The "union hack" is treacherous and needs far more investigation
than it appears.
So treacherous that people have been managing to use it for the last 30 years or so?
Are we going to mandate that the reintepretation of
the bits cannot possibly yield trap representation?
I wouldn't. This is for non-portable machine-specific code. If you blow a trap representation or an alignment access, you are on your own.
If I wanted to do this via copying, I'd just use Java.
Assuming no alignment, packing, endian or malformed packet issues, here is an example of what people want to do:
enum class Protocol : char
{
UDP = 17,
// ...
};
struct IPHeader
{
//...
Protocol protocol;
// ...
};
struct UDPHeader : IPHeader
{
//...
uint16_t length;
//...
char data[1];
};
union Header
{
IPHeader ip;
UDPHeader udp;
//...
};
// Deliver payload to f without copying
Payload(dynarray<char> const& packet, std::function<void(char const*, char const*)> const& f)
{
// Want to overlay Header on packet.data()
auto& header = *static_cast<Header const*>(static_cast<void const*>(packet.data()));
switch (header.ip.protocol)
{
case Protocol::UDP:
f(&header.udp.data[0], &header.udp.data[header.udp.length - 8]);
break;
//...
}
}
Q1: How many places has undefined behavior been invoked in the above?
Q2: What is the correct way to write this code?