[Headers][X86] Group related AVX512VL FMA intrinsics together (NFC) (#156794)

Follow-up of #156385.

Author: ckoparkar

clang/lib/Headers/avx512vlintrin.h

@@ -925,13 +925,25 @@ _mm_mask_fmsub_pd(__m128d __A, __mmask8 __U, __m128d __B, __m128d __C) {
       (__mmask8)__U, (__v2df)_mm_fmsub_pd(__A, __B, __C), (__v2df)__A);
 }
 
+static __inline__ __m128d __DEFAULT_FN_ATTRS128_CONSTEXPR
+_mm_mask3_fmsub_pd(__m128d __A, __m128d __B, __m128d __C, __mmask8 __U) {
+  return (__m128d)__builtin_ia32_selectpd_128(
+      (__mmask8)__U, (__v2df)_mm_fmsub_pd(__A, __B, __C), (__v2df)__C);
+}
+
 static __inline__ __m128d __DEFAULT_FN_ATTRS128_CONSTEXPR
 _mm_maskz_fmsub_pd(__mmask8 __U, __m128d __A, __m128d __B, __m128d __C) {
   return (__m128d)__builtin_ia32_selectpd_128(
       (__mmask8)__U, (__v2df)_mm_fmsub_pd(__A, __B, __C),
       (__v2df)_mm_setzero_pd());
 }
 
+static __inline__ __m128d __DEFAULT_FN_ATTRS128_CONSTEXPR
+_mm_mask_fnmadd_pd(__m128d __A, __mmask8 __U, __m128d __B, __m128d __C) {
+  return (__m128d)__builtin_ia32_selectpd_128(
+      (__mmask8)__U, (__v2df)_mm_fnmadd_pd(__A, __B, __C), (__v2df)__A);
+}
+
 static __inline__ __m128d __DEFAULT_FN_ATTRS128_CONSTEXPR
 _mm_mask3_fnmadd_pd(__m128d __A, __m128d __B, __m128d __C, __mmask8 __U) {
   return (__m128d)__builtin_ia32_selectpd_128(
@@ -945,6 +957,18 @@ _mm_maskz_fnmadd_pd(__mmask8 __U, __m128d __A, __m128d __B, __m128d __C) {
       (__v2df)_mm_setzero_pd());
 }
 
+static __inline__ __m128d __DEFAULT_FN_ATTRS128_CONSTEXPR
+_mm_mask_fnmsub_pd(__m128d __A, __mmask8 __U, __m128d __B, __m128d __C) {
+  return (__m128d)__builtin_ia32_selectpd_128(
+      (__mmask8)__U, (__v2df)_mm_fnmsub_pd(__A, __B, __C), (__v2df)__A);
+}
+
+static __inline__ __m128d __DEFAULT_FN_ATTRS128_CONSTEXPR
+_mm_mask3_fnmsub_pd(__m128d __A, __m128d __B, __m128d __C, __mmask8 __U) {
+  return (__m128d)__builtin_ia32_selectpd_128(
+      (__mmask8)__U, (__v2df)_mm_fnmsub_pd(__A, __B, __C), (__v2df)__C);
+}
+
 static __inline__ __m128d __DEFAULT_FN_ATTRS128_CONSTEXPR
 _mm_maskz_fnmsub_pd(__mmask8 __U, __m128d __A, __m128d __B, __m128d __C) {
   return (__m128d)__builtin_ia32_selectpd_128(
@@ -977,13 +1001,25 @@ _mm256_mask_fmsub_pd(__m256d __A, __mmask8 __U, __m256d __B, __m256d __C) {
       (__mmask8)__U, (__v4df)_mm256_fmsub_pd(__A, __B, __C), (__v4df)__A);
 }
 
+static __inline__ __m256d __DEFAULT_FN_ATTRS256_CONSTEXPR
+_mm256_mask3_fmsub_pd(__m256d __A, __m256d __B, __m256d __C, __mmask8 __U) {
+  return (__m256d)__builtin_ia32_selectpd_256(
+      (__mmask8)__U, (__v4df)_mm256_fmsub_pd(__A, __B, __C), (__v4df)__C);
+}
+
 static __inline__ __m256d __DEFAULT_FN_ATTRS256_CONSTEXPR
 _mm256_maskz_fmsub_pd(__mmask8 __U, __m256d __A, __m256d __B, __m256d __C) {
   return (__m256d)__builtin_ia32_selectpd_256(
       (__mmask8)__U, (__v4df)_mm256_fmsub_pd(__A, __B, __C),
       (__v4df)_mm256_setzero_pd());
 }
 
