|
150
|
1 ; RUN: llc -march=amdgcn -mcpu=bonaire -verify-machineinstrs < %s | FileCheck -check-prefix=GCN -check-prefix=CI %s
|
|
|
2
|
|
|
3 ; The memory operand was dropped from the buffer_load_dword_offset
|
|
|
4 ; when replaced with the addr64 during operand legalization, resulting
|
|
|
5 ; in the global loads not being scheduled together.
|
|
|
6
|
|
|
7 ; GCN-LABEL: {{^}}reschedule_global_load_lds_store:
|
|
|
8 ; GCN: buffer_load_dword
|
|
|
9 ; GCN: buffer_load_dword
|
|
|
10 ; GCN: ds_write2_b32
|
|
|
11 ; GCN: s_endpgm
|
|
|
12 define amdgpu_kernel void @reschedule_global_load_lds_store(i32 addrspace(1)* noalias %gptr0, i32 addrspace(1)* noalias %gptr1, i32 addrspace(3)* noalias %lptr, i32 %c) #0 {
|
|
|
13 entry:
|
|
|
14 %tid = tail call i32 @llvm.amdgcn.workitem.id.x() #1
|
|
|
15 %idx = shl i32 %tid, 2
|
|
|
16 %gep0 = getelementptr i32, i32 addrspace(1)* %gptr0, i32 %idx
|
|
|
17 %gep1 = getelementptr i32, i32 addrspace(1)* %gptr1, i32 %idx
|
|
|
18 %gep2 = getelementptr i32, i32 addrspace(3)* %lptr, i32 %tid
|
|
|
19 %cmp0 = icmp eq i32 %c, 0
|
|
|
20 br i1 %cmp0, label %for.body, label %exit
|
|
|
21
|
|
|
22 for.body: ; preds = %for.body, %entry
|
|
|
23 %i = phi i32 [ 0, %entry ], [ %i.inc, %for.body ]
|
|
|
24 %gptr0.phi = phi i32 addrspace(1)* [ %gep0, %entry ], [ %gep0.inc, %for.body ]
|
|
|
25 %gptr1.phi = phi i32 addrspace(1)* [ %gep1, %entry ], [ %gep1.inc, %for.body ]
|
|
|
26 %lptr0.phi = phi i32 addrspace(3)* [ %gep2, %entry ], [ %gep2.inc, %for.body ]
|
|
|
27 %lptr1 = getelementptr i32, i32 addrspace(3)* %lptr0.phi, i32 2
|
|
|
28 %val0 = load i32, i32 addrspace(1)* %gep0
|
|
|
29 store i32 %val0, i32 addrspace(3)* %lptr0.phi
|
|
|
30 %val1 = load i32, i32 addrspace(1)* %gep1
|
|
|
31 store i32 %val1, i32 addrspace(3)* %lptr1
|
|
|
32 %gep0.inc = getelementptr i32, i32 addrspace(1)* %gptr0.phi, i32 4
|
|
|
33 %gep1.inc = getelementptr i32, i32 addrspace(1)* %gptr1.phi, i32 4
|
|
|
34 %gep2.inc = getelementptr i32, i32 addrspace(3)* %lptr0.phi, i32 4
|
|
|
35 %i.inc = add nsw i32 %i, 1
|
|
|
36 %cmp1 = icmp ne i32 %i, 256
|
|
|
37 br i1 %cmp1, label %for.body, label %exit
|
|
|
38
|
|
|
39 exit: ; preds = %for.body, %entry
|
|
|
40 ret void
|
|
|
41 }
|
|
|
42
|
|
|
43 ; Function Attrs: nounwind readnone
|
|
|
44 declare i32 @llvm.amdgcn.workitem.id.x() #1
|
|
|
45
|
|
|
46 attributes #0 = { nounwind }
|
|
|
47 attributes #1 = { nounwind readnone }
|
|
|
48 attributes #2 = { convergent nounwind }
|
