From: Andrei Tatar <andrei@unikraft.io>
Date: Mon, 17 Jul 2023 19:06:27 +0000 (+0200)
Subject: Add compatibility with gcc <= 12
X-Git-Tag: RELEASE-0.14.0~6
X-Git-Url: http://xenbits.xensource.com/gitweb?a=commitdiff_plain;h=ee86c1d8503929eedf9a96b7b87ffec0fd140cdf;p=unikraft%2Flibs%2Fintel-intrinsics.git

Add compatibility with gcc <= 12

Implementations based off headers in GCC release 12.3.0.

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-gcc/avx512bf16intrin.h b/include-gcc/avx512bf16intrin.h
index 107f4a4..cfe7acc 100644
--- a/include-gcc/avx512bf16intrin.h
+++ b/include-gcc/avx512bf16intrin.h
@@ -34,6 +34,24 @@
 #define __DISABLE_AVX512BF16__
 #endif /* __AVX512BF16__ */
 
+#if (__GNUC__ < 13)
+/* Internal data types for implementing the intrinsics.  */
+typedef short __v32bh __attribute__ ((__vector_size__ (64)));
+
+/* The Intel API is flexible enough that we must allow aliasing with other
+   vector types, and their scalar components.  */
+typedef short __m512bh __attribute__ ((__vector_size__ (64), __may_alias__));
+
+/* Convert One BF16 Data to One Single Float Data.  */
+extern __inline float
+__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
+_mm_cvtsbh_ss (__bfloat16 __A)
+{
+  union{ float __a; unsigned int __b;} __tmp;
+  __tmp.__b = ((unsigned int)(__A)) << 16;
+  return __tmp.__a;
+}
+#else
 /* Internal data types for implementing the intrinsics.  */
 typedef __bf16 __v32bf __attribute__ ((__vector_size__ (64)));
 
@@ -48,6 +66,7 @@ _mm_cvtsbh_ss (__bf16 __A)
 {
   return __builtin_ia32_cvtbf2sf (__A);
 }
+#endif
 
 /* vcvtne2ps2bf16 */
 
@@ -55,21 +74,33 @@ extern __inline __m512bh
 __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 _mm512_cvtne2ps_pbh (__m512 __A, __m512 __B)
 {
+#if (__GNUC__ < 13)
+  return (__m512bh)__builtin_ia32_cvtne2ps2bf16_v32hi(__A, __B);
+#else
   return (__m512bh)__builtin_ia32_cvtne2ps2bf16_v32bf(__A, __B);
+#endif
 }
 
 extern __inline __m512bh
 __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 _mm512_mask_cvtne2ps_pbh (__m512bh __A, __mmask32 __B, __m512 __C, __m512 __D)
 {
+#if (__GNUC__ < 13)
+  return (__m512bh)__builtin_ia32_cvtne2ps2bf16_v32hi_mask(__C, __D, __A, __B);
+#else
   return (__m512bh)__builtin_ia32_cvtne2ps2bf16_v32bf_mask(__C, __D, __A, __B);
+#endif
 }
 
 extern __inline __m512bh
 __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 _mm512_maskz_cvtne2ps_pbh (__mmask32 __A, __m512 __B, __m512 __C)
 {
+#if (__GNUC__ < 13)
+  return (__m512bh)__builtin_ia32_cvtne2ps2bf16_v32hi_maskz(__B, __C, __A);
+#else
   return (__m512bh)__builtin_ia32_cvtne2ps2bf16_v32bf_maskz(__B, __C, __A);
+#endif
 }
 
 /* vcvtneps2bf16 */
diff --git a/include-gcc/avx512bf16vlintrin.h b/include-gcc/avx512bf16vlintrin.h
index 6e8a6a0..861931a 100644
--- a/include-gcc/avx512bf16vlintrin.h
+++ b/include-gcc/avx512bf16vlintrin.h
@@ -34,6 +34,18 @@
 #define __DISABLE_AVX512BF16VL__
 #endif /* __AVX512BF16__ */
 
+#if (__GNUC__ < 13)
+/* Internal data types for implementing the intrinsics.  */
+typedef short __v16bh __attribute__ ((__vector_size__ (32)));
+typedef short __v8bh __attribute__ ((__vector_size__ (16)));
+
+/* The Intel API is flexible enough that we must allow aliasing with other
+   vector types, and their scalar components.  */
+typedef short __m256bh __attribute__ ((__vector_size__ (32), __may_alias__));
+typedef short __m128bh __attribute__ ((__vector_size__ (16), __may_alias__));
+
+typedef unsigned short __bfloat16;
+#else
 /* Internal data types for implementing the intrinsics.  */
 typedef __bf16 __v16bf __attribute__ ((__vector_size__ (32)));
 typedef __bf16 __v8bf __attribute__ ((__vector_size__ (16)));
@@ -44,6 +56,7 @@ typedef __bf16 __m256bh __attribute__ ((__vector_size__ (32), __may_alias__));
 typedef __bf16 __m128bh __attribute__ ((__vector_size__ (16), __may_alias__));
 
