From: Andrei Tatar <andrei@unikraft.io>
Date: Mon, 17 Jul 2023 21:14:01 +0000 (+0200)
Subject: Add compatibility with clang <= 11
X-Git-Tag: RELEASE-0.14.0~3
X-Git-Url: http://xenbits.xensource.com/gitweb?a=commitdiff_plain;h=8964455b8ac46977168d58e0cf3c7d696aebb95c;p=unikraft%2Flibs%2Fintel-intrinsics.git

Add compatibility with clang <= 11

Implementations based off headers in LLVM 11.0.0.

Signed-off-by: Andrei Tatar <andrei@unikraft.io>
Reviewed-by: Maria Sfiraiala <maria.sfiraiala@gmail.com>
Reviewed-by: Radu Nichita <radunichita99@gmail.com>
Approved-by: Razvan Deaconescu <razvand@unikraft.io>
Tested-by: Unikraft CI <monkey@unikraft.io>
GitHub-Closes: #3
---

diff --git a/include-llvm/amxintrin.h b/include-llvm/amxintrin.h
index 51948fc..aaf82fb 100644
--- a/include-llvm/amxintrin.h
+++ b/include-llvm/amxintrin.h
@@ -226,6 +226,7 @@ static __inline__ void __DEFAULT_FN_ATTRS_TILE _tile_release(void) {
 #define _tile_dpbf16ps(dst, src0, src1)                                        \
   __builtin_ia32_tdpbf16ps((dst), (src0), (src1))
 
+#if (__clang_major__ >= 12)
 /// AMX tile register size can be configured, the maximum size is 16x64=1024
 /// bytes. Since there is no 2D type in llvm IR, we use vector type to
 /// represent 2D tile and the fixed size is maximum amx tile register size.
@@ -496,6 +497,7 @@ static __inline__ void __tile_dpbf16ps(__tile1024i *dst, __tile1024i src0,
                                       src0.tile, src1.tile);
 }
 #endif
+#endif
 
 #undef __DEFAULT_FN_ATTRS_TILE
 #undef __DEFAULT_FN_ATTRS_INT8
diff --git a/include-llvm/avx512fintrin.h b/include-llvm/avx512fintrin.h
index 3ffe999..0bc512d 100644
--- a/include-llvm/avx512fintrin.h
+++ b/include-llvm/avx512fintrin.h
@@ -9355,35 +9355,56 @@ _mm512_mask_abs_pd(__m512d __W, __mmask8 __K, __m512d __A)
  * This takes log2(n) steps where n is the number of elements in the vector.
  */
 
+#if (__clang_major__ <= 11)
+#define _mm512_mask_reduce_operator(op) \
+  __v4du __t1 = (__v4du)_mm512_extracti64x4_epi64(__W, 0); \
+  __v4du __t2 = (__v4du)_mm512_extracti64x4_epi64(__W, 1); \
+  __m256i __t3 = (__m256i)(__t1 op __t2); \
+  __v2du __t4 = (__v2du)_mm256_extracti128_si256(__t3, 0); \
+  __v2du __t5 = (__v2du)_mm256_extracti128_si256(__t3, 1); \
+  __v2du __t6 = __t4 op __t5; \
+  __v2du __t7 = __builtin_shufflevector(__t6, __t6, 1, 0); \
+  __v2du __t8 = __t6 op __t7; \
+  return __t8[0]
+#endif
+
 static __inline__ long long __DEFAULT_FN_ATTRS512 _mm512_reduce_add_epi64(__m512i __W) {
 #if (__clang_major__ > 14)
   return __builtin_reduce_add((__v8di)__W);
-#else
+#elif (__clang_major__ > 11)
   return __builtin_ia32_reduce_add_q512(__W);
+#else
+  _mm512_mask_reduce_operator(+);
 #endif
 }
 
 static __inline__ long long __DEFAULT_FN_ATTRS512 _mm512_reduce_mul_epi64(__m512i __W) {
 #if (__clang_major__ > 14)
   return __builtin_reduce_mul((__v8di)__W);
-#else
+#elif (__clang_major__ > 11)
   return __builtin_ia32_reduce_mul_q512(__W);
+#else
+  _mm512_mask_reduce_operator(*);
 #endif
 }
 
