276 lines · plain
1; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py UTC_ARGS: --version 42; RUN: llc -global-isel=0 -mtriple=amdgcn-amd-amdhsa -mcpu=gfx900 < %s | FileCheck -check-prefixes=GCN,SDAG %s3; RUN: not --crash llc -global-isel=1 -mtriple=amdgcn-amd-amdhsa -mcpu=gfx900 < %s 2>&1 | FileCheck -check-prefix=GISEL %s4 5; FIXME: GISEL can't handle the "fptrunc float to bfloat" that expand-fp emits.6 7; GISEL: unable to translate instruction: fptrunc8 9define bfloat @sitofp_i128_to_bf16(i128 %x) {10; GCN-LABEL: sitofp_i128_to_bf16:11; GCN: ; %bb.0: ; %itofp-entry12; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)13; GCN-NEXT: v_or_b32_e32 v5, v1, v314; GCN-NEXT: v_or_b32_e32 v4, v0, v215; GCN-NEXT: v_cmp_ne_u64_e32 vcc, 0, v[4:5]16; GCN-NEXT: v_mov_b32_e32 v4, 017; GCN-NEXT: s_and_saveexec_b64 s[6:7], vcc18; GCN-NEXT: s_cbranch_execz .LBB0_1419; GCN-NEXT: ; %bb.1: ; %itofp-if-end20; GCN-NEXT: v_sub_co_u32_e32 v4, vcc, 0, v021; GCN-NEXT: v_subb_co_u32_e32 v5, vcc, 0, v1, vcc22; GCN-NEXT: v_subb_co_u32_e32 v6, vcc, 0, v2, vcc23; GCN-NEXT: v_subb_co_u32_e32 v7, vcc, 0, v3, vcc24; GCN-NEXT: v_cmp_gt_i64_e32 vcc, 0, v[2:3]25; GCN-NEXT: ; implicit-def: $vgpr826; GCN-NEXT: v_cndmask_b32_e32 v0, v0, v4, vcc27; GCN-NEXT: v_cndmask_b32_e32 v4, v2, v6, vcc28; GCN-NEXT: v_cndmask_b32_e32 v1, v1, v5, vcc29; GCN-NEXT: v_cndmask_b32_e32 v5, v3, v7, vcc30; GCN-NEXT: v_ffbh_u32_e32 v2, v431; GCN-NEXT: v_add_u32_e32 v2, 32, v232; GCN-NEXT: v_ffbh_u32_e32 v6, v533; GCN-NEXT: v_min_u32_e32 v2, v2, v634; GCN-NEXT: v_ffbh_u32_e32 v6, v035; GCN-NEXT: v_add_u32_e32 v6, 32, v636; GCN-NEXT: v_ffbh_u32_e32 v7, v137; GCN-NEXT: v_min_u32_e32 v6, v6, v738; GCN-NEXT: v_cmp_ne_u64_e32 vcc, 0, v[4:5]39; GCN-NEXT: v_add_u32_e32 v6, 64, v640; GCN-NEXT: v_cndmask_b32_e32 v7, v6, v2, vcc41; GCN-NEXT: v_sub_u32_e32 v2, 0x80, v742; GCN-NEXT: v_cmp_gt_i32_e32 vcc, 25, v243; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc44; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5]45; GCN-NEXT: ; %bb.2: ; %itofp-if-else46; GCN-NEXT: v_add_u32_e32 v2, 0xffffff98, v747; GCN-NEXT: v_lshlrev_b64 v[0:1], v2, v[0:1]48; GCN-NEXT: v_cmp_gt_u32_e32 vcc, 64, v249; GCN-NEXT: v_cndmask_b32_e32 v8, 0, v0, vcc50; GCN-NEXT: ; implicit-def: $vgpr251; GCN-NEXT: ; implicit-def: $vgpr0_vgpr152; GCN-NEXT: ; implicit-def: $vgpr4_vgpr553; GCN-NEXT: ; %bb.3: ; %Flow354; GCN-NEXT: s_or_saveexec_b64 s[8:9], s[4:5]55; GCN-NEXT: v_sub_u32_e32 v6, 0x7f, v756; GCN-NEXT: s_xor_b64 exec, exec, s[8:9]57; GCN-NEXT: s_cbranch_execz .LBB0_1358; GCN-NEXT: ; %bb.4: ; %NodeBlock59; GCN-NEXT: v_cmp_lt_i32_e32 vcc, 25, v260; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc61; GCN-NEXT: s_xor_b64 s[10:11], exec, s[4:5]62; GCN-NEXT: s_cbranch_execz .LBB0_863; GCN-NEXT: ; %bb.5: ; %LeafBlock64; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 26, v265; GCN-NEXT: s_and_saveexec_b64 s[12:13], vcc66; GCN-NEXT: s_cbranch_execz .LBB0_767; GCN-NEXT: ; %bb.6: ; %itofp-sw-default68; GCN-NEXT: v_sub_u32_e32 v12, 0x66, v769; GCN-NEXT: v_sub_u32_e32 v10, 64, v1270; GCN-NEXT: v_lshrrev_b64 v[8:9], v12, v[0:1]71; GCN-NEXT: v_lshlrev_b64 v[10:11], v10, v[4:5]72; GCN-NEXT: v_sub_u32_e32 v13, 38, v773; GCN-NEXT: v_or_b32_e32 v11, v9, v1174; GCN-NEXT: v_or_b32_e32 v10, v8, v1075; GCN-NEXT: v_lshrrev_b64 v[8:9], v13, v[4:5]76; GCN-NEXT: v_cmp_gt_u32_e32 vcc, 64, v1277; GCN-NEXT: v_add_u32_e32 v14, 26, v778; GCN-NEXT: v_cndmask_b32_e32 v9, v9, v11, vcc79; GCN-NEXT: v_cmp_eq_u32_e64 s[4:5], 0, v1280; GCN-NEXT: v_cndmask_b32_e32 v8, v8, v10, vcc81; GCN-NEXT: v_lshrrev_b64 v[10:11], v13, v[0:1]82; GCN-NEXT: v_lshlrev_b64 v[12:13], v14, v[4:5]83; GCN-NEXT: v_subrev_u32_e32 v7, 38, v784; GCN-NEXT: v_cndmask_b32_e64 v15, v8, v0, s[4:5]85; GCN-NEXT: v_lshlrev_b64 v[7:8], v7, v[0:1]86; GCN-NEXT: v_cndmask_b32_e64 v9, v9, v1, s[4:5]87; GCN-NEXT: v_or_b32_e32 v11, v13, v1188; GCN-NEXT: v_or_b32_e32 v10, v12, v1089; GCN-NEXT: v_cmp_gt_u32_e32 vcc, 64, v1490; GCN-NEXT: v_lshlrev_b64 v[0:1], v14, v[0:1]91; GCN-NEXT: v_cndmask_b32_e32 v8, v8, v11, vcc92; GCN-NEXT: v_cmp_eq_u32_e64 s[4:5], 0, v1493; GCN-NEXT: v_cndmask_b32_e32 v7, v7, v10, vcc94; GCN-NEXT: v_cndmask_b32_e64 v5, v8, v5, s[4:5]95; GCN-NEXT: v_cndmask_b32_e64 v4, v7, v4, s[4:5]96; GCN-NEXT: v_cndmask_b32_e32 v1, 0, v1, vcc97; GCN-NEXT: v_cndmask_b32_e32 v0, 0, v0, vcc98; GCN-NEXT: v_or_b32_e32 v1, v1, v599; GCN-NEXT: v_or_b32_e32 v0, v0, v4100; GCN-NEXT: v_cmp_ne_u64_e32 vcc, 0, v[0:1]101; GCN-NEXT: v_cndmask_b32_e64 v0, 0, 1, vcc102; GCN-NEXT: v_or_b32_e32 v8, v15, v0103; GCN-NEXT: v_mov_b32_e32 v0, v8104; GCN-NEXT: v_mov_b32_e32 v1, v9105; GCN-NEXT: .LBB0_7: ; %Flow1106; GCN-NEXT: s_or_b64 exec, exec, s[12:13]107; GCN-NEXT: .LBB0_8: ; %Flow2108; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[10:11]109; GCN-NEXT: ; %bb.9: ; %itofp-sw-bb110; GCN-NEXT: v_lshlrev_b64 v[0:1], 1, v[0:1]111; GCN-NEXT: ; %bb.10: ; %itofp-sw-epilog112; GCN-NEXT: s_or_b64 exec, exec, s[4:5]113; GCN-NEXT: v_lshrrev_b32_e32 v4, 2, v0114; GCN-NEXT: v_and_or_b32 v0, v4, 1, v0115; GCN-NEXT: v_add_co_u32_e32 v0, vcc, 1, v0116; GCN-NEXT: v_addc_co_u32_e32 v1, vcc, 0, v1, vcc117; GCN-NEXT: v_and_b32_e32 v4, 0x4000000, v0118; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v4119; GCN-NEXT: v_alignbit_b32 v8, v1, v0, 2120; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc121; GCN-NEXT: ; %bb.11: ; %itofp-if-then20122; GCN-NEXT: v_alignbit_b32 v8, v1, v0, 3123; GCN-NEXT: v_mov_b32_e32 v6, v2124; GCN-NEXT: ; %bb.12: ; %Flow125; GCN-NEXT: s_or_b64 exec, exec, s[4:5]126; GCN-NEXT: .LBB0_13: ; %Flow4127; GCN-NEXT: s_or_b64 exec, exec, s[8:9]128; GCN-NEXT: v_and_b32_e32 v0, 0x80000000, v3129; GCN-NEXT: v_lshl_add_u32 v1, v6, 23, 1.0130; GCN-NEXT: v_and_b32_e32 v2, 0x7fffff, v8131; GCN-NEXT: v_or3_b32 v0, v2, v0, v1132; GCN-NEXT: v_bfe_u32 v1, v8, 16, 1133; GCN-NEXT: s_movk_i32 s4, 0x7fff134; GCN-NEXT: v_add3_u32 v1, v1, v0, s4135; GCN-NEXT: v_or_b32_e32 v2, 0x400000, v0136; GCN-NEXT: v_cmp_u_f32_e32 vcc, v0, v0137; GCN-NEXT: v_cndmask_b32_e32 v0, v1, v2, vcc138; GCN-NEXT: v_lshrrev_b32_e32 v4, 16, v0139; GCN-NEXT: .LBB0_14: ; %Flow5140; GCN-NEXT: s_or_b64 exec, exec, s[6:7]141; GCN-NEXT: v_mov_b32_e32 v0, v4142; GCN-NEXT: s_setpc_b64 s[30:31]143 %cvt = sitofp i128 %x to bfloat144 ret bfloat %cvt145}146 147define bfloat @uitofp_i128_to_bf16(i128 %x) {148; GCN-LABEL: uitofp_i128_to_bf16:149; GCN: ; %bb.0: ; %itofp-entry150; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)151; GCN-NEXT: v_or_b32_e32 v5, v1, v3152; GCN-NEXT: v_or_b32_e32 v4, v0, v2153; GCN-NEXT: v_cmp_ne_u64_e32 vcc, 0, v[4:5]154; GCN-NEXT: v_mov_b32_e32 v4, 0155; GCN-NEXT: s_and_saveexec_b64 s[6:7], vcc156; GCN-NEXT: s_cbranch_execz .LBB1_14157; GCN-NEXT: ; %bb.1: ; %itofp-if-end158; GCN-NEXT: v_ffbh_u32_e32 v4, v2159; GCN-NEXT: v_add_u32_e32 v4, 32, v4160; GCN-NEXT: v_ffbh_u32_e32 v5, v3161; GCN-NEXT: v_min_u32_e32 v4, v4, v5162; GCN-NEXT: v_ffbh_u32_e32 v5, v0163; GCN-NEXT: v_add_u32_e32 v5, 32, v5164; GCN-NEXT: v_ffbh_u32_e32 v6, v1165; GCN-NEXT: v_min_u32_e32 v5, v5, v6166; GCN-NEXT: