[X86] Add the AVX512 exp2a23 intrinsics.

git-svn-id: https://llvm.org/svn/llvm-project/cfe/trunk@227769 91177308-0d34-0410-b5e6-96231b3b80d8

diff --git a/include/clang/Basic/BuiltinsX86.def b/include/clang/Basic/BuiltinsX86.def
index c5897124edb1cbc05c4f778b5f5fe884bc91e4fc..a246c2516910654c734fc3c4f1008fbdea1e15b6 100644 (file)
--- a/include/clang/Basic/BuiltinsX86.def
+++ b/include/clang/Basic/BuiltinsX86.def
@@ -838,6 +838,8 @@ BUILTIN(__builtin_ia32_rcp28sd_mask, "V2dV2dV2dV2dUcIi", "")
 BUILTIN(__builtin_ia32_rcp28ss_mask, "V4fV4fV4fV4fUcIi", "")
 BUILTIN(__builtin_ia32_rcp28pd_mask, "V8dV8dV8dUcIi", "")
 BUILTIN(__builtin_ia32_rcp28ps_mask, "V16fV16fV16fUsIi", "")
+BUILTIN(__builtin_ia32_exp2pd_mask, "V8dV8dV8dUcIi", "")
+BUILTIN(__builtin_ia32_exp2ps_mask, "V16fV16fV16fUsIi", "")
 BUILTIN(__builtin_ia32_cvttps2dq512_mask, "V16iV16fV16iUsIi", "")
 BUILTIN(__builtin_ia32_cvttps2udq512_mask, "V16iV16fV16iUsIi", "")
 BUILTIN(__builtin_ia32_cvttpd2dq512_mask, "V8iV8dV8iUcIi", "")

diff --git a/lib/Headers/avx512erintrin.h b/lib/Headers/avx512erintrin.h
index 323f4f0d6728e1718d6c67569bb2fe469c200386..57c61aa0e112ee4df7f8710988dfa2ecda5c9d4f 100644 (file)
--- a/lib/Headers/avx512erintrin.h
+++ b/lib/Headers/avx512erintrin.h
@@ -28,6 +28,55 @@
 #define __AVX512ERINTRIN_H
 
 
+// exp2a23
+#define _mm512_exp2a23_round_pd(A, R) __extension__ ({ \
+  (__m512d)__builtin_ia32_exp2pd_mask((__v8df)(__m512d)(A), \
+                                      (__v8df)_mm512_setzero_pd(), \
+                                      (__mmask8)-1, (R)); })
+
+#define _mm512_mask_exp2a23_round_pd(S, M, A, R) __extension__ ({ \
+  (__m512d)__builtin_ia32_exp2pd_mask((__v8df)(__m512d)(A), \
+                                      (__v8df)(__m512d)(S), \
+                                      (__mmask8)(M), (R)); })
+
+#define _mm512_maskz_exp2a23_round_pd(M, A, R) __extension__ ({ \
+  (__m512d)__builtin_ia32_exp2pd_mask((__v8df)(__m512d)(A), \
+                                      (__v8df)_mm512_setzero_pd(), \
+                                      (__mmask8)(M), (R)); })
+
+#define _mm512_exp2a23_pd(A) \
+   _mm512_exp2a23_round_pd((A), _MM_FROUND_CUR_DIRECTION)
+
+#define _mm512_mask_exp2a23_pd(S, M, A) \
+   _mm512_mask_exp2a23_round_pd((S), (M), (A), _MM_FROUND_CUR_DIRECTION)
+
+#define _mm512_maskz_exp2a23_pd(M, A) \
+   _mm512_maskz_exp2a23_round_pd((M), (A), _MM_FROUND_CUR_DIRECTION)
+
+#define _mm512_exp2a23_round_ps(A, R) __extension__ ({ \
+  (__m512)__builtin_ia32_exp2ps_mask((__v16sf)(__m512)(A), \
+                                     (__v16sf)_mm512_setzero_ps(), \
+                                     (__mmask8)-1, (R)); })
+
+#define _mm512_mask_exp2a23_round_ps(S, M, A, R) __extension__ ({ \
+  (__m512)__builtin_ia32_exp2ps_mask((__v16sf)(__m512)(A), \
+                                     (__v16sf)(__m512)(S), \
+                                     (__mmask8)(M), (R)); })
+
+#define _mm512_maskz_exp2a23_round_ps(M, A, R) __extension__ ({ \
+  (__m512)__builtin_ia32_exp2ps_mask((__v16sf)(__m512)(A), \
+                                     (__v16sf)_mm512_setzero_ps(), \
+                                     (__mmask8)(M), (R)); })
+
+#define _mm512_exp2a23_ps(A) \
+   _mm512_exp2a23_round_ps((A), _MM_FROUND_CUR_DIRECTION)
+
+#define _mm512_mask_exp2a23_ps(S, M, A) \
+   _mm512_mask_exp2a23_round_ps((S), (M), (A), _MM_FROUND_CUR_DIRECTION)
+
+#define _mm512_maskz_exp2a23_ps(M, A) \
+   _mm512_maskz_exp2a23_round_ps((M), (A), _MM_FROUND_CUR_DIRECTION)
+
 // rsqrt28
 #define _mm512_rsqrt28_round_pd(A, R) __extension__ ({ \
   (__m512d)__builtin_ia32_rsqrt28pd_mask((__v8df)(__m512d)(A), \

