114 lines · plain
1; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py UTC_ARGS: --version 52; RUN: llc -mtriple=amdgcn < %s | FileCheck -check-prefix=GCN %s3 4define amdgpu_kernel void @test_i128_vreg(ptr addrspace(1) noalias %out, ptr addrspace(1) noalias %inA, ptr addrspace(1) noalias %inB) {5; GCN-LABEL: test_i128_vreg:6; GCN: ; %bb.0:7; GCN-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x98; GCN-NEXT: s_load_dwordx2 s[4:5], s[4:5], 0xd9; GCN-NEXT: s_mov_b32 s11, 0xf00010; GCN-NEXT: s_mov_b32 s14, 011; GCN-NEXT: v_lshlrev_b32_e32 v4, 4, v012; GCN-NEXT: v_mov_b32_e32 v5, 013; GCN-NEXT: s_mov_b32 s15, s1114; GCN-NEXT: s_mov_b64 s[6:7], s[14:15]15; GCN-NEXT: s_waitcnt lgkmcnt(0)16; GCN-NEXT: s_mov_b64 s[12:13], s[2:3]17; GCN-NEXT: buffer_load_dwordx4 v[0:3], v[4:5], s[12:15], 0 addr6418; GCN-NEXT: buffer_load_dwordx4 v[4:7], v[4:5], s[4:7], 0 addr6419; GCN-NEXT: s_mov_b32 s10, -120; GCN-NEXT: s_mov_b32 s8, s021; GCN-NEXT: s_mov_b32 s9, s122; GCN-NEXT: s_waitcnt vmcnt(0)23; GCN-NEXT: v_add_i32_e32 v0, vcc, v0, v424; GCN-NEXT: v_addc_u32_e32 v1, vcc, v1, v5, vcc25; GCN-NEXT: v_addc_u32_e32 v2, vcc, v2, v6, vcc26; GCN-NEXT: v_addc_u32_e32 v3, vcc, v3, v7, vcc27; GCN-NEXT: buffer_store_dwordx4 v[0:3], off, s[8:11], 028; GCN-NEXT: s_endpgm29 %tid = call i32 @llvm.amdgcn.workitem.id.x() readnone30 %a_ptr = getelementptr i128, ptr addrspace(1) %inA, i32 %tid31 %b_ptr = getelementptr i128, ptr addrspace(1) %inB, i32 %tid32 %a = load i128, ptr addrspace(1) %a_ptr33 %b = load i128, ptr addrspace(1) %b_ptr34 %result = add i128 %a, %b35 store i128 %result, ptr addrspace(1) %out36 ret void37}38 39; Check that the SGPR add operand is correctly moved to a VGPR.40define amdgpu_kernel void @sgpr_operand(ptr addrspace(1) noalias %out, ptr addrspace(1) noalias %in, i128 %a) {41; GCN-LABEL: sgpr_operand:42; GCN: ; %bb.0:43; GCN-NEXT: s_load_dwordx8 s[0:7], s[4:5], 0x944; GCN-NEXT: s_waitcnt lgkmcnt(0)45; GCN-NEXT: s_load_dwordx4 s[8:11], s[2:3], 0x046; GCN-NEXT: s_mov_b32 s3, 0xf00047; GCN-NEXT: s_waitcnt lgkmcnt(0)48; GCN-NEXT: s_add_u32 s4, s8, s449; GCN-NEXT: s_addc_u32 s5, s9, s550; GCN-NEXT: s_addc_u32 s6, s10, s651; GCN-NEXT: s_addc_u32 s7, s11, s752; GCN-NEXT: s_mov_b32 s2, -153; GCN-NEXT: v_mov_b32_e32 v0, s454; GCN-NEXT: v_mov_b32_e32 v1, s555; GCN-NEXT: v_mov_b32_e32 v2, s656; GCN-NEXT: v_mov_b32_e32 v3, s757; GCN-NEXT: buffer_store_dwordx4 v[0:3], off, s[0:3], 058; GCN-NEXT: s_endpgm59 %foo = load i128, ptr addrspace(1) %in, align 860 %result = add i128 %foo, %a61 store i128 %result, ptr addrspace(1) %out62 ret void63}64 65define amdgpu_kernel void @sgpr_operand_reversed(ptr addrspace(1) noalias %out, ptr addrspace(1) noalias %in, i128 %a) {66; GCN-LABEL: sgpr_operand_reversed:67; GCN: ; %bb.0:68; GCN-NEXT: s_load_dwordx8 s[0:7], s[4:5], 0x969; GCN-NEXT: s_waitcnt lgkmcnt(0)70; GCN-NEXT: s_load_dwordx4 s[8:11], s[2:3], 0x071; GCN-NEXT: s_mov_b32 s3, 0xf00072; GCN-NEXT: s_waitcnt lgkmcnt(0)73; GCN-NEXT: s_add_u32 s4, s4, s874; GCN-NEXT: s_addc_u32 s5, s5, s975; GCN-NEXT: s_addc_u32 s6, s6, s1076; GCN-NEXT: s_addc_u32 s7, s7, s1177; GCN-NEXT: s_mov_b32 s2, -178; GCN-NEXT: v_mov_b32_e32 v0, s479; GCN-NEXT: v_mov_b32_e32 v1, s580; GCN-NEXT: v_mov_b32_e32 v2, s681; GCN-NEXT: v_mov_b32_e32 v3, s782; GCN-NEXT: buffer_store_dwordx4 v[0:3], off, s[0:3], 083; GCN-NEXT: s_endpgm84 %foo = load i128, ptr addrspace(1) %in, align 885 %result = add i128 %a, %foo86 store i128 %result, ptr addrspace(1) %out87 ret void88}89 90define amdgpu_kernel void @test_sreg(ptr addrspace(1) noalias %out, i128 %a, i128 %b) {91; GCN-LABEL: test_sreg:92; GCN: ; %bb.0:93; GCN-NEXT: s_load_dwordx8 s[8:15], s[4:5], 0xb94; GCN-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x995; GCN-NEXT: s_mov_b32 s3, 0xf00096; GCN-NEXT: s_waitcnt lgkmcnt(0)97; GCN-NEXT: s_add_u32 s4, s8, s1298; GCN-NEXT: s_addc_u32 s5, s9, s1399; GCN-NEXT: s_addc_u32 s6, s10, s14100; GCN-NEXT: s_addc_u32 s7, s11, s15101; GCN-NEXT: s_mov_b32 s2, -1102; GCN-NEXT: v_mov_b32_e32 v0, s4103; GCN-NEXT: v_mov_b32_e32 v1, s5104; GCN-NEXT: v_mov_b32_e32 v2, s6105; GCN-NEXT: v_mov_b32_e32 v3, s7106; GCN-NEXT: buffer_store_dwordx4 v[0:3], off, s[0:3], 0107; GCN-NEXT: s_endpgm108 %result = add i128 %a, %b109 store i128 %result, ptr addrspace(1) %out110 ret void111}112 113declare i32 @llvm.amdgcn.workitem.id.x() readnone114