Name Mangling Scheme

The hash values below are made up, but should eventually come from some decent algorithm for short hashes.

1. package/module
   • Fully qualified name of module is used, with each part mangled individually
   • Includes the mangled package name (e.g. std)
   • Hash for each part is created using "mod" + module name
   • :: becomes __
   • Example: std::num becomes std_493ef8a1__num_839a1b34
2. type
   • Mangled name of containing module + __ + mangled name of type
   • Hash for type is created using "type" + type name
   • Example: std::num::NonZeroU32 becomes std_493ef8a1__num_839a1b34__NonZeroU32_6743fa12
3. method
   • Mangled name of type + __ + unmangled name of method
   • Method names are unique for a given type so this works
   • Example: std::num::NonZeroU32::new becomes std_493ef8a1__num_839a1b34__NonZeroU32_6743fa12__new
4. function
   • Mangled name of containing module/function + __ + mangled name of function
   • Hash for function name is created using "fn" + function name
   • Don't want type names and function names to be able to conflict, so both are mangled with a different prefix for the hash
   • Example: std::io::copy becomes std_493ef8a1__io_15ad8c91__copy_ab77d31a
5. variable/function parameter
   • Hash of variable name
   • Needed to prevent collisions with temporaries

The extra string included in the hash is because different namespaces can contain the same names. That is, you can technically have a type and a function named Foo.

Short Hashes

use std::io::Read;

use adler32::adler32;

fn mangle_function_name(mod_mangled_name: &str, name: &str) -> String {
    let prefix = "fn".as_bytes();
    // https://doc.rust-lang.org/nightly/std/io/trait.Read.html#method.chain
    let mangle_input = prefix.chain(name.as_bytes());
    // unwrap() is safe because Read for &[u8] can never result in an error.
    let hash = adler32(mangle_input).unwrap();
    format!("{}__{}_{:x}", mod_mangled_name, name, hash)
}

Update: Apparently adler32 isn't great for short strings. Might just want to truncate md5 to 32-bits. This may have issues too.

from dino.

Even simpler scheme that just uses the length and not a hash: https://rust-lang.github.io/rfcs/2603-symbol-name-mangling-v2.html#the-mangling-scheme-by-example

from dino.