1; RUN: llc -march=amdgcn -mcpu=tahiti -verify-machineinstrs < %s | FileCheck -check-prefix=GCN %s
2
3; Test that when extracting the same unknown vector index from an
4; insertelement the dynamic indexing is folded away.
5
6declare i32 @llvm.amdgcn.workitem.id.x() #0
7
8; No dynamic indexing required
9; GCN-LABEL: {{^}}extract_insert_same_dynelt_v4i32:
10; GCN: s_load_dword [[VAL:s[0-9]+]], s{{\[[0-9]+:[0-9]+\]}}, 0xd{{$}}
11; GCN-NOT buffer_load_dword
12; GCN-NOT: [[VAL]]
13; GCN: v_mov_b32_e32 [[VVAL:v[0-9]+]], [[VAL]]
14; GCN-NOT: [[VVAL]]
15; GCN: buffer_store_dword [[VVAL]]
16define amdgpu_kernel void @extract_insert_same_dynelt_v4i32(i32 addrspace(1)* %out, <4 x i32> addrspace(1)* %in, i32 %val, i32 %idx) #1 {
17  %id = call i32 @llvm.amdgcn.workitem.id.x()
18  %id.ext = sext i32 %id to i64
19  %gep.in = getelementptr inbounds <4 x i32>, <4 x i32> addrspace(1)* %in, i64 %id.ext
20  %gep.out = getelementptr inbounds i32, i32 addrspace(1)* %out, i64 %id.ext
21  %vec = load <4 x i32>, <4 x i32> addrspace(1)* %gep.in
22  %insert = insertelement <4 x i32> %vec, i32 %val, i32 %idx
23  %extract = extractelement <4 x i32> %insert, i32 %idx
24  store i32 %extract, i32 addrspace(1)* %gep.out
25  ret void
26}
27
28; GCN-LABEL: {{^}}extract_insert_different_dynelt_v4i32:
29; GCN: buffer_load_dwordx4
30; GCN: v_movreld_b32
31; GCN: v_cndmask_b32
32; GCN: v_cndmask_b32
33; GCN: v_cndmask_b32
34; GCN: buffer_store_dword v
35define amdgpu_kernel void @extract_insert_different_dynelt_v4i32(i32 addrspace(1)* %out, <4 x i32> addrspace(1)* %in, i32 %val, i32 %idx0, i32 %idx1) #1 {
36  %id = call i32 @llvm.amdgcn.workitem.id.x()
37  %id.ext = sext i32 %id to i64
38  %gep.in = getelementptr inbounds <4 x i32>, <4 x i32> addrspace(1)* %in, i64 %id.ext
39  %gep.out = getelementptr inbounds i32, i32 addrspace(1)* %out, i64 %id.ext
40  %vec = load <4 x i32>, <4 x i32> addrspace(1)* %gep.in
41  %insert = insertelement <4 x i32> %vec, i32 %val, i32 %idx0
42  %extract = extractelement <4 x i32> %insert, i32 %idx1
43  store i32 %extract, i32 addrspace(1)* %gep.out
44  ret void
45}
46
47; GCN-LABEL: {{^}}extract_insert_same_elt2_v4i32:
48; GCN: s_load_dword [[VAL:s[0-9]+]], s{{\[[0-9]+:[0-9]+\]}}, 0xd{{$}}
49; GCN-NOT buffer_load_dword
50; GCN-NOT: [[VAL]]
51; GCN: v_mov_b32_e32 [[VVAL:v[0-9]+]], [[VAL]]
52; GCN-NOT: [[VVAL]]
53; GCN: buffer_store_dword [[VVAL]]
54define amdgpu_kernel void @extract_insert_same_elt2_v4i32(i32 addrspace(1)* %out, <4 x i32> addrspace(1)* %in, i32 %val, i32 %idx) #1 {
55  %id = call i32 @llvm.amdgcn.workitem.id.x()
56  %id.ext = sext i32 %id to i64
57  %gep.in = getelementptr inbounds <4 x i32>, <4 x i32> addrspace(1)* %in, i64 %id.ext
58  %gep.out = getelementptr inbounds i32, i32 addrspace(1)* %out, i64 %id.ext
59  %vec = load <4 x i32>, <4 x i32> addrspace(1)* %gep.in
60  %insert = insertelement <4 x i32> %vec, i32 %val, i32 %idx
61  %extract = extractelement <4 x i32> %insert, i32 %idx
62  store i32 %extract, i32 addrspace(1)* %gep.out
63  ret void
64}
65
66; GCN-LABEL: {{^}}extract_insert_same_dynelt_v4f32:
67; GCN: s_load_dword [[VAL:s[0-9]+]], s{{\[[0-9]+:[0-9]+\]}}, 0xd{{$}}
68; GCN-NOT buffer_load_dword
69; GCN-NOT: [[VAL]]
70; GCN: v_mov_b32_e32 [[VVAL:v[0-9]+]], [[VAL]]
71; GCN-NOT: [[VVAL]]
72; GCN: buffer_store_dword [[VVAL]]
73define amdgpu_kernel void @extract_insert_same_dynelt_v4f32(float addrspace(1)* %out, <4 x float> addrspace(1)* %in, float %val, i32 %idx) #1 {
74  %id = call i32 @llvm.amdgcn.workitem.id.x()
75  %id.ext = sext i32 %id to i64
76  %gep.in = getelementptr inbounds <4 x float>, <4 x float> addrspace(1)* %in, i64 %id.ext
77  %gep.out = getelementptr inbounds float, float addrspace(1)* %out, i64 %id.ext
78  %vec = load volatile <4 x float>, <4 x float> addrspace(1)* %gep.in
79  %insert = insertelement <4 x float> %vec, float %val, i32 %idx
80  %extract = extractelement <4 x float> %insert, i32 %idx
81  store float %extract, float addrspace(1)* %gep.out
82  ret void
83}
84
85attributes #0 = { nounwind readnone }
86attributes #1 = { nounwind }
87