+static __inline__ __m256d __DEFAULT_FN_ATTRS256_CONSTEXPR
+_mm256_mask_fnmadd_pd(__m256d __A, __mmask8 __U, __m256d __B, __m256d __C) {
+  return (__m256d)__builtin_ia32_selectpd_256(
+      (__mmask8)__U, (__v4df)_mm256_fnmadd_pd(__A, __B, __C), (__v4df)__A);
+}
+
 static __inline__ __m256d __DEFAULT_FN_ATTRS256_CONSTEXPR
 _mm256_mask3_fnmadd_pd(__m256d __A, __m256d __B, __m256d __C, __mmask8 __U) {
   return (__m256d)__builtin_ia32_selectpd_256(
@@ -997,6 +1033,18 @@ _mm256_maskz_fnmadd_pd(__mmask8 __U, __m256d __A, __m256d __B, __m256d __C) {
       (__v4df)_mm256_setzero_pd());
 }
 
+static __inline__ __m256d __DEFAULT_FN_ATTRS256_CONSTEXPR
+_mm256_mask_fnmsub_pd(__m256d __A, __mmask8 __U, __m256d __B, __m256d __C) {
+  return (__m256d)__builtin_ia32_selectpd_256(
+      (__mmask8)__U, (__v4df)_mm256_fnmsub_pd(__A, __B, __C), (__v4df)__A);
+}
+
+static __inline__ __m256d __DEFAULT_FN_ATTRS256_CONSTEXPR
+_mm256_mask3_fnmsub_pd(__m256d __A, __m256d __B, __m256d __C, __mmask8 __U) {
+  return (__m256d)__builtin_ia32_selectpd_256(
+      (__mmask8)__U, (__v4df)_mm256_fnmsub_pd(__A, __B, __C), (__v4df)__C);
+}
+
 static __inline__ __m256d __DEFAULT_FN_ATTRS256_CONSTEXPR
 _mm256_maskz_fnmsub_pd(__mmask8 __U, __m256d __A, __m256d __B, __m256d __C) {
   return (__m256d)__builtin_ia32_selectpd_256(
@@ -1029,13 +1077,25 @@ _mm_mask_fmsub_ps(__m128 __A, __mmask8 __U, __m128 __B, __m128 __C) {
       (__mmask8)__U, (__v4sf)_mm_fmsub_ps(__A, __B, __C), (__v4sf)__A);
 }
 
+static __inline__ __m128 __DEFAULT_FN_ATTRS128_CONSTEXPR
+_mm_mask3_fmsub_ps(__m128 __A, __m128 __B, __m128 __C, __mmask8 __U) {
+  return (__m128)__builtin_ia32_selectps_128(
+      (__mmask8)__U, (__v4sf)_mm_fmsub_ps(__A, __B, __C), (__v4sf)__C);
+}
+
 static __inline__ __m128 __DEFAULT_FN_ATTRS128_CONSTEXPR
 _mm_maskz_fmsub_ps(__mmask8 __U, __m128 __A, __m128 __B, __m128 __C) {
   return (__m128)__builtin_ia32_selectps_128(
       (__mmask8)__U, (__v4sf)_mm_fmsub_ps(__A, __B, __C),
       (__v4sf)_mm_setzero_ps());
 }
 
+static __inline__ __m128 __DEFAULT_FN_ATTRS128_CONSTEXPR
+_mm_mask_fnmadd_ps(__m128 __A, __mmask8 __U, __m128 __B, __m128 __C) {
+  return (__m128)__builtin_ia32_selectps_128(
+      (__mmask8)__U, (__v4sf)_mm_fnmadd_ps(__A, __B, __C), (__v4sf)__A);
+}
+
 static __inline__ __m128 __DEFAULT_FN_ATTRS128_CONSTEXPR
 _mm_mask3_fnmadd_ps(__m128 __A, __m128 __B, __m128 __C, __mmask8 __U) {
   return (__m128)__builtin_ia32_selectps_128(
@@ -1049,6 +1109,18 @@ _mm_maskz_fnmadd_ps(__mmask8 __U, __m128 __A, __m128 __B, __m128 __C) {
       (__v4sf)_mm_setzero_ps());
 }
 
