[PATCH] pr13012: Support -fpcc-struct-return for x86-32

[John McCall]

On Jun 7, 2013, at 2:12 PM, Arthur O'Dwyer <arthur.j.odwyer at gmail.com> wrote:
> On Fri, Jun 7, 2013 at 1:58 PM, John McCall <rjmccall at apple.com> wrote:
>> On Jun 7, 2013, at 12:07 PM, Arthur O'Dwyer <arthur.j.odwyer at gmail.com> wrote:
>>> On Wed, Jun 5, 2013 at 1:23 PM, Arthur O'Dwyer
>>> <arthur.j.odwyer at gmail.com> wrote:
>>>> 
>>>> This is my first post to cfe-commits, so my apologies if I get
>>>> something wrong.
>>>> A while back I filed an issue on Bugzilla requesting support for
>>>> -fpcc-struct-return:
>>>> http://llvm.org/bugs/show_bug.cgi?id=13012
>>>> 
>>>> Recently I wrote a real patch [...] I believe I don't have the necessary
>>>> credentials to commit the patch anyway, so basically I'm looking for
>>>> a patron to adopt this patch. :)
>>>> 
>>>> To observe what the patch does:
>>>> echo "struct S {int i;} foo() {return (struct S){42};}" >foo.c
>>>> clang -m32 -O3 -fomit-frame-pointer foo.c -S -o reg.s
>>>> clang -m32 -O3 -fomit-frame-pointer foo.c -S -o pcc.s -fpcc-struct-return
>>>> diff reg.s pcc.s
>>> 
>>> Ping? Any takers? Any comments at all?
>> 
>> I would really like to avoid adding support for all these little ABI-tweaking
>> options, and frankly, unless you're just being unassuming, it doesn't sound
>> like this is all that important to you.
> 
> I do tend to be unassuming. Would yelling about it help? ;) It's
> important to my employer, in that the only alternative we're aware of
> is for us to maintain our own little Clang fork in perpetuity. If
> there's a better alternative, I'd like to hear it.
> 
> For example, maybe it would be more acceptable to add a new target
> triple specifically for "x86-32 Linux with pcc struct returns"?
> Philosophically, I vastly prefer to add the -fpcc-struct-return option
> so that *everybody* can get the benefit, no matter what platform
> they're on (and in a GCC-compatible way).
> 
>> ABI-changing options are also a Real Big Hammer for things like making
>> a couple assembly stubs have a slightly simpler CC.
> 
> In our case it's not "a couple assembly stubs", it's an entire
> application program written in Objective-C; and we need it to use the
> pcc calling convention so that we can link it against dynamic
> libraries that use the pcc calling convention. We need a hammer big
> enough to change the CC across the entire program. On GCC, that hammer
> is named "-fpcc-struct-return", and it works like this.

Alright.  My main worry is the sheer number of tiny ABI variants out there,
as you can joyfully witness by, say, reading the GCC manual cover to cover.
Maybe it is our destiny to support every last one of them, but if, by being
slightly obstinate, I can force one or two of them to fade into oblivion, then
I think the world will be better for it.

Since we effectively already support this and just need to expose an
explicit lever of control, it's fine by me.  I'll review the patch.

John.