Index: include/clang/Basic/BuiltinsX86.def
===================================================================
--- include/clang/Basic/BuiltinsX86.def
+++ include/clang/Basic/BuiltinsX86.def
@@ -1794,6 +1794,25 @@
 TARGET_BUILTIN(__builtin_ia32_loadupd256_mask, "V4dV4d*V4dUc","","avx512vl")
 TARGET_BUILTIN(__builtin_ia32_loadups128_mask, "V4fV4f*V4fUc","","avx512vl")
 TARGET_BUILTIN(__builtin_ia32_loadups256_mask, "V8fV8f*V8fUc","","avx512vl")
-
+TARGET_BUILTIN(__builtin_ia32_shuf_f32x4_mask, "V16fV16fV16fIiV16fUs","","avx512f")
+TARGET_BUILTIN(__builtin_ia32_shuf_f64x2_mask, "V8dV8dV8dIiV8dUc","","avx512f")
+TARGET_BUILTIN(__builtin_ia32_shuf_i32x4_mask, "V16iV16iV16iIiV16iUs","","avx512f")
+TARGET_BUILTIN(__builtin_ia32_shuf_i64x2_mask, "V8LLiV8LLiV8LLiIiV8LLiUc","","avx512f")
+TARGET_BUILTIN(__builtin_ia32_shufpd512_mask, "V8dV8dV8dIiV8dUc","","avx512f")
+TARGET_BUILTIN(__builtin_ia32_shufps512_mask, "V16fV16fV16fIiV16fUs","","avx512f")
+TARGET_BUILTIN(__builtin_ia32_shuf_f32x4_256_mask, "V8fV8fV8fIiV8fUc","","avx512vl")
+TARGET_BUILTIN(__builtin_ia32_shuf_f64x2_256_mask, "V4dV4dV4dIiV4dUc","","avx512vl")
+TARGET_BUILTIN(__builtin_ia32_shuf_i32x4_256_mask, "V8iV8iV8iIiV8iUc","","avx512vl")
+TARGET_BUILTIN(__builtin_ia32_shuf_i64x2_256_mask, "V4LLiV4LLiV4LLiIiV4LLiUc","","avx512vl")
+TARGET_BUILTIN(__builtin_ia32_shufpd128_mask, "V2dV2dV2dIiV2dUc","","avx512vl")
+TARGET_BUILTIN(__builtin_ia32_shufpd256_mask, "V4dV4dV4dIiV4dUc","","avx512vl")
+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")
 #undef BUILTIN
 #undef TARGET_BUILTIN
Index: lib/Headers/avx512fintrin.h
===================================================================
--- lib/Headers/avx512fintrin.h
+++ lib/Headers/avx512fintrin.h
@@ -3974,6 +3974,228 @@
 __builtin_ia32_vcomiss ((__v4sf) (__A), (__v4sf) (__B), ( __P), ( __R));\
 })
 
+#define _mm512_shuffle_f32x4( __A, __B, __imm) __extension__ ({ \
+__builtin_ia32_shuf_f32x4_mask ((__v16sf)( __A),\
+              (__v16sf)( __B),( __imm),\
+              (__v16sf)\
+              _mm512_undefined_ps (),\
+              (__mmask16) -1);\
+})
+
+#define _mm512_mask_shuffle_f32x4( __W, __U, __A, __B, __imm) __extension__ ({ \
+__builtin_ia32_shuf_f32x4_mask ((__v16sf)( __A),\
+              (__v16sf)( __B),( __imm),\
+              (__v16sf)( __W),\
+              (__mmask16)( __U));\
+})
+
+#define _mm512_maskz_shuffle_f32x4( __U, __A, __B, __imm) __extension__ ({ \
+__builtin_ia32_shuf_f32x4_mask ((__v16sf)( __A),\
+              (__v16sf)( __B),( __imm),\
+              (__v16sf)\
+              _mm512_setzero_ps (),\
+              (__mmask16)( __U));\
+})
+
+#define _mm512_shuffle_f64x2( __A, __B, __imm) __extension__ ({ \
+__builtin_ia32_shuf_f64x2_mask ((__v8df)( __A),\
+               (__v8df)( __B),( __imm),\
+               (__v8df)\
+               _mm512_undefined_pd (),\
+               (__mmask8) -1);\
+})
+
+#define _mm512_mask_shuffle_f64x2( __W, __U, __A, __B, __imm) __extension__ ({ \
+__builtin_ia32_shuf_f64x2_mask ((__v8df)( __A),\
+               (__v8df)( __B),( __imm),\
+               (__v8df)( __W),\
+               (__mmask8)( __U));\
+})
+
+#define _mm512_maskz_shuffle_f64x2( __U, __A, __B, __imm) __extension__ ({ \
+__builtin_ia32_shuf_f64x2_mask ((__v8df)( __A),\
+               (__v8df)( __B),( __imm),\
+               (__v8df)\
+               _mm512_setzero_pd (),\
+               (__mmask8)( __U));\
+})
+
+#define _mm512_shuffle_i32x4( __A, __B, __imm) __extension__ ({ \
+__builtin_ia32_shuf_i32x4_mask ((__v16si)( __A),\
+               (__v16si)( __B),\
+              ( __imm),\
+               (__v16si)\
+               _mm512_setzero_si512 (),\
+               (__mmask16) -1);\
+})
+
+#define _mm512_mask_shuffle_i32x4( __W, __U, __A, __B, __imm) __extension__ ({ \
+__builtin_ia32_shuf_i32x4_mask ((__v16si)( __A),\
+               (__v16si)( __B),\
+              ( __imm),\
+               (__v16si)( __W),\
+               (__mmask16)( __U));\
