[X86][AVX512F] Add FP scalar intrinsics

[AsafBadouh]

Summary

intrinsics for: add/sub/mul/div/min/max in their FP scalar versions

[AsafBadouh]

change builtin name: _round -> _mask

BE change in http://reviews.llvm.org/D11454

[delena]

LGTM

include/clang/Basic/BuiltinsX86.def

 BUILTIN(__builtin_ia32_divpd512_mask, "V8dV8dV8dV8dUcIi", "")
 BUILTIN(__builtin_ia32_divps512_mask, "V16fV16fV16fV16fUsIi", "")
 BUILTIN(__builtin_ia32_mulpd512_mask, "V8dV8dV8dV8dUcIi", "")
 BUILTIN(__builtin_ia32_mulps512_mask, "V16fV16fV16fV16fUsIi", "")
 BUILTIN(__builtin_ia32_subpd512_mask, "V8dV8dV8dV8dUcIi", "")
 BUILTIN(__builtin_ia32_subps512_mask, "V16fV16fV16fV16fUsIi", "")
 BUILTIN(__builtin_ia32_pmaddubsw512_mask, "V32sV64cV64cV32sUi", "")
 BUILTIN(__builtin_ia32_pmaddwd512_mask, "V16iV32sV32sV16iUs", "")
+BUILTIN(__builtin_ia32_addss_mask, "V4fV4fV4fV4fUcIi", "")
+BUILTIN(__builtin_ia32_divss_mask, "V4fV4fV4fV4fUcIi", "")
+BUILTIN(__builtin_ia32_mulss_mask, "V4fV4fV4fV4fUcIi", "")
+BUILTIN(__builtin_ia32_subss_mask, "V4fV4fV4fV4fUcIi", "")
+BUILTIN(__builtin_ia32_maxss_mask, "V4fV4fV4fV4fUcIi", "")
+BUILTIN(__builtin_ia32_minss_mask, "V4fV4fV4fV4fUcIi", "")
+BUILTIN(__builtin_ia32_addsd_mask, "V2dV2dV2dV2dUcIi", "")
+BUILTIN(__builtin_ia32_divsd_mask, "V2dV2dV2dV2dUcIi", "")
+BUILTIN(__builtin_ia32_mulsd_mask, "V2dV2dV2dV2dUcIi", "")
+BUILTIN(__builtin_ia32_subsd_mask, "V2dV2dV2dV2dUcIi", "")
+BUILTIN(__builtin_ia32_maxsd_mask, "V2dV2dV2dV2dUcIi", "")
+BUILTIN(__builtin_ia32_minsd_mask, "V2dV2dV2dV2dUcIi", "")
 
 #undef BUILTIN

lib/Headers/avx512fintrin.h

_mm512_max_ps(__m512 __A, __m512 __B)
   return (__m512) __builtin_ia32_maxps512_mask ((__v16sf) __A,
             (__v16sf) __B,
             (__v16sf)
             _mm512_setzero_ps (),
             (__mmask16) -1,
             _MM_FROUND_CUR_DIRECTION);
 }
 
+static __inline__ __m128 __DEFAULT_FN_ATTRS
+_mm_mask_max_ss(__m128 __W, __mmask8 __U,__m128 __A, __m128 __B) {
+  return (__m128) __builtin_ia32_maxss_mask ((__v4sf) __A,
+                (__v4sf) __B,
+                (__v4sf) __W,
+                (__mmask8) __U,
+                _MM_FROUND_CUR_DIRECTION);
+}
+
+static __inline__ __m128 __DEFAULT_FN_ATTRS
+_mm_maskz_max_ss(__mmask8 __U,__m128 __A, __m128 __B) {
+  return (__m128) __builtin_ia32_maxss_mask ((__v4sf) __A,
+                (__v4sf) __B,
+                (__v4sf)  _mm_setzero_ps (),
+                (__mmask8) __U,
+                _MM_FROUND_CUR_DIRECTION);
+}
+
+#define _mm_max_round_ss(__A, __B, __R) __extension__ ({ \
+  (__m128) __builtin_ia32_maxss_mask ((__v4sf) __A, (__v4sf) __B, \
+                (__v4sf) _mm_setzero_ps(), (__mmask8) -1, __R); })
+
+#define _mm_mask_max_round_ss(__W, __U, __A, __B, __R) __extension__ ({ \
+  (__m128) __builtin_ia32_maxss_mask ((__v4sf) __A, (__v4sf) __B, \
+                (__v4sf)  __W, (__mmask8) __U,__R); })
+
+#define _mm_maskz_max_round_ss(__U, __A, __B, __R) __extension__ ({ \
+  (__m128) __builtin_ia32_maxss_mask ((__v4sf) __A, (__v4sf) __B, \
+                (__v4sf)  _mm_setzero_ps(), (__mmask8) __U,__R); })
+
+static __inline__ __m128d __DEFAULT_FN_ATTRS
+_mm_mask_max_sd(__m128d __W, __mmask8 __U,__m128d __A, __m128d __B) {
+  return (__m128d) __builtin_ia32_maxsd_mask ((__v2df) __A,
+                (__v2df) __B,
+                (__v2df) __W,
+                (__mmask8) __U,
+                _MM_FROUND_CUR_DIRECTION);
+}
+
+static __inline__ __m128d __DEFAULT_FN_ATTRS
+_mm_maskz_max_sd(__mmask8 __U,__m128d __A, __m128d __B) {
+  return (__m128d) __builtin_ia32_maxsd_mask ((__v2df) __A,
+                (__v2df) __B,
+                (__v2df)  _mm_setzero_pd (),
+                (__mmask8) __U,
+                _MM_FROUND_CUR_DIRECTION);
+}
+
+#define _mm_max_round_sd(__A, __B, __R) __extension__ ({ \
+  (__m128d) __builtin_ia32_maxsd_mask ((__v2df) __A, (__v2df) __B, \
+                (__v2df) _mm_setzero_pd(), (__mmask8) -1, __R); })
+
+#define _mm_mask_max_round_sd(__W, __U, __A, __B, __R) __extension__ ({ \
+  (__m128d) __builtin_ia32_maxsd_mask ((__v2df) __A, (__v2df) __B, \
+                (__v2df)  __W, (__mmask8) __U,__R); })
+
+#define _mm_maskz_max_round_sd(__U, __A, __B, __R) __extension__ ({ \
+  (__m128d) __builtin_ia32_maxsd_mask ((__v2df) __A, (__v2df) __B, \
+                (__v2df)  _mm_setzero_pd(), (__mmask8) __U,__R); })
+
 static __inline __m512i
 __DEFAULT_FN_ATTRS
 _mm512_max_epi32(__m512i __A, __m512i __B)
 {
   return (__m512i) __builtin_ia32_pmaxsd512_mask ((__v16si) __A,
               (__v16si) __B,
               (__v16si)
               _mm512_setzero_si512 (),
_mm512_min_ps(__m512 __A, __m512 __B)
   return (__m512) __builtin_ia32_minps512_mask ((__v16sf) __A,
             (__v16sf) __B,
             (__v16sf)
             _mm512_setzero_ps (),
             (__mmask16) -1,
             _MM_FROUND_CUR_DIRECTION);
 }
 
