Patches

[LemonBoy]

No description provided.

[andrewrk]

The bug fix needs work. The other commit looks good.

[andrewrk]

I think a call to ir_resolve_const_val is appropriate here. Maybe like this (changing some lines above as well):

if (orig_array_ptr_val->special != ConstValSpecialRuntime) {
    if ((err = ir_resolve_const_val(ira->codegen, ira->new_irb.exec, value->base.source_node,
                    orig_array_ptr_val, UndefBad)))
    {
        return ira->codegen->invalid_inst_gen;
    }
    if (orig_array_ptr_val->data.x_ptr.special != ConstPtrSpecialHardCodedAddr) {
        orig_array_ptr_val = const_ptr_pointee(ira, ira->codegen, orig_array_ptr_val,
            elem_ptr_instruction->base.base.source_node);
        if (orig_array_ptr_val == nullptr)
            return ira->codegen->invalid_inst_gen; 
        // and then if we need to make sure the new orig_array_ptr_val is also not
        // undefined then another call to ir_resolve_const_val here.
    }
}

[andrewrk]

This test case isn't correct - ram_u32 is a global variable, and it might have been reassigned before entry is executed, so it can't be a compile error detecting use of undefined here without some advanced analysis that Zig is not currently capable of. Also the "use of undefined value" as a compile error is deprecated (See #1947), only certain kinds of use of undefined is invalid, and it gives the message, "use of undefined value here causes undefined behavior", which should be emitted via a call to ir_resolve_const_val (or ir_resolve_const) with the UndefAllowed parameter set to UndefBad.

It's a bit tricky - when obtaining the pointer to a global variable, it is considered to be comptime known, however the memory that it points to is runtime known. So if ir_analyze_instruction_elem_ptr is seeing ConstValSpecialUndef then I think the bug has already occurred elsewhere. ir_analyze_instruction_elem_ptr should be seeing ConstValSpecialRuntime.

[LemonBoy]

Indeed, I got sidetracked by the complex logic and forgot that it was a var. The new patch allows the code to compile (no idea how to test that) and I've added a check to make sure the use of undefined is caught at compile-time.