[Mesa-dev] [PATCH 1/3] util: Cope with LONG_BIT not being defined on Windows.

[Jose Fonseca]

On 09/08/15 21:30, Jose Fonseca wrote:
> On 09/08/15 21:21, Roland Scheidegger wrote:
>> Am 09.08.2015 um 20:08 schrieb Jose Fonseca:
>>> On 09/08/15 17:47, Matt Turner wrote:
>>>> On Sun, Aug 9, 2015 at 3:57 AM, Jose Fonseca <jfonseca at vmware.com>
>>>> wrote:
>>>>> Neither MSVC nor MinGW defines LONG_BIT.  For MSVC this was not a
>>>>> problem as
>>>>> it doesn't define __x86_64__ macro (it's GCC specific.)
>>>>>
>>>>> However on Windows long type is guaranteed to be 32bits.
>>>>>
>>>>> Also add an #error, as GCC will just warn, not throw any error,
>>>>> when no
>>>>> value is returned.
>>>>>
>>>>> Trivial.
>>>>> ---
>>>>>    src/util/rounding.h | 8 ++++++--
>>>>>    1 file changed, 6 insertions(+), 2 deletions(-)
>>>>>
>>>>> diff --git a/src/util/rounding.h b/src/util/rounding.h
>>>>> index b0c9918..ec31b47 100644
>>>>> --- a/src/util/rounding.h
>>>>> +++ b/src/util/rounding.h
>>>>> @@ -96,8 +96,10 @@ _mesa_lroundevenf(float x)
>>>>>    #ifdef __x86_64__
>>>>>    #if LONG_BIT == 64
>>>>>       return _mm_cvtss_si64(_mm_load_ss(&x));
>>>>> -#elif LONG_BIT == 32
>>>>> +#elif LONG_BIT == 32 || defined(_WIN32)
>>>>
>>>> Not clear to me that you want to do this, since on x86-32 using SSE
>>>> intrinsics will force a few memory transfers to and from the x87 FPU.
>>>
>>> Our Windows builds use SSE math for x86. (ie., -fpmath=sse for GCC, no
>>> option required for MSVC, as it's the standard.)  So no x87 should be
>>> ever used.
>>>
>>> (Probably it should make sense to use -msse2 and -fpmath=sse on all
>>> 32bits x86.)
>>>
>>>>
>>>>>       return _mm_cvtss_si32(_mm_load_ss(&x));
>>>>> +#else
>>>>> +#error "Unsupported or undefined LONG_BIT"
>>>>
>>>> MSVC... sigh.
>>>
>>> The problem was actually with MinGW, not MSVC.  As the commit
>>> description said, MSVC was fine (because it doesn't define __x86_64__).
>>>
>>> AFAICT, LONG_BIT is not standard C.  So I don't think that MSVC or MinGW
>>> is at fault.
>>>
>>> Jose
>>
>> Apparently it's not even always defined on linux...
>> I wonder if soething like
>> #if LONG_MAX == 0x7FFFFFFFL would be more portable...
>
> Yes, that seems a good idea.

One potential problem is that when long is 32bits the constant in

   #if LONG_MAX == 0x7fffffffffffffffL

might overflow and give wrong results.

So it might be better to use LL suffix there.

Or just use stdint.h's INT32_MAX/INT64_MAX constants.

Jose