Re: c++/7112: Regression: ICE on C++ code involving templates and sizeof

["M.E. O'Neill" <oneill at cs dot hmc dot edu>]

Kriang Lerdsuwanakij writes:
> This has to do with the ABI.  Mangling of 'sizeof expr' is specified
> in the ABI and implemented in GCC, but not its companion 'sizeof(type)'.

This ABI classification seems curious to me -- if I simplify the test
by changing Foobar<sizeof(Wrapper<A>)> to Foobar<sizeof(A)> everything
compiles okay.

i.e., 

    template <typename A>
    Foobar<sizeof(A)> *
    compiler_bug (A)
    {
        return 0;
    }

instantiated with A=int compiles as:

    __Z12compiler_bugIiEP6FoobarIXszT_EET_

which demangles as:

    Foobar< sizeof(int)>* compiler_bug<int>(int)

This seems to indicate that at least some of the time, name mangling
*can* represent sizeof(TYPE) [although I'm unclear why sizeof(TYPE) or
sizeof(EXPR) would be necessary anyway -- why encode sizeof(int)
symbolically when we could say (int)4, and *would* say (int)4 in other
resonably similar circumstances?]

Similarly:

    template <typename A>
    Foobar<sizeof(Wrapper<int>)> *
    compiler_bug (A)
    {
        return 0;
    }

compiles with the signature I'd actually expect for this code,

    __Z12compiler_bugIiEP6FoobarILi1EET_ 
      i.e. Foobar<(int)1>* compiler_bug<int>(int)

After all, sizeof() can be evaluated at compile time, so I would
expect sizeof(Wrapper<A>) and sizeof(Wrapper<int>) to both evaluate as
1 when A=int.  (Perhaps there is some subtle issue I'm missing here?)

    M.E.O.

P.S.  I'd no idea how to check this comment into the GNATS database,
maybe I can't (?).