[Clang][AVX512][Builtin] Adding intrinsics of pmovus{d|q}{b|w|d}{128|256|512} instruction set

[m_zuckerman]

Ping

[igorb]

lib/Headers/avx512fintrin.h

5626

could you please put casting on the same line with argument ( the same regarding the code below)

[igorb]

LGTM

Thanks!

include/clang/Basic/BuiltinsX86.def

 TARGET_BUILTIN(__builtin_ia32_shufps128_mask, "V4fV4fV4fIiV4fUc","","avx512vl")
 TARGET_BUILTIN(__builtin_ia32_shufps256_mask, "V8fV8fV8fIiV8fUc","","avx512vl")
 TARGET_BUILTIN(__builtin_ia32_sqrtsd_round_mask, "V2dV2dV2dV2dUcIi","","avx512f")
 TARGET_BUILTIN(__builtin_ia32_sqrtss_round_mask, "V4fV4fV4fV4fUcIi","","avx512f")
 TARGET_BUILTIN(__builtin_ia32_rsqrt14pd128_mask, "V2dV2dV2dUc","","avx512vl")
 TARGET_BUILTIN(__builtin_ia32_rsqrt14pd256_mask, "V4dV4dV4dUc","","avx512vl")
 TARGET_BUILTIN(__builtin_ia32_rsqrt14ps128_mask, "V4fV4fV4fUc","","avx512vl")
 TARGET_BUILTIN(__builtin_ia32_rsqrt14ps256_mask, "V8fV8fV8fUc","","avx512vl")
+TARGET_BUILTIN(__builtin_ia32_pmovusdb512_mask, "V16cV16iV16cUs","","avx512f")
+TARGET_BUILTIN(__builtin_ia32_pmovusdb512mem_mask, "vV16c*V16iUs","","avx512f")
+TARGET_BUILTIN(__builtin_ia32_pmovusdw512_mask, "V16sV16iV16sUs","","avx512f")
+TARGET_BUILTIN(__builtin_ia32_pmovusdw512mem_mask, "vV16s*V16iUs","","avx512f")
+TARGET_BUILTIN(__builtin_ia32_pmovusqb512_mask, "V16cV8LLiV16cUc","","avx512f")
+TARGET_BUILTIN(__builtin_ia32_pmovusqb512mem_mask, "vV16c*V8LLiUc","","avx512f")
+TARGET_BUILTIN(__builtin_ia32_pmovusqd512_mask, "V8iV8LLiV8iUc","","avx512f")
+TARGET_BUILTIN(__builtin_ia32_pmovusqd512mem_mask, "vV8i*V8LLiUc","","avx512f")
+TARGET_BUILTIN(__builtin_ia32_pmovusqw512_mask, "V8sV8LLiV8sUc","","avx512f")
+TARGET_BUILTIN(__builtin_ia32_pmovusqw512mem_mask, "vV8s*V8LLiUc","","avx512f")
+TARGET_BUILTIN(__builtin_ia32_pmovusdb128_mask, "V16cV4iV16cUc","","avx512vl")
+TARGET_BUILTIN(__builtin_ia32_pmovusdb128mem_mask, "vV16c*V4iUc","","avx512vl")
+TARGET_BUILTIN(__builtin_ia32_pmovusdb256_mask, "V16cV8iV16cUc","","avx512vl")
+TARGET_BUILTIN(__builtin_ia32_pmovusdb256mem_mask, "vV16c*V8iUc","","avx512vl")
+TARGET_BUILTIN(__builtin_ia32_pmovusdw128_mask, "V8sV4iV8sUc","","avx512vl")
+TARGET_BUILTIN(__builtin_ia32_pmovusdw128mem_mask, "vV8s*V4iUc","","avx512vl")
+TARGET_BUILTIN(__builtin_ia32_pmovusdw256_mask, "V8sV8iV8sUc","","avx512vl")
+TARGET_BUILTIN(__builtin_ia32_pmovusdw256mem_mask, "vV8s*V8iUc","","avx512vl")
+TARGET_BUILTIN(__builtin_ia32_pmovusqb128_mask, "V16cV2LLiV16cUc","","avx512vl")
+TARGET_BUILTIN(__builtin_ia32_pmovusqb128mem_mask, "vV16c*V2LLiUc","","avx512vl")
+TARGET_BUILTIN(__builtin_ia32_pmovusqb256_mask, "V16cV4LLiV16cUc","","avx512vl")
+TARGET_BUILTIN(__builtin_ia32_pmovusqb256mem_mask, "vV16c*V4LLiUc","","avx512vl")
+TARGET_BUILTIN(__builtin_ia32_pmovusqd128_mask, "V4iV2LLiV4iUc","","avx512vl")
+TARGET_BUILTIN(__builtin_ia32_pmovusqd128mem_mask, "vV4i*V2LLiUc","","avx512vl")
+TARGET_BUILTIN(__builtin_ia32_pmovusqd256_mask, "V4iV4LLiV4iUc","","avx512vl")
+TARGET_BUILTIN(__builtin_ia32_pmovusqd256mem_mask, "vV4i*V4LLiUc","","avx512vl")
+TARGET_BUILTIN(__builtin_ia32_pmovusqw128_mask, "V8sV2LLiV8sUc","","avx512vl")
+TARGET_BUILTIN(__builtin_ia32_pmovusqw128mem_mask, "vV8s*V2LLiUc","","avx512vl")
+TARGET_BUILTIN(__builtin_ia32_pmovusqw256_mask, "V8sV4LLiV8sUc","","avx512vl")
+TARGET_BUILTIN(__builtin_ia32_pmovusqw256mem_mask, "vV8s*V4LLiUc","","avx512vl")
 
 #undef BUILTIN
 #undef TARGET_BUILTIN

lib/Headers/avx512fintrin.h

 
 #define _mm_maskz_sqrt_round_ss( __U, __A, __B, __R) __extension__ ({ \
 __builtin_ia32_sqrtss_round_mask ((__v4sf)( __B),\
                  (__v4sf)( __A),(__v4sf) _mm_setzero_ps(),\
                  (__mmask8) __U,\
                  __R);\
 })
 
