763 lines · plain
1; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py2; RUN: llc < %s -mtriple=riscv32 \3; RUN: | FileCheck %s --check-prefixes=RV32-BOTH,RV324; RUN: llc < %s -mtriple=riscv64 \5; RUN: | FileCheck %s --check-prefixes=RV64-BOTH,RV646; RUN: llc < %s -mtriple=riscv32 -mattr=+unaligned-scalar-mem \7; RUN: | FileCheck %s --check-prefixes=RV32-BOTH,RV32-FAST8; RUN: llc < %s -mtriple=riscv64 -mattr=+unaligned-scalar-mem \9; RUN: | FileCheck %s --check-prefixes=RV64-BOTH,RV64-FAST10 11; ----------------------------------------------------------------------12; Fully unaligned cases13 14define void @unaligned_memcpy0(ptr nocapture %dest, ptr %src) nounwind {15; RV32-BOTH-LABEL: unaligned_memcpy0:16; RV32-BOTH: # %bb.0: # %entry17; RV32-BOTH-NEXT: ret18;19; RV64-BOTH-LABEL: unaligned_memcpy0:20; RV64-BOTH: # %bb.0: # %entry21; RV64-BOTH-NEXT: ret22entry:23 tail call void @llvm.memcpy.p0.p0.i64(ptr %dest, ptr %src, i64 0, i1 false)24 ret void25}26 27define void @unaligned_memcpy1(ptr nocapture %dest, ptr %src) nounwind {28; RV32-BOTH-LABEL: unaligned_memcpy1:29; RV32-BOTH: # %bb.0: # %entry30; RV32-BOTH-NEXT: lbu a1, 0(a1)31; RV32-BOTH-NEXT: sb a1, 0(a0)32; RV32-BOTH-NEXT: ret33;34; RV64-BOTH-LABEL: unaligned_memcpy1:35; RV64-BOTH: # %bb.0: # %entry36; RV64-BOTH-NEXT: lbu a1, 0(a1)37; RV64-BOTH-NEXT: sb a1, 0(a0)38; RV64-BOTH-NEXT: ret39entry:40 tail call void @llvm.memcpy.p0.p0.i64(ptr %dest, ptr %src, i64 1, i1 false)41 ret void42}43 44define void @unaligned_memcpy2(ptr nocapture %dest, ptr %src) nounwind {45; RV32-LABEL: unaligned_memcpy2:46; RV32: # %bb.0: # %entry47; RV32-NEXT: lbu a2, 1(a1)48; RV32-NEXT: sb a2, 1(a0)49; RV32-NEXT: lbu a1, 0(a1)50; RV32-NEXT: sb a1, 0(a0)51; RV32-NEXT: ret52;53; RV64-LABEL: unaligned_memcpy2:54; RV64: # %bb.0: # %entry55; RV64-NEXT: lbu a2, 1(a1)56; RV64-NEXT: sb a2, 1(a0)57; RV64-NEXT: lbu a1, 0(a1)58; RV64-NEXT: sb a1, 0(a0)59; RV64-NEXT: ret60;61; RV32-FAST-LABEL: unaligned_memcpy2:62; RV32-FAST: # %bb.0: # %entry63; RV32-FAST-NEXT: lh a1, 0(a1)64; RV32-FAST-NEXT: sh a1, 0(a0)65; RV32-FAST-NEXT: ret66;67; RV64-FAST-LABEL: unaligned_memcpy2:68; RV64-FAST: # %bb.0: # %entry69; RV64-FAST-NEXT: lh a1, 0(a1)70; RV64-FAST-NEXT: sh a1, 0(a0)71; RV64-FAST-NEXT: ret72entry:73 tail call void @llvm.memcpy.p0.p0.i64(ptr %dest, ptr %src, i64 2, i1 false)74 ret void75}76 77define void @unaligned_memcpy3(ptr nocapture %dest, ptr %src) nounwind {78; RV32-LABEL: unaligned_memcpy3:79; RV32: # %bb.0: # %entry80; RV32-NEXT: lbu a2, 2(a1)81; RV32-NEXT: sb a2, 2(a0)82; RV32-NEXT: lbu a2, 1(a1)83; RV32-NEXT: sb a2, 1(a0)84; RV32-NEXT: lbu a1, 0(a1)85; RV32-NEXT: sb a1, 0(a0)86; RV32-NEXT: ret87;88; RV64-LABEL: unaligned_memcpy3:89; RV64: # %bb.0: # %entry90; RV64-NEXT: lbu a2, 2(a1)91; RV64-NEXT: sb a2, 2(a0)92; RV64-NEXT: lbu a2, 