diff --git a/config/compilers/compiler_clang.h.in b/config/compilers/compiler_clang.h.in
index 91f8398ef4..f62824ef0e 100644
--- a/config/compilers/compiler_clang.h.in
+++ b/config/compilers/compiler_clang.h.in
@@ -62,3 +62,15 @@
 
 #undef DUK_USE_GCC_PRAGMAS
 #define DUK_USE_PACK_CLANG_ATTR
+
+#if defined(__clang__) && defined(__has_builtin)
+#if __has_builtin(__builtin_bswap64)
+#define DUK_BSWAP64(x) ((duk_uint64_t) __builtin_bswap64((duk_uint64_t) (x)))
+#endif
+#if __has_builtin(__builtin_bswap32)
+#define DUK_BSWAP32(x) ((duk_uint32_t) __builtin_bswap32((duk_uint32_t) (x)))
+#endif
+#if __has_builtin(__builtin_bswap16)
+#define DUK_BSWAP16(x) ((duk_uint16_t) __builtin_bswap16((duk_uint16_t) (x)))
+#endif
+#endif
diff --git a/config/compilers/compiler_gcc.h.in b/config/compilers/compiler_gcc.h.in
index 0a9b7d7af6..9e7f62048a 100644
--- a/config/compilers/compiler_gcc.h.in
+++ b/config/compilers/compiler_gcc.h.in
@@ -82,3 +82,13 @@
 #endif
 
 #define DUK_USE_PACK_GCC_ATTR
+
+/* Availability varies based on platform (between GCC 4.4 and 4.8), and there
+ * are apparently some bugs in GCC 4.x.  Check for GCC 5.0 before enabling
+ * these to be safe.
+ */
+#if defined(DUK_F_GCC_VERSION) && (DUK_F_GCC_VERSION >= 50000L)
+#define DUK_BSWAP64(x) ((duk_uint64_t) __builtin_bswap64((duk_uint64_t) (x)))
+#define DUK_BSWAP32(x) ((duk_uint32_t) __builtin_bswap32((duk_uint32_t) (x)))
+#define DUK_BSWAP16(x) ((duk_uint16_t) __builtin_bswap16((duk_uint16_t) (x)))
+#endif
diff --git a/config/header-snippets/compiler_fillins.h.in b/config/header-snippets/compiler_fillins.h.in
index 706b539633..431a87981e 100644
--- a/config/header-snippets/compiler_fillins.h.in
+++ b/config/header-snippets/compiler_fillins.h.in
@@ -1,3 +1,10 @@
+#if !defined(DUK_U64_CONSTANT)
+#define DUK_U64_CONSTANT(x) x##ULL
+#endif
+#if !defined(DUK_I64_CONSTANT)
+#define DUK_I64_CONSTANT(x) x##LL
+#endif
+
 #if !defined(DUK_VA_COPY)
 /* We need va_copy() which is defined in C99 / C++11, so an awkward
  * replacement is needed for pre-C99 / pre-C++11 environments.  This
@@ -148,17 +155,30 @@
 #endif
 #endif
 
+#if defined(DUK_F_HAVE_64BIT)
+#if !defined(DUK_BSWAP64)
+#define DUK_BSWAP64(x) \
+	((((duk_uint64_t) (x)) >> 56U) | \
+	 ((((duk_uint64_t) (x)) >> 40U) & DUK_U64_CONSTANT(0xff00)) | \
+	 ((((duk_uint64_t) (x)) >> 24U) & DUK_U64_CONSTANT(0xff0000)) | \
+	 ((((duk_uint64_t) (x)) >> 8U) & DUK_U64_CONSTANT(0xff000000)) | \
+	 ((((duk_uint64_t) (x)) << 8U) & DUK_U64_CONSTANT(0xff00000000)) | \
+	 ((((duk_uint64_t) (x)) << 24U) & DUK_U64_CONSTANT(0xff0000000000)) | \
+	 ((((duk_uint64_t) (x)) << 40U) & DUK_U64_CONSTANT(0xff000000000000)) | \
+	 (((duk_uint64_t) (x)) << 56U))
+#endif
+#endif
 #if !defined(DUK_BSWAP32)
 #define DUK_BSWAP32(x) \
-	((((duk_uint32_t) (x)) >> 24) | \
-	 ((((duk_uint32_t) (x)) >> 8) & 0xff00UL) | \
-	 ((((duk_uint32_t) (x)) << 8) & 0xff0000UL) | \
-	 (((duk_uint32_t) (x)) << 24))
+	((((duk_uint32_t) (x)) >> 24U) | \
+	 ((((duk_uint32_t) (x)) >> 8U) & 0xff00UL) | \
+	 ((((duk_uint32_t) (x)) << 8U) & 0xff0000UL) | \
+	 (((duk_uint32_t) (x)) << 24U))
 #endif
 #if !defined(DUK_BSWAP16)
 #define DUK_BSWAP16(x) \
-	((duk_uint16_t) (x) >> 8) | \
-	((duk_uint16_t) (x) << 8)
+	((duk_uint16_t) (x) >> 8U) | \
+	((duk_uint16_t) (x) << 8U)
 #endif
 