 static __inline__ long long __DEFAULT_FN_ATTRS512 _mm512_reduce_and_epi64(__m512i __W) {
-#if (__clang_major__ < 14)
+#if (__clang_major__ > 13)
+  return __builtin_reduce_and((__v8di)__W);
+#elif (__clang_major__ > 11)
   return __builtin_ia32_reduce_and_q512(__W);
 #else
-  return __builtin_reduce_and((__v8di)__W);
+  _mm512_mask_reduce_operator(&);
 #endif
 }
 
 static __inline__ long long __DEFAULT_FN_ATTRS512 _mm512_reduce_or_epi64(__m512i __W) {
-#if (__clang_major__ < 14)
+#if (__clang_major__ > 13)
+  return __builtin_reduce_or((__v8di)__W);
+#elif (__clang_major__ > 11)
   return __builtin_ia32_reduce_or_q512(__W);
 #else
-  return __builtin_reduce_or((__v8di)__W);
+  _mm512_mask_reduce_operator(|);
 #endif
 }
 
@@ -9392,8 +9413,10 @@ _mm512_mask_reduce_add_epi64(__mmask8 __M, __m512i __W) {
   __W = _mm512_maskz_mov_epi64(__M, __W);
 #if (__clang_major__ > 14)
   return __builtin_reduce_add((__v8di)__W);
-#else
+#elif (__clang_major__ > 11)
   return __builtin_ia32_reduce_add_q512(__W);
+#else
+  _mm512_mask_reduce_operator(+);
 #endif
 }
 
@@ -9402,60 +9425,114 @@ _mm512_mask_reduce_mul_epi64(__mmask8 __M, __m512i __W) {
   __W = _mm512_mask_mov_epi64(_mm512_set1_epi64(1), __M, __W);
 #if (__clang_major__ > 14)
   return __builtin_reduce_mul((__v8di)__W);
-#else
+#elif (__clang_major__ > 11)
   return __builtin_ia32_reduce_mul_q512(__W);
+#else
+  _mm512_mask_reduce_operator(*);
 #endif
 }
 
 static __inline__ long long __DEFAULT_FN_ATTRS512
 _mm512_mask_reduce_and_epi64(__mmask8 __M, __m512i __W) {
   __W = _mm512_mask_mov_epi64(_mm512_set1_epi64(~0ULL), __M, __W);
-#if (__clang_major__ < 14)
+#if (__clang_major__ > 13)
+  return __builtin_reduce_and((__v8di)__W);
+#elif (__clang_major__ > 11)
   return __builtin_ia32_reduce_and_q512(__W);
 #else
-  return __builtin_reduce_and((__v8di)__W);
+  _mm512_mask_reduce_operator(&);
 #endif
 }
 
 static __inline__ long long __DEFAULT_FN_ATTRS512
 _mm512_mask_reduce_or_epi64(__mmask8 __M, __m512i __W) {
   __W = _mm512_maskz_mov_epi64(__M, __W);
-#if (__clang_major__ < 14)
+#if (__clang_major__ > 13)
+  return __builtin_reduce_or((__v8di)__W);
+#elif (__clang_major__ > 11)
   return __builtin_ia32_reduce_or_q512(__W);
 #else
-  return __builtin_reduce_or((__v8di)__W);
+  _mm512_mask_reduce_operator(|);
 #endif
 }
+#if (__clang_major__ <= 11)
+#undef _mm512_mask_reduce_operator
+
+#define _mm512_mask_reduce_operator(op) \
+  __m256d __t1 = _mm512_extractf64x4_pd(__W, 0); \
+  __m256d __t2 = _mm512_extractf64x4_pd(__W, 1); \
+  __m256d __t3 = __t1 op __t2; \
+  __m128d __t4 = _mm256_extractf128_pd(__t3, 0); \
+  __m128d __t5 = _mm256_extractf128_pd(__t3, 1); \
+  __m128d __t6 = __t4 op __t5; \
+  __m128d __t7 = __builtin_shufflevector(__t6, __t6, 1, 0); \
+  __m128d __t8 = __t6 op __t7; \
+  return __t8[0]
+#endif
 