v_cmp_ne_u64_e32 vcc, 0, v[2:3]167; GCN-NEXT: v_add_u32_e32 v5, 64, v5168; GCN-NEXT: v_cndmask_b32_e32 v6, v5, v4, vcc169; GCN-NEXT: v_sub_u32_e32 v4, 0x80, v6170; GCN-NEXT: v_cmp_gt_i32_e32 vcc, 25, v4171; GCN-NEXT: ; implicit-def: $vgpr7172; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc173; GCN-NEXT: s_xor_b64 s[4:5], exec, s[4:5]174; GCN-NEXT: ; %bb.2: ; %itofp-if-else175; GCN-NEXT: v_add_u32_e32 v2, 0xffffff98, v6176; GCN-NEXT: v_lshlrev_b64 v[0:1], v2, v[0:1]177; GCN-NEXT: v_cmp_gt_u32_e32 vcc, 64, v2178; GCN-NEXT: v_cndmask_b32_e32 v7, 0, v0, vcc179; GCN-NEXT: ; implicit-def: $vgpr4180; GCN-NEXT: ; implicit-def: $vgpr0_vgpr1181; GCN-NEXT: ; implicit-def: $vgpr2_vgpr3182; GCN-NEXT: ; %bb.3: ; %Flow3183; GCN-NEXT: s_or_saveexec_b64 s[8:9], s[4:5]184; GCN-NEXT: v_sub_u32_e32 v5, 0x7f, v6185; GCN-NEXT: s_xor_b64 exec, exec, s[8:9]186; GCN-NEXT: s_cbranch_execz .LBB1_13187; GCN-NEXT: ; %bb.4: ; %NodeBlock188; GCN-NEXT: v_cmp_lt_i32_e32 vcc, 25, v4189; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc190; GCN-NEXT: s_xor_b64 s[10:11], exec, s[4:5]191; GCN-NEXT: s_cbranch_execz .LBB1_8192; GCN-NEXT: ; %bb.5: ; %LeafBlock193; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 26, v4194; GCN-NEXT: s_and_saveexec_b64 s[12:13], vcc195; GCN-NEXT: s_cbranch_execz .LBB1_7196; GCN-NEXT: ; %bb.6: ; %itofp-sw-default197; GCN-NEXT: v_sub_u32_e32 v11, 0x66, v6198; GCN-NEXT: v_sub_u32_e32 v9, 64, v11199; GCN-NEXT: v_lshrrev_b64 v[7:8], v11, v[0:1]200; GCN-NEXT: v_lshlrev_b64 v[9:10], v9, v[2:3]201; GCN-NEXT: v_sub_u32_e32 v12, 38, v6202; GCN-NEXT: v_or_b32_e32 v10, v8, v10203; GCN-NEXT: v_or_b32_e32 v9, v7, v9204; GCN-NEXT: v_lshrrev_b64 v[7:8], v12, v[2:3]205; GCN-NEXT: v_cmp_gt_u32_e32 vcc, 64, v11206; GCN-NEXT: v_add_u32_e32 v13, 26, v6207; GCN-NEXT: v_cndmask_b32_e32 v8, v8, v10, vcc208; GCN-NEXT: v_cmp_eq_u32_e64 s[4:5], 0, v11209; GCN-NEXT: v_cndmask_b32_e32 v7, v7, v9, vcc210; GCN-NEXT: v_lshrrev_b64 v[9:10], v12, v[0:1]211; GCN-NEXT: v_lshlrev_b64 v[11:12], v13, v[2:3]212; GCN-NEXT: v_subrev_u32_e32 v6, 38, v6213; GCN-NEXT: v_cndmask_b32_e64 v14, v7, v0, s[4:5]214; GCN-NEXT: v_lshlrev_b64 v[6:7], v6, v[0:1]215; GCN-NEXT: v_cndmask_b32_e64 v8, v8, v1, s[4:5]216; GCN-NEXT: v_or_b32_e32 v10, v12, v10217; GCN-NEXT: v_or_b32_e32 v9, v11, v9218; GCN-NEXT: v_cmp_gt_u32_e32 vcc, 64, v13219; GCN-NEXT: v_lshlrev_b64 