+static __inline__ __m128 __DEFAULT_FN_ATTRS128_CONSTEXPR
+_mm_mask_fnmsub_ps(__m128 __A, __mmask8 __U, __m128 __B, __m128 __C) {
+  return (__m128)__builtin_ia32_selectps_128(
+      (__mmask8)__U, (__v4sf)_mm_fnmsub_ps(__A, __B, __C), (__v4sf)__A);
+}
+
+static __inline__ __m128 __DEFAULT_FN_ATTRS128_CONSTEXPR
+_mm_mask3_fnmsub_ps(__m128 __A, __m128 __B, __m128 __C, __mmask8 __U) {
+  return (__m128)__builtin_ia32_selectps_128(
+      (__mmask8)__U, (__v4sf)_mm_fnmsub_ps(__A, __B, __C), (__v4sf)__C);
+}
+
 static __inline__ __m128 __DEFAULT_FN_ATTRS128_CONSTEXPR
 _mm_maskz_fnmsub_ps(__mmask8 __U, __m128 __A, __m128 __B, __m128 __C) {
   return (__m128)__builtin_ia32_selectps_128(
@@ -1081,13 +1153,25 @@ _mm256_mask_fmsub_ps(__m256 __A, __mmask8 __U, __m256 __B, __m256 __C) {
       (__mmask8)__U, (__v8sf)_mm256_fmsub_ps(__A, __B, __C), (__v8sf)__A);
 }
 
+static __inline__ __m256 __DEFAULT_FN_ATTRS256_CONSTEXPR
+_mm256_mask3_fmsub_ps(__m256 __A, __m256 __B, __m256 __C, __mmask8 __U) {
+  return (__m256)__builtin_ia32_selectps_256(
+      (__mmask8)__U, (__v8sf)_mm256_fmsub_ps(__A, __B, __C), (__v8sf)__C);
+}
+
 static __inline__ __m256 __DEFAULT_FN_ATTRS256_CONSTEXPR
 _mm256_maskz_fmsub_ps(__mmask8 __U, __m256 __A, __m256 __B, __m256 __C) {
   return (__m256)__builtin_ia32_selectps_256(
       (__mmask8)__U, (__v8sf)_mm256_fmsub_ps(__A, __B, __C),
       (__v8sf)_mm256_setzero_ps());
 }
 
+static __inline__ __m256 __DEFAULT_FN_ATTRS256_CONSTEXPR
+_mm256_mask_fnmadd_ps(__m256 __A, __mmask8 __U, __m256 __B, __m256 __C) {
+  return (__m256)__builtin_ia32_selectps_256(
+      (__mmask8)__U, (__v8sf)_mm256_fnmadd_ps(__A, __B, __C), (__v8sf)__A);
+}
+
 static __inline__ __m256 __DEFAULT_FN_ATTRS256_CONSTEXPR
 _mm256_mask3_fnmadd_ps(__m256 __A, __m256 __B, __m256 __C, __mmask8 __U) {
   return (__m256)__builtin_ia32_selectps_256(
@@ -1101,6 +1185,18 @@ _mm256_maskz_fnmadd_ps(__mmask8 __U, __m256 __A, __m256 __B, __m256 __C) {
       (__v8sf)_mm256_setzero_ps());
 }
 