 // -0.0 is used to ignore the start value since it is the neutral value of
 // floating point addition. For more information, please refer to
 // https://llvm.org/docs/LangRef.html#llvm-vector-reduce-fadd-intrinsic
 static __inline__ double __DEFAULT_FN_ATTRS512 _mm512_reduce_add_pd(__m512d __W) {
+#if (__clang_major__ > 11)
   return __builtin_ia32_reduce_fadd_pd512(-0.0, __W);
+#else
+  _mm512_mask_reduce_operator(+);
+#endif
 }
 
 static __inline__ double __DEFAULT_FN_ATTRS512 _mm512_reduce_mul_pd(__m512d __W) {
+#if (__clang_major__ > 11)
   return __builtin_ia32_reduce_fmul_pd512(1.0, __W);
+#else
+  _mm512_mask_reduce_operator(*);
+#endif
 }
 
 static __inline__ double __DEFAULT_FN_ATTRS512
 _mm512_mask_reduce_add_pd(__mmask8 __M, __m512d __W) {
   __W = _mm512_maskz_mov_pd(__M, __W);
+#if (__clang_major__ > 11)
   return __builtin_ia32_reduce_fadd_pd512(-0.0, __W);
+#else
+  _mm512_mask_reduce_operator(+);
+#endif
 }
 
 static __inline__ double __DEFAULT_FN_ATTRS512
 _mm512_mask_reduce_mul_pd(__mmask8 __M, __m512d __W) {
   __W = _mm512_mask_mov_pd(_mm512_set1_pd(1.0), __M, __W);
+#if (__clang_major__ > 11)
   return __builtin_ia32_reduce_fmul_pd512(1.0, __W);
+#else
+  _mm512_mask_reduce_operator(*);
+#endif
 }
+#if (__clang_major__ <= 11)
+#undef _mm512_mask_reduce_operator
+
+#define _mm512_mask_reduce_operator(op) \
+  __v8su __t1 = (__v8su)_mm512_extracti64x4_epi64(__W, 0); \
+  __v8su __t2 = (__v8su)_mm512_extracti64x4_epi64(__W, 1); \
+  __m256i __t3 = (__m256i)(__t1 op __t2); \
+  __v4su __t4 = (__v4su)_mm256_extracti128_si256(__t3, 0); \
+  __v4su __t5 = (__v4su)_mm256_extracti128_si256(__t3, 1); \
+  __v4su __t6 = __t4 op __t5; \
+  __v4su __t7 = __builtin_shufflevector(__t6, __t6, 2, 3, 0, 1); \
+  __v4su __t8 = __t6 op __t7; \
+  __v4su __t9 = __builtin_shufflevector(__t8, __t8, 1, 0, 3, 2); \
+  __v4su __t10 = __t8 op __t9; \
+  return __t10[0]
+#endif
 
 static __inline__ int __DEFAULT_FN_ATTRS512
 _mm512_reduce_add_epi32(__m512i __W) {
 #if (__clang_major__ > 14)
   return __builtin_reduce_add((__v16si)__W);
-#else
+#elif (__clang_major__ > 11)
   return __builtin_ia32_reduce_add_d512((__v16si)__W);
+#else
+  _mm512_mask_reduce_operator(+);
 #endif
 }
 
