[AVX512] Add parentheses around macro arguments in AVX512VLBW intrinsics. Remove leading underscores from macro argument names. Add explicit typecasts to all macro arguments and return values. And finally reformat after all the adjustments.

This is a mostly mechanical change accomplished with a script. I tried to split out any changes to the typecasts that already existed into separate commits.

git-svn-id: https://llvm.org/svn/llvm-project/cfe/trunk@269743 91177308-0d34-0410-b5e6-96231b3b80d8

1 files changed, 184 insertions, 240 deletions

diff --git a/lib/Headers/avx512vlbwintrin.h b/lib/Headers/avx512vlbwintrin.h
index d8e67fc15a..d35dc7e3bf 100644
--- a/lib/Headers/avx512vlbwintrin.h
+++ b/lib/Headers/avx512vlbwintrin.h
@@ -2328,144 +2328,128 @@ _mm256_maskz_cvtepu8_epi16 (__mmask16 __U, __m128i __A)
 
 #define _mm_cmp_epi8_mask(a, b, p) __extension__ ({ \
   (__mmask16)__builtin_ia32_cmpb128_mask((__v16qi)(__m128i)(a), \
-                                         (__v16qi)(__m128i)(b), \
-                                         (p), (__mmask16)-1); })
+                                         (__v16qi)(__m128i)(b), (int)(p), \
+                                         (__mmask16)-1); })
 
 #define _mm_mask_cmp_epi8_mask(m, a, b, p) __extension__ ({ \
   (__mmask16)__builtin_ia32_cmpb128_mask((__v16qi)(__m128i)(a), \
-                                         (__v16qi)(__m128i)(b), \
-                                         (p), (__mmask16)(m)); })
+                                         (__v16qi)(__m128i)(b), (int)(p), \
+                                         (__mmask16)(m)); })
 
 #define _mm_cmp_epu8_mask(a, b, p) __extension__ ({ \
   (__mmask16)__builtin_ia32_ucmpb128_mask((__v16qi)(__m128i)(a), \
-                                          (__v16qi)(__m128i)(b), \
-                                          (p), (__mmask16)-1); })
+                                          (__v16qi)(__m128i)(b), (int)(p), \
+                                          (__mmask16)-1); })
 
 #define _mm_mask_cmp_epu8_mask(m, a, b, p) __extension__ ({ \
   (__mmask16)__builtin_ia32_ucmpb128_mask((__v16qi)(__m128i)(a), \
-                                          (__v16qi)(__m128i)(b), \
-                                          (p), (__mmask16)(m)); })
+                                          (__v16qi)(__m128i)(b), (int)(p), \
+                                          (__mmask16)(m)); })
 
 #define _mm256_cmp_epi8_mask(a, b, p) __extension__ ({ \
   (__mmask32)__builtin_ia32_cmpb256_mask((__v32qi)(__m256i)(a), \
-                                         (__v32qi)(__m256i)(b), \
-                                         (p), (__mmask32)-1); })
+                                         (__v32qi)(__m256i)(b), (int)(p), \
+                                         (__mmask32)-1); })
 
 #define _mm256_mask_cmp_epi8_mask(m, a, b, p) __extension__ ({ \
   (__mmask32)__builtin_ia32_cmpb256_mask((__v32qi)(__m256i)(a), \
-                                         (__v32qi)(__m256i)(b), \
-                                         (p), (__mmask32)(m)); })
+                                         (__v32qi)(__m256i)(b), (int)(p), \
+                                         (__mmask32)(m)); })
 
 #define _mm256_cmp_epu8_mask(a, b, p) __extension__ ({ \
   (__mmask32)__builtin_ia32_ucmpb256_mask((__v32qi)(__m256i)(a), \
-                                          (__v32qi)(__m256i)(b), \
-                                          (p), (__mmask32)-1); })
+                                          (__v32qi)(__m256i)(b), (int)(p), \
+                                          (__mmask32)-1); })
 
 #define _mm256_mask_cmp_epu8_mask(m, a, b, p) __extension__ ({ \
   (__mmask32)__builtin_ia32_ucmpb256_mask((__v32qi)(__m256i)(a), \
-                                          (__v32qi)(__m256i)(b), \
-                                          (p), (__mmask32)(m)); })
+                                          (__v32qi)(__m256i)(b), (int)(p), \
+                                          (__mmask32)(m)); })
 
