[clang] [NFC] Use macros only when __AVX512IFMA__ and __AVXIFMA__ undefined (PR #162760)

[Phoebe Wang via cfe-commits]

https://github.com/phoebewang updated https://github.com/llvm/llvm-project/pull/162760

>From 04217fe5454a10b57f18014131f15e6fe17f1967 Mon Sep 17 00:00:00 2001
From: Phoebe Wang <phoebe.wang at intel.com>
Date: Fri, 10 Oct 2025 09:48:06 +0800
Subject: [PATCH] [NFC] Use macros only when __AVX512IFMA__ and __AVXIFMA__
 undefined

---
 clang/lib/Headers/avx512ifmavlintrin.h | 72 +++++++++++++++++---------
 clang/lib/Headers/avxifmaintrin.h      |  7 +++
 2 files changed, 55 insertions(+), 24 deletions(-)

diff --git a/clang/lib/Headers/avx512ifmavlintrin.h b/clang/lib/Headers/avx512ifmavlintrin.h
index a72b56113a12b..ddee9636f20f7 100644
--- a/clang/lib/Headers/avx512ifmavlintrin.h
+++ b/clang/lib/Headers/avx512ifmavlintrin.h
@@ -24,6 +24,7 @@
                  __target__("avx512ifma,avx512vl"),                            \
                  __min_vector_width__(256)))
 
+#if !(defined(__AVXIFMA__) || defined(__AVX512IFMA__))
 #define _mm_madd52hi_epu64(X, Y, Z)                                            \
   ((__m128i)__builtin_ia32_vpmadd52huq128((__v2di)(X), (__v2di)(Y),            \
                                           (__v2di)(Z)))
@@ -39,69 +40,92 @@
 #define _mm256_madd52lo_epu64(X, Y, Z)                                         \
   ((__m256i)__builtin_ia32_vpmadd52luq256((__v4di)(X), (__v4di)(Y),            \
                                           (__v4di)(Z)))
+#endif
+
+#if defined(__AVX512IFMA__)
+static __inline__ __m128i __DEFAULT_FN_ATTRS128
+_mm_madd52hi_epu64(__m128i __X, __m128i __Y, __m128i __Z) {
+  return (__m128i)__builtin_ia32_vpmadd52huq128((__v2di)__X, (__v2di)__Y,
+                                                (__v2di)__Z);
+}
+
+static __inline__ __m256i __DEFAULT_FN_ATTRS256
+_mm256_madd52hi_epu64(__m256i __X, __m256i __Y, __m256i __Z) {
+  return (__m256i)__builtin_ia32_vpmadd52huq256((__v4di)__X, (__v4di)__Y,
+                                                (__v4di)__Z);
+}
+
+static __inline__ __m128i __DEFAULT_FN_ATTRS128
+_mm_madd52lo_epu64(__m128i __X, __m128i __Y, __m128i __Z) {
+  return (__m128i)__builtin_ia32_vpmadd52luq128((__v2di)__X, (__v2di)__Y,
+                                                (__v2di)__Z);
+}
+
+static __inline__ __m256i __DEFAULT_FN_ATTRS256
+_mm256_madd52lo_epu64(__m256i __X, __m256i __Y, __m256i __Z) {
+  return (__m256i)__builtin_ia32_vpmadd52luq256((__v4di)__X, (__v4di)__Y,
+                                                (__v4di)__Z);
+}
+#endif
 
 static __inline__ __m128i __DEFAULT_FN_ATTRS128
 _mm_mask_madd52hi_epu64 (__m128i __W, __mmask8 __M, __m128i __X, __m128i __Y)
 {
-  return (__m128i)__builtin_ia32_selectq_128(__M,
-                                      (__v2di)_mm_madd52hi_epu64(__W, __X, __Y),
-                                      (__v2di)__W);
+  return (__m128i)__builtin_ia32_selectq_128(
+      __M, (__v2di)__builtin_ia32_vpmadd52huq128(__W, __X, __Y), (__v2di)__W);
 }
 
 static __inline__ __m128i __DEFAULT_FN_ATTRS128
 _mm_maskz_madd52hi_epu64 (__mmask8 __M, __m128i __X, __m128i __Y, __m128i __Z)
 {
-  return (__m128i)__builtin_ia32_selectq_128(__M,
-                                      (__v2di)_mm_madd52hi_epu64(__X, __Y, __Z),
-                                      (__v2di)_mm_setzero_si128());
+  return (__m128i)__builtin_ia32_selectq_128(
+      __M, (__v2di)__builtin_ia32_vpmadd52huq128(__X, __Y, __Z),
+      (__v2di)_mm_setzero_si128());
 }
 