1(a1)93; RV64-NEXT: sb a2, 1(a0)94; RV64-NEXT: lbu a1, 0(a1)95; RV64-NEXT: sb a1, 0(a0)96; RV64-NEXT: ret97;98; RV32-FAST-LABEL: unaligned_memcpy3:99; RV32-FAST: # %bb.0: # %entry100; RV32-FAST-NEXT: lbu a2, 2(a1)101; RV32-FAST-NEXT: sb a2, 2(a0)102; RV32-FAST-NEXT: lh a1, 0(a1)103; RV32-FAST-NEXT: sh a1, 0(a0)104; RV32-FAST-NEXT: ret105;106; RV64-FAST-LABEL: unaligned_memcpy3:107; RV64-FAST: # %bb.0: # %entry108; RV64-FAST-NEXT: lbu a2, 2(a1)109; RV64-FAST-NEXT: sb a2, 2(a0)110; RV64-FAST-NEXT: lh a1, 0(a1)111; RV64-FAST-NEXT: sh a1, 0(a0)112; RV64-FAST-NEXT: ret113entry:114 tail call void @llvm.memcpy.p0.p0.i64(ptr %dest, ptr %src, i64 3, i1 false)115 ret void116}117 118define void @unaligned_memcpy4(ptr nocapture %dest, ptr %src) nounwind {119; RV32-LABEL: unaligned_memcpy4:120; RV32: # %bb.0: # %entry121; RV32-NEXT: lbu a2, 3(a1)122; RV32-NEXT: sb a2, 3(a0)123; RV32-NEXT: lbu a2, 2(a1)124; RV32-NEXT: sb a2, 2(a0)125; RV32-NEXT: lbu a2, 1(a1)126; RV32-NEXT: sb a2, 1(a0)127; RV32-NEXT: lbu a1, 0(a1)128; RV32-NEXT: sb a1, 0(a0)129; RV32-NEXT: ret130;131; RV64-LABEL: unaligned_memcpy4:132; RV64: # %bb.0: # %entry133; RV64-NEXT: lbu a2, 3(a1)134; RV64-NEXT: sb a2, 3(a0)135; RV64-NEXT: lbu a2, 2(a1)136; RV64-NEXT: sb a2, 2(a0)137; RV64-NEXT: lbu a2, 1(a1)138; RV64-NEXT: sb a2, 1(a0)139; RV64-NEXT: lbu a1, 0(a1)140; RV64-NEXT: sb a1, 0(a0)141; RV64-NEXT: ret142;143; RV32-FAST-LABEL: unaligned_memcpy4:144; RV32-FAST: # %bb.0: # %entry145; RV32-FAST-NEXT: lw a1, 0(a1)146; RV32-FAST-NEXT: sw a1, 0(a0)147; RV32-FAST-NEXT: ret148;149; RV64-FAST-LABEL: unaligned_memcpy4:150; RV64-FAST: # %bb.0: # %entry151; RV64-FAST-NEXT: lw a1, 0(a1)152; RV64-FAST-NEXT: sw a1, 0(a0)153; RV64-FAST-NEXT: ret154entry:155 tail call void @llvm.memcpy.p0.p0.i64(ptr %dest, ptr %src, i64 4, i1 false)156 ret void157}158 159define void @unaligned_memcpy7(ptr nocapture %dest, ptr %src) nounwind {160; RV32-LABEL: unaligned_memcpy7:161; RV32: # %bb.0: # %entry162; RV32-NEXT: lbu a2, 6(a1)163; RV32-NEXT: sb a2, 6(a0)164; RV32-NEXT: lbu a2, 5(a1)165; RV32-NEXT: sb a2, 5(a0)166; RV32-NEXT: lbu a2, 4(a1)167; RV32-NEXT: sb a2, 4(a0)168; RV32-NEXT: lbu a2, 3(a1)169; RV32-NEXT: sb a2, 3(a0)170; RV32-NEXT: lbu a2, 2(a1)171; RV32-NEXT: sb a2, 2(a0)172; RV32-NEXT: lbu a2, 1(a1)173; RV32-NEXT: sb a2, 1(a0)174; RV32-NEXT: lbu a1, 0(a1)175; RV32-NEXT: sb a1, 0(a0)176; RV32-NEXT: ret177;178; RV64-LABEL: unaligned_memcpy7:179; RV64: # %bb.0: # %entry180; RV64-NEXT: lbu a2, 6(a1)181; RV64-NEXT: sb a2, 6(a0)182; RV64-NEXT: lbu a2, 5(a1)183; RV64-NEXT: sb a2, 5(a0)184; RV64-NEXT: lbu a2, 4(a1)185; RV64-NEXT: sb a2, 4(a0)186; RV64-NEXT: lbu a2, 3(a1)187; RV64-NEXT: sb a2, 3(a0)188; RV64-NEXT: lbu a2, 2(a1)189; RV64-NEXT: sb a2, 2(a0)190; RV64-NEXT: lbu a2, 1(a1)191; RV64-NEXT: sb a2, 1(a0)192; RV64-NEXT: lbu a1, 