| ; RUN: llc -march=amdgcn -mcpu=kaveri -mtriple=amdgcn-unknown-amdhsa -mattr=-flat-for-global < %s | FileCheck -check-prefix=GCN %s |
| |
| ; Check that when mubuf addr64 instruction is handled in moveToVALU |
| ; from the pointer, dead register writes are not emitted. |
| |
| ; FIXME: We should be able to use the SGPR directly as src0 to v_add_i32 |
| |
| ; GCN-LABEL: {{^}}clobber_vgpr_pair_pointer_add: |
| ; GCN: s_load_dwordx2 s{{\[}}[[ARG1LO:[0-9]+]]:[[ARG1HI:[0-9]+]]{{\]}}, s{{\[[0-9]+:[0-9]+\]}}, 0x0{{$}} |
| |
| ; GCN-NOT: v_mov_b32 |
| ; GCN: v_mov_b32_e32 v[[VARG1LO:[0-9]+]], s[[ARG1LO]] |
| ; GCN: buffer_load_dwordx2 v{{\[}}[[LDPTRLO:[0-9]+]]:[[LDPTRHI:[0-9]+]]{{\]}} |
| ; GCN-NOT: v_mov_b32 |
| ; GCN: v_mov_b32_e32 v[[VARG1HI:[0-9]+]], s[[ARG1HI]] |
| ; GCN-NOT: v_mov_b32 |
| |
| ; GCN: v_add_i32_e32 v[[PTRLO:[0-9]+]], vcc, v[[LDPTRLO]], v[[VARG1LO]] |
| ; GCN: v_addc_u32_e32 v[[PTRHI:[0-9]+]], vcc, v[[LDPTRHI]], v[[VARG1HI]] |
| ; GCN: buffer_load_ubyte v{{[0-9]+}}, v{{\[}}[[PTRLO]]:[[PTRHI]]{{\]}}, |
| |
| define amdgpu_kernel void @clobber_vgpr_pair_pointer_add(i64 %arg1, [8 x i32], i8 addrspace(1)* addrspace(1)* %ptrarg, i32 %arg3) #0 { |
| bb: |
| %tmp = icmp sgt i32 %arg3, 0 |
| br i1 %tmp, label %bb4, label %bb17 |
| |
| bb4: |
| %tmp14 = load volatile i8 addrspace(1)*, i8 addrspace(1)* addrspace(1)* %ptrarg |
| %tmp15 = getelementptr inbounds i8, i8 addrspace(1)* %tmp14, i64 %arg1 |
| %tmp16 = load volatile i8, i8 addrspace(1)* %tmp15 |
| br label %bb17 |
| |
| bb17: |
| ret void |
| } |
| |
| attributes #0 = { nounwind } |
