[Headers][X86] Allow BITALG vpopcntw/vpopcntb intrinsics to be used in constexpr (#152701)

RKSimon · web-flow · commit f169893cbf6e · 2025-08-08T16:09:26.000+01:00
Matches VPOPCNTDQ handling
diff --git a/clang/lib/Headers/avx512bitalgintrin.h b/clang/lib/Headers/avx512bitalgintrin.h
@@ -20,7 +20,13 @@
                  __target__("avx512bitalg,evex512"),                           \
                  __min_vector_width__(512)))
 
-static __inline__ __m512i __DEFAULT_FN_ATTRS
+#if defined(__cplusplus) && (__cplusplus >= 201103L)
+#define __DEFAULT_FN_ATTRS_CONSTEXPR __DEFAULT_FN_ATTRS constexpr
+#else
+#define __DEFAULT_FN_ATTRS_CONSTEXPR __DEFAULT_FN_ATTRS
+#endif
+
+static __inline__ __m512i __DEFAULT_FN_ATTRS_CONSTEXPR
 _mm512_popcnt_epi16(__m512i __A)
 {
   return (__m512i)__builtin_elementwise_popcount((__v32hu)__A);
@@ -42,7 +48,7 @@ _mm512_maskz_popcnt_epi16(__mmask32 __U, __m512i __B)
               __B);
 }
 
-static __inline__ __m512i __DEFAULT_FN_ATTRS
+static __inline__ __m512i __DEFAULT_FN_ATTRS_CONSTEXPR
 _mm512_popcnt_epi8(__m512i __A)
 {
   return (__m512i)__builtin_elementwise_popcount((__v64qu)__A);
@@ -80,7 +86,7 @@ _mm512_bitshuffle_epi64_mask(__m512i __A, __m512i __B)
               __B);
 }
 
-
 #undef __DEFAULT_FN_ATTRS
+#undef __DEFAULT_FN_ATTRS_CONSTEXPR
 
 #endif
diff --git a/clang/lib/Headers/avx512vlbitalgintrin.h b/clang/lib/Headers/avx512vlbitalgintrin.h
@@ -24,7 +24,15 @@
                  __target__("avx512vl,avx512bitalg,no-evex512"),               \
                  __min_vector_width__(256)))
 
-static __inline__ __m256i __DEFAULT_FN_ATTRS256
+#if defined(__cplusplus) && (__cplusplus >= 201103L)
+#define __DEFAULT_FN_ATTRS128_CONSTEXPR __DEFAULT_FN_ATTRS128 constexpr
+#define __DEFAULT_FN_ATTRS256_CONSTEXPR __DEFAULT_FN_ATTRS256 constexpr
+#else
+#define __DEFAULT_FN_ATTRS128_CONSTEXPR __DEFAULT_FN_ATTRS128
+#define __DEFAULT_FN_ATTRS256_CONSTEXPR __DEFAULT_FN_ATTRS256
+#endif
+
+static __inline__ __m256i __DEFAULT_FN_ATTRS256_CONSTEXPR
 _mm256_popcnt_epi16(__m256i __A)
 {
   return (__m256i)__builtin_elementwise_popcount((__v16hu)__A);
@@ -46,7 +54,7 @@ _mm256_maskz_popcnt_epi16(__mmask16 __U, __m256i __B)
               __B);
 }
 
-static __inline__ __m128i __DEFAULT_FN_ATTRS128
+static __inline__ __m128i __DEFAULT_FN_ATTRS128_CONSTEXPR
 _mm_popcnt_epi16(__m128i __A)
 {
   return (__m128i)__builtin_elementwise_popcount((__v8hu)__A);
@@ -68,7 +76,7 @@ _mm_maskz_popcnt_epi16(__mmask8 __U, __m128i __B)
               __B);
 }
 
-static __inline__ __m256i __DEFAULT_FN_ATTRS256
+static __inline__ __m256i __DEFAULT_FN_ATTRS256_CONSTEXPR
 _mm256_popcnt_epi8(__m256i __A)
 {
   return (__m256i)__builtin_elementwise_popcount((__v32qu)__A);
@@ -90,7 +98,7 @@ _mm256_maskz_popcnt_epi8(__mmask32 __U, __m256i __B)
               __B);
 }
 