@@ -9463,26 +9540,32 @@ static __inline__ int __DEFAULT_FN_ATTRS512
 _mm512_reduce_mul_epi32(__m512i __W) {
 #if (__clang_major__ > 14)
   return __builtin_reduce_mul((__v16si)__W);
-#else
+#elif (__clang_major__ > 11)
   return __builtin_ia32_reduce_mul_d512((__v16si)__W);
+#else
+  _mm512_mask_reduce_operator(*);
 #endif
 }
 
 static __inline__ int __DEFAULT_FN_ATTRS512
 _mm512_reduce_and_epi32(__m512i __W) {
-#if (__clang_major__ < 14)
+#if (__clang_major__ > 13)
+  return __builtin_reduce_and((__v16si)__W);
+#elif (__clang_major__ > 11)
   return __builtin_ia32_reduce_and_d512((__v16si)__W);
 #else
-  return __builtin_reduce_and((__v16si)__W);
+  _mm512_mask_reduce_operator(&);
 #endif
 }
 
 static __inline__ int __DEFAULT_FN_ATTRS512
 _mm512_reduce_or_epi32(__m512i __W) {
-#if (__clang_major__ < 14)
+#if (__clang_major__ > 13)
+  return __builtin_reduce_or((__v16si)__W);
+#elif (__clang_major__ > 11)
   return __builtin_ia32_reduce_or_d512((__v16si)__W);
 #else
-  return __builtin_reduce_or((__v16si)__W);
+  _mm512_mask_reduce_operator(|);
 #endif
 }
 
@@ -9491,8 +9574,10 @@ _mm512_mask_reduce_add_epi32( __mmask16 __M, __m512i __W) {
   __W = _mm512_maskz_mov_epi32(__M, __W);
 #if (__clang_major__ > 14)
   return __builtin_reduce_add((__v16si)__W);
-#else
+#elif (__clang_major__ > 11)
   return __builtin_ia32_reduce_add_d512((__v16si)__W);
+#else
+  _mm512_mask_reduce_operator(+);
 #endif
 }
 
@@ -9501,205 +9586,305 @@ _mm512_mask_reduce_mul_epi32( __mmask16 __M, __m512i __W) {
   __W = _mm512_mask_mov_epi32(_mm512_set1_epi32(1), __M, __W);
 #if (__clang_major__ > 14)
   return __builtin_reduce_mul((__v16si)__W);
-#else
+#elif (__clang_major__ > 11)
   return __builtin_ia32_reduce_mul_d512((__v16si)__W);
+#else
+  _mm512_mask_reduce_operator(*);
 #endif
 }
 
 static __inline__ int __DEFAULT_FN_ATTRS512
 _mm512_mask_reduce_and_epi32( __mmask16 __M, __m512i __W) {
   __W = _mm512_mask_mov_epi32(_mm512_set1_epi32(~0U), __M, __W);
-#if (__clang_major__ < 14)
+#if (__clang_major__ > 13)
+  return __builtin_reduce_and((__v16si)__W);
+#elif (__clang_major__ > 11)
   return __builtin_ia32_reduce_and_d512((__v16si)__W);
 #else
-  return __builtin_reduce_and((__v16si)__W);
+  _mm512_mask_reduce_operator(&);
 #endif
 }
 
 static __inline__ int __DEFAULT_FN_ATTRS512
 _mm512_mask_reduce_or_epi32(__mmask16 __M, __m512i __W) {
   __W = _mm512_maskz_mov_epi32(__M, __W);
-#if (__clang_major__ < 14)
+#if (__clang_major__ > 13)
+  return __builtin_reduce_or((__v16si)__W);
+#elif (__clang_major__ > 11)
   return __builtin_ia32_reduce_or_d512((__v16si)__W);
 #else
-  return __builtin_reduce_or((__v16si)__W);
+  _mm512_mask_reduce_operator(|);
 #endif
 }
+#if (__clang_major__ <= 11)
+#undef _mm512_mask_reduce_operator
+
+#define _mm512_mask_reduce_operator(op) \
+  __m256 __t1 = (__m256)_mm512_extractf64x4_pd((__m512d)__W, 0); \
+  __m256 __t2 = (__m256)_mm512_extractf64x4_pd((__m512d)__W, 1); \
+  __m256 __t3 = __t1 op __t2; \
+  __m128 __t4 = _mm256_extractf128_ps(__t3, 0); \
+  __m128 __t5 = _mm256_extractf128_ps(__t3, 1); \
+  __m128 __t6 = __t4 op __t5; \
+  __m128 __t7 = __builtin_shufflevector(__t6, __t6, 2, 3, 0, 1); \
+  __m128 __t8 = __t6 op __t7; \
+  __m128 __t9 = __builtin_shufflevector(__t8, __t8, 1, 0, 3, 2); \
+  __m128 __t10 = __t8 op __t9; \
+  return __t10[0]
+#endif
 