+static __inline__ __m128 __DEFAULT_FN_ATTRS
+_mm_mask_min_ss(__m128 __W, __mmask8 __U,__m128 __A, __m128 __B) {
+  return (__m128) __builtin_ia32_minss_mask ((__v4sf) __A,
+                (__v4sf) __B,
+                (__v4sf) __W,
+                (__mmask8) __U,
+                _MM_FROUND_CUR_DIRECTION);
+}
+
+static __inline__ __m128 __DEFAULT_FN_ATTRS
+_mm_maskz_min_ss(__mmask8 __U,__m128 __A, __m128 __B) {
+  return (__m128) __builtin_ia32_minss_mask ((__v4sf) __A,
+                (__v4sf) __B,
+                (__v4sf)  _mm_setzero_ps (),
+                (__mmask8) __U,
+                _MM_FROUND_CUR_DIRECTION);
+}
+
+#define _mm_min_round_ss(__A, __B, __R) __extension__ ({ \
+  (__m128) __builtin_ia32_minss_mask ((__v4sf) __A, (__v4sf) __B, \
+                (__v4sf) _mm_setzero_ps(), (__mmask8) -1, __R); })
+
+#define _mm_mask_min_round_ss(__W, __U, __A, __B, __R) __extension__ ({ \
+  (__m128) __builtin_ia32_minss_mask ((__v4sf) __A, (__v4sf) __B, \
+                (__v4sf)  __W, (__mmask8) __U,__R); })
+
+#define _mm_maskz_min_round_ss(__U, __A, __B, __R) __extension__ ({ \
+  (__m128) __builtin_ia32_minss_mask ((__v4sf) __A, (__v4sf) __B, \
+                (__v4sf)  _mm_setzero_ps(), (__mmask8) __U,__R); })
+
+static __inline__ __m128d __DEFAULT_FN_ATTRS
+_mm_mask_min_sd(__m128d __W, __mmask8 __U,__m128d __A, __m128d __B) {
+  return (__m128d) __builtin_ia32_minsd_mask ((__v2df) __A,
+                (__v2df) __B,
+                (__v2df) __W,
+                (__mmask8) __U,
+                _MM_FROUND_CUR_DIRECTION);
+}
+
+static __inline__ __m128d __DEFAULT_FN_ATTRS
+_mm_maskz_min_sd(__mmask8 __U,__m128d __A, __m128d __B) {
+  return (__m128d) __builtin_ia32_minsd_mask ((__v2df) __A,
+                (__v2df) __B,
+                (__v2df)  _mm_setzero_pd (),
+                (__mmask8) __U,
+                _MM_FROUND_CUR_DIRECTION);
+}
+
+#define _mm_min_round_sd(__A, __B, __R) __extension__ ({ \
+  (__m128d) __builtin_ia32_minsd_mask ((__v2df) __A, (__v2df) __B, \
+                (__v2df) _mm_setzero_pd(), (__mmask8) -1, __R); })
+
+#define _mm_mask_min_round_sd(__W, __U, __A, __B, __R) __extension__ ({ \
+  (__m128d) __builtin_ia32_minsd_mask ((__v2df) __A, (__v2df) __B, \
+                (__v2df)  __W, (__mmask8) __U,__R); })
+
+#define _mm_maskz_min_round_sd(__U, __A, __B, __R) __extension__ ({ \
+  (__m128d) __builtin_ia32_minsd_mask ((__v2df) __A, (__v2df) __B, \
+                (__v2df)  _mm_setzero_pd(), (__mmask8) __U,__R); })
+
 static __inline __m512i
 __DEFAULT_FN_ATTRS
 _mm512_min_epi32(__m512i __A, __m512i __B)
 {
   return (__m512i) __builtin_ia32_pminsd512_mask ((__v16si) __A,
               (__v16si) __B,
               (__v16si)
               _mm512_setzero_si512 (),

 _mm512_abs_epi32(__m512i __A)
 {
   return (__m512i) __builtin_ia32_pabsd512_mask ((__v16si) __A,
              (__v16si)
              _mm512_setzero_si512 (),
              (__mmask16) -1);
 }
 
+static __inline__ __m128 __DEFAULT_FN_ATTRS
+_mm_mask_add_ss(__m128 __W, __mmask8 __U,__m128 __A, __m128 __B) {
+  return (__m128) __builtin_ia32_addss_mask ((__v4sf) __A,
+                (__v4sf) __B,
+                (__v4sf) __W,
+                (__mmask8) __U,
+                _MM_FROUND_CUR_DIRECTION);
+}
+
+static __inline__ __m128 __DEFAULT_FN_ATTRS
+_mm_maskz_add_ss(__mmask8 __U,__m128 __A, __m128 __B) {
+  return (__m128) __builtin_ia32_addss_mask ((__v4sf) __A,
+                (__v4sf) __B,
+                (__v4sf)  _mm_setzero_ps (),
+                (__mmask8) __U,
+                _MM_FROUND_CUR_DIRECTION);
+}
+
+#define _mm_add_round_ss(__A, __B, __R) __extension__ ({ \
+  (__m128) __builtin_ia32_addss_mask ((__v4sf) __A, (__v4sf) __B, \
+                (__v4sf) _mm_setzero_ps(), (__mmask8) -1, __R); })
+
+#define _mm_mask_add_round_ss(__W, __U, __A, __B, __R) __extension__ ({ \
+  (__m128) __builtin_ia32_addss_mask ((__v4sf) __A, (__v4sf) __B, \
+                (__v4sf)  __W, (__mmask8) __U,__R); })
+
+#define _mm_maskz_add_round_ss(__U, __A, __B, __R) __extension__ ({ \
+  (__m128) __builtin_ia32_addss_mask ((__v4sf) __A, (__v4sf) __B, \
+                (__v4sf)  _mm_setzero_ps(), (__mmask8) __U,__R); })
+
+static __inline__ __m128d __DEFAULT_FN_ATTRS
+_mm_mask_add_sd(__m128d __W, __mmask8 __U,__m128d __A, __m128d __B) {
+  return (__m128d) __builtin_ia32_addsd_mask ((__v2df) __A,
+                (__v2df) __B,
+                (__v2df) __W,
+                (__mmask8) __U,
+                _MM_FROUND_CUR_DIRECTION);
+}
+
+static __inline__ __m128d __DEFAULT_FN_ATTRS
+_mm_maskz_add_sd(__mmask8 __U,__m128d __A, __m128d __B) {
+  return (__m128d) __builtin_ia32_addsd_mask ((__v2df) __A,
+                (__v2df) __B,
+                (__v2df)  _mm_setzero_pd (),
+                (__mmask8) __U,
+                _MM_FROUND_CUR_DIRECTION);
+}
+#define _mm_add_round_sd(__A, __B, __R) __extension__ ({ \
+  (__m128d) __builtin_ia32_addsd_mask ((__v2df) __A, (__v2df) __B, \
+                (__v2df) _mm_setzero_pd(), (__mmask8) -1, __R); })
+
+#define _mm_mask_add_round_sd(__W, __U, __A, __B, __R) __extension__ ({ \
+  (__m128d) __builtin_ia32_addsd_mask ((__v2df) __A, (__v2df) __B, \
+                (__v2df)  __W, (__mmask8) __U,__R); })
+
+#define _mm_maskz_add_round_sd(__U, __A, __B, __R) __extension__ ({ \
+  (__m128d) __builtin_ia32_addsd_mask ((__v2df) __A, (__v2df) __B, \
+                (__v2df)  _mm_setzero_pd(), (__mmask8) __U,__R); })
+
 static __inline__ __m512d __DEFAULT_FN_ATTRS
 _mm512_mask_add_pd(__m512d __W, __mmask8 __U, __m512d __A, __m512d __B) {
   return (__m512d) __builtin_ia32_addpd512_mask ((__v8df) __A,
              (__v8df) __B,
              (__v8df) __W,
              (__mmask8) __U,
              _MM_FROUND_CUR_DIRECTION);
 }

