249 lines · plain
1; Test memcmp using CLC, with i32 results.2;3; RUN: llc < %s -mtriple=s390x-linux-gnu | FileCheck %s4 5declare signext i32 @memcmp(ptr %src1, ptr %src2, i64 %size)6 7; Zero-length comparisons should be optimized away.8define i32 @f1(ptr %src1, ptr %src2) {9; CHECK-LABEL: f1:10; CHECK: lhi %r2, 011; CHECK: br %r1412 %res = call i32 @memcmp(ptr %src1, ptr %src2, i64 0)13 ret i32 %res14}15 16; Check a case where the result is used as an integer.17define i32 @f2(ptr %src1, ptr %src2) {18; CHECK-LABEL: f2:19; CHECK: clc 0(2,%r3), 0(%r2)20; CHECK: ipm %r221; CHECK: sll %r2, 222; CHECK: sra %r2, 3023; CHECK: br %r1424 %res = call i32 @memcmp(ptr %src1, ptr %src2, i64 2)25 ret i32 %res26}27 28; Check a case where the result is tested for equality.29define void @f3(ptr %src1, ptr %src2, ptr %dest) {30; CHECK-LABEL: f3:31; CHECK: clc 0(3,%r3), 0(%r2)32; CHECK-NEXT: ber %r1433; CHECK: br %r1434 %res = call i32 @memcmp(ptr %src1, ptr %src2, i64 3)35 %cmp = icmp eq i32 %res, 036 br i1 %cmp, label %exit, label %store37 38store:39 store i32 0, ptr %dest40 br label %exit41 42exit:43 ret void44}45 46; Check a case where the result is tested for inequality.47define void @f4(ptr %src1, ptr %src2, ptr %dest) {48; CHECK-LABEL: f4:49; CHECK: clc 0(4,%r3), 0(%r2)50; CHECK-NEXT: blhr %r1451; CHECK: br %r1452entry:53 %res = call i32 @memcmp(ptr %src1, ptr %src2, i64 4)54 %cmp = icmp ne i32 %res, 055 br i1 %cmp, label %exit, label %store56 57store:58 store i32 0, ptr %dest59 br label %exit60 61exit:62 ret void63}64 65; Check a case where the result is tested via slt.66define void @f5(ptr %src1, ptr %src2, ptr %dest) {67; CHECK-LABEL: f5:68; CHECK: clc 0(5,%r3), 0(%r2)69; CHECK-NEXT: bhr %r1470; CHECK: br %r1471entry:72 %res = call i32 @memcmp(ptr %src1, ptr %src2, i64 5)73 %cmp = icmp slt i32 %res, 074 br i1 %cmp, label %exit, label %store75 76store:77 store i32 0, ptr %dest78 br label %exit79 80exit:81 ret void82}83 84; Check a case where the result is tested for sgt.85define void @f6(ptr %src1, ptr %src2, ptr %dest) {86; CHECK-LABEL: f6:87; CHECK: clc 0(6,%r3), 0(%r2)88; CHECK-NEXT: blr %r1489; CHECK: br %r1490entry:91 %res = call i32 @memcmp(ptr %src1, ptr %src2, i64 6)92 %cmp = icmp sgt i32 %res, 093 br i1 %cmp, label %exit, label %store94 95store:96 store i32 0, ptr %dest97 br label %exit98 99exit:100 ret void101}102 103; Check the upper end of the CLC range. Here the result is used both as104; an integer and for branching.105define i32 @f7(ptr %src1, ptr %src2, ptr %dest) {106; CHECK-LABEL: f7:107; CHECK: clc 0(256,%r3), 0(%r2)108; CHECK: ipm %r2109; CHECK: sll %r2, 2110; CHECK: sra %r2, 30111; CHECK: blr %r14112; CHECK: br %r14113entry:114 %res = call i32 @memcmp(ptr %src1, ptr %src2, i64 256)115 %cmp = icmp slt i32 %res, 0116 br i1 %cmp, label %exit, label %store117 118store:119 store i32 0, ptr %dest120 br label %exit121 122exit:123 ret i32 %res124}125 126; 257 bytes needs two CLCs.127define i32 @f8(ptr %src1, ptr %src2) {128; CHECK-LABEL: f8:129; CHECK: clc 0(256,%r3), 0(%r2)130; CHECK: jlh [[LABEL:\..*]]131; CHECK: clc 256(1,%r3), 256(%r2)132; CHECK: [[LABEL]]:133; CHECK: ipm [[REG:%r[0-5]]]134; CHECK: br %r14135 %res = call i32 @memcmp(ptr %src1, ptr %src2, i64 257)136 ret i32 %res137}138