1 ; RUN: llc -amdgpu-scalarize-global-loads=false -mtriple=amdgcn -verify-machineinstrs < %s | FileCheck -check-prefix=GCN %s
3 ; GCN-LABEL: {{^}}alignbit_shr_pat:
4 ; GCN-DAG: s_load_dword s[[SHR:[0-9]+]]
5 ; GCN-DAG: load_dwordx2 v[[[LO:[0-9]+]]:[[HI:[0-9]+]]]
6 ; GCN: v_alignbit_b32 v{{[0-9]+}}, v[[HI]], v[[LO]], s[[SHR]]
8 define amdgpu_kernel void @alignbit_shr_pat(ptr addrspace(1) nocapture readonly %arg, ptr addrspace(1) nocapture %arg1, i32 %arg2) {
10 %tmp = load i64, ptr addrspace(1) %arg, align 8
11 %tmp3 = and i32 %arg2, 31
12 %tmp4 = zext i32 %tmp3 to i64
13 %tmp5 = lshr i64 %tmp, %tmp4
14 %tmp6 = trunc i64 %tmp5 to i32
15 store i32 %tmp6, ptr addrspace(1) %arg1, align 4
19 ; GCN-LABEL: {{^}}alignbit_shr_pat_v:
20 ; GCN-DAG: load_dword v[[SHR:[0-9]+]],
21 ; GCN-DAG: load_dwordx2 v[[[LO:[0-9]+]]:[[HI:[0-9]+]]]
22 ; GCN: v_alignbit_b32 v{{[0-9]+}}, v[[HI]], v[[LO]], v[[SHR]]
24 define amdgpu_kernel void @alignbit_shr_pat_v(ptr addrspace(1) nocapture readonly %arg, ptr addrspace(1) nocapture %arg1) {
26 %tid = tail call i32 @llvm.amdgcn.workitem.id.x()
27 %gep1 = getelementptr inbounds i64, ptr addrspace(1) %arg, i32 %tid
28 %tmp = load i64, ptr addrspace(1) %gep1, align 8
29 %gep2 = getelementptr inbounds i32, ptr addrspace(1) %arg1, i32 %tid
30 %amt = load i32, ptr addrspace(1) %gep2, align 4
31 %tmp3 = and i32 %amt, 31
32 %tmp4 = zext i32 %tmp3 to i64
33 %tmp5 = lshr i64 %tmp, %tmp4
34 %tmp6 = trunc i64 %tmp5 to i32
35 store i32 %tmp6, ptr addrspace(1) %gep2, align 4
39 ; GCN-LABEL: {{^}}alignbit_shr_pat_wrong_and30:
40 ; Negative test, wrong constant
42 ; GCN-NOT: v_alignbit_b32
44 define amdgpu_kernel void @alignbit_shr_pat_wrong_and30(ptr addrspace(1) nocapture readonly %arg, ptr addrspace(1) nocapture %arg1, i32 %arg2) {
46 %tmp = load i64, ptr addrspace(1) %arg, align 8
47 %tmp3 = and i32 %arg2, 30
48 %tmp4 = zext i32 %tmp3 to i64
49 %tmp5 = lshr i64 %tmp, %tmp4
50 %tmp6 = trunc i64 %tmp5 to i32
51 store i32 %tmp6, ptr addrspace(1) %arg1, align 4
55 ; GCN-LABEL: {{^}}alignbit_shr_pat_wrong_and63:
56 ; Negative test, wrong constant
58 ; GCN-NOT: v_alignbit_b32
60 define amdgpu_kernel void @alignbit_shr_pat_wrong_and63(ptr addrspace(1) nocapture readonly %arg, ptr addrspace(1) nocapture %arg1, i32 %arg2) {
62 %tmp = load i64, ptr addrspace(1) %arg, align 8
63 %tmp3 = and i32 %arg2, 63
64 %tmp4 = zext i32 %tmp3 to i64
65 %tmp5 = lshr i64 %tmp, %tmp4
66 %tmp6 = trunc i64 %tmp5 to i32
67 store i32 %tmp6, ptr addrspace(1) %arg1, align 4
71 ; GCN-LABEL: {{^}}alignbit_shr_pat_const30:
72 ; GCN: load_dwordx2 v[[[LO:[0-9]+]]:[[HI:[0-9]+]]]
73 ; GCN: v_alignbit_b32 v{{[0-9]+}}, v[[HI]], v[[LO]], 30
75 define amdgpu_kernel void @alignbit_shr_pat_const30(ptr addrspace(1) nocapture readonly %arg, ptr addrspace(1) nocapture %arg1) {
77 %tmp = load i64, ptr addrspace(1) %arg, align 8
78 %tmp5 = lshr i64 %tmp, 30
79 %tmp6 = trunc i64 %tmp5 to i32
80 store i32 %tmp6, ptr addrspace(1) %arg1, align 4
84 ; GCN-LABEL: {{^}}alignbit_shr_pat_wrong_const33:
85 ; Negative test, shift amount more than 31
86 ; GCN: v_lshrrev_b32_e32 v{{[0-9]+}}, 1, v{{[0-9]+}}
87 ; GCN-NOT: v_alignbit_b32
89 define amdgpu_kernel void @alignbit_shr_pat_wrong_const33(ptr addrspace(1) nocapture readonly %arg, ptr addrspace(1) nocapture %arg1) {
91 %tmp = load i64, ptr addrspace(1) %arg, align 8
92 %tmp5 = lshr i64 %tmp, 33
93 %tmp6 = trunc i64 %tmp5 to i32
94 store i32 %tmp6, ptr addrspace(1) %arg1, align 4
98 declare i32 @llvm.amdgcn.workitem.id.x() #0
100 attributes #0 = { nounwind readnone speculatable }