+static __inline__ __m128i __DEFAULT_FN_ATTRS
+_mm512_cvtusepi32_epi8 (__m512i __A)
+{
+  return (__m128i) __builtin_ia32_pmovusdb512_mask ((__v16si) __A,
+                (__v16qi) _mm_undefined_si128 (),
+                (__mmask16) -1);
+}
+
+static __inline__ __m128i __DEFAULT_FN_ATTRS
+_mm512_mask_cvtusepi32_epi8 (__m128i __O, __mmask16 __M, __m512i __A)
+{
+  return (__m128i) __builtin_ia32_pmovusdb512_mask ((__v16si) __A,
+                (__v16qi) __O,
+                __M);
+}
+
+static __inline__ __m128i __DEFAULT_FN_ATTRS
+_mm512_maskz_cvtusepi32_epi8 (__mmask16 __M, __m512i __A)
+{
+  return (__m128i) __builtin_ia32_pmovusdb512_mask ((__v16si) __A,
+                (__v16qi) _mm_setzero_si128 (),
+                __M);
+}
+
+static __inline__ void __DEFAULT_FN_ATTRS
+_mm512_mask_cvtusepi32_storeu_epi8 (void * __P, __mmask16 __M, __m512i __A)
+{
+  __builtin_ia32_pmovusdb512mem_mask ((__v16qi *) __P, (__v16si) __A, __M);
+}
+
+static __inline__ __m256i __DEFAULT_FN_ATTRS
+_mm512_cvtusepi32_epi16 (__m512i __A)
+{
+  return (__m256i) __builtin_ia32_pmovusdw512_mask ((__v16si) __A,
+                (__v16hi) _mm256_undefined_si256 (),
+                (__mmask16) -1);
+}
+
+static __inline__ __m256i __DEFAULT_FN_ATTRS
+_mm512_mask_cvtusepi32_epi16 (__m256i __O, __mmask16 __M, __m512i __A)
+{
+  return (__m256i) __builtin_ia32_pmovusdw512_mask ((__v16si) __A,
+                (__v16hi) __O,
+                __M);
+}
+
+static __inline__ __m256i __DEFAULT_FN_ATTRS
+_mm512_maskz_cvtusepi32_epi16 (__mmask16 __M, __m512i __A)
+{
+  return (__m256i) __builtin_ia32_pmovusdw512_mask ((__v16si) __A,
+                (__v16hi) _mm256_setzero_si256 (),
+                __M);
+}
+
+static __inline__ void __DEFAULT_FN_ATTRS
+_mm512_mask_cvtusepi32_storeu_epi16 (void *__P, __mmask16 __M, __m512i __A)
+{
+  __builtin_ia32_pmovusdw512mem_mask ((__v16hi*) __P, (__v16si) __A, __M);
+}
+
+static __inline__ __m128i __DEFAULT_FN_ATTRS
+_mm512_cvtusepi64_epi8 (__m512i __A)
+{
+  return (__m128i) __builtin_ia32_pmovusqb512_mask ((__v8di) __A,
+                (__v16qi) _mm_undefined_si128 (),
+                (__mmask8) -1);
+}
+
+static __inline__ __m128i __DEFAULT_FN_ATTRS
+_mm512_mask_cvtusepi64_epi8 (__m128i __O, __mmask8 __M, __m512i __A)
+{
+  return (__m128i) __builtin_ia32_pmovusqb512_mask ((__v8di) __A,
+                (__v16qi) __O,
+                __M);
+}
+
+static __inline__ __m128i __DEFAULT_FN_ATTRS
+_mm512_maskz_cvtusepi64_epi8 (__mmask8 __M, __m512i __A)
+{
+  return (__m128i) __builtin_ia32_pmovusqb512_mask ((__v8di) __A,
+                (__v16qi) _mm_setzero_si128 (),
+                __M);
+}
+
+static __inline__ void __DEFAULT_FN_ATTRS
+_mm512_mask_cvtusepi64_storeu_epi8 (void * __P, __mmask8 __M, __m512i __A)
+{
+  __builtin_ia32_pmovusqb512mem_mask ((__v16qi *) __P, (__v8di) __A, __M);
+}
+
+static __inline__ __m256i __DEFAULT_FN_ATTRS
+_mm512_cvtusepi64_epi32 (__m512i __A)
+{
+  return (__m256i) __builtin_ia32_pmovusqd512_mask ((__v8di) __A,
+                (__v8si) _mm256_undefined_si256 (),
+                (__mmask8) -1);
+}
+
+static __inline__ __m256i __DEFAULT_FN_ATTRS
+_mm512_mask_cvtusepi64_epi32 (__m256i __O, __mmask8 __M, __m512i __A)
+{
+  return (__m256i) __builtin_ia32_pmovusqd512_mask ((__v8di) __A,
+                (__v8si) __O, __M);
+}
+
+static __inline__ __m256i __DEFAULT_FN_ATTRS
+_mm512_maskz_cvtusepi64_epi32 (__mmask8 __M, __m512i __A)
+{
+  return (__m256i) __builtin_ia32_pmovusqd512_mask ((__v8di) __A,
+                (__v8si) _mm256_setzero_si256 (),
+                __M);
+}
+
+static __inline__ void __DEFAULT_FN_ATTRS
+_mm512_mask_cvtusepi64_storeu_epi32 (void* __P, __mmask8 __M, __m512i __A)
+{
+  __builtin_ia32_pmovusqd512mem_mask ((__v8si*) __P, (__v8di) __A, __M);
+}
+
+static __inline__ __m128i __DEFAULT_FN_ATTRS
+_mm512_cvtusepi64_epi16 (__m512i __A)
+{
+  return (__m128i) __builtin_ia32_pmovusqw512_mask ((__v8di) __A,
+                (__v8hi) _mm_undefined_si128 (),
+                (__mmask8) -1);
+}
+
+static __inline__ __m128i __DEFAULT_FN_ATTRS
+_mm512_mask_cvtusepi64_epi16 (__m128i __O, __mmask8 __M, __m512i __A)
+{
+  return (__m128i) __builtin_ia32_pmovusqw512_mask ((__v8di) __A,
+                (__v8hi) __O, __M);
+}
+
+static __inline__ __m128i __DEFAULT_FN_ATTRS
+_mm512_maskz_cvtusepi64_epi16 (__mmask8 __M, __m512i __A)
+{
+  return (__m128i) __builtin_ia32_pmovusqw512_mask ((__v8di) __A,
+                (__v8hi) _mm_setzero_si128 (),
+                __M);
+}
+
+static __inline__ void __DEFAULT_FN_ATTRS
+_mm512_mask_cvtusepi64_storeu_epi16 (void *__P, __mmask8 __M, __m512i __A)
+{
+  __builtin_ia32_pmovusqw512mem_mask ((__v8hi*) __P, (__v8di) __A, __M);
+}
+
 #undef __DEFAULT_FN_ATTRS
 
