1 ; RUN: llc -march=amdgcn -verify-machineinstrs -mattr=+load-store-opt < %s | FileCheck -check-prefix=SI --check-prefix=CHECK %s
2 ; RUN: llc -march=amdgcn -mcpu=bonaire -verify-machineinstrs -mattr=+load-store-opt < %s | FileCheck -check-prefix=CI --check-prefix=CHECK %s
3 ; RUN: llc -march=amdgcn -verify-machineinstrs -mattr=+load-store-opt,+unsafe-ds-offset-folding < %s | FileCheck -check-prefix=CI --check-prefix=CHECK %s
5 declare i32 @llvm.amdgcn.workitem.id.x() #0
6 declare void @llvm.amdgcn.s.barrier() #1
8 ; Function Attrs: nounwind
9 ; CHECK-LABEL: {{^}}signed_ds_offset_addressing_loop:
11 ; CHECK: v_add_i32_e32 [[VADDR:v[0-9]+]],
12 ; SI-DAG: ds_read_b32 v{{[0-9]+}}, [[VADDR]]
13 ; SI-DAG: v_add_i32_e32 [[VADDR8:v[0-9]+]], vcc, 8, [[VADDR]]
14 ; SI-DAG: ds_read_b32 v{{[0-9]+}}, [[VADDR8]]
15 ; SI-DAG: v_add_i32_e32 [[VADDR0x80:v[0-9]+]], vcc, 0x80, [[VADDR]]
16 ; SI-DAG: ds_read_b32 v{{[0-9]+}}, [[VADDR0x80]]
17 ; SI-DAG: v_add_i32_e32 [[VADDR0x88:v[0-9]+]], vcc, 0x88, [[VADDR]]
18 ; SI-DAG: ds_read_b32 v{{[0-9]+}}, [[VADDR0x88]]
19 ; SI-DAG: v_add_i32_e32 [[VADDR0x100:v[0-9]+]], vcc, 0x100, [[VADDR]]
20 ; SI-DAG: ds_read_b32 v{{[0-9]+}}, [[VADDR0x100]]
22 ; CI-DAG: ds_read2_b32 v{{\[[0-9]+:[0-9]+\]}}, [[VADDR]] offset1:2
23 ; CI-DAG: ds_read2_b32 v{{\[[0-9]+:[0-9]+\]}}, [[VADDR]] offset0:32 offset1:34
24 ; CI-DAG: ds_read_b32 v{{[0-9]+}}, [[VADDR]] offset:256
26 define amdgpu_kernel void @signed_ds_offset_addressing_loop(float addrspace(1)* noalias nocapture %out, float addrspace(3)* noalias nocapture readonly %lptr, i32 %n) #2 {
28 %x.i = tail call i32 @llvm.amdgcn.workitem.id.x() #0
29 %mul = shl nsw i32 %x.i, 1
32 for.body: ; preds = %for.body, %entry
33 %sum.03 = phi float [ 0.000000e+00, %entry ], [ %add13, %for.body ]
34 %offset.02 = phi i32 [ %mul, %entry ], [ %add14, %for.body ]
35 %k.01 = phi i32 [ 0, %entry ], [ %inc, %for.body ]
36 tail call void @llvm.amdgcn.s.barrier() #1
37 %arrayidx = getelementptr inbounds float, float addrspace(3)* %lptr, i32 %offset.02
38 %tmp = load float, float addrspace(3)* %arrayidx, align 4
39 %add1 = add nsw i32 %offset.02, 2
40 %arrayidx2 = getelementptr inbounds float, float addrspace(3)* %lptr, i32 %add1
41 %tmp1 = load float, float addrspace(3)* %arrayidx2, align 4
42 %add3 = add nsw i32 %offset.02, 32
43 %arrayidx4 = getelementptr inbounds float, float addrspace(3)* %lptr, i32 %add3
44 %tmp2 = load float, float addrspace(3)* %arrayidx4, align 4
45 %add5 = add nsw i32 %offset.02, 34
46 %arrayidx6 = getelementptr inbounds float, float addrspace(3)* %lptr, i32 %add5
47 %tmp3 = load float, float addrspace(3)* %arrayidx6, align 4
48 %add7 = add nsw i32 %offset.02, 64
49 %arrayidx8 = getelementptr inbounds float, float addrspace(3)* %lptr, i32 %add7
50 %tmp4 = load float, float addrspace(3)* %arrayidx8, align 4
51 %add9 = fadd float %tmp, %tmp1
52 %add10 = fadd float %add9, %tmp2
53 %add11 = fadd float %add10, %tmp3
54 %add12 = fadd float %add11, %tmp4
55 %add13 = fadd float %sum.03, %add12
56 %inc = add nsw i32 %k.01, 1
57 %add14 = add nsw i32 %offset.02, 97
58 %exitcond = icmp eq i32 %inc, 8
59 br i1 %exitcond, label %for.end, label %for.body
61 for.end: ; preds = %for.body
62 %tmp5 = sext i32 %x.i to i64
63 %arrayidx15 = getelementptr inbounds float, float addrspace(1)* %out, i64 %tmp5
64 store float %add13, float addrspace(1)* %arrayidx15, align 4
68 attributes #0 = { nounwind readnone }
69 attributes #1 = { convergent nounwind }
70 attributes #2 = { nounwind }