0(a1)193; RV64-NEXT: sb a1, 0(a0)194; RV64-NEXT: ret195;196; RV32-FAST-LABEL: unaligned_memcpy7:197; RV32-FAST: # %bb.0: # %entry198; RV32-FAST-NEXT: lw a2, 3(a1)199; RV32-FAST-NEXT: sw a2, 3(a0)200; RV32-FAST-NEXT: lw a1, 0(a1)201; RV32-FAST-NEXT: sw a1, 0(a0)202; RV32-FAST-NEXT: ret203;204; RV64-FAST-LABEL: unaligned_memcpy7:205; RV64-FAST: # %bb.0: # %entry206; RV64-FAST-NEXT: lw a2, 3(a1)207; RV64-FAST-NEXT: sw a2, 3(a0)208; RV64-FAST-NEXT: lw a1, 0(a1)209; RV64-FAST-NEXT: sw a1, 0(a0)210; RV64-FAST-NEXT: ret211entry:212 tail call void @llvm.memcpy.p0.p0.i64(ptr %dest, ptr %src, i64 7, i1 false)213 ret void214}215 216define void @unaligned_memcpy8(ptr nocapture %dest, ptr %src) nounwind {217; RV32-LABEL: unaligned_memcpy8:218; RV32: # %bb.0: # %entry219; RV32-NEXT: lbu a2, 7(a1)220; RV32-NEXT: sb a2, 7(a0)221; RV32-NEXT: lbu a2, 6(a1)222; RV32-NEXT: sb a2, 6(a0)223; RV32-NEXT: lbu a2, 5(a1)224; RV32-NEXT: sb a2, 5(a0)225; RV32-NEXT: lbu a2, 4(a1)226; RV32-NEXT: sb a2, 4(a0)227; RV32-NEXT: lbu a2, 3(a1)228; RV32-NEXT: sb a2, 3(a0)229; RV32-NEXT: lbu a2, 2(a1)230; RV32-NEXT: sb a2, 2(a0)231; RV32-NEXT: lbu a2, 1(a1)232; RV32-NEXT: sb a2, 1(a0)233; RV32-NEXT: lbu a1, 0(a1)234; RV32-NEXT: sb a1, 0(a0)235; RV32-NEXT: ret236;237; RV64-LABEL: unaligned_memcpy8:238; RV64: # %bb.0: # %entry239; RV64-NEXT: lbu a2, 7(a1)240; RV64-NEXT: sb a2, 7(a0)241; RV64-NEXT: lbu a2, 6(a1)242; RV64-NEXT: sb a2, 6(a0)243; RV64-NEXT: lbu a2, 5(a1)244; RV64-NEXT: sb a2, 5(a0)245; RV64-NEXT: lbu a2, 4(a1)246; RV64-NEXT: sb a2, 4(a0)247; RV64-NEXT: lbu a2, 3(a1)248; RV64-NEXT: sb a2, 3(a0)249; RV64-NEXT: lbu a2, 2(a1)250; RV64-NEXT: sb a2, 2(a0)251; RV64-NEXT: lbu a2, 1(a1)252; RV64-NEXT: sb a2, 1(a0)253; RV64-NEXT: lbu a1, 0(a1)254; RV64-NEXT: sb a1, 0(a0)255; RV64-NEXT: ret256;257; RV32-FAST-LABEL: unaligned_memcpy8:258; RV32-FAST: # %bb.0: # %entry259; RV32-FAST-NEXT: lw a2, 4(a1)260; RV32-FAST-NEXT: sw a2, 4(a0)261; RV32-FAST-NEXT: lw a1, 0(a1)262; RV32-FAST-NEXT: sw a1, 0(a0)263; RV32-FAST-NEXT: ret264;265; RV64-FAST-LABEL: unaligned_memcpy8:266; RV64-FAST: # %bb.0: # %entry267; RV64-FAST-NEXT: ld a1, 0(a1)268; RV64-FAST-NEXT: sd a1, 0(a0)269; RV64-FAST-NEXT: ret270entry:271 tail call void @llvm.memcpy.p0.p0.i64(ptr %dest, ptr %src, i64 8, i1 false)272 ret void273}274 275define void @unaligned_memcpy15(ptr nocapture %dest, ptr %src) nounwind {276; RV32-LABEL: unaligned_memcpy15:277; RV32: # %bb.0: # %entry278; RV32-NEXT: li a2, 15279; RV32-NEXT: tail memcpy280;281; RV64-LABEL: unaligned_memcpy15:282; RV64: # %bb.0: # %entry283; RV64-NEXT: li a2, 15284; RV64-NEXT: tail memcpy285;286; RV32-FAST-LABEL: unaligned_memcpy15:287; RV32-FAST: # %bb.0: # %entry288; RV32-FAST-NEXT: lw a2, 11(a1)289; RV32-FAST-NEXT: sw a2, 11(a0)290; RV32-FAST-NEXT: lw a2, 8(a1)291; RV32-FAST-NEXT: sw a2, 