169 lines · plain
1; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py2; RUN: llc -mtriple=amdgcn < %s | FileCheck -check-prefix=GCN %s3 4define amdgpu_kernel void @negated_cond(ptr addrspace(1) %arg1) {5; GCN-LABEL: negated_cond:6; GCN: ; %bb.0: ; %bb7; GCN-NEXT: s_load_dwordx2 s[4:5], s[4:5], 0x98; GCN-NEXT: s_mov_b32 s7, 0xf0009; GCN-NEXT: s_mov_b32 s10, -110; GCN-NEXT: s_mov_b32 s6, 011; GCN-NEXT: s_mov_b32 s11, s712; GCN-NEXT: s_waitcnt lgkmcnt(0)13; GCN-NEXT: s_mov_b32 s8, s414; GCN-NEXT: s_mov_b32 s9, s515; GCN-NEXT: v_mov_b32_e32 v0, 016; GCN-NEXT: s_branch .LBB0_217; GCN-NEXT: .LBB0_1: ; %loop.exit.guard18; GCN-NEXT: ; in Loop: Header=BB0_2 Depth=119; GCN-NEXT: s_and_b64 vcc, exec, s[14:15]20; GCN-NEXT: s_cbranch_vccnz .LBB0_921; GCN-NEXT: .LBB0_2: ; %bb122; GCN-NEXT: ; =>This Loop Header: Depth=123; GCN-NEXT: ; Child Loop BB0_4 Depth 224; GCN-NEXT: buffer_load_dword v1, off, s[8:11], 025; GCN-NEXT: s_waitcnt vmcnt(0)26; GCN-NEXT: v_cmp_ne_u32_e64 s[2:3], 0, v127; GCN-NEXT: v_cmp_eq_u32_e32 vcc, 0, v128; GCN-NEXT: v_cndmask_b32_e64 v1, 0, 1, vcc29; GCN-NEXT: v_cmp_ne_u32_e64 s[0:1], 1, v130; GCN-NEXT: s_mov_b32 s12, s631; GCN-NEXT: s_branch .LBB0_432; GCN-NEXT: .LBB0_3: ; %Flow133; GCN-NEXT: ; in Loop: Header=BB0_4 Depth=234; GCN-NEXT: s_andn2_b64 vcc, exec, s[16:17]35; GCN-NEXT: s_cbranch_vccz .LBB0_136; GCN-NEXT: .LBB0_4: ; %bb237; GCN-NEXT: ; Parent Loop BB0_2 Depth=138; GCN-NEXT: ; => This Inner Loop Header: Depth=239; GCN-NEXT: s_and_b64 vcc, exec, s[0:1]40; GCN-NEXT: s_lshl_b32 s12, s12, 541; GCN-NEXT: s_cbranch_vccz .LBB0_642; GCN-NEXT: ; %bb.5: ; in Loop: Header=BB0_4 Depth=243; GCN-NEXT: s_mov_b64 s[16:17], s[2:3]44; GCN-NEXT: s_branch .LBB0_745; GCN-NEXT: .LBB0_6: ; %bb346; GCN-NEXT: ; in Loop: Header=BB0_4 Depth=247; GCN-NEXT: s_add_i32 s12, s12, 148; GCN-NEXT: s_mov_b64 s[16:17], -149; GCN-NEXT: .LBB0_7: ; %Flow50; GCN-NEXT: ; in Loop: Header=BB0_4 Depth=251; GCN-NEXT: s_mov_b64 s[14:15], -152; GCN-NEXT: s_andn2_b64 vcc, exec, s[16:17]53; GCN-NEXT: s_mov_b64 s[16:17], -154; GCN-NEXT: s_cbranch_vccnz .LBB0_355; GCN-NEXT: ; %bb.8: ; %bb456; GCN-NEXT: ; in Loop: Header=BB0_4 Depth=257; GCN-NEXT: s_ashr_i32 s13, s12, 3158; GCN-NEXT: s_lshl_b64 s[16:17], s[12:13], 259; GCN-NEXT: s_mov_b64 s[14:15], 060; GCN-NEXT: v_mov_b32_e32 v1, s1661; GCN-NEXT: v_mov_b32_e32 v2, s1762; GCN-NEXT: buffer_store_dword v0, v[1:2], s[4:7], 0 addr6463; GCN-NEXT: s_cmp_eq_u32 s12, 3264; GCN-NEXT: s_cselect_b64 s[16:17], -1, 065; GCN-NEXT: s_branch .LBB0_366; GCN-NEXT: .LBB0_9: ; %DummyReturnBlock67; GCN-NEXT: s_endpgm68bb:69 br label %bb170 71bb1:72 %tmp1 = load i32, ptr addrspace(1) %arg173 %tmp2 = icmp eq i32 %tmp1, 074 br label %bb275 76bb2:77 %tmp3 = phi i32 [ 0, %bb1 ], [ %tmp6, %bb4 ]78 %tmp4 = shl i32 %tmp3, 579 br i1 %tmp2, label %bb3, label %bb480 81bb3:82 %tmp5 = add i32 %tmp4, 183 br label %bb484 85bb4:86 %tmp6 = phi i32 [ %tmp5, %bb3 ], [ %tmp4, %bb2 ]87 %gep = getelementptr inbounds i32, ptr addrspace(1) %arg1, i32 %tmp688 store i32 0, ptr addrspace(1) %gep89 %tmp7 = icmp eq i32 %tmp6, 3290 br i1 %tmp7, label %bb1, label %bb291}92 93define amdgpu_kernel void @negated_cond_dominated_blocks(ptr addrspace(1) %arg1) {94; GCN-LABEL: negated_cond_dominated_blocks:95; GCN: ; %bb.0: ; %bb96; GCN-NEXT: s_load_dwordx2 s[4:5], s[4:5], 0x997; GCN-NEXT: s_waitcnt lgkmcnt(0)98; GCN-NEXT: s_load_dword s0, s[4:5], 0x099; GCN-NEXT: s_mov_b32 s6, 0100; GCN-NEXT: s_mov_b32 s7, 0xf000101; GCN-NEXT: s_waitcnt lgkmcnt(0)102; GCN-NEXT: s_cmp_lg_u32 s0, 0103; GCN-NEXT: s_cselect_b64 s[0:1], -1, 0104; GCN-NEXT: s_and_b64 s[0:1], exec, s[0:1]105; GCN-NEXT: v_mov_b32_e32 v0, 0106; GCN-NEXT: s_mov_b32 s3, s6107; GCN-NEXT: s_branch .LBB1_2108; GCN-NEXT: .LBB1_1: ; %bb7109; GCN-NEXT: ; in Loop: Header=BB1_2 Depth=1110; GCN-NEXT: s_ashr_i32 s3, s2, 31111; GCN-NEXT: s_lshl_b64 s[8:9], s[2:3], 2112; GCN-NEXT: v_mov_b32_e32 v1, s8113; GCN-NEXT: v_mov_b32_e32 v2, s9114; GCN-NEXT: s_cmp_eq_u32 s2, 32115; GCN-NEXT: buffer_store_dword v0, v[1:2], s[4:7], 0 addr64116; GCN-NEXT: s_mov_b32 s3, s2117; GCN-NEXT: s_cbranch_scc1 .LBB1_6118; GCN-NEXT: .LBB1_2: ; %bb4119; GCN-NEXT: ; =>This Inner Loop Header: Depth=1120; GCN-NEXT: s_mov_b64 vcc, s[0:1]121; GCN-NEXT: s_cbranch_vccz .LBB1_4122; GCN-NEXT: ; %bb.3: ; %bb6123; GCN-NEXT: ; in Loop: Header=BB1_2 Depth=1124; GCN-NEXT: s_add_i32 s2, s3, 1125; GCN-NEXT: s_mov_b64 vcc, exec126; GCN-NEXT: s_cbranch_execnz .LBB1_1127; GCN-NEXT: s_branch .LBB1_5128; GCN-NEXT: .LBB1_4: ; in Loop: Header=BB1_2 Depth=1129; GCN-NEXT: ; implicit-def: $sgpr2130; GCN-NEXT: s_mov_b64 vcc, 0131; GCN-NEXT: .LBB1_5: ; %bb5132; GCN-NEXT: ; in Loop: Header=BB1_2 Depth=1133; GCN-NEXT: s_lshl_b32 s2, s3, 5134; GCN-NEXT: s_or_b32 s2, s2, 1135; GCN-NEXT: s_branch .LBB1_1136; GCN-NEXT: .LBB1_6: ; %bb3137; GCN-NEXT: s_endpgm138bb:139 br label %bb2140 141bb2:142 %tmp1 = load i32, ptr addrspace(1) %arg1143 %tmp2 = icmp eq i32 %tmp1, 0144 br label %bb4145 146bb3:147 ret void148 149bb4:150 %tmp3 = phi i32 [ 0, %bb2 ], [ %tmp7, %bb7 ]151 %tmp4 = shl i32 %tmp3, 5152 br i1 %tmp2, label %bb5, label %bb6153 154bb5:155 %tmp5 = add i32 %tmp4, 1156 br label %bb7157 158bb6:159 %tmp6 = add i32 %tmp3, 1160 br label %bb7161 162bb7:163 %tmp7 = phi i32 [ %tmp5, %bb5 ], [ %tmp6, %bb6 ]164 %gep = getelementptr inbounds i32, ptr addrspace(1) %arg1, i32 %tmp7165 store i32 0, ptr addrspace(1) %gep166 %tmp8 = icmp eq i32 %tmp7, 32167 br i1 %tmp8, label %bb3, label %bb4168}169