343 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 a1, a1, a212; RV32-NEXT: add a0, a0, a113; RV32-NEXT: slt a2, a0, a314; RV32-NEXT: srli a1, a1, 3115; 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: addw a2, a0, a128; RV64I-NEXT: add 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: add 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.sadd.sat.i32(i32 %x, i32 %a)49 ret i32 %tmp50}51 52define i64 @func64(i64 %x, i64 %y, i64 %z) nounwind {53; RV32I-LABEL: func64:54; RV32I: # %bb.0:55; RV32I-NEXT: mv a2, a156; RV32I-NEXT: mv a1, a057; RV32I-NEXT: add a3, a2, a558; RV32I-NEXT: add a0, a0, a459; RV32I-NEXT: xor a5, a2, a560; RV32I-NEXT: sltu a1, a0, a161; RV32I-NEXT: add a1, a3, a162; RV32I-NEXT: xor a2, a2, a163; RV32I-NEXT: not a3, a564; RV32I-NEXT: and a2, a3, a265; RV32I-NEXT: bgez a2, .LBB1_266; RV32I-NEXT: # %bb.1:67; RV32I-NEXT: srai a0, a1, 3168; RV32I-NEXT: lui a1, 52428869; RV32I-NEXT: xor a1, a0, a170; RV32I-NEXT: .LBB1_2:71; RV32I-NEXT: ret72;73; RV64-LABEL: func64:74; RV64: # %bb.0:75; RV64-NEXT: mv a1, a076; RV64-NEXT: add a0, a0, a277; RV64-NEXT: slt a1, a0, a178; RV64-NEXT: srli a2, a2, 6379; RV64-NEXT: beq a2, a1, .LBB1_280; RV64-NEXT: # %bb.1:81; RV64-NEXT: srai a0, a0, 6382; RV64-NEXT: li a1, -183; RV64-NEXT: slli a1, a1, 6384; RV64-NEXT: xor a0, a0, a185; RV64-NEXT: .LBB1_2:86; RV64-NEXT: ret87;88; RV32IZbb-LABEL: func64:89; RV32IZbb: # %bb.0:90; RV32IZbb-NEXT: mv a2, a191; RV32IZbb-NEXT: mv a1, a092; RV32IZbb-NEXT: add a3, a2, a593; RV32IZbb-NEXT: add a0, a0, a494; RV32IZbb-NEXT: sltu a1, a0, a195; RV32IZbb-NEXT: add a1, a3, a196; RV32IZbb-NEXT: xor a3, a2, a197; RV32IZbb-NEXT: xor a2, a2, a598; RV32IZbb-NEXT: andn a2, a3, a299; RV32IZbb-NEXT: bgez a2, .LBB1_2100; RV32IZbb-NEXT: # %bb.1:101; RV32IZbb-NEXT: srai a0, a1, 31102; RV32IZbb-NEXT: lui a1, 524288103; RV32IZbb-NEXT: xor a1, a0, a1104; RV32IZbb-NEXT: .LBB1_2:105; RV32IZbb-NEXT: ret106 %a = mul i64 %y, %z107 %tmp = call i64 @llvm.sadd.sat.i64(i64 %x, i64 %z)108 ret i64 %tmp109}110 111define i16 @func16(i16 %x, i16 %y, i16 %z) nounwind {112; RV32I-LABEL: func16:113; RV32I: # %bb.0:114; RV32I-NEXT: slli a0, a0, 16115; RV32I-NEXT: mul a1, a1, a2116; RV32I-NEXT: lui a2, 8117; RV32I-NEXT: srai a0, a0, 16118; RV32I-NEXT: slli a1, a1, 16119; RV32I-NEXT: srai a1, a1, 16120; RV32I-NEXT: add a0, a0, a1121; RV32I-NEXT: addi a1, a2, -1122; RV32I-NEXT: bge a0, a1, .LBB2_3123; RV32I-NEXT: # %bb.1:124; RV32I-NEXT: lui a1, 1048568125; RV32I-NEXT: bge a1, a0, .LBB2_4126; RV32I-NEXT: .LBB2_2:127; RV32I-NEXT: ret128; RV32I-NEXT: .LBB2_3:129; RV32I-NEXT: mv a0, a1130; RV32I-NEXT: lui a1, 1048568131; RV32I-NEXT: blt a1, a0, .LBB2_2132; RV32I-NEXT: .LBB2_4:133; RV32I-NEXT: lui a0, 1048568134; RV32I-NEXT: ret135;136; RV64I-LABEL: func16:137; RV64I: # %bb.0:138; RV64I-NEXT: slli a0, a0, 48139; RV64I-NEXT: mul