670 lines · plain
1; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py2; RUN: llc -mtriple=amdgcn -mcpu=tahiti < %s | FileCheck -check-prefix=SI %s3; RUN: llc -mtriple=amdgcn -mcpu=hawaii < %s | FileCheck -check-prefix=CI %s4 5define amdgpu_kernel void @round_f64(ptr addrspace(1) %out, double %x) #0 {6; SI-LABEL: round_f64:7; SI: ; %bb.0:8; SI-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x99; SI-NEXT: s_mov_b32 s6, -110; SI-NEXT: s_mov_b32 s5, 0xfffff11; SI-NEXT: s_mov_b32 s4, s612; SI-NEXT: s_waitcnt lgkmcnt(0)13; SI-NEXT: s_bfe_u32 s7, s3, 0xb001414; SI-NEXT: s_addk_i32 s7, 0xfc0115; SI-NEXT: s_lshr_b64 s[4:5], s[4:5], s716; SI-NEXT: s_and_b32 s8, s3, 0x8000000017; SI-NEXT: s_andn2_b64 s[4:5], s[2:3], s[4:5]18; SI-NEXT: s_cmp_lt_i32 s7, 019; SI-NEXT: s_cselect_b32 s4, 0, s420; SI-NEXT: s_cselect_b32 s5, s8, s521; SI-NEXT: s_cmp_gt_i32 s7, 5122; SI-NEXT: s_cselect_b32 s8, s2, s423; SI-NEXT: s_cselect_b32 s9, s3, s524; SI-NEXT: v_mov_b32_e32 v0, s825; SI-NEXT: v_mov_b32_e32 v1, s926; SI-NEXT: v_add_f64 v[0:1], s[2:3], -v[0:1]27; SI-NEXT: s_mov_b32 s4, s028; SI-NEXT: v_cmp_ge_f64_e64 s[10:11], |v[0:1]|, 0.529; SI-NEXT: s_brev_b32 s2, -230; SI-NEXT: s_and_b64 s[10:11], s[10:11], exec31; SI-NEXT: s_cselect_b32 s0, 0x3ff00000, 032; SI-NEXT: v_mov_b32_e32 v0, s033; SI-NEXT: v_mov_b32_e32 v1, s334; SI-NEXT: v_bfi_b32 v1, s2, v0, v135; SI-NEXT: v_mov_b32_e32 v0, 036; SI-NEXT: v_add_f64 v[0:1], s[8:9], v[0:1]37; SI-NEXT: s_mov_b32 s7, 0xf00038; SI-NEXT: s_mov_b32 s5, s139; SI-NEXT: buffer_store_dwordx2 v[0:1], off, s[4:7], 040; SI-NEXT: s_endpgm41;42; CI-LABEL: round_f64:43; CI: ; %bb.0:44; CI-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x945; CI-NEXT: s_brev_b32 s5, -246; CI-NEXT: s_mov_b32 s7, 0xf00047; CI-NEXT: s_mov_b32 s6, -148; CI-NEXT: s_waitcnt lgkmcnt(0)49; CI-NEXT: v_trunc_f64_e32 v[0:1], s[2:3]50; CI-NEXT: s_mov_b32 s4, s051; CI-NEXT: v_add_f64 v[2:3], s[2:3], -v[0:1]52; CI-NEXT: v_cmp_ge_f64_e64 s[8:9], |v[2:3]|, 0.553; CI-NEXT: v_mov_b32_e32 v2, s354; CI-NEXT: s_and_b64 s[2:3], s[8:9], exec55; CI-NEXT: s_cselect_b32 s0, 0x3ff00000, 056; CI-NEXT: v_mov_b32_e32 v3, s057; CI-NEXT: v_bfi_b32 v3, s5, v3, v258; CI-NEXT: v_mov_b32_e32 v2, 059; CI-NEXT: v_add_f64 v[0:1], v[0:1], v[2:3]60; CI-NEXT: s_mov_b32 s5, s161; CI-NEXT: buffer_store_dwordx2 v[0:1], off, s[4:7], 062; CI-NEXT: s_endpgm63 %result = call double @llvm.round.f64(double %x) #164 store double %result, ptr addrspace(1) %out65 ret void66}67 68define amdgpu_kernel void @v_round_f64(ptr addrspace(1) %out, ptr addrspace(1) %in) #0 {69; SI-LABEL: v_round_f64:70; SI: ; %bb.0:71; SI-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x972; SI-NEXT: s_mov_b32 s7, 0xf00073; SI-NEXT: s_mov_b32 s6, 074; SI-NEXT: v_lshlrev_b32_e32 v0, 3, v075; SI-NEXT: v_mov_b32_e32 v1, 076; SI-NEXT: s_waitcnt lgkmcnt(0)77; SI-NEXT: s_mov_b64 s[4:5], s[2:3]78; SI-NEXT: buffer_load_dwordx2 v[2:3], v[0:1], s[4:7], 0 addr6479; SI-NEXT: s_mov_b32 s2, -180; SI-NEXT: s_mov_b32 s3, 0xfffff81; SI-NEXT: s_waitcnt vmcnt(0)82; SI-NEXT: v_bfe_u32 v4, v3, 20, 1183; SI-NEXT: v_add_i32_e32 v6, vcc, 0xfffffc01, v484; SI-NEXT: v_lshr_b64 v[4:5], s[2:3], v685; SI-NEXT: v_and_b32_e32 v7, 0x80000000, v386; SI-NEXT: v_bfi_b32 v5, v5, 0, v387; SI-NEXT: v_bfi_b32 v4, v4, 0, v288; SI-NEXT: v_cmp_gt_i32_e32 vcc, 0, v689; SI-NEXT: v_cndmask_b32_e32 v5, v5, v7, vcc90; SI-NEXT: v_cndmask_b32_e64 