 #define _mm_cmp_epi16_mask(a, b, p) __extension__ ({ \
   (__mmask8)__builtin_ia32_cmpw128_mask((__v8hi)(__m128i)(a), \
-                                        (__v8hi)(__m128i)(b), \
-                                        (p), (__mmask8)-1); })
+                                        (__v8hi)(__m128i)(b), (int)(p), \
+                                        (__mmask8)-1); })
 
 #define _mm_mask_cmp_epi16_mask(m, a, b, p) __extension__ ({ \
   (__mmask8)__builtin_ia32_cmpw128_mask((__v8hi)(__m128i)(a), \
-                                        (__v8hi)(__m128i)(b), \
-                                        (p), (__mmask8)(m)); })
+                                        (__v8hi)(__m128i)(b), (int)(p), \
+                                        (__mmask8)(m)); })
 
 #define _mm_cmp_epu16_mask(a, b, p) __extension__ ({ \
   (__mmask8)__builtin_ia32_ucmpw128_mask((__v8hi)(__m128i)(a), \
-                                         (__v8hi)(__m128i)(b), \
-                                         (p), (__mmask8)-1); })
+                                         (__v8hi)(__m128i)(b), (int)(p), \
+                                         (__mmask8)-1); })
 
 #define _mm_mask_cmp_epu16_mask(m, a, b, p) __extension__ ({ \
   (__mmask8)__builtin_ia32_ucmpw128_mask((__v8hi)(__m128i)(a), \
-                                         (__v8hi)(__m128i)(b), \
-                                         (p), (__mmask8)(m)); })
+                                         (__v8hi)(__m128i)(b), (int)(p), \
+                                         (__mmask8)(m)); })
 
 #define _mm256_cmp_epi16_mask(a, b, p) __extension__ ({ \
   (__mmask16)__builtin_ia32_cmpw256_mask((__v16hi)(__m256i)(a), \
-                                         (__v16hi)(__m256i)(b), \
-                                         (p), (__mmask16)-1); })
+                                         (__v16hi)(__m256i)(b), (int)(p), \
+                                         (__mmask16)-1); })
 
 #define _mm256_mask_cmp_epi16_mask(m, a, b, p) __extension__ ({ \
   (__mmask16)__builtin_ia32_cmpw256_mask((__v16hi)(__m256i)(a), \
-                                         (__v16hi)(__m256i)(b), \
-                                         (p), (__mmask16)(m)); })
+                                         (__v16hi)(__m256i)(b), (int)(p), \
+                                         (__mmask16)(m)); })
 
 #define _mm256_cmp_epu16_mask(a, b, p) __extension__ ({ \
   (__mmask16)__builtin_ia32_ucmpw256_mask((__v16hi)(__m256i)(a), \
-                                          (__v16hi)(__m256i)(b), \
-                                          (p), (__mmask16)-1); })
+                                          (__v16hi)(__m256i)(b), (int)(p), \
+                                          (__mmask16)-1); })
 
