From 92121319976d96081428ec58e1164e6e1d6f116f Mon Sep 17 00:00:00 2001
From: hyung-hwan <hyunghwan.chung@gmail.com>
Date: Fri, 3 May 2019 07:36:53 +0000
Subject: [PATCH] enhanced mio_bswap16/32/64()

---
 mio/lib/mio-cmn.h | 23 +++++++++++++++++++++++
 mio/lib/mio-utl.h | 21 +++++++++++++++++++++
 2 files changed, 44 insertions(+)

diff --git a/mio/lib/mio-cmn.h b/mio/lib/mio-cmn.h
index faaf3ef..94d2fcc 100644
--- a/mio/lib/mio-cmn.h
+++ b/mio/lib/mio-cmn.h
@@ -802,6 +802,19 @@ struct mio_cmgr_t
 		#define MIO_HAVE_SYNC_VAL_COMPARE_AND_SWAP
 	#endif
 
+	#if __has_builtin(__builtin_bswap16)
+		#define MIO_HAVE_BUILTIN_BSWAP16
+	#endif
+	#if __has_builtin(__builtin_bswap32)
+		#define MIO_HAVE_BUILTIN_BSWAP32
+	#endif
+	#if __has_builtin(__builtin_bswap64)
+		#define MIO_HAVE_BUILTIN_BSWAP64
+	#endif
+	#if __has_builtin(__builtin_bswap128)
+		#define MIO_HAVE_BUILTIN_BSWAP128
+	#endif
+
 #elif defined(__GNUC__) && defined(__GNUC_MINOR__)
 
 	#if (__GNUC__ >= 4) 
@@ -836,6 +849,16 @@ struct mio_cmgr_t
 		#define MIO_HAVE_BUILTIN_SMULLL_OVERFLOW
 	#endif
 
+	#if (__GNUC__ >= 5) || (__GNUC__ == 4 && __GNUC_MINOR__ >= 8)
+		/* 4.8.0 or later */
+		#define MIO_HAVE_BUILTIN_BSWAP16
+	#endif
+	#if (__GNUC__ >= 5) || (__GNUC__ == 4 && __GNUC_MINOR__ >= 3)
+		/* 4.3.0 or later */
+		#define MIO_HAVE_BUILTIN_BSWAP32
+		#define MIO_HAVE_BUILTIN_BSWAP64
+		/*#define MIO_HAVE_BUILTIN_BSWAP128*/
+	#endif
 #endif
 
 #if defined(MIO_HAVE_BUILTIN_EXPECT)
diff --git a/mio/lib/mio-utl.h b/mio/lib/mio-utl.h
index b1e6a5d..897f8ac 100644
--- a/mio/lib/mio-utl.h
+++ b/mio/lib/mio-utl.h
@@ -517,23 +517,43 @@ MIO_EXPORT mio_oow_t mio_utf8_to_uc (
 #if defined(MIO_HAVE_UINT16_T)
 static MIO_INLINE mio_uint16_t mio_bswap16 (mio_uint16_t x)
 {
+#if defined(MIO_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));
+	return x;
+#else
 	return (x << 8) | (x >> 8);
+#endif
 }
 #endif
 
 #if defined(MIO_HAVE_UINT32_T)
 static MIO_INLINE mio_uint32_t mio_bswap32 (mio_uint32_t x)
 {
+#if defined(MIO_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));
+	return x;
+#else
 	return ((x >> 24)) | 
 	       ((x >>  8) & ((mio_uint32_t)0xff << 8)) | 
 	       ((x <<  8) & ((mio_uint32_t)0xff << 16)) | 
 	       ((x << 24));
+#endif
 }
 #endif
 
 #if defined(MIO_HAVE_UINT64_T)
 static MIO_INLINE mio_uint64_t mio_bswap64 (mio_uint64_t x)
 {
+#if defined(MIO_HAVE_BUILTIN_BSWAP64)
+	return __builtin_bswap64(x);
+#elif defined(__GNUC__) && (defined(__x86_64) || defined(__amd64))
+	__asm__ volatile ("bswapq %0" : "=r"(x) : "0"(x));
+	return x;
+#else
 	return ((x >> 56)) | 
 	       ((x >> 40) & ((mio_uint64_t)0xff << 8)) | 
 	       ((x >> 24) & ((mio_uint64_t)0xff << 16)) | 
@@ -542,6 +562,7 @@ static MIO_INLINE mio_uint64_t mio_bswap64 (mio_uint64_t x)
 	       ((x << 24) & ((mio_uint64_t)0xff << 40)) | 
 	       ((x << 40) & ((mio_uint64_t)0xff << 48)) | 
 	       ((x << 56));
+#endif
 }
 #endif