251 lines · plain
1; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py UTC_ARGS: --version 52; RUN: llc -mtriple=amdgcn -mcpu=gfx900 < %s | FileCheck -check-prefixes=GFX9 %s3; RUN: llc -mtriple=amdgcn -mcpu=gfx1010 < %s | FileCheck -check-prefixes=GFX10 %s4; RUN: llc -mtriple=amdgcn -mcpu=gfx1100 < %s | FileCheck -check-prefixes=GFX11 %s5 6define amdgpu_kernel void @add_var_var_i1(ptr addrspace(1) %out, ptr addrspace(1) %in0, ptr addrspace(1) %in1) {7; GFX9-LABEL: add_var_var_i1:8; GFX9: ; %bb.0:9; GFX9-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x2410; GFX9-NEXT: s_load_dwordx2 s[6:7], s[4:5], 0x3411; GFX9-NEXT: v_mov_b32_e32 v0, 012; GFX9-NEXT: s_waitcnt lgkmcnt(0)13; GFX9-NEXT: global_load_ubyte v1, v0, s[2:3] glc14; GFX9-NEXT: s_waitcnt vmcnt(0)15; GFX9-NEXT: global_load_ubyte v2, v0, s[6:7] glc16; GFX9-NEXT: s_waitcnt vmcnt(0)17; GFX9-NEXT: v_xor_b32_e32 v1, v1, v218; GFX9-NEXT: v_and_b32_e32 v1, 1, v119; GFX9-NEXT: global_store_byte v0, v1, s[0:1]20; GFX9-NEXT: s_endpgm21;22; GFX10-LABEL: add_var_var_i1:23; GFX10: ; %bb.0:24; GFX10-NEXT: s_clause 0x125; GFX10-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x2426; GFX10-NEXT: s_load_dwordx2 s[6:7], s[4:5], 0x3427; GFX10-NEXT: v_mov_b32_e32 v0, 028; GFX10-NEXT: s_waitcnt lgkmcnt(0)29; GFX10-NEXT: global_load_ubyte v1, v0, s[2:3] glc dlc30; GFX10-NEXT: s_waitcnt vmcnt(0)31; GFX10-NEXT: global_load_ubyte v2, v0, s[6:7] glc dlc32; GFX10-NEXT: s_waitcnt vmcnt(0)33; GFX10-NEXT: v_xor_b32_e32 v1, v1, v234; GFX10-NEXT: v_and_b32_e32 v1, 1, v135; GFX10-NEXT: global_store_byte v0, v1, s[0:1]36; GFX10-NEXT: s_endpgm37;38; GFX11-LABEL: add_var_var_i1:39; GFX11: ; %bb.0:40; GFX11-NEXT: s_clause 0x141; GFX11-NEXT: s_load_b128 s[0:3], s[4:5], 0x2442; GFX11-NEXT: s_load_b64 s[4:5], s[4:5], 0x3443; GFX11-NEXT: v_mov_b32_e32 v0, 044; GFX11-NEXT: s_waitcnt lgkmcnt(0)45; GFX11-NEXT: global_load_u8 v1, v0, s[2:3] glc dlc46; GFX11-NEXT: s_waitcnt vmcnt(0)47; GFX11-NEXT: global_load_u8 v2, v0, s[4:5] glc dlc48; GFX11-NEXT: s_waitcnt vmcnt(0)49; GFX11-NEXT: v_xor_b32_e32 v1, v1, v250; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)51; GFX11-NEXT: v_and_b32_e32 v1, 1, v152; GFX11-NEXT: global_store_b8 v0, v1, s[0:1]53; GFX11-NEXT: s_endpgm54 %a = load volatile i1, ptr addrspace(1) %in055 %b = load volatile i1, ptr addrspace(1) %in156 %add = add i1 %a, %b57 store i1 %add, ptr addrspace(1) %out58 ret void59}60 61define