[Clang][AVX512][BuiltIn] Adding avx512 ( vperm{i|t}2var, vpermil{var}{ps|pd}{256|512} ) builtin to clang.

Differential Revision: http://reviews.llvm.org/D18933

llvm-svn: 265915

2 files changed, 320 insertions, 0 deletions

diff --git a/clang/lib/Headers/avx512fintrin.h b/clang/lib/Headers/avx512fintrin.h
index 23f5ac1b335..fe27795abf6 100644
--- a/clang/lib/Headers/avx512fintrin.h
+++ b/clang/lib/Headers/avx512fintrin.h
@@ -3985,6 +3985,16 @@ _mm512_mask_unpackhi_pd (__m512d __W, __mmask8 __U, __m512d __A, __m512d __B)
 #define _mm_cvt_roundsd_si64( __A, __R) __extension__ ({ \
 __builtin_ia32_vcvtsd2si64 ((__v2df)( __A),( __R));\
 })
+static __inline__ __m512i __DEFAULT_FN_ATTRS
+_mm512_mask2_permutex2var_epi32 (__m512i __A, __m512i __I,
+         __mmask16 __U, __m512i __B)
+{
+  return (__m512i) __builtin_ia32_vpermi2vard512_mask ((__v16si) __A,
+                   (__v16si) __I
+                   /* idx */ ,
+                   (__v16si) __B,
+                   (__mmask16) __U);
+}
 
 static __inline__ __m512d __DEFAULT_FN_ATTRS
 _mm512_maskz_unpackhi_pd (__mmask8 __U, __m512d __A, __m512d __B)
@@ -4233,6 +4243,181 @@ _mm_cvttss_u64 (__m128 __A)
                   _MM_FROUND_CUR_DIRECTION);
 }
 