 #define _mm512_mask_add_round_ps(__W, __U, __A, __B, __R) __extension__ ({ \
   (__m512) __builtin_ia32_addps512_mask ((__v16sf) __A, (__v16sf) __B, \
                 (__v16sf) __W, (__mmask16)__U, __R); })
 
 #define _mm512_maskz_add_round_ps(__U, __A, __B, __R) __extension__ ({ \
   (__m512) __builtin_ia32_addps512_mask ((__v16sf) __A, (__v16sf) __B, \
                 (__v16sf) _mm512_setzero_ps(), (__mmask16)__U, __R); })
 
+static __inline__ __m128 __DEFAULT_FN_ATTRS
+_mm_mask_sub_ss(__m128 __W, __mmask8 __U,__m128 __A, __m128 __B) {
+  return (__m128) __builtin_ia32_subss_mask ((__v4sf) __A,
+                (__v4sf) __B,
+                (__v4sf) __W,
+                (__mmask8) __U,
+                _MM_FROUND_CUR_DIRECTION);
+}
+
+static __inline__ __m128 __DEFAULT_FN_ATTRS
+_mm_maskz_sub_ss(__mmask8 __U,__m128 __A, __m128 __B) {
+  return (__m128) __builtin_ia32_subss_mask ((__v4sf) __A,
+                (__v4sf) __B,
+                (__v4sf)  _mm_setzero_ps (),
+                (__mmask8) __U,
+                _MM_FROUND_CUR_DIRECTION);
+}
+#define _mm_sub_round_ss(__A, __B, __R) __extension__ ({ \
+  (__m128) __builtin_ia32_subss_mask ((__v4sf) __A, (__v4sf) __B, \
+                (__v4sf) _mm_setzero_ps(), (__mmask8) -1, __R); })
+
+#define _mm_mask_sub_round_ss(__W, __U, __A, __B, __R) __extension__ ({ \
+  (__m128) __builtin_ia32_subss_mask ((__v4sf) __A, (__v4sf) __B, \
+                (__v4sf)  __W, (__mmask8) __U,__R); })
+
+#define _mm_maskz_sub_round_ss(__U, __A, __B, __R) __extension__ ({ \
+  (__m128) __builtin_ia32_subss_mask ((__v4sf) __A, (__v4sf) __B, \
+                (__v4sf)  _mm_setzero_ps(), (__mmask8) __U,__R); })
+
+static __inline__ __m128d __DEFAULT_FN_ATTRS
+_mm_mask_sub_sd(__m128d __W, __mmask8 __U,__m128d __A, __m128d __B) {
+  return (__m128d) __builtin_ia32_subsd_mask ((__v2df) __A,
+                (__v2df) __B,
+                (__v2df) __W,
+                (__mmask8) __U,
+                _MM_FROUND_CUR_DIRECTION);
+}
+
+static __inline__ __m128d __DEFAULT_FN_ATTRS
+_mm_maskz_sub_sd(__mmask8 __U,__m128d __A, __m128d __B) {
+  return (__m128d) __builtin_ia32_subsd_mask ((__v2df) __A,
+                (__v2df) __B,
+                (__v2df)  _mm_setzero_pd (),
+                (__mmask8) __U,
+                _MM_FROUND_CUR_DIRECTION);
+}
+
+#define _mm_sub_round_sd(__A, __B, __R) __extension__ ({ \
+  (__m128d) __builtin_ia32_subsd_mask ((__v2df) __A, (__v2df) __B, \
+                (__v2df) _mm_setzero_pd(), (__mmask8) -1, __R); })
+
+#define _mm_mask_sub_round_sd(__W, __U, __A, __B, __R) __extension__ ({ \
+  (__m128d) __builtin_ia32_subsd_mask ((__v2df) __A, (__v2df) __B, \
+                (__v2df)  __W, (__mmask8) __U,__R); })
+
+#define _mm_maskz_sub_round_sd(__U, __A, __B, __R) __extension__ ({ \
+  (__m128d) __builtin_ia32_subsd_mask ((__v2df) __A, (__v2df) __B, \
+                (__v2df)  _mm_setzero_pd(), (__mmask8) __U,__R); })
+
 static __inline__ __m512d __DEFAULT_FN_ATTRS
 _mm512_mask_sub_pd(__m512d __W, __mmask8 __U, __m512d __A, __m512d __B) {
   return (__m512d) __builtin_ia32_subpd512_mask ((__v8df) __A,
              (__v8df) __B,
              (__v8df) __W,
              (__mmask8) __U,
              _MM_FROUND_CUR_DIRECTION);
 }

 #define _mm512_mask_sub_round_ps(__W, __U, __A, __B, __R)  __extension__ ({ \
   (__m512) __builtin_ia32_subps512_mask ((__v16sf) __A, (__v16sf) __B, \
             (__v16sf) __W, (__mmask16) __U, __R); });
 
 #define _mm512_maskz_sub_round_ps(__U, __A, __B, __R)  __extension__ ({ \
   (__m512) __builtin_ia32_subps512_mask ((__v16sf) __A, (__v16sf) __B, \
             (__v16sf) _mm512_setzero_ps (), (__mmask16) __U, __R);});
 