v4, v4, 0, vcc91; SI-NEXT: v_cmp_lt_i32_e32 vcc, 51, v692; SI-NEXT: v_cndmask_b32_e32 v5, v5, v3, vcc93; SI-NEXT: v_cndmask_b32_e32 v4, v4, v2, vcc94; SI-NEXT: v_add_f64 v[6:7], v[2:3], -v[4:5]95; SI-NEXT: v_mov_b32_e32 v2, 0x3ff0000096; SI-NEXT: v_cmp_ge_f64_e64 vcc, |v[6:7]|, 0.597; SI-NEXT: s_brev_b32 s2, -298; SI-NEXT: v_cndmask_b32_e32 v2, 0, v2, vcc99; SI-NEXT: v_bfi_b32 v2, s2, v2, v3100; SI-NEXT: v_add_f64 v[2:3], v[4:5], v[1:2]101; SI-NEXT: s_mov_b64 s[2:3], s[6:7]102; SI-NEXT: buffer_store_dwordx2 v[2:3], v[0:1], s[0:3], 0 addr64103; SI-NEXT: s_endpgm104;105; CI-LABEL: v_round_f64:106; CI: ; %bb.0:107; CI-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x9108; CI-NEXT: s_mov_b32 s7, 0xf000109; CI-NEXT: s_mov_b32 s6, 0110; CI-NEXT: v_lshlrev_b32_e32 v0, 3, v0111; CI-NEXT: v_mov_b32_e32 v1, 0112; CI-NEXT: s_waitcnt lgkmcnt(0)113; CI-NEXT: s_mov_b64 s[4:5], s[2:3]114; CI-NEXT: buffer_load_dwordx2 v[2:3], v[0:1], s[4:7], 0 addr64115; CI-NEXT: s_brev_b32 s2, -2116; CI-NEXT: s_waitcnt vmcnt(0)117; CI-NEXT: v_trunc_f64_e32 v[4:5], v[2:3]118; CI-NEXT: v_add_f64 v[6:7], v[2:3], -v[4:5]119; CI-NEXT: v_mov_b32_e32 v2, 0x3ff00000120; CI-NEXT: v_cmp_ge_f64_e64 vcc, |v[6:7]|, 0.5121; CI-NEXT: v_cndmask_b32_e32 v2, 0, v2, vcc122; CI-NEXT: v_bfi_b32 v2, s2, v2, v3123; CI-NEXT: v_add_f64 v[2:3], v[4:5], v[1:2]124; CI-NEXT: s_mov_b64 s[2:3], s[6:7]125; CI-NEXT: buffer_store_dwordx2 v[2:3], v[0:1], s[0:3], 0 addr64126; CI-NEXT: s_endpgm127 %tid = call i32 @llvm.amdgcn.workitem.id.x() #1128 %gep = getelementptr double, ptr addrspace(1) %in, i32 %tid129 %out.gep = getelementptr double, ptr addrspace(1) %out, i32 %tid130 %x = load double, ptr addrspace(1) %gep131 %result = call double @llvm.round.f64(double %x) #1132 store double %result, ptr addrspace(1) %out.gep133 ret void134}135 136define amdgpu_kernel void @round_v2f64(ptr addrspace(1) %out, <2 x double> %in) #0 {137; SI-LABEL: round_v2f64:138; SI: ; %bb.0:139; SI-NEXT: s_load_dwordx4 s[8:11], s[4:5], 0xd140; SI-NEXT: s_mov_b32 s2, -1141; SI-NEXT: s_mov_b32 s7, 0xfffff142; SI-NEXT: s_mov_b32 s6, s2143; SI-NEXT: s_waitcnt lgkmcnt(0)144; SI-NEXT: s_bfe_u32 s0, s11, 0xb0014145; SI-NEXT: s_add_i32 s12, s0, 0xfffffc01146; SI-NEXT: s_lshr_b64 s[0:1], s[6:7], s12147; SI-NEXT: s_and_b32 s3, s11, 0x80000000148; SI-NEXT: s_andn2_b64 s[0:1], s[10:11], s[0:1]149; SI-NEXT: s_cmp_lt_i32 s12, 0150; SI-NEXT: s_cselect_b32 s0, 0, s0151; SI-NEXT: s_cselect_b32 s1, s3, s1152; SI-NEXT: s_cmp_gt_i32 s12, 51153; SI-NEXT: s_cselect_b32 s12, s10, s0154; SI-NEXT: s_cselect_b32 s13, s11, s1155; SI-NEXT: v_mov_b32_e32 v0, s12156; SI-NEXT: v_mov_b32_e32 v1, s13157; SI-NEXT: v_add_f64 v[0:1], s[10:11], -v[0:1]158; SI-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x9159; SI-NEXT: v_cmp_ge_f64_e64 s[14:15], |v[0:1]|, 0.5160; SI-NEXT: s_brev_b32 s10, -2161; SI-NEXT: s_and_b64 s[4:5], s[14:15], exec162; SI-NEXT: s_cselect_b32 s3, 0x3ff00000, 0163; SI-NEXT: v_mov_b32_e32 v0, s3164; SI-NEXT: s_bfe_u32 s3, s9, 0xb0014165; SI-NEXT: s_addk_i32 s3, 0xfc01166; SI-NEXT: s_lshr_b64 s[4:5], s[6:7], s3167; SI-NEXT: s_andn2_b64 s[4:5], s[8:9], s[4:5]168; SI-NEXT: s_and_b32 s6, s9, 0x80000000169; SI-NEXT: s_cmp_lt_i32 s3, 0170; SI-NEXT: s_cselect_b32 s4, 0, s4171; SI-NEXT: s_cselect_b32 s5, s6, s5172; SI-NEXT: s_cmp_gt_i32 s3, 51173; SI-NEXT: s_cselect_b32 s4, s8, s4174; SI-NEXT: s_cselect_b32 