llvm dev - Feb 2015 - [LLVMdev] RFC: Removing ABIArgInfo::Expand from Clang ABI lowering

Hey, all.  I’d like to remove ABIArgInfo::Expand from Clang’s function lowering
logic.  It is essentially ABIArgInfo::Direct except (1) it is less flexible
about the actual types to coerce to and (2) it requires some awkward redundant
code to do recursive walks through types.  The redundancy can be fixed, but I’d
rather just remove the code completely.

In principle, Expand can be used to expand structures with internal padding, but
in practice nobody seems to be using it that way, probably because structures
like that tend to be passed on the stack or need some other careful handling.

I’ve looked at our currently-supported in-tree uses of Expand, and I feel
confident that they can all be easily switched to Direct.  However, I am willing
to maintain Expand if anyone is maintaining an out-of-tree target and cannot
simply switch to Direct, or if anyone has some other compelling reason Expand
needs to stay.  Otherwise, I’ll be removing it relatively soon.

John.

Hi John,

On 3 February 2015 at 21:40, John McCall <rjmccall at apple.com>
wrote:
> I’ve looked at our currently-supported in-tree uses of Expand, and I feel
confident
> that they can all be easily switched to Direct.  However, I am willing to
maintain
> Expand if anyone is maintaining an out-of-tree target and cannot simply
switch to
> Direct, or if anyone has some other compelling reason Expand needs to stay.
> Otherwise, I’ll be removing it relatively soon.
I think removing it is a good idea too (assuming no OOT user has a
compelling scenario).

By default a struct coercion gets expanded to something equivalent
before lib/Target in the backend sees it anyway (modulo possible weird
union behaviour, which would almost certainly be unintended).

Cheers.

Tim.