+static __inline__ __m128 __DEFAULT_FN_ATTRS
+_mm_mask_mul_ss(__m128 __W, __mmask8 __U,__m128 __A, __m128 __B) {
+  return (__m128) __builtin_ia32_mulss_mask ((__v4sf) __A,
+                (__v4sf) __B,
+                (__v4sf) __W,
+                (__mmask8) __U,
+                _MM_FROUND_CUR_DIRECTION);
+}
+
+static __inline__ __m128 __DEFAULT_FN_ATTRS
+_mm_maskz_mul_ss(__mmask8 __U,__m128 __A, __m128 __B) {
+  return (__m128) __builtin_ia32_mulss_mask ((__v4sf) __A,
+                (__v4sf) __B,
+                (__v4sf)  _mm_setzero_ps (),
+                (__mmask8) __U,
+                _MM_FROUND_CUR_DIRECTION);
+}
+#define _mm_mul_round_ss(__A, __B, __R) __extension__ ({ \
+  (__m128) __builtin_ia32_mulss_mask ((__v4sf) __A, (__v4sf) __B, \
+                (__v4sf) _mm_setzero_ps(), (__mmask8) -1, __R); })
+
+#define _mm_mask_mul_round_ss(__W, __U, __A, __B, __R) __extension__ ({ \
+  (__m128) __builtin_ia32_mulss_mask ((__v4sf) __A, (__v4sf) __B, \
+                (__v4sf)  __W, (__mmask8) __U,__R); })
+
+#define _mm_maskz_mul_round_ss(__U, __A, __B, __R) __extension__ ({ \
+  (__m128) __builtin_ia32_mulss_mask ((__v4sf) __A, (__v4sf) __B, \
+                (__v4sf)  _mm_setzero_ps(), (__mmask8) __U,__R); })
+
+static __inline__ __m128d __DEFAULT_FN_ATTRS
+_mm_mask_mul_sd(__m128d __W, __mmask8 __U,__m128d __A, __m128d __B) {
+  return (__m128d) __builtin_ia32_mulsd_mask ((__v2df) __A,
+                (__v2df) __B,
+                (__v2df) __W,
+                (__mmask8) __U,
+                _MM_FROUND_CUR_DIRECTION);
+}
+
+static __inline__ __m128d __DEFAULT_FN_ATTRS
+_mm_maskz_mul_sd(__mmask8 __U,__m128d __A, __m128d __B) {
+  return (__m128d) __builtin_ia32_mulsd_mask ((__v2df) __A,
+                (__v2df) __B,
+                (__v2df)  _mm_setzero_pd (),
+                (__mmask8) __U,
+                _MM_FROUND_CUR_DIRECTION);
+}
+
+#define _mm_mul_round_sd(__A, __B, __R) __extension__ ({ \
+  (__m128d) __builtin_ia32_mulsd_mask ((__v2df) __A, (__v2df) __B, \
+                (__v2df) _mm_setzero_pd(), (__mmask8) -1, __R); })
+
+#define _mm_mask_mul_round_sd(__W, __U, __A, __B, __R) __extension__ ({ \
+  (__m128d) __builtin_ia32_mulsd_mask ((__v2df) __A, (__v2df) __B, \
+                (__v2df)  __W, (__mmask8) __U,__R); })
+
+#define _mm_maskz_mul_round_sd(__U, __A, __B, __R) __extension__ ({ \
+  (__m128d) __builtin_ia32_mulsd_mask ((__v2df) __A, (__v2df) __B, \
+                (__v2df)  _mm_setzero_pd(), (__mmask8) __U,__R); })
+
 static __inline__ __m512d __DEFAULT_FN_ATTRS
 _mm512_mask_mul_pd(__m512d __W, __mmask8 __U, __m512d __A, __m512d __B) {
   return (__m512d) __builtin_ia32_mulpd512_mask ((__v8df) __A,
              (__v8df) __B,
              (__v8df) __W,
              (__mmask8) __U,
              _MM_FROUND_CUR_DIRECTION);
 }

 #define _mm512_mask_mul_round_ps(__W, __U, __A, __B, __R)  __extension__ ({ \
   (__m512) __builtin_ia32_mulps512_mask ((__v16sf) __A, (__v16sf) __B, \
             (__v16sf) __W, (__mmask16) __U, __R); });
 
 #define _mm512_maskz_mul_round_ps(__U, __A, __B, __R)  __extension__ ({ \
   (__m512) __builtin_ia32_mulps512_mask ((__v16sf) __A, (__v16sf) __B, \
             (__v16sf) _mm512_setzero_ps (), (__mmask16) __U, __R);});
 
+static __inline__ __m128 __DEFAULT_FN_ATTRS
+_mm_mask_div_ss(__m128 __W, __mmask8 __U,__m128 __A, __m128 __B) {
+  return (__m128) __builtin_ia32_divss_mask ((__v4sf) __A,
+                (__v4sf) __B,
+                (__v4sf) __W,
+                (__mmask8) __U,
+                _MM_FROUND_CUR_DIRECTION);
+}
+
+static __inline__ __m128 __DEFAULT_FN_ATTRS
+_mm_maskz_div_ss(__mmask8 __U,__m128 __A, __m128 __B) {
+  return (__m128) __builtin_ia32_divss_mask ((__v4sf) __A,
+                (__v4sf) __B,
+                (__v4sf)  _mm_setzero_ps (),
+                (__mmask8) __U,
+                _MM_FROUND_CUR_DIRECTION);
+}
+
+#define _mm_div_round_ss(__A, __B, __R) __extension__ ({ \
+  (__m128) __builtin_ia32_divss_mask ((__v4sf) __A, (__v4sf) __B, \
+                (__v4sf) _mm_setzero_ps(), (__mmask8) -1, __R); })
+
+#define _mm_mask_div_round_ss(__W, __U, __A, __B, __R) __extension__ ({ \
+  (__m128) __builtin_ia32_divss_mask ((__v4sf) __A, (__v4sf) __B, \
+                (__v4sf)  __W, (__mmask8) __U,__R); })
+
+#define _mm_maskz_div_round_ss(__U, __A, __B, __R) __extension__ ({ \
+  (__m128) __builtin_ia32_divss_mask ((__v4sf) __A, (__v4sf) __B, \
+                (__v4sf)  _mm_setzero_ps(), (__mmask8) __U,__R); })
+
+static __inline__ __m128d __DEFAULT_FN_ATTRS
+_mm_mask_div_sd(__m128d __W, __mmask8 __U,__m128d __A, __m128d __B) {
+  return (__m128d) __builtin_ia32_divsd_mask ((__v2df) __A,
+                (__v2df) __B,
+                (__v2df) __W,
+                (__mmask8) __U,
+                _MM_FROUND_CUR_DIRECTION);
+}
+
+static __inline__ __m128d __DEFAULT_FN_ATTRS
+_mm_maskz_div_sd(__mmask8 __U,__m128d __A, __m128d __B) {
+  return (__m128d) __builtin_ia32_divsd_mask ((__v2df) __A,
+                (__v2df) __B,
+                (__v2df)  _mm_setzero_pd (),
+                (__mmask8) __U,
+                _MM_FROUND_CUR_DIRECTION);
+}
+
+#define _mm_div_round_sd(__A, __B, __R) __extension__ ({ \
+  (__m128d) __builtin_ia32_divsd_mask ((__v2df) __A, (__v2df) __B, \
+                (__v2df) _mm_setzero_pd(), (__mmask8) -1, __R); })
+
+#define _mm_mask_div_round_sd(__W, __U, __A, __B, __R) __extension__ ({ \
+  (__m128d) __builtin_ia32_divsd_mask ((__v2df) __A, (__v2df) __B, \
+                (__v2df)  __W, (__mmask8) __U,__R); })
+
+#define _mm_maskz_div_round_sd(__U, __A, __B, __R) __extension__ ({ \
+  (__m128d) __builtin_ia32_divsd_mask ((__v2df) __A, (__v2df) __B, \
+                (__v2df)  _mm_setzero_pd(), (__mmask8) __U,__R); })
+
 static __inline__ __m512d __DEFAULT_FN_ATTRS
 _mm512_mask_div_pd(__m512d __W, __mmask8 __U, __m512d __A, __m512d __B) {
   return (__m512d) __builtin_ia32_divpd512_mask ((__v8df) __A,
              (__v8df) __B,
              (__v8df) __W,
              (__mmask8) __U,
              _MM_FROUND_CUR_DIRECTION);
 }

