323 lines · plain
1; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py2; RUN: llc < %s -mtriple=riscv32 -mattr=+m | FileCheck %s --check-prefixes=RV32,RV32I3; RUN: llc < %s -mtriple=riscv64 -mattr=+m | FileCheck %s --check-prefixes=RV64,RV64I4; RUN: llc < %s -mtriple=riscv32 -mattr=+m,+zbb | FileCheck %s --check-prefixes=RV32,RV32IZbb5; RUN: llc < %s -mtriple=riscv64 -mattr=+m,+zbb | FileCheck %s --check-prefixes=RV64,RV64IZbb6 7define i32 @func32(i32 %x, i32 %y, i32 %z) nounwind {8; RV32-LABEL: func32:9; RV32: # %bb.0:10; RV32-NEXT: mv a3, a011; RV32-NEXT: mul a0, a1, a212; RV32-NEXT: sgtz a1, a013; RV32-NEXT: sub a0, a3, a014; RV32-NEXT: slt a2, a0, a315; RV32-NEXT: beq a1, a2, .LBB0_216; RV32-NEXT: # %bb.1:17; RV32-NEXT: srai a0, a0, 3118; RV32-NEXT: lui a1, 52428819; RV32-NEXT: xor a0, a0, a120; RV32-NEXT: .LBB0_2:21; RV32-NEXT: ret22;23; RV64I-LABEL: func32:24; RV64I: # %bb.0:25; RV64I-NEXT: sext.w a0, a026; RV64I-NEXT: mulw a1, a1, a227; RV64I-NEXT: subw a2, a0, a128; RV64I-NEXT: sub a0, a0, a129; RV64I-NEXT: beq a2, a0, .LBB0_230; RV64I-NEXT: # %bb.1:31; RV64I-NEXT: sraiw a0, a0, 3132; RV64I-NEXT: lui a1, 52428833; RV64I-NEXT: xor a0, a0, a134; RV64I-NEXT: .LBB0_2:35; RV64I-NEXT: ret36;37; RV64IZbb-LABEL: func32:38; RV64IZbb: # %bb.0:39; RV64IZbb-NEXT: sext.w a0, a040; RV64IZbb-NEXT: mulw a1, a1, a241; RV64IZbb-NEXT: lui a2, 52428842; RV64IZbb-NEXT: sub a0, a0, a143; RV64IZbb-NEXT: addiw a1, a2, -144; RV64IZbb-NEXT: min a0, a0, a145; RV64IZbb-NEXT: max a0, a0, a246; RV64IZbb-NEXT: ret47 %a = mul i32 %y, %z48 %tmp = call i32 @llvm.ssub.sat.i32(i32 %x, i32 %a)49 ret i32 %tmp50}51 52define i64 @func64(i64 %x, i64 %y, i64 %z) nounwind {53; RV32-LABEL: func64:54; RV32: # %bb.0:55; RV32-NEXT: mv a2, a156; RV32-NEXT: sltu a1, a0, a457; RV32-NEXT: sub a3, a2, a558; RV32-NEXT: sub a1, a3, a159; RV32-NEXT: xor a3, a2, a160; RV32-NEXT: xor a2, a2, a561; RV32-NEXT: and a2, a2, a362; RV32-NEXT: bltz a2, .LBB1_263; RV32-NEXT: # %bb.1:64; RV32-NEXT: sub a0, a0, a465; RV32-NEXT: ret66; RV32-NEXT: .LBB1_2:67; RV32-NEXT: srai a0, a1, 3168; RV32-NEXT: lui a1, 52428869; RV32-NEXT: xor a1, a0, a170; RV32-NEXT: ret71;72; RV64-LABEL: func64:73; RV64: # %bb.0:74; RV64-NEXT: mv a1, a075; RV64-NEXT: sgtz a3, a276; RV64-NEXT: sub a0, a0, a277; RV64-NEXT: slt a1, a0, a178; RV64-NEXT: beq a3, a1, .LBB1_279; RV64-NEXT: # %bb.1:80; RV64-NEXT: srai a0, a0, 6381; RV64-NEXT: li a1, -182; RV64-NEXT: slli a1, a1, 6383; RV64-NEXT: xor a0, a0, a184; RV64-NEXT: .LBB1_2:85; RV64-NEXT: ret86 %a = mul i64 %y, %z87 %tmp = call i64 @llvm.ssub.sat.i64(i64 %x, i64 %z)88 ret i64 %tmp89}90 91define i16 @func16(i16 %x, i16 %y, i16 %z) nounwind {92; RV32I-LABEL: func16:93; RV32I: # %bb.0:94; RV32I-NEXT: slli a0, a0, 1695; RV32I-NEXT: mul a1, a1, a296; RV32I-NEXT: lui a2, 897; RV32I-NEXT: srai a0, a0, 1698; RV32I-NEXT: slli a1, a1, 1699; RV32I-NEXT: srai a1, a1, 16100; RV32I-NEXT: sub a0, a0, a1101; RV32I-NEXT: addi a1, a2, -1102; RV32I-NEXT: bge a0, a1, .LBB2_3103; RV32I-NEXT: # %bb.1:104; RV32I-NEXT: lui a1, 1048568105; RV32I-NEXT: bge a1, a0, .LBB2_4106; RV32I-NEXT: .LBB2_2:107; RV32I-NEXT: ret108; RV32I-NEXT: .LBB2_3:109; RV32I-NEXT: mv a0, a1110; RV32I-NEXT: lui a1, 1048568111; RV32I-NEXT: blt a1, a0, .LBB2_2112; RV32I-NEXT: .LBB2_4:113; RV32I-NEXT: lui a0, 1048568114; RV32I-NEXT: ret115;116; RV64I-LABEL: func16:117; RV64I: # %bb.0:118; RV64I-NEXT: slli a0, a0, 48119; RV64I-NEXT: mul a1, a1, a2120; RV64I-NEXT: lui a2, 8121; RV64I-NEXT: srai a0, a0, 48122; RV64I-NEXT: slli a1, a1, 48123; RV64I-NEXT: srai a1, a1, 48124; RV64I-NEXT: sub a0, a0, a1125; RV64I-NEXT: addi a1, a2, -1126; RV64I-NEXT: bge a0, a1, .LBB2_3127; RV64I-NEXT: # %bb.1:128; RV64I-NEXT: lui a1, 1048568129; RV64I-NEXT: bge a1, a0, .LBB2_4130; RV64I-NEXT: .LBB2_2:131; RV64I-NEXT: ret132; RV64I-NEXT: .LBB2_3:133; RV64I-NEXT: mv a0, a1134; RV64I-NEXT: lui a1, 1048568135; RV64I-NEXT: blt a1, a0, .LBB2_2136; RV64I-NEXT: .LBB2_4:137; RV64I-NEXT: lui a0, 1048568138; RV64I-NEXT: ret139;140; RV32IZbb-LABEL: func16:141; RV32IZbb: # %bb.0:142; RV32IZbb-NEXT: sext.h a0, a0143; RV32IZbb-NEXT: mul a1, a1, a2144; RV32IZbb-NEXT: lui a2, 8145; RV32IZbb-NEXT: sext.h a1, a1146; RV32IZbb-NEXT: addi a2, a2, -1147; RV32IZbb-NEXT: sub a0, a0, a1148; RV32IZbb-NEXT: min a0, a0, a2149; RV32IZbb-NEXT: lui a1, 1048568150; RV32IZbb-NEXT: max a0, a0, a1151; RV32IZbb-NEXT: ret152;153; RV64IZbb-LABEL: func16:154; RV64IZbb: # %bb.0:155; RV64IZbb-NEXT: sext.h a0, a0156; RV64IZbb-NEXT: mul a1, a1, a2157; RV64IZbb-NEXT: lui a2, 8158; RV64IZbb-NEXT: sext.h