 #endif // __AVX512FINTRIN_H

lib/Headers/avx512vlintrin.h

 _mm256_maskz_rsqrt14_ps (__mmask8 __U, __m256 __A)
 {
   return (__m256) __builtin_ia32_rsqrt14ps256_mask ((__v8sf) __A,
                 (__v8sf)
                 _mm256_setzero_ps (),
                 (__mmask8) __U);
 }
 
+static __inline__ __m128i __DEFAULT_FN_ATTRS
+_mm_cvtusepi32_epi8 (__m128i __A)
+{
+  return (__m128i) __builtin_ia32_pmovusdb128_mask ((__v4si) __A,
+                (__v16qi)_mm_undefined_si128(),
+                (__mmask8) -1);
+}
+
+static __inline__ __m128i __DEFAULT_FN_ATTRS
+_mm_mask_cvtusepi32_epi8 (__m128i __O, __mmask8 __M, __m128i __A)
+{
+  return (__m128i) __builtin_ia32_pmovusdb128_mask ((__v4si) __A,
+                (__v16qi) __O,
+                __M);
+}
+
+static __inline__ __m128i __DEFAULT_FN_ATTRS
+_mm_maskz_cvtusepi32_epi8 (__mmask8 __M, __m128i __A)
+{
+  return (__m128i) __builtin_ia32_pmovusdb128_mask ((__v4si) __A,
+                (__v16qi) _mm_setzero_si128 (),
+                __M);
+}
+
+static __inline__ void __DEFAULT_FN_ATTRS
+_mm_mask_cvtusepi32_storeu_epi8 (void * __P, __mmask8 __M, __m128i __A)
+{
+  __builtin_ia32_pmovusdb128mem_mask ((__v16qi *) __P, (__v4si) __A, __M);
+}
+
+static __inline__ __m128i __DEFAULT_FN_ATTRS
+_mm256_cvtusepi32_epi8 (__m256i __A)
+{
+  return (__m128i) __builtin_ia32_pmovusdb256_mask ((__v8si) __A,
+                (__v16qi)_mm_undefined_si128(),
+                (__mmask8) -1);
+}
+
+static __inline__ __m128i __DEFAULT_FN_ATTRS
+_mm256_mask_cvtusepi32_epi8 (__m128i __O, __mmask8 __M, __m256i __A)
+{
+  return (__m128i) __builtin_ia32_pmovusdb256_mask ((__v8si) __A,
+                (__v16qi) __O,
+                __M);
+}
+
+static __inline__ __m128i __DEFAULT_FN_ATTRS
+_mm256_maskz_cvtusepi32_epi8 (__mmask8 __M, __m256i __A)
+{
+  return (__m128i) __builtin_ia32_pmovusdb256_mask ((__v8si) __A,
+                (__v16qi) _mm_setzero_si128 (),
+                __M);
+}
+
+static __inline__ void __DEFAULT_FN_ATTRS
+_mm256_mask_cvtusepi32_storeu_epi8 (void * __P, __mmask8 __M, __m256i __A)
+{
+  __builtin_ia32_pmovusdb256mem_mask ((__v16qi*) __P, (__v8si) __A, __M);
+}
+
+static __inline__ __m128i __DEFAULT_FN_ATTRS
+_mm_cvtusepi32_epi16 (__m128i __A)
+{
+  return (__m128i) __builtin_ia32_pmovusdw128_mask ((__v4si) __A,
+                (__v8hi)_mm_undefined_si128(),
+                (__mmask8) -1);
+}
+
+static __inline__ __m128i __DEFAULT_FN_ATTRS
+_mm_mask_cvtusepi32_epi16 (__m128i __O, __mmask8 __M, __m128i __A)
+{
+  return (__m128i) __builtin_ia32_pmovusdw128_mask ((__v4si) __A,
+                (__v8hi) __O, __M);
+}
+
+static __inline__ __m128i __DEFAULT_FN_ATTRS
+_mm_maskz_cvtusepi32_epi16 (__mmask8 __M, __m128i __A)
+{
+  return (__m128i) __builtin_ia32_pmovusdw128_mask ((__v4si) __A,
+                (__v8hi) _mm_setzero_si128 (),
+                __M);
+}
+
+static __inline__ void __DEFAULT_FN_ATTRS
+_mm_mask_cvtusepi32_storeu_epi16 (void * __P, __mmask8 __M, __m128i __A)
+{
+  __builtin_ia32_pmovusdw128mem_mask ((__v8hi *) __P, (__v4si) __A, __M);
+}
+
+static __inline__ __m128i __DEFAULT_FN_ATTRS
+_mm256_cvtusepi32_epi16 (__m256i __A)
+{
+  return (__m128i) __builtin_ia32_pmovusdw256_mask ((__v8si) __A,
+                (__v8hi) _mm_undefined_si128(),
+                (__mmask8) -1);
+}
+
+static __inline__ __m128i __DEFAULT_FN_ATTRS
+_mm256_mask_cvtusepi32_epi16 (__m128i __O, __mmask8 __M, __m256i __A)
+{
+  return (__m128i) __builtin_ia32_pmovusdw256_mask ((__v8si) __A,
+                (__v8hi) __O, __M);
+}
+
+static __inline__ __m128i __DEFAULT_FN_ATTRS
+_mm256_maskz_cvtusepi32_epi16 (__mmask8 __M, __m256i __A)
+{
+  return (__m128i) __builtin_ia32_pmovusdw256_mask ((__v8si) __A,
+                (__v8hi) _mm_setzero_si128 (),
+                __M);
+}
+
+static __inline__ void __DEFAULT_FN_ATTRS
+_mm256_mask_cvtusepi32_storeu_epi16 (void * __P, __mmask8 __M, __m256i __A)
+{
+  __builtin_ia32_pmovusdw256mem_mask ((__v8hi *) __P, (__v8si) __A, __M);
+}
+
+static __inline__ __m128i __DEFAULT_FN_ATTRS
+_mm_cvtusepi64_epi8 (__m128i __A)
+{
+  return (__m128i) __builtin_ia32_pmovusqb128_mask ((__v2di) __A,
+                (__v16qi)_mm_undefined_si128(),
+                (__mmask8) -1);
+}
+
+static __inline__ __m128i __DEFAULT_FN_ATTRS
+_mm_mask_cvtusepi64_epi8 (__m128i __O, __mmask8 __M, __m128i __A)
+{
+  return (__m128i) __builtin_ia32_pmovusqb128_mask ((__v2di) __A,
+                (__v16qi) __O,
+                __M);
+}
+
+static __inline__ __m128i __DEFAULT_FN_ATTRS
+_mm_maskz_cvtusepi64_epi8 (__mmask8 __M, __m128i __A)
+{
+  return (__m128i) __builtin_ia32_pmovusqb128_mask ((__v2di) __A,
+                (__v16qi) _mm_setzero_si128 (),
+                __M);
+}
+
+static __inline__ void __DEFAULT_FN_ATTRS
+_mm_mask_cvtusepi64_storeu_epi8 (void * __P, __mmask8 __M, __m128i __A)