test/CodeGen/avx512f-builtins.c

__m512 test_mm512_mask_add_ps(__m512 __W, __mmask16 __U, __m512 __A, __m512 __B) {
   // CHECK: @llvm.x86.avx512.mask.add.ps.512
   return _mm512_mask_add_ps(__W,__U,__A,__B); 
 }
 __m512 test_mm512_maskz_add_ps(__mmask16 __U, __m512 __A, __m512 __B) {
   // CHECK-LABEL: @test_mm512_maskz_add_ps
   // CHECK: @llvm.x86.avx512.mask.add.ps.512
   return _mm512_maskz_add_ps(__U,__A,__B); 
 }
+__m128 test_mm_add_round_ss(__m128 __A, __m128 __B) {
+  // CHECK-LABEL: @test_mm_add_round_ss
+  // CHECK: @llvm.x86.avx512.mask.add.ss.mask
+  return _mm_add_round_ss(__A,__B,_MM_FROUND_TO_NEAREST_INT); 
+}
+__m128 test_mm_mask_add_round_ss(__m128 __W, __mmask8 __U, __m128 __A, __m128 __B) {
+  // CHECK-LABEL: @test_mm_mask_add_round_ss
+  // CHECK: @llvm.x86.avx512.mask.add.ss.mask
+  return _mm_mask_add_round_ss(__W,__U,__A,__B,_MM_FROUND_TO_NEAREST_INT); 
+}
+__m128 test_mm_maskz_add_round_ss(__mmask8 __U, __m128 __A, __m128 __B) {
+  // CHECK-LABEL: @test_mm_maskz_add_round_ss
+  // CHECK: @llvm.x86.avx512.mask.add.ss.mask
+  return _mm_maskz_add_round_ss(__U,__A,__B,_MM_FROUND_TO_NEAREST_INT); 
+}
+__m128 test_mm_mask_add_ss(__m128 __W, __mmask8 __U, __m128 __A, __m128 __B) {
+  // CHECK-LABEL: @test_mm_mask_add_ss
+  // CHECK: @llvm.x86.avx512.mask.add.ss.mask
+  return _mm_mask_add_ss(__W,__U,__A,__B); 
+}
+__m128 test_mm_maskz_add_ss(__mmask8 __U, __m128 __A, __m128 __B) {
+  // CHECK-LABEL: @test_mm_maskz_add_ss
+  // CHECK: @llvm.x86.avx512.mask.add.ss.mask
+  return _mm_maskz_add_ss(__U,__A,__B); 
+}
+__m128d test_mm_add_round_sd(__m128d __A, __m128d __B) {
+  // CHECK-LABEL: @test_mm_add_round_sd
+  // CHECK: @llvm.x86.avx512.mask.add.sd.mask
+  return _mm_add_round_sd(__A,__B,_MM_FROUND_TO_NEAREST_INT); 
+}
+__m128d test_mm_mask_add_round_sd(__m128d __W, __mmask8 __U, __m128d __A, __m128d __B) {
+  // CHECK-LABEL: @test_mm_mask_add_round_sd
+  // CHECK: @llvm.x86.avx512.mask.add.sd.mask
+  return _mm_mask_add_round_sd(__W,__U,__A,__B,_MM_FROUND_TO_NEAREST_INT); 
+}
+__m128d test_mm_maskz_add_round_sd(__mmask8 __U, __m128d __A, __m128d __B) {
+  // CHECK-LABEL: @test_mm_maskz_add_round_sd
+  // CHECK: @llvm.x86.avx512.mask.add.sd.mask
+  return _mm_maskz_add_round_sd(__U,__A,__B,_MM_FROUND_TO_NEAREST_INT); 
+}
+__m128d test_mm_mask_add_sd(__m128d __W, __mmask8 __U, __m128d __A, __m128d __B) {
+  // CHECK-LABEL: @test_mm_mask_add_sd
+  // CHECK: @llvm.x86.avx512.mask.add.sd.mask
+  return _mm_mask_add_sd(__W,__U,__A,__B); 
+}
+__m128d test_mm_maskz_add_sd(__mmask8 __U, __m128d __A, __m128d __B) {
+  // CHECK-LABEL: @test_mm_maskz_add_sd
+  // CHECK: @llvm.x86.avx512.mask.add.sd.mask
+  return _mm_maskz_add_sd(__U,__A,__B); 
+}
 __m512d test_mm512_sub_round_pd(__m512d __A, __m512d __B) {
   // CHECK-LABEL: @test_mm512_sub_round_pd
   // CHECK: @llvm.x86.avx512.mask.sub.pd.512
   return _mm512_sub_round_pd(__A,__B,_MM_FROUND_TO_NEAREST_INT); 
 }
 __m512d test_mm512_mask_sub_round_pd(__m512d __W, __mmask8 __U, __m512d __A, __m512d __B) {
   // CHECK-LABEL: @test_mm512_mask_sub_round_pd
   // CHECK: @llvm.x86.avx512.mask.sub.pd.512
__m512 test_mm512_mask_sub_ps(__m512 __W, __mmask16 __U, __m512 __A, __m512 __B) {
   // CHECK: @llvm.x86.avx512.mask.sub.ps.512
   return _mm512_mask_sub_ps(__W,__U,__A,__B); 
 }
 __m512 test_mm512_maskz_sub_ps(__mmask16 __U, __m512 __A, __m512 __B) {
   // CHECK-LABEL: @test_mm512_maskz_sub_ps
   // CHECK: @llvm.x86.avx512.mask.sub.ps.512
   return _mm512_maskz_sub_ps(__U,__A,__B); 
 }
+__m128 test_mm_sub_round_ss(__m128 __A, __m128 __B) {
+  // CHECK-LABEL: @test_mm_sub_round_ss
+  // CHECK: @llvm.x86.avx512.mask.sub.ss.mask
+  return _mm_sub_round_ss(__A,__B,_MM_FROUND_TO_NEAREST_INT); 
+}
+__m128 test_mm_mask_sub_round_ss(__m128 __W, __mmask8 __U, __m128 __A, __m128 __B) {
+  // CHECK-LABEL: @test_mm_mask_sub_round_ss
+  // CHECK: @llvm.x86.avx512.mask.sub.ss.mask
+  return _mm_mask_sub_round_ss(__W,__U,__A,__B,_MM_FROUND_TO_NEAREST_INT); 
+}
+__m128 test_mm_maskz_sub_round_ss(__mmask8 __U, __m128 __A, __m128 __B) {
