1 ; RUN: llc -march=amdgcn -mcpu=SI < %s | FileCheck -check-prefix=SI %s
2 ; RUN: llc -march=amdgcn -mcpu=tonga < %s | FileCheck -check-prefix=SI %s
4 ; SI-LABEL: {{^}}no_reorder_v2f64_global_load_store:
5 ; SI: buffer_load_dwordx2
6 ; SI: buffer_load_dwordx2
7 ; SI: buffer_load_dwordx2
8 ; SI: buffer_load_dwordx2
9 ; SI: buffer_store_dwordx2
10 ; SI: buffer_store_dwordx2
11 ; SI: buffer_store_dwordx2
12 ; SI: buffer_store_dwordx2
14 define void @no_reorder_v2f64_global_load_store(<2 x double> addrspace(1)* nocapture %x, <2 x double> addrspace(1)* nocapture %y) nounwind {
15 %tmp1 = load <2 x double>, <2 x double> addrspace(1)* %x, align 16
16 %tmp4 = load <2 x double>, <2 x double> addrspace(1)* %y, align 16
17 store <2 x double> %tmp4, <2 x double> addrspace(1)* %x, align 16
18 store <2 x double> %tmp1, <2 x double> addrspace(1)* %y, align 16
22 ; SI-LABEL: {{^}}no_reorder_scalarized_v2f64_local_load_store:
28 define void @no_reorder_scalarized_v2f64_local_load_store(<2 x double> addrspace(3)* nocapture %x, <2 x double> addrspace(3)* nocapture %y) nounwind {
29 %tmp1 = load <2 x double>, <2 x double> addrspace(3)* %x, align 16
30 %tmp4 = load <2 x double>, <2 x double> addrspace(3)* %y, align 16
31 store <2 x double> %tmp4, <2 x double> addrspace(3)* %x, align 16
32 store <2 x double> %tmp1, <2 x double> addrspace(3)* %y, align 16
36 ; SI-LABEL: {{^}}no_reorder_split_v8i32_global_load_store:
37 ; SI: buffer_load_dwordx4
38 ; SI: buffer_load_dwordx4
39 ; SI: buffer_load_dwordx4
40 ; SI: buffer_load_dwordx4
43 ; SI: buffer_store_dwordx4
44 ; SI: buffer_store_dwordx4
45 ; SI: buffer_store_dwordx4
46 ; SI: buffer_store_dwordx4
48 define void @no_reorder_split_v8i32_global_load_store(<8 x i32> addrspace(1)* nocapture %x, <8 x i32> addrspace(1)* nocapture %y) nounwind {
49 %tmp1 = load <8 x i32>, <8 x i32> addrspace(1)* %x, align 32
50 %tmp4 = load <8 x i32>, <8 x i32> addrspace(1)* %y, align 32
51 store <8 x i32> %tmp4, <8 x i32> addrspace(1)* %x, align 32
52 store <8 x i32> %tmp1, <8 x i32> addrspace(1)* %y, align 32
56 ; SI-LABEL: {{^}}no_reorder_extload_64:
63 define void @no_reorder_extload_64(<2 x i32> addrspace(3)* nocapture %x, <2 x i32> addrspace(3)* nocapture %y) nounwind {
64 %tmp1 = load <2 x i32>, <2 x i32> addrspace(3)* %x, align 8
65 %tmp4 = load <2 x i32>, <2 x i32> addrspace(3)* %y, align 8
66 %tmp1ext = zext <2 x i32> %tmp1 to <2 x i64>
67 %tmp4ext = zext <2 x i32> %tmp4 to <2 x i64>
68 %tmp7 = add <2 x i64> %tmp1ext, <i64 1, i64 1>
69 %tmp9 = add <2 x i64> %tmp4ext, <i64 1, i64 1>
70 %trunctmp9 = trunc <2 x i64> %tmp9 to <2 x i32>
71 %trunctmp7 = trunc <2 x i64> %tmp7 to <2 x i32>
72 store <2 x i32> %trunctmp9, <2 x i32> addrspace(3)* %x, align 8
73 store <2 x i32> %trunctmp7, <2 x i32> addrspace(3)* %y, align 8