 #define _mm256_mask_cmp_epu16_mask(m, a, b, p) __extension__ ({ \
   (__mmask16)__builtin_ia32_ucmpw256_mask((__v16hi)(__m256i)(a), \
-                                          (__v16hi)(__m256i)(b), \
-                                          (p), (__mmask16)(m)); })
-
-#define _mm_mask_shufflehi_epi16( __W, __U, __A, __imm) __extension__ ({ \
-__builtin_ia32_pshufhw128_mask ((__v8hi) __A, (__imm),\
-               (__v8hi)( __W),\
-               (__mmask8)( __U));\
-})
-
-#define _mm_maskz_shufflehi_epi16( __U, __A, __imm) __extension__ ({ \
-__builtin_ia32_pshufhw128_mask ((__v8hi)( __A),( __imm),\
-               (__v8hi)\
-               _mm_setzero_hi (),\
-               (__mmask8)( __U));\
-})
-
-
-#define _mm256_mask_shufflehi_epi16( __W, __U, __A, __imm) __extension__ ({ \
-__builtin_ia32_pshufhw256_mask ((__v16hi) (__A),\
-               (__imm),\
-               (__v16hi)( __W),\
-               (__mmask16)( __U));\
-})
-
-
-#define _mm256_maskz_shufflehi_epi16( __U, __A, __imm) __extension__ ({ \
-__builtin_ia32_pshufhw256_mask ((__v16hi) (__A),\
-               (__imm),\
-               (__v16hi)\
-               _mm256_setzero_si256 (),\
-               (__mmask16)( __U));\
-})
-
-
-#define _mm_mask_shufflelo_epi16( __W, __U, __A, __imm) __extension__ ({ \
-__builtin_ia32_pshuflw128_mask ((__v8hi) __A, (__imm),\
-               (__v8hi)( __W),\
-               (__mmask8)( __U));\
-})
-
-#define _mm_maskz_shufflelo_epi16( __U, __A, __imm) __extension__ ({ \
-__builtin_ia32_pshuflw128_mask ((__v8hi)( __A),( __imm),\
-               (__v8hi)\
-               _mm_setzero_hi (),\
-               (__mmask8)( __U));\
-})
-
-
-#define _mm256_mask_shufflelo_epi16( __W, __U, __A, __imm) __extension__ ({ \
-__builtin_ia32_pshuflw256_mask ((__v16hi) (__A),\
-               (__imm),\
-               (__v16hi)( __W),\
-               (__mmask16)( __U));\
-})
-
-
-#define _mm256_maskz_shufflelo_epi16( __U, __A, __imm) __extension__ ({ \
-__builtin_ia32_pshuflw256_mask ((__v16hi) (__A),\
-               (__imm),\
-               (__v16hi)\
-               _mm256_setzero_si256 (),\
-               (__mmask16)( __U));\
-})
+                                          (__v16hi)(__m256i)(b), (int)(p), \
+                                          (__mmask16)(m)); })
+
+#define _mm_mask_shufflehi_epi16(W, U, A, imm) __extension__ ({ \
+  (__m128i)__builtin_ia32_pshufhw128_mask((__v8hi)(__m128i)(A), (int)(imm), \
+                                          (__v8hi)(__m128i)(W), \
+                                          (__mmask8)(U)); })
+
+#define _mm_maskz_shufflehi_epi16(U, A, imm) __extension__ ({ \
+  (__m128i)__builtin_ia32_pshufhw128_mask((__v8hi)(__m128i)(A), (int)(imm), \
+                                          (__v8hi)_mm_setzero_hi(), \
+                                          (__mmask8)(U)); })
+
+
+#define _mm256_mask_shufflehi_epi16(W, U, A, imm) __extension__ ({ \
+  (__m256i)__builtin_ia32_pshufhw256_mask((__v16hi)(__m256i)(A), (int)(imm), \
+                                          (__v16hi)(__m256i)(W), \
+                                          (__mmask16)(U)); })
+
+
+#define _mm256_maskz_shufflehi_epi16(U, A, imm) __extension__ ({ \
+  (__m256i)__builtin_ia32_pshufhw256_mask((__v16hi)(__m256i)(A), (int)(imm), \
+                                          (__v16hi)_mm256_setzero_si256(), \
+                                          (__mmask16)(U)); })
+
+
+#define _mm_mask_shufflelo_epi16(W, U, A, imm) __extension__ ({ \
+  (__m128i)__builtin_ia32_pshuflw128_mask((__v8hi)(__m128i)(A), (int)(imm), \
+                                          (__v8hi)(__m128i)(W), \
+                                          (__mmask8)(U)); })
+
+#define _mm_maskz_shufflelo_epi16(U, A, imm) __extension__ ({ \
+  (__m128i)__builtin_ia32_pshuflw128_mask((__v8hi)(__m128i)(A), (int)(imm), \
+                                          (__v8hi)_mm_setzero_hi(), \
+                                          (__mmask8)(U)); })
+
+
+#define _mm256_mask_shufflelo_epi16(W, U, A, imm) __extension__ ({ \
+  (__m256i)__builtin_ia32_pshuflw256_mask((__v16hi)(__m256i)(A), (int)(imm), \
+                                          (__v16hi)(__m256i)(W), \
+                                          (__mmask16)(U)); })
+
+
+#define _mm256_maskz_shufflelo_epi16(U, A, imm) __extension__ ({ \
+  (__m256i)__builtin_ia32_pshuflw256_mask((__v16hi)(__m256i)(A), (int)(imm), \
+                                          (__v16hi)_mm256_setzero_si256(), \
+                                          (__mmask16)(U)); })
 