-static __inline__ __m128i __DEFAULT_FN_ATTRS128
+static __inline__ __m128i __DEFAULT_FN_ATTRS128_CONSTEXPR
 _mm_popcnt_epi8(__m128i __A)
 {
   return (__m128i)__builtin_elementwise_popcount((__v16qu)__A);
@@ -147,5 +155,7 @@ _mm_bitshuffle_epi64_mask(__m128i __A, __m128i __B)
 
 #undef __DEFAULT_FN_ATTRS128
 #undef __DEFAULT_FN_ATTRS256
+#undef __DEFAULT_FN_ATTRS128_CONSTEXPR
+#undef __DEFAULT_FN_ATTRS256_CONSTEXPR
 
 #endif
diff --git a/clang/test/CodeGen/X86/avx512bitalg-builtins.c b/clang/test/CodeGen/X86/avx512bitalg-builtins.c
@@ -4,12 +4,14 @@
 // RUN: %clang_cc1 -x c++ -flax-vector-conversions=none -ffreestanding %s -triple=i386-apple-darwin -target-feature +avx512bitalg -emit-llvm -o - -Wall -Werror | FileCheck %s
 
 #include <immintrin.h>
+#include "builtin_test_helpers.h"
 
 __m512i test_mm512_popcnt_epi16(__m512i __A) {
   // CHECK-LABEL: test_mm512_popcnt_epi16
   // CHECK: @llvm.ctpop.v32i16
   return _mm512_popcnt_epi16(__A);
 }
+TEST_CONSTEXPR(match_v32hi(_mm512_popcnt_epi16((__m512i)(__v32hi){+5, -3, -10, +8, 0, -256, +256, -128, +3, +9, +15, +33, +63, +129, +511, +1025, +5, -3, -10, +8, 0, -256, +256, -128, +3, +9, +15, +33, +63, +129, +511, +1025}), 2, 15, 14, 1, 0, 8, 1, 9, 2, 2, 4, 2, 6, 2, 9, 2, 2, 15, 14, 1, 0, 8, 1, 9, 2, 2, 4, 2, 6, 2, 9, 2));
 
 __m512i test_mm512_mask_popcnt_epi16(__m512i __A, __mmask32 __U, __m512i __B) {
   // CHECK-LABEL: test_mm512_mask_popcnt_epi16
@@ -29,6 +31,7 @@ __m512i test_mm512_popcnt_epi8(__m512i __A) {
   // CHECK: @llvm.ctpop.v64i8
   return _mm512_popcnt_epi8(__A);
 }
+TEST_CONSTEXPR(match_v64qi(_mm512_popcnt_epi8((__m512i)(__v64qi){+5, -3, -10, +8, 0, -16, +16, -16, +3, +9, +15, +33, +63, +33, +53, +73, +5, -3, -10, +8, 0, -16, +16, -16, +3, +9, +15, +33, +63, +33, +53, +73, +5, -3, -10, +8, 0, -16, +16, -16, +3, +9, +15, +33, +63, +33, +53, +73, +5, -3, -10, +8, 0, -16, +16, -16, +3, +9, +15, +33, +63, +33, +53, +73}), 2, 7, 6, 1, 0, 4, 1, 4, 2, 2, 4, 2, 6, 2, 4, 3, 2, 7, 6, 1, 0, 4, 1, 4, 2, 2, 4, 2, 6, 2, 4, 3, 2, 7, 6, 1, 0, 4, 1, 4, 2, 2, 4, 2, 6, 2, 4, 3, 2, 7, 6, 1, 0, 4, 1, 4, 2, 2, 4, 2, 6, 2, 4, 3));
 
 __m512i test_mm512_mask_popcnt_epi8(__m512i __A, __mmask64 __U, __m512i __B) {
   // CHECK-LABEL: test_mm512_mask_popcnt_epi8
diff --git a/clang/test/CodeGen/X86/avx512vlbitalg-builtins.c b/clang/test/CodeGen/X86/avx512vlbitalg-builtins.c
@@ -4,12 +4,14 @@
 // RUN: %clang_cc1 -x c++ -flax-vector-conversions=none -ffreestanding %s -triple=i386-apple-darwin -target-feature +avx512bitalg -target-feature +avx512vl -emit-llvm -o - -Wall -Werror | FileCheck %s
 