a1, a1159; RV64IZbb-NEXT: addi a2, a2, -1160; RV64IZbb-NEXT: sub a0, a0, a1161; RV64IZbb-NEXT: min a0, a0, a2162; RV64IZbb-NEXT: lui a1, 1048568163; RV64IZbb-NEXT: max a0, a0, a1164; RV64IZbb-NEXT: ret165 %a = mul i16 %y, %z166 %tmp = call i16 @llvm.ssub.sat.i16(i16 %x, i16 %a)167 ret i16 %tmp168}169 170define i8 @func8(i8 %x, i8 %y, i8 %z) nounwind {171; RV32I-LABEL: func8:172; RV32I: # %bb.0:173; RV32I-NEXT: slli a0, a0, 24174; RV32I-NEXT: mul a1, a1, a2175; RV32I-NEXT: srai a0, a0, 24176; RV32I-NEXT: slli a1, a1, 24177; RV32I-NEXT: srai a1, a1, 24178; RV32I-NEXT: sub a0, a0, a1179; RV32I-NEXT: li a1, 127180; RV32I-NEXT: bge a0, a1, .LBB3_3181; RV32I-NEXT: # %bb.1:182; RV32I-NEXT: li a1, -128183; RV32I-NEXT: bge a1, a0, .LBB3_4184; RV32I-NEXT: .LBB3_2:185; RV32I-NEXT: ret186; RV32I-NEXT: .LBB3_3:187; RV32I-NEXT: li a0, 127188; RV32I-NEXT: li a1, -128189; RV32I-NEXT: blt a1, a0, .LBB3_2190; RV32I-NEXT: .LBB3_4:191; RV32I-NEXT: li a0, -128192; RV32I-NEXT: ret193;194; RV64I-LABEL: func8:195; RV64I: # %bb.0:196; RV64I-NEXT: slli a0, a0, 56197; RV64I-NEXT: mul a1, a1, a2198; RV64I-NEXT: srai a0, a0, 56199; RV64I-NEXT: slli a1, a1, 56200; RV64I-NEXT: srai a1, a1, 56201; RV64I-NEXT: sub a0, a0, a1202; RV64I-NEXT: li a1, 127203; RV64I-NEXT: bge a0, a1, .LBB3_3204; RV64I-NEXT: # %bb.1:205; RV64I-NEXT: li a1, -128206; RV64I-NEXT: bge a1, a0, .LBB3_4207; RV64I-NEXT: .LBB3_2:208; RV64I-NEXT: ret209; RV64I-NEXT: .LBB3_3:210; RV64I-NEXT: li a0, 127211; RV64I-NEXT: li a1, -128212; RV64I-NEXT: blt a1, a0, .LBB3_2213; RV64I-NEXT: .LBB3_4:214; RV64I-NEXT: li a0, -128215; RV64I-NEXT: ret216;217; RV32IZbb-LABEL: func8:218; RV32IZbb: # %bb.0:219; RV32IZbb-NEXT: sext.b a0, a0220; RV32IZbb-NEXT: mul a1, a1, a2221; RV32IZbb-NEXT: li a2, 127222; RV32IZbb-NEXT: sext.b a1, a1223; RV32IZbb-NEXT: sub a0, a0, a1224; RV32IZbb-NEXT: min a0, a0, a2225; RV32IZbb-NEXT: li a1, -128226; RV32IZbb-NEXT: max a0, a0, a1227; RV32IZbb-NEXT: ret228;229; RV64IZbb-LABEL: func8:230; RV64IZbb: # %bb.0:231; RV64IZbb-NEXT: sext.b a0, a0232; RV64IZbb-NEXT: mul a1, a1, a2233; RV64IZbb-NEXT: li a2, 127234; RV64IZbb-NEXT: sext.b a1, a1235; RV64IZbb-NEXT: sub a0, a0, a1236; RV64IZbb-NEXT: min a0, a0, a2237; RV64IZbb-NEXT: li a1, -128238; RV64IZbb-NEXT: max a0, a0, a1239; RV64IZbb-NEXT: ret240 %a = mul i8 %y, %z241 %tmp = call i8 @llvm.ssub.sat.i8(i8 %x, i8 %a)242 ret i8 %tmp243}244 245define i4 @func4(i4 %x, i4 %y, i4 %z) nounwind {246; RV32I-LABEL: func4:247; RV32I: # %bb.0:248; RV32I-NEXT: slli a0, a0, 28249; RV32I-NEXT: mul a1, a1, a2250; RV32I-NEXT: srai a0, a0, 28251; RV32I-NEXT: slli a1, a1, 28252; RV32I-NEXT: srai a1, a1, 28253; RV32I-NEXT: sub a0, a0, a1254; RV32I-NEXT: li a1, 7255; RV32I-NEXT: bge a0, a1, .LBB4_3256; RV32I-NEXT: # %bb.1:257; RV32I-NEXT: li a1, -8258; RV32I-NEXT: bge a1, a0, .LBB4_4259; RV32I-NEXT: .LBB4_2:260; RV32I-NEXT: ret261; RV32I-NEXT: .LBB4_3:262; RV32I-NEXT: li a0, 7263; RV32I-NEXT: li a1, -8264; RV32I-NEXT: blt a1, a0, .LBB4_2265; RV32I-NEXT: .LBB4_4:266; RV32I-NEXT: li a0, -8267; RV32I-NEXT: ret268;269; RV64I-LABEL: func4:270; RV64I: # %bb.0:271; RV64I-NEXT: slli a0, a0, 60272; RV64I-NEXT: mul a1, a1, a2273; RV64I-NEXT: srai a0, a0, 60274; RV64I-NEXT: slli a1, a1, 60275; RV64I-NEXT: srai a1, a1, 60276; RV64I-NEXT: sub a0, a0, a1277; RV64I-NEXT: li a1, 7278; RV64I-NEXT: bge a0, a1, .LBB4_3279; RV64I-NEXT: # %bb.1:280; RV64I-NEXT: li a1, -8281; RV64I-NEXT: bge a1, a0, .LBB4_4282; RV64I-NEXT: .LBB4_2:283; RV64I-NEXT: ret284; RV64I-NEXT: .LBB4_3:285; RV64I-NEXT: li a0, 7286; RV64I-NEXT: li a1, -8287; RV64I-NEXT: blt a1, a0, .LBB4_2288; RV64I-NEXT: .LBB4_4:289; RV64I-NEXT: li a0, -8290; RV64I-NEXT: ret291;292; RV32IZbb-LABEL: func4:293; RV32IZbb: # %bb.0:294; RV32IZbb-NEXT: slli a0, a0, 28295; RV32IZbb-NEXT: mul a1, a1, a2296; RV32IZbb-NEXT: li a2, 7297; RV32IZbb-NEXT: srai a0, a0, 28298; RV32IZbb-NEXT: slli a1, a1, 28299; RV32IZbb-NEXT: srai a1, a1, 28300; RV32IZbb-NEXT: sub a0, a0, a1301; RV32IZbb-NEXT: min a0, a0, a2302; RV32IZbb-NEXT: li a1, -8303; RV32IZbb-NEXT: max a0, a0, a1304; RV32IZbb-NEXT: ret305;306; RV64IZbb-LABEL: func4:307; RV64IZbb: # %bb.0:308; RV64IZbb-NEXT: slli a0, a0, 60309; RV64IZbb-NEXT: mul a1, a1, a2310; RV64IZbb-NEXT: li a2, 7311; RV64IZbb-NEXT: srai a0, a0, 60312; RV64IZbb-NEXT: slli a1, a1, 60313; RV64IZbb-NEXT: srai a1, a1, 60314; RV64IZbb-NEXT: sub a0, a0, a1315; RV64IZbb-NEXT: min a0, a0, a2316; RV64IZbb-NEXT: li a1, -8317; RV64IZbb-NEXT: max a0, a0, a1318; RV64IZbb-NEXT: ret319 %a = mul i4 %y, %z320 %tmp = call i4 @llvm.ssub.sat.i4(i4 %x, i4 %a)321 ret i4 %tmp322}323