 static __inline__ __m256i __DEFAULT_FN_ATTRS
 _mm256_sllv_epi16 (__m256i __A, __m256i __B)
@@ -2567,31 +2551,25 @@ _mm256_maskz_sll_epi16 (__mmask16 __U, __m256i __A, __m128i __B)
              (__mmask16) __U);
 }
 
-#define _mm_mask_slli_epi16( __W, __U, __A, __B) __extension__ ({ \
-__builtin_ia32_psllwi128_mask ((__v8hi)( __A),( __B),\
-              (__v8hi)( __W),\
-              (__mmask8)( __U));\
-})
+#define _mm_mask_slli_epi16(W, U, A, B) __extension__ ({ \
+  (__m128i)__builtin_ia32_psllwi128_mask((__v8hi)(__m128i)(A), (int)(B), \
+                                         (__v8hi)(__m128i)(W), \
+                                         (__mmask8)(U)); })
 
-#define _mm_maskz_slli_epi16( __U, __A, __B) __extension__ ({ \
-__builtin_ia32_psllwi128_mask ((__v8hi)( __A),( __B),\
-              (__v8hi)\
-              _mm_setzero_si128 (),\
-              (__mmask8)( __U));\
-})
+#define _mm_maskz_slli_epi16(U, A, B) __extension__ ({ \
+  (__m128i)__builtin_ia32_psllwi128_mask((__v8hi)(__m128i)(A), (int)(B), \
+                                         (__v8hi)_mm_setzero_si128(), \
+                                         (__mmask8)(U)); })
 
-#define _mm256_mask_slli_epi16( __W, __U, __A, __B) __extension__ ({ \
-__builtin_ia32_psllwi256_mask ((__v16hi)( __A),( __B),\
-              (__v16hi)( __W),\
-              (__mmask16)( __U));\
-})
+#define _mm256_mask_slli_epi16(W, U, A, B) __extension__ ({ \
+  (__m256i)__builtin_ia32_psllwi256_mask((__v16hi)(__m256i)(A), (int)(B), \
+                                         (__v16hi)(__m256i)(W), \
+                                         (__mmask16)(U)); })
 
-#define _mm256_maskz_slli_epi16( __U, __A, __B) __extension__ ({ \
-__builtin_ia32_psllwi256_mask ((__v16hi)( __A),( __B),\
-              (__v16hi)\
-              _mm256_setzero_si256 (),\
-              (__mmask16)( __U));\
-})
+#define _mm256_maskz_slli_epi16(U, A, B) __extension__ ({ \
+  (__m256i)__builtin_ia32_psllwi256_mask((__v16hi)(__m256i)(A), (int)(B), \
+                                         (__v16hi)_mm256_setzero_si256(), \
+                                         (__mmask16)(U)); })
 
 
 
@@ -2755,31 +2733,25 @@ _mm256_maskz_sra_epi16 (__mmask16 __U, __m256i __A, __m128i __B)
              (__mmask16) __U);
 }
 
-#define _mm_mask_srai_epi16( __W, __U, __A, __imm) __extension__ ({ \
-__builtin_ia32_psrawi128_mask ((__v8hi)( __A),( __imm),\
-              (__v8hi)( __W),\
-              (__mmask8)( __U));\
-})
+#define _mm_mask_srai_epi16(W, U, A, imm) __extension__ ({ \
+  (__m128i)__builtin_ia32_psrawi128_mask((__v8hi)(__m128i)(A), (int)(imm), \
+                                         (__v8hi)(__m128i)(W), \
+                                         (__mmask8)(U)); })
 
