[clang] Invalid 32bit pointer (-fms-extensions __ptr32 __uptr) on non-Windows platforms

[d-mozulyov]

Problem

32-bit pointer does not work correctly on non-Windows platforms. See the Godbolt listing.

Sample

int param_std(int* foo) {
    int temp = *foo;
    *foo = 1;
    return temp;
}

int param_ptr32(int* __ptr32 __uptr bar) {
    int temp = *bar;
    *bar = 2;
    return temp;
}

int emulate_ptr32(unsigned baz) {
    unsigned long long baz64 = baz;
    int* bazptr = (int*)baz64;
    int temp = *bazptr;
    *bazptr = 3;
    return temp;
}

struct Qux {
    int* __ptr32 __uptr value;
};

int struct_ptr32(Qux& qux) {
    int temp = *qux.value;
    *qux.value = 4;
    return temp;
}

Conclusions

The param_ptr32 disassembler must be equal to emulate_ptr32, like struct_ptr32, and different from param_std. Based on the Godbolt listing, the MSVC and Clang for Windows target (x86_64-windows-gnu) compilers do just that. But on non-Windows platforms, Clang may not do the important conversion from a 32-bit pointer to a 64-bit one. I have localized this behavior in two cases. First, if the pointer appears as a function argument (param_ptr32). Second, if it is a member of a small structure that is passed as an argument (change Qux& qux to Qux qux).

[EugeneZelenko]

@d-mozulyov: Please elaborate.

[d-mozulyov]

@EugeneZelenko

It looks like a 32 bit pointer works on POSIX x64. The difference is in the ABI - on POSIX x64 the 32-bit pointer-argument is already unpacked.

https://godbolt.org/z/WvosqdW96

[d-mozulyov]

@EugeneZelenko

But on ARM 64 still not working: #62536

[llvmbot]

@llvm/issue-subscribers-backend-x86