s5, s9, s5175; SI-NEXT: v_mov_b32_e32 v2, s4176; SI-NEXT: v_mov_b32_e32 v3, s5177; SI-NEXT: v_add_f64 v[2:3], s[8:9], -v[2:3]178; SI-NEXT: v_mov_b32_e32 v1, s11179; SI-NEXT: v_cmp_ge_f64_e64 s[6:7], |v[2:3]|, 0.5180; SI-NEXT: v_bfi_b32 v1, s10, v0, v1181; SI-NEXT: s_and_b64 s[6:7], s[6:7], exec182; SI-NEXT: v_mov_b32_e32 v0, 0183; SI-NEXT: s_cselect_b32 s3, 0x3ff00000, 0184; SI-NEXT: v_add_f64 v[2:3], s[12:13], v[0:1]185; SI-NEXT: v_mov_b32_e32 v1, s3186; SI-NEXT: v_mov_b32_e32 v4, s9187; SI-NEXT: v_bfi_b32 v1, s10, v1, v4188; SI-NEXT: v_add_f64 v[0:1], s[4:5], v[0:1]189; SI-NEXT: s_mov_b32 s3, 0xf000190; SI-NEXT: s_waitcnt lgkmcnt(0)191; SI-NEXT: buffer_store_dwordx4 v[0:3], off, s[0:3], 0192; SI-NEXT: s_endpgm193;194; CI-LABEL: round_v2f64:195; CI: ; %bb.0:196; CI-NEXT: s_load_dwordx4 s[8:11], s[4:5], 0xd197; CI-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x9198; CI-NEXT: s_brev_b32 s2, -2199; CI-NEXT: v_mov_b32_e32 v0, 0200; CI-NEXT: s_mov_b32 s3, 0xf000201; CI-NEXT: s_waitcnt lgkmcnt(0)202; CI-NEXT: v_trunc_f64_e32 v[2:3], s[10:11]203; CI-NEXT: v_trunc_f64_e32 v[6:7], s[8:9]204; CI-NEXT: v_add_f64 v[4:5], s[10:11], -v[2:3]205; CI-NEXT: v_mov_b32_e32 v1, s11206; CI-NEXT: v_cmp_ge_f64_e64 s[4:5], |v[4:5]|, 0.5207; CI-NEXT: v_add_f64 v[4:5], s[8:9], -v[6:7]208; CI-NEXT: s_and_b64 s[4:5], s[4:5], exec209; CI-NEXT: s_cselect_b32 s4, 0x3ff00000, 0210; CI-NEXT: v_mov_b32_e32 v8, s4211; CI-NEXT: v_cmp_ge_f64_e64 s[4:5], |v[4:5]|, 0.5212; CI-NEXT: v_bfi_b32 v1, s2, v8, v1213; CI-NEXT: s_and_b64 s[4:5], s[4:5], exec214; CI-NEXT: s_cselect_b32 s4, 0x3ff00000, 0215; CI-NEXT: v_add_f64 v[2:3], v[2:3], v[0:1]216; CI-NEXT: v_mov_b32_e32 v1, s4217; CI-NEXT: v_mov_b32_e32 v4, s9218; CI-NEXT: v_bfi_b32 v1, s2, v1, v4219; CI-NEXT: v_add_f64 v[0:1], v[6:7], v[0:1]220; CI-NEXT: s_mov_b32 s2, -1221; CI-NEXT: buffer_store_dwordx4 v[0:3], off, s[0:3], 0222; CI-NEXT: s_endpgm223 %result = call <2 x double> @llvm.round.v2f64(<2 x double> %in) #1224 store <2 x double> %result, ptr addrspace(1) %out225 ret void226}227 228define amdgpu_kernel void @round_v4f64(ptr addrspace(1) %out, <4 x double> %in) #0 {229; SI-LABEL: round_v4f64:230; SI: ; %bb.0:231; SI-NEXT: s_load_dwordx8 s[8:15], s[4:5], 0x11232; SI-NEXT: s_mov_b32 s2, -1233; SI-NEXT: s_mov_b32 s7, 0xfffff234; SI-NEXT: s_mov_b32 s6, s2235; SI-NEXT: v_mov_b32_e32 v4, 0236; SI-NEXT: s_waitcnt lgkmcnt(0)237; SI-NEXT: s_bfe_u32 s0, s11, 0xb0014238; SI-NEXT: s_add_i32 s16, s0, 0xfffffc01239; SI-NEXT: s_lshr_b64 s[0:1], s[6:7], s16240; SI-NEXT: s_and_b32 s3, s11, 0x80000000241; SI-NEXT: s_andn2_b64 s[0:1], s[10:11], s[0:1]242; SI-NEXT: s_cmp_lt_i32 s16, 0243; SI-NEXT: s_cselect_b32 s0, 0, s0244; SI-NEXT: s_cselect_b32 s1, s3, s1245; SI-NEXT: s_cmp_gt_i32 s16, 51246; SI-NEXT: s_cselect_b32 s16, s10, s0247; SI-NEXT: s_cselect_b32 s17, s11, s1248; SI-NEXT: v_mov_b32_e32 v0, s16249; SI-NEXT: v_mov_b32_e32 v1, s17250; SI-NEXT: v_add_f64 v[0:1], s[10:11], -v[0:1]251; SI-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x9252; SI-NEXT: v_cmp_ge_f64_e64 s[18:19], |v[0:1]|, 0.5253; SI-NEXT: v_mov_b32_e32 v1, s11254; SI-NEXT: s_and_b64 s[4:5], s[18:19], exec255; SI-NEXT: s_cselect_b32 s3, 0x3ff00000, 0256; SI-NEXT: v_mov_b32_e32 v0, s3257; SI-NEXT: s_bfe_u32 s3, s9, 0xb0014258; SI-NEXT: s_addk_i32 s3, 0xfc01259; SI-NEXT: s_lshr_b64 s[4:5], s[6:7], s3260; SI-NEXT: s_andn2_b64 