[PATCH] IR: Move the complex address expression field out of DIVariable and into an extra argument of the dbg.declare/dbg.value intrinsics.

[David Blaikie]

>>! In D4919#10, @aprantl wrote:
> If we also bump the metadata version number with this patch, the debug info will be stripped. Originally I thought that AsmParser would still throw an error because the intrinsics are missing an argument, but this is not the case, I just verified that. So this is good, that removes the autoupgrade requirement!

Excellent!

> My rationale for DW_TAG_user_lo is that this way we can guarantee that all DIDescriptor tags (except line table entries) are within the range [DW_TAG_array_type | LLVM_DebugVersion .. DW_TAG_user_hi|LLVM_DebugVersion].

In what way is this beneficial? I suppose it helps us reserve different namespaces for different metadata annotations. Are we annotating any other metadata? How are we identifying that metadata? If we are, I assume we'd want an overall scheme, but I doubt we are.

> My current suggestion is to use the end of the DW_TAG_user range for this purpose:
> ```
>     enum ComplexAddrKind {
>       OpFirst = dwarf::DW_TAG_hi_user - dwarf::DW_OP_hi_user,
>       OpPlus  = OpFirst + dwarf::DW_OP_plus,
>       OpDeref = OpFirst + dwarf::DW_OP_deref,
>       OpPiece = OpFirst + dwarf::DW_OP_piece,
>       OpLast = OpPiece
>     };

The reason I shy away from this is that I think, as I mentioned, it confuses what these values are - they're not tags, they're just identifiers. (I mean I suppose the same is true of DW_TAG_arg_variable/param_variable or whatever it is we added that are non-standard 'tags' - but that's a bit more justified because they're in the same part of the schema - they have to be unique there/in the same range)

> 
> ```

http://reviews.llvm.org/D4919