 static __inline__ float __DEFAULT_FN_ATTRS512
 _mm512_reduce_add_ps(__m512 __W) {
+#if (__clang_major__ > 11)
   return __builtin_ia32_reduce_fadd_ps512(-0.0f, __W);
+#else
+  _mm512_mask_reduce_operator(+);
+#endif
 }
 
 static __inline__ float __DEFAULT_FN_ATTRS512
 _mm512_reduce_mul_ps(__m512 __W) {
+#if (__clang_major__ > 11)
   return __builtin_ia32_reduce_fmul_ps512(1.0f, __W);
+#else
+  _mm512_mask_reduce_operator(*);
+#endif
 }
 
 static __inline__ float __DEFAULT_FN_ATTRS512
 _mm512_mask_reduce_add_ps(__mmask16 __M, __m512 __W) {
   __W = _mm512_maskz_mov_ps(__M, __W);
+#if (__clang_major__ > 11)
   return __builtin_ia32_reduce_fadd_ps512(-0.0f, __W);
+#else
+  _mm512_mask_reduce_operator(+);
+#endif
 }
 
 static __inline__ float __DEFAULT_FN_ATTRS512
 _mm512_mask_reduce_mul_ps(__mmask16 __M, __m512 __W) {
   __W = _mm512_mask_mov_ps(_mm512_set1_ps(1.0f), __M, __W);
+#if (__clang_major__ > 11)
   return __builtin_ia32_reduce_fmul_ps512(1.0f, __W);
+#else
+  _mm512_mask_reduce_operator(*);
+#endif
 }
+#if (__clang_major__ <= 11)
+#undef _mm512_mask_reduce_operator
+
+#define _mm512_mask_reduce_operator(op) \
+  __m512i __t1 = (__m512i)__builtin_shufflevector((__v8di)__V, (__v8di)__V, 4, 5, 6, 7, 0, 1, 2, 3); \
+  __m512i __t2 = _mm512_##op(__V, __t1); \
+  __m512i __t3 = (__m512i)__builtin_shufflevector((__v8di)__t2, (__v8di)__t2, 2, 3, 0, 1, 6, 7, 4, 5); \
+  __m512i __t4 = _mm512_##op(__t2, __t3); \
+  __m512i __t5 = (__m512i)__builtin_shufflevector((__v8di)__t4, (__v8di)__t4, 1, 0, 3, 2, 5, 4, 7, 6); \
+  __v8di __t6 = (__v8di)_mm512_##op(__t4, __t5); \
+  return __t6[0]
+#endif
 
 static __inline__ long long __DEFAULT_FN_ATTRS512
 _mm512_reduce_max_epi64(__m512i __V) {
-#if (__clang_major__ < 14)
+#if (__clang_major__ > 13)
+  return __builtin_reduce_max((__v8di)__V);
+#elif (__clang_major__ > 11)
   return __builtin_ia32_reduce_smax_q512(__V);
 #else
-  return __builtin_reduce_max((__v8di)__V);
+  _mm512_mask_reduce_operator(max_epi64);
 #endif
 }
 
 static __inline__ unsigned long long __DEFAULT_FN_ATTRS512
 _mm512_reduce_max_epu64(__m512i __V) {
-#if (__clang_major__ < 14)
+#if (__clang_major__ > 13)
+  return __builtin_reduce_max((__v8du)__V);
+#elif (__clang_major__ > 11)
   return __builtin_ia32_reduce_umax_q512(__V);
 #else
-  return __builtin_reduce_max((__v8du)__V);
+  _mm512_mask_reduce_operator(max_epu64);
 #endif
 }
 
