1; RUN: llc -march=amdgcn -mcpu=verde -verify-machineinstrs < %s | FileCheck -check-prefix=GCN -check-prefix=FUNC %s 2; RUN: llc -march=amdgcn -mcpu=bonaire -verify-machineinstrs < %s | FileCheck -check-prefix=GCN -check-prefix=FUNC %s 3; RUN: llc -march=r600 -mcpu=redwood < %s | FileCheck -check-prefix=EG -check-prefix=FUNC %s 4 5@local_memory.local_mem = internal unnamed_addr addrspace(3) global [128 x i32] undef, align 4 6 7@lds = addrspace(3) global [512 x i32] undef, align 4 8 9; On SI we need to make sure that the base offset is a register and 10; not an immediate. 11 12; FUNC-LABEL: {{^}}load_i32_local_const_ptr: 13; GCN: v_mov_b32_e32 v[[PTR:[0-9]+]], lds@abs32@lo 14; GCN: ds_read_b32 v{{[0-9]+}}, v[[PTR]] offset:4 15 16; R600: LDS_READ_RET 17define amdgpu_kernel void @load_i32_local_const_ptr(i32 addrspace(1)* %out, i32 addrspace(3)* %in) #0 { 18entry: 19 %tmp0 = getelementptr [512 x i32], [512 x i32] addrspace(3)* @lds, i32 0, i32 1 20 %tmp1 = load i32, i32 addrspace(3)* %tmp0 21 %tmp2 = getelementptr i32, i32 addrspace(1)* %out, i32 1 22 store i32 %tmp1, i32 addrspace(1)* %tmp2 23 ret void 24} 25 26; Test loading a i32 and v2i32 value from the same base pointer. 27; FUNC-LABEL: {{^}}load_i32_v2i32_local: 28; R600: LDS_READ_RET 29; R600: LDS_READ_RET 30; R600: LDS_READ_RET 31; GCN-DAG: ds_read_b32 32; GCN-DAG: ds_read2_b32 33define amdgpu_kernel void @load_i32_v2i32_local(<2 x i32> addrspace(1)* %out, i32 addrspace(3)* %in) #0 { 34 %scalar = load i32, i32 addrspace(3)* %in 35 %tmp0 = bitcast i32 addrspace(3)* %in to <2 x i32> addrspace(3)* 36 %vec_ptr = getelementptr <2 x i32>, <2 x i32> addrspace(3)* %tmp0, i32 2 37 %vec0 = load <2 x i32>, <2 x i32> addrspace(3)* %vec_ptr, align 4 38 %vec1 = insertelement <2 x i32> <i32 0, i32 0>, i32 %scalar, i32 0 39 %vec = add <2 x i32> %vec0, %vec1 40 store <2 x i32> %vec, <2 x i32> addrspace(1)* %out 41 ret void 42} 43 44attributes #0 = { nounwind } 45