+{
+  __builtin_ia32_pmovusqb128mem_mask ((__v16qi *) __P, (__v2di) __A, __M);
+}
+
+static __inline__ __m128i __DEFAULT_FN_ATTRS
+_mm256_cvtusepi64_epi8 (__m256i __A)
+{
+  return (__m128i) __builtin_ia32_pmovusqb256_mask ((__v4di) __A,
+                (__v16qi)_mm_undefined_si128(),
+                (__mmask8) -1);
+}
+
+static __inline__ __m128i __DEFAULT_FN_ATTRS
+_mm256_mask_cvtusepi64_epi8 (__m128i __O, __mmask8 __M, __m256i __A)
+{
+  return (__m128i) __builtin_ia32_pmovusqb256_mask ((__v4di) __A,
+                (__v16qi) __O,
+                __M);
+}
+
+static __inline__ __m128i __DEFAULT_FN_ATTRS
+_mm256_maskz_cvtusepi64_epi8 (__mmask8 __M, __m256i __A)
+{
+  return (__m128i) __builtin_ia32_pmovusqb256_mask ((__v4di) __A,
+                (__v16qi) _mm_setzero_si128 (),
+                __M);
+}
+
+static __inline__ void __DEFAULT_FN_ATTRS
+_mm256_mask_cvtusepi64_storeu_epi8 (void * __P, __mmask8 __M, __m256i __A)
+{
+  __builtin_ia32_pmovusqb256mem_mask ((__v16qi *) __P, (__v4di) __A, __M);
+}
+
+static __inline__ __m128i __DEFAULT_FN_ATTRS
+_mm_cvtusepi64_epi32 (__m128i __A)
+{
+  return (__m128i) __builtin_ia32_pmovusqd128_mask ((__v2di) __A,
+                (__v4si)_mm_undefined_si128(),
+                (__mmask8) -1);
+}
+
+static __inline__ __m128i __DEFAULT_FN_ATTRS
+_mm_mask_cvtusepi64_epi32 (__m128i __O, __mmask8 __M, __m128i __A)
+{
+  return (__m128i) __builtin_ia32_pmovusqd128_mask ((__v2di) __A,
+                (__v4si) __O, __M);
+}
+
+static __inline__ __m128i __DEFAULT_FN_ATTRS
+_mm_maskz_cvtusepi64_epi32 (__mmask8 __M, __m128i __A)
+{
+  return (__m128i) __builtin_ia32_pmovusqd128_mask ((__v2di) __A,
+                (__v4si) _mm_setzero_si128 (),
+                __M);
+}
+
+static __inline__ void __DEFAULT_FN_ATTRS
+_mm_mask_cvtusepi64_storeu_epi32 (void * __P, __mmask8 __M, __m128i __A)
+{
+  __builtin_ia32_pmovusqd128mem_mask ((__v4si *) __P, (__v2di) __A, __M);
+}
+
+static __inline__ __m128i __DEFAULT_FN_ATTRS
+_mm256_cvtusepi64_epi32 (__m256i __A)
+{
+  return (__m128i) __builtin_ia32_pmovusqd256_mask ((__v4di) __A,
+                (__v4si)_mm_undefined_si128(),
+                (__mmask8) -1);
+}
+
+static __inline__ __m128i __DEFAULT_FN_ATTRS
+_mm256_mask_cvtusepi64_epi32 (__m128i __O, __mmask8 __M, __m256i __A)
+{
+  return (__m128i) __builtin_ia32_pmovusqd256_mask ((__v4di) __A,
+                (__v4si) __O, __M);
+}
+
+static __inline__ __m128i __DEFAULT_FN_ATTRS
+_mm256_maskz_cvtusepi64_epi32 (__mmask8 __M, __m256i __A)
+{
+  return (__m128i) __builtin_ia32_pmovusqd256_mask ((__v4di) __A,
+                (__v4si) _mm_setzero_si128 (),
+                __M);
+}
+
+static __inline__ void __DEFAULT_FN_ATTRS
+_mm256_mask_cvtusepi64_storeu_epi32 (void * __P, __mmask8 __M, __m256i __A)
+{
+  __builtin_ia32_pmovusqd256mem_mask ((__v4si *) __P, (__v4di) __A, __M);
+}
+
+static __inline__ __m128i __DEFAULT_FN_ATTRS
+_mm_cvtusepi64_epi16 (__m128i __A)
+{
+  return (__m128i) __builtin_ia32_pmovusqw128_mask ((__v2di) __A,
+                (__v8hi)_mm_undefined_si128(),
+                (__mmask8) -1);
+}
+
+static __inline__ __m128i __DEFAULT_FN_ATTRS
+_mm_mask_cvtusepi64_epi16 (__m128i __O, __mmask8 __M, __m128i __A)
+{
+  return (__m128i) __builtin_ia32_pmovusqw128_mask ((__v2di) __A,
+                (__v8hi) __O, __M);
+}
+
+static __inline__ __m128i __DEFAULT_FN_ATTRS
+_mm_maskz_cvtusepi64_epi16 (__mmask8 __M, __m128i __A)
+{
+  return (__m128i) __builtin_ia32_pmovusqw128_mask ((__v2di) __A,
+                (__v8hi) _mm_setzero_si128 (),
+                __M);
+}
+
+static __inline__ void __DEFAULT_FN_ATTRS
+_mm_mask_cvtusepi64_storeu_epi16 (void * __P, __mmask8 __M, __m128i __A)
+{
+  __builtin_ia32_pmovusqw128mem_mask ((__v8hi *) __P, (__v2di) __A, __M);
+}
+
+static __inline__ __m128i __DEFAULT_FN_ATTRS
+_mm256_cvtusepi64_epi16 (__m256i __A)
+{
+  return (__m128i) __builtin_ia32_pmovusqw256_mask ((__v4di) __A,
+                (__v8hi)_mm_undefined_si128(),
+                (__mmask8) -1);
+}
+
+static __inline__ __m128i __DEFAULT_FN_ATTRS
+_mm256_mask_cvtusepi64_epi16 (__m128i __O, __mmask8 __M, __m256i __A)
+{
+  return (__m128i) __builtin_ia32_pmovusqw256_mask ((__v4di) __A,
+                (__v8hi) __O, __M);
+}
+
+static __inline__ __m128i __DEFAULT_FN_ATTRS
+_mm256_maskz_cvtusepi64_epi16 (__mmask8 __M, __m256i __A)
+{
+  return (__m128i) __builtin_ia32_pmovusqw256_mask ((__v4di) __A,
+                (__v8hi) _mm_setzero_si128 (),
+                __M);
+}
+
+static __inline__ void __DEFAULT_FN_ATTRS
+_mm256_mask_cvtusepi64_storeu_epi16 (void * __P, __mmask8 __M, __m256i __A)
+{
+  return __builtin_ia32_pmovusqw256mem_mask ((__v8hi *) __P, (__v4di) __A, __M);
+}
+
 #undef __DEFAULT_FN_ATTRS
 #undef __DEFAULT_FN_ATTRS_BOTH
 