8(a0)292; RV32-FAST-NEXT: lw a2, 4(a1)293; RV32-FAST-NEXT: sw a2, 4(a0)294; RV32-FAST-NEXT: lw a1, 0(a1)295; RV32-FAST-NEXT: sw a1, 0(a0)296; RV32-FAST-NEXT: ret297;298; RV64-FAST-LABEL: unaligned_memcpy15:299; RV64-FAST: # %bb.0: # %entry300; RV64-FAST-NEXT: ld a2, 7(a1)301; RV64-FAST-NEXT: sd a2, 7(a0)302; RV64-FAST-NEXT: ld a1, 0(a1)303; RV64-FAST-NEXT: sd a1, 0(a0)304; RV64-FAST-NEXT: ret305entry:306 tail call void @llvm.memcpy.p0.p0.i64(ptr %dest, ptr %src, i64 15, i1 false)307 ret void308}309 310define void @unaligned_memcpy16(ptr nocapture %dest, ptr %src) nounwind {311; RV32-LABEL: unaligned_memcpy16:312; RV32: # %bb.0: # %entry313; RV32-NEXT: li a2, 16314; RV32-NEXT: tail memcpy315;316; RV64-LABEL: unaligned_memcpy16:317; RV64: # %bb.0: # %entry318; RV64-NEXT: li a2, 16319; RV64-NEXT: tail memcpy320;321; RV32-FAST-LABEL: unaligned_memcpy16:322; RV32-FAST: # %bb.0: # %entry323; RV32-FAST-NEXT: lw a2, 12(a1)324; RV32-FAST-NEXT: sw a2, 12(a0)325; RV32-FAST-NEXT: lw a2, 8(a1)326; RV32-FAST-NEXT: sw a2, 8(a0)327; RV32-FAST-NEXT: lw a2, 4(a1)328; RV32-FAST-NEXT: sw a2, 4(a0)329; RV32-FAST-NEXT: lw a1, 0(a1)330; RV32-FAST-NEXT: sw a1, 0(a0)331; RV32-FAST-NEXT: ret332;333; RV64-FAST-LABEL: unaligned_memcpy16:334; RV64-FAST: # %bb.0: # %entry335; RV64-FAST-NEXT: ld a2, 8(a1)336; RV64-FAST-NEXT: sd a2, 8(a0)337; RV64-FAST-NEXT: ld a1, 0(a1)338; RV64-FAST-NEXT: sd a1, 0(a0)339; RV64-FAST-NEXT: ret340entry:341 tail call void @llvm.memcpy.p0.p0.i64(ptr %dest, ptr %src, i64 16, i1 false)342 ret void343}344 345define void @unaligned_memcpy31(ptr nocapture %dest, ptr %src) nounwind {346; RV32-LABEL: unaligned_memcpy31:347; RV32: # %bb.0: # %entry348; RV32-NEXT: li a2, 31349; RV32-NEXT: tail memcpy350;351; RV64-LABEL: unaligned_memcpy31:352; RV64: # %bb.0: # %entry353; RV64-NEXT: li a2, 31354; RV64-NEXT: tail memcpy355;356; RV32-FAST-LABEL: unaligned_memcpy31:357; RV32-FAST: # %bb.0: # %entry358; RV32-FAST-NEXT: lw a2, 27(a1)359; RV32-FAST-NEXT: sw a2, 27(a0)360; RV32-FAST-NEXT: lw a2, 24(a1)361; RV32-FAST-NEXT: sw a2, 24(a0)362; RV32-FAST-NEXT: lw a2, 20(a1)363; RV32-FAST-NEXT: sw a2, 20(a0)364; RV32-FAST-NEXT: lw a2, 16(a1)365; RV32-FAST-NEXT: sw a2, 16(a0)366; RV32-FAST-NEXT: lw a2, 12(a1)367; RV32-FAST-NEXT: sw a2, 12(a0)368; RV32-FAST-NEXT: lw a2, 8(a1)369; RV32-FAST-NEXT: sw a2, 8(a0)370; RV32-FAST-NEXT: lw a2, 4(a1)371; RV32-FAST-NEXT: sw a2, 4(a0)372; RV32-FAST-NEXT: lw a1, 0(a1)373; RV32-FAST-NEXT: sw a1, 0(a0)374; RV32-FAST-NEXT: ret375;376; RV64-FAST-LABEL: unaligned_memcpy31:377; RV64-FAST: # %bb.0: # %entry378; RV64-FAST-NEXT: ld a2, 23(a1)379; RV64-FAST-NEXT: sd a2, 23(a0)380; RV64-FAST-NEXT: ld a2, 16(a1)381; RV64-FAST-NEXT: sd a2, 16(a0)382; RV64-FAST-NEXT: ld a2, 8(a1)383; RV64-FAST-NEXT: sd a2, 8(a0)384; RV64-FAST-NEXT: ld a1, 0(a1)385; RV64-FAST-NEXT: sd