 #include <immintrin.h>
+#include "builtin_test_helpers.h"
 
 __m256i test_mm256_popcnt_epi16(__m256i __A) {
   // CHECK-LABEL: test_mm256_popcnt_epi16
   // CHECK: @llvm.ctpop.v16i16
   return _mm256_popcnt_epi16(__A);
 }
+TEST_CONSTEXPR(match_v16hi(_mm256_popcnt_epi16((__m256i)(__v16hi){+5, -3, -10, +8, 0, -256, +256, -128, +3, +9, +15, +33, +63, +129, +511, +1025}), 2, 15, 14, 1, 0, 8, 1, 9, 2, 2, 4, 2, 6, 2, 9, 2));
 
 __m256i test_mm256_mask_popcnt_epi16(__m256i __A, __mmask16 __U, __m256i __B) {
   // CHECK-LABEL: test_mm256_mask_popcnt_epi16
@@ -29,6 +31,7 @@ __m128i test_mm_popcnt_epi16(__m128i __A) {
   // CHECK: @llvm.ctpop.v8i16
   return _mm_popcnt_epi16(__A);
 }
+TEST_CONSTEXPR(match_v8hi(_mm_popcnt_epi16((__m128i)(__v8hi){+5, -3, -10, +8, 0, -256, +256, -128}), 2, 15, 14, 1, 0, 8, 1, 9));
 
 __m128i test_mm_mask_popcnt_epi16(__m128i __A, __mmask8 __U, __m128i __B) {
   // CHECK-LABEL: test_mm_mask_popcnt_epi16
@@ -48,6 +51,7 @@ __m256i test_mm256_popcnt_epi8(__m256i __A) {
   // CHECK: @llvm.ctpop.v32i8
   return _mm256_popcnt_epi8(__A);
 }
+TEST_CONSTEXPR(match_v32qi(_mm256_popcnt_epi8((__m256i)(__v32qi){+5, -3, -10, +8, 0, -16, +16, -16, +3, +9, +15, +33, +63, +33, +53, +73, +5, -3, -10, +8, 0, -16, +16, -16, +3, +9, +15, +33, +63, +33, +53, +73}), 2, 7, 6, 1, 0, 4, 1, 4, 2, 2, 4, 2, 6, 2, 4, 3, 2, 7, 6, 1, 0, 4, 1, 4, 2, 2, 4, 2, 6, 2, 4, 3));
 
 __m256i test_mm256_mask_popcnt_epi8(__m256i __A, __mmask32 __U, __m256i __B) {
   // CHECK-LABEL: test_mm256_mask_popcnt_epi8
@@ -67,6 +71,7 @@ __m128i test_mm_popcnt_epi8(__m128i __A) {
   // CHECK: @llvm.ctpop.v16i8
   return _mm_popcnt_epi8(__A);
 }
+TEST_CONSTEXPR(match_v16qi(_mm_popcnt_epi8((__m128i)(__v16qi){+5, -3, -10, +8, 0, -16, +16, -16, +3, +9, +15, +33, +63, +33, +53, +73}), 2, 7, 6, 1, 0, 4, 1, 4, 2, 2, 4, 2, 6, 2, 4, 3));
 
 __m128i test_mm_mask_popcnt_epi8(__m128i __A, __mmask16 __U, __m128i __B) {
   // CHECK-LABEL: test_mm_mask_popcnt_epi8
diff --git a/clang/test/CodeGen/X86/builtin_test_helpers.h b/clang/test/CodeGen/X86/builtin_test_helpers.h
@@ -122,6 +122,36 @@ constexpr bool match_v16si(__m512i _v, int a, int b, int c, int d, int e, int f,
   return v[0] == a && v[1] == b && v[2] == c && v[3] == d && v[4] == e && v[5] == f && v[6] == g && v[7] == h && v[8] == i && v[9] == j && v[10] == k && v[11] == l && v[12] == m && v[13] == n && v[14] == o && v[15] == p;
 }
 
