1*67e74705SXin Li // RUN: %clang_cc1 %s -triple=x86_64-apple-darwin -target-feature +f16c -emit-llvm -o - -Werror | FileCheck %s
2*67e74705SXin Li
3*67e74705SXin Li // Don't include mm_malloc.h, it's system specific.
4*67e74705SXin Li #define __MM_MALLOC_H
5*67e74705SXin Li
6*67e74705SXin Li #include <x86intrin.h>
7*67e74705SXin Li
test_cvtsh_ss(unsigned short a)8*67e74705SXin Li float test_cvtsh_ss(unsigned short a) {
9*67e74705SXin Li // CHECK-LABEL: test_cvtsh_ss
10*67e74705SXin Li // CHECK: insertelement <8 x i16> undef, i16 %{{.*}}, i32 0
11*67e74705SXin Li // CHECK: insertelement <8 x i16> %{{.*}}, i16 0, i32 1
12*67e74705SXin Li // CHECK: insertelement <8 x i16> %{{.*}}, i16 0, i32 2
13*67e74705SXin Li // CHECK: insertelement <8 x i16> %{{.*}}, i16 0, i32 3
14*67e74705SXin Li // CHECK: insertelement <8 x i16> %{{.*}}, i16 0, i32 4
15*67e74705SXin Li // CHECK: insertelement <8 x i16> %{{.*}}, i16 0, i32 5
16*67e74705SXin Li // CHECK: insertelement <8 x i16> %{{.*}}, i16 0, i32 6
17*67e74705SXin Li // CHECK: insertelement <8 x i16> %{{.*}}, i16 0, i32 7
18*67e74705SXin Li // CHECK: call <4 x float> @llvm.x86.vcvtph2ps.128(<8 x i16> %{{.*}})
19*67e74705SXin Li // CHECK: extractelement <4 x float> %{{.*}}, i32 0
20*67e74705SXin Li return _cvtsh_ss(a);
21*67e74705SXin Li }
22*67e74705SXin Li
test_cvtss_sh(float a)23*67e74705SXin Li unsigned short test_cvtss_sh(float a) {
24*67e74705SXin Li // CHECK-LABEL: test_cvtss_sh
25*67e74705SXin Li // CHECK: insertelement <4 x float> undef, float %{{.*}}, i32 0
26*67e74705SXin Li // CHECK: insertelement <4 x float> %{{.*}}, float 0.000000e+00, i32 1
27*67e74705SXin Li // CHECK: insertelement <4 x float> %{{.*}}, float 0.000000e+00, i32 2
28*67e74705SXin Li // CHECK: insertelement <4 x float> %{{.*}}, float 0.000000e+00, i32 3
29*67e74705SXin Li // CHECK: call <8 x i16> @llvm.x86.vcvtps2ph.128(<4 x float> %{{.*}}, i32 0)
30*67e74705SXin Li // CHECK: extractelement <8 x i16> %{{.*}}, i32 0
31*67e74705SXin Li return _cvtss_sh(a, 0);
32*67e74705SXin Li }
33*67e74705SXin Li
test_mm_cvtph_ps(__m128i a)34*67e74705SXin Li __m128 test_mm_cvtph_ps(__m128i a) {
35*67e74705SXin Li // CHECK-LABEL: test_mm_cvtph_ps
36*67e74705SXin Li // CHECK: call <4 x float> @llvm.x86.vcvtph2ps.128(<8 x i16> %{{.*}})
37*67e74705SXin Li return _mm_cvtph_ps(a);
38*67e74705SXin Li }
39*67e74705SXin Li
test_mm256_cvtph_ps(__m128i a)40*67e74705SXin Li __m256 test_mm256_cvtph_ps(__m128i a) {
41*67e74705SXin Li // CHECK-LABEL: test_mm256_cvtph_ps
42*67e74705SXin Li // CHECK: call <8 x float> @llvm.x86.vcvtph2ps.256(<8 x i16> %{{.*}})
43*67e74705SXin Li return _mm256_cvtph_ps(a);
44*67e74705SXin Li }
45*67e74705SXin Li
test_mm_cvtps_ph(__m128 a)46*67e74705SXin Li __m128i test_mm_cvtps_ph(__m128 a) {
47*67e74705SXin Li // CHECK-LABEL: test_mm_cvtps_ph
48*67e74705SXin Li // CHECK: call <8 x i16> @llvm.x86.vcvtps2ph.128(<4 x float> %{{.*}}, i32 0)
49*67e74705SXin Li return _mm_cvtps_ph(a, 0);
50*67e74705SXin Li }
51*67e74705SXin Li
test_mm256_cvtps_ph(__m256 a)52*67e74705SXin Li __m128i test_mm256_cvtps_ph(__m256 a) {
53*67e74705SXin Li // CHECK-LABEL: test_mm256_cvtps_ph
54*67e74705SXin Li // CHECK: call <8 x i16> @llvm.x86.vcvtps2ph.256(<8 x float> %{{.*}}, i32 0)
55*67e74705SXin Li return _mm256_cvtps_ph(a, 0);
56*67e74705SXin Li }
57