 static __inline__ __m256i __DEFAULT_FN_ATTRS256
 _mm256_mask_madd52hi_epu64 (__m256i __W, __mmask8 __M, __m256i __X, __m256i __Y)
 {
-  return (__m256i)__builtin_ia32_selectq_256(__M,
-                                   (__v4di)_mm256_madd52hi_epu64(__W, __X, __Y),
-                                   (__v4di)__W);
+  return (__m256i)__builtin_ia32_selectq_256(
+      __M, (__v4di)__builtin_ia32_vpmadd52huq256(__W, __X, __Y), (__v4di)__W);
 }
 
 static __inline__ __m256i __DEFAULT_FN_ATTRS256
 _mm256_maskz_madd52hi_epu64 (__mmask8 __M, __m256i __X, __m256i __Y, __m256i __Z)
 {
-  return (__m256i)__builtin_ia32_selectq_256(__M,
-                                   (__v4di)_mm256_madd52hi_epu64(__X, __Y, __Z),
-                                   (__v4di)_mm256_setzero_si256());
+  return (__m256i)__builtin_ia32_selectq_256(
+      __M, (__v4di)__builtin_ia32_vpmadd52huq256(__X, __Y, __Z),
+      (__v4di)_mm256_setzero_si256());
 }
 
 static __inline__ __m128i __DEFAULT_FN_ATTRS128
 _mm_mask_madd52lo_epu64 (__m128i __W, __mmask8 __M, __m128i __X, __m128i __Y)
 {
-  return (__m128i)__builtin_ia32_selectq_128(__M,
-                                      (__v2di)_mm_madd52lo_epu64(__W, __X, __Y),
-                                      (__v2di)__W);
+  return (__m128i)__builtin_ia32_selectq_128(
+      __M, (__v2di)__builtin_ia32_vpmadd52luq128(__W, __X, __Y), (__v2di)__W);
 }
 
 static __inline__ __m128i __DEFAULT_FN_ATTRS128
 _mm_maskz_madd52lo_epu64 (__mmask8 __M, __m128i __X, __m128i __Y, __m128i __Z)
 {
-  return (__m128i)__builtin_ia32_selectq_128(__M,
-                                      (__v2di)_mm_madd52lo_epu64(__X, __Y, __Z),
-                                      (__v2di)_mm_setzero_si128());
+  return (__m128i)__builtin_ia32_selectq_128(
+      __M, (__v2di)__builtin_ia32_vpmadd52luq128(__X, __Y, __Z),
+      (__v2di)_mm_setzero_si128());
 }
 
 static __inline__ __m256i __DEFAULT_FN_ATTRS256
 _mm256_mask_madd52lo_epu64 (__m256i __W, __mmask8 __M, __m256i __X, __m256i __Y)
 {
-  return (__m256i)__builtin_ia32_selectq_256(__M,
-                                   (__v4di)_mm256_madd52lo_epu64(__W, __X, __Y),
-                                   (__v4di)__W);
+  return (__m256i)__builtin_ia32_selectq_256(
+      __M, (__v4di)__builtin_ia32_vpmadd52luq256(__W, __X, __Y), (__v4di)__W);
 }
 
 static __inline__ __m256i __DEFAULT_FN_ATTRS256
 _mm256_maskz_madd52lo_epu64 (__mmask8 __M, __m256i __X, __m256i __Y, __m256i __Z)
 {
-  return (__m256i)__builtin_ia32_selectq_256(__M,
-                                   (__v4di)_mm256_madd52lo_epu64(__X, __Y, __Z),
-                                   (__v4di)_mm256_setzero_si256());
+  return (__m256i)__builtin_ia32_selectq_256(
+      __M, (__v4di)__builtin_ia32_vpmadd52luq256(__X, __Y, __Z),
+      (__v4di)_mm256_setzero_si256());
 }
 
 
diff --git a/clang/lib/Headers/avxifmaintrin.h b/clang/lib/Headers/avxifmaintrin.h
index 5c782d2a5b865..c218b7ddb3960 100644
--- a/clang/lib/Headers/avxifmaintrin.h
+++ b/clang/lib/Headers/avxifmaintrin.h
@@ -22,6 +22,13 @@
   __attribute__((__always_inline__, __nodebug__, __target__("avxifma"),        \
                  __min_vector_width__(256)))
 
+#if !defined(__AVX512IFMA__) && defined(__AVXIFMA__)
+#define _mm_madd52hi_epu64(X, Y, Z) _mm_madd52hi_avx_epu64(X, Y, Z)
+#define _mm_madd52lo_epu64(X, Y, Z) _mm_madd52lo_avx_epu64(X, Y, Z)
+#define _mm256_madd52hi_epu64(X, Y, Z) _mm256_madd52hi_avx_epu64(X, Y, Z)
+#define _mm256_madd52lo_epu64(X, Y, Z) _mm256_madd52lo_avx_epu64(X, Y, Z)
+#endif
+
 // must vex-encoding
 
 /// Multiply packed unsigned 52-bit integers in each 64-bit element of \a __Y