diff --git a/test/CodeGen/avx512er-builtins.c b/test/CodeGen/avx512er-builtins.c
index 4aa8c30acb23cbe58927129370b5296f655b0224..993f177a1168af589e52c13fd44524e396e50a4f 100644 (file)
--- a/test/CodeGen/avx512er-builtins.c
+++ b/test/CodeGen/avx512er-builtins.c
@@ -271,3 +271,76 @@ __m128d test_mm_maskz_rcp28_sd(__mmask8 m, __m128d a, __m128d b) {
   // CHECK: @llvm.x86.avx512.rcp28.sd
   return _mm_maskz_rcp28_sd(m, a, b);
 }
+
+__m512d test_mm512_exp2a23_round_pd(__m512d a) {
+  // CHECK-LABEL: @test_mm512_exp2a23_round_pd
+  // CHECK: @llvm.x86.avx512.exp2.pd
+  return _mm512_exp2a23_round_pd(a, _MM_FROUND_TO_NEAREST_INT);
+}
+
+__m512d test_mm512_mask_exp2a23_round_pd(__m512d s, __mmask8 m, __m512d a) {
+  // check-label: @test_mm512_mask_exp2a23_round_pd
+  // check: @llvm.x86.avx512.exp2.pd
+  return _mm512_mask_exp2a23_round_pd(s, m, a, _MM_FROUND_TO_NEAREST_INT);
+}
+
+__m512d test_mm512_maskz_exp2a23_round_pd(__mmask8 m, __m512d a) {
+  // check-label: @test_mm512_maskz_exp2a23_round_pd
+  // check: @llvm.x86.avx512.exp2.pd
+  return _mm512_maskz_exp2a23_round_pd(m, a, _MM_FROUND_TO_NEAREST_INT);
+}
+
+__m512d test_mm512_exp2a23_pd(__m512d a) {
+  // CHECK-LABEL: @test_mm512_exp2a23_pd
+  // CHECK: @llvm.x86.avx512.exp2.pd
+  return _mm512_exp2a23_pd(a);
+}
+
+__m512d test_mm512_mask_exp2a23_pd(__m512d s, __mmask8 m, __m512d a) {
+  // check-label: @test_mm512_mask_exp2a23_pd
+  // check: @llvm.x86.avx512.exp2.pd
+  return _mm512_mask_exp2a23_pd(s, m, a);
+}
+
+__m512d test_mm512_maskz_exp2a23_pd(__mmask8 m, __m512d a) {
+  // check-label: @test_mm512_maskz_exp2a23_pd
+  // check: @llvm.x86.avx512.exp2.pd
+  return _mm512_maskz_exp2a23_pd(m, a);
+}
+
+__m512 test_mm512_exp2a23_round_ps(__m512 a) {
+  // CHECK-LABEL: @test_mm512_exp2a23_round_ps
+  // CHECK: @llvm.x86.avx512.exp2.ps
+  return _mm512_exp2a23_round_ps(a, _MM_FROUND_TO_NEAREST_INT);
+}
+
+__m512 test_mm512_mask_exp2a23_round_ps(__m512 s, __mmask16 m, __m512 a) {
+  // CHECK-LABEL: @test_mm512_mask_exp2a23_round_ps
+  // CHECK: @llvm.x86.avx512.exp2.ps
+  return _mm512_mask_exp2a23_round_ps(s, m, a, _MM_FROUND_TO_NEAREST_INT);
+}
+
+__m512 test_mm512_maskz_exp2a23_round_ps(__mmask16 m, __m512 a) {
+  // CHECK-LABEL: @test_mm512_maskz_exp2a23_round_ps
+  // CHECK: @llvm.x86.avx512.exp2.ps
+  return _mm512_maskz_exp2a23_round_ps(m, a, _MM_FROUND_TO_NEAREST_INT);
+}
+
+__m512 test_mm512_exp2a23_ps(__m512 a) {
+  // CHECK-LABEL: @test_mm512_exp2a23_ps
+  // CHECK: @llvm.x86.avx512.exp2.ps
+  return _mm512_exp2a23_ps(a);
+}
+
+__m512 test_mm512_mask_exp2a23_ps(__m512 s, __mmask16 m, __m512 a) {
+  // CHECK-LABEL: @test_mm512_mask_exp2a23_ps
+  // CHECK: @llvm.x86.avx512.exp2.ps
+  return _mm512_mask_exp2a23_ps(s, m, a);
+}
+
+__m512 test_mm512_maskz_exp2a23_ps(__mmask16 m, __m512 a) {
+  // CHECK-LABEL: @test_mm512_maskz_exp2a23_ps
+  // CHECK: @llvm.x86.avx512.exp2.ps
+  return _mm512_maskz_exp2a23_ps(m, a);
+}
+