49 lines · plain
1; RUN: llc -mtriple=amdgcn -mcpu=bonaire < %s | FileCheck -check-prefix=GCN %s2 3; The memory operand was dropped from the buffer_load_dword_offset4; when replaced with the addr64 during operand legalization, resulting5; in the global loads not being scheduled together.6 7; GCN-LABEL: {{^}}reschedule_global_load_lds_store:8; GCN: buffer_load_dword9; GCN: buffer_load_dword10; GCN: ds_write2_b3211; GCN: s_endpgm12define amdgpu_kernel void @reschedule_global_load_lds_store(ptr addrspace(1) noalias %gptr0, ptr addrspace(1) noalias %gptr1, ptr addrspace(3) noalias %lptr, i32 %c) #0 {13entry:14 %tid = tail call i32 @llvm.amdgcn.workitem.id.x() #115 %idx = shl i32 %tid, 216 %gep0 = getelementptr i32, ptr addrspace(1) %gptr0, i32 %idx17 %gep1 = getelementptr i32, ptr addrspace(1) %gptr1, i32 %idx18 %gep2 = getelementptr i32, ptr addrspace(3) %lptr, i32 %tid19 %cmp0 = icmp eq i32 %c, 020 br i1 %cmp0, label %for.body, label %exit21 22for.body: ; preds = %for.body, %entry23 %i = phi i32 [ 0, %entry ], [ %i.inc, %for.body ]24 %gptr0.phi = phi ptr addrspace(1) [ %gep0, %entry ], [ %gep0.inc, %for.body ]25 %gptr1.phi = phi ptr addrspace(1) [ %gep1, %entry ], [ %gep1.inc, %for.body ]26 %lptr0.phi = phi ptr addrspace(3) [ %gep2, %entry ], [ %gep2.inc, %for.body ]27 %lptr1 = getelementptr i32, ptr addrspace(3) %lptr0.phi, i32 228 %val0 = load i32, ptr addrspace(1) %gep029 store i32 %val0, ptr addrspace(3) %lptr0.phi30 %val1 = load i32, ptr addrspace(1) %gep131 store i32 %val1, ptr addrspace(3) %lptr132 %gep0.inc = getelementptr i32, ptr addrspace(1) %gptr0.phi, i32 433 %gep1.inc = getelementptr i32, ptr addrspace(1) %gptr1.phi, i32 434 %gep2.inc = getelementptr i32, ptr addrspace(3) %lptr0.phi, i32 435 %i.inc = add nsw i32 %i, 136 %cmp1 = icmp ne i32 %i, 25637 br i1 %cmp1, label %for.body, label %exit38 39exit: ; preds = %for.body, %entry40 ret void41}42 43; Function Attrs: nounwind readnone44declare i32 @llvm.amdgcn.workitem.id.x() #145 46attributes #0 = { nounwind }47attributes #1 = { nounwind readnone }48attributes #2 = { convergent nounwind }49