1; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
2; RUN: llc -march=amdgcn -mcpu=tahiti -verify-machineinstrs < %s | FileCheck -check-prefix=GCN %s
3
4; Test that when extracting the same unknown vector index from an
5; insertelement the dynamic indexing is folded away.
6
7declare i32 @llvm.amdgcn.workitem.id.x() #0
8
9; No dynamic indexing required
10define amdgpu_kernel void @extract_insert_same_dynelt_v4i32(i32 addrspace(1)* %out, <4 x i32> addrspace(1)* %in, i32 %val, i32 %idx) #1 {
11; GCN-LABEL: extract_insert_same_dynelt_v4i32:
12; GCN:       ; %bb.0:
13; GCN-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x9
14; GCN-NEXT:    s_load_dword s0, s[0:1], 0xd
15; GCN-NEXT:    s_waitcnt lgkmcnt(0)
16; GCN-NEXT:    s_mov_b32 s7, 0xf000
17; GCN-NEXT:    s_mov_b32 s6, 0
18; GCN-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
19; GCN-NEXT:    v_mov_b32_e32 v1, 0
20; GCN-NEXT:    v_mov_b32_e32 v2, s0
21; GCN-NEXT:    buffer_store_dword v2, v[0:1], s[4:7], 0 addr64
22; GCN-NEXT:    s_endpgm
23  %id = call i32 @llvm.amdgcn.workitem.id.x()
24  %id.ext = sext i32 %id to i64
25  %gep.in = getelementptr inbounds <4 x i32>, <4 x i32> addrspace(1)* %in, i64 %id.ext
26  %gep.out = getelementptr inbounds i32, i32 addrspace(1)* %out, i64 %id.ext
27  %vec = load <4 x i32>, <4 x i32> addrspace(1)* %gep.in
28  %insert = insertelement <4 x i32> %vec, i32 %val, i32 %idx
29  %extract = extractelement <4 x i32> %insert, i32 %idx
30  store i32 %extract, i32 addrspace(1)* %gep.out
31  ret void
32}
33
34define 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 {
35; GCN-LABEL: extract_insert_different_dynelt_v4i32:
36; GCN:       ; %bb.0:
37; GCN-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x9
38; GCN-NEXT:    s_load_dwordx4 s[8:11], s[0:1], 0xd
39; GCN-NEXT:    v_mov_b32_e32 v5, 0
40; GCN-NEXT:    s_mov_b32 s3, 0xf000
41; GCN-NEXT:    s_mov_b32 s2, 0
42; GCN-NEXT:    s_waitcnt lgkmcnt(0)
43; GCN-NEXT:    s_mov_b64 s[0:1], s[6:7]
44; GCN-NEXT:    v_lshlrev_b32_e32 v4, 4, v0
45; GCN-NEXT:    buffer_load_dwordx4 v[1:4], v[4:5], s[0:3], 0 addr64
46; GCN-NEXT:    v_lshlrev_b32_e32 v6, 2, v0
47; GCN-NEXT:    v_mov_b32_e32 v0, s8
48; GCN-NEXT:    v_cmp_eq_u32_e64 vcc, s9, 3
49; GCN-NEXT:    v_mov_b32_e32 v7, v5
50; GCN-NEXT:    s_mov_b64 s[6:7], s[2:3]
51; GCN-NEXT:    s_waitcnt vmcnt(0)
52; GCN-NEXT:    v_cndmask_b32_e32 v4, v4, v0, vcc
53; GCN-NEXT:    v_cmp_eq_u32_e64 vcc, s9, 2
54; GCN-NEXT:    v_cndmask_b32_e32 v3, v3, v0, vcc
55; GCN-NEXT:    v_cmp_eq_u32_e64 vcc, s9, 1
56; GCN-NEXT:    v_cndmask_b32_e32 v2, v2, v0, vcc
57; GCN-NEXT:    v_cmp_eq_u32_e64 vcc, s9, 0
58; GCN-NEXT:    v_cndmask_b32_e32 v0, v1, v0, vcc
59; GCN-NEXT:    v_cmp_eq_u32_e64 vcc, s10, 1
60; GCN-NEXT:    v_cndmask_b32_e32 v0, v0, v2, vcc
61; GCN-NEXT:    v_cmp_eq_u32_e64 vcc, s10, 2
62; GCN-NEXT:    v_cndmask_b32_e32 v0, v0, v3, vcc
63; GCN-NEXT:    v_cmp_eq_u32_e64 vcc, s10, 3
64; GCN-NEXT:    v_cndmask_b32_e32 v0, v0, v4, vcc
65; GCN-NEXT:    buffer_store_dword v0, v[6:7], s[4:7], 0 addr64
66; GCN-NEXT:    s_endpgm
67  %id = call i32 @llvm.amdgcn.workitem.id.x()
68  %id.ext = sext i32 %id to i64