+})
+
+#define _mm512_maskz_shuffle_i32x4( __U, __A, __B, __imm) __extension__ ({ \
+__builtin_ia32_shuf_i32x4_mask ((__v16si)( __A),\
+               (__v16si)( __B),\
+              ( __imm),\
+               (__v16si)\
+               _mm512_setzero_si512 (),\
+               (__mmask16)( __U));\
+})
+
+#define _mm512_shuffle_i64x2( __A, __B, __imm) __extension__ ({ \
+__builtin_ia32_shuf_i64x2_mask ((__v8di)( __A),\
+               (__v8di)( __B),( __imm),\
+               (__v8di)\
+               _mm512_setzero_si512 (),\
+               (__mmask8) -1);\
+})
+
+#define _mm512_mask_shuffle_i64x2( __W, __U, __A, __B, __imm) __extension__ ({ \
+__builtin_ia32_shuf_i64x2_mask ((__v8di)( __A),\
+               (__v8di)( __B),( __imm),\
+               (__v8di)( __W),\
+               (__mmask8)( __U));\
+})
+
+#define _mm512_maskz_shuffle_i64x2( __U, __A, __B, __imm) __extension__ ({ \
+__builtin_ia32_shuf_i64x2_mask ((__v8di)( __A),\
+               (__v8di)( __B),( __imm),\
+               (__v8di)\
+               _mm512_setzero_si512 (),\
+               (__mmask8)( __U));\
+})
+
+#define _mm512_shuffle_pd( __M, __V, __imm) __extension__ ({ \
+__builtin_ia32_shufpd512_mask ((__v8df)( __M),\
+              (__v8df)( __V),( __imm),\
+              (__v8df)\
+              _mm512_undefined_pd (),\
+              (__mmask8) -1);\
+})
+
+#define _mm512_mask_shuffle_pd( __W, __U, __M, __V, __imm) __extension__ ({ \
+__builtin_ia32_shufpd512_mask ((__v8df)( __M),\
+              (__v8df)( __V),( __imm),\
+              (__v8df)( __W),\
+              (__mmask8)( __U));\
+})
+
+#define _mm512_maskz_shuffle_pd( __U, __M, __V, __imm) __extension__ ({ \
+__builtin_ia32_shufpd512_mask ((__v8df)( __M),\
+              (__v8df)( __V),( __imm),\
+              (__v8df)\
+              _mm512_setzero_pd (),\
+              (__mmask8)( __U));\
+})
+
+#define _mm512_shuffle_ps( __M, __V, __imm) __extension__ ({ \
+__builtin_ia32_shufps512_mask ((__v16sf)( __M),\
+             (__v16sf)( __V),( __imm),\
+             (__v16sf)\
+             _mm512_undefined_ps (),\
+             (__mmask16) -1);\
+})
+
+#define _mm512_mask_shuffle_ps( __W, __U, __M, __V, __imm) __extension__ ({ \
+__builtin_ia32_shufps512_mask ((__v16sf)( __M),\
+             (__v16sf)( __V),( __imm),\
+             (__v16sf)( __W),\
+             (__mmask16)( __U));\
+})
+
+#define _mm512_maskz_shuffle_ps( __U, __M, __V, __imm) __extension__ ({ \
+__builtin_ia32_shufps512_mask ((__v16sf)( __M),\
+             (__v16sf)( __V),( __imm),\
+             (__v16sf)\
+             _mm512_setzero_ps (),\
+             (__mmask16)( __U));\
+})
+
+#define _mm_sqrt_round_sd( __A, __B, __R) __extension__ ({ \
+__builtin_ia32_sqrtsd_round_mask ((__v2df)( __B),\
+            (__v2df)( __A),(__v2df) _mm_setzero_pd(),\
+            (__mmask8) -1,\
+           ( __R));\
+})
+
+static __inline__ __m128d __DEFAULT_FN_ATTRS
+_mm_mask_sqrt_sd (__m128d __W, __mmask8 __U, __m128d __A, __m128d __B) 
+{
+ return (__m128d) __builtin_ia32_sqrtsd_round_mask ( (__v2df) __B,
+                 (__v2df) __A,
+                (__v2df) __W,
+                (__mmask8) __U,
+                _MM_FROUND_CUR_DIRECTION);
+}
+
+#define _mm_mask_sqrt_round_sd( __W, __U, __A, __B, __R) __extension__ ({ \
+__builtin_ia32_sqrtsd_round_mask ((__v2df)( __B),\
+            (__v2df)( __A),(__v2df) __W,\
+            (__mmask8) __U,\
+            ( __R));\
+})
+
+static __inline__ __m128d __DEFAULT_FN_ATTRS
+_mm_maskz_sqrt_sd (__mmask8 __U, __m128d __A, __m128d __B) 
+{
+ return (__m128d) __builtin_ia32_sqrtsd_round_mask ( (__v2df) __B,
+                 (__v2df) __A,
+                (__v2df) _mm_setzero_pd (),
+                (__mmask8) __U,
+                _MM_FROUND_CUR_DIRECTION);
+}
+
+#define _mm_maskz_sqrt_round_sd( __U, __A, __B, __R) __extension__ ({ \
+__builtin_ia32_sqrtsd_round_mask ((__v2df)( __B),\
+            (__v2df)( __A),(__v2df) _mm_setzero_pd(),\
+            (__mmask8) __U,\