 #endif /* __AVX512VLINTRIN_H */

test/CodeGen/avx512f-builtins.c

// CHECK: @llvm.x86.avx512.mask.sqrt.ss
     return _mm_maskz_sqrt_ss(__U,__A,__B);
 }
 
 __m128 test_mm_maskz_sqrt_round_ss(__mmask8 __U, __m128 __A, __m128 __B){
   // CHECK: @llvm.x86.avx512.mask.sqrt.ss
     return _mm_maskz_sqrt_round_ss(__U,__A,__B,_MM_FROUND_CUR_DIRECTION);
 }
 
+__m128i test_mm512_cvtusepi32_epi8(__m512i __A) {
+  // CHECK-LABEL: @test_mm512_cvtusepi32_epi8
+  // CHECK: @llvm.x86.avx512.mask.pmovus.db.512
+  return _mm512_cvtusepi32_epi8(__A); 
+}
+
+__m128i test_mm512_mask_cvtusepi32_epi8(__m128i __O, __mmask16 __M, __m512i __A) {
+  // CHECK-LABEL: @test_mm512_mask_cvtusepi32_epi8
+  // CHECK: @llvm.x86.avx512.mask.pmovus.db.512
+  return _mm512_mask_cvtusepi32_epi8(__O, __M, __A); 
+}
+
+__m128i test_mm512_maskz_cvtusepi32_epi8(__mmask16 __M, __m512i __A) {
+  // CHECK-LABEL: @test_mm512_maskz_cvtusepi32_epi8
+  // CHECK: @llvm.x86.avx512.mask.pmovus.db.512
+  return _mm512_maskz_cvtusepi32_epi8(__M, __A); 
+}
+
+void test_mm512_mask_cvtusepi32_storeu_epi8(void * __P, __mmask16 __M, __m512i __A) {
+  // CHECK-LABEL: @test_mm512_mask_cvtusepi32_storeu_epi8
+  // CHECK: @llvm.x86.avx512.mask.pmovus.db.mem.512
+  return _mm512_mask_cvtusepi32_storeu_epi8(__P, __M, __A); 
+}
+
+__m256i test_mm512_cvtusepi32_epi16(__m512i __A) {
+  // CHECK-LABEL: @test_mm512_cvtusepi32_epi16
+  // CHECK: @llvm.x86.avx512.mask.pmovus.dw.512
+  return _mm512_cvtusepi32_epi16(__A); 
+}
+
+__m256i test_mm512_mask_cvtusepi32_epi16(__m256i __O, __mmask16 __M, __m512i __A) {
+  // CHECK-LABEL: @test_mm512_mask_cvtusepi32_epi16
+  // CHECK: @llvm.x86.avx512.mask.pmovus.dw.512
+  return _mm512_mask_cvtusepi32_epi16(__O, __M, __A); 
+}
+
+__m256i test_mm512_maskz_cvtusepi32_epi16(__mmask16 __M, __m512i __A) {
+  // CHECK-LABEL: @test_mm512_maskz_cvtusepi32_epi16
+  // CHECK: @llvm.x86.avx512.mask.pmovus.dw.512
+  return _mm512_maskz_cvtusepi32_epi16(__M, __A); 
+}
+
+void test_mm512_mask_cvtusepi32_storeu_epi16(void *__P, __mmask16 __M, __m512i __A) {
+  // CHECK-LABEL: @test_mm512_mask_cvtusepi32_storeu_epi16
+  // CHECK: @llvm.x86.avx512.mask.pmovus.dw.mem.512
+  return _mm512_mask_cvtusepi32_storeu_epi16(__P, __M, __A); 
+}
+
+__m128i test_mm512_cvtusepi64_epi8(__m512i __A) {
+  // CHECK-LABEL: @test_mm512_cvtusepi64_epi8
+  // CHECK: @llvm.x86.avx512.mask.pmovus.qb.512
+  return _mm512_cvtusepi64_epi8(__A); 
+}
+
+__m128i test_mm512_mask_cvtusepi64_epi8(__m128i __O, __mmask8 __M, __m512i __A) {
+  // CHECK-LABEL: @test_mm512_mask_cvtusepi64_epi8
+  // CHECK: @llvm.x86.avx512.mask.pmovus.qb.512
+  return _mm512_mask_cvtusepi64_epi8(__O, __M, __A); 
+}
 
