| ; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py | |
| ; RUN: llc -global-isel -O0 -mtriple=amdgpu10.31-amd-amdhsa < %s | FileCheck %s | |
| ; Testcase which happened to trigger a liveness verifier error | |
| define amdgpu_kernel void @test_long_add4(<4 x i64> %arg) #0 { | |
| ; CHECK-LABEL: test_long_add4: | |
| ; CHECK: ; %bb.0: ; %entry | |
| ; CHECK-NEXT: s_mov_b64 s[6:7], 0 | |
| ; CHECK-NEXT: v_mov_b32_e32 v0, s6 | |
| ; CHECK-NEXT: v_mov_b32_e32 v1, s7 | |
| ; CHECK-NEXT: global_load_dwordx4 v[10:13], v[0:1], off | |
| ; CHECK-NEXT: s_mov_b64 s[4:5], 16 | |
| ; CHECK-NEXT: v_mov_b32_e32 v0, s4 | |
| ; CHECK-NEXT: v_mov_b32_e32 v1, s5 | |
| ; CHECK-NEXT: global_load_dwordx4 v[0:3], v[0:1], off | |
| ; CHECK-NEXT: ; kill: def $vgpr10_vgpr11_vgpr12_vgpr13 killed $vgpr10_vgpr11_vgpr12_vgpr13 def $vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15_vgpr16_vgpr17 killed $exec | |
| ; CHECK-NEXT: s_waitcnt vmcnt(0) | |
| ; CHECK-NEXT: v_mov_b32_e32 v17, v3 | |
| ; CHECK-NEXT: v_mov_b32_e32 v16, v2 | |
| ; CHECK-NEXT: v_mov_b32_e32 v15, v1 | |
| ; CHECK-NEXT: v_mov_b32_e32 v14, v0 | |
| ; CHECK-NEXT: v_mov_b32_e32 v8, v10 | |
| ; CHECK-NEXT: v_mov_b32_e32 v9, v11 | |
| ; CHECK-NEXT: v_mov_b32_e32 v6, v12 | |
| ; CHECK-NEXT: v_mov_b32_e32 v7, v13 | |
| ; CHECK-NEXT: v_mov_b32_e32 v4, v14 | |
| ; CHECK-NEXT: v_mov_b32_e32 v5, v15 | |
| ; CHECK-NEXT: v_mov_b32_e32 v2, v16 | |
| ; CHECK-NEXT: v_mov_b32_e32 v3, v17 | |
| ; CHECK-NEXT: v_mov_b32_e32 v1, v8 | |
| ; CHECK-NEXT: v_mov_b32_e32 v0, v9 | |
| ; CHECK-NEXT: v_readfirstlane_b32 s12, v1 | |
| ; CHECK-NEXT: v_readfirstlane_b32 s13, v0 | |
| ; CHECK-NEXT: v_mov_b32_e32 v1, v6 | |
| ; CHECK-NEXT: v_mov_b32_e32 v0, v7 | |
| ; CHECK-NEXT: v_readfirstlane_b32 s10, v1 | |
| ; CHECK-NEXT: v_readfirstlane_b32 s11, v0 | |
| ; CHECK-NEXT: v_mov_b32_e32 v1, v4 | |
| ; CHECK-NEXT: v_mov_b32_e32 v0, v5 | |
| ; CHECK-NEXT: v_readfirstlane_b32 s8, v1 | |
| ; CHECK-NEXT: v_readfirstlane_b32 s9, v0 | |
| ; CHECK-NEXT: v_mov_b32_e32 v1, v2 | |
| ; CHECK-NEXT: v_mov_b32_e32 v0, v3 | |
| ; CHECK-NEXT: v_readfirstlane_b32 s16, v1 | |
| ; CHECK-NEXT: v_readfirstlane_b32 s14, v0 | |
| ; CHECK-NEXT: v_mov_b32_e32 v0, s6 | |
| ; CHECK-NEXT: v_mov_b32_e32 v1, s7 | |
| ; CHECK-NEXT: global_load_dwordx4 v[10:13], v[0:1], off | |
| ; CHECK-NEXT: v_mov_b32_e32 v0, s4 | |
| ; CHECK-NEXT: v_mov_b32_e32 v1, s5 | |
| ; CHECK-NEXT: global_load_dwordx4 v[0:3], v[0:1], off | |
| ; CHECK-NEXT: ; kill: def $vgpr10_vgpr11_vgpr12_vgpr13 killed $vgpr10_vgpr11_vgpr12_vgpr13 def $vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15_vgpr16_vgpr17 killed $exec | |
| ; CHECK-NEXT: s_waitcnt vmcnt(0) | |
| ; CHECK-NEXT: v_mov_b32_e32 v17, v3 | |
| ; CHECK-NEXT: v_mov_b32_e32 v16, v2 | |
| ; CHECK-NEXT: v_mov_b32_e32 v15, v1 | |
| ; CHECK-NEXT: v_mov_b32_e32 v14, v0 | |
| ; CHECK-NEXT: v_mov_b32_e32 v8, v10 | |
| ; CHECK-NEXT: v_mov_b32_e32 v9, v11 | |
| ; CHECK-NEXT: v_mov_b32_e32 v6, v12 | |
| ; CHECK-NEXT: v_mov_b32_e32 v7, v13 | |
| ; CHECK-NEXT: v_mov_b32_e32 v4, v14 | |
| ; CHECK-NEXT: v_mov_b32_e32 v5, v15 | |
| ; CHECK-NEXT: v_mov_b32_e32 v2, v16 | |