+            ( __R));\
+})
+
+#define _mm_sqrt_round_ss( __A, __B, __R) __extension__ ({ \
+__builtin_ia32_sqrtss_round_mask ((__v4sf)( __B),\
+                 (__v4sf)( __A),(__v4sf) _mm_setzero_ps(),\
+                 (__mmask8) -1,\
+                 ( __R));\
+})
+
+static __inline__ __m128 __DEFAULT_FN_ATTRS
+_mm_mask_sqrt_ss (__m128 __W, __mmask8 __U, __m128 __A, __m128 __B) 
+{
+ return (__m128) __builtin_ia32_sqrtss_round_mask ( (__v4sf) __B,
+                 (__v4sf) __A,
+                (__v4sf) __W,
+                (__mmask8) __U,
+                _MM_FROUND_CUR_DIRECTION);
+}
+
+#define _mm_mask_sqrt_round_ss( __W, __U, __A, __B, __R) __extension__ ({ \
+__builtin_ia32_sqrtss_round_mask ((__v4sf)( __B),\
+                 (__v4sf)( __A),(__v4sf) __W,\
+                 (__mmask8) __U,\
+                 ( __R));\
+})
+
+static __inline__ __m128 __DEFAULT_FN_ATTRS
+_mm_maskz_sqrt_ss (__mmask8 __U, __m128 __A, __m128 __B) 
+{
+ return (__m128) __builtin_ia32_sqrtss_round_mask ( (__v4sf) __A,
+                 (__v4sf) __B,
+                (__v4sf) _mm_setzero_ps (),
+                (__mmask8) __U,
+                _MM_FROUND_CUR_DIRECTION);
+}
+
+#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,\
+                 _MM_FROUND_CUR_DIRECTION);\
+})
 
 #undef __DEFAULT_FN_ATTRS
 
Index: lib/Headers/avx512vlintrin.h
===================================================================
--- lib/Headers/avx512vlintrin.h
+++ lib/Headers/avx512vlintrin.h
@@ -6360,6 +6360,275 @@
               _mm256_setzero_ps (),
               (__mmask8) __U);
 }
+
+#define _mm256_shuffle_f32x4( __A, __B, __imm) __extension__ ({ \
+__builtin_ia32_shuf_f32x4_256_mask ((__v8sf)( __A),\
+                  (__v8sf)( __B),\
+                 ( __imm),\
+                  (__v8sf)\
+                  _mm256_setzero_ps (),\
+                  (__mmask8) -1);\
+})
+
+#define _mm256_mask_shuffle_f32x4( __W, __U, __A, __B, __imm) __extension__ ({ \
+__builtin_ia32_shuf_f32x4_256_mask ((__v8sf)( __A),\
+                  (__v8sf)( __B),\
+                 ( __imm),\
+                  (__v8sf)( __W),\
+                  (__mmask8)( __U));\
+})
+
+#define _mm256_maskz_shuffle_f32x4( __U, __A, __B, __imm) __extension__ ({ \
+__builtin_ia32_shuf_f32x4_256_mask ((__v8sf)( __A),\
+                  (__v8sf)( __B),\
+                 ( __imm),\
+                  (__v8sf)\
+                  _mm256_setzero_ps (),\
+                  (__mmask8)( __U));\
+})
+
+#define _mm256_shuffle_f64x2( __A, __B, __imm) __extension__ ({ \
+__builtin_ia32_shuf_f64x2_256_mask ((__v4df)( __A),\
+                   (__v4df)( __B),\
+                  ( __imm),\
+                   (__v4df)\
+                   _mm256_setzero_pd (),\
+                   (__mmask8) -1);\
+})
+
+#define _mm256_mask_shuffle_f64x2( __W, __U, __A, __B, __imm) __extension__ ({ \
+__builtin_ia32_shuf_f64x2_256_mask ((__v4df)( __A),\
+                   (__v4df)( __B),\
+                  ( __imm),\
+                   (__v4df)( __W),\
+                   (__mmask8)( __U));\
+})
+
+#define _mm256_maskz_shuffle_f64x2( __U, __A, __B, __imm) __extension__ ({ \
+__builtin_ia32_shuf_f64x2_256_mask ((__v4df)( __A),\
+                   (__v4df)( __B),\
+                  ( __imm),\
+                   (__v4df)\
+                   _mm256_setzero_pd (),\
+                   (__mmask8)( __U));\
+})
+
+#define _mm256_shuffle_i32x4( __A, __B, __imm) __extension__ ({ \
+__builtin_ia32_shuf_i32x4_256_mask ((__v8si)( __A),\
+                   (__v8si)( __B),\
+                  ( __imm),\
+                   (__v8si)\
+                   _mm256_setzero_si256 (),\
+                   (__mmask8) -1);\
+})
+
+#define _mm256_mask_shuffle_i32x4( __W, __U, __A, __B, __imm) __extension__ ({ \
+__builtin_ia32_shuf_i32x4_256_mask ((__v8si)( __A),\
+                   (__v8si)( __B),\
+                  ( __imm),\
+                   (__v8si)( __W),\
+                   (__mmask8)( __U));\
