From 727865dcc538bac378331e15ff2b1b89b2ca47e7 Mon Sep 17 00:00:00 2001
From: hyung-hwan <hyunghwan.chung@gmail.com>
Date: Fri, 3 May 2019 07:39:01 +0000
Subject: [PATCH] enhanced hcl_bswap16/32/64()

---
 lib/hcl-cmn.h | 24 ++++++++++++++++++++++++
 lib/hcl-utl.h | 22 ++++++++++++++++++++++
 2 files changed, 46 insertions(+)

diff --git a/lib/hcl-cmn.h b/lib/hcl-cmn.h
index 8e72ff1..3ea7208 100644
--- a/lib/hcl-cmn.h
+++ b/lib/hcl-cmn.h
@@ -912,6 +912,19 @@ typedef struct hcl_t hcl_t;
 		#define HCL_HAVE_SYNC_VAL_COMPARE_AND_SWAP
 	#endif
 
+	#if __has_builtin(__builtin_bswap16)
+		#define HCL_HAVE_BUILTIN_BSWAP16
+	#endif
+	#if __has_builtin(__builtin_bswap32)
+		#define HCL_HAVE_BUILTIN_BSWAP32
+	#endif
+	#if __has_builtin(__builtin_bswap64)
+		#define HCL_HAVE_BUILTIN_BSWAP64
+	#endif
+	#if __has_builtin(__builtin_bswap128)
+		#define HCL_HAVE_BUILTIN_BSWAP128
+	#endif
+
 #elif defined(__GNUC__) && defined(__GNUC_MINOR__)
 
 	#if (__GNUC__ >= 4) 
@@ -949,6 +962,17 @@ typedef struct hcl_t hcl_t;
 		#define HCL_HAVE_BUILTIN_SMULLL_OVERFLOW
 	#endif
 
+	#if (__GNUC__ >= 5) || (__GNUC__ == 4 && __GNUC_MINOR__ >= 8)
+		/* 4.8.0 or later */
+		#define HCL_HAVE_BUILTIN_BSWAP16
+	#endif
+	#if (__GNUC__ >= 5) || (__GNUC__ == 4 && __GNUC_MINOR__ >= 3)
+		/* 4.3.0 or later */
+		#define HCL_HAVE_BUILTIN_BSWAP32
+		#define HCL_HAVE_BUILTIN_BSWAP64
+		/*#define HCL_HAVE_BUILTIN_BSWAP128*/
+	#endif
+
 #endif
 
 #if defined(HCL_HAVE_BUILTIN_EXPECT)
diff --git a/lib/hcl-utl.h b/lib/hcl-utl.h
index c427fba..8fd60f9 100644
--- a/lib/hcl-utl.h
+++ b/lib/hcl-utl.h
@@ -692,23 +692,43 @@ HCL_EXPORT int hcl_ucwidth (
 #if defined(HCL_HAVE_UINT16_T)
 static HCL_INLINE hcl_uint16_t hcl_bswap16 (hcl_uint16_t x)
 {
+#if defined(HCL_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(HCL_HAVE_UINT32_T)
 static HCL_INLINE hcl_uint32_t hcl_bswap32 (hcl_uint32_t x)
 {
+#if defined(HCL_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) & ((hcl_uint32_t)0xff << 8)) | 
 	       ((x <<  8) & ((hcl_uint32_t)0xff << 16)) | 
 	       ((x << 24));
+#endif
 }
 #endif
 
 #if defined(HCL_HAVE_UINT64_T)
 static HCL_INLINE hcl_uint64_t hcl_bswap64 (hcl_uint64_t x)
 {
+#if defined(HCL_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) & ((hcl_uint64_t)0xff << 8)) | 
 	       ((x >> 24) & ((hcl_uint64_t)0xff << 16)) | 
@@ -717,9 +737,11 @@ static HCL_INLINE hcl_uint64_t hcl_bswap64 (hcl_uint64_t x)
 	       ((x << 24) & ((hcl_uint64_t)0xff << 40)) | 
 	       ((x << 40) & ((hcl_uint64_t)0xff << 48)) | 
 	       ((x << 56));
+#endif
 }
 #endif
 
+
 #if defined(HCL_HAVE_UINT128_T)
 static HCL_INLINE hcl_uint128_t hcl_bswap128 (hcl_uint128_t x)
 {