endian\t%u\n", cu_endian);
+    printf("%x\t%x\n", 0x12345678, cu_bswap32(0x12345678));
+    printf("%lx\t%lx\n", 0x12345678abcd8765, cu_bswap64(0x12345678abcd8765));
 
     cutil_exit();
     return 0;