+})
+
+#define _mm256_maskz_shuffle_i32x4( __U, __A, __B, __imm) __extension__ ({ \
+__builtin_ia32_shuf_i32x4_256_mask ((__v8si)( __A),\
+                   (__v8si)( __B),\
+                  ( __imm),\
+                   (__v8si)\
+                   _mm256_setzero_si256 (),\
+                   (__mmask8)( __U));\
+})
+
+#define _mm256_shuffle_i64x2( __A, __B, __imm) __extension__ ({ \
+__builtin_ia32_shuf_i64x2_256_mask ((__v4di)( __A),\
+                   (__v4di)( __B),\
+                  ( __imm),\
+                   (__v4di)\
+                   _mm256_setzero_si256 (),\
+                   (__mmask8) -1);\
+})
+
+#define _mm256_mask_shuffle_i64x2( __W, __U, __A, __B, __imm) __extension__ ({ \
+__builtin_ia32_shuf_i64x2_256_mask ((__v4di)( __A),\
+                   (__v4di)( __B),\
+                  ( __imm),\
+                   (__v4di)( __W),\
+                   (__mmask8)( __U));\
+})
+
+#define _mm256_maskz_shuffle_i64x2( __U, __A, __B, __imm) __extension__ ({ \
+__builtin_ia32_shuf_i64x2_256_mask ((__v4di)( __A),\
+                   (__v4di)( __B),\
+                  ( __imm),\
+                   (__v4di)\
+                   _mm256_setzero_si256 (),\
+                   (__mmask8)( __U));\
+})
+
+#define _mm_mask_shuffle_pd( __W, __U, __A, __B, __imm) __extension__ ({ \
+__builtin_ia32_shufpd128_mask ((__v2df)( __A),\
+              (__v2df)( __B),( __imm),\
+              (__v2df)( __W),\
+              (__mmask8)( __U));\
+})
+
+#define _mm_maskz_shuffle_pd( __U, __A, __B, __imm) __extension__ ({ \
+__builtin_ia32_shufpd128_mask ((__v2df)( __A),\
+              (__v2df)( __B),( __imm),\
+              (__v2df)\
+              _mm_setzero_pd (),\
+              (__mmask8)( __U));\
+})
+
+#define _mm256_mask_shuffle_pd( __W, __U, __A, __B, __imm) __extension__ ({ \
+__builtin_ia32_shufpd256_mask ((__v4df)( __A),\
+              (__v4df)( __B),( __imm),\
+              (__v4df)( __W),\
+              (__mmask8)( __U));\
+})
+
+#define _mm256_maskz_shuffle_pd( __U, __A, __B, __imm) __extension__ ({ \
+__builtin_ia32_shufpd256_mask ((__v4df)( __A),\
+              (__v4df)( __B),( __imm),\
+              (__v4df)\
+              _mm256_setzero_pd (),\
+              (__mmask8)( __U));\
+})
+
+#define _mm_mask_shuffle_ps( __W, __U, __A, __B, __imm) __extension__ ({ \
+__builtin_ia32_shufps128_mask ((__v4sf)( __A),\
+             (__v4sf)( __B),( __imm),\
+             (__v4sf)( __W),\
+             (__mmask8)( __U));\
+})
+
+#define _mm_maskz_shuffle_ps( __U, __A, __B, __imm) __extension__ ({ \
+__builtin_ia32_shufps128_mask ((__v4sf)( __A),\
+             (__v4sf)( __B),( __imm),\
+             (__v4sf)\
+             _mm_setzero_ps (),\
+             (__mmask8)( __U));\
+})
+
+#define _mm256_mask_shuffle_ps( __W, __U, __A, __B, __imm) __extension__ ({ \
+__builtin_ia32_shufps256_mask ((__v8sf)( __A),\
+             (__v8sf)( __B),( __imm),\
+             (__v8sf)( __W),\
+             (__mmask8)( __U));\
+})
+
+#define _mm256_maskz_shuffle_ps( __U, __A, __B, __imm) __extension__ ({ \
+__builtin_ia32_shufps256_mask ((__v8sf)( __A),\
+             (__v8sf)( __B),( __imm),\
+             (__v8sf)\
+             _mm256_setzero_ps (),\
+             (__mmask8)( __U));\
+})
+
+static __inline__ __m128d __DEFAULT_FN_ATTRS
+_mm_rsqrt14_pd (__m128d __A)
+{
+  return (__m128d) __builtin_ia32_rsqrt14pd128_mask ((__v2df) __A,
+                 (__v2df)
+                 _mm_setzero_pd (),
+                 (__mmask8) -1);
+}
+
+static __inline__ __m128d __DEFAULT_FN_ATTRS
+_mm_mask_rsqrt14_pd (__m128d __W, __mmask8 __U, __m128d __A)
+{
+  return (__m128d) __builtin_ia32_rsqrt14pd128_mask ((__v2df) __A,
+                 (__v2df) __W,
+                 (__mmask8) __U);
+}
+
+static __inline__ __m128d __DEFAULT_FN_ATTRS
+_mm_maskz_rsqrt14_pd (__mmask8 __U, __m128d __A)
+{
+  return (__m128d) __builtin_ia32_rsqrt14pd128_mask ((__v2df) __A,
+                 (__v2df)
