1; RUN: llc < %s -march=r600 -mcpu=redwood | FileCheck --check-prefix=EG %s
2; RUN: llc < %s -march=amdgcn -mcpu=verde -verify-machineinstrs | FileCheck --check-prefix=GCN --check-prefix=SI %s
3; RUN: llc < %s -march=amdgcn -mcpu=bonaire -verify-machineinstrs | FileCheck --check-prefix=GCN --check-prefix=CI %s
4
5@local_memory_two_objects.local_mem0 = internal unnamed_addr addrspace(3) global [4 x i32] undef, align 4
6@local_memory_two_objects.local_mem1 = internal unnamed_addr addrspace(3) global [4 x i32] undef, align 4
7
8
9; Check that the LDS size emitted correctly
10; EG: .long 166120
11; EG-NEXT: .long 8
12; GCN: .long 47180
13; GCN-NEXT: .long 32900
14
15; EG: {{^}}local_memory_two_objects:
16
17; We would like to check the lds writes are using different
18; addresses, but due to variations in the scheduler, we can't do
19; this consistently on evergreen GPUs.
20; EG: LDS_WRITE
21; EG: LDS_WRITE
22; GCN: ds_write_b32 {{v[0-9]*}}, v[[ADDRW:[0-9]*]]
23; GCN-NOT: ds_write_b32 {{v[0-9]*}}, v[[ADDRW]]
24
25; GROUP_BARRIER must be the last instruction in a clause
26; EG: GROUP_BARRIER
27; EG-NEXT: ALU clause
28
29; Make sure the lds reads are using different addresses, at different
30; constant offsets.
31; EG: LDS_READ_RET {{[*]*}} OQAP, {{PV|T}}[[ADDRR:[0-9]*\.[XYZW]]]
32; EG-NOT: LDS_READ_RET {{[*]*}} OQAP, T[[ADDRR]]
33; SI: v_add_i32_e32 [[SIPTR:v[0-9]+]], vcc, 16, v{{[0-9]+}}
34; SI: ds_read_b32 {{v[0-9]+}}, [[SIPTR]]
35; CI: ds_read_b32 {{v[0-9]+}}, [[ADDRR:v[0-9]+]] offset:16
36; CI: ds_read_b32 {{v[0-9]+}}, [[ADDRR]]
37
38define void @local_memory_two_objects(i32 addrspace(1)* %out) {
39entry:
40  %x.i = call i32 @llvm.r600.read.tidig.x() #0
41  %arrayidx = getelementptr inbounds [4 x i32], [4 x i32] addrspace(3)* @local_memory_two_objects.local_mem0, i32 0, i32 %x.i
42  store i32 %x.i, i32 addrspace(3)* %arrayidx, align 4
43  %mul = shl nsw i32 %x.i, 1
44  %arrayidx1 = getelementptr inbounds [4 x i32], [4 x i32] addrspace(3)* @local_memory_two_objects.local_mem1, i32 0, i32 %x.i
45  store i32 %mul, i32 addrspace(3)* %arrayidx1, align 4
46  %sub = sub nsw i32 3, %x.i
47  call void @llvm.AMDGPU.barrier.local()
48  %arrayidx2 = getelementptr inbounds [4 x i32], [4 x i32] addrspace(3)* @local_memory_two_objects.local_mem0, i32 0, i32 %sub
49  %0 = load i32, i32 addrspace(3)* %arrayidx2, align 4
50  %arrayidx3 = getelementptr inbounds i32, i32 addrspace(1)* %out, i32 %x.i
51  store i32 %0, i32 addrspace(1)* %arrayidx3, align 4
52  %arrayidx4 = getelementptr inbounds [4 x i32], [4 x i32] addrspace(3)* @local_memory_two_objects.local_mem1, i32 0, i32 %sub
53  %1 = load i32, i32 addrspace(3)* %arrayidx4, align 4
54  %add = add nsw i32 %x.i, 4
55  %arrayidx5 = getelementptr inbounds i32, i32 addrspace(1)* %out, i32 %add
56  store i32 %1, i32 addrspace(1)* %arrayidx5, align 4
57  ret void
58}
59
60declare i32 @llvm.r600.read.tidig.x() #0
61declare void @llvm.AMDGPU.barrier.local()
62
63attributes #0 = { readnone }
64