Extended Description

Currently global dtor callback registration stub has the prototype that ignores pointer address spaces

i32 @​__cxa_atexit(void (i8*), i8, i8 addrspace(1)*)

The address space of object parameter of destructor and argument of __cxa_atexit is not taken into account and therefore when address space appears in the destructor or an object the generated IR is not functional.

Some ideas to fix this is proposed in https://reviews.llvm.org/D62413#1583079.

This suggests that for OpenCL we could generate a custom function (one per each global object to be destoried). This function will be passed into func ptr parameter of __cxa_atexit instead of dtor itself how it is done now. The new function will contain code to invoke dtor on the right object directly i.e. for the following example

struct S {
~S(){};
};

S s;

The proposed IR generated equivalent should be:

void __dtor_s(){ // The naming scheme is to be determined
~S(s); // NOTE if address spaces are mismatched we need a conversion. Sema should already check its validity (?)
}

void @​__cxx_global_var_init()
{
@​__cxa_atexit(__dtor_S, null, null);
}

Note that currently clang doesn't generate __dtor_s() and it passes dtor itself as an argument to @​__cxa_atexit, see https://godbolt.org/z/4soT5r

The advantage of the proposed approach is that it can work for any address space uniformly and it makes implementation of @​__cxa_atexit for various addres spaces very straight forward. There is no need to change default ABI at all. However it might be good to standardise what clang generates in a long term. This works for global variables and static local destruction too both in gloabl and constant address spaces.

The drawback is that we need to generate an extra function for each global object. To avoid this we could introduce some sort of overloading for @​__cxa_atexit that could take dtor function with different parameters as well as object argument in different address spaces. It seems like a large change to ABI and potentially not backward compatible. There could also be some hybrid schemes where only @​__cxa_atexit with address spaces is to be mangled. However both options makes @​__cxa_atexit implementation non-trivial.