69  %gep.in = getelementptr inbounds <4 x i32>, <4 x i32> addrspace(1)* %in, i64 %id.ext
70  %gep.out = getelementptr inbounds i32, i32 addrspace(1)* %out, i64 %id.ext
71  %vec = load <4 x i32>, <4 x i32> addrspace(1)* %gep.in
72  %insert = insertelement <4 x i32> %vec, i32 %val, i32 %idx0
73  %extract = extractelement <4 x i32> %insert, i32 %idx1
74  store i32 %extract, i32 addrspace(1)* %gep.out
75  ret void
76}
77
78define amdgpu_kernel void @extract_insert_same_elt2_v4i32(i32 addrspace(1)* %out, <4 x i32> addrspace(1)* %in, i32 %val, i32 %idx) #1 {
79; GCN-LABEL: extract_insert_same_elt2_v4i32:
80; GCN:       ; %bb.0:
81; GCN-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x9
82; GCN-NEXT:    s_load_dword s0, s[0:1], 0xd
83; GCN-NEXT:    s_waitcnt lgkmcnt(0)
84; GCN-NEXT:    s_mov_b32 s7, 0xf000
85; GCN-NEXT:    s_mov_b32 s6, 0
86; GCN-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
87; GCN-NEXT:    v_mov_b32_e32 v1, 0
88; GCN-NEXT:    v_mov_b32_e32 v2, s0
89; GCN-NEXT:    buffer_store_dword v2, v[0:1], s[4:7], 0 addr64
90; GCN-NEXT:    s_endpgm
91  %id = call i32 @llvm.amdgcn.workitem.id.x()
92  %id.ext = sext i32 %id to i64
93  %gep.in = getelementptr inbounds <4 x i32>, <4 x i32> addrspace(1)* %in, i64 %id.ext
94  %gep.out = getelementptr inbounds i32, i32 addrspace(1)* %out, i64 %id.ext
95  %vec = load <4 x i32>, <4 x i32> addrspace(1)* %gep.in
96  %insert = insertelement <4 x i32> %vec, i32 %val, i32 %idx
97  %extract = extractelement <4 x i32> %insert, i32 %idx
98  store i32 %extract, i32 addrspace(1)* %gep.out
99  ret void
100}
101
102define amdgpu_kernel void @extract_insert_same_dynelt_v4f32(float addrspace(1)* %out, <4 x float> addrspace(1)* %in, float %val, i32 %idx) #1 {
103; GCN-LABEL: extract_insert_same_dynelt_v4f32:
104; GCN:       ; %bb.0:
105; GCN-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x9
106; GCN-NEXT:    s_load_dword s8, s[0:1], 0xd
107; GCN-NEXT:    v_mov_b32_e32 v2, 0
108; GCN-NEXT:    s_mov_b32 s3, 0xf000
109; GCN-NEXT:    s_mov_b32 s2, 0
110; GCN-NEXT:    s_waitcnt lgkmcnt(0)
111; GCN-NEXT:    s_mov_b64 s[0:1], s[6:7]
112; GCN-NEXT:    v_lshlrev_b32_e32 v1, 4, v0
113; GCN-NEXT:    v_lshlrev_b32_e32 v4, 2, v0
114; GCN-NEXT:    v_mov_b32_e32 v5, v2
115; GCN-NEXT:    buffer_load_dwordx4 v[0:3], v[1:2], s[0:3], 0 addr64
116; GCN-NEXT:    s_mov_b64 s[6:7], s[2:3]
117; GCN-NEXT:    s_waitcnt vmcnt(0)
118; GCN-NEXT:    v_mov_b32_e32 v0, s8
119; GCN-NEXT:    buffer_store_dword v0, v[4:5], s[4:7], 0 addr64
120; GCN-NEXT:    s_endpgm
121  %id = call i32 @llvm.amdgcn.workitem.id.x()
122  %id.ext = sext i32 %id to i64
123  %gep.in = getelementptr inbounds <4 x float>, <4 x float> addrspace(1)* %in, i64 %id.ext
124  %gep.out = getelementptr inbounds float, float addrspace(1)* %out, i64 %id.ext
125  %vec = load volatile <4 x float>, <4 x float> addrspace(1)* %gep.in
126  %insert = insertelement <4 x float> %vec, float %val, i32 %idx
127  %extract = extractelement <4 x float> %insert, i32 %idx
128  store float %extract, float addrspace(1)* %gep.out
129  ret void
130}
131
132attributes #0 = { nounwind readnone }
133attributes #1 = { nounwind }
134