Re: [Qemu-devel] [PATCH v2 2/4] softfloat: Avoid uint16 type conflict on Darwin

[Andreas Färber]

Am 01.11.2011 20:06, schrieb Eric Sunshine:
> 
> On Nov 1, 2011, at 2:52 PM, Andreas Färber wrote:
> 
>> Am 01.11.2011 19:47, schrieb Eric Sunshine:
>>> On Nov 1, 2011, at 12:37 PM, Andreas Färber wrote:
>>>> Am 01.11.2011 09:09, schrieb Eric Sunshine:
>>>>> Perhaps the following alternative solution would be more palatable?
>>>>> It's
>>>>> still tremendously ugly, but is localized to cocoa.m, thus less
>>>>> intrusive.
>>>>>
>>>>> -- >8 --
>>>>> Subject: [PATCH] softfloat: Avoid uint16 type conflict on Darwin
>>>>>
>>>>> cocoa.m includes <Security/cssmconfig.h> indirectly via
>>>>> <Cocoa/Cocoa.h>.
>>>>> cssmconfig.h defines type uint16 which unfortunately conflicts with
>>>>> the
>>>>> definition in qemu's softfloat.h, thus resulting in compilation
>>>>> failure.
>>>>> To work around the problem, #define _UINT16, which informs
>>>>> cssmconfig.h
>>>>> that uint16 is already defined and that it should not apply its own
>>>>> definition.
>>>>
>>>> Thanks for the suggestion! _UINT16 is an interesting suggestion,
>>>> however
>>>> softfloat's uint16 is not uint16_t but int, so I'd rather not do it
>>>> that
>>>> way around.
>>>>
>>>> (I had also decided against the AIX path of never defining uint16 and
>>>> always using system definitions, since that wouldn't work outside Cocoa
>>>> code.)
>>>>
>>>> Do you have any thoughts about the include path issue? If we could keep
>>>> QEMU code from getting into #import <Cocoa/Cocoa.h> then we could
>>>> redefine the system type instead, in cocoa.m.
>>>
>>> Is the intention to trust uint16 from <Security/cssmconfig.h> over the
>>> one softfloat.h? If so, shouldn't we be taking as many type definitions
>>> from <Security/cssmconfig.h> as we can rather than just this one? (I'm
>>> not recommending it; just trying to understand the goal.)
>>
>> Short-term goal: make Darwin build 1.0 without breaking others
>> Long-term goal: not use uint16 etc. in QEMU at all
>>
>> Don't see what you mean with "taking as many type definitions". After
>> uint16 I get no further conflicts for --enable-system --disable-user,
>> so what is there to take?
> 
> Sorry for not being clear. My question was not about build errors but
> about semantics. What I meant was that, with this patch, you are giving
> special preference only to Darwin's definition of uint16, but then
> contrarily preferring softfloat's definition of int16. Likewise,
> softfloat's uint32, int32, uint64, int64 from softfloat are trusted over
> the definitions from Darwin.
> 
> Other than the fact that only uint16 led to a compilation error, it does
> not make sense semantically to single out Darwin's definition of only
> this one type. I would think that we should be trusting either _all_
> Darwin type definitions or _none_. Singling out just this one seems
> anomalous.

Listen, I dont have time for this. We have three options:

1) I can say, "I'm the Cocoa maintainer for multiple years now, I don't
care if someone pops up day before the deadline and complains" and just
push my version of preference.

2) We disagree on the solution, so I'm fair and send a pull request for
the three other non-controversial patches only and 1.0 remains broken on
Darwin.

3) You send a patch based on this one, detailing what additional changes
you suggest and we'll see clearer what exactly you mean.

I'm not preferring any definition of int16, uint32, etc., there simply
is no conflict, so why would I clutter softfloat.h with unnecessary
workarounds that we want to go away anyway.

Feel free to refactor fpu/* instead to not use uint16 in the first
place. I did so once and it was rejected, so I'm not too inclined to do
that again unless we decide on how exactly to proceed with that!

Andreas