From 8935beba49687044fe6f7ad510360e0ab1e6183f Mon Sep 17 00:00:00 2001
From: Adam Nemet <anemet@apple.com>
Date: Mon, 28 Jul 2014 17:14:45 +0000
Subject: [PATCH] [AVX512] Add non-masking FP store intrinsics

Part of <rdar://problem/17688758>

git-svn-id: https://llvm.org/svn/llvm-project/cfe/trunk@214099 91177308-0d34-0410-b5e6-96231b3b80d8
---
 lib/Headers/avx512fintrin.h     | 24 ++++++++++++++++++++++++
 test/CodeGen/avx512f-builtins.c | 28 ++++++++++++++++++++++++++++
 2 files changed, 52 insertions(+)

diff --git a/lib/Headers/avx512fintrin.h b/lib/Headers/avx512fintrin.h
index 785d5281bc..cebb0d5a05 100644
--- a/lib/Headers/avx512fintrin.h
+++ b/lib/Headers/avx512fintrin.h
@@ -760,6 +760,12 @@ _mm512_mask_storeu_pd(void *__P, __mmask8 __U, __m512d __A)
   __builtin_ia32_storeupd512_mask ((__v8df *)__P, (__v8df) __A, (__mmask8) __U);
 }
 
+static __inline void __attribute__ ((__always_inline__, __nodebug__))
+_mm512_storeu_pd(void *__P, __m512d __A)
+{
+  __builtin_ia32_storeupd512_mask((__v8df *)__P, (__v8df)__A, (__mmask8)-1);
+}
+
 static __inline void __attribute__ ((__always_inline__, __nodebug__))
 _mm512_mask_storeu_ps(void *__P, __mmask16 __U, __m512 __A)
 {
@@ -767,4 +773,22 @@ _mm512_mask_storeu_ps(void *__P, __mmask16 __U, __m512 __A)
                                    (__mmask16) __U);
 }
 
+static __inline void __attribute__ ((__always_inline__, __nodebug__))
+_mm512_storeu_ps(void *__P, __m512 __A)
+{
+  __builtin_ia32_storeups512_mask((__v16sf *)__P, (__v16sf)__A, (__mmask16)-1);
+}
+
+static __inline void __attribute__ ((__always_inline__, __nodebug__))
+_mm512_store_ps(void *__P, __m512 __A)
+{
+  *(__m512*)__P = __A;
+}
+
+static __inline void __attribute__ ((__always_inline__, __nodebug__))
+_mm512_store_pd(void *__P, __m512d __A)
+{
+  *(__m512d*)__P = __A;
+}
+
 #endif // __AVX512FINTRIN_H
diff --git a/test/CodeGen/avx512f-builtins.c b/test/CodeGen/avx512f-builtins.c
index 968c015c87..ea84583734 100644
--- a/test/CodeGen/avx512f-builtins.c
+++ b/test/CodeGen/avx512f-builtins.c
@@ -60,3 +60,31 @@ __m512d test_mm512_mul_pd(__m512d a, __m512d b)
   // CHECK: fmul <8 x double>
   return _mm512_mul_pd(a, b);
 }
+
+void test_mm512_storeu_ps(void *p, __m512 a)
+{
+  // CHECK-LABEL: @test_mm512_storeu_ps
+  // CHECK: @llvm.x86.avx512.mask.storeu.ps.512
+  _mm512_storeu_ps(p, a);
+}
+
+void test_mm512_storeu_pd(void *p, __m512d a)
+{
+  // CHECK-LABEL: @test_mm512_storeu_pd
+  // CHECK: @llvm.x86.avx512.mask.storeu.pd.512
+  _mm512_storeu_pd(p, a);
+}
+
+void test_mm512_store_ps(void *p, __m512 a)
+{
+  // CHECK-LABEL: @test_mm512_store_ps
+  // CHECK: store <16 x float>
+  _mm512_store_ps(p, a);
+}
+
+void test_mm512_store_pd(void *p, __m512d a)
+{
+  // CHECK-LABEL: @test_mm512_store_pd
+  // CHECK: store <8 x double>
+  _mm512_store_pd(p, a);
+}
-- 
2.40.0