+static __inline__ __m512d __DEFAULT_FN_ATTRS
+_mm512_mask2_permutex2var_pd (__m512d __A, __m512i __I, __mmask8 __U,
+            __m512d __B)
+{
+  return (__m512d) __builtin_ia32_vpermi2varpd512_mask ((__v8df) __A,
+              (__v8di) __I
+              /* idx */ ,
+              (__v8df) __B,
+              (__mmask8) __U);
+}
+
+static __inline__ __m512 __DEFAULT_FN_ATTRS
+_mm512_mask2_permutex2var_ps (__m512 __A, __m512i __I, __mmask16 __U,
+            __m512 __B)
+{
+  return (__m512) __builtin_ia32_vpermi2varps512_mask ((__v16sf) __A,
+                   (__v16si) __I
+                   /* idx */ ,
+                   (__v16sf) __B,
+                   (__mmask16) __U);
+}
+
+static __inline__ __m512i __DEFAULT_FN_ATTRS
+_mm512_mask2_permutex2var_epi64 (__m512i __A, __m512i __I,
+         __mmask8 __U, __m512i __B)
+{
+  return (__m512i) __builtin_ia32_vpermi2varq512_mask ((__v8di) __A,
+                   (__v8di) __I
+                   /* idx */ ,
+                   (__v8di) __B,
+                   (__mmask8) __U);
+}
+
+#define _mm512_permute_pd( __X, __C) __extension__ ({ \
+__builtin_ia32_vpermilpd512_mask ((__v8df)( __X),( __C),\
+                 (__v8df)\
+                 _mm512_undefined_pd (),\
+                 (__mmask8) -1);\
+})
+
+#define _mm512_mask_permute_pd( __W, __U, __X, __C) __extension__ ({ \
+__builtin_ia32_vpermilpd512_mask ((__v8df)( __X),( __C),\
+                 (__v8df)( __W),\
+                 (__mmask8)( __U));\
+})
+
+#define _mm512_maskz_permute_pd( __U, __X, __C) __extension__ ({ \
+__builtin_ia32_vpermilpd512_mask ((__v8df)( __X),( __C),\
+                 (__v8df)\
+                 _mm512_setzero_pd (),\
+                 (__mmask8)( __U));\
+})
+
+#define _mm512_permute_ps( __X, __C) __extension__ ({ \
+__builtin_ia32_vpermilps512_mask ((__v16sf)( __X),( __C),\
+                (__v16sf)\
+                _mm512_undefined_ps (),\
+                (__mmask16) -1);\
+})
+
+#define _mm512_mask_permute_ps( __W, __U, __X, __C) __extension__ ({ \
+__builtin_ia32_vpermilps512_mask ((__v16sf)( __X),( __C),\
+                (__v16sf)( __W),\
+                (__mmask16)( __U));\
+})
+
+#define _mm512_maskz_permute_ps( __U, __X, __C) __extension__ ({ \
+__builtin_ia32_vpermilps512_mask ((__v16sf)( __X),( __C),\
+                (__v16sf)\
+                _mm512_setzero_ps (),\
+                (__mmask16)( __U));\
+})
+
+static __inline__ __m512d __DEFAULT_FN_ATTRS
+_mm512_permutevar_pd (__m512d __A, __m512i __C)
+{
+  return (__m512d) __builtin_ia32_vpermilvarpd512_mask ((__v8df) __A,
+              (__v8di) __C,
+              (__v8df)
+              _mm512_undefined_pd (),
+              (__mmask8) -1);
+}
+
+static __inline__ __m512d __DEFAULT_FN_ATTRS
+_mm512_mask_permutevar_pd (__m512d __W, __mmask8 __U, __m512d __A, __m512i __C)
+{
