Fixed cnt* overloads when uint64_t and size_t are not the same type.

include/bx/inline/uint32_t.inl

@@ -336,7 +336,7 @@ namespace bx
 	}
 
 	template<>
-	inline BX_CONSTEXPR_FUNC uint32_t uint32_cntbits(uint64_t _val)
+	inline BX_CONSTEXPR_FUNC uint32_t uint32_cntbits(unsigned long long _val)
 	{
 #if BX_COMPILER_GCC || BX_COMPILER_CLANG
 		return __builtin_popcountll(_val);
@@ -344,12 +344,18 @@ namespace bx
 		const uint32_t lo = uint32_t(_val&UINT32_MAX);
 		const uint32_t hi = uint32_t(_val>>32);
 
-		const uint32_t total = uint32_cntbits(lo)
-							 + uint32_cntbits(hi);
-		return total;
+		return uint32_cntbits(lo)
+			+  uint32_cntbits(hi)
+			;
 #endif // BX_COMPILER_*
 	}
 
+	template<>
+	inline BX_CONSTEXPR_FUNC uint32_t uint32_cntbits(unsigned long _val)
+	{
+		return uint32_cntbits<unsigned long long>(_val);
+	}
+
 	template<> inline BX_CONSTEXPR_FUNC uint32_t uint32_cntbits(uint8_t  _val) { return uint32_cntbits<uint32_t>(_val); }
 	template<> inline BX_CONSTEXPR_FUNC uint32_t uint32_cntbits(int8_t   _val) { return uint32_cntbits<uint8_t >(_val); }
 	template<> inline BX_CONSTEXPR_FUNC uint32_t uint32_cntbits(uint16_t _val) { return uint32_cntbits<uint32_t>(_val); }
@@ -381,7 +387,7 @@ namespace bx
 	}
 
 	template<>
-	inline BX_CONSTEXPR_FUNC uint32_t uint32_cntlz(uint64_t _val)
+	inline BX_CONSTEXPR_FUNC uint32_t uint32_cntlz(unsigned long long _val)
 	{
 #if BX_COMPILER_GCC || BX_COMPILER_CLANG
 		return 0 == _val ? 64 : __builtin_clzll(_val);
@@ -393,6 +399,12 @@ namespace bx
 #endif // BX_COMPILER_*
 	}
 
+	template<>
+	inline BX_CONSTEXPR_FUNC uint32_t uint32_cntlz(unsigned long _val)
+	{
+		return uint32_cntlz<unsigned long long>(_val);
+	}
+
 	template<> inline BX_CONSTEXPR_FUNC uint32_t uint32_cntlz(uint8_t  _val) { return uint32_cntlz<uint32_t>(_val)-24; }
 	template<> inline BX_CONSTEXPR_FUNC uint32_t uint32_cntlz(int8_t   _val) { return uint32_cntlz<uint8_t >(_val);    }
 	template<> inline BX_CONSTEXPR_FUNC uint32_t uint32_cntlz(uint16_t _val) { return uint32_cntlz<uint32_t>(_val)-16; }
@@ -416,7 +428,7 @@ namespace bx
 	}
 
 	template<>
-	inline BX_CONSTEXPR_FUNC uint32_t uint32_cnttz(uint64_t _val)
+	inline BX_CONSTEXPR_FUNC uint32_t uint32_cnttz(unsigned long long _val)
 	{
 #if BX_COMPILER_GCC || BX_COMPILER_CLANG
 		return 0 == _val ? 64 : __builtin_ctzll(_val);
@@ -428,6 +440,12 @@ namespace bx
 #endif // BX_COMPILER_*
 	}
 
+	template<>
+	inline BX_CONSTEXPR_FUNC uint32_t uint32_cnttz(unsigned long _val)
+	{
+		return uint32_cnttz<unsigned long long>(_val);
+	}
+
 	template<> inline BX_CONSTEXPR_FUNC uint32_t uint32_cnttz(uint8_t  _val) { return bx::min(8u,  uint32_cnttz<uint32_t>(_val) ); }
 	template<> inline BX_CONSTEXPR_FUNC uint32_t uint32_cnttz(int8_t   _val) { return              uint32_cnttz<uint8_t >(_val);   }
 	template<> inline BX_CONSTEXPR_FUNC uint32_t uint32_cnttz(uint16_t _val) { return bx::min(16u, uint32_cnttz<uint32_t>(_val) ); }