 /* DUK_USE_VARIADIC_MACROS: required from compilers, so no fill-in. */
@@ -180,10 +200,3 @@
 #if 0  /* not defined by default */
 #undef DUK_USE_GCC_PRAGMAS
 #endif
-
-#if !defined(DUK_U64_CONSTANT)
-#define DUK_U64_CONSTANT(x) x##ULL
-#endif
-#if !defined(DUK_I64_CONSTANT)
-#define DUK_I64_CONSTANT(x) x##LL
-#endif
diff --git a/src-input/duk_selftest.c b/src-input/duk_selftest.c
index 5c759464a6..db3665761d 100644
--- a/src-input/duk_selftest.c
+++ b/src-input/duk_selftest.c
@@ -187,10 +187,28 @@ DUK_LOCAL duk_uint_t duk__selftest_byte_order(void) {
 
 DUK_LOCAL duk_uint_t duk__selftest_bswap_macros(void) {
 	duk_uint_t error_count = 0;
+	volatile duk_uint32_t x32_input, x32_output;
 	duk_uint32_t x32;
+	volatile duk_uint16_t x16_input, x16_output;
 	duk_uint16_t x16;
 	duk_double_union du;
 	duk_double_t du_diff;
+#if defined(DUK_BSWAP64)
+	volatile duk_uint64_t x64_input, x64_output;
+	duk_uint64_t x64;
+#endif
+
+	/* Cover both compile time and runtime bswap operations, as these
+	 * may have different bugs.
+	 */
+
+	x16_input = 0xbeefUL;
+	x16 = x16_input;
+	x16 = DUK_BSWAP16(x16);
+	x16_output = x16;
+	if (x16_output != (duk_uint16_t) 0xefbeUL) {
+		DUK__FAILED("DUK_BSWAP16");
+	}
 
 	x16 = 0xbeefUL;
 	x16 = DUK_BSWAP16(x16);
@@ -198,12 +216,36 @@ DUK_LOCAL duk_uint_t duk__selftest_bswap_macros(void) {
 		DUK__FAILED("DUK_BSWAP16");
 	}
 
+	x32_input = 0xdeadbeefUL;
+	x32 = x32_input;
+	x32 = DUK_BSWAP32(x32);
+	x32_output = x32;
+	if (x32_output != (duk_uint32_t) 0xefbeaddeUL) {
+		DUK__FAILED("DUK_BSWAP32");
+	}
+
 	x32 = 0xdeadbeefUL;
 	x32 = DUK_BSWAP32(x32);
 	if (x32 != (duk_uint32_t) 0xefbeaddeUL) {
 		DUK__FAILED("DUK_BSWAP32");
 	}
 
+#if defined(DUK_BSWAP64)
+	x64_input = DUK_U64_CONSTANT(0x8899aabbccddeeff);
+	x64 = x64_input;
+	x64 = DUK_BSWAP64(x64);
+	x64_output = x64;
+	if (x64_output != (duk_uint64_t) DUK_U64_CONSTANT(0xffeeddccbbaa9988)) {
+		DUK__FAILED("DUK_BSWAP64");
+	}
+
+	x64 = DUK_U64_CONSTANT(0x8899aabbccddeeff);
+	x64 = DUK_BSWAP64(x64);
+	if (x64 != (duk_uint64_t) DUK_U64_CONSTANT(0xffeeddccbbaa9988)) {
+		DUK__FAILED("DUK_BSWAP64");
+	}
+#endif
+
 	/* >>> struct.unpack('>d', '4000112233445566'.decode('hex'))
 	 * (2.008366013071895,)
 	 */