 static __inline__ long long __DEFAULT_FN_ATTRS512
 _mm512_reduce_min_epi64(__m512i __V) {
-#if (__clang_major__ < 14)
+#if (__clang_major__ > 13)
+  return __builtin_reduce_min((__v8di)__V);
+#elif (__clang_major__ > 11)
   return __builtin_ia32_reduce_smin_q512(__V);
 #else
-  return __builtin_reduce_min((__v8di)__V);
+  _mm512_mask_reduce_operator(min_epi64);
 #endif
 }
 
 static __inline__ unsigned long long __DEFAULT_FN_ATTRS512
 _mm512_reduce_min_epu64(__m512i __V) {
-#if (__clang_major__ < 14)
+#if (__clang_major__ > 13)
+  return __builtin_reduce_min((__v8du)__V);
+#elif (__clang_major__ > 11)
   return __builtin_ia32_reduce_umin_q512(__V);
 #else
-  return __builtin_reduce_min((__v8du)__V);
+  _mm512_mask_reduce_operator(min_epu64);
 #endif
 }
 
 static __inline__ long long __DEFAULT_FN_ATTRS512
 _mm512_mask_reduce_max_epi64(__mmask8 __M, __m512i __V) {
   __V = _mm512_mask_mov_epi64(_mm512_set1_epi64(-__LONG_LONG_MAX__ - 1LL), __M, __V);
-#if (__clang_major__ < 14)
+#if (__clang_major__ > 13)
+  return __builtin_reduce_max((__v8di)__V);
+#elif (__clang_major__ > 11)
   return __builtin_ia32_reduce_smax_q512(__V);
 #else
-  return __builtin_reduce_max((__v8di)__V);
+  _mm512_mask_reduce_operator(max_epi64);
 #endif
 }
 
 static __inline__ unsigned long long __DEFAULT_FN_ATTRS512
 _mm512_mask_reduce_max_epu64(__mmask8 __M, __m512i __V) {
   __V = _mm512_maskz_mov_epi64(__M, __V);
-#if (__clang_major__ < 14)
+#if (__clang_major__ > 13)
+  return __builtin_reduce_max((__v8du)__V);
+#elif (__clang_major__ > 11)
   return __builtin_ia32_reduce_umax_q512(__V);
 #else
-  return __builtin_reduce_max((__v8du)__V);
+  _mm512_mask_reduce_operator(max_epu64);
 #endif
 }
 
 static __inline__ long long __DEFAULT_FN_ATTRS512
 _mm512_mask_reduce_min_epi64(__mmask8 __M, __m512i __V) {
   __V = _mm512_mask_mov_epi64(_mm512_set1_epi64(__LONG_LONG_MAX__), __M, __V);
-#if (__clang_major__ < 14)
+#if (__clang_major__ > 13)
+  return __builtin_reduce_min((__v8di)__V);
+#elif (__clang_major__ > 11)
   return __builtin_ia32_reduce_smin_q512(__V);
 #else
-  return __builtin_reduce_min((__v8di)__V);
+  _mm512_mask_reduce_operator(min_epi64);
 #endif
 }
 
 static __inline__ unsigned long long __DEFAULT_FN_ATTRS512
 _mm512_mask_reduce_min_epu64(__mmask8 __M, __m512i __V) {
   __V = _mm512_mask_mov_epi64(_mm512_set1_epi64(~0ULL), __M, __V);
-#if (__clang_major__ < 14)
+#if (__clang_major__ > 13)
+  return __builtin_reduce_min((__v8du)__V);
+#elif (__clang_major__ > 11)
   return __builtin_ia32_reduce_umin_q512(__V);
 #else
-  return __builtin_reduce_min((__v8du)__V);
+  _mm512_mask_reduce_operator(min_epu64);
 #endif
 }
+#if (__clang_major__ <= 11)
+#undef _mm512_mask_reduce_operator
+
+#define _mm512_mask_reduce_operator(op) \
+  __m256i __t1 = _mm512_extracti64x4_epi64(__V, 0); \
+  __m256i __t2 = _mm512_extracti64x4_epi64(__V, 1); \
+  __m256i __t3 = _mm256_##op(__t1, __t2); \
+  __m128i __t4 = _mm256_extracti128_si256(__t3, 0); \
+  __m128i __t5 = _mm256_extracti128_si256(__t3, 1); \
+  __m128i __t6 = _mm_##op(__t4, __t5); \
+  __m128i __t7 = (__m128i)__builtin_shufflevector((__v4si)__t6, (__v4si)__t6, 2, 3, 0, 1); \
+  __m128i __t8 = _mm_##op(__t6, __t7); \
+  __m128i __t9 = (__m128i)__builtin_shufflevector((__v4si)__t8, (__v4si)__t8, 1, 0, 3, 2); \
+  __v4si __t10 = (__v4si)_mm_##op(__t8, __t9); \
+  return __t10[0]
+#endif
+
 static __inline__ int __DEFAULT_FN_ATTRS512
 _mm512_reduce_max_epi32(__m512i __V) {
-#if (__clang_major__ < 14)
+#if (__clang_major__ > 13)
+  return __builtin_reduce_max((__v16si)__V);
+#elif (__clang_major__ > 11)
   return __builtin_ia32_reduce_smax_d512((__v16si)__V);
 #else
-  return __builtin_reduce_max((__v16si)__V);
+  _mm512_mask_reduce_operator(max_epi32);
 #endif
 }
 