-#define _mm_maskz_srai_epi16( __U, __A, __imm) __extension__ ({ \
-__builtin_ia32_psrawi128_mask ((__v8hi)( __A),( __imm),\
-              (__v8hi)\
-              _mm_setzero_si128 (),\
-              (__mmask8)( __U));\
-})
+#define _mm_maskz_srai_epi16(U, A, imm) __extension__ ({ \
+  (__m128i)__builtin_ia32_psrawi128_mask((__v8hi)(__m128i)(A), (int)(imm), \
+                                         (__v8hi)_mm_setzero_si128(), \
+                                         (__mmask8)(U)); })
 
-#define _mm256_mask_srai_epi16( __W, __U, __A, __imm) __extension__ ({ \
-__builtin_ia32_psrawi256_mask ((__v16hi)( __A),( __imm),\
-              (__v16hi)( __W),\
-              (__mmask16)( __U));\
-})
+#define _mm256_mask_srai_epi16(W, U, A, imm) __extension__ ({ \
+  (__m256i)__builtin_ia32_psrawi256_mask((__v16hi)(__m256i)(A), (int)(imm), \
+                                         (__v16hi)(__m256i)(W), \
+                                         (__mmask16)(U)); })
 
-#define _mm256_maskz_srai_epi16( __U, __A, __imm) __extension__ ({ \
-__builtin_ia32_psrawi256_mask ((__v16hi)( __A),( __imm),\
-              (__v16hi)\
-              _mm256_setzero_si256 (),\
-              (__mmask16)( __U));\
-})
+#define _mm256_maskz_srai_epi16(U, A, imm) __extension__ ({ \
+  (__m256i)__builtin_ia32_psrawi256_mask((__v16hi)(__m256i)(A), (int)(imm), \
+                                         (__v16hi)_mm256_setzero_si256(), \
+                                         (__mmask16)(U)); })
 
 static __inline__ __m128i __DEFAULT_FN_ATTRS
 _mm_mask_srl_epi16 (__m128i __W, __mmask8 __U, __m128i __A,
@@ -2821,31 +2793,25 @@ _mm256_maskz_srl_epi16 (__mmask16 __U, __m256i __A, __m128i __B)
              (__mmask16) __U);
 }
 
-#define _mm_mask_srli_epi16( __W, __U, __A, __imm) __extension__ ({ \
-__builtin_ia32_psrlwi128_mask ((__v8hi)( __A),( __imm),\
-              (__v8hi)( __W),\
-              (__mmask8)( __U));\
-})
+#define _mm_mask_srli_epi16(W, U, A, imm) __extension__ ({ \
+  (__m128i)__builtin_ia32_psrlwi128_mask((__v8hi)(__m128i)(A), (int)(imm), \
+                                         (__v8hi)(__m128i)(W), \
+                                         (__mmask8)(U)); })
 
-#define _mm_maskz_srli_epi16( __U, __A, __imm) __extension__ ({ \
-__builtin_ia32_psrlwi128_mask ((__v8hi)( __A),( __imm),\
-              (__v8hi)\
-              _mm_setzero_si128 (),\
-              (__mmask8)( __U));\
-})
+#define _mm_maskz_srli_epi16(U, A, imm) __extension__ ({ \
+  (__m128i)__builtin_ia32_psrlwi128_mask((__v8hi)(__m128i)(A), (int)(imm), \
+                                         (__v8hi)_mm_setzero_si128(), \
+                                         (__mmask8)(U)); })
 
-#define _mm256_mask_srli_epi16( __W, __U, __A, __imm) __extension__ ({ \
-__builtin_ia32_psrlwi256_mask ((__v16hi)( __A),( __imm),\
-              (__v16hi)( __W),\
-              (__mmask16)( __U));\
-})
+#define _mm256_mask_srli_epi16(W, U, A, imm) __extension__ ({ \
+  (__m256i)__builtin_ia32_psrlwi256_mask((__v16hi)(__m256i)(A), (int)(imm), \
+                                         (__v16hi)(__m256i)(W), \
+                                         (__mmask16)(U)); })
 