 typedef __bf16 __bfloat16;
+#endif
 
 #define _mm256_cvtneps_pbh(A) \
   (__m128bh) __builtin_ia32_cvtneps2bf16_v8sf (A)
@@ -56,42 +69,66 @@ extern __inline __m256bh
 __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 _mm256_cvtne2ps_pbh (__m256 __A, __m256 __B)
 {
+#if (__GNUC__ < 13)
+  return (__m256bh)__builtin_ia32_cvtne2ps2bf16_v16hi(__A, __B);
+#else
   return (__m256bh)__builtin_ia32_cvtne2ps2bf16_v16bf(__A, __B);
+#endif
 }
 
 extern __inline __m256bh
 __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 _mm256_mask_cvtne2ps_pbh (__m256bh __A, __mmask16 __B, __m256 __C, __m256 __D)
 {
+#if (__GNUC__ < 13)
+  return (__m256bh)__builtin_ia32_cvtne2ps2bf16_v16hi_mask(__C, __D, __A, __B);
+#else
   return (__m256bh)__builtin_ia32_cvtne2ps2bf16_v16bf_mask(__C, __D, __A, __B);
+#endif
 }
 
 extern __inline __m256bh
 __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 _mm256_maskz_cvtne2ps_pbh (__mmask16 __A, __m256 __B, __m256 __C)
 {
+#if (__GNUC__ < 13)
+  return (__m256bh)__builtin_ia32_cvtne2ps2bf16_v16hi_maskz(__B, __C, __A);
+#else
   return (__m256bh)__builtin_ia32_cvtne2ps2bf16_v16bf_maskz(__B, __C, __A);
+#endif
 }
 
 extern __inline __m128bh
 __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 _mm_cvtne2ps_pbh (__m128 __A, __m128 __B)
 {
+#if (__GNUC__ < 13)
+  return (__m128bh)__builtin_ia32_cvtne2ps2bf16_v8hi(__A, __B);
+#else
   return (__m128bh)__builtin_ia32_cvtne2ps2bf16_v8bf(__A, __B);
+#endif
 }
 
 extern __inline __m128bh
 __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 _mm_mask_cvtne2ps_pbh (__m128bh __A, __mmask8 __B, __m128 __C, __m128 __D)
 {
+#if (__GNUC__ < 13)
+  return (__m128bh)__builtin_ia32_cvtne2ps2bf16_v8hi_mask(__C, __D, __A, __B);
+#else
   return (__m128bh)__builtin_ia32_cvtne2ps2bf16_v8bf_mask(__C, __D, __A, __B);
+#endif
 }
 
 extern __inline __m128bh
 __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 _mm_maskz_cvtne2ps_pbh (__mmask8 __A, __m128 __B, __m128 __C)
 {
+#if (__GNUC__ < 13)
+  return (__m128bh)__builtin_ia32_cvtne2ps2bf16_v8hi_maskz(__B, __C, __A);
+#else
   return (__m128bh)__builtin_ia32_cvtne2ps2bf16_v8bf_maskz(__B, __C, __A);
+#endif
 }
 
 /* vcvtneps2bf16 */
@@ -168,13 +205,18 @@ _mm_maskz_dpbf16_ps (__mmask8 __A, __m128 __B, __m128bh __C, __m128bh __D)
   return (__m128)__builtin_ia32_dpbf16ps_v4sf_maskz(__B, __C, __D, __A);
 }
 
-extern __inline __bf16
+extern __inline __bfloat16
 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
 _mm_cvtness_sbh (float __A)
 {
   __v4sf __V = {__A, 0, 0, 0};
+#if (__GNUC__ < 13)
+  __v8hi __R = __builtin_ia32_cvtneps2bf16_v4sf_mask ((__v4sf)__V,
+	       (__v8hi)_mm_undefined_si128 (), (__mmask8)-1);
+#else
   __v8bf __R = __builtin_ia32_cvtneps2bf16_v4sf_mask ((__v4sf)__V,
 	       (__v8bf)_mm_undefined_si128 (), (__mmask8)-1);
+#endif
   return __R[0];
 }
 
diff --git a/include-gcc/avx512fp16intrin.h b/include-gcc/avx512fp16intrin.h
index dd083e5..31de9ba 100644
--- a/include-gcc/avx512fp16intrin.h
+++ b/include-gcc/avx512fp16intrin.h
@@ -183,21 +183,33 @@ extern __inline __m128h
 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
 _mm_setzero_ph (void)
 {
+#if (__GNUC__ < 13)
+  return _mm_set1_ph (0.0f);
+#else
   return _mm_set1_ph (0.0f16);
+#endif
 }
 