 static __inline__ unsigned int __DEFAULT_FN_ATTRS512
 _mm512_reduce_max_epu32(__m512i __V) {
-#if (__clang_major__ < 14)
+#if (__clang_major__ > 13)
+  return __builtin_reduce_max((__v16su)__V);
+#elif (__clang_major__ > 11)
   return __builtin_ia32_reduce_umax_d512((__v16si)__V);
 #else
-  return __builtin_reduce_max((__v16su)__V);
+  _mm512_mask_reduce_operator(max_epu32);
 #endif
 }
 
 static __inline__ int __DEFAULT_FN_ATTRS512
 _mm512_reduce_min_epi32(__m512i __V) {
-#if (__clang_major__ < 14)
+#if (__clang_major__ > 13)
+  return __builtin_reduce_min((__v16si)__V);
+#elif (__clang_major__ > 11)
   return __builtin_ia32_reduce_smin_d512((__v16si)__V);
 #else
-  return __builtin_reduce_min((__v16si)__V);
+  _mm512_mask_reduce_operator(min_epi32);
 #endif
 }
 
 static __inline__ unsigned int __DEFAULT_FN_ATTRS512
 _mm512_reduce_min_epu32(__m512i __V) {
-#if (__clang_major__ < 14)
+#if (__clang_major__ > 13)
+  return __builtin_reduce_min((__v16su)__V);
+#elif (__clang_major__ > 11)
   return __builtin_ia32_reduce_umin_d512((__v16si)__V);
 #else
-  return __builtin_reduce_min((__v16su)__V);
+  _mm512_mask_reduce_operator(min_epu32);
 #endif
 }
 
 static __inline__ int __DEFAULT_FN_ATTRS512
 _mm512_mask_reduce_max_epi32(__mmask16 __M, __m512i __V) {
   __V = _mm512_mask_mov_epi32(_mm512_set1_epi32(-__INT_MAX__ - 1), __M, __V);
-#if (__clang_major__ < 14)
+#if (__clang_major__ > 13)
+  return __builtin_reduce_max((__v16si)__V);
+#elif (__clang_major__ > 11)
   return __builtin_ia32_reduce_smax_d512((__v16si)__V);
 #else
-  return __builtin_reduce_max((__v16si)__V);
+  _mm512_mask_reduce_operator(max_epi32);
 #endif
 }
 
 static __inline__ unsigned int __DEFAULT_FN_ATTRS512
 _mm512_mask_reduce_max_epu32(__mmask16 __M, __m512i __V) {
   __V = _mm512_maskz_mov_epi32(__M, __V);
-#if (__clang_major__ < 14)
+#if (__clang_major__ > 13)
+  return __builtin_reduce_max((__v16su)__V);
+#elif (__clang_major__ > 11)
   return __builtin_ia32_reduce_umax_d512((__v16si)__V);
 #else
-  return __builtin_reduce_max((__v16su)__V);
+  _mm512_mask_reduce_operator(max_epu32);
 #endif
 }
 