a1, 0(a0)386; RV64-FAST-NEXT: ret387entry:388 tail call void @llvm.memcpy.p0.p0.i64(ptr %dest, ptr %src, i64 31, i1 false)389 ret void390}391 392; ----------------------------------------------------------------------393; Fully aligned cases394 395define void @aligned_memcpy0(ptr nocapture %dest, ptr %src) nounwind {396; RV32-BOTH-LABEL: aligned_memcpy0:397; RV32-BOTH: # %bb.0: # %entry398; RV32-BOTH-NEXT: ret399;400; RV64-BOTH-LABEL: aligned_memcpy0:401; RV64-BOTH: # %bb.0: # %entry402; RV64-BOTH-NEXT: ret403entry:404 tail call void @llvm.memcpy.p0.p0.i64(ptr align 8 %dest, ptr align 8 %src, i64 0, i1 false)405 ret void406}407 408define void @aligned_memcpy1(ptr nocapture %dest, ptr %src) nounwind {409; RV32-BOTH-LABEL: aligned_memcpy1:410; RV32-BOTH: # %bb.0: # %entry411; RV32-BOTH-NEXT: lbu a1, 0(a1)412; RV32-BOTH-NEXT: sb a1, 0(a0)413; RV32-BOTH-NEXT: ret414;415; RV64-BOTH-LABEL: aligned_memcpy1:416; RV64-BOTH: # %bb.0: # %entry417; RV64-BOTH-NEXT: lbu a1, 0(a1)418; RV64-BOTH-NEXT: sb a1, 0(a0)419; RV64-BOTH-NEXT: ret420entry:421 tail call void @llvm.memcpy.p0.p0.i64(ptr align 8 %dest, ptr align 8 %src, i64 1, i1 false)422 ret void423}424 425define void @aligned_memcpy2(ptr nocapture %dest, ptr %src) nounwind {426; RV32-BOTH-LABEL: aligned_memcpy2:427; RV32-BOTH: # %bb.0: # %entry428; RV32-BOTH-NEXT: lh a1, 0(a1)429; RV32-BOTH-NEXT: sh a1, 0(a0)430; RV32-BOTH-NEXT: ret431;432; RV64-BOTH-LABEL: aligned_memcpy2:433; RV64-BOTH: # %bb.0: # %entry434; RV64-BOTH-NEXT: lh a1, 0(a1)435; RV64-BOTH-NEXT: sh a1, 0(a0)436; RV64-BOTH-NEXT: ret437entry:438 tail call void @llvm.memcpy.p0.p0.i64(ptr align 8 %dest, ptr align 8 %src, i64 2, i1 false)439 ret void440}441 442define void @aligned_memcpy3(ptr nocapture %dest, ptr %src) nounwind {443; RV32-BOTH-LABEL: aligned_memcpy3:444; RV32-BOTH: # %bb.0: # %entry445; RV32-BOTH-NEXT: lbu a2, 2(a1)446; RV32-BOTH-NEXT: sb a2, 2(a0)447; RV32-BOTH-NEXT: lh a1, 0(a1)448; RV32-BOTH-NEXT: sh a1, 0(a0)449; RV32-BOTH-NEXT: ret450;451; RV64-BOTH-LABEL: aligned_memcpy3:452; RV64-BOTH: # %bb.0: # %entry453; RV64-BOTH-NEXT: lbu a2, 2(a1)454; RV64-BOTH-NEXT: sb a2, 2(a0)455; RV64-BOTH-NEXT: lh a1, 0(a1)456; RV64-BOTH-NEXT: sh a1, 0(a0)457; RV64-BOTH-NEXT: ret458entry:459 tail call void @llvm.memcpy.p0.p0.i64(ptr align 8 %dest, ptr align 8 %src, i64 3, i1 false)460 ret void461}462 463define void @aligned_memcpy4(ptr nocapture %dest, ptr %src) nounwind {464; RV32-BOTH-LABEL: aligned_memcpy4:465; RV32-BOTH: # %bb.0: # %entry466; RV32-BOTH-NEXT: lw a1, 0(a1)467; RV32-BOTH-NEXT: sw a1, 0(a0)468; RV32-BOTH-NEXT: ret469;470; RV64-BOTH-LABEL: aligned_memcpy4:471; RV64-BOTH: # %bb.0: # %entry472; RV64-BOTH-NEXT: lw a1, 0(a1)473; RV64-BOTH-NEXT: sw a1, 0(a0)474; RV64-BOTH-NEXT: ret475entry:476 tail call void @llvm.memcpy.p0.p0.i64(ptr align 8 %dest, ptr