1; RUN: opt -mtriple=nvptx64-nvidia-cuda -load-store-vectorizer -S -o - %s | FileCheck %s
2
3define void @ldg_f16(half* nocapture align 16 %rd0) {
4  %in1b = bitcast half* %rd0 to <2 x half>*
5  %load1 = load <2 x half>, <2 x half>* %in1b, align 4
6  %p1 = fcmp ogt <2 x half> %load1, zeroinitializer
7  %s1 = select <2 x i1> %p1, <2 x half> %load1, <2 x half> zeroinitializer
8  store <2 x half> %s1, <2 x half>* %in1b, align 4
9  %in2 = getelementptr half, half* %rd0, i64 2
10  %in2b = bitcast half* %in2 to <2 x half>*
11  %load2 = load <2 x half>, <2 x half>* %in2b, align 4
12  %p2 = fcmp ogt <2 x half> %load2, zeroinitializer
13  %s2 = select <2 x i1> %p2, <2 x half> %load2, <2 x half> zeroinitializer
14  store <2 x half> %s2, <2 x half>* %in2b, align 4
15  %in3 = getelementptr half, half* %rd0, i64 4
16  %in3b = bitcast half* %in3 to <2 x half>*
17  %load3 = load <2 x half>, <2 x half>* %in3b, align 4
18  %p3 = fcmp ogt <2 x half> %load3, zeroinitializer
19  %s3 = select <2 x i1> %p3, <2 x half> %load3, <2 x half> zeroinitializer
20  store <2 x half> %s3, <2 x half>* %in3b, align 4
21  %in4 = getelementptr half, half* %rd0, i64 6
22  %in4b = bitcast half* %in4 to <2 x half>*
23  %load4 = load <2 x half>, <2 x half>* %in4b, align 4
24  %p4 = fcmp ogt <2 x half> %load4, zeroinitializer
25  %s4 = select <2 x i1> %p4, <2 x half> %load4, <2 x half> zeroinitializer
26  store <2 x half> %s4, <2 x half>* %in4b, align 4
27  ret void
28
29; CHECK-LABEL: @ldg_f16
30; CHECK: %[[LD:.*]] = load <8 x half>, <8 x half>*
31; CHECK: shufflevector <8 x half> %[[LD]], <8 x half> poison, <2 x i32> <i32 0, i32 1>
32; CHECK: shufflevector <8 x half> %[[LD]], <8 x half> poison, <2 x i32> <i32 2, i32 3>
33; CHECK: shufflevector <8 x half> %[[LD]], <8 x half> poison, <2 x i32> <i32 4, i32 5>
34; CHECK: shufflevector <8 x half> %[[LD]], <8 x half> poison, <2 x i32> <i32 6, i32 7>
35; CHECK: store <8 x half>
36}
37
38define void @no_nonpow2_vector(half* nocapture align 16 %rd0) {
39  %in1b = bitcast half* %rd0 to <3 x half>*
40  %load1 = load <3 x half>, <3 x half>* %in1b, align 4
41  %p1 = fcmp ogt <3 x half> %load1, zeroinitializer
42  %s1 = select <3 x i1> %p1, <3 x half> %load1, <3 x half> zeroinitializer
43  store <3 x half> %s1, <3 x half>* %in1b, align 4
44  %in2 = getelementptr half, half* %rd0, i64 3
45  %in2b = bitcast half* %in2 to <3 x half>*
46  %load2 = load <3 x half>, <3 x half>* %in2b, align 4
47  %p2 = fcmp ogt <3 x half> %load2, zeroinitializer
48  %s2 = select <3 x i1> %p2, <3 x half> %load2, <3 x half> zeroinitializer
49  store <3 x half> %s2, <3 x half>* %in2b, align 4
50  %in3 = getelementptr half, half* %rd0, i64 6
51  %in3b = bitcast half* %in3 to <3 x half>*
52  %load3 = load <3 x half>, <3 x half>* %in3b, align 4
53  %p3 = fcmp ogt <3 x half> %load3, zeroinitializer
54  %s3 = select <3 x i1> %p3, <3 x half> %load3, <3 x half> zeroinitializer
55  store <3 x half> %s3, <3 x half>* %in3b, align 4
56  %in4 = getelementptr half, half* %rd0, i64 9
57  %in4b = bitcast half* %in4 to <3 x half>*
58  %load4 = load <3 x half>, <3 x half>* %in4b, align 4
59  %p4 = fcmp ogt <3 x half> %load4, zeroinitializer
60  %s4 = select <3 x i1> %p4, <3 x half> %load4, <3 x half> zeroinitializer
61  store <3 x half> %s4, <3 x half>* %in4b, align 4
62  ret void
63
64; CHECK-LABEL: @no_nonpow2_vector
65; CHECK-NOT: shufflevector
66}
67
68define void @no_pointer_vector(half** nocapture align 16 %rd0) {
69  %in1b = bitcast half** %rd0 to <2 x half*>*
70  %load1 = load <2 x half*>, <2 x half*>* %in1b, align 4
71  %p1 = icmp ne <2 x half*> %load1, zeroinitializer
72  %s1 = select <2 x i1> %p1, <2 x half*> %load1, <2 x half*> zeroinitializer
73  store <2 x half*> %s1, <2 x half*>* %in1b, align 4
74  %in2 = getelementptr half*, half** %rd0, i64 2
75  %in2b = bitcast half** %in2 to <2 x half*>*
76  %load2 = load <2 x half*>, <2 x half*>* %in2b, align 4
77  %p2 = icmp ne <2 x half*> %load2, zeroinitializer
78  %s2 = select <2 x i1> %p2, <2 x half*> %load2, <2 x half*> zeroinitializer
79  store <2 x half*> %s2, <2 x half*>* %in2b, align 4
80  %in3 = getelementptr half*, half** %rd0, i64 4
81  %in3b = bitcast half** %in3 to <2 x half*>*
82  %load3 = load <2 x half*>, <2 x half*>* %in3b, align 4
83  %p3 = icmp ne <2 x half*> %load3, zeroinitializer
84  %s3 = select <2 x i1> %p3, <2 x half*> %load3, <2 x half*> zeroinitializer
85  store <2 x half*> %s3, <2 x half*>* %in3b, align 4
86  %in4 = getelementptr half*, half** %rd0, i64 6
87  %in4b = bitcast half** %in4 to <2 x half*>*
88  %load4 = load <2 x half*>, <2 x half*>* %in4b, align 4
89  %p4 = icmp ne <2 x half*> %load4, zeroinitializer
90  %s4 = select <2 x i1> %p4, <2 x half*> %load4, <2 x half*> zeroinitializer
91  store <2 x half*> %s4, <2 x half*>* %in4b, align 4
92  ret void
93
94; CHECK-LABEL: @no_pointer_vector
95; CHECK-NOT: shufflevector
96}
97