229 lines · plain
1; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py UTC_ARGS: --version 52; RUN: llc -mtriple=riscv32 -mattr=+zilsd -verify-machineinstrs < %s \3; RUN: | FileCheck -check-prefixes=CHECK,SLOW %s4; RUN: llc -mtriple=riscv32 -mattr=+zilsd,+unaligned-scalar-mem -verify-machineinstrs < %s \5; RUN: | FileCheck -check-prefixes=CHECK,FAST %s6; RUN: llc -mtriple=riscv32 -mattr=+zilsd,+zilsd-4byte-align -verify-machineinstrs < %s \7; RUN: | FileCheck -check-prefixes=CHECK,4BYTEALIGN %s8 9define i64 @load(ptr %a) nounwind {10; CHECK-LABEL: load:11; CHECK: # %bb.0:12; CHECK-NEXT: mv a2, a013; CHECK-NEXT: ld a0, 80(a0)14; CHECK-NEXT: ld zero, 0(a2)15; CHECK-NEXT: ret16 %1 = getelementptr i64, ptr %a, i32 1017 %2 = load i64, ptr %118 %3 = load volatile i64, ptr %a19 ret i64 %220}21 22define i64 @load_align4(ptr %a) nounwind {23; SLOW-LABEL: load_align4:24; SLOW: # %bb.0:25; SLOW-NEXT: lw a2, 80(a0)26; SLOW-NEXT: lw a1, 84(a0)27; SLOW-NEXT: mv a0, a228; SLOW-NEXT: ret29;30; FAST-LABEL: load_align4:31; FAST: # %bb.0:32; FAST-NEXT: ld a0, 80(a0)33; FAST-NEXT: ret34;35; 4BYTEALIGN-LABEL: load_align4:36; 4BYTEALIGN: # %bb.0:37; 4BYTEALIGN-NEXT: ld a0, 80(a0)38; 4BYTEALIGN-NEXT: ret39 %1 = getelementptr i64, ptr %a, i32 1040 %2 = load i64, ptr %1, align 441 ret i64 %242}43 44define void @store(ptr %a, i64 %b) nounwind {45; CHECK-LABEL: store:46; CHECK: # %bb.0:47; CHECK-NEXT: mv a3, a248; CHECK-NEXT: mv a2, a149; CHECK-NEXT: sd a2, 0(a0)50; CHECK-NEXT: sd a2, 88(a0)51; CHECK-NEXT: ret52 store i64 %b, ptr %a53 %1 = getelementptr i64, ptr %a, i32 1154 store i64 %b, ptr %155 ret void56}57 58define void @store_align4(ptr %a, i64 %b) nounwind {59; SLOW-LABEL: store_align4:60; SLOW: # %bb.0:61; SLOW-NEXT: sw a1, 88(a0)62; SLOW-NEXT: sw a2, 92(a0)63; SLOW-NEXT: ret64;65; FAST-LABEL: store_align4:66; FAST: # %bb.0:67; FAST-NEXT: mv a3, a268; FAST-NEXT: mv a2, a169; FAST-NEXT: sd a2, 88(a0)70; FAST-NEXT: ret71;72; 4BYTEALIGN-LABEL: store_align4:73; 4BYTEALIGN: # %bb.0:74; 4BYTEALIGN-NEXT: mv a3, a275; 4BYTEALIGN-NEXT: mv a2, a176; 4BYTEALIGN-NEXT: sd a2, 88(a0)77; 4BYTEALIGN-NEXT: ret78 %1 = getelementptr i64, ptr %a, i32 1179 store i64 %b, ptr %1, align 480 ret void81}82 83define i64 @load_unaligned(ptr %p) {84; SLOW-LABEL: load_unaligned:85; SLOW: # %bb.0:86; SLOW-NEXT: lbu a1, 1(a0)87; SLOW-NEXT: lbu a2, 2(a0)88; SLOW-NEXT: lbu a3, 3(a0)89; SLOW-NEXT: lbu a4, 0(a0)90; SLOW-NEXT: slli a1, a1, 891; SLOW-NEXT: slli a2, a2, 1692; SLOW-NEXT: slli a3, a3, 2493; SLOW-NEXT: or a1, a1, a494; SLOW-NEXT: or a2, a3, a295; SLOW-NEXT: lbu a3, 5(a0)96; SLOW-NEXT: lbu a4, 4(a0)97; SLOW-NEXT: lbu a5, 6(a0)98; SLOW-NEXT: lbu a0, 7(a0)99; SLOW-NEXT: slli a3, a3, 8100; SLOW-NEXT: or a3, a3, a4101; SLOW-NEXT: slli a5, a5, 16102; SLOW-NEXT: slli a0, a0, 24103; SLOW-NEXT: or a5, a0, a5104; SLOW-NEXT: or a0, a2, a1105; SLOW-NEXT: or a1, a5, a3106; SLOW-NEXT: ret107;108; FAST-LABEL: load_unaligned:109; FAST: # %bb.0:110; FAST-NEXT: ld a0, 0(a0)111; FAST-NEXT: ret112;113; 4BYTEALIGN-LABEL: load_unaligned:114; 4BYTEALIGN: # %bb.0:115; 4BYTEALIGN-NEXT: lbu a1, 1(a0)116; 4BYTEALIGN-NEXT: lbu a2, 2(a0)117; 4BYTEALIGN-NEXT: lbu a3, 3(a0)118; 4BYTEALIGN-NEXT: lbu a4, 0(a0)119; 4BYTEALIGN-NEXT: slli a1, a1, 8120; 4BYTEALIGN-NEXT: slli a2, a2, 16121; 4BYTEALIGN-NEXT: slli a3, a3, 24122; 4BYTEALIGN-NEXT: or a1, a1, a4123; 4BYTEALIGN-NEXT: or a2, a3, a2124; 4BYTEALIGN-NEXT: lbu a3, 5(a0)125; 4BYTEALIGN-NEXT: lbu a4, 4(a0)126; 4BYTEALIGN-NEXT: lbu a5, 6(a0)127; 4BYTEALIGN-NEXT: lbu a0, 7(a0)128; 4BYTEALIGN-NEXT: slli a3, a3, 8129; 4BYTEALIGN-NEXT: or a3, a3, a4130; 4BYTEALIGN-NEXT: slli a5, a5, 16131; 4BYTEALIGN-NEXT: slli a0, a0, 24132; 4BYTEALIGN-NEXT: or a5, a0, a5133; 4BYTEALIGN-NEXT: or a0, a2, a1134; 4BYTEALIGN-NEXT: or a1, a5, a3135; 4BYTEALIGN-NEXT: ret136 %res = load i64, ptr %p, align 1137 ret i64 %res138}139 140define void @store_unaligned(ptr %p, i64 %v) {141; SLOW-LABEL: store_unaligned:142; SLOW: # %bb.0:143; SLOW-NEXT: srli a3, a2, 24144; SLOW-NEXT: srli a4, a2, 16145; SLOW-NEXT: srli a5, a2, 8146; SLOW-NEXT: srli a6, a1, 24147; SLOW-NEXT: srli a7, a1, 16148; SLOW-NEXT: sb a2, 4(a0)149; SLOW-NEXT: sb a5, 5(a0)150; SLOW-NEXT: sb a4, 6(a0)151; SLOW-NEXT: sb a3, 7(a0)152; SLOW-NEXT: srli a2, a1, 8153; SLOW-NEXT: sb a1, 0(a0)154; SLOW-NEXT: sb a2, 1(a0)155; SLOW-NEXT: sb a7, 2(a0)156; SLOW-NEXT: sb a6, 3(a0)157; SLOW-NEXT: ret158;159; FAST-LABEL: store_unaligned:160; FAST: # %bb.0:161; FAST-NEXT: mv a3, a2162; FAST-NEXT: mv a2, a1163; FAST-NEXT: sd a2, 0(a0)164; FAST-NEXT: ret165;166; 4BYTEALIGN-LABEL: store_unaligned:167; 4BYTEALIGN: # %bb.0:168; 4BYTEALIGN-NEXT: srli a3, a2, 24169; 4BYTEALIGN-NEXT: srli a4, a2, 16170; 4BYTEALIGN-NEXT: srli a5, a2, 8171; 4BYTEALIGN-NEXT: srli a6, a1, 24172; 4BYTEALIGN-NEXT: srli a7, a1, 16173; 4BYTEALIGN-NEXT: sb a2, 4(a0)174; 4BYTEALIGN-NEXT: sb a5, 5(a0)175; 4BYTEALIGN-NEXT: sb a4, 6(a0)176; 4BYTEALIGN-NEXT: sb a3, 7(a0)177; 4BYTEALIGN-NEXT: srli a2, a1, 8178; 4BYTEALIGN-NEXT: sb a1, 0(a0)179; 4BYTEALIGN-NEXT: sb a2, 1(a0)180; 4BYTEALIGN-NEXT: sb a7, 2(a0)181; 4BYTEALIGN-NEXT: sb a6, 3(a0)182; 4BYTEALIGN-NEXT: ret183 store i64 %v, ptr %p, align 1184 ret void185}186 187@g = dso_local global i64 0, align 8188 189define i64 @load_g() nounwind {190; CHECK-LABEL: load_g:191; CHECK: # %bb.0: # %entry192; CHECK-NEXT: lui a0, %hi(g)193; CHECK-NEXT: ld a0, %lo(g)(a0)194; CHECK-NEXT: ret195entry:196 %0 = load i64, ptr @g197 ret i64 %0198}199 200define void @store_g() nounwind {201; CHECK-LABEL: store_g:202; CHECK: # %bb.0: # %entyr203; CHECK-NEXT: lui a0, %hi(g)204; CHECK-NEXT: sd zero, %lo(g)(a0)205; CHECK-NEXT: ret206entyr:207 store i64 0, ptr @g208 ret void209}210 211define void @large_offset(ptr nocapture %p, i64 %d) nounwind {212; CHECK-LABEL: large_offset:213; CHECK: # %bb.0: # %entry214; CHECK-NEXT: lui a1, 4215; CHECK-NEXT: add a0, a0, a1216; CHECK-NEXT: ld a2, -384(a0)217; CHECK-NEXT: addi a2, a2, 1218; CHECK-NEXT: seqz a1, a2219; CHECK-NEXT: add a3, a3, a1220; CHECK-NEXT: sd a2, -384(a0)221; CHECK-NEXT: ret222entry:223 %add.ptr = getelementptr inbounds i64, ptr %p, i64 2000224 %a = load i64, ptr %add.ptr, align 8225 %b = add i64 %a, 1226 store i64 %b, ptr %add.ptr, align 8227 ret void228}229