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[8:11], s[0:1], 0x9
38; GCN-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0xd
39; GCN-NEXT:    v_mov_b32_e32 v5, 0
40; GCN-NEXT:    s_mov_b32 s7, 0xf000
41; GCN-NEXT:    s_mov_b32 s6, 0
42; GCN-NEXT:    s_waitcnt lgkmcnt(0)
43; GCN-NEXT:    s_mov_b64 s[4:5], s[10:11]
44; GCN-NEXT:    v_lshlrev_b32_e32 v4, 4, v0
45; GCN-NEXT:    buffer_load_dwordx4 v[1:4], v[4:5], s[4:7], 0 addr64
46; GCN-NEXT:    s_load_dword s0, s[0:1], 0xf
47; GCN-NEXT:    v_lshlrev_b32_e32 v6, 2, v0
48; GCN-NEXT:    v_mov_b32_e32 v0, s2
49; GCN-NEXT:    v_cmp_eq_u32_e64 vcc, s3, 3
50; GCN-NEXT:    v_mov_b32_e32 v7, v5
51; GCN-NEXT:    s_mov_b64 s[10:11], s[6:7]
52; GCN-NEXT:    s_waitcnt vmcnt(0)
53; GCN-NEXT:    v_cndmask_b32_e32 v4, v4, v0, vcc
54; GCN-NEXT:    v_cmp_eq_u32_e64 vcc, s3, 2
55; GCN-NEXT:    v_cndmask_b32_e32 v3, v3, v0, vcc
56; GCN-NEXT:    v_cmp_eq_u32_e64 vcc, s3, 1
57; GCN-NEXT:    v_cndmask_b32_e32 v2, v2, v0, vcc
58; GCN-NEXT:    v_cmp_eq_u32_e64 vcc, s3, 0
59; GCN-NEXT:    v_cndmask_b32_e32 v0, v1, v0, vcc
60; GCN-NEXT:    s_waitcnt lgkmcnt(0)
61; GCN-NEXT:    v_cmp_eq_u32_e64 vcc, s0, 1
62; GCN-NEXT:    v_cndmask_b32_e32 v0, v0, v2, vcc
63; GCN-NEXT:    v_cmp_eq_u32_e64 vcc, s0, 2
64; GCN-NEXT:    v_cndmask_b32_e32 v0, v0, v3, vcc
65; GCN-NEXT:    v_cmp_eq_u32_e64 vcc, s0, 3
66; GCN-NEXT:    v_cndmask_b32_e32 v0, v0, v4, vcc
67; GCN-NEXT:    buffer_store_dword v0, v[6:7], s[8:11], 0 addr64
68; GCN-NEXT:    s_endpgm
69  %id = call i32 @llvm.amdgcn.workitem.id.x()
70  %id.ext = sext i32 %id to i64
71  %gep.in = getelementptr inbounds <4 x i32>, <4 x i32> addrspace(1)* %in, i64 %id.ext
72  %gep.out = getelementptr inbounds i32, i32 addrspace(1)* %out, i64 %id.ext
73  %vec = load <4 x i32>, <4 x i32> addrspace(1)* %gep.in
74  %insert = insertelement <4 x i32> %vec, i32 %val, i32 %idx0
75  %extract = extractelement <4 x i32> %insert, i32 %idx1
76  store i32 %extract, i32 addrspace(1)* %gep.out
77  ret void
78}
79
80define amdgpu_kernel void @extract_insert_same_elt2_v4i32(i32 addrspace(1)* %out, <4 x i32> addrspace(1)* %in, i32 %val, i32 %idx) #1 {
81; GCN-LABEL: extract_insert_same_elt2_v4i32:
82; GCN:       ; %bb.0:
83; GCN-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x9
84; GCN-NEXT:    s_load_dword s0, s[0:1], 0xd
85; GCN-NEXT:    s_waitcnt lgkmcnt(0)
86; GCN-NEXT:    s_mov_b32 s7, 0xf000
87; GCN-NEXT:    s_mov_b32 s6, 0
88; GCN-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
89; GCN-NEXT:    v_mov_b32_e32 v1, 0
90; GCN-NEXT:    v_mov_b32_e32 v2, s0
91; GCN-NEXT:    buffer_store_dword v2, v[0:1], s[4:7], 0 addr64
92; GCN-NEXT:    s_endpgm
93  %id = call i32 @llvm.amdgcn.workitem.id.x()
94  %id.ext = sext i32 %id to i64
95  %gep.in = getelementptr inbounds <4 x i32>, <4 x i32> addrspace(1)* %in, i64 %id.ext
96  %gep.out = getelementptr inbounds i32, i32 addrspace(1)* %out, i64 %id.ext
97  %vec = load <4 x i32>, <4 x i32> addrspace(1)* %gep.in
98  %insert = insertelement <4 x i32> %vec, i32 %val, i32 %idx
99  %extract = extractelement <4 x i32> %insert, i32 %idx
100  store i32 %extract, i32 addrspace(1)* %gep.out
101  ret void
102}
103
104define amdgpu_kernel void @extract_insert_same_dynelt_v4f32(float addrspace(1)* %out, <4 x float> addrspace(1)* %in, float %val, i32 %idx) #1 {
105; GCN-LABEL: extract_insert_same_dynelt_v4f32:
106; GCN:       ; %bb.0:
107; GCN-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x9
108; GCN-NEXT:    s_load_dword s8, s[0:1], 0xd
109; GCN-NEXT:    v_mov_b32_e32 v2, 0
110; GCN-NEXT:    s_mov_b32 s3, 0xf000
111; GCN-NEXT:    s_mov_b32 s2, 0
112; GCN-NEXT:    s_waitcnt lgkmcnt(0)
113; GCN-NEXT:    s_mov_b64 s[0:1], s[6:7]
114; GCN-NEXT:    v_lshlrev_b32_e32 v1, 4, v0
115; GCN-NEXT:    v_lshlrev_b32_e32 v4, 2, v0
116; GCN-NEXT:    v_mov_b32_e32 v5, v2
117; GCN-NEXT:    buffer_load_dwordx4 v[0:3], v[1:2], s[0:3], 0 addr64
118; GCN-NEXT:    s_mov_b64 s[6:7], s[2:3]
119; GCN-NEXT:    s_waitcnt vmcnt(0)
120; GCN-NEXT:    v_mov_b32_e32 v0, s8
121; GCN-NEXT:    buffer_store_dword v0, v[4:5], s[4:7], 0 addr64
122; GCN-NEXT:    s_endpgm
123  %id = call i32 @llvm.amdgcn.workitem.id.x()
124  %id.ext = sext i32 %id to i64
125  %gep.in = getelementptr inbounds <4 x float>, <4 x float> addrspace(1)* %in, i64 %id.ext
126  %gep.out = getelementptr inbounds float, float addrspace(1)* %out, i64 %id.ext
127  %vec = load volatile <4 x float>, <4 x float> addrspace(1)* %gep.in
128  %insert = insertelement <4 x float> %vec, float %val, i32 %idx
129  %extract = extractelement <4 x float> %insert, i32 %idx
130  store float %extract, float addrspace(1)* %gep.out
131  ret void
132}
133
134attributes #0 = { nounwind readnone }
135attributes #1 = { nounwind }
136