+                 _mm_setzero_pd (),
+                 (__mmask8) __U);
+}
+
+static __inline__ __m256d __DEFAULT_FN_ATTRS
+_mm256_rsqrt14_pd (__m256d __A)
+{
+  return (__m256d) __builtin_ia32_rsqrt14pd256_mask ((__v4df) __A,
+                 (__v4df)
+                 _mm256_setzero_pd (),
+                 (__mmask8) -1);
+}
+
+static __inline__ __m256d __DEFAULT_FN_ATTRS
+_mm256_mask_rsqrt14_pd (__m256d __W, __mmask8 __U, __m256d __A)
+{
+  return (__m256d) __builtin_ia32_rsqrt14pd256_mask ((__v4df) __A,
+                 (__v4df) __W,
+                 (__mmask8) __U);
+}
+
+static __inline__ __m256d __DEFAULT_FN_ATTRS
+_mm256_maskz_rsqrt14_pd (__mmask8 __U, __m256d __A)
+{
+  return (__m256d) __builtin_ia32_rsqrt14pd256_mask ((__v4df) __A,
+                 (__v4df)
+                 _mm256_setzero_pd (),
+                 (__mmask8) __U);
+}
+
+static __inline__ __m128 __DEFAULT_FN_ATTRS
+_mm_rsqrt14_ps (__m128 __A)
+{
+  return (__m128) __builtin_ia32_rsqrt14ps128_mask ((__v4sf) __A,
+                (__v4sf)
+                _mm_setzero_ps (),
+                (__mmask8) -1);
+}
+
+static __inline__ __m128 __DEFAULT_FN_ATTRS
+_mm_mask_rsqrt14_ps (__m128 __W, __mmask8 __U, __m128 __A)
+{
+  return (__m128) __builtin_ia32_rsqrt14ps128_mask ((__v4sf) __A,
+                (__v4sf) __W,
+                (__mmask8) __U);
+}
+
+static __inline__ __m128 __DEFAULT_FN_ATTRS
+_mm_maskz_rsqrt14_ps (__mmask8 __U, __m128 __A)
+{
+  return (__m128) __builtin_ia32_rsqrt14ps128_mask ((__v4sf) __A,
+                (__v4sf)
+                _mm_setzero_ps (),
+                (__mmask8) __U);
+}
+
+static __inline__ __m256 __DEFAULT_FN_ATTRS
+_mm256_rsqrt14_ps (__m256 __A)
+{
+  return (__m256) __builtin_ia32_rsqrt14ps256_mask ((__v8sf) __A,
+                (__v8sf)
+                _mm256_setzero_ps (),
+                (__mmask8) -1);
+}
+
+static __inline__ __m256 __DEFAULT_FN_ATTRS
+_mm256_mask_rsqrt14_ps (__m256 __W, __mmask8 __U, __m256 __A)
+{
+  return (__m256) __builtin_ia32_rsqrt14ps256_mask ((__v8sf) __A,
+                (__v8sf) __W,
+                (__mmask8) __U);
+}
+
+static __inline__ __m256 __DEFAULT_FN_ATTRS
+_mm256_maskz_rsqrt14_ps (__mmask8 __U, __m256 __A)
+{
+  return (__m256) __builtin_ia32_rsqrt14ps256_mask ((__v8sf) __A,
+                (__v8sf)
+                _mm256_setzero_ps (),
+                (__mmask8) __U);
+}
+
 #undef __DEFAULT_FN_ATTRS
 #undef __DEFAULT_FN_ATTRS_BOTH
 
Index: test/CodeGen/avx512f-builtins.c
===================================================================
--- test/CodeGen/avx512f-builtins.c
+++ test/CodeGen/avx512f-builtins.c
@@ -2576,5 +2576,164 @@
   return _mm512_kmov(__A); 
 }
 
+__m512 test_mm512_shuffle_f32x4(__m512 __A, __m512 __B) {
+  // CHECK-LABEL: @test_mm512_shuffle_f32x4
+  // CHECK: @llvm.x86.avx512.mask.shuf
+  return _mm512_shuffle_f32x4(__A, __B, 4); 
+}
+
+__m512 test_mm512_mask_shuffle_f32x4(__m512 __W, __mmask16 __U, __m512 __A, __m512 __B) {
+  // CHECK-LABEL: @test_mm512_mask_shuffle_f32x4
+  // CHECK: @llvm.x86.avx512.mask.shuf
+  return _mm512_mask_shuffle_f32x4(__W, __U, __A, __B, 4); 
+}
+
+__m512 test_mm512_maskz_shuffle_f32x4(__mmask16 __U, __m512 __A, __m512 __B) {
+  // CHECK-LABEL: @test_mm512_maskz_shuffle_f32x4
+  // CHECK: @llvm.x86.avx512.mask.shuf
+  return _mm512_maskz_shuffle_f32x4(__U, __A, __B, 4); 
+}
+
+__m512d test_mm512_shuffle_f64x2(__m512d __A, __m512d __B) {
+  // CHECK-LABEL: @test_mm512_shuffle_f64x2
+  // CHECK: @llvm.x86.avx512.mask.shuf
+  return _mm512_shuffle_f64x2(__A, __B, 4); 
+}
+
+__m512d test_mm512_mask_shuffle_f64x2(__m512d __W, __mmask8 __U, __m512d __A, __m512d __B) {
+  // CHECK-LABEL: @test_mm512_mask_shuffle_f64x2
+  // CHECK: @llvm.x86.avx512.mask.shuf
+  return _mm512_mask_shuffle_f64x2(__W, __U, __A, __B, 4); 
+}
+
+__m512d test_mm512_maskz_shuffle_f64x2(__mmask8 __U, __m512d __A, __m512d __B) {
+  // CHECK-LABEL: @test_mm512_maskz_shuffle_f64x2
+  // CHECK: @llvm.x86.avx512.mask.shuf