+  return (__m512d) __builtin_ia32_vpermilvarpd512_mask ((__v8df) __A,
+              (__v8di) __C,
+              (__v8df) __W,
+              (__mmask8) __U);
+}
+
+static __inline__ __m512d __DEFAULT_FN_ATTRS
+_mm512_maskz_permutevar_pd (__mmask8 __U, __m512d __A, __m512i __C)
+{
+  return (__m512d) __builtin_ia32_vpermilvarpd512_mask ((__v8df) __A,
+              (__v8di) __C,
+              (__v8df)
+              _mm512_setzero_pd (),
+              (__mmask8) __U);
+}
+
+static __inline__ __m512 __DEFAULT_FN_ATTRS
+_mm512_permutevar_ps (__m512 __A, __m512i __C)
+{
+  return (__m512) __builtin_ia32_vpermilvarps512_mask ((__v16sf) __A,
+                   (__v16si) __C,
+                   (__v16sf)
+                   _mm512_undefined_ps (),
+                   (__mmask16) -1);
+}
+
+static __inline__ __m512 __DEFAULT_FN_ATTRS
+_mm512_mask_permutevar_ps (__m512 __W, __mmask16 __U, __m512 __A, __m512i __C)
+{
+  return (__m512) __builtin_ia32_vpermilvarps512_mask ((__v16sf) __A,
+                   (__v16si) __C,
+                   (__v16sf) __W,
+                   (__mmask16) __U);
+}
+
+static __inline__ __m512 __DEFAULT_FN_ATTRS
+_mm512_maskz_permutevar_ps (__mmask16 __U, __m512 __A, __m512i __C)
+{
+  return (__m512) __builtin_ia32_vpermilvarps512_mask ((__v16sf) __A,
+                   (__v16si) __C,
+                   (__v16sf)
+                   _mm512_setzero_ps (),
+                   (__mmask16) __U);
+}
+
+static __inline__ __m512i __DEFAULT_FN_ATTRS
+_mm512_maskz_permutex2var_epi32 (__mmask16 __U, __m512i __A,
+         __m512i __I, __m512i __B)
+{
+  return (__m512i) __builtin_ia32_vpermt2vard512_maskz ((__v16si) __I
+              /* idx */ ,
+              (__v16si) __A,
+              (__v16si) __B,
+              (__mmask16) __U);
+}
+
+static __inline__ __m512d __DEFAULT_FN_ATTRS
+_mm512_maskz_permutex2var_pd (__mmask8 __U, __m512d __A, __m512i __I,
+            __m512d __B)
+{
+  return (__m512d) __builtin_ia32_vpermt2varpd512_maskz ((__v8di) __I
+               /* idx */ ,
+               (__v8df) __A,
+               (__v8df) __B,
+               (__mmask8) __U);
+}
+
+static __inline__ __m512 __DEFAULT_FN_ATTRS
+_mm512_maskz_permutex2var_ps (__mmask16 __U, __m512 __A, __m512i __I,
+            __m512 __B)
+{
+  return (__m512) __builtin_ia32_vpermt2varps512_maskz ((__v16si) __I
+              /* idx */ ,
+              (__v16sf) __A,
+              (__v16sf) __B,
+              (__mmask16) __U);
+}
+
+static __inline__ __m512i __DEFAULT_FN_ATTRS
+_mm512_maskz_permutex2var_epi64 (__mmask8 __U, __m512i __A,
+         __m512i __I, __m512i __B)
+{
+  return (__m512i) __builtin_ia32_vpermt2varq512_maskz ((__v8di) __I
+              /* idx */ ,
+              (__v8di) __A,
+              (__v8di) __B,
+              (__mmask8) __U);
+}
+
 #undef __DEFAULT_FN_ATTRS
 
