On Oct 9, 2010, at 9:58 PM, Argyrios Kyrtzidis wrote:
>>> Why can't they use the right intrinsic from xmmintrin.h? We prefer not to
>>> implement builtins like this, people are supposed to use the official
>>> documented interfaces in the *mmintrin.h files.
>>
>> Oh... unless this is an MMX builtin. In which case we *do* want to support
>> it, but we *have* to map it onto an LLVM IR intrinsic.
>
> They are using _mm_cvtsi64_si32 from mmintrin.h, here it is for reference:
>
> static __inline__ int __attribute__((__always_inline__, __nodebug__))
> _mm_cvtsi64_si32(__m64 __m)
> {
> return __builtin_ia32_vec_ext_v2si((__v2si)__m, 0);
> }
>
> The IR emitted for it is the same as llvm-gcc, if it needs mapping onto an
> LLVM IR intrinsic feel free to make the necessary changes or let me know what
> they should be.
Ok, cool. Bill, can you look into hooking up the intrinsic?
-Chris
_______________________________________________
cfe-commits mailing list
[email protected]
http://lists.cs.uiuc.edu/mailman/listinfo/cfe-commits