+static __inline__ __m256 __DEFAULT_FN_ATTRS256_CONSTEXPR
+_mm256_mask_fnmsub_ps(__m256 __A, __mmask8 __U, __m256 __B, __m256 __C) {
+  return (__m256)__builtin_ia32_selectps_256(
+      (__mmask8)__U, (__v8sf)_mm256_fnmsub_ps(__A, __B, __C), (__v8sf)__A);
+}
+
+static __inline__ __m256 __DEFAULT_FN_ATTRS256_CONSTEXPR
+_mm256_mask3_fnmsub_ps(__m256 __A, __m256 __B, __m256 __C, __mmask8 __U) {
+  return (__m256)__builtin_ia32_selectps_256(
+      (__mmask8)__U, (__v8sf)_mm256_fnmsub_ps(__A, __B, __C), (__v8sf)__C);
+}
+
 static __inline__ __m256 __DEFAULT_FN_ATTRS256_CONSTEXPR
 _mm256_maskz_fnmsub_ps(__mmask8 __U, __m256 __A, __m256 __B, __m256 __C) {
   return (__m256)__builtin_ia32_selectps_256(
@@ -1308,30 +1404,6 @@ _mm256_maskz_fmsubadd_ps(__mmask8 __U, __m256 __A, __m256 __B, __m256 __C)
                     (__v8sf)_mm256_setzero_ps());
 }
 
-static __inline__ __m128d __DEFAULT_FN_ATTRS128_CONSTEXPR
-_mm_mask3_fmsub_pd(__m128d __A, __m128d __B, __m128d __C, __mmask8 __U) {
-  return (__m128d)__builtin_ia32_selectpd_128(
-      (__mmask8)__U, (__v2df)_mm_fmsub_pd(__A, __B, __C), (__v2df)__C);
-}
-
-static __inline__ __m256d __DEFAULT_FN_ATTRS256_CONSTEXPR
-_mm256_mask3_fmsub_pd(__m256d __A, __m256d __B, __m256d __C, __mmask8 __U) {
-  return (__m256d)__builtin_ia32_selectpd_256(
-      (__mmask8)__U, (__v4df)_mm256_fmsub_pd(__A, __B, __C), (__v4df)__C);
-}
-
-static __inline__ __m128 __DEFAULT_FN_ATTRS128_CONSTEXPR
-_mm_mask3_fmsub_ps(__m128 __A, __m128 __B, __m128 __C, __mmask8 __U) {
-  return (__m128)__builtin_ia32_selectps_128(
-      (__mmask8)__U, (__v4sf)_mm_fmsub_ps(__A, __B, __C), (__v4sf)__C);
-}
-
-static __inline__ __m256 __DEFAULT_FN_ATTRS256_CONSTEXPR
-_mm256_mask3_fmsub_ps(__m256 __A, __m256 __B, __m256 __C, __mmask8 __U) {
-  return (__m256)__builtin_ia32_selectps_256(
-      (__mmask8)__U, (__v8sf)_mm256_fmsub_ps(__A, __B, __C), (__v8sf)__C);
-}
-
 static __inline__ __m128d __DEFAULT_FN_ATTRS128
 _mm_mask3_fmsubadd_pd(__m128d __A, __m128d __B, __m128d __C, __mmask8 __U)
 {
@@ -1372,78 +1444,6 @@ _mm256_mask3_fmsubadd_ps(__m256 __A, __m256 __B, __m256 __C, __mmask8 __U)
                     (__v8sf) __C);
 }
 