amdgpu_kernel void @add_var_imm_i1(ptr addrspace(1) %out, ptr addrspace(1) %in) {62; GFX9-LABEL: add_var_imm_i1:63; GFX9: ; %bb.0:64; GFX9-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x2465; GFX9-NEXT: v_mov_b32_e32 v0, 066; GFX9-NEXT: s_waitcnt lgkmcnt(0)67; GFX9-NEXT: global_load_ubyte v1, v0, s[2:3] glc68; GFX9-NEXT: s_waitcnt vmcnt(0)69; GFX9-NEXT: v_and_b32_e32 v1, 1, v170; GFX9-NEXT: v_cmp_eq_u32_e32 vcc, 1, v171; GFX9-NEXT: s_xor_b64 s[2:3], vcc, -172; GFX9-NEXT: v_cndmask_b32_e64 v1, 0, 1, s[2:3]73; GFX9-NEXT: global_store_byte v0, v1, s[0:1]74; GFX9-NEXT: s_endpgm75;76; GFX10-LABEL: add_var_imm_i1:77; GFX10: ; %bb.0:78; GFX10-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x2479; GFX10-NEXT: v_mov_b32_e32 v0, 080; GFX10-NEXT: s_waitcnt lgkmcnt(0)81; GFX10-NEXT: global_load_ubyte v1, v0, s[2:3] glc dlc82; GFX10-NEXT: s_waitcnt vmcnt(0)83; GFX10-NEXT: v_and_b32_e32 v1, 1, v184; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, 1, v185; GFX10-NEXT: s_xor_b32 s2, vcc_lo, -186; GFX10-NEXT: v_cndmask_b32_e64 v1, 0, 1, s287; GFX10-NEXT: global_store_byte v0, v1, s[0:1]88; GFX10-NEXT: s_endpgm89;90; GFX11-LABEL: add_var_imm_i1:91; GFX11: ; %bb.0:92; GFX11-NEXT: s_load_b128 s[0:3], s[4:5], 0x2493; GFX11-NEXT: v_mov_b32_e32 v0, 094; GFX11-NEXT: s_waitcnt lgkmcnt(0)95; GFX11-NEXT: global_load_u8 v1, v0, s[2:3] glc dlc96; GFX11-NEXT: s_waitcnt vmcnt(0)97; GFX11-NEXT: v_and_b32_e32 v1, 1, v198; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)99; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, 1, v1100; GFX11-NEXT: s_xor_b32 s2, vcc_lo, -1101; GFX11-NEXT: v_cndmask_b32_e64 v1, 0, 1, s2102; GFX11-NEXT: global_store_b8 v0, v1, s[0:1]103; GFX11-NEXT: s_endpgm104 %a = load volatile i1, ptr addrspace(1) %in105 %add = add i1 %a, 1106 store i1 %add, ptr addrspace(1) %out107 ret void108}109 110define amdgpu_kernel void @add_i1_cf(ptr addrspace(1) %out, ptr addrspace(1) %a, ptr addrspace(1) %b) {111; GFX9-LABEL: add_i1_cf:112; GFX9: ; %bb.0: ; %entry113; GFX9-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x24114; GFX9-NEXT: s_load_dwordx2 s[8:9], s[4:5], 0x34115; GFX9-NEXT: v_cmp_lt_u32_e32 vcc, 15, v0116; GFX9-NEXT: ; implicit-def: $sgpr4_sgpr5117; GFX9-NEXT: s_and_saveexec_b64 s[6:7], vcc118; GFX9-NEXT: s_xor_b64 s[6:7], exec, s[6:7]119; GFX9-NEXT: s_cbranch_execz .LBB2_2120; GFX9-NEXT: ; %bb.1: ; %else121; GFX9-NEXT: v_mov_b32_e32 v0, 0122; GFX9-NEXT: s_waitcnt lgkmcnt(0)123; GFX9-NEXT: global_load_ubyte v0, v0, s[8:9] glc124; GFX9-NEXT: s_waitcnt vmcnt(0)125; GFX9-NEXT: v_and_b32_e32 v0, 1, v0126; GFX9-NEXT: v_cmp_eq_u32_e64 s[4:5], 1, v0127; GFX9-NEXT: .LBB2_2: ; %Flow128; GFX9-NEXT: s_andn2_saveexec_b64 s[6:7], s[6:7]129; GFX9-NEXT: s_cbranch_execz .LBB2_4130; GFX9-NEXT: ; %bb.3: ; %if131; GFX9-NEXT: v_mov_b32_e32 v0, 0132; GFX9-NEXT: s_waitcnt lgkmcnt(0)133; GFX9-NEXT: global_load_ubyte v0, v0, s[2:3] glc134; GFX9-NEXT: s_waitcnt vmcnt(0)135; GFX9-NEXT: s_andn2_b64 s[2:3], s[4:5], exec136; GFX9-NEXT: v_and_b32_e32 v0, 1, v0137; GFX9-NEXT: v_cmp_eq_u32_e32 vcc, 1, v0138; GFX9-NEXT: s_and_b64 s[4:5], vcc, exec139; GFX9-NEXT: s_or_b64 s[4:5], s[2:3], s[4:5]140; GFX9-NEXT: .LBB2_4: ; %endif141; GFX9-NEXT: s_or_b64 exec, exec, s[6:7]142; GFX9-NEXT: s_waitcnt lgkmcnt(0)143; GFX9-NEXT: s_xor_b64 s[2:3], s[4:5], -1144; GFX9-NEXT: v_mov_b32_e32 v0, 0145; GFX9-NEXT: v_cndmask_b32_e64 v1, 0, 1, s[2:3]146; GFX9-NEXT: global_store_byte v0, v1, s[0:1]147; GFX9-NEXT: s_endpgm148;149; GFX10-LABEL: add_i1_cf:150; GFX10: ; %bb.0: ; %entry151; GFX10-NEXT: s_clause 0x1152; GFX10-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x24153; GFX10-NEXT: s_load_dwordx2 s[6:7], s[4:5], 0x34154; GFX10-NEXT: v_cmp_lt_u32_e32 vcc_lo, 15, v0155; GFX10-NEXT: ; implicit-def: $sgpr4156; GFX10-NEXT: s_and_saveexec_b32 s5, vcc_lo157; GFX10-NEXT: s_xor_b32 s5, exec_lo, s5158; GFX10-NEXT: s_cbranch_execz .LBB2_2159; GFX10-NEXT: ; %bb.1: ; %else160; GFX10-NEXT: v_mov_b32_e32 v0, 0161; GFX10-NEXT: s_waitcnt lgkmcnt(0)162; GFX10-NEXT: global_load_ubyte v0, v0, s[6:7] glc dlc163; GFX10-NEXT: s_waitcnt vmcnt(0)164; GFX10-NEXT: v_and_b32_e32 v0, 1, v0165; GFX10-NEXT: v_cmp_eq_u32_e64 s4, 1, v0166; GFX10-NEXT: .LBB2_2: ; %Flow167; GFX10-NEXT: s_andn2_saveexec_b32 s5, s5168; GFX10-NEXT: s_cbranch_execz .LBB2_4169; GFX10-NEXT: ; %bb.3: ; %if170; GFX10-NEXT: v_mov_b32_e32 v0, 0171; GFX10-NEXT: s_waitcnt lgkmcnt(0)172; GFX10-NEXT: global_load_ubyte v0, v0, s[2:3] glc dlc173; GFX10-NEXT: s_waitcnt vmcnt(0)174; GFX10-NEXT: s_waitcnt_depctr depctr_vm_vsrc(0)175; GFX10-NEXT: s_andn2_b32 s2, s4, exec_lo176; GFX10-NEXT: v_and_b32_e32 v0, 1, v0177; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, 1, v0178; GFX10-NEXT: s_and_b32 s3, vcc_lo, exec_lo179; GFX10-NEXT: s_or_b32 s4, s2, s3180; GFX10-NEXT: .LBB2_4: ; %endif181; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s5182; GFX10-NEXT: s_waitcnt lgkmcnt(0)183; GFX10-NEXT: s_xor_b32 s2, s4, -1184; GFX10-NEXT: v_mov_b32_e32 v0, 0185; GFX10-NEXT: v_cndmask_b32_e64 v1, 0, 1, s2186; GFX10-NEXT: global_store_byte v0, v1, s[0:1]187; GFX10-NEXT: s_endpgm188;189; GFX11-LABEL: add_i1_cf:190; GFX11: ; %bb.0: ; %entry191; GFX11-NEXT: s_clause 0x1192; GFX11-NEXT: s_load_b128 s[0:3], s[4:5], 0x24193; GFX11-NEXT: s_load_b64 s[4:5], s[4:5], 0x34194; GFX11-NEXT: v_and_b32_e32 v0, 0x3ff, v0195; GFX11-NEXT: s_mov_b32 s7, exec_lo196; GFX11-NEXT: ; implicit-def: $sgpr6197; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)198; GFX11-NEXT: v_cmpx_lt_u32_e32 15, v0199; GFX11-NEXT: s_xor_b32 s7, exec_lo, s7200; GFX11-NEXT: s_cbranch_execz .LBB2_2201; GFX11-NEXT: ; %bb.1: ; %else202; GFX11-NEXT: v_mov_b32_e32 v0, 0203; GFX11-NEXT: s_waitcnt lgkmcnt(0)204; GFX11-NEXT: global_load_u8 v0, v0, s[4:5] glc dlc205; GFX11-NEXT: s_waitcnt vmcnt(0)206; GFX11-NEXT: v_and_b32_e32 v0, 1, v0207; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)208; GFX11-NEXT: v_cmp_eq_u32_e64 s6, 1, v0209; GFX11-NEXT: .LBB2_2: ; %Flow210; GFX11-NEXT: s_waitcnt lgkmcnt(0)211; GFX11-NEXT: s_and_not1_saveexec_b32 s4, s7212; GFX11-NEXT: s_cbranch_execz .LBB2_4213; GFX11-NEXT: ; %bb.3: ; %if214; GFX11-NEXT: v_mov_b32_e32 v0, 0215; GFX11-NEXT: global_load_u8 v0, v0, s[2:3] glc dlc216; GFX11-NEXT: s_waitcnt vmcnt(0)217; GFX11-NEXT: s_and_not1_b32 s2, s6, exec_lo218; GFX11-NEXT: v_and_b32_e32 v0, 1, v0219; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)220; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, 1, v0221; GFX11-NEXT: s_and_b32 s3, vcc_lo, exec_lo222; GFX11-NEXT: s_or_b32 s6, s2, s3223; GFX11-NEXT: .LBB2_4: ; %endif224; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s4225; GFX11-NEXT: s_xor_b32 s2, s6, -1226; GFX11-NEXT: v_mov_b32_e32 v0, 0227; GFX11-NEXT: v_cndmask_b32_e64 v1, 0, 1, s2228; GFX11-NEXT: global_store_b8 v0, v1, s[0:1]229; GFX11-NEXT: s_endpgm230entry:231 %tid = call i32 @llvm.amdgcn.workitem.id.x()232 %d_cmp = icmp ult i32 %tid, 16233 br i1 %d_cmp, label %if, label %else234 235if:236 %0 = load volatile i1, ptr addrspace(1) %a237 br label %endif238 239else:240 %1 = load volatile i1, ptr addrspace(1) %b241 br label %endif242 243endif:244 %2 = phi i1 [%0, %if], [%1, %else]245 %3 = add i1 %2, -1246 store i1 %3, ptr addrspace(1) %out247 ret void248}249 250declare i32 @llvm.amdgcn.workitem.id.x()251