ldc.gccbuiltins_x86 are missing __builtin_ia32_pmullw128

[MartinNowak]

There is __builtin_ia32_pmulhw128 but no correspondig pmullw.
I tried to add the following declaration, but it results in a link error.

pragma(LDC_intrinsic, "llvm.x86.sse2.pmull.w")
    short8 __builtin_ia32_pmullw128(short8, short8)nothrow pure @safe;

function _Dmain: error: undefined reference to 'llvm.x86.sse2.pmull.w'

[dnadlinger]

This means LLVM doesn't recognize the intrinsic at all. Usually, this happens if there is a natural way to express this in terms of LLVM IR vector operations. This specific intrinsic doesn't ring a bell with me, but maybe an option would be to use the inline IR facilities. http://clang.llvm.org/builtins.py also seems to indicate that there is possibly a rewrite.

[dnadlinger]

Okay, Clang defines this in the library (lib/Headers/emmintrin.h) as

static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
_mm_mullo_epi16(__m128i __a, __m128i __b)
{
  return (__m128i)((__v8hi)__a * (__v8hi)__b);
}

So, the best way to implement this would probably be either directly using D vector ops, or using inline IR. See https://github.com/ldc-developers/druntime/blob/ldc/src/ldc/simd.di for an example of the latter.

[Safety0ff]

Correct link: https://github.com/ldc-developers/druntime/blob/ldc/src/ldc/simd.di

[dnadlinger]

Thanks – don't know what happened there. ;)

[dnadlinger]

Closing, since the solution seems to be to use vector type multiplication.