v[0:1], v13, v[0:1]220; GCN-NEXT: v_cndmask_b32_e32 v7, v7, v10, vcc221; GCN-NEXT: v_cmp_eq_u32_e64 s[4:5], 0, v13222; GCN-NEXT: v_cndmask_b32_e32 v6, v6, v9, vcc223; GCN-NEXT: v_cndmask_b32_e64 v3, v7, v3, s[4:5]224; GCN-NEXT: v_cndmask_b32_e64 v2, v6, v2, s[4:5]225; GCN-NEXT: v_cndmask_b32_e32 v1, 0, v1, vcc226; GCN-NEXT: v_cndmask_b32_e32 v0, 0, v0, vcc227; GCN-NEXT: v_or_b32_e32 v1, v1, v3228; GCN-NEXT: v_or_b32_e32 v0, v0, v2229; GCN-NEXT: v_cmp_ne_u64_e32 vcc, 0, v[0:1]230; GCN-NEXT: v_cndmask_b32_e64 v0, 0, 1, vcc231; GCN-NEXT: v_or_b32_e32 v7, v14, v0232; GCN-NEXT: v_mov_b32_e32 v0, v7233; GCN-NEXT: v_mov_b32_e32 v1, v8234; GCN-NEXT: .LBB1_7: ; %Flow1235; GCN-NEXT: s_or_b64 exec, exec, s[12:13]236; GCN-NEXT: .LBB1_8: ; %Flow2237; GCN-NEXT: s_andn2_saveexec_b64 s[4:5], s[10:11]238; GCN-NEXT: ; %bb.9: ; %itofp-sw-bb239; GCN-NEXT: v_lshlrev_b64 v[0:1], 1, v[0:1]240; GCN-NEXT: ; %bb.10: ; %itofp-sw-epilog241; GCN-NEXT: s_or_b64 exec, exec, s[4:5]242; GCN-NEXT: v_lshrrev_b32_e32 v2, 2, v0243; GCN-NEXT: v_and_or_b32 v0, v2, 1, v0244; GCN-NEXT: v_add_co_u32_e32 v0, vcc, 1, v0245; GCN-NEXT: v_addc_co_u32_e32 v1, vcc, 0, v1, vcc246; GCN-NEXT: v_and_b32_e32 v2, 0x4000000, v0247; GCN-NEXT: v_cmp_ne_u32_e32 vcc, 0, v2248; GCN-NEXT: v_alignbit_b32 v7, v1, v0, 2249; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc250; GCN-NEXT: ; %bb.11: ; %itofp-if-then20251; GCN-NEXT: v_alignbit_b32 v7, v1, v0, 3252; GCN-NEXT: v_mov_b32_e32 v5, v4253; GCN-NEXT: ; %bb.12: ; %Flow254; GCN-NEXT: s_or_b64 exec, exec, s[4:5]255; GCN-NEXT: .LBB1_13: ; %Flow4256; GCN-NEXT: s_or_b64 exec, exec, s[8:9]257; GCN-NEXT: v_and_b32_e32 v0, 0x7fffff, v7258; GCN-NEXT: v_lshl_or_b32 v0, v5, 23, v0259; GCN-NEXT: v_add_u32_e32 v0, 1.0, v0260; GCN-NEXT: v_bfe_u32 v1, v0, 16, 1261; GCN-NEXT: s_movk_i32 s4, 0x7fff262; GCN-NEXT: v_add3_u32 v1, v1, v0, s4263; GCN-NEXT: v_or_b32_e32 v2, 0x400000, v0264; GCN-NEXT: v_cmp_u_f32_e32 vcc, v0, v0265; GCN-NEXT: v_cndmask_b32_e32 v0, v1, v2, vcc266; GCN-NEXT: v_lshrrev_b32_e32 v4, 16, v0267; GCN-NEXT: .LBB1_14: ; %Flow5268; GCN-NEXT: s_or_b64 exec, exec, s[6:7]269; GCN-NEXT: v_mov_b32_e32 v0, v4270; GCN-NEXT: s_setpc_b64 s[30:31]271 %cvt = uitofp i128 %x to bfloat272 ret bfloat %cvt273}274;; NOTE: These prefixes are unused and the list is autogenerated. Do not add tests below this line:275; SDAG: {{.*}}276