s[4:5], s[8:9], s[4:5]261; SI-NEXT: s_and_b32 s10, s9, 0x80000000262; SI-NEXT: s_cmp_lt_i32 s3, 0263; SI-NEXT: s_cselect_b32 s4, 0, s4264; SI-NEXT: s_cselect_b32 s5, s10, s5265; SI-NEXT: s_cmp_gt_i32 s3, 51266; SI-NEXT: s_brev_b32 s18, -2267; SI-NEXT: s_cselect_b32 s4, s8, s4268; SI-NEXT: v_bfi_b32 v5, s18, v0, v1269; SI-NEXT: s_cselect_b32 s5, s9, s5270; SI-NEXT: v_mov_b32_e32 v0, s4271; SI-NEXT: v_mov_b32_e32 v1, s5272; SI-NEXT: v_add_f64 v[0:1], s[8:9], -v[0:1]273; SI-NEXT: v_add_f64 v[2:3], s[16:17], v[4:5]274; SI-NEXT: v_cmp_ge_f64_e64 s[10:11], |v[0:1]|, 0.5275; SI-NEXT: v_mov_b32_e32 v6, s9276; SI-NEXT: s_and_b64 s[10:11], s[10:11], exec277; SI-NEXT: s_cselect_b32 s3, 0x3ff00000, 0278; SI-NEXT: v_mov_b32_e32 v5, s3279; SI-NEXT: s_bfe_u32 s3, s15, 0xb0014280; SI-NEXT: s_addk_i32 s3, 0xfc01281; SI-NEXT: s_lshr_b64 s[8:9], s[6:7], s3282; SI-NEXT: s_andn2_b64 s[8:9], s[14:15], s[8:9]283; SI-NEXT: s_and_b32 s10, s15, 0x80000000284; SI-NEXT: s_cmp_lt_i32 s3, 0285; SI-NEXT: s_cselect_b32 s8, 0, s8286; SI-NEXT: s_cselect_b32 s9, s10, s9287; SI-NEXT: s_cmp_gt_i32 s3, 51288; SI-NEXT: s_cselect_b32 s8, s14, s8289; SI-NEXT: s_cselect_b32 s9, s15, s9290; SI-NEXT: v_mov_b32_e32 v0, s8291; SI-NEXT: v_mov_b32_e32 v1, s9292; SI-NEXT: v_add_f64 v[0:1], s[14:15], -v[0:1]293; SI-NEXT: v_bfi_b32 v5, s18, v5, v6294; SI-NEXT: v_cmp_ge_f64_e64 s[10:11], |v[0:1]|, 0.5295; SI-NEXT: v_add_f64 v[0:1], s[4:5], v[4:5]296; SI-NEXT: s_and_b64 s[4:5], s[10:11], exec297; SI-NEXT: s_cselect_b32 s3, 0x3ff00000, 0298; SI-NEXT: v_mov_b32_e32 v8, s3299; SI-NEXT: s_bfe_u32 s3, s13, 0xb0014300; SI-NEXT: s_addk_i32 s3, 0xfc01301; SI-NEXT: s_lshr_b64 s[4:5], s[6:7], s3302; SI-NEXT: s_andn2_b64 s[4:5], s[12:13], s[4:5]303; SI-NEXT: s_and_b32 s6, s13, 0x80000000304; SI-NEXT: s_cmp_lt_i32 s3, 0305; SI-NEXT: s_cselect_b32 s4, 0, s4306; SI-NEXT: s_cselect_b32 s5, s6, s5307; SI-NEXT: s_cmp_gt_i32 s3, 51308; SI-NEXT: s_cselect_b32 s5, s13, s5309; SI-NEXT: s_cselect_b32 s4, s12, s4310; SI-NEXT: v_mov_b32_e32 v6, s5311; SI-NEXT: v_mov_b32_e32 v5, s4312; SI-NEXT: v_add_f64 v[6:7], s[12:13], -v[5:6]313; SI-NEXT: v_mov_b32_e32 v9, s15314; SI-NEXT: v_cmp_ge_f64_e64 s[6:7], |v[6:7]|, 0.5315; SI-NEXT: v_bfi_b32 v5, s18, v8, v9316; SI-NEXT: s_and_b64 s[6:7], s[6:7], exec317; SI-NEXT: s_cselect_b32 s3, 0x3ff00000, 0318; SI-NEXT: v_add_f64 v[6:7], s[8:9], v[4:5]319; SI-NEXT: v_mov_b32_e32 v5, s3320; SI-NEXT: v_mov_b32_e32 v8, s13321; SI-NEXT: v_bfi_b32 v5, s18, v5, v8322; SI-NEXT: v_add_f64 v[4:5], s[4:5], v[4:5]323; SI-NEXT: s_mov_b32 s3, 0xf000324; SI-NEXT: s_waitcnt lgkmcnt(0)325; SI-NEXT: buffer_store_dwordx4 v[4:7], off, s[0:3], 0 offset:16326; SI-NEXT: buffer_store_dwordx4 v[0:3], off, s[0:3], 0327; SI-NEXT: s_endpgm328;329; CI-LABEL: round_v4f64:330; CI: ; %bb.0:331; CI-NEXT: s_load_dwordx8 s[8:15], s[4:5], 0x11332; CI-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x9333; CI-NEXT: s_brev_b32 s2, -2334; CI-NEXT: v_mov_b32_e32 v4, 0335; CI-NEXT: s_mov_b32 s3, 0xf000336; CI-NEXT: s_waitcnt lgkmcnt(0)337; CI-NEXT: v_trunc_f64_e32 v[0:1], s[10:11]338; CI-NEXT: v_trunc_f64_e32 v[6:7], s[8:9]339; CI-NEXT: v_add_f64 v[2:3], s[10:11], -v[0:1]340; CI-NEXT: v_mov_b32_e32 v5, s11341; CI-NEXT: v_cmp_ge_f64_e64 s[4:5], |v[2:3]|, 0.5342; CI-NEXT: v_add_f64 v[2:3], s[8:9], -v[6:7]343; CI-NEXT: s_and_b64 s[4:5], s[4:5], exec344; CI-NEXT: s_cselect_b32 s4, 0x3ff00000, 0345; CI-NEXT: v_mov_b32_e32 v8, s4346; CI-NEXT: v_cmp_ge_f64_e64 s[4:5], |v[2:3]|, 0.5347; CI-NEXT: v_bfi_b32 v5, s2, v8, v5348; CI-NEXT: v_trunc_f64_e32 v[8:9], s[14:15]349; CI-NEXT: s_and_b64 s[4:5], s[4:5], exec350; CI-NEXT: v_add_f64 v[2:3], v[0:1], v[4:5]351; CI-NEXT: s_cselect_b32 s4, 0x3ff00000, 0352; CI-NEXT: v_add_f64 v[0:1], s[14:15], -v[8:9]353; CI-NEXT: v_mov_b32_e32 v5, s4354; CI-NEXT: v_mov_b32_e32 v10, s9355; CI-NEXT: v_bfi_b32 v5, s2, v5, v10356; CI-NEXT: v_cmp_ge_f64_e64 s[4:5], |v[0:1]|, 0.5357; CI-NEXT: v_trunc_f64_e32 v[10:11], s[12:13]358; CI-NEXT: v_add_f64 v[0:1], v[6:7], v[4:5]359; CI-NEXT: s_and_b64 s[4:5], s[4:5], exec360; CI-NEXT: v_add_f64 v[6:7], s[12:13], -v[10:11]361; CI-NEXT: s_cselect_b32 s4, 0x3ff00000, 0362; CI-NEXT: v_mov_b32_e32 v5, s4363; CI-NEXT: v_cmp_ge_f64_e64 s[4:5], |v[6:7]|, 0.5364; CI-NEXT: v_mov_b32_e32 v12, s15365; CI-NEXT: s_and_b64 s[4:5], s[4:5], exec366; CI-NEXT: v_bfi_b32 v5, s2, v5, v12367; CI-NEXT: s_cselect_b32 s4, 0x3ff00000, 0368; CI-NEXT: v_add_f64 v[6:7], v[8:9], v[4:5]369; CI-NEXT: v_mov_b32_e32 v5, s4370; CI-NEXT: v_mov_b32_e32 v8, s13371; CI-NEXT: v_bfi_b32 v5, s2, v5, v8372; CI-NEXT: v_add_f64 v[4:5], v[10:11], v[4:5]373; CI-NEXT: s_mov_b32 s2, -1374; CI-NEXT: buffer_store_dwordx4 v[4:7], off, s[0:3], 0 offset:16375; CI-NEXT: buffer_store_dwordx4 v[0:3], off, s[0:3], 0376; CI-NEXT: s_endpgm377 %result = call <4 x double> @llvm.round.v4f64(<4 x double> %in) #1378 store <4 x double> %result, ptr addrspace(1) %out379 ret void380}381 382define amdgpu_kernel void @round_v8f64(ptr addrspace(1) %out, <8 x double> %in) #0 {383; SI-LABEL: round_v8f64:384; SI: ; %bb.0:385; SI-NEXT: s_load_dwordx16 s[8:23], s[4:5], 0x19386; SI-NEXT: s_mov_b32 s2, -1387; SI-NEXT: s_mov_b32 s7, 0xfffff388; SI-NEXT: s_mov_b32 s6, s2389; SI-NEXT: v_mov_b32_e32 v8, 0390; SI-NEXT: s_waitcnt lgkmcnt(0)391; SI-NEXT: s_bfe_u32 s0, s11, 0xb0014392; SI-NEXT: s_add_i32 s24, s0, 0xfffffc01393; SI-NEXT: s_lshr_b64 s[0:1], s[6:7], s24394; SI-NEXT: s_and_b32 s3, s11, 0x80000000395; SI-NEXT: s_andn2_b64 s[0:1], s[10:11], s[0:1]396; SI-NEXT: s_cmp_lt_i32 s24, 0397; SI-NEXT: s_cselect_b32 s0, 0, s0398; SI-NEXT: s_cselect_b32 s1, s3, s1399; SI-NEXT: s_cmp_gt_i32 s24, 51400; SI-NEXT: s_cselect_b32 s24, s10, s0401; SI-NEXT: s_cselect_b32 s25, s11, s1402; SI-NEXT: v_mov_b32_e32 v0, s24403; SI-NEXT: v_mov_b32_e32 v1, s25404; SI-NEXT: v_add_f64 v[0:1], s[10:11], -v[0:1]405; SI-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x9406; SI-NEXT: v_cmp_ge_f64_e64 s[26:27], |v[0:1]|, 0.5407; SI-NEXT: v_mov_b32_e32 v1, s11408; SI-NEXT: s_and_b64 s[4:5], s[26:27], exec409; SI-NEXT: s_cselect_b32 s4, 0x3ff00000, 0410; SI-NEXT: v_mov_b32_e32 v0, s4411; SI-NEXT: s_bfe_u32 s4, s9, 0xb0014412; SI-NEXT: s_add_i32 s10, s4, 0xfffffc01413; SI-NEXT: s_lshr_b64 s[4:5], s[6:7], s10414; SI-NEXT: s_andn2_b64 s[4:5], s[8:9], s[4:5]415; SI-NEXT: s_and_b32 s11, s9, 0x80000000416; SI-NEXT: s_cmp_lt_i32 s10, 0417; SI-NEXT: s_cselect_b32 s4, 0, s4418; SI-NEXT: s_cselect_b32 s5, s11, s5419; SI-NEXT: s_cmp_gt_i32 s10, 51420; SI-NEXT: s_brev_b32 s3, -2421; SI-NEXT: s_cselect_b32 s4, s8, s4422; SI-NEXT: v_bfi_b32 v9, s3, v0, v1423; SI-NEXT: s_cselect_b32 s5, s9, s5424; SI-NEXT: v_mov_b32_e32 v0, s4425; SI-NEXT: v_mov_b32_e32 v1, s5426; SI-NEXT: v_add_f64 v[0:1], s[8:9], -v[0:1]427; SI-NEXT: v_mov_b32_e32 v5, s9428; SI-NEXT: v_cmp_ge_f64_e64 