+  // CHECK-LABEL: @test_mm_maskz_sub_round_ss
+  // CHECK: @llvm.x86.avx512.mask.sub.ss.mask
+  return _mm_maskz_sub_round_ss(__U,__A,__B,_MM_FROUND_TO_NEAREST_INT); 
+}
+__m128 test_mm_mask_sub_ss(__m128 __W, __mmask8 __U, __m128 __A, __m128 __B) {
+  // CHECK-LABEL: @test_mm_mask_sub_ss
+  // CHECK: @llvm.x86.avx512.mask.sub.ss.mask
+  return _mm_mask_sub_ss(__W,__U,__A,__B); 
+}
+__m128 test_mm_maskz_sub_ss(__mmask8 __U, __m128 __A, __m128 __B) {
+  // CHECK-LABEL: @test_mm_maskz_sub_ss
+  // CHECK: @llvm.x86.avx512.mask.sub.ss.mask
+  return _mm_maskz_sub_ss(__U,__A,__B); 
+}
+__m128d test_mm_sub_round_sd(__m128d __A, __m128d __B) {
+  // CHECK-LABEL: @test_mm_sub_round_sd
+  // CHECK: @llvm.x86.avx512.mask.sub.sd.mask
+  return _mm_sub_round_sd(__A,__B,_MM_FROUND_TO_NEAREST_INT); 
+}
+__m128d test_mm_mask_sub_round_sd(__m128d __W, __mmask8 __U, __m128d __A, __m128d __B) {
+  // CHECK-LABEL: @test_mm_mask_sub_round_sd
+  // CHECK: @llvm.x86.avx512.mask.sub.sd.mask
+  return _mm_mask_sub_round_sd(__W,__U,__A,__B,_MM_FROUND_TO_NEAREST_INT); 
+}
+__m128d test_mm_maskz_sub_round_sd(__mmask8 __U, __m128d __A, __m128d __B) {
+  // CHECK-LABEL: @test_mm_maskz_sub_round_sd
+  // CHECK: @llvm.x86.avx512.mask.sub.sd.mask
+  return _mm_maskz_sub_round_sd(__U,__A,__B,_MM_FROUND_TO_NEAREST_INT); 
+}
+__m128d test_mm_mask_sub_sd(__m128d __W, __mmask8 __U, __m128d __A, __m128d __B) {
+  // CHECK-LABEL: @test_mm_mask_sub_sd
+  // CHECK: @llvm.x86.avx512.mask.sub.sd.mask
+  return _mm_mask_sub_sd(__W,__U,__A,__B); 
+}
+__m128d test_mm_maskz_sub_sd(__mmask8 __U, __m128d __A, __m128d __B) {
+  // CHECK-LABEL: @test_mm_maskz_sub_sd
+  // CHECK: @llvm.x86.avx512.mask.sub.sd.mask
+  return _mm_maskz_sub_sd(__U,__A,__B); 
+}
 __m512d test_mm512_mul_round_pd(__m512d __A, __m512d __B) {
   // CHECK-LABEL: @test_mm512_mul_round_pd
   // CHECK: @llvm.x86.avx512.mask.mul.pd.512
   return _mm512_mul_round_pd(__A,__B,_MM_FROUND_TO_NEAREST_INT); 
 }
 __m512d test_mm512_mask_mul_round_pd(__m512d __W, __mmask8 __U, __m512d __A, __m512d __B) {
   // CHECK-LABEL: @test_mm512_mask_mul_round_pd
   // CHECK: @llvm.x86.avx512.mask.mul.pd.512
__m512 test_mm512_mask_mul_ps(__m512 __W, __mmask16 __U, __m512 __A, __m512 __B) {
   // CHECK: @llvm.x86.avx512.mask.mul.ps.512
   return _mm512_mask_mul_ps(__W,__U,__A,__B); 
 }
 __m512 test_mm512_maskz_mul_ps(__mmask16 __U, __m512 __A, __m512 __B) {
   // CHECK-LABEL: @test_mm512_maskz_mul_ps
   // CHECK: @llvm.x86.avx512.mask.mul.ps.512
   return _mm512_maskz_mul_ps(__U,__A,__B); 
 }
+__m128 test_mm_mul_round_ss(__m128 __A, __m128 __B) {
+  // CHECK-LABEL: @test_mm_mul_round_ss
+  // CHECK: @llvm.x86.avx512.mask.mul.ss.mask
+  return _mm_mul_round_ss(__A,__B,_MM_FROUND_TO_NEAREST_INT); 
+}
+__m128 test_mm_mask_mul_round_ss(__m128 __W, __mmask8 __U, __m128 __A, __m128 __B) {
+  // CHECK-LABEL: @test_mm_mask_mul_round_ss
+  // CHECK: @llvm.x86.avx512.mask.mul.ss.mask
+  return _mm_mask_mul_round_ss(__W,__U,__A,__B,_MM_FROUND_TO_NEAREST_INT); 
+}
+__m128 test_mm_maskz_mul_round_ss(__mmask8 __U, __m128 __A, __m128 __B) {
+  // CHECK-LABEL: @test_mm_maskz_mul_round_ss
+  // CHECK: @llvm.x86.avx512.mask.mul.ss.mask
+  return _mm_maskz_mul_round_ss(__U,__A,__B,_MM_FROUND_TO_NEAREST_INT); 
+}
+__m128 test_mm_mask_mul_ss(__m128 __W, __mmask8 __U, __m128 __A, __m128 __B) {
+  // CHECK-LABEL: @test_mm_mask_mul_ss
+  // CHECK: @llvm.x86.avx512.mask.mul.ss.mask
+  return _mm_mask_mul_ss(__W,__U,__A,__B); 
+}
+__m128 test_mm_maskz_mul_ss(__mmask8 __U, __m128 __A, __m128 __B) {
+  // CHECK-LABEL: @test_mm_maskz_mul_ss
+  // CHECK: @llvm.x86.avx512.mask.mul.ss.mask
+  return _mm_maskz_mul_ss(__U,__A,__B); 
+}
+__m128d test_mm_mul_round_sd(__m128d __A, __m128d __B) {
+  // CHECK-LABEL: @test_mm_mul_round_sd
+  // CHECK: @llvm.x86.avx512.mask.mul.sd.mask
+  return _mm_mul_round_sd(__A,__B,_MM_FROUND_TO_NEAREST_INT); 
+}
+__m128d test_mm_mask_mul_round_sd(__m128d __W, __mmask8 __U, __m128d __A, __m128d __B) {
+  // CHECK-LABEL: @test_mm_mask_mul_round_sd
+  // CHECK: @llvm.x86.avx512.mask.mul.sd.mask