a1, a1, a2140; RV64I-NEXT: lui a2, 8141; RV64I-NEXT: srai a0, a0, 48142; RV64I-NEXT: slli a1, a1, 48143; RV64I-NEXT: srai a1, a1, 48144; RV64I-NEXT: add a0, a0, a1145; RV64I-NEXT: addi a1, a2, -1146; RV64I-NEXT: bge a0, a1, .LBB2_3147; RV64I-NEXT: # %bb.1:148; RV64I-NEXT: lui a1, 1048568149; RV64I-NEXT: bge a1, a0, .LBB2_4150; RV64I-NEXT: .LBB2_2:151; RV64I-NEXT: ret152; RV64I-NEXT: .LBB2_3:153; RV64I-NEXT: mv a0, a1154; RV64I-NEXT: lui a1, 1048568155; RV64I-NEXT: blt a1, a0, .LBB2_2156; RV64I-NEXT: .LBB2_4:157; RV64I-NEXT: lui a0, 1048568158; RV64I-NEXT: ret159;160; RV32IZbb-LABEL: func16:161; RV32IZbb: # %bb.0:162; RV32IZbb-NEXT: sext.h a0, a0163; RV32IZbb-NEXT: mul a1, a1, a2164; RV32IZbb-NEXT: lui a2, 8165; RV32IZbb-NEXT: sext.h a1, a1166; RV32IZbb-NEXT: addi a2, a2, -1167; RV32IZbb-NEXT: add a0, a0, a1168; RV32IZbb-NEXT: min a0, a0, a2169; RV32IZbb-NEXT: lui a1, 1048568170; RV32IZbb-NEXT: max a0, a0, a1171; RV32IZbb-NEXT: ret172;173; RV64IZbb-LABEL: func16:174; RV64IZbb: # %bb.0:175; RV64IZbb-NEXT: sext.h a0, a0176; RV64IZbb-NEXT: mul a1, a1, a2177; RV64IZbb-NEXT: lui a2, 8178; RV64IZbb-NEXT: sext.h a1, a1179; RV64IZbb-NEXT: addi a2, a2, -1180; RV64IZbb-NEXT: add a0, a0, a1181; RV64IZbb-NEXT: min a0, a0, a2182; RV64IZbb-NEXT: lui a1, 1048568183; RV64IZbb-NEXT: max a0, a0, a1184; RV64IZbb-NEXT: ret185 %a = mul i16 %y, %z186 %tmp = call i16 @llvm.sadd.sat.i16(i16 %x, i16 %a)187 ret i16 %tmp188}189 190define i8 @func8(i8 %x, i8 %y, i8 %z) nounwind {191; RV32I-LABEL: func8:192; RV32I: # %bb.0:193; RV32I-NEXT: slli a0, a0, 24194; RV32I-NEXT: mul a1, a1, a2195; RV32I-NEXT: srai a0, a0, 24196; RV32I-NEXT: slli a1, a1, 24197; RV32I-NEXT: srai a1, a1, 24198; RV32I-NEXT: add a0, a0, a1199; RV32I-NEXT: li a1, 127200; RV32I-NEXT: bge a0, a1, .LBB3_3201; RV32I-NEXT: # %bb.1:202; RV32I-NEXT: li a1, -128203; RV32I-NEXT: bge a1, a0, .LBB3_4204; RV32I-NEXT: .LBB3_2:205; RV32I-NEXT: ret206; RV32I-NEXT: .LBB3_3:207; RV32I-NEXT: li a0, 127208; RV32I-NEXT: li a1, -128209; RV32I-NEXT: blt a1, a0, .LBB3_2210; RV32I-NEXT: .LBB3_4:211; RV32I-NEXT: li a0, -128212; RV32I-NEXT: ret213;214; RV64I-LABEL: func8:215; RV64I: # %bb.0:216; RV64I-NEXT: slli a0, a0, 56217; RV64I-NEXT: mul a1, a1, a2218; RV64I-NEXT: srai a0, a0, 56219; RV64I-NEXT: slli a1, a1, 56220; RV64I-NEXT: srai a1, a1, 56221; RV64I-NEXT: add a0, a0, a1222; RV64I-NEXT: li a1, 127223; RV64I-NEXT: bge a0, a1, .LBB3_3224; RV64I-NEXT: # %bb.1:225; RV64I-NEXT: li a1, -128226; RV64I-NEXT: bge a1, a0, .LBB3_4227; RV64I-NEXT: .LBB3_2:228; RV64I-NEXT: ret229; RV64I-NEXT: .LBB3_3:230; RV64I-NEXT: li a0, 127231; RV64I-NEXT: li a1, -128232; RV64I-NEXT: blt a1, a0, .LBB3_2233; RV64I-NEXT: .LBB3_4:234; RV64I-NEXT: li a0, -128235; RV64I-NEXT: ret236;237; RV32IZbb-LABEL: func8:238; RV32IZbb: # %bb.0:239; RV32IZbb-NEXT: sext.b a0, a0240; RV32IZbb-NEXT: mul