I've got the following C++ code :
template <int isBigEndian, typename val>
struct EndiannessConv
{
inline static val fromLittleEndianToHost( val v )
{
union
{
val outVal __attribute__ ((used));
uint8_t bytes[ sizeof( val ) ] __attribute__ ((used));
} ;
outVal = v;
std::reverse( &bytes[0], &bytes[ sizeof(val) ] );
return outVal;
}
inline static void convertArray( val v[], uint32_t size )
{
// TODO : find a way to map the array for (uint32_t i = 0; i < size; i++)
for (uint32_t i = 0; i < size; i++)
v[i] = fromLittleEndianToHost( v[i] );
}
};
Which work and has been tested (without the used attributes). When compiling I obtain the following errors from g++ (version 4.4.1)
|| g++ -Wall -Wextra -O3 -o t t.cc
|| t.cc: In static member function 'static val EndiannessConv<isBigEndian, val>::fromLittleEndianToHost(val)':
t.cc|98| warning: 'used' attribute ignored
t.cc|99| warning: 'used' attribute ignored
|| t.cc: In static member function 'static val EndiannessConv<isBigEndian, val>::fromLittleEndianToHost(val) [with int isBigEndian = 1, val = double]':
t.cc|148| instantiated from here
t.cc|100| warning: unused variable 'outVal'
t.cc|100| warning: unused variable 'bytes'
I've tried to use the following code :
template <int size, typename valType>
struct EndianInverser { /* should not compile */ };
template <typename valType>
struct EndianInverser<4, valType>
{
static inline valType reverseEndianness( const valType &val )
{
uint32_t castedVal =
*reinterpret_cast<const uint32_t*>( &val );
castedVal = (castedVal & 0x000000FF << (3 * 8))
| (castedVal & 0x0000FF00 << (1 * 8))
| (castedVal & 0x00FF0000 >> (1 * 8))
| (castedVal & 0xFF000000 >> (3 * 8));
return *reinterpret_cast<valType*>( &castedVal );
}
};
but it break when enabling optimizations due to the type punning.
So, why does my used
attribute got ignored?
Is there a workaround to convert endianness (I rely on the enum to avoid type punning) in templates?
I only have gcc 4.2.1 but if I get rid of the attribute ((used)) and give the union a name it compiles without warnings for me.
From what I've read the 'union' technique works on gcc but is not guaranteed in the standard, the other 'reinterpret_cast' method is wrong (because of type aliasing). However I think this applies to C, not sure about C++. Hope that helps.