+  return _mm512_maskz_shuffle_f64x2(__U, __A, __B, 4); 
+}
+
+__m512i test_mm512_shuffle_i32x4(__m512i __A, __m512i __B) {
+  // CHECK-LABEL: @test_mm512_shuffle_i32x4
+  // CHECK: @llvm.x86.avx512.mask.shuf
+  return _mm512_shuffle_i32x4(__A, __B, 4); 
+}
+
+__m512i test_mm512_mask_shuffle_i32x4(__m512i __W, __mmask16 __U, __m512i __A, __m512i __B) {
+  // CHECK-LABEL: @test_mm512_mask_shuffle_i32x4
+  // CHECK: @llvm.x86.avx512.mask.shuf
+  return _mm512_mask_shuffle_i32x4(__W, __U, __A, __B, 4); 
+}
+
+__m512i test_mm512_maskz_shuffle_i32x4(__mmask16 __U, __m512i __A, __m512i __B) {
+  // CHECK-LABEL: @test_mm512_maskz_shuffle_i32x4
+  // CHECK: @llvm.x86.avx512.mask.shuf
+  return _mm512_maskz_shuffle_i32x4(__U, __A, __B, 4); 
+}
+
+__m512i test_mm512_shuffle_i64x2(__m512i __A, __m512i __B) {
+  // CHECK-LABEL: @test_mm512_shuffle_i64x2
+  // CHECK: @llvm.x86.avx512.mask.shuf
+  return _mm512_shuffle_i64x2(__A, __B, 4); 
+}
+
+__m512i test_mm512_mask_shuffle_i64x2(__m512i __W, __mmask8 __U, __m512i __A, __m512i __B) {
+  // CHECK-LABEL: @test_mm512_mask_shuffle_i64x2
+  // CHECK: @llvm.x86.avx512.mask.shuf
+  return _mm512_mask_shuffle_i64x2(__W, __U, __A, __B, 4); 
+}
+
+__m512i test_mm512_maskz_shuffle_i64x2(__mmask8 __U, __m512i __A, __m512i __B) {
+  // CHECK-LABEL: @test_mm512_maskz_shuffle_i64x2
+  // CHECK: @llvm.x86.avx512.mask.shuf
+  return _mm512_maskz_shuffle_i64x2(__U, __A, __B, 4); 
+}
+
+__m512d test_mm512_shuffle_pd(__m512d __M, __m512d __V) {
+  // CHECK-LABEL: @test_mm512_shuffle_pd
+  // CHECK: @llvm.x86.avx512.mask.shuf.pd.512
+  return _mm512_shuffle_pd(__M, __V, 4); 
+}
+
+__m512d test_mm512_mask_shuffle_pd(__m512d __W, __mmask8 __U, __m512d __M, __m512d __V) {
+  // CHECK-LABEL: @test_mm512_mask_shuffle_pd
+  // CHECK: @llvm.x86.avx512.mask.shuf.pd.512
+  return _mm512_mask_shuffle_pd(__W, __U, __M, __V, 4); 
+}
+
+__m512d test_mm512_maskz_shuffle_pd(__mmask8 __U, __m512d __M, __m512d __V) {
+  // CHECK-LABEL: @test_mm512_maskz_shuffle_pd
+  // CHECK: @llvm.x86.avx512.mask.shuf.pd.512
+  return _mm512_maskz_shuffle_pd(__U, __M, __V, 4); 
+}
+
+__m512 test_mm512_shuffle_ps(__m512 __M, __m512 __V) {
+  // CHECK-LABEL: @test_mm512_shuffle_ps
+  // CHECK: @llvm.x86.avx512.mask.shuf.ps.512
+  return _mm512_shuffle_ps(__M, __V, 4); 
+}
+
+__m512 test_mm512_mask_shuffle_ps(__m512 __W, __mmask16 __U, __m512 __M, __m512 __V) {
+  // CHECK-LABEL: @test_mm512_mask_shuffle_ps
+  // CHECK: @llvm.x86.avx512.mask.shuf.ps.512
+  return _mm512_mask_shuffle_ps(__W, __U, __M, __V, 4); 
+}
+
+__m512 test_mm512_maskz_shuffle_ps(__mmask16 __U, __m512 __M, __m512 __V) {
+  // CHECK-LABEL: @test_mm512_maskz_shuffle_ps
+  // CHECK: @llvm.x86.avx512.mask.shuf.ps.512
+  return _mm512_maskz_shuffle_ps(__U, __M, __V, 4); 
+}
+
+__m128d test_mm_sqrt_round_sd(__m128d __A, __m128d __B) {
+  // CHECK-LABEL: @test_mm_sqrt_round_sd
+  // CHECK: @llvm.x86.avx512.mask.sqrt.sd
+  return _mm_sqrt_round_sd(__A, __B, 4); 
+}
+
+__m128d test_mm_mask_sqrt_sd(__m128d __W, __mmask8 __U, __m128d __A, __m128d __B){
+  // CHECK: @llvm.x86.avx512.mask.sqrt.sd
+    return _mm_mask_sqrt_sd(__W,__U,__A,__B);
+}
+
+__m128d test_mm_mask_sqrt_round_sd(__m128d __W, __mmask8 __U, __m128d __A, __m128d __B){
+  // CHECK: @llvm.x86.avx512.mask.sqrt.sd
+    return _mm_mask_sqrt_round_sd(__W,__U,__A,__B,_MM_FROUND_CUR_DIRECTION);
+}
+
+__m128d test_mm_maskz_sqrt_sd(__mmask8 __U, __m128d __A, __m128d __B){
+  // CHECK: @llvm.x86.avx512.mask.sqrt.sd
+    return _mm_maskz_sqrt_sd(__U,__A,__B);
+}
+
+__m128d test_mm_maskz_sqrt_round_sd(__mmask8 __U, __m128d __A, __m128d __B){
+  // CHECK: @llvm.x86.avx512.mask.sqrt.sd
+    return _mm_maskz_sqrt_round_sd(__U,__A,__B,_MM_FROUND_CUR_DIRECTION);
+}
+
+__m128 test_mm_sqrt_round_ss(__m128 __A, __m128 __B) {
+  // CHECK-LABEL: @test_mm_sqrt_round_ss
+  // CHECK: @llvm.x86.avx512.mask.sqrt.ss
+  return _mm_sqrt_round_ss(__A, __B, 4); 
+}
+
+__m128 test_mm_mask_sqrt_ss(__m128 __W, __mmask8 __U, __m128 __A, __m128 __B){