-#define _mm256_maskz_srli_epi16( __U, __A, __imm) __extension__ ({ \
-__builtin_ia32_psrlwi256_mask ((__v16hi)( __A),( __imm),\
-              (__v16hi)\
-              _mm256_setzero_si256 (),\
-              (__mmask16)( __U));\
-})
+#define _mm256_maskz_srli_epi16(U, A, imm) __extension__ ({ \
+  (__m256i)__builtin_ia32_psrlwi256_mask((__v16hi)(__m256i)(A), (int)(imm), \
+                                         (__v16hi)_mm256_setzero_si256(), \
+                                         (__mmask16)(U)); })
 
 static __inline__ __m128i __DEFAULT_FN_ATTRS
 _mm_mask_mov_epi16 (__m128i __W, __mmask8 __U, __m128i __A)
@@ -3370,87 +3336,65 @@ _mm256_mask_permutexvar_epi16 (__m256i __W, __mmask16 __M, __m256i __A,
                  (__mmask16) __M);
 }
 
-#define _mm_mask_alignr_epi8( __W, __U, __A, __B, __N) __extension__ ({ \
-__builtin_ia32_palignr128_mask ((__v2di)( __A),\
-               (__v2di)( __B),\
-               ( __N) * 8,\
-               (__v2di)( __W),\
-               (__mmask16)( __U));\
-})
-
-#define _mm_maskz_alignr_epi8( __U, __A, __B, __N) __extension__ ({ \
-__builtin_ia32_palignr128_mask ((__v2di)( __A),\
-               (__v2di)( __B),\
-               ( __N) * 8,\
-               (__v2di)\
-               _mm_setzero_si128 (),\
-               (__mmask16)( __U));\
-})
-
-#define _mm256_mask_alignr_epi8( __W, __U, __A, __B, __N) __extension__ ({ \
-__builtin_ia32_palignr256_mask ((__v4di)( __A),\
-               (__v4di)( __B),\
-               ( __N) * 8,\
-               (__v4di)( __W),\
-               (__mmask32)( __U));\
-})
-
-#define _mm256_maskz_alignr_epi8( __U, __A, __B, __N) __extension__ ({ \
-__builtin_ia32_palignr256_mask ((__v4di)( __A),\
-               (__v4di)( __B),\
-               ( __N) * 8,\
-               (__v4di)\
-               _mm256_setzero_si256 (),\
-               (__mmask32)( __U));\
-})
-
-#define _mm_dbsad_epu8( __A, __B, __imm) __extension__ ({ \
-__builtin_ia32_dbpsadbw128_mask ((__v16qi)( __A),\
-                (__v16qi)( __B),\
-                ( __imm),\
-                (__v8hi) _mm_setzero_hi (),\
-                (__mmask8) -1);\
-})
-
-#define _mm_mask_dbsad_epu8( __W, __U, __A, __B, __imm) __extension__ ({ \
-__builtin_ia32_dbpsadbw128_mask ((__v16qi)( __A),\
-                (__v16qi)( __B),\
-                ( __imm),\
-                (__v8hi)( __W),\
-                (__mmask8)( __U));\
-})
-
-#define _mm_maskz_dbsad_epu8( __U, __A, __B, __imm) __extension__ ({ \
-__builtin_ia32_dbpsadbw128_mask ((__v16qi)( __A),\
-                (__v16qi)( __B),\
-                ( __imm),\
-                (__v8hi) _mm_setzero_si128 (),\
-                (__mmask8)( __U));\
-})
-
-#define _mm256_dbsad_epu8( __A, __B, __imm) __extension__ ({ \
-__builtin_ia32_dbpsadbw256_mask ((__v32qi)( __A),\
-                (__v32qi)( __B),\
-                ( __imm),\
-                (__v16hi) _mm256_setzero_si256 (),\
-                (__mmask16) -1);\
-})
-
-#define _mm256_mask_dbsad_epu8( __W, __U, __A, __B, __imm) __extension__ ({ \
-__builtin_ia32_dbpsadbw256_mask ((__v32qi)( __A),\
-                (__v32qi)( __B),\
-                ( __imm),\
-                (__v16hi)( __W),\
-                (__mmask16)( __U));\
-})
-
-#define _mm256_maskz_dbsad_epu8( __U, __A, __B, __imm) __extension__ ({ \
-__builtin_ia32_dbpsadbw256_mask ((__v32qi)( __A),\
-                (__v32qi)( __B),\
-                ( __imm),\
-                (__v16hi) _mm256_setzero_si256 (),\
-                (__mmask16)( __U));\
-})