s[10:11], |v[0:1]|, 0.5429; SI-NEXT: v_add_f64 v[2:3], s[24:25], v[8:9]430; SI-NEXT: s_and_b64 s[10:11], s[10:11], exec431; SI-NEXT: s_cselect_b32 s8, 0x3ff00000, 0432; SI-NEXT: v_mov_b32_e32 v4, s8433; SI-NEXT: s_bfe_u32 s8, s15, 0xb0014434; SI-NEXT: s_add_i32 s10, s8, 0xfffffc01435; SI-NEXT: s_lshr_b64 s[8:9], s[6:7], s10436; SI-NEXT: s_andn2_b64 s[8:9], s[14:15], s[8:9]437; SI-NEXT: s_and_b32 s11, s15, 0x80000000438; SI-NEXT: s_cmp_lt_i32 s10, 0439; SI-NEXT: s_cselect_b32 s8, 0, s8440; SI-NEXT: s_cselect_b32 s9, s11, s9441; SI-NEXT: s_cmp_gt_i32 s10, 51442; SI-NEXT: s_cselect_b32 s8, s14, s8443; SI-NEXT: s_cselect_b32 s9, s15, s9444; SI-NEXT: v_mov_b32_e32 v0, s8445; SI-NEXT: v_mov_b32_e32 v1, s9446; SI-NEXT: v_add_f64 v[0:1], s[14:15], -v[0:1]447; SI-NEXT: v_bfi_b32 v9, s3, v4, v5448; SI-NEXT: v_cmp_ge_f64_e64 s[10:11], |v[0:1]|, 0.5449; SI-NEXT: v_add_f64 v[0:1], s[4:5], v[8:9]450; SI-NEXT: s_and_b64 s[4:5], s[10:11], exec451; SI-NEXT: s_cselect_b32 s4, 0x3ff00000, 0452; SI-NEXT: v_mov_b32_e32 v6, s4453; SI-NEXT: s_bfe_u32 s4, s13, 0xb0014454; SI-NEXT: s_add_i32 s10, s4, 0xfffffc01455; SI-NEXT: s_lshr_b64 s[4:5], s[6:7], s10456; SI-NEXT: s_andn2_b64 s[4:5], s[12:13], s[4:5]457; SI-NEXT: s_and_b32 s11, s13, 0x80000000458; SI-NEXT: s_cmp_lt_i32 s10, 0459; SI-NEXT: s_cselect_b32 s4, 0, s4460; SI-NEXT: s_cselect_b32 s5, s11, s5461; SI-NEXT: s_cmp_gt_i32 s10, 51462; SI-NEXT: s_cselect_b32 s4, s12, s4463; SI-NEXT: s_cselect_b32 s5, s13, s5464; SI-NEXT: v_mov_b32_e32 v4, s4465; SI-NEXT: v_mov_b32_e32 v5, s5466; SI-NEXT: v_add_f64 v[4:5], s[12:13], -v[4:5]467; SI-NEXT: v_mov_b32_e32 v7, s15468; SI-NEXT: v_cmp_ge_f64_e64 s[10:11], |v[4:5]|, 0.5469; SI-NEXT: v_bfi_b32 v9, s3, v6, v7470; SI-NEXT: v_add_f64 v[6:7], s[8:9], v[8:9]471; SI-NEXT: s_and_b64 s[8:9], s[10:11], exec472; SI-NEXT: s_cselect_b32 s8, 0x3ff00000, 0473; SI-NEXT: v_mov_b32_e32 v9, s8474; SI-NEXT: s_bfe_u32 s8, s19, 0xb0014475; SI-NEXT: s_add_i32 s10, s8, 0xfffffc01476; SI-NEXT: s_lshr_b64 s[8:9], s[6:7], s10477; SI-NEXT: s_andn2_b64 s[8:9], s[18:19], s[8:9]478; SI-NEXT: s_and_b32 s11, s19, 0x80000000479; SI-NEXT: s_cmp_lt_i32 s10, 0480; SI-NEXT: s_cselect_b32 s8, 0, s8481; SI-NEXT: s_cselect_b32 s9, s11, s9482; SI-NEXT: s_cmp_gt_i32 s10, 51483; SI-NEXT: s_cselect_b32 s8, s18, s8484; SI-NEXT: s_cselect_b32 s9, s19, s9485; SI-NEXT: v_mov_b32_e32 v4, s8486; SI-NEXT: v_mov_b32_e32 v5, s9487; SI-NEXT: v_add_f64 v[4:5], s[18:19], -v[4:5]488; SI-NEXT: v_mov_b32_e32 v10, s13489; SI-NEXT: v_cmp_ge_f64_e64 s[10:11], |v[4:5]|, 0.5490; SI-NEXT: v_bfi_b32 v9, s3, v9, v10491; SI-NEXT: v_add_f64 v[4:5], s[4:5], v[8:9]492; SI-NEXT: s_and_b64 s[4:5], s[10:11], exec493; SI-NEXT: s_cselect_b32 s4, 0x3ff00000, 0494; SI-NEXT: v_mov_b32_e32 v12, s4495; SI-NEXT: s_bfe_u32 s4, s17, 0xb0014496; SI-NEXT: s_add_i32 s10, s4, 0xfffffc01497; SI-NEXT: s_lshr_b64 s[4:5], s[6:7], s10498; SI-NEXT: s_andn2_b64 s[4:5], s[16:17], s[4:5]499; SI-NEXT: s_and_b32 s11, s17, 0x80000000500; SI-NEXT: s_cmp_lt_i32 s10, 0501; SI-NEXT: s_cselect_b32 s4, 0, s4502; SI-NEXT: s_cselect_b32 s5, s11, s5503; SI-NEXT: s_cmp_gt_i32 s10, 51504; SI-NEXT: s_cselect_b32 s5, s17, s5505; SI-NEXT: s_cselect_b32 s4, s16, s4506; SI-NEXT: v_mov_b32_e32 v10, s5507; SI-NEXT: v_mov_b32_e32 v9, s4508; SI-NEXT: v_add_f64 v[10:11], s[16:17], -v[9:10]509; SI-NEXT: v_mov_b32_e32 v13, s19510; SI-NEXT: v_cmp_ge_f64_e64 s[10:11], |v[10:11]|, 