 static __inline__ int __DEFAULT_FN_ATTRS512
 _mm512_mask_reduce_min_epi32(__mmask16 __M, __m512i __V) {
   __V = _mm512_mask_mov_epi32(_mm512_set1_epi32(__INT_MAX__), __M, __V);
-#if (__clang_major__ < 14)
+#if (__clang_major__ > 13)
+  return __builtin_reduce_min((__v16si)__V);
+#elif (__clang_major__ > 11)
   return __builtin_ia32_reduce_smin_d512((__v16si)__V);
 #else
-  return __builtin_reduce_min((__v16si)__V);
+  _mm512_mask_reduce_operator(min_epi32);
 #endif
 }
 
 static __inline__ unsigned int __DEFAULT_FN_ATTRS512
 _mm512_mask_reduce_min_epu32(__mmask16 __M, __m512i __V) {
   __V = _mm512_mask_mov_epi32(_mm512_set1_epi32(~0U), __M, __V);
-#if (__clang_major__ < 14)
+#if (__clang_major__ > 13)
+  return __builtin_reduce_min((__v16su)__V);
+#elif (__clang_major__ > 11)
   return __builtin_ia32_reduce_umin_d512((__v16si)__V);
 #else
-  return __builtin_reduce_min((__v16su)__V);
+  _mm512_mask_reduce_operator(min_epu32);
 #endif
 }
-
 #if (__clang_major__ < 13)
+#undef _mm512_mask_reduce_operator
+
 #define _mm512_mask_reduce_operator(op) \
   __m256d __t1 = _mm512_extractf64x4_pd(__V, 0); \
   __m256d __t2 = _mm512_extractf64x4_pd(__V, 1); \
diff --git a/include-llvm/immintrin.h b/include-llvm/immintrin.h
index e5174f8..0cc7260 100644
--- a/include-llvm/immintrin.h
+++ b/include-llvm/immintrin.h
@@ -14,7 +14,9 @@
 #error "This header is only meant to be used on x86 and x64 architecture"
 #endif
 
+#if (__clang_major__ >= 12)
 #include <x86gprintrin.h>
+#endif
 
 #if !(defined(_MSC_VER) || defined(__SCE__)) || __has_feature(modules) ||      \
     defined(__MMX__)
@@ -144,10 +146,12 @@
 #include <avx512vlvnniintrin.h>
 #endif
 
+#if (__clang_major__ >= 12)
 #if !(defined(_MSC_VER) || defined(__SCE__)) || __has_feature(modules) ||      \
     defined(__AVXVNNI__)
 #include <avxvnniintrin.h>
 #endif
+#endif
 
 #if !(defined(_MSC_VER) || defined(__SCE__)) || __has_feature(modules) ||      \
     defined(__AVX512DQ__)
@@ -214,6 +218,7 @@
 #include <avx512pfintrin.h>
 #endif
 
+#if (__clang_major__ >= 12)
 /*
  * FIXME: _Float16 type is legal only when HW support float16 operation.
  * We use __AVX512FP16__ to identify if float16 is supported or not, so
@@ -227,6 +232,7 @@
 #if defined(__AVX512FP16__) && defined(__AVX512VL__)
 #include <avx512vlfp16intrin.h>
 #endif
+#endif
 
 #if !(defined(_MSC_VER) || defined(__SCE__)) || __has_feature(modules) ||      \
     defined(__AVX512BF16__)
@@ -496,10 +502,12 @@ _storebe_i64(void * __P, long long __D) {
 #include <invpcidintrin.h>
 #endif
 
+#if (__clang_major__ >= 12)
 #if !(defined(_MSC_VER) || defined(__SCE__)) || __has_feature(modules) ||      \
     defined(__KL__) || defined(__WIDEKL__)
 #include <keylockerintrin.h>
 #endif
+#endif
 
 #if !(defined(_MSC_VER) || defined(__SCE__)) || __has_feature(modules) ||      \
     defined(__AMXTILE__) || defined(__AMXINT8__) || defined(__AMXBF16__)