+#define _mm_mask_alignr_epi8(W, U, A, B, N) __extension__ ({ \
+  (__m128i)__builtin_ia32_palignr128_mask((__v16qi)(__m128i)(A), \
+                                          (__v16qi)(__m128i)(B), (int)(N) * 8, \
+                                          (__v16qi)(__m128i)(W), \
+                                          (__mmask16)(U)); })
+
+#define _mm_maskz_alignr_epi8(U, A, B, N) __extension__ ({ \
+  (__m128i)__builtin_ia32_palignr128_mask((__v16qi)(__m128i)(A), \
+                                          (__v16qi)(__m128i)(B), (int)(N) * 8, \
+                                          (__v16qi)_mm_setzero_si128(), \
+                                          (__mmask16)(U)); })
+
+#define _mm256_mask_alignr_epi8(W, U, A, B, N) __extension__ ({ \
+  (__m256i)__builtin_ia32_palignr256_mask((__v32qi)(__m256i)(A), \
+                                          (__v32qi)(__m256i)(B), (int)(N) * 8, \
+                                          (__v32qi)(__m256i)(W), \
+                                          (__mmask32)(U)); })
+
+#define _mm256_maskz_alignr_epi8(U, A, B, N) __extension__ ({ \
+  (__m256i)__builtin_ia32_palignr256_mask((__v32qi)(__m256i)(A), \
+                                          (__v32qi)(__m256i)(B), (int)(N) * 8, \
+                                          (__v32qi)_mm256_setzero_si256(), \
+                                          (__mmask32)(U)); })
+
+#define _mm_dbsad_epu8(A, B, imm) __extension__ ({ \
+  (__m128i)__builtin_ia32_dbpsadbw128_mask((__v16qi)(__m128i)(A), \
+                                           (__v16qi)(__m128i)(B), (int)(imm), \
+                                           (__v8hi)_mm_setzero_hi(), \
+                                           (__mmask8)-1); })
+
+#define _mm_mask_dbsad_epu8(W, U, A, B, imm) __extension__ ({ \
+  (__m128i)__builtin_ia32_dbpsadbw128_mask((__v16qi)(__m128i)(A), \
+                                           (__v16qi)(__m128i)(B), (int)(imm), \
+                                           (__v8hi)(__m128i)(W), \
+                                           (__mmask8)(U)); })
+
+#define _mm_maskz_dbsad_epu8(U, A, B, imm) __extension__ ({ \
+  (__m128i)__builtin_ia32_dbpsadbw128_mask((__v16qi)(__m128i)(A), \
+                                           (__v16qi)(__m128i)(B), (int)(imm), \
+                                           (__v8hi)_mm_setzero_si128(), \
+                                           (__mmask8)(U)); })
+
+#define _mm256_dbsad_epu8(A, B, imm) __extension__ ({ \
+  (__m256i)__builtin_ia32_dbpsadbw256_mask((__v32qi)(__m256i)(A), \
+                                           (__v32qi)(__m256i)(B), (int)(imm), \
+                                           (__v16hi)_mm256_setzero_si256(), \
+                                           (__mmask16)-1); })
+
+#define _mm256_mask_dbsad_epu8(W, U, A, B, imm) __extension__ ({ \
+  (__m256i)__builtin_ia32_dbpsadbw256_mask((__v32qi)(__m256i)(A), \
+                                           (__v32qi)(__m256i)(B), (int)(imm), \
+                                           (__v16hi)(__m256i)(W), \
+                                           (__mmask16)(U)); })
+
+#define _mm256_maskz_dbsad_epu8(U, A, B, imm) __extension__ ({ \
+  (__m256i)__builtin_ia32_dbpsadbw256_mask((__v32qi)(__m256i)(A), \
+                                           (__v32qi)(__m256i)(B), (int)(imm), \
+                                           (__v16hi)_mm256_setzero_si256(), \
+                                           (__mmask16)(U)); })
 
 #undef __DEFAULT_FN_ATTRS