+constexpr bool match_v32hi(__m512i _v, short __e00, short __e01, short __e02, short __e03, short __e04, short __e05, short __e06, short __e07,
+                                       short __e08, short __e09, short __e10, short __e11, short __e12, short __e13, short __e14, short __e15,
+                                       short __e16, short __e17, short __e18, short __e19, short __e20, short __e21, short __e22, short __e23,
+                                       short __e24, short __e25, short __e26, short __e27, short __e28, short __e29, short __e30, short __e31) {
+  __v32hi v = (__v32hi)_v;
+  return v[ 0] == __e00 && v[ 1] == __e01 && v[ 2] == __e02 && v[ 3] == __e03 && v[ 4] == __e04 && v[ 5] == __e05 && v[ 6] == __e06 && v[ 7] ==  __e07 &&
+         v[ 8] == __e08 && v[ 9] == __e09 && v[10] == __e10 && v[11] == __e11 && v[12] == __e12 && v[13] == __e13 && v[14] == __e14 && v[15] ==  __e15 &&
+         v[16] == __e16 && v[17] == __e17 && v[18] == __e18 && v[19] == __e19 && v[20] == __e20 && v[21] == __e21 && v[22] == __e22 && v[23] ==  __e23 &&
+         v[24] == __e24 && v[25] == __e25 && v[26] == __e26 && v[27] == __e27 && v[28] == __e28 && v[29] == __e29 && v[30] == __e30 && v[31] ==  __e31;
+}
+
+constexpr bool match_v64qi(__m512i _v, char __e00, char __e01, char __e02, char __e03, char __e04, char __e05, char __e06, char __e07,
+                                       char __e08, char __e09, char __e10, char __e11, char __e12, char __e13, char __e14, char __e15,
+                                       char __e16, char __e17, char __e18, char __e19, char __e20, char __e21, char __e22, char __e23,
+                                       char __e24, char __e25, char __e26, char __e27, char __e28, char __e29, char __e30, char __e31,
+                                       char __e32, char __e33, char __e34, char __e35, char __e36, char __e37, char __e38, char __e39,
+                                       char __e40, char __e41, char __e42, char __e43, char __e44, char __e45, char __e46, char __e47,
+                                       char __e48, char __e49, char __e50, char __e51, char __e52, char __e53, char __e54, char __e55,
+                                       char __e56, char __e57, char __e58, char __e59, char __e60, char __e61, char __e62, char __e63) {
+  __v64qi v = (__v64qi)_v;
+  return v[ 0] == __e00 && v[ 1] == __e01 && v[ 2] == __e02 && v[ 3] == __e03 && v[ 4] == __e04 && v[ 5] == __e05 && v[ 6] == __e06 && v[ 7] == __e07 &&
+         v[ 8] == __e08 && v[ 9] == __e09 && v[10] == __e10 && v[11] == __e11 && v[12] == __e12 && v[13] == __e13 && v[14] == __e14 && v[15] == __e15 &&
+         v[16] == __e16 && v[17] == __e17 && v[18] == __e18 && v[19] == __e19 && v[20] == __e20 && v[21] == __e21 && v[22] == __e22 && v[23] == __e23 &&
+         v[24] == __e24 && v[25] == __e25 && v[26] == __e26 && v[27] == __e27 && v[28] == __e28 && v[29] == __e29 && v[30] == __e30 && v[31] == __e31 &&
+         v[32] == __e32 && v[33] == __e33 && v[34] == __e34 && v[35] == __e35 && v[36] == __e36 && v[37] == __e37 && v[38] == __e38 && v[39] == __e39 &&
+         v[40] == __e40 && v[41] == __e41 && v[42] == __e42 && v[43] == __e43 && v[44] == __e44 && v[45] == __e45 && v[46] == __e46 && v[47] == __e47 &&
+         v[48] == __e48 && v[49] == __e49 && v[50] == __e50 && v[51] == __e51 && v[52] == __e52 && v[53] == __e53 && v[54] == __e54 && v[55] == __e55 &&
+         v[56] == __e56 && v[57] == __e57 && v[58] == __e58 && v[59] == __e59 && v[60] == __e60 && v[61] == __e61 && v[62] == __e62 && v[63] == __e63;
+}
+
 #define TEST_CONSTEXPR(...) static_assert(__VA_ARGS__)
 
 #else