 extern __inline __m256h
 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
 _mm256_setzero_ph (void)
 {
+#if (__GNUC__ < 13)
+  return _mm256_set1_ph (0.0f);
+#else
   return _mm256_set1_ph (0.0f16);
+#endif
 }
 
 extern __inline __m512h
 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
 _mm512_setzero_ph (void)
 {
+#if (__GNUC__ < 13)
+  return _mm512_set1_ph (0.0f);
+#else
   return _mm512_set1_ph (0.0f16);
+#endif
 }
 
 extern __inline __m128h
@@ -367,8 +379,12 @@ extern __inline __m128h
 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
 _mm_set_sh (_Float16 __F)
 {
+#if (__GNUC__ < 13)
+  return _mm_set_ph (0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, __F);
+#else
   return _mm_set_ph (0.0f16, 0.0f16, 0.0f16, 0.0f16, 0.0f16, 0.0f16, 0.0f16,
 		     __F);
+#endif
 }
 
 /* Create a vector with element 0 as *P and the rest zero.  */
@@ -376,8 +392,13 @@ extern __inline __m128h
 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
 _mm_load_sh (void const *__P)
 {
+#if (__GNUC__ < 13)
+  return _mm_set_ph (0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f,
+		     *(_Float16 const *) __P);
+#else
   return _mm_set_ph (0.0f16, 0.0f16, 0.0f16, 0.0f16, 0.0f16, 0.0f16, 0.0f16,
 		     *(_Float16 const *) __P);
+#endif
 }
 
 extern __inline __m512h
diff --git a/include-gcc/immintrin.h b/include-gcc/immintrin.h
index b220d87..ca1c509 100644
--- a/include-gcc/immintrin.h
+++ b/include-gcc/immintrin.h
@@ -44,9 +44,11 @@
 
 #include <avxvnniintrin.h>
 
+#if (__GNUC__ >= 13)
 #include <avxifmaintrin.h>
 
 #include <avxvnniint8intrin.h>
+#endif
 
 #include <avx2intrin.h>
 
@@ -123,8 +125,10 @@
 
 #include <avx512bf16intrin.h>
 
+#if (__GNUC__ >= 13)
 #include <avxneconvertintrin.h>
 #endif
+#endif
 
 #include <amxtileintrin.h>
 
@@ -132,12 +136,16 @@
 
 #include <amxbf16intrin.h>
 
+#if (__GNUC__ >= 13)
 #include <amxcomplexintrin.h>
+#endif
 
 #include <prfchwintrin.h>
 
 #include <keylockerintrin.h>
 
+#if (__GNUC__ >= 13)
 #include <amxfp16intrin.h>
+#endif
 
 #endif /* _IMMINTRIN_H_INCLUDED */
diff --git a/include-gcc/x86gprintrin.h b/include-gcc/x86gprintrin.h
index f41be3f..6719e62 100644
--- a/include-gcc/x86gprintrin.h
+++ b/include-gcc/x86gprintrin.h
@@ -52,7 +52,9 @@
 
 #include <clzerointrin.h>
 
+#if (__GNUC__ >= 13)
 #include <cmpccxaddintrin.h>
+#endif
 
 #include <enqcmdintrin.h>
 
@@ -74,9 +76,11 @@
 
 #include <pkuintrin.h>
 
+#if (__GNUC__ >= 13)
 #include <prfchiintrin.h>
 
 #include <raointintrin.h>
+#endif
 
 #include <rdseedintrin.h>
 
diff --git a/include-gcc/xmmintrin.h b/include-gcc/xmmintrin.h
index cb518fc..1885c13 100644
--- a/include-gcc/xmmintrin.h
+++ b/include-gcc/xmmintrin.h
@@ -53,13 +53,22 @@ enum _mm_hint
 extern __inline void __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 _mm_prefetch (const void *__P, enum _mm_hint __I)
 {
+#if (__GNUC__ < 13)
+  __builtin_prefetch (__P, (__I & 0x4) >> 2, __I & 0x3);
+#else
   __builtin_ia32_prefetch (__P, (__I & 0x4) >> 2,
 			   __I & 0x3, (__I & 0x10) >> 4);
+#endif
 }
 #else
+#if (__GNUC__ < 13)
+#define _mm_prefetch(P, I) \
+  __builtin_prefetch ((P), ((I & 0x4) >> 2), (I & 0x3))
+#else
 #define _mm_prefetch(P, I) \
   __builtin_ia32_prefetch ((P), ((I) & 0x4) >> 2, ((I) & 0x3), ((I) & 0x10) >> 4)
 #endif
+#endif
 
 #ifndef __SSE__
 #pragma GCC push_options