+  // CHECK: @llvm.x86.avx512.mask.sqrt.ss
+    return _mm_mask_sqrt_ss(__W,__U,__A,__B);
+}
+
+__m128 test_mm_mask_sqrt_round_ss(__m128 __W, __mmask8 __U, __m128 __A, __m128 __B){
+  // CHECK: @llvm.x86.avx512.mask.sqrt.ss
+    return _mm_mask_sqrt_round_ss(__W,__U,__A,__B,_MM_FROUND_CUR_DIRECTION);
+}
+
+__m128 test_mm_maskz_sqrt_ss(__mmask8 __U, __m128 __A, __m128 __B){
+  // 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);
+}
 
 
Index: test/CodeGen/avx512vl-builtins.c
===================================================================
--- test/CodeGen/avx512vl-builtins.c
+++ test/CodeGen/avx512vl-builtins.c
@@ -4296,3 +4296,194 @@
   return _mm256_maskz_loadu_ps(__U, __P); 
 }
 
+__m256 test_mm256_shuffle_f32x4(__m256 __A, __m256 __B) {
+  // CHECK-LABEL: @test_mm256_shuffle_f32x4
+  // CHECK: @llvm.x86.avx512.mask.shuf
+  return _mm256_shuffle_f32x4(__A, __B, _MM_FROUND_CUR_DIRECTION); 
+}
+
+__m256 test_mm256_mask_shuffle_f32x4(__m256 __W, __mmask8 __U, __m256 __A, __m256 __B) {
+  // CHECK-LABEL: @test_mm256_mask_shuffle_f32x4
+  // CHECK: @llvm.x86.avx512.mask.shuf
+  return _mm256_mask_shuffle_f32x4(__W, __U, __A, __B, _MM_FROUND_CUR_DIRECTION); 
+}
+
+__m256 test_mm256_maskz_shuffle_f32x4(__mmask8 __U, __m256 __A, __m256 __B) {
+  // CHECK-LABEL: @test_mm256_maskz_shuffle_f32x4
+  // CHECK: @llvm.x86.avx512.mask.shuf
+  return _mm256_maskz_shuffle_f32x4(__U, __A, __B, _MM_FROUND_CUR_DIRECTION); 
+}
+
+__m256d test_mm256_shuffle_f64x2(__m256d __A, __m256d __B) {
+  // CHECK-LABEL: @test_mm256_shuffle_f64x2
+  // CHECK: @llvm.x86.avx512.mask.shuf
+  return _mm256_shuffle_f64x2(__A, __B, _MM_FROUND_CUR_DIRECTION); 
+}
+
+__m256d test_mm256_mask_shuffle_f64x2(__m256d __W, __mmask8 __U, __m256d __A, __m256d __B) {
+  // CHECK-LABEL: @test_mm256_mask_shuffle_f64x2
+  // CHECK: @llvm.x86.avx512.mask.shuf
+  return _mm256_mask_shuffle_f64x2(__W, __U, __A, __B, _MM_FROUND_CUR_DIRECTION); 
+}
+
+__m256d test_mm256_maskz_shuffle_f64x2(__mmask8 __U, __m256d __A, __m256d __B) {
+  // CHECK-LABEL: @test_mm256_maskz_shuffle_f64x2
+  // CHECK: @llvm.x86.avx512.mask.shuf
+  return _mm256_maskz_shuffle_f64x2(__U, __A, __B, _MM_FROUND_CUR_DIRECTION); 
+}
+
+__m256i test_mm256_shuffle_i32x4(__m256i __A, __m256i __B) {
+  // CHECK-LABEL: @test_mm256_shuffle_i32x4
+  // CHECK: @llvm.x86.avx512.mask.shuf
+  return _mm256_shuffle_i32x4(__A, __B, _MM_FROUND_CUR_DIRECTION); 
+}
+
+__m256i test_mm256_mask_shuffle_i32x4(__m256i __W, __mmask8 __U, __m256i __A, __m256i __B) {
+  // CHECK-LABEL: @test_mm256_mask_shuffle_i32x4
+  // CHECK: @llvm.x86.avx512.mask.shuf
+  return _mm256_mask_shuffle_i32x4(__W, __U, __A, __B, _MM_FROUND_CUR_DIRECTION); 
+}
+
+__m256i test_mm256_maskz_shuffle_i32x4(__mmask8 __U, __m256i __A, __m256i __B) {
+  // CHECK-LABEL: @test_mm256_maskz_shuffle_i32x4
+  // CHECK: @llvm.x86.avx512.mask.shuf
+  return _mm256_maskz_shuffle_i32x4(__U, __A, __B, _MM_FROUND_CUR_DIRECTION); 
+}
+
+__m256i test_mm256_shuffle_i64x2(__m256i __A, __m256i __B) {
+  // CHECK-LABEL: @test_mm256_shuffle_i64x2
+  // CHECK: @llvm.x86.avx512.mask.shuf
+  return _mm256_shuffle_i64x2(__A, __B, _MM_FROUND_CUR_DIRECTION); 
+}
+
+__m256i test_mm256_mask_shuffle_i64x2(__m256i __W, __mmask8 __U, __m256i __A, __m256i __B) {
+  // CHECK-LABEL: @test_mm256_mask_shuffle_i64x2
+  // CHECK: @llvm.x86.avx512.mask.shuf
+  return _mm256_mask_shuffle_i64x2(__W, __U, __A, __B, _MM_FROUND_CUR_DIRECTION); 
+}
+
+__m256i test_mm256_maskz_shuffle_i64x2(__mmask8 __U, __m256i __A, __m256i __B) {
+  // CHECK-LABEL: @test_mm256_maskz_shuffle_i64x2
+  // CHECK: @llvm.x86.avx512.mask.shuf
+  return _mm256_maskz_shuffle_i64x2(__U, __A, __B, _MM_FROUND_CUR_DIRECTION); 
+}
+
+__m128d test_mm_mask_shuffle_pd(__m128d __W, __mmask8 __U, __m128d __A, __m128d __B) {
+  // CHECK-LABEL: @test_mm_mask_shuffle_pd
+  // CHECK: @llvm.x86.avx512.mask.shuf.pd.128
+  return _mm_mask_shuffle_pd(__W, __U, __A, __B, _MM_FROUND_CUR_DIRECTION); 