a1, a1, a2241; RV32IZbb-NEXT: li a2, 127242; RV32IZbb-NEXT: sext.b a1, a1243; RV32IZbb-NEXT: add a0, a0, a1244; RV32IZbb-NEXT: min a0, a0, a2245; RV32IZbb-NEXT: li a1, -128246; RV32IZbb-NEXT: max a0, a0, a1247; RV32IZbb-NEXT: ret248;249; RV64IZbb-LABEL: func8:250; RV64IZbb: # %bb.0:251; RV64IZbb-NEXT: sext.b a0, a0252; RV64IZbb-NEXT: mul a1, a1, a2253; RV64IZbb-NEXT: li a2, 127254; RV64IZbb-NEXT: sext.b a1, a1255; RV64IZbb-NEXT: add a0, a0, a1256; RV64IZbb-NEXT: min a0, a0, a2257; RV64IZbb-NEXT: li a1, -128258; RV64IZbb-NEXT: max a0, a0, a1259; RV64IZbb-NEXT: ret260 %a = mul i8 %y, %z261 %tmp = call i8 @llvm.sadd.sat.i8(i8 %x, i8 %a)262 ret i8 %tmp263}264 265define i4 @func4(i4 %x, i4 %y, i4 %z) nounwind {266; RV32I-LABEL: func4:267; RV32I: # %bb.0:268; RV32I-NEXT: slli a0, a0, 28269; RV32I-NEXT: mul a1, a1, a2270; RV32I-NEXT: srai a0, a0, 28271; RV32I-NEXT: slli a1, a1, 28272; RV32I-NEXT: srai a1, a1, 28273; RV32I-NEXT: add a0, a0, a1274; RV32I-NEXT: li a1, 7275; RV32I-NEXT: bge a0, a1, .LBB4_3276; RV32I-NEXT: # %bb.1:277; RV32I-NEXT: li a1, -8278; RV32I-NEXT: bge a1, a0, .LBB4_4279; RV32I-NEXT: .LBB4_2:280; RV32I-NEXT: ret281; RV32I-NEXT: .LBB4_3:282; RV32I-NEXT: li a0, 7283; RV32I-NEXT: li a1, -8284; RV32I-NEXT: blt a1, a0, .LBB4_2285; RV32I-NEXT: .LBB4_4:286; RV32I-NEXT: li a0, -8287; RV32I-NEXT: ret288;289; RV64I-LABEL: func4:290; RV64I: # %bb.0:291; RV64I-NEXT: slli a0, a0, 60292; RV64I-NEXT: mul a1, a1, a2293; RV64I-NEXT: srai a0, a0, 60294; RV64I-NEXT: slli a1, a1, 60295; RV64I-NEXT: srai a1, a1, 60296; RV64I-NEXT: add a0, a0, a1297; RV64I-NEXT: li a1, 7298; RV64I-NEXT: bge a0, a1, .LBB4_3299; RV64I-NEXT: # %bb.1:300; RV64I-NEXT: li a1, -8301; RV64I-NEXT: bge a1, a0, .LBB4_4302; RV64I-NEXT: .LBB4_2:303; RV64I-NEXT: ret304; RV64I-NEXT: .LBB4_3:305; RV64I-NEXT: li a0, 7306; RV64I-NEXT: li a1, -8307; RV64I-NEXT: blt a1, a0, .LBB4_2308; RV64I-NEXT: .LBB4_4:309; RV64I-NEXT: li a0, -8310; RV64I-NEXT: ret311;312; RV32IZbb-LABEL: func4:313; RV32IZbb: # %bb.0:314; RV32IZbb-NEXT: slli a0, a0, 28315; RV32IZbb-NEXT: mul a1, a1, a2316; RV32IZbb-NEXT: li a2, 7317; RV32IZbb-NEXT: srai a0, a0, 28318; RV32IZbb-NEXT: slli a1, a1, 28319; RV32IZbb-NEXT: srai a1, a1, 28320; RV32IZbb-NEXT: add a0, a0, a1321; RV32IZbb-NEXT: min a0, a0, a2322; RV32IZbb-NEXT: li a1, -8323; RV32IZbb-NEXT: max a0, a0, a1324; RV32IZbb-NEXT: ret325;326; RV64IZbb-LABEL: func4:327; RV64IZbb: # %bb.0:328; RV64IZbb-NEXT: slli a0, a0, 60329; RV64IZbb-NEXT: mul a1, a1, a2330; RV64IZbb-NEXT: li a2, 7331; RV64IZbb-NEXT: srai a0, a0, 60332; RV64IZbb-NEXT: slli a1, a1, 60333; RV64IZbb-NEXT: srai a1, a1, 60334; RV64IZbb-NEXT: add a0, a0, a1335; RV64IZbb-NEXT: min a0, a0, a2336; RV64IZbb-NEXT: li a1, -8337; RV64IZbb-NEXT: max a0, a0, a1338; RV64IZbb-NEXT: ret339 %a = mul i4 %y, %z340 %tmp = call i4 @llvm.sadd.sat.i4(i4 %x, i4 %a)341 ret i4 %tmp342}343