| ; CHECK-NEXT: v_mov_b32_e32 v3, v17 | |
| ; CHECK-NEXT: v_mov_b32_e32 v1, v8 | |
| ; CHECK-NEXT: v_mov_b32_e32 v0, v9 | |
| ; CHECK-NEXT: v_readfirstlane_b32 s23, v1 | |
| ; CHECK-NEXT: v_readfirstlane_b32 s22, v0 | |
| ; CHECK-NEXT: v_mov_b32_e32 v1, v6 | |
| ; CHECK-NEXT: v_mov_b32_e32 v0, v7 | |
| ; CHECK-NEXT: v_readfirstlane_b32 s21, v1 | |
| ; CHECK-NEXT: v_readfirstlane_b32 s20, v0 | |
| ; CHECK-NEXT: v_mov_b32_e32 v1, v4 | |
| ; CHECK-NEXT: v_mov_b32_e32 v0, v5 | |
| ; CHECK-NEXT: v_readfirstlane_b32 s19, v1 | |
| ; CHECK-NEXT: v_readfirstlane_b32 s18, v0 | |
| ; CHECK-NEXT: v_mov_b32_e32 v1, v2 | |
| ; CHECK-NEXT: v_mov_b32_e32 v0, v3 | |
| ; CHECK-NEXT: v_readfirstlane_b32 s17, v1 | |
| ; CHECK-NEXT: v_readfirstlane_b32 s15, v0 | |
| ; CHECK-NEXT: s_add_u32 s12, s12, s23 | |
| ; CHECK-NEXT: s_cselect_b32 s23, 1, 0 | |
| ; CHECK-NEXT: s_cmp_lg_u32 s23, 0 | |
| ; CHECK-NEXT: s_addc_u32 s22, s13, s22 | |
| ; CHECK-NEXT: ; kill: def $sgpr12 killed $sgpr12 def $sgpr12_sgpr13 | |
| ; CHECK-NEXT: s_mov_b32 s13, s22 | |
| ; CHECK-NEXT: s_add_u32 s10, s10, s21 | |
| ; CHECK-NEXT: s_cselect_b32 s21, 1, 0 | |
| ; CHECK-NEXT: s_cmp_lg_u32 s21, 0 | |
| ; CHECK-NEXT: s_addc_u32 s20, s11, s20 | |
| ; CHECK-NEXT: ; kill: def $sgpr10 killed $sgpr10 def $sgpr10_sgpr11 | |
| ; CHECK-NEXT: s_mov_b32 s11, s20 | |
| ; CHECK-NEXT: s_add_u32 s8, s8, s19 | |
| ; CHECK-NEXT: s_cselect_b32 s19, 1, 0 | |
| ; CHECK-NEXT: s_cmp_lg_u32 s19, 0 | |
| ; CHECK-NEXT: s_addc_u32 s18, s9, s18 | |
| ; CHECK-NEXT: ; kill: def $sgpr8 killed $sgpr8 def $sgpr8_sgpr9 | |
| ; CHECK-NEXT: s_mov_b32 s9, s18 | |
| ; CHECK-NEXT: s_add_u32 s16, s16, s17 | |
| ; CHECK-NEXT: s_cselect_b32 s17, 1, 0 | |
| ; CHECK-NEXT: s_cmp_lg_u32 s17, 0 | |
| ; CHECK-NEXT: s_addc_u32 s14, s14, s15 | |
| ; CHECK-NEXT: ; kill: def $sgpr16 killed $sgpr16 def $sgpr16_sgpr17 | |
| ; CHECK-NEXT: s_mov_b32 s17, s14 | |
| ; CHECK-NEXT: ; kill: def $sgpr12_sgpr13 killed $sgpr12_sgpr13 def $sgpr12_sgpr13_sgpr14_sgpr15 | |
| ; CHECK-NEXT: s_mov_b64 s[14:15], s[10:11] | |
| ; CHECK-NEXT: ; kill: def $sgpr8_sgpr9 killed $sgpr8_sgpr9 def $sgpr8_sgpr9_sgpr10_sgpr11 | |
| ; CHECK-NEXT: s_mov_b64 s[10:11], s[16:17] | |
| ; CHECK-NEXT: v_mov_b32_e32 v2, s12 | |
| ; CHECK-NEXT: v_mov_b32_e32 v3, s13 | |
| ; CHECK-NEXT: v_mov_b32_e32 v4, s14 | |
| ; CHECK-NEXT: v_mov_b32_e32 v5, s15 | |
| ; CHECK-NEXT: v_mov_b32_e32 v0, s6 | |
| ; CHECK-NEXT: v_mov_b32_e32 v1, s7 | |
| ; CHECK-NEXT: global_store_dwordx4 v[0:1], v[2:5], off | |
| ; CHECK-NEXT: v_mov_b32_e32 v2, s8 | |
| ; CHECK-NEXT: v_mov_b32_e32 v3, s9 | |
| ; CHECK-NEXT: v_mov_b32_e32 v4, s10 | |
| ; CHECK-NEXT: v_mov_b32_e32 v5, s11 | |
| ; CHECK-NEXT: v_mov_b32_e32 v0, s4 | |
| ; CHECK-NEXT: v_mov_b32_e32 v1, s5 | |
| ; CHECK-NEXT: global_store_dwordx4 v[0:1], v[2:5], off | |
| ; CHECK-NEXT: s_endpgm | |
| entry: | |
| %load0 = load <4 x i64>, ptr addrspace(1) null, align 32 | |
| %load1 = load <4 x i64>, ptr addrspace(1) null, align 32 | |
| %add = add <4 x i64> %load0, %load1 | |
| store <4 x i64> %add, ptr addrspace(1) null, align 32 | |
| ret void | |
| } | |
| attributes #0 = { noinline optnone } |