diff --git a/include/bx/bx.h b/include/bx/bx.h
index 161270a..ea97d0d 100644
--- a/include/bx/bx.h
+++ b/include/bx/bx.h
@@ -15,40 +15,37 @@
 #include "config.h"
 #include "macros.h"
 
-namespace bx
-{
-	// http://cnicholson.net/2011/01/stupid-c-tricks-a-better-sizeof_array/
-	template<typename T, size_t N> char (&COUNTOF_REQUIRES_ARRAY_ARGUMENT(const T(&)[N]) )[N];
+///
 #define BX_COUNTOF(_x) sizeof(bx::COUNTOF_REQUIRES_ARRAY_ARGUMENT(_x) )
 
-	// Template for avoiding MSVC: C4127: conditional expression is constant
-	template<bool>
-	inline bool isEnabled()
-	{
-		return true;
-	}
-
-	template<>
-	inline bool isEnabled<false>()
-	{
-		return false;
-	}
-#define BX_ENABLED(_x) bx::isEnabled<!!(_x)>()
-
-	inline bool ignoreC4127(bool _x)
-	{
-		return _x;
-	}
+///
 #define BX_IGNORE_C4127(_x) bx::ignoreC4127(!!(_x) )
 
-	template<typename Ty>
-	inline void xchg(Ty& _a, Ty& _b)
-	{
-		Ty tmp = _a; _a = _b; _b = tmp;
-	}
+///
+#define BX_ENABLED(_x) bx::isEnabled<!!(_x)>()
+
+namespace bx
+{
+	/// Template for avoiding MSVC: C4127: conditional expression is constant
+	template<bool>
+	bool isEnabled();
 
 	///
-	void* memCopy(void* _dst, const void* _src, size_t _numBytes);
+	bool ignoreC4127(bool _x);
+
+	///
+	template<typename Ty>
+	void xchg(Ty& _a, Ty& _b);
+
+	///
+	void xchg(void* _a, void* _b, size_t _numBytes);
+
+	// http://cnicholson.net/2011/01/stupid-c-tricks-a-better-sizeof_array/
+	template<typename T, size_t N>
+	char (&COUNTOF_REQUIRES_ARRAY_ARGUMENT(const T(&)[N]) )[N];
+
+	///
+	void memCopy(void* _dst, const void* _src, size_t _numBytes);
 
 	///
 	void memCopy(void* _dst, const void* _src, uint32_t _size, uint32_t _num, uint32_t _srcPitch, uint32_t _dstPitch);
@@ -60,11 +57,13 @@ namespace bx
 	void scatter(void* _dst, const void* _src, uint32_t _size, uint32_t _num, uint32_t _dstPitch);
 
 	///
-	void* memMove(void* _dst, const void* _src, size_t _numBytes);
+	void memMove(void* _dst, const void* _src, size_t _numBytes);
 
 	///
-	void* memSet(void* _dst, uint8_t _ch, size_t _numBytes);
+	void memSet(void* _dst, uint8_t _ch, size_t _numBytes);
 
 } // namespace bx
 
+#include "bx.inl"
+
 #endif // BX_H_HEADER_GUARD
diff --git a/include/bx/bx.inl b/include/bx/bx.inl
new file mode 100644
index 0000000..f8e9c63
--- /dev/null
+++ b/include/bx/bx.inl
@@ -0,0 +1,35 @@
+/*
+ * Copyright 2010-2017 Branimir Karadzic. All rights reserved.
+ * License: https://github.com/bkaradzic/bx#license-bsd-2-clause
+ */
+
+#ifndef BX_H_HEADER_GUARD
+#	error "Must be included from bx/bx.h!"
+#endif // BX_H_HEADER_GUARD
+
+namespace bx
+{
+	template<bool>
+	inline bool isEnabled()
+	{
+		return true;
+	}
+
+	template<>
+	inline bool isEnabled<false>()
+	{
+		return false;
+	}
+
+	inline bool ignoreC4127(bool _x)
+	{
+		return _x;
+	}
+
+	template<typename Ty>
+	inline void xchg(Ty& _a, Ty& _b)
+	{
+		Ty tmp = _a; _a = _b; _b = tmp;
+	}
+
+} // namespace bx
diff --git a/src/crt.cpp b/src/crt.cpp
index fbc91c7..66c09bb 100644
--- a/src/crt.cpp
+++ b/src/crt.cpp
@@ -13,7 +13,18 @@
 
 namespace bx
 {
-	void* memCopyRef(void* _dst, const void* _src, size_t _numBytes)
+	void xchg(void* _a, void* _b, size_t _numBytes)
+	{
+		uint8_t* lhs = (uint8_t*)_a;
+		uint8_t* rhs = (uint8_t*)_b;
+		const uint8_t* end = rhs + _numBytes;
+		while (rhs != end)
+		{
+			xchg(*lhs++, *rhs++);
+		}
+	}
+
+	void memCopyRef(void* _dst, const void* _src, size_t _numBytes)
 	{
 		uint8_t* dst = (uint8_t*)_dst;
 		const uint8_t* end = dst + _numBytes;
@@ -22,16 +33,14 @@ namespace bx
 		{
 			*dst++ = *src++;
 		}
-
-		return _dst;
 	}
 
-	void* memCopy(void* _dst, const void* _src, size_t _numBytes)
+	void memCopy(void* _dst, const void* _src, size_t _numBytes)
 	{
 #if BX_CRT_NONE
-		return memCopyRef(_dst, _src, _numBytes);
+		memCopyRef(_dst, _src, _numBytes);
 #else
-		return ::memcpy(_dst, _src, _numBytes);
+		::memcpy(_dst, _src, _numBytes);
 #endif // BX_CRT_NONE
 	}
 
@@ -60,7 +69,7 @@ namespace bx
 		memCopy(_dst, _src, _size, _num, _size, _dstPitch);
 	}
 
-	void* memMoveRef(void* _dst, const void* _src, size_t _numBytes)
+	void memMoveRef(void* _dst, const void* _src, size_t _numBytes)
 	{
 		uint8_t* dst = (uint8_t*)_dst;
 		const uint8_t* src = (const uint8_t*)_src;
@@ -68,33 +77,32 @@ namespace bx
 		if (_numBytes == 0
 		||  dst == src)
 		{
-			return dst;
+			return;
 		}
 
 		//	if (src+_numBytes <= dst || end <= src)
 		if (dst < src)
 		{
-			return memCopy(_dst, _src, _numBytes);
+			memCopy(_dst, _src, _numBytes);
+			return;
 		}
 
 		for (intptr_t ii = _numBytes-1; ii >= 0; --ii)
 		{
 			dst[ii] = src[ii];
 		}
-
-		return _dst;
 	}
 
-	void* memMove(void* _dst, const void* _src, size_t _numBytes)
+	void memMove(void* _dst, const void* _src, size_t _numBytes)
 	{
 #if BX_CRT_NONE
-		return memMoveRef(_dst, _src, _numBytes);
+		memMoveRef(_dst, _src, _numBytes);
 #else
-		return ::memmove(_dst, _src, _numBytes);
+		::memmove(_dst, _src, _numBytes);
 #endif // BX_CRT_NONE
 	}
 
-	void* memSetRef(void* _dst, uint8_t _ch, size_t _numBytes)
+	void memSetRef(void* _dst, uint8_t _ch, size_t _numBytes)
 	{
 		uint8_t* dst = (uint8_t*)_dst;
 		const uint8_t* end = dst + _numBytes;
@@ -102,16 +110,14 @@ namespace bx
 		{
 			*dst++ = char(_ch);
 		}
-
-		return _dst;
 	}
 
-	void* memSet(void* _dst, uint8_t _ch, size_t _numBytes)
+	void memSet(void* _dst, uint8_t _ch, size_t _numBytes)
 	{
 #if BX_CRT_NONE
-		return memSetRef(_dst, _ch, _numBytes);
+		memSetRef(_dst, _ch, _numBytes);
 #else
-		return ::memset(_dst, _ch, _numBytes);
+		::memset(_dst, _ch, _numBytes);
 #endif // BX_CRT_NONE
 	}
 
@@ -396,16 +402,19 @@ namespace bx
 #if BX_CRT_NONE
 extern "C" void* memcpy(void* _dst, const void* _src, size_t _numBytes)
 {
-	return bx::memCopy(_dst, _src, _numBytes);
+	bx::memCopy(_dst, _src, _numBytes);
+	return _dst;
 }
 
 extern "C" void* memmove(void* _dst, const void* _src, size_t _numBytes)
 {
-	return bx::memMove(_dst, _src, _numBytes);
+	bx::memMove(_dst, _src, _numBytes);
+	return _dst;
 }
 
 extern "C" void* memset(void* _dst, int _ch, size_t _numBytes)
 {
-	return bx::memSet(_dst, uint8_t(_ch), _numBytes);
+	bx::memSet(_dst, uint8_t(_ch), _numBytes);
+	return _dst;
 }
 #endif // BX_CRT_NONE
diff --git a/tests/crt_test.cpp b/tests/crt_test.cpp
index 7c4688c..fedb925 100644
--- a/tests/crt_test.cpp
+++ b/tests/crt_test.cpp
@@ -10,10 +10,10 @@ TEST_CASE("memSet", "")
 {
 	char temp[] =  { 1, 2, 3, 4, 5, 6, 7, 8, 9, 0 };
 
-	REQUIRE(temp == bx::memSet(temp, 0, 0) );
+	bx::memSet(temp, 0, 0);
 	REQUIRE(temp[0] == 1);
 
-	REQUIRE(temp == bx::memSet(temp, 0, 5) );
+	bx::memSet(temp, 0, 5);
 	REQUIRE(temp[0] == 0);
 	REQUIRE(temp[1] == 0);
 	REQUIRE(temp[2] == 0);
@@ -24,17 +24,21 @@ TEST_CASE("memSet", "")
 
 TEST_CASE("memMove", "")
 {
+	const char* orignal = "xxxxabvgd";
 	char str[] = { 'x', 'x', 'x', 'x', 'a', 'b', 'v', 'g', 'd' };
 
-	REQUIRE(bx::memMove(&str[4], &str[4], 0) );
-	REQUIRE(bx::memMove(&str[4], &str[4], 5) );
+	bx::memMove(&str[4], &str[4], 0);
+	REQUIRE(0 == bx::strncmp(str, orignal) );
 
-	REQUIRE(&str[0] == bx::memMove(str, &str[4], 5) );
-	REQUIRE(      0 == bx::strncmp(str, "abvgd", 5) );
+	bx::memMove(&str[4], &str[4], 5);
+	REQUIRE(0 == bx::strncmp(str, orignal) );
 
-	REQUIRE(&str[4] == bx::memMove(&str[4], str, 5) );
-	REQUIRE( str[4] == 'a' );
+	bx::memMove(str, &str[4], 5);
+	REQUIRE(0 == bx::strncmp(str, "abvgd", 5) );
+
+	bx::memMove(&str[4], str, 5);
+	REQUIRE(str[4] == 'a' );
 
 	bx::memSet(str, 'x', 4);
-	REQUIRE(0 == bx::strncmp(str, "xxxxabvgd", 9) );
+	REQUIRE(0 == bx::strncmp(str, orignal) );
 }