1; RUN: opt -S -mtriple=nvptx64-nvidia-cuda -infer-address-spaces %s | FileCheck %s
2
3%struct.S = type { [5 x i32] }
4
5$g1 = comdat any
6
7@g1 = linkonce_odr addrspace(3) global %struct.S zeroinitializer, comdat, align 4
8
9; CHECK-LABEL: @foo(
10; CHECK:  %x0 = tail call i32 @llvm.nvvm.read.ptx.sreg.tid.x() #2
11; CHECK:  %idxprom.i = zext i32 %x0 to i64
12; CHECK:  %arrayidx.i = getelementptr %struct.S, %struct.S* addrspacecast (%struct.S addrspace(3)* @g1 to %struct.S*), i64 0, i32 0, i64 %idxprom.i
13; CHECK:  tail call void @f1(i32* %arrayidx.i, i32 undef) #0
14; CHECK:  %x1 = load i32, i32 addrspace(3)* getelementptr inbounds (%struct.S, %struct.S addrspace(3)* @g1, i64 0, i32 0, i64 0), align 4
15; CHECK:  %L.sroa.0.0.insert.ext.i = zext i32 %x1 to i64
16; CHECK:  tail call void @f2(i64* null, i64 %L.sroa.0.0.insert.ext.i) #0
17; CHECK:  ret void
18define void @foo() local_unnamed_addr #0 {
19entry:
20  %x0 = tail call i32 @llvm.nvvm.read.ptx.sreg.tid.x() #2
21  %idxprom.i = zext i32 %x0 to i64
22  %arrayidx.i = getelementptr %struct.S, %struct.S* addrspacecast (%struct.S addrspace(3)* @g1 to %struct.S*), i64 0, i32 0, i64 %idxprom.i
23  tail call void @f1(i32* %arrayidx.i, i32 undef) #0
24  %x1 = load i32, i32* getelementptr (%struct.S, %struct.S* addrspacecast (%struct.S addrspace(3)* @g1 to %struct.S*), i64 0, i32 0, i64 0), align 4
25  %L.sroa.0.0.insert.ext.i = zext i32 %x1 to i64
26  tail call void @f2(i64* null, i64 %L.sroa.0.0.insert.ext.i) #0
27  ret void
28}
29
30declare void @f1(i32*, i32) local_unnamed_addr #0
31declare void @f2(i64*, i64) local_unnamed_addr #0
32declare i32 @llvm.nvvm.read.ptx.sreg.tid.x() #1
33
34; Make sure we can clone GEP which uses complex constant expressions as indices.
35; https://bugs.llvm.org/show_bug.cgi?id=51099
36@g2 = internal addrspace(3) global [128 x i8] undef, align 1
37
38; CHECK-LABEL: @complex_ce(
39; CHECK: %0 = load float, float addrspace(3)* bitcast
40; CHECK-SAME:   i8 addrspace(3)* getelementptr (i8,
41; CHECK-SAME:     i8 addrspace(3)* getelementptr inbounds ([128 x i8], [128 x i8] addrspace(3)* @g2, i64 0, i64 0),
42; CHECK-SAME:     i64 sub (
43; CHECK-SAME        i64 ptrtoint (
44; CHECK-SAME          i8 addrspace(3)* getelementptr inbounds ([128 x i8], [128 x i8] addrspace(3)* @g2, i64 0, i64 123) to i64),
45; CHECK-SAME:       i64 ptrtoint (
46; CHECK-SAME:         i8 addrspace(3)* getelementptr inbounds ([128 x i8], [128 x i8] addrspace(3)* @g2, i64 2, i64 0) to i64)))
47; CHECK-SAME:   to float addrspace(3)*)
48; Function Attrs: norecurse nounwind
49define float @complex_ce(i8* nocapture readnone %a, i8* nocapture readnone %b, i8* nocapture readnone %c) local_unnamed_addr #0 {
50entry:
51  %0 = load float, float* bitcast (
52       i8* getelementptr (
53         i8, i8* getelementptr inbounds (
54           [128 x i8],
55           [128 x i8]* addrspacecast ([128 x i8] addrspace(3)* @g2 to [128 x i8]*),
56           i64 0,
57           i64 0),
58         i64 sub (
59           i64 ptrtoint (
60             i8* getelementptr inbounds (
61               [128 x i8],
62               [128 x i8]* addrspacecast ([128 x i8] addrspace(3)* @g2 to [128 x i8]*),
63               i64 0,
64               i64 123)
65             to i64),
66           i64 ptrtoint (
67             i8* getelementptr inbounds (
68               [128 x i8],
69               [128 x i8]* addrspacecast ([128 x i8] addrspace(3)* @g2 to [128 x i8]*),
70               i64 2,
71               i64 0)
72             to i64)))
73        to float*), align 4
74  ret float %0
75}
76
77
78
79attributes #0 = { convergent nounwind }
80attributes #1 = { nounwind readnone }
81attributes #2 = { nounwind }
82