+__m128i test_mm512_maskz_cvtusepi64_epi8(__mmask8 __M, __m512i __A) {
+  // CHECK-LABEL: @test_mm512_maskz_cvtusepi64_epi8
+  // CHECK: @llvm.x86.avx512.mask.pmovus.qb.512
+  return _mm512_maskz_cvtusepi64_epi8(__M, __A); 
+}
+
+void test_mm512_mask_cvtusepi64_storeu_epi8(void * __P, __mmask8 __M, __m512i __A) {
+  // CHECK-LABEL: @test_mm512_mask_cvtusepi64_storeu_epi8
+  // CHECK: @llvm.x86.avx512.mask.pmovus.qb.mem.512
+  return _mm512_mask_cvtusepi64_storeu_epi8(__P, __M, __A); 
+}
+
+__m256i test_mm512_cvtusepi64_epi32(__m512i __A) {
+  // CHECK-LABEL: @test_mm512_cvtusepi64_epi32
+  // CHECK: @llvm.x86.avx512.mask.pmovus.qd.512
+  return _mm512_cvtusepi64_epi32(__A); 
+}
+
+__m256i test_mm512_mask_cvtusepi64_epi32(__m256i __O, __mmask8 __M, __m512i __A) {
+  // CHECK-LABEL: @test_mm512_mask_cvtusepi64_epi32
+  // CHECK: @llvm.x86.avx512.mask.pmovus.qd.512
+  return _mm512_mask_cvtusepi64_epi32(__O, __M, __A); 
+}
+
+__m256i test_mm512_maskz_cvtusepi64_epi32(__mmask8 __M, __m512i __A) {
+  // CHECK-LABEL: @test_mm512_maskz_cvtusepi64_epi32
+  // CHECK: @llvm.x86.avx512.mask.pmovus.qd.512
+  return _mm512_maskz_cvtusepi64_epi32(__M, __A); 
+}
+
+void test_mm512_mask_cvtusepi64_storeu_epi32(void* __P, __mmask8 __M, __m512i __A) {
+  // CHECK-LABEL: @test_mm512_mask_cvtusepi64_storeu_epi32
+  // CHECK: @llvm.x86.avx512.mask.pmovus.qd.mem.512
+  return _mm512_mask_cvtusepi64_storeu_epi32(__P, __M, __A); 
+}
+
+__m128i test_mm512_cvtusepi64_epi16(__m512i __A) {
+  // CHECK-LABEL: @test_mm512_cvtusepi64_epi16
+  // CHECK: @llvm.x86.avx512.mask.pmovus.qw.512
+  return _mm512_cvtusepi64_epi16(__A); 
+}
+
+__m128i test_mm512_mask_cvtusepi64_epi16(__m128i __O, __mmask8 __M, __m512i __A) {
+  // CHECK-LABEL: @test_mm512_mask_cvtusepi64_epi16
+  // CHECK: @llvm.x86.avx512.mask.pmovus.qw.512
+  return _mm512_mask_cvtusepi64_epi16(__O, __M, __A); 
+}
+
+__m128i test_mm512_maskz_cvtusepi64_epi16(__mmask8 __M, __m512i __A) {
+  // CHECK-LABEL: @test_mm512_maskz_cvtusepi64_epi16
+  // CHECK: @llvm.x86.avx512.mask.pmovus.qw.512
+  return _mm512_maskz_cvtusepi64_epi16(__M, __A); 
+}
+
+void test_mm512_mask_cvtusepi64_storeu_epi16(void *__P, __mmask8 __M, __m512i __A) {
+  // CHECK-LABEL: @test_mm512_mask_cvtusepi64_storeu_epi16
+  // CHECK: @llvm.x86.avx512.mask.pmovus.qw.mem.512
+  return _mm512_mask_cvtusepi64_storeu_epi16(__P, __M, __A); 
+}

test/CodeGen/avx512vl-builtins.c

__m256 test_mm256_mask_rsqrt14_ps(__m256 __W, __mmask8 __U, __m256 __A) {
   return _mm256_mask_rsqrt14_ps(__W, __U, __A); 
 }
 