+}
+
+__m128d test_mm_maskz_shuffle_pd(__mmask8 __U, __m128d __A, __m128d __B) {
+  // CHECK-LABEL: @test_mm_maskz_shuffle_pd
+  // CHECK: @llvm.x86.avx512.mask.shuf.pd.128
+  return _mm_maskz_shuffle_pd(__U, __A, __B, _MM_FROUND_CUR_DIRECTION); 
+}
+
+__m256d test_mm256_mask_shuffle_pd(__m256d __W, __mmask8 __U, __m256d __A, __m256d __B) {
+  // CHECK-LABEL: @test_mm256_mask_shuffle_pd
+  // CHECK: @llvm.x86.avx512.mask.shuf.pd.256
+  return _mm256_mask_shuffle_pd(__W, __U, __A, __B, _MM_FROUND_CUR_DIRECTION); 
+}
+
+__m256d test_mm256_maskz_shuffle_pd(__mmask8 __U, __m256d __A, __m256d __B) {
+  // CHECK-LABEL: @test_mm256_maskz_shuffle_pd
+  // CHECK: @llvm.x86.avx512.mask.shuf.pd.256
+  return _mm256_maskz_shuffle_pd(__U, __A, __B, _MM_FROUND_CUR_DIRECTION); 
+}
+
+__m128 test_mm_mask_shuffle_ps(__m128 __W, __mmask8 __U, __m128 __A, __m128 __B) {
+  // CHECK-LABEL: @test_mm_mask_shuffle_ps
+  // CHECK: @llvm.x86.avx512.mask.shuf.ps.128
+  return _mm_mask_shuffle_ps(__W, __U, __A, __B, _MM_FROUND_CUR_DIRECTION); 
+}
+
+__m128 test_mm_maskz_shuffle_ps(__mmask8 __U, __m128 __A, __m128 __B) {
+  // CHECK-LABEL: @test_mm_maskz_shuffle_ps
+  // CHECK: @llvm.x86.avx512.mask.shuf.ps.128
+  return _mm_maskz_shuffle_ps(__U, __A, __B, _MM_FROUND_CUR_DIRECTION); 
+}
+
+__m256 test_mm256_mask_shuffle_ps(__m256 __W, __mmask8 __U, __m256 __A, __m256 __B) {
+  // CHECK-LABEL: @test_mm256_mask_shuffle_ps
+  // CHECK: @llvm.x86.avx512.mask.shuf.ps.256
+  return _mm256_mask_shuffle_ps(__W, __U, __A, __B, _MM_FROUND_CUR_DIRECTION); 
+}
+
+__m256 test_mm256_maskz_shuffle_ps(__mmask8 __U, __m256 __A, __m256 __B) {
+  // CHECK-LABEL: @test_mm256_maskz_shuffle_ps
+  // CHECK: @llvm.x86.avx512.mask.shuf.ps.256
+  return _mm256_maskz_shuffle_ps(__U, __A, __B, _MM_FROUND_CUR_DIRECTION); 
+}
+
+__m128d test_mm_rsqrt14_pd(__m128d __A) {
+  // CHECK-LABEL: @test_mm_rsqrt14_pd
+  // CHECK: @llvm.x86.avx512.rsqrt14.pd.128
+  return _mm_rsqrt14_pd(__A); 
+}
+
+__m128d test_mm_mask_rsqrt14_pd(__m128d __W, __mmask8 __U, __m128d __A) {
+  // CHECK-LABEL: @test_mm_mask_rsqrt14_pd
+  // CHECK: @llvm.x86.avx512.rsqrt14.pd.128
+  return _mm_mask_rsqrt14_pd(__W, __U, __A); 
+}
+
+__m128d test_mm_maskz_rsqrt14_pd(__mmask8 __U, __m128d __A) {
+  // CHECK-LABEL: @test_mm_maskz_rsqrt14_pd
+  // CHECK: @llvm.x86.avx512.rsqrt14.pd.128
+  return _mm_maskz_rsqrt14_pd(__U, __A); 
+}
+
+__m256d test_mm256_rsqrt14_pd(__m256d __A) {
+  // CHECK-LABEL: @test_mm256_rsqrt14_pd
+  // CHECK: @llvm.x86.avx512.rsqrt14.pd.256
+  return _mm256_rsqrt14_pd(__A); 
+}
+
+__m256d test_mm256_mask_rsqrt14_pd(__m256d __W, __mmask8 __U, __m256d __A) {
+  // CHECK-LABEL: @test_mm256_mask_rsqrt14_pd
+  // CHECK: @llvm.x86.avx512.rsqrt14.pd.256
+  return _mm256_mask_rsqrt14_pd(__W, __U, __A); 
+}
+
+__m256d test_mm256_maskz_rsqrt14_pd(__mmask8 __U, __m256d __A) {
+  // CHECK-LABEL: @test_mm256_maskz_rsqrt14_pd
+  // CHECK: @llvm.x86.avx512.rsqrt14.pd.256
+  return _mm256_maskz_rsqrt14_pd(__U, __A); 
+}
+
+__m128 test_mm_rsqrt14_ps(__m128 __A) {
+  // CHECK-LABEL: @test_mm_rsqrt14_ps
+  // CHECK: @llvm.x86.avx512.rsqrt14.ps.128
+  return _mm_rsqrt14_ps(__A); 
+}
+
+__m128 test_mm_mask_rsqrt14_ps(__m128 __W, __mmask8 __U, __m128 __A) {
+  // CHECK-LABEL: @test_mm_mask_rsqrt14_ps
+  // CHECK: @llvm.x86.avx512.rsqrt14.ps.128
+  return _mm_mask_rsqrt14_ps(__W, __U, __A); 
+}
+
+__m128 test_mm_maskz_rsqrt14_ps(__mmask8 __U, __m128 __A) {
+  // CHECK-LABEL: @test_mm_maskz_rsqrt14_ps
+  // CHECK: @llvm.x86.avx512.rsqrt14.ps.128
+  return _mm_maskz_rsqrt14_ps(__U, __A); 
+}
+
+__m256 test_mm256_rsqrt14_ps(__m256 __A) {
+  // CHECK-LABEL: @test_mm256_rsqrt14_ps
+  // CHECK: @llvm.x86.avx512.rsqrt14.ps.256
+  return _mm256_rsqrt14_ps(__A); 
+}
+
+__m256 test_mm256_mask_rsqrt14_ps(__m256 __W, __mmask8 __U, __m256 __A) {
+  // CHECK-LABEL: @test_mm256_mask_rsqrt14_ps
+  // CHECK: @llvm.x86.avx512.rsqrt14.ps.256
+  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); 
+}