+  return _mm_mask_mul_round_sd(__W,__U,__A,__B,_MM_FROUND_TO_NEAREST_INT); 
+}
+__m128d test_mm_maskz_mul_round_sd(__mmask8 __U, __m128d __A, __m128d __B) {
+  // CHECK-LABEL: @test_mm_maskz_mul_round_sd
+  // CHECK: @llvm.x86.avx512.mask.mul.sd.mask
+  return _mm_maskz_mul_round_sd(__U,__A,__B,_MM_FROUND_TO_NEAREST_INT); 
+}
+__m128d test_mm_mask_mul_sd(__m128d __W, __mmask8 __U, __m128d __A, __m128d __B) {
+  // CHECK-LABEL: @test_mm_mask_mul_sd
+  // CHECK: @llvm.x86.avx512.mask.mul.sd.mask
+  return _mm_mask_mul_sd(__W,__U,__A,__B); 
+}
+__m128d test_mm_maskz_mul_sd(__mmask8 __U, __m128d __A, __m128d __B) {
+  // CHECK-LABEL: @test_mm_maskz_mul_sd
+  // CHECK: @llvm.x86.avx512.mask.mul.sd.mask
+  return _mm_maskz_mul_sd(__U,__A,__B); 
+}
 __m512d test_mm512_div_round_pd(__m512d __A, __m512d __B) {
   // CHECK-LABEL: @test_mm512_div_round_pd
   // CHECK: @llvm.x86.avx512.mask.div.pd.512
   return _mm512_div_round_pd(__A,__B,_MM_FROUND_TO_NEAREST_INT); 
 }
 __m512d test_mm512_mask_div_round_pd(__m512d __W, __mmask8 __U, __m512d __A, __m512d __B) {
   // CHECK-LABEL: @test_mm512_mask_div_round_pd
   // CHECK: @llvm.x86.avx512.mask.div.pd.512
__m512 test_mm512_mask_div_ps(__m512 __W, __mmask16 __U, __m512 __A, __m512 __B) {
   // CHECK: @llvm.x86.avx512.mask.div.ps.512
   return _mm512_mask_div_ps(__W,__U,__A,__B); 
 }
 __m512 test_mm512_maskz_div_ps(__mmask16 __U, __m512 __A, __m512 __B) {
   // CHECK-LABEL: @test_mm512_maskz_div_ps
   // CHECK: @llvm.x86.avx512.mask.div.ps.512
   return _mm512_maskz_div_ps(__U,__A,__B); 
 }
+__m128 test_mm_div_round_ss(__m128 __A, __m128 __B) {
+  // CHECK-LABEL: @test_mm_div_round_ss
+  // CHECK: @llvm.x86.avx512.mask.div.ss.mask
+  return _mm_div_round_ss(__A,__B,_MM_FROUND_TO_NEAREST_INT); 
+}
+__m128 test_mm_mask_div_round_ss(__m128 __W, __mmask8 __U, __m128 __A, __m128 __B) {
+  // CHECK-LABEL: @test_mm_mask_div_round_ss
+  // CHECK: @llvm.x86.avx512.mask.div.ss.mask
+  return _mm_mask_div_round_ss(__W,__U,__A,__B,_MM_FROUND_TO_NEAREST_INT); 
+}
+__m128 test_mm_maskz_div_round_ss(__mmask8 __U, __m128 __A, __m128 __B) {
+  // CHECK-LABEL: @test_mm_maskz_div_round_ss
+  // CHECK: @llvm.x86.avx512.mask.div.ss.mask
+  return _mm_maskz_div_round_ss(__U,__A,__B,_MM_FROUND_TO_NEAREST_INT); 
+}
+__m128 test_mm_mask_div_ss(__m128 __W, __mmask8 __U, __m128 __A, __m128 __B) {
+  // CHECK-LABEL: @test_mm_mask_div_ss
+  // CHECK: @llvm.x86.avx512.mask.div.ss.mask
+  return _mm_mask_div_ss(__W,__U,__A,__B); 
+}
+__m128 test_mm_maskz_div_ss(__mmask8 __U, __m128 __A, __m128 __B) {
+  // CHECK-LABEL: @test_mm_maskz_div_ss
+  // CHECK: @llvm.x86.avx512.mask.div.ss.mask
+  return _mm_maskz_div_ss(__U,__A,__B); 
+}
+__m128d test_mm_div_round_sd(__m128d __A, __m128d __B) {
+  // CHECK-LABEL: @test_mm_div_round_sd
+  // CHECK: @llvm.x86.avx512.mask.div.sd.mask
+  return _mm_div_round_sd(__A,__B,_MM_FROUND_TO_NEAREST_INT); 
+}
+__m128d test_mm_mask_div_round_sd(__m128d __W, __mmask8 __U, __m128d __A, __m128d __B) {
+  // CHECK-LABEL: @test_mm_mask_div_round_sd
+  // CHECK: @llvm.x86.avx512.mask.div.sd.mask
+  return _mm_mask_div_round_sd(__W,__U,__A,__B,_MM_FROUND_TO_NEAREST_INT); 
+}
+__m128d test_mm_maskz_div_round_sd(__mmask8 __U, __m128d __A, __m128d __B) {
+  // CHECK-LABEL: @test_mm_maskz_div_round_sd
+  // CHECK: @llvm.x86.avx512.mask.div.sd.mask
+  return _mm_maskz_div_round_sd(__U,__A,__B,_MM_FROUND_TO_NEAREST_INT); 
+}
+__m128d test_mm_mask_div_sd(__m128d __W, __mmask8 __U, __m128d __A, __m128d __B) {
+  // CHECK-LABEL: @test_mm_mask_div_sd
+  // CHECK: @llvm.x86.avx512.mask.div.sd.mask
+  return _mm_mask_div_sd(__W,__U,__A,__B); 
+}
+__m128d test_mm_maskz_div_sd(__mmask8 __U, __m128d __A, __m128d __B) {
+  // CHECK-LABEL: @test_mm_maskz_div_sd
+  // CHECK: @llvm.x86.avx512.mask.div.sd.mask
+  return _mm_maskz_div_sd(__U,__A,__B); 
+}
+__m128 test_mm_max_round_ss(__m128 __A, __m128 __B) {
+  // CHECK-LABEL: @test_mm_max_round_ss
+  // CHECK: @llvm.x86.avx512.mask.max.ss.mask
+  return _mm_max_round_ss(__A,__B,0x08); 
+}
+__m128 test_mm_mask_max_round_ss(__m128 __W, __mmask8 __U, __m128 __A, __m128 __B) {