 __m256 test_mm256_maskz_rsqrt14_ps(__mmask8 __U, __m256 __A) {
   // CHECK-LABEL: @test_mm256_maskz_rsqrt14_ps
   // CHECK: @llvm.x86.avx512.rsqrt14.ps.256
   return _mm256_maskz_rsqrt14_ps(__U, __A); 
 }
+
+__m128i test_mm_cvtusepi32_epi8(__m128i __A) {
+  // CHECK-LABEL: @test_mm_cvtusepi32_epi8
+  // CHECK: @llvm.x86.avx512.mask.pmovus.db.128
+  return _mm_cvtusepi32_epi8(__A); 
+}
+
+__m128i test_mm_mask_cvtusepi32_epi8(__m128i __O, __mmask8 __M, __m128i __A) {
+  // CHECK-LABEL: @test_mm_mask_cvtusepi32_epi8
+  // CHECK: @llvm.x86.avx512.mask.pmovus.db.128
+  return _mm_mask_cvtusepi32_epi8(__O, __M, __A); 
+}
+
+__m128i test_mm_maskz_cvtusepi32_epi8(__mmask8 __M, __m128i __A) {
+  // CHECK-LABEL: @test_mm_maskz_cvtusepi32_epi8
+  // CHECK: @llvm.x86.avx512.mask.pmovus.db.128
+  return _mm_maskz_cvtusepi32_epi8(__M, __A); 
+}
+
+void test_mm_mask_cvtusepi32_storeu_epi8(void * __P, __mmask8 __M, __m128i __A) {
+  // CHECK-LABEL: @test_mm_mask_cvtusepi32_storeu_epi8
+  // CHECK: @llvm.x86.avx512.mask.pmovus.db.mem.128
+  return _mm_mask_cvtusepi32_storeu_epi8(__P, __M, __A); 
+}
+
+__m128i test_mm256_cvtusepi32_epi8(__m256i __A) {
+  // CHECK-LABEL: @test_mm256_cvtusepi32_epi8
+  // CHECK: @llvm.x86.avx512.mask.pmovus.db.256
+  return _mm256_cvtusepi32_epi8(__A); 
+}
+
+__m128i test_mm256_mask_cvtusepi32_epi8(__m128i __O, __mmask8 __M, __m256i __A) {
+  // CHECK-LABEL: @test_mm256_mask_cvtusepi32_epi8
+  // CHECK: @llvm.x86.avx512.mask.pmovus.db.256
+  return _mm256_mask_cvtusepi32_epi8(__O, __M, __A); 
+}
+
+__m128i test_mm256_maskz_cvtusepi32_epi8(__mmask8 __M, __m256i __A) {
+  // CHECK-LABEL: @test_mm256_maskz_cvtusepi32_epi8
+  // CHECK: @llvm.x86.avx512.mask.pmovus.db.256
+  return _mm256_maskz_cvtusepi32_epi8(__M, __A); 
+}
+
+void test_mm256_mask_cvtusepi32_storeu_epi8(void * __P, __mmask8 __M, __m256i __A) {
+  // CHECK-LABEL: @test_mm256_mask_cvtusepi32_storeu_epi8
+  // CHECK: @llvm.x86.avx512.mask.pmovus.db.mem.256
+  return _mm256_mask_cvtusepi32_storeu_epi8(__P, __M, __A); 
+}
+
+__m128i test_mm_cvtusepi32_epi16(__m128i __A) {
+  // CHECK-LABEL: @test_mm_cvtusepi32_epi16
+  // CHECK: @llvm.x86.avx512.mask.pmovus.dw.128
+  return _mm_cvtusepi32_epi16(__A); 
+}
+
+__m128i test_mm_mask_cvtusepi32_epi16(__m128i __O, __mmask8 __M, __m128i __A) {
+  // CHECK-LABEL: @test_mm_mask_cvtusepi32_epi16
+  // CHECK: @llvm.x86.avx512.mask.pmovus.dw.128
+  return _mm_mask_cvtusepi32_epi16(__O, __M, __A); 
+}
+
+__m128i test_mm_maskz_cvtusepi32_epi16(__mmask8 __M, __m128i __A) {
+  // CHECK-LABEL: @test_mm_maskz_cvtusepi32_epi16
+  // CHECK: @llvm.x86.avx512.mask.pmovus.dw.128
+  return _mm_maskz_cvtusepi32_epi16(__M, __A); 
+}
+
+void test_mm_mask_cvtusepi32_storeu_epi16(void * __P, __mmask8 __M, __m128i __A) {
+  // CHECK-LABEL: @test_mm_mask_cvtusepi32_storeu_epi16
+  // CHECK: @llvm.x86.avx512.mask.pmovus.dw.mem.128
+  return _mm_mask_cvtusepi32_storeu_epi16(__P, __M, __A); 
+}
+
+__m128i test_mm256_cvtusepi32_epi16(__m256i __A) {
+  // CHECK-LABEL: @test_mm256_cvtusepi32_epi16
+  // CHECK: @llvm.x86.avx512.mask.pmovus.dw.256
+  return _mm256_cvtusepi32_epi16(__A); 
+}
+
+__m128i test_mm256_mask_cvtusepi32_epi16(__m128i __O, __mmask8 __M, __m256i __A) {
+  // CHECK-LABEL: @test_mm256_mask_cvtusepi32_epi16
+  // CHECK: @llvm.x86.avx512.mask.pmovus.dw.256
+  return _mm256_mask_cvtusepi32_epi16(__O, __M, __A); 
+}
+
+__m128i test_mm256_maskz_cvtusepi32_epi16(__mmask8 __M, __m256i __A) {
+  // CHECK-LABEL: @test_mm256_maskz_cvtusepi32_epi16
+  // CHECK: @llvm.x86.avx512.mask.pmovus.dw.256
+  return _mm256_maskz_cvtusepi32_epi16(__M, __A); 
+}
+
+void test_mm256_mask_cvtusepi32_storeu_epi16(void * __P, __mmask8 __M, __m256i __A) {
+  // CHECK-LABEL: @test_mm256_mask_cvtusepi32_storeu_epi16
+  // CHECK: @llvm.x86.avx512.mask.pmovus.dw.mem.256
+  return _mm256_mask_cvtusepi32_storeu_epi16(__P, __M, __A); 
+}
+
+__m128i test_mm_cvtusepi64_epi8(__m128i __A) {
+  // CHECK-LABEL: @test_mm_cvtusepi64_epi8
+  // CHECK: @llvm.x86.avx512.mask.pmovus.qb.128
+  return _mm_cvtusepi64_epi8(__A); 
+}
+
+__m128i test_mm_mask_cvtusepi64_epi8(__m128i __O, __mmask8 __M, __m128i __A) {
+  // CHECK-LABEL: @test_mm_mask_cvtusepi64_epi8
+  // CHECK: @llvm.x86.avx512.mask.pmovus.qb.128
+  return _mm_mask_cvtusepi64_epi8(__O, __M, __A); 
+}
+
+__m128i test_mm_maskz_cvtusepi64_epi8(__mmask8 __M, __m128i __A) {
+  // CHECK-LABEL: @test_mm_maskz_cvtusepi64_epi8
+  // CHECK: @llvm.x86.avx512.mask.pmovus.qb.128
+  return _mm_maskz_cvtusepi64_epi8(__M, __A); 
+}
+
+void test_mm_mask_cvtusepi64_storeu_epi8(void * __P, __mmask8 __M, __m128i __A) {
+  // CHECK-LABEL: @test_mm_mask_cvtusepi64_storeu_epi8
+  // CHECK: @llvm.x86.avx512.mask.pmovus.qb.mem.128