align 8 %src, i64 4, i1 false)477 ret void478}479 480define void @aligned_memcpy7(ptr nocapture %dest, ptr %src) nounwind {481; RV32-LABEL: aligned_memcpy7:482; RV32: # %bb.0: # %entry483; RV32-NEXT: lbu a2, 6(a1)484; RV32-NEXT: sb a2, 6(a0)485; RV32-NEXT: lh a2, 4(a1)486; RV32-NEXT: sh a2, 4(a0)487; RV32-NEXT: lw a1, 0(a1)488; RV32-NEXT: sw a1, 0(a0)489; RV32-NEXT: ret490;491; RV64-LABEL: aligned_memcpy7:492; RV64: # %bb.0: # %entry493; RV64-NEXT: lbu a2, 6(a1)494; RV64-NEXT: sb a2, 6(a0)495; RV64-NEXT: lh a2, 4(a1)496; RV64-NEXT: sh a2, 4(a0)497; RV64-NEXT: lw a1, 0(a1)498; RV64-NEXT: sw a1, 0(a0)499; RV64-NEXT: ret500;501; RV32-FAST-LABEL: aligned_memcpy7:502; RV32-FAST: # %bb.0: # %entry503; RV32-FAST-NEXT: lw a2, 3(a1)504; RV32-FAST-NEXT: sw a2, 3(a0)505; RV32-FAST-NEXT: lw a1, 0(a1)506; RV32-FAST-NEXT: sw a1, 0(a0)507; RV32-FAST-NEXT: ret508;509; RV64-FAST-LABEL: aligned_memcpy7:510; RV64-FAST: # %bb.0: # %entry511; RV64-FAST-NEXT: lw a2, 3(a1)512; RV64-FAST-NEXT: sw a2, 3(a0)513; RV64-FAST-NEXT: lw a1, 0(a1)514; RV64-FAST-NEXT: sw a1, 0(a0)515; RV64-FAST-NEXT: ret516entry:517 tail call void @llvm.memcpy.p0.p0.i64(ptr align 8 %dest, ptr align 8 %src, i64 7, i1 false)518 ret void519}520 521define void @aligned_memcpy8(ptr nocapture %dest, ptr %src) nounwind {522; RV32-BOTH-LABEL: aligned_memcpy8:523; RV32-BOTH: # %bb.0: # %entry524; RV32-BOTH-NEXT: lw a2, 4(a1)525; RV32-BOTH-NEXT: sw a2, 4(a0)526; RV32-BOTH-NEXT: lw a1, 0(a1)527; RV32-BOTH-NEXT: sw a1, 0(a0)528; RV32-BOTH-NEXT: ret529;530; RV64-BOTH-LABEL: aligned_memcpy8:531; RV64-BOTH: # %bb.0: # %entry532; RV64-BOTH-NEXT: ld a1, 0(a1)533; RV64-BOTH-NEXT: sd a1, 0(a0)534; RV64-BOTH-NEXT: ret535entry:536 tail call void @llvm.memcpy.p0.p0.i64(ptr align 8 %dest, ptr align 8 %src, i64 8, i1 false)537 ret void538}539 540define void @aligned_memcpy15(ptr nocapture %dest, ptr %src) nounwind {541; RV32-LABEL: aligned_memcpy15:542; RV32: # %bb.0: # %entry543; RV32-NEXT: lbu a2, 14(a1)544; RV32-NEXT: sb a2, 14(a0)545; RV32-NEXT: lh a2, 12(a1)546; RV32-NEXT: sh a2, 12(a0)547; RV32-NEXT: lw a2, 8(a1)548; RV32-NEXT: sw a2, 8(a0)549; RV32-NEXT: lw a2, 4(a1)550; RV32-NEXT: sw a2, 4(a0)551; RV32-NEXT: lw a1, 0(a1)552; RV32-NEXT: sw a1, 0(a0)553; RV32-NEXT: ret554;555; RV64-LABEL: aligned_memcpy15:556; RV64: # %bb.0: # %entry557; RV64-NEXT: lbu a2, 14(a1)558; RV64-NEXT: sb a2, 14(a0)559; RV64-NEXT: lh a2, 12(a1)560; RV64-NEXT: sh a2, 12(a0)561; RV64-NEXT: lw a2, 8(a1)562; RV64-NEXT: sw a2, 8(a0)563; RV64-NEXT: ld a1, 0(a1)564; RV64-NEXT: sd a1, 0(a0)565; RV64-NEXT: ret566;567; RV32-FAST-LABEL: aligned_memcpy15:568; RV32-FAST: # %bb.0: # %entry569; RV32-FAST-NEXT: lw a2, 11(a1)570; RV32-FAST-NEXT: sw a2, 11(a0)571; RV32-FAST-NEXT: lw a2, 8(a1)572; RV32-FAST-NEXT: sw a2, 8(a0)573; RV32-FAST-NEXT: lw a2, 4(a1)574; RV32-FAST-NEXT: sw a2, 4(a0)575; RV32-FAST-NEXT: lw a1, 0(a1)576; RV32-FAST-NEXT: sw a1, 0(a0)577; RV32-FAST-NEXT: ret578;579; RV64-FAST-LABEL: aligned_memcpy15:580; RV64-FAST: # %bb.0: # %entry581; RV64-FAST-NEXT: ld a2, 7(a1)582; RV64-FAST-NEXT: sd a2, 7(a0)583; RV64-FAST-NEXT: ld a1, 0(a1)584; RV64-FAST-NEXT: sd a1, 0(a0)585; RV64-FAST-NEXT: ret586entry:587 tail call void @llvm.memcpy.p0.p0.i64(ptr align 8 %dest, ptr align 8 %src, i64 15, i1 false)588 ret void589}590 591define void @aligned_memcpy16(ptr nocapture %dest, ptr %src) nounwind {592; RV32-BOTH-LABEL: aligned_memcpy16:593; RV32-BOTH: # %bb.0: # %entry594; RV32-BOTH-NEXT: lw a2, 12(a1)595; RV32-BOTH-NEXT: sw a2, 12(a0)596; RV32-BOTH-NEXT: lw a2, 8(a1)597; RV32-BOTH-NEXT: sw a2, 8(a0)598; RV32-BOTH-NEXT: lw a2, 4(a1)599; RV32-BOTH-NEXT: sw a2, 4(a0)600; RV32-BOTH-NEXT: lw a1, 0(a1)601; RV32-BOTH-NEXT: sw a1, 0(a0)602; RV32-BOTH-NEXT: ret603;604; RV64-BOTH-LABEL: aligned_memcpy16:605; RV64-BOTH: # %bb.0: # %entry606; RV64-BOTH-NEXT: ld a2, 8(a1)607; RV64-BOTH-NEXT: sd a2, 8(a0)608; RV64-BOTH-NEXT: ld a1, 0(a1)609; RV64-BOTH-NEXT: sd a1, 0(a0)610; RV64-BOTH-NEXT: ret611entry:612 tail call void @llvm.memcpy.p0.p0.i64(ptr align 8 %dest, ptr align 8 %src, i64 16, i1 false)613 ret void614}615 616define void @aligned_memcpy31(ptr nocapture %dest, ptr %src) nounwind {617; RV32-LABEL: aligned_memcpy31:618; RV32: # %bb.0: # %entry619; RV32-NEXT: li a2, 31620; RV32-NEXT: tail memcpy621;622; RV64-LABEL: aligned_memcpy31:623; RV64: # %bb.0: # %entry624; RV64-NEXT: lbu a2, 30(a1)625; RV64-NEXT: sb a2, 30(a0)626; RV64-NEXT: lh a2, 28(a1)627; RV64-NEXT: sh a2, 28(a0)628; RV64-NEXT: lw a2, 24(a1)629; RV64-NEXT: sw a2, 24(a0)630; RV64-NEXT: ld a2, 16(a1)631; RV64-NEXT: sd a2, 16(a0)632; RV64-NEXT: ld a2, 8(a1)633; RV64-NEXT: sd a2, 8(a0)634; RV64-NEXT: ld a1, 0(a1)635; RV64-NEXT: sd a1, 0(a0)636; RV64-NEXT: ret637;638; RV32-FAST-LABEL: aligned_memcpy31:639; RV32-FAST: # %bb.0: # %entry640; RV32-FAST-NEXT: lw a2, 27(a1)641; RV32-FAST-NEXT: sw a2, 27(a0)642; RV32-FAST-NEXT: lw a2, 24(a1)643; RV32-FAST-NEXT: sw a2, 24(a0)644; RV32-FAST-NEXT: lw a2, 20(a1)645; RV32-FAST-NEXT: sw a2, 20(a0)646; RV32-FAST-NEXT: lw a2, 16(a1)647; RV32-FAST-NEXT: sw a2, 16(a0)648; RV32-FAST-NEXT: lw a2, 12(a1)649; RV32-FAST-NEXT: sw a2, 12(a0)650; RV32-FAST-NEXT: lw a2, 8(a1)651; RV32-FAST-NEXT: sw a2, 8(a0)652; RV32-FAST-NEXT: lw a2, 4(a1)653; RV32-FAST-NEXT: sw a2, 4(a0)654; RV32-FAST-NEXT: lw a1, 0(a1)655; RV32-FAST-NEXT: sw a1, 0(a0)656; RV32-FAST-NEXT: ret657;658; RV64-FAST-LABEL: aligned_memcpy31:659; RV64-FAST: # %bb.0: # %entry660; RV64-FAST-NEXT: ld a2, 23(a1)661; RV64-FAST-NEXT: sd a2, 23(a0)662; RV64-FAST-NEXT: ld a2, 16(a1)663; RV64-FAST-NEXT: sd a2, 16(a0)664; RV64-FAST-NEXT: ld a2, 8(a1)665; RV64-FAST-NEXT: sd a2, 8(a0)666; RV64-FAST-NEXT: ld a1, 0(a1)667; RV64-FAST-NEXT: sd a1, 0(a0)668; RV64-FAST-NEXT: ret669entry:670 tail call void @llvm.memcpy.p0.p0.i64(ptr align 8 %dest, ptr align 8 %src, i64 31, i1 false)671 ret void672}673 674; ------------------------------------------------------------------------675; A few partially aligned cases676 677define void @memcpy16_align4(ptr nocapture %dest, ptr nocapture %src) nounwind {678; RV32-BOTH-LABEL: memcpy16_align4:679; RV32-BOTH: # %bb.0: # %entry680; RV32-BOTH-NEXT: lw a2, 12(a1)681; RV32-BOTH-NEXT: sw a2, 12(a0)682; RV32-BOTH-NEXT: lw a2, 8(a1)683; RV32-BOTH-NEXT: sw a2, 8(a0)684; RV32-BOTH-NEXT: lw a2, 4(a1)685; RV32-BOTH-NEXT: sw a2, 4(a0)686; RV32-BOTH-NEXT: lw a1, 0(a1)687; RV32-BOTH-NEXT: sw a1, 0(a0)688; RV32-BOTH-NEXT: ret689;690; RV64-LABEL: memcpy16_align4:691; RV64: # %bb.0: # %entry692; RV64-NEXT: lw a2, 12(a1)693; RV64-NEXT: sw a2, 12(a0)694; RV64-NEXT: lw a2, 8(a1)695; RV64-NEXT: sw a2, 8(a0)696; RV64-NEXT: lw a2, 4(a1)697; RV64-NEXT: sw a2, 4(a0)698; RV64-NEXT: lw a1, 0(a1)699; RV64-NEXT: sw a1, 0(a0)700; RV64-NEXT: ret701;702; RV64-FAST-LABEL: memcpy16_align4:703; RV64-FAST: # %bb.0: # %entry704; RV64-FAST-NEXT: ld a2, 8(a1)705; RV64-FAST-NEXT: sd a2, 8(a0)706; RV64-FAST-NEXT: ld a1, 0(a1)707; RV64-FAST-NEXT: sd a1, 0(a0)708; RV64-FAST-NEXT: ret709entry:710 tail call void @llvm.memcpy.p0.p0.i32(ptr align 4 %dest, ptr align 4 %src, i32 16, i1 false)711 ret void712}713 714define i32 @memcpy11_align8(ptr nocapture %dest, ptr %src) {715; RV32-LABEL: memcpy11_align8:716; RV32: # %bb.0: # %entry717; RV32-NEXT: lbu a2, 10(a1)718; RV32-NEXT: sb a2, 10(a0)719; RV32-NEXT: lh a2, 8(a1)720; RV32-NEXT: sh a2, 8(a0)721; RV32-NEXT: lw a2, 4(a1)722; RV32-NEXT: sw a2, 4(a0)723; RV32-NEXT: lw a1, 0(a1)724; RV32-NEXT: sw a1, 0(a0)725; RV32-NEXT: li a0, 0726; RV32-NEXT: ret727;728; RV64-LABEL: memcpy11_align8:729; RV64: # %bb.0: # %entry730; RV64-NEXT: lbu a2, 10(a1)731; RV64-NEXT: sb a2, 10(a0)732; RV64-NEXT: lh a2, 8(a1)733; RV64-NEXT: sh a2, 8(a0)734; RV64-NEXT: ld a1, 0(a1)735; RV64-NEXT: sd a1, 0(a0)736; RV64-NEXT: li a0, 0737; RV64-NEXT: ret738;739; RV32-FAST-LABEL: memcpy11_align8:740; RV32-FAST: # %bb.0: # %entry741; RV32-FAST-NEXT: lw a2, 7(a1)742; RV32-FAST-NEXT: sw a2, 7(a0)743; RV32-FAST-NEXT: lw a2, 4(a1)744; RV32-FAST-NEXT: sw a2, 4(a0)745; RV32-FAST-NEXT: lw a1, 0(a1)746; RV32-FAST-NEXT: sw a1, 0(a0)747; RV32-FAST-NEXT: li a0, 0748; RV32-FAST-NEXT: ret749;750; RV64-FAST-LABEL: memcpy11_align8:751; RV64-FAST: # %bb.0: # %entry752; RV64-FAST-NEXT: lw a2, 7(a1)753; RV64-FAST-NEXT: sw a2, 7(a0)754; RV64-FAST-NEXT: ld a1, 0(a1)755; RV64-FAST-NEXT: sd a1, 0(a0)756; RV64-FAST-NEXT: li a0, 0757; RV64-FAST-NEXT: ret758entry:759 call void @llvm.memcpy.p0.p0.i32(ptr align 8 %dest, ptr align 8 %src, i32 11, i1 false)760 ret i32 0761}762 763