0.5511; SI-NEXT: v_bfi_b32 v9, s3, v12, v13512; SI-NEXT: v_add_f64 v[12:13], s[8:9], v[8:9]513; SI-NEXT: s_and_b64 s[8:9], s[10:11], exec514; SI-NEXT: s_cselect_b32 s8, 0x3ff00000, 0515; SI-NEXT: v_mov_b32_e32 v14, s8516; SI-NEXT: s_bfe_u32 s8, s23, 0xb0014517; SI-NEXT: s_add_i32 s10, s8, 0xfffffc01518; SI-NEXT: s_lshr_b64 s[8:9], s[6:7], s10519; SI-NEXT: s_andn2_b64 s[8:9], s[22:23], s[8:9]520; SI-NEXT: s_and_b32 s11, s23, 0x80000000521; SI-NEXT: s_cmp_lt_i32 s10, 0522; SI-NEXT: s_cselect_b32 s8, 0, s8523; SI-NEXT: s_cselect_b32 s9, s11, s9524; SI-NEXT: s_cmp_gt_i32 s10, 51525; SI-NEXT: s_cselect_b32 s9, s23, s9526; SI-NEXT: s_cselect_b32 s8, s22, s8527; SI-NEXT: v_mov_b32_e32 v10, s9528; SI-NEXT: v_mov_b32_e32 v9, s8529; SI-NEXT: v_add_f64 v[10:11], s[22:23], -v[9:10]530; SI-NEXT: v_mov_b32_e32 v15, s17531; SI-NEXT: v_cmp_ge_f64_e64 s[10:11], |v[10:11]|, 0.5532; SI-NEXT: v_bfi_b32 v9, s3, v14, v15533; SI-NEXT: v_add_f64 v[10:11], s[4:5], v[8:9]534; SI-NEXT: s_and_b64 s[4:5], s[10:11], exec535; SI-NEXT: s_cselect_b32 s4, 0x3ff00000, 0536; SI-NEXT: v_mov_b32_e32 v9, s4537; SI-NEXT: s_bfe_u32 s4, s21, 0xb0014538; SI-NEXT: s_add_i32 s10, s4, 0xfffffc01539; SI-NEXT: s_lshr_b64 s[4:5], s[6:7], s10540; SI-NEXT: s_andn2_b64 s[4:5], s[20:21], s[4:5]541; SI-NEXT: s_and_b32 s6, s21, 0x80000000542; SI-NEXT: s_cmp_lt_i32 s10, 0543; SI-NEXT: s_cselect_b32 s4, 0, s4544; SI-NEXT: s_cselect_b32 s5, s6, s5545; SI-NEXT: s_cmp_gt_i32 s10, 51546; SI-NEXT: s_cselect_b32 s5, s21, s5547; SI-NEXT: s_cselect_b32 s4, s20, s4548; SI-NEXT: v_mov_b32_e32 v15, s5549; SI-NEXT: v_mov_b32_e32 v14, s4550; SI-NEXT: v_add_f64 v[14:15], s[20:21], -v[14:15]551; SI-NEXT: v_mov_b32_e32 v16, s23552; SI-NEXT: v_cmp_ge_f64_e64 s[6:7], |v[14:15]|, 0.5553; SI-NEXT: v_bfi_b32 v9, s3, v9, v16554; SI-NEXT: s_and_b64 s[6:7], s[6:7], exec555; SI-NEXT: s_cselect_b32 s6, 0x3ff00000, 0556; SI-NEXT: v_add_f64 v[16:17], s[8:9], v[8:9]557; SI-NEXT: v_mov_b32_e32 v9, s6558; SI-NEXT: v_mov_b32_e32 v14, s21559; SI-NEXT: v_bfi_b32 v9, s3, v9, v14560; SI-NEXT: v_add_f64 v[14:15], s[4:5], v[8:9]561; SI-NEXT: s_mov_b32 s3, 0xf000562; SI-NEXT: s_waitcnt lgkmcnt(0)563; SI-NEXT: buffer_store_dwordx4 v[14:17], off, s[0:3], 0 offset:48564; SI-NEXT: buffer_store_dwordx4 v[10:13], off, s[0:3], 0 offset:32565; SI-NEXT: buffer_store_dwordx4 v[4:7], off, s[0:3], 0 offset:16566; SI-NEXT: buffer_store_dwordx4 v[0:3], off, s[0:3], 0567; SI-NEXT: s_endpgm568;569; CI-LABEL: round_v8f64:570; CI: ; %bb.0:571; CI-NEXT: s_load_dwordx16 s[8:23], s[4:5], 0x19572; CI-NEXT: s_brev_b32 s6, -2573; CI-NEXT: v_mov_b32_e32 v12, 0574; CI-NEXT: s_waitcnt lgkmcnt(0)575; CI-NEXT: v_trunc_f64_e32 v[0:1], s[10:11]576; CI-NEXT: v_trunc_f64_e32 v[4:5], s[8:9]577; CI-NEXT: v_add_f64 v[2:3], s[10:11], -v[0:1]578; CI-NEXT: v_add_f64 v[6:7], s[8:9], -v[4:5]579; CI-NEXT: v_cmp_ge_f64_e64 s[0:1], |v[2:3]|, 0.5580; CI-NEXT: v_cmp_ge_f64_e64 s[2:3], |v[6:7]|, 0.5581; CI-NEXT: s_and_b64 s[0:1], s[0:1], exec582; CI-NEXT: s_cselect_b32 s7, 0x3ff00000, 0583; CI-NEXT: v_mov_b32_e32 v8, s11584; CI-NEXT: s_and_b64 s[0:1], s[2:3], exec585; CI-NEXT: v_mov_b32_e32 v2, s7586; CI-NEXT: v_trunc_f64_e32 v[6:7], s[14:15]587; CI-NEXT: v_bfi_b32 v13, s6, v2, v8588; CI-NEXT: s_cselect_b32 s0, 0x3ff00000, 0589; CI-NEXT: v_add_f64 v[2:3], v[0:1], v[12:13]590; CI-NEXT: v_mov_b32_e32 v8, s0591; CI-NEXT: v_mov_b32_e32 v9, s9592; CI-NEXT: v_add_f64 v[0:1], s[14:15], -v[6:7]593; CI-NEXT: v_bfi_b32 v13, s6, v8, v9594; CI-NEXT: v_cmp_ge_f64_e64 s[0:1], |v[0:1]|, 0.5595; CI-NEXT: v_add_f64 v[0:1], v[4:5], v[12:13]596; CI-NEXT: v_trunc_f64_e32 v[4:5], s[12:13]597; CI-NEXT: s_and_b64 s[0:1], s[0:1], exec598; CI-NEXT: v_add_f64 v[8:9], s[12:13], -v[4:5]599; CI-NEXT: s_cselect_b32 s0, 0x3ff00000, 0600; CI-NEXT: v_mov_b32_e32 v10, s0601; CI-NEXT: v_cmp_ge_f64_e64 s[0:1], |v[8:9]|, 0.5602; CI-NEXT: v_trunc_f64_e32 v[8:9], s[18:19]603; CI-NEXT: v_mov_b32_e32 v11, s15604; CI-NEXT: v_bfi_b32 v13, s6, v10, v11605; CI-NEXT: s_and_b64 s[0:1], s[0:1], exec606; CI-NEXT: v_add_f64 v[10:11], s[18:19], -v[8:9]607; CI-NEXT: s_cselect_b32 s0, 0x3ff00000, 0608; CI-NEXT: v_add_f64 v[6:7], v[6:7], v[12:13]609; CI-NEXT: v_mov_b32_e32 v13, s0610; CI-NEXT: v_mov_b32_e32 v14, s13611; CI-NEXT: v_cmp_ge_f64_e64 s[0:1], |v[10:11]|, 0.5612; CI-NEXT: v_bfi_b32 v13, s6, v13, v14613; CI-NEXT: v_trunc_f64_e32 v[14:15], s[16:17]614; CI-NEXT: s_and_b64 s[0:1], s[0:1], exec615; CI-NEXT: s_cselect_b32 s0, 0x3ff00000, 0616; CI-NEXT: v_add_f64 v[10:11], s[16:17], -v[14:15]617; CI-NEXT: v_add_f64 v[4:5], v[4:5], v[12:13]618; CI-NEXT: v_mov_b32_e32 v13, s0619; CI-NEXT: v_mov_b32_e32 v16, s19620; CI-NEXT: v_bfi_b32 v13, s6, v13, v16621; CI-NEXT: v_cmp_ge_f64_e64 s[0:1], |v[10:11]|, 0.5622; CI-NEXT: v_trunc_f64_e32 v[16:17], s[22:23]623; CI-NEXT: s_and_b64 s[0:1], s[0:1], exec624; CI-NEXT: v_add_f64 v[18:19], s[22:23], -v[16:17]625; CI-NEXT: s_cselect_b32 s0, 0x3ff00000, 0626; CI-NEXT: v_add_f64 v[10:11], v[8:9], v[12:13]627; CI-NEXT: v_mov_b32_e32 v8, s0628; CI-NEXT: v_mov_b32_e32 v9, s17629; CI-NEXT: v_cmp_ge_f64_e64 s[0:1], |v[18:19]|, 0.5630; CI-NEXT: v_trunc_f64_e32 v[18:19], s[20:21]631; CI-NEXT: v_bfi_b32 v13, s6, v8, v9632; CI-NEXT: v_add_f64 v[8:9], v[14:15], v[12:13]633; CI-NEXT: v_add_f64 v[13:14], s[20:21], -v[18:19]634; CI-NEXT: s_and_b64 s[0:1], s[0:1], exec635; CI-NEXT: v_cmp_ge_f64_e64 s[0:1], |v[13:14]|, 0.5636; CI-NEXT: s_cselect_b32 s2, 0x3ff00000, 0637; CI-NEXT: s_and_b64 s[0:1], s[0:1], exec638; CI-NEXT: s_cselect_b32 s0, 0x3ff00000, 0639; CI-NEXT: v_mov_b32_e32 v13, s2640; CI-NEXT: v_mov_b32_e32 v14, s23641; CI-NEXT: v_mov_b32_e32 v20, s0642; CI-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x9643; CI-NEXT: v_bfi_b32 v13, s6, v13, v14644; CI-NEXT: v_mov_b32_e32 v21, s21645; CI-NEXT: v_add_f64 v[14:15], v[16:17], v[12:13]646; CI-NEXT: v_bfi_b32 v13, s6, v20, v21647; CI-NEXT: v_add_f64 v[12:13], v[18:19], v[12:13]648; CI-NEXT: s_mov_b32 s3, 0xf000649; CI-NEXT: s_mov_b32 s2, -1650; CI-NEXT: s_waitcnt lgkmcnt(0)651; CI-NEXT: buffer_store_dwordx4 v[12:15], off, s[0:3], 0 offset:48652; CI-NEXT: buffer_store_dwordx4 v[8:11], off, s[0:3], 0 offset:32653; CI-NEXT: buffer_store_dwordx4 v[4:7], off, s[0:3], 0 offset:16654; CI-NEXT: buffer_store_dwordx4 v[0:3], off, s[0:3], 0655; CI-NEXT: s_endpgm656 %result = call <8 x double> @llvm.round.v8f64(<8 x double> %in) #1657 store <8 x double> %result, ptr addrspace(1) %out658 ret void659}660 661declare i32 @llvm.amdgcn.workitem.id.x() #1662 663declare double @llvm.round.f64(double) #1664declare <2 x double> @llvm.round.v2f64(<2 x double>) #1665declare <4 x double> @llvm.round.v4f64(<4 x double>) #1666declare <8 x double> @llvm.round.v8f64(<8 x double>) #1667 668attributes #0 = { nounwind }669attributes #1 = { nounwind readnone }670