From: Craig Topper <craig.topper@gmail.com>
Date: Sun, 10 Jun 2012 07:47:32 +0000 (+0000)
Subject: Add XOP vprot* instruction intrinsics
X-Git-Url: https://granicus.if.org/sourcecode?a=commitdiff_plain;h=3335eb7f7c5b82f109c585fb4009ff8240b11a84;p=clang

Add XOP vprot* instruction intrinsics

git-svn-id: https://llvm.org/svn/llvm-project/cfe/trunk@158292 91177308-0d34-0410-b5e6-96231b3b80d8
---

diff --git a/include/clang/Basic/BuiltinsX86.def b/include/clang/Basic/BuiltinsX86.def
index 35739e9649..5d1175fa93 100644
--- a/include/clang/Basic/BuiltinsX86.def
+++ b/include/clang/Basic/BuiltinsX86.def
@@ -676,4 +676,13 @@ BUILTIN(__builtin_ia32_vpcmov_256, "V4LLiV4LLiV4LLiV4LLi", "")
 
 BUILTIN(__builtin_ia32_vpperm, "V16cV16cV16cV16c", "")
 
+BUILTIN(__builtin_ia32_vprotb, "V16cV16cV16c", "")
+BUILTIN(__builtin_ia32_vprotw, "V8sV8sV8s", "")
+BUILTIN(__builtin_ia32_vprotd, "V4iV4iV4i", "")
+BUILTIN(__builtin_ia32_vprotq, "V2LLiV2LLiV2LLi", "")
+BUILTIN(__builtin_ia32_vprotbi, "V16cV16cIc", "")
+BUILTIN(__builtin_ia32_vprotwi, "V8sV8sIc", "")
+BUILTIN(__builtin_ia32_vprotdi, "V4iV4iIc", "")
+BUILTIN(__builtin_ia32_vprotqi, "V2LLiV2LLiIc", "")
+
 #undef BUILTIN
diff --git a/lib/Headers/xopintrin.h b/lib/Headers/xopintrin.h
index 099fca995f..108f123ca5 100644
--- a/lib/Headers/xopintrin.h
+++ b/lib/Headers/xopintrin.h
@@ -214,6 +214,46 @@ _mm_perm_epi8(__m128i __A, __m128i __B, __m128i __C)
   return (__m128i)__builtin_ia32_vpperm((__v16qi)__A, (__v16qi)__B, (__v16qi)__C);
 }
 
+static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
+_mm_rot_epi8(__m128i __A, __m128 __B)
+{
+  return (__m128i)__builtin_ia32_vprotb((__v16qi)__A, (__v16qi)__B);
+}
+
+static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
+_mm_rot_epi16(__m128i __A, __m128 __B)
+{
+  return (__m128i)__builtin_ia32_vprotw((__v8hi)__A, (__v8hi)__B);
+}
+
+static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
+_mm_rot_epi32(__m128i __A, __m128 __B)
+{
+  return (__m128i)__builtin_ia32_vprotd((__v4si)__A, (__v4si)__B);
+}
+
+static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
+_mm_rot_epi64(__m128i __A, __m128 __B)
+{
+  return (__m128i)__builtin_ia32_vprotq((__v2di)__A, (__v2di)__B);
+}
+
+#define _mm_roti_epi8(A, N) __extension__ ({ \
+  __m128i __A = (A); \
+  (__m128i)__builtin_ia32_vprotbi((__v16qi)__A, (N)); })
+
+#define _mm_roti_epi16(A, N) __extension__ ({ \
+  __m128i __A = (A); \
+  (__m128i)__builtin_ia32_vprotwi((__v8hi)__A, (N)); })
+
+#define _mm_roti_epi32(A, N) __extension__ ({ \
+  __m128i __A = (A); \
+  (__m128i)__builtin_ia32_vprotdi((__v4si)__A, (N)); })
+
+#define _mm_roti_epi64(A, N) __extension__ ({ \
+  __m128i __A = (A); \
+  (__m128i)__builtin_ia32_vprotqi((__v2di)__A, (N)); })
+
 #endif /* __XOP__ */
 
 #endif /* __XOPINTRIN_H */
diff --git a/test/CodeGen/xop-builtins.c b/test/CodeGen/xop-builtins.c
index 28708ab9cf..af7734809f 100644
--- a/test/CodeGen/xop-builtins.c
+++ b/test/CodeGen/xop-builtins.c
@@ -154,3 +154,43 @@ __m128i test_mm_perm_epi8(__m128i a, __m128i b, __m128i c) {
   // CHECK: @llvm.x86.xop.vpperm
   return _mm_perm_epi8(a, b, c);
 }
+
+__m128i test_mm_rot_epi8(__m128i a, __m128i b) {
+  // CHECK: @llvm.x86.xop.vprotb
+  return _mm_rot_epi8(a, b);
+}
+
+__m128i test_mm_rot_epi16(__m128i a, __m128i b) {
+  // CHECK: @llvm.x86.xop.vprotw
+  return _mm_rot_epi16(a, b);
+}
+
+__m128i test_mm_rot_epi32(__m128i a, __m128i b) {
+  // CHECK: @llvm.x86.xop.vprotd
+  return _mm_rot_epi32(a, b);
+}
+
+__m128i test_mm_rot_epi64(__m128i a, __m128i b) {
+  // CHECK: @llvm.x86.xop.vprotq
+  return _mm_rot_epi64(a, b);
+}
+
+__m128i test_mm_roti_epi8(__m128i a) {
+  // CHECK: @llvm.x86.xop.vprotbi
+  return _mm_roti_epi8(a, 1);
+}
+
+__m128i test_mm_roti_epi16(__m128i a) {
+  // CHECK: @llvm.x86.xop.vprotwi
+  return _mm_roti_epi16(a, 50);
+}
+
+__m128i test_mm_roti_epi32(__m128i a) {
+  // CHECK: @llvm.x86.xop.vprotdi
+  return _mm_roti_epi32(a, -30);
+}
+
+__m128i test_mm_roti_epi64(__m128i a) {
+  // CHECK: @llvm.x86.xop.vprotqi
+  return _mm_roti_epi64(a, 100);
+}