1; NOTE: Assertions have been autogenerated by utils/update_test_checks.py
2; RUN: opt -S -mtriple=nvptx64-nvidia-cuda -infer-address-spaces %s | FileCheck %s
3
4%struct.S = type { [5 x i32] }
5
6$g1 = comdat any
7
8@g1 = linkonce_odr addrspace(3) global %struct.S zeroinitializer, comdat, align 4
9
10define void @foo() local_unnamed_addr #0 {
11; CHECK-LABEL: @foo(
12; CHECK-NEXT:  entry:
13; CHECK-NEXT:    [[X0:%.*]] = tail call i32 @llvm.nvvm.read.ptx.sreg.tid.x() #[[ATTR2:[0-9]+]]
14; CHECK-NEXT:    [[IDXPROM_I:%.*]] = zext i32 [[X0]] to i64
15; CHECK-NEXT:    [[ARRAYIDX_I:%.*]] = getelementptr [[STRUCT_S:%.*]], %struct.S* addrspacecast ([[STRUCT_S]] addrspace(3)* @g1 to %struct.S*), i64 0, i32 0, i64 [[IDXPROM_I]]
16; CHECK-NEXT:    tail call void @f1(i32* [[ARRAYIDX_I]], i32 undef) #[[ATTR0:[0-9]+]]
17; CHECK-NEXT:    [[X1:%.*]] = load i32, i32 addrspace(3)* getelementptr inbounds ([[STRUCT_S]], [[STRUCT_S]] addrspace(3)* @g1, i64 0, i32 0, i64 0), align 4
18; CHECK-NEXT:    [[L_SROA_0_0_INSERT_EXT_I:%.*]] = zext i32 [[X1]] to i64
19; CHECK-NEXT:    tail call void @f2(i64* null, i64 [[L_SROA_0_0_INSERT_EXT_I]]) #[[ATTR0]]
20; CHECK-NEXT:    ret void
21;
22entry:
23  %x0 = tail call i32 @llvm.nvvm.read.ptx.sreg.tid.x() #2
24  %idxprom.i = zext i32 %x0 to i64
25  %arrayidx.i = getelementptr %struct.S, %struct.S* addrspacecast (%struct.S addrspace(3)* @g1 to %struct.S*), i64 0, i32 0, i64 %idxprom.i
26  tail call void @f1(i32* %arrayidx.i, i32 undef) #0
27  %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
28  %L.sroa.0.0.insert.ext.i = zext i32 %x1 to i64
29  tail call void @f2(i64* null, i64 %L.sroa.0.0.insert.ext.i) #0
30  ret void
31}
32
33declare void @f1(i32*, i32) local_unnamed_addr #0
34declare void @f2(i64*, i64) local_unnamed_addr #0
35declare i32 @llvm.nvvm.read.ptx.sreg.tid.x() #1
36
37; Make sure we can clone GEP which uses complex constant expressions as indices.
38; https://bugs.llvm.org/show_bug.cgi?id=51099
39@g2 = internal addrspace(3) global [128 x i8] undef, align 1
40
41define float @complex_ce(i8* nocapture readnone %a, i8* nocapture readnone %b, i8* nocapture readnone %c) local_unnamed_addr #0 {
42; CHECK-LABEL: @complex_ce(
43; CHECK-NEXT:  entry:
44; CHECK-NEXT:    [[TMP0:%.*]] = load float, float addrspace(3)* bitcast (i8 addrspace(3)* getelementptr (i8, i8 addrspace(3)* getelementptr inbounds ([128 x i8], [128 x i8] addrspace(3)* @g2, i64 0, i64 0), i64 sub (i64 ptrtoint (i8 addrspace(3)* getelementptr inbounds ([128 x i8], [128 x i8] addrspace(3)* @g2, i64 0, i64 123) to i64), i64 ptrtoint (i8 addrspace(3)* getelementptr inbounds ([128 x i8], [128 x i8] addrspace(3)* @g2, i64 2, i64 0) to i64))) to float addrspace(3)*), align 4
45; CHECK-NEXT:    ret float [[TMP0]]
46;
47entry:
48  %0 = load float, float* bitcast (
49       i8* getelementptr (
50         i8, i8* getelementptr inbounds (
51           [128 x i8],
52           [128 x i8]* addrspacecast ([128 x i8] addrspace(3)* @g2 to [128 x i8]*),
53           i64 0,
54           i64 0),
55         i64 sub (
56           i64 ptrtoint (
57             i8* getelementptr inbounds (
58               [128 x i8],
59               [128 x i8]* addrspacecast ([128 x i8] addrspace(3)* @g2 to [128 x i8]*),
60               i64 0,
61               i64 123)
62             to i64),
63           i64 ptrtoint (
64             i8* getelementptr inbounds (
65               [128 x i8],
66               [128 x i8]* addrspacecast ([128 x i8] addrspace(3)* @g2 to [128 x i8]*),
67               i64 2,
68               i64 0)
69             to i64)))
70        to float*), align 4
71  ret float %0
72}
73
74
75
76attributes #0 = { convergent nounwind }
77attributes #1 = { nounwind readnone }
78attributes #2 = { nounwind }
79