 #endif // __AVX512FINTRIN_H
diff --git a/clang/lib/Headers/avx512vlintrin.h b/clang/lib/Headers/avx512vlintrin.h
index fd18e7177b7..aa7c172ecaf 100644
--- a/clang/lib/Headers/avx512vlintrin.h
+++ b/clang/lib/Headers/avx512vlintrin.h
@@ -6824,6 +6824,141 @@ _mm256_maskz_lzcnt_epi64 (__mmask8 __U, __m256i __A)
                  (__mmask8) __U);
 }
 
+
+#define _mm_mask_permute_pd( __W, __U, __X, __C) __extension__ ({ \
+__builtin_ia32_vpermilpd_mask ((__v2df)( __X),( __C),\
+              (__v2df)( __W),\
+              (__mmask8)( __U));\
+})
+
+#define _mm_maskz_permute_pd( __U, __X, __C) __extension__ ({ \
+__builtin_ia32_vpermilpd_mask ((__v2df)( __X),( __C),\
+              (__v2df)\
+              _mm_setzero_pd (),\
+              (__mmask8)( __U));\
+})
+
+#define _mm256_mask_permute_pd( __W, __U, __X, __C) __extension__ ({ \
+__builtin_ia32_vpermilpd256_mask ((__v4df)( __X),( __C),\
+                 (__v4df)( __W),\
+                 (__mmask8)( __U));\
+})
+
+#define _mm256_maskz_permute_pd( __U, __X, __C) __extension__ ({ \
+__builtin_ia32_vpermilpd256_mask ((__v4df)( __X),( __C),\
+                 (__v4df)\
+                 _mm256_setzero_pd (),\
+                 (__mmask8)( __U));\
+})
+
+#define _mm_mask_permute_ps( __W, __U, __X, __C) __extension__ ({ \
+__builtin_ia32_vpermilps_mask ((__v4sf)( __X),( __C),\
+             (__v4sf)( __W),\
+             (__mmask8)( __U));\
+})
+
+#define _mm_maskz_permute_ps( __U, __X, __C) __extension__ ({ \
+__builtin_ia32_vpermilps_mask ((__v4sf)( __X),( __C),\
+             (__v4sf)\
+             _mm_setzero_ps (),\
+             (__mmask8)( __U));\
+})
+
+#define _mm256_mask_permute_ps( __W, __U, __X, __C) __extension__ ({ \
+__builtin_ia32_vpermilps256_mask ((__v8sf)( __X),( __C),\
+                (__v8sf)( __W),\
+                (__mmask8)( __U));\
+})
+
+#define _mm256_maskz_permute_ps( __U, __X, __C) __extension__ ({ \
+__builtin_ia32_vpermilps256_mask ((__v8sf)( __X),( __C),\
+                (__v8sf)\
+                _mm256_setzero_ps (),\
+                (__mmask8)( __U));\
+})
+
+static __inline__ __m128d __DEFAULT_FN_ATTRS
+_mm_mask_permutevar_pd (__m128d __W, __mmask8 __U, __m128d __A,
+      __m128i __C)
+{
+  return (__m128d) __builtin_ia32_vpermilvarpd_mask ((__v2df) __A,
+                 (__v2di) __C,
+                 (__v2df) __W,
+                 (__mmask8) __U);
+}
+
+static __inline__ __m128d __DEFAULT_FN_ATTRS
+_mm_maskz_permutevar_pd (__mmask8 __U, __m128d __A, __m128i __C)
+{
+  return (__m128d) __builtin_ia32_vpermilvarpd_mask ((__v2df) __A,
+                 (__v2di) __C,
+                 (__v2df)
+                 _mm_setzero_pd (),
+                 (__mmask8) __U);
+}
+
+static __inline__ __m256d __DEFAULT_FN_ATTRS
+_mm256_mask_permutevar_pd (__m256d __W, __mmask8 __U, __m256d __A,
+         __m256i __C)
+{
+  return (__m256d) __builtin_ia32_vpermilvarpd256_mask ((__v4df) __A,
+              (__v4di) __C,
+              (__v4df) __W,
+              (__mmask8)
+              __U);
+}
+
+static __inline__ __m256d __DEFAULT_FN_ATTRS
+_mm256_maskz_permutevar_pd (__mmask8 __U, __m256d __A, __m256i __C)
+{
+  return (__m256d) __builtin_ia32_vpermilvarpd256_mask ((__v4df) __A,
+              (__v4di) __C,
+              (__v4df)
+              _mm256_setzero_pd (),
+              (__mmask8)
+              __U);
+}
+
+static __inline__ __m128 __DEFAULT_FN_ATTRS
+_mm_mask_permutevar_ps (__m128 __W, __mmask8 __U, __m128 __A,
+      __m128i __C)
+{
+  return (__m128) __builtin_ia32_vpermilvarps_mask ((__v4sf) __A,
+                (__v4si) __C,
+                (__v4sf) __W,
+                (__mmask8) __U);
+}
+
+static __inline__ __m128 __DEFAULT_FN_ATTRS
+_mm_maskz_permutevar_ps (__mmask8 __U, __m128 __A, __m128i __C)
+{
+  return (__m128) __builtin_ia32_vpermilvarps_mask ((__v4sf) __A,
+                (__v4si) __C,
+                (__v4sf)
+                _mm_setzero_ps (),
+                (__mmask8) __U);
+}
+
+static __inline__ __m256 __DEFAULT_FN_ATTRS
+_mm256_mask_permutevar_ps (__m256 __W, __mmask8 __U, __m256 __A,
+         __m256i __C)
+{
+  return (__m256) __builtin_ia32_vpermilvarps256_mask ((__v8sf) __A,
+                   (__v8si) __C,
+                   (__v8sf) __W,
+                   (__mmask8) __U);
+}
+
+static __inline__ __m256 __DEFAULT_FN_ATTRS
+_mm256_maskz_permutevar_ps (__mmask8 __U, __m256 __A, __m256i __C)
+{
+  return (__m256) __builtin_ia32_vpermilvarps256_mask ((__v8sf) __A,
+                   (__v8si) __C,
+                   (__v8sf)
+                   _mm256_setzero_ps (),
+                   (__mmask8) __U);
+}
+
 #undef __DEFAULT_FN_ATTRS
 #undef __DEFAULT_FN_ATTRS_BOTH