+  // CHECK-LABEL: @test_mm_mask_max_round_ss
+  // CHECK: @llvm.x86.avx512.mask.max.ss.mask
+  return _mm_mask_max_round_ss(__W,__U,__A,__B,0x08); 
+}
+__m128 test_mm_maskz_max_round_ss(__mmask8 __U, __m128 __A, __m128 __B) {
+  // CHECK-LABEL: @test_mm_maskz_max_round_ss
+  // CHECK: @llvm.x86.avx512.mask.max.ss.mask
+  return _mm_maskz_max_round_ss(__U,__A,__B,0x08); 
+}
+__m128 test_mm_mask_max_ss(__m128 __W, __mmask8 __U, __m128 __A, __m128 __B) {
+  // CHECK-LABEL: @test_mm_mask_max_ss
+  // CHECK: @llvm.x86.avx512.mask.max.ss.mask
+  return _mm_mask_max_ss(__W,__U,__A,__B); 
+}
+__m128 test_mm_maskz_max_ss(__mmask8 __U, __m128 __A, __m128 __B) {
+  // CHECK-LABEL: @test_mm_maskz_max_ss
+  // CHECK: @llvm.x86.avx512.mask.max.ss.mask
+  return _mm_maskz_max_ss(__U,__A,__B); 
+}
+__m128d test_mm_max_round_sd(__m128d __A, __m128d __B) {
+  // CHECK-LABEL: @test_mm_max_round_sd
+  // CHECK: @llvm.x86.avx512.mask.max.sd.mask
+  return _mm_max_round_sd(__A,__B,0x08); 
+}
+__m128d test_mm_mask_max_round_sd(__m128d __W, __mmask8 __U, __m128d __A, __m128d __B) {
+  // CHECK-LABEL: @test_mm_mask_max_round_sd
+  // CHECK: @llvm.x86.avx512.mask.max.sd.mask
+  return _mm_mask_max_round_sd(__W,__U,__A,__B,0x08); 
+}
+__m128d test_mm_maskz_max_round_sd(__mmask8 __U, __m128d __A, __m128d __B) {
+  // CHECK-LABEL: @test_mm_maskz_max_round_sd
+  // CHECK: @llvm.x86.avx512.mask.max.sd.mask
+  return _mm_maskz_max_round_sd(__U,__A,__B,0x08); 
+}
+__m128d test_mm_mask_max_sd(__m128d __W, __mmask8 __U, __m128d __A, __m128d __B) {
+  // CHECK-LABEL: @test_mm_mask_max_sd
+  // CHECK: @llvm.x86.avx512.mask.max.sd.mask
+  return _mm_mask_max_sd(__W,__U,__A,__B); 
+}
+__m128d test_mm_maskz_max_sd(__mmask8 __U, __m128d __A, __m128d __B) {
+  // CHECK-LABEL: @test_mm_maskz_max_sd
+  // CHECK: @llvm.x86.avx512.mask.max.sd.mask
+  return _mm_maskz_max_sd(__U,__A,__B); 
+}
+__m128 test_mm_min_round_ss(__m128 __A, __m128 __B) {
+  // CHECK-LABEL: @test_mm_min_round_ss
+  // CHECK: @llvm.x86.avx512.mask.min.ss.mask
+  return _mm_min_round_ss(__A,__B,0x08); 
+}
+__m128 test_mm_mask_min_round_ss(__m128 __W, __mmask8 __U, __m128 __A, __m128 __B) {
+  // CHECK-LABEL: @test_mm_mask_min_round_ss
+  // CHECK: @llvm.x86.avx512.mask.min.ss.mask
+  return _mm_mask_min_round_ss(__W,__U,__A,__B,0x08); 
+}
+__m128 test_mm_maskz_min_round_ss(__mmask8 __U, __m128 __A, __m128 __B) {
+  // CHECK-LABEL: @test_mm_maskz_min_round_ss
+  // CHECK: @llvm.x86.avx512.mask.min.ss.mask
+  return _mm_maskz_min_round_ss(__U,__A,__B,0x08); 
+}
+__m128 test_mm_mask_min_ss(__m128 __W, __mmask8 __U, __m128 __A, __m128 __B) {
+  // CHECK-LABEL: @test_mm_mask_min_ss
+  // CHECK: @llvm.x86.avx512.mask.min.ss.mask
+  return _mm_mask_min_ss(__W,__U,__A,__B); 
+}
+__m128 test_mm_maskz_min_ss(__mmask8 __U, __m128 __A, __m128 __B) {
+  // CHECK-LABEL: @test_mm_maskz_min_ss
+  // CHECK: @llvm.x86.avx512.mask.min.ss.mask
+  return _mm_maskz_min_ss(__U,__A,__B); 
+}
+__m128d test_mm_min_round_sd(__m128d __A, __m128d __B) {
+  // CHECK-LABEL: @test_mm_min_round_sd
+  // CHECK: @llvm.x86.avx512.mask.min.sd.mask
+  return _mm_min_round_sd(__A,__B,0x08); 
+}
+__m128d test_mm_mask_min_round_sd(__m128d __W, __mmask8 __U, __m128d __A, __m128d __B) {
+  // CHECK-LABEL: @test_mm_mask_min_round_sd
+  // CHECK: @llvm.x86.avx512.mask.min.sd.mask
+  return _mm_mask_min_round_sd(__W,__U,__A,__B,0x08); 
+}
+__m128d test_mm_maskz_min_round_sd(__mmask8 __U, __m128d __A, __m128d __B) {
+  // CHECK-LABEL: @test_mm_maskz_min_round_sd
+  // CHECK: @llvm.x86.avx512.mask.min.sd.mask
+  return _mm_maskz_min_round_sd(__U,__A,__B,0x08); 
+}
+__m128d test_mm_mask_min_sd(__m128d __W, __mmask8 __U, __m128d __A, __m128d __B) {
+  // CHECK-LABEL: @test_mm_mask_min_sd
+  // CHECK: @llvm.x86.avx512.mask.min.sd.mask
+  return _mm_mask_min_sd(__W,__U,__A,__B); 
+}
+__m128d test_mm_maskz_min_sd(__mmask8 __U, __m128d __A, __m128d __B) {
+  // CHECK-LABEL: @test_mm_maskz_min_sd
+  // CHECK: @llvm.x86.avx512.mask.min.sd.mask
+  return _mm_maskz_min_sd(__U,__A,__B); 
+}