+  return _mm_mask_cvtusepi64_storeu_epi8(__P, __M, __A); 
+}
+
+__m128i test_mm256_cvtusepi64_epi8(__m256i __A) {
+  // CHECK-LABEL: @test_mm256_cvtusepi64_epi8
+  // CHECK: @llvm.x86.avx512.mask.pmovus.qb.256
+  return _mm256_cvtusepi64_epi8(__A); 
+}
+
+__m128i test_mm256_mask_cvtusepi64_epi8(__m128i __O, __mmask8 __M, __m256i __A) {
+  // CHECK-LABEL: @test_mm256_mask_cvtusepi64_epi8
+  // CHECK: @llvm.x86.avx512.mask.pmovus.qb.256
+  return _mm256_mask_cvtusepi64_epi8(__O, __M, __A); 
+}
+
+__m128i test_mm256_maskz_cvtusepi64_epi8(__mmask8 __M, __m256i __A) {
+  // CHECK-LABEL: @test_mm256_maskz_cvtusepi64_epi8
+  // CHECK: @llvm.x86.avx512.mask.pmovus.qb.256
+  return _mm256_maskz_cvtusepi64_epi8(__M, __A); 
+}
+
+void test_mm256_mask_cvtusepi64_storeu_epi8(void * __P, __mmask8 __M, __m256i __A) {
+  // CHECK-LABEL: @test_mm256_mask_cvtusepi64_storeu_epi8
+  // CHECK: @llvm.x86.avx512.mask.pmovus.qb.mem.256
+  return _mm256_mask_cvtusepi64_storeu_epi8(__P, __M, __A); 
+}
+
+__m128i test_mm_cvtusepi64_epi32(__m128i __A) {
+  // CHECK-LABEL: @test_mm_cvtusepi64_epi32
+  // CHECK: @llvm.x86.avx512.mask.pmovus.qd.128
+  return _mm_cvtusepi64_epi32(__A); 
+}
+
+__m128i test_mm_mask_cvtusepi64_epi32(__m128i __O, __mmask8 __M, __m128i __A) {
+  // CHECK-LABEL: @test_mm_mask_cvtusepi64_epi32
+  // CHECK: @llvm.x86.avx512.mask.pmovus.qd.128
+  return _mm_mask_cvtusepi64_epi32(__O, __M, __A); 
+}
+
+__m128i test_mm_maskz_cvtusepi64_epi32(__mmask8 __M, __m128i __A) {
+  // CHECK-LABEL: @test_mm_maskz_cvtusepi64_epi32
+  // CHECK: @llvm.x86.avx512.mask.pmovus.qd.128
+  return _mm_maskz_cvtusepi64_epi32(__M, __A); 
+}
+
+void test_mm_mask_cvtusepi64_storeu_epi32(void * __P, __mmask8 __M, __m128i __A) {
+  // CHECK-LABEL: @test_mm_mask_cvtusepi64_storeu_epi32
+  // CHECK: @llvm.x86.avx512.mask.pmovus.qd.mem.128
+  return _mm_mask_cvtusepi64_storeu_epi32(__P, __M, __A); 
+}
+
+__m128i test_mm256_cvtusepi64_epi32(__m256i __A) {
+  // CHECK-LABEL: @test_mm256_cvtusepi64_epi32
+  // CHECK: @llvm.x86.avx512.mask.pmovus.qd.256
+  return _mm256_cvtusepi64_epi32(__A); 
+}
+
+__m128i test_mm256_mask_cvtusepi64_epi32(__m128i __O, __mmask8 __M, __m256i __A) {
+  // CHECK-LABEL: @test_mm256_mask_cvtusepi64_epi32
+  // CHECK: @llvm.x86.avx512.mask.pmovus.qd.256
+  return _mm256_mask_cvtusepi64_epi32(__O, __M, __A); 
+}
+
+__m128i test_mm256_maskz_cvtusepi64_epi32(__mmask8 __M, __m256i __A) {
+  // CHECK-LABEL: @test_mm256_maskz_cvtusepi64_epi32
+  // CHECK: @llvm.x86.avx512.mask.pmovus.qd.256
+  return _mm256_maskz_cvtusepi64_epi32(__M, __A); 
+}
+
+void test_mm256_mask_cvtusepi64_storeu_epi32(void * __P, __mmask8 __M, __m256i __A) {
+  // CHECK-LABEL: @test_mm256_mask_cvtusepi64_storeu_epi32
+  // CHECK: @llvm.x86.avx512.mask.pmovus.qd.mem.256
+  return _mm256_mask_cvtusepi64_storeu_epi32(__P, __M, __A); 
+}
+
+__m128i test_mm_cvtusepi64_epi16(__m128i __A) {
+  // CHECK-LABEL: @test_mm_cvtusepi64_epi16
+  // CHECK: @llvm.x86.avx512.mask.pmovus.qw.128
+  return _mm_cvtusepi64_epi16(__A); 
+}
+
+__m128i test_mm_mask_cvtusepi64_epi16(__m128i __O, __mmask8 __M, __m128i __A) {
+  // CHECK-LABEL: @test_mm_mask_cvtusepi64_epi16
+  // CHECK: @llvm.x86.avx512.mask.pmovus.qw.128
+  return _mm_mask_cvtusepi64_epi16(__O, __M, __A); 
+}
+
+__m128i test_mm_maskz_cvtusepi64_epi16(__mmask8 __M, __m128i __A) {
+  // CHECK-LABEL: @test_mm_maskz_cvtusepi64_epi16
+  // CHECK: @llvm.x86.avx512.mask.pmovus.qw.128
+  return _mm_maskz_cvtusepi64_epi16(__M, __A); 
+}
+
+void test_mm_mask_cvtusepi64_storeu_epi16(void * __P, __mmask8 __M, __m128i __A) {
+  // CHECK-LABEL: @test_mm_mask_cvtusepi64_storeu_epi16
+  // CHECK: @llvm.x86.avx512.mask.pmovus.qw.mem.128
+  return _mm_mask_cvtusepi64_storeu_epi16(__P, __M, __A); 
+}
+
+__m128i test_mm256_cvtusepi64_epi16(__m256i __A) {
+  // CHECK-LABEL: @test_mm256_cvtusepi64_epi16
+  // CHECK: @llvm.x86.avx512.mask.pmovus.qw.256
+  return _mm256_cvtusepi64_epi16(__A); 
+}
+
+__m128i test_mm256_mask_cvtusepi64_epi16(__m128i __O, __mmask8 __M, __m256i __A) {
+  // CHECK-LABEL: @test_mm256_mask_cvtusepi64_epi16
+  // CHECK: @llvm.x86.avx512.mask.pmovus.qw.256
+  return _mm256_mask_cvtusepi64_epi16(__O, __M, __A); 
+}
+
+__m128i test_mm256_maskz_cvtusepi64_epi16(__mmask8 __M, __m256i __A) {
+  // CHECK-LABEL: @test_mm256_maskz_cvtusepi64_epi16
+  // CHECK: @llvm.x86.avx512.mask.pmovus.qw.256
+  return _mm256_maskz_cvtusepi64_epi16(__M, __A); 
+}
+
+void test_mm256_mask_cvtusepi64_storeu_epi16(void * __P, __mmask8 __M, __m256i __A) {
+  // CHECK-LABEL: @test_mm256_mask_cvtusepi64_storeu_epi16
+  // CHECK: @llvm.x86.avx512.mask.pmovus.qw.mem.256
+  return _mm256_mask_cvtusepi64_storeu_epi16(__P, __M, __A); 
+}