-static __inline__ __m128d __DEFAULT_FN_ATTRS128_CONSTEXPR
-_mm_mask_fnmadd_pd(__m128d __A, __mmask8 __U, __m128d __B, __m128d __C) {
-  return (__m128d)__builtin_ia32_selectpd_128(
-      (__mmask8)__U, (__v2df)_mm_fnmadd_pd(__A, __B, __C), (__v2df)__A);
-}
-
-static __inline__ __m256d __DEFAULT_FN_ATTRS256_CONSTEXPR
-_mm256_mask_fnmadd_pd(__m256d __A, __mmask8 __U, __m256d __B, __m256d __C) {
-  return (__m256d)__builtin_ia32_selectpd_256(
-      (__mmask8)__U, (__v4df)_mm256_fnmadd_pd(__A, __B, __C), (__v4df)__A);
-}
-
-static __inline__ __m128 __DEFAULT_FN_ATTRS128_CONSTEXPR
-_mm_mask_fnmadd_ps(__m128 __A, __mmask8 __U, __m128 __B, __m128 __C) {
-  return (__m128)__builtin_ia32_selectps_128(
-      (__mmask8)__U, (__v4sf)_mm_fnmadd_ps(__A, __B, __C), (__v4sf)__A);
-}
-
-static __inline__ __m256 __DEFAULT_FN_ATTRS256_CONSTEXPR
-_mm256_mask_fnmadd_ps(__m256 __A, __mmask8 __U, __m256 __B, __m256 __C) {
-  return (__m256)__builtin_ia32_selectps_256(
-      (__mmask8)__U, (__v8sf)_mm256_fnmadd_ps(__A, __B, __C), (__v8sf)__A);
-}
-
-static __inline__ __m128d __DEFAULT_FN_ATTRS128_CONSTEXPR
-_mm_mask_fnmsub_pd(__m128d __A, __mmask8 __U, __m128d __B, __m128d __C) {
-  return (__m128d)__builtin_ia32_selectpd_128(
-      (__mmask8)__U, (__v2df)_mm_fnmsub_pd(__A, __B, __C), (__v2df)__A);
-}
-
-static __inline__ __m128d __DEFAULT_FN_ATTRS128_CONSTEXPR
-_mm_mask3_fnmsub_pd(__m128d __A, __m128d __B, __m128d __C, __mmask8 __U) {
-  return (__m128d)__builtin_ia32_selectpd_128(
-      (__mmask8)__U, (__v2df)_mm_fnmsub_pd(__A, __B, __C), (__v2df)__C);
-}
-
-static __inline__ __m256d __DEFAULT_FN_ATTRS256_CONSTEXPR
-_mm256_mask_fnmsub_pd(__m256d __A, __mmask8 __U, __m256d __B, __m256d __C) {
-  return (__m256d)__builtin_ia32_selectpd_256(
-      (__mmask8)__U, (__v4df)_mm256_fnmsub_pd(__A, __B, __C), (__v4df)__A);
-}
-
-static __inline__ __m256d __DEFAULT_FN_ATTRS256_CONSTEXPR
-_mm256_mask3_fnmsub_pd(__m256d __A, __m256d __B, __m256d __C, __mmask8 __U) {
-  return (__m256d)__builtin_ia32_selectpd_256(
-      (__mmask8)__U, (__v4df)_mm256_fnmsub_pd(__A, __B, __C), (__v4df)__C);
-}
-
-static __inline__ __m128 __DEFAULT_FN_ATTRS128_CONSTEXPR
-_mm_mask_fnmsub_ps(__m128 __A, __mmask8 __U, __m128 __B, __m128 __C) {
-  return (__m128)__builtin_ia32_selectps_128(
-      (__mmask8)__U, (__v4sf)_mm_fnmsub_ps(__A, __B, __C), (__v4sf)__A);
-}
-
-static __inline__ __m128 __DEFAULT_FN_ATTRS128_CONSTEXPR
-_mm_mask3_fnmsub_ps(__m128 __A, __m128 __B, __m128 __C, __mmask8 __U) {
-  return (__m128)__builtin_ia32_selectps_128(
-      (__mmask8)__U, (__v4sf)_mm_fnmsub_ps(__A, __B, __C), (__v4sf)__C);
-}
-
-static __inline__ __m256 __DEFAULT_FN_ATTRS256_CONSTEXPR
-_mm256_mask_fnmsub_ps(__m256 __A, __mmask8 __U, __m256 __B, __m256 __C) {
-  return (__m256)__builtin_ia32_selectps_256(
-      (__mmask8)__U, (__v8sf)_mm256_fnmsub_ps(__A, __B, __C), (__v8sf)__A);
-}
-
-static __inline__ __m256 __DEFAULT_FN_ATTRS256_CONSTEXPR
-_mm256_mask3_fnmsub_ps(__m256 __A, __m256 __B, __m256 __C, __mmask8 __U) {
-  return (__m256)__builtin_ia32_selectps_256(
-      (__mmask8)__U, (__v8sf)_mm256_fnmsub_ps(__A, __B, __C), (__v8sf)__C);
-}
-
 static __inline__ __m128d __DEFAULT_FN_ATTRS128
 _mm_mask_add_pd(__m128d __W, __mmask8 __U, __m128d __A, __m128d __B) {
   return (__m128d)__builtin_ia32_selectpd_128((__mmask8)__U,