From 80b05e9d95e24c00da2eec8e4271fc95f1a3938c Mon Sep 17 00:00:00 2001
From: "hyunghwan.chung" <hyunghwan.chung@c9d47f8d-4fed-4a2f-8cad-691f5fbe3904>
Date: Sat, 4 May 2019 02:16:25 +0000
Subject: [PATCH] used arm inline assembly in some byte swap functions

---
 moo/lib/bigint.c  |  4 ++--
 moo/lib/moo-cmn.h | 17 +++++++++++++++++
 moo/lib/moo-utl.h | 22 +++++++++++++++++++---
 3 files changed, 38 insertions(+), 5 deletions(-)

diff --git a/moo/lib/bigint.c b/moo/lib/bigint.c
index 74812db..5abc831 100644
--- a/moo/lib/bigint.c
+++ b/moo/lib/bigint.c
@@ -161,7 +161,7 @@ static MOO_INLINE int get_pos_of_msb_set_pow2 (moo_oow_t x)
 	);
 #endif
 	return (int)pos;
-#elif defined(__GNUC__) && defined(__arm__) && (defined(__ARM_ARCH_5__) || defined(__ARM_ARCH_6__) || defined(__ARM_ARCH_7__) || defined(__ARM_ARCH_8__))
+#elif defined(__GNUC__) && defined(__arm__) && (defined(__ARM_ARCH) && (__ARM_ARCH >= 5))
 	moo_oow_t n;
 	/* CLZ is available in ARMv5T and above. there is no instruction to
 	 * count trailing zeros or something similar. using RBIT with CLZ
@@ -199,7 +199,7 @@ static MOO_INLINE int get_pos_of_msb_set (moo_oow_t x)
 		: "r"(x) /* input */
 	);
 	return (int)pos;
-#elif defined(__GNUC__) && defined(__arm__) && (defined(__ARM_ARCH_5__) || defined(__ARM_ARCH_6__) || defined(__ARM_ARCH_7__) || defined(__ARM_ARCH_8__))
+#elif defined(__GNUC__) && defined(__arm__) && (defined(__ARM_ARCH) && (__ARM_ARCH >= 5))
 	moo_oow_t n;
 	__asm__ volatile (
 		"clz %0,%1\n\t"
diff --git a/moo/lib/moo-cmn.h b/moo/lib/moo-cmn.h
index 6f3f227..de27633 100644
--- a/moo/lib/moo-cmn.h
+++ b/moo/lib/moo-cmn.h
@@ -45,6 +45,11 @@
 #	error UNSUPPORTED SYSTEM
 #endif
 
+
+/* =========================================================================
+ * ARCHITECTURE/COMPILER TWEAKS
+ * ========================================================================= */
+
 #if defined(EMSCRIPTEN)
 #	if defined(MOO_SIZEOF___INT128)
 #		undef MOO_SIZEOF___INT128 
@@ -57,6 +62,18 @@
 #	endif
 #endif
 
+#if defined(__GNUC__) && defined(__arm__)  && !defined(__ARM_ARCH)
+#	if defined(__ARM_ARCH_8__)
+#		define __ARM_ARCH 8
+#	elif defined(__ARM_ARCH_7__)
+#		define __ARM_ARCH 7
+#	elif defined(__ARM_ARCH_6__)
+#		define __ARM_ARCH 6
+#	elif defined(__ARM_ARCH_5__)
+#		define __ARM_ARCH 5
+#	endif
+#endif
+
 /* =========================================================================
  * PRIMITIVE TYPE DEFINTIONS
  * ========================================================================= */
diff --git a/moo/lib/moo-utl.h b/moo/lib/moo-utl.h
index 509e179..a7a8278 100644
--- a/moo/lib/moo-utl.h
+++ b/moo/lib/moo-utl.h
@@ -746,8 +746,10 @@ static MOO_INLINE moo_uint16_t moo_bswap16 (moo_uint16_t x)
 #if defined(MOO_HAVE_BUILTIN_BSWAP16)
 	return __builtin_bswap16(x);
 #elif defined(__GNUC__) && (defined(__x86_64) || defined(__amd64) || defined(__i386) || defined(i386))
-	__asm__ volatile ("xchgb %b0, %h0" : "=Q"(x): "0"(x));
+	__asm__ /*volatile*/ ("xchgb %b0, %h0" : "=Q"(x): "0"(x));
 	return x;
+#elif defined(__GNUC__) && (defined(__ARM_ARCH) && (__ARM_ARCH >= 6))
+	__asm__ /*volatile*/ ("rev16 %0, %0" : "+r"(x));
 #else
 	return (x << 8) | (x >> 8);
 #endif
@@ -760,7 +762,19 @@ static MOO_INLINE moo_uint32_t moo_bswap32 (moo_uint32_t x)
 #if defined(MOO_HAVE_BUILTIN_BSWAP32)
 	return __builtin_bswap32(x);
 #elif defined(__GNUC__) && (defined(__x86_64) || defined(__amd64) || defined(__i386) || defined(i386))
-	__asm__ volatile ("bswapl %0" : "=r"(x) : "0"(x));
+	__asm__ /*volatile*/ ("bswapl %0" : "=r"(x) : "0"(x));
+	return x;
+#elif defined(__GNUC__) && (defined(__ARM_ARCH) && (__ARM_ARCH >= 6))
+	__asm__ /*volatile*/ ("rev32 %0, %0" : "+r"(x));
+#elif defined(__GNUC__) && defined(__ARM_ARCH)
+	moo_uint32_t tmp;
+	__asm__ /*volatile*/ (
+		"eor %1, %0, %0, ror #16\n\t"
+		"bic %1, %1, #0x00ff0000\n\t"
+		"mov %0, %0, ror #8\n\t"
+		"eor %0, %0, %1, lsr #8\n\t"
+		:"+r"(x), "=&r"(tmp)
+	);
 	return x;
 #else
 	return ((x >> 24)) | 
@@ -777,8 +791,10 @@ static MOO_INLINE moo_uint64_t moo_bswap64 (moo_uint64_t x)
 #if defined(MOO_HAVE_BUILTIN_BSWAP64)
 	return __builtin_bswap64(x);
 #elif defined(__GNUC__) && (defined(__x86_64) || defined(__amd64))
-	__asm__ volatile ("bswapq %0" : "=r"(x) : "0"(x));
+	__asm__ /*volatile*/ ("bswapq %0" : "=r"(x) : "0"(x));
 	return x;
+#elif defined(__GNUC__) && defined(__aarch64__)
+	__asm__ /*volatile*/ ("rev64 %0, %0" : "+r"(x));
 #else
 	return ((x >> 56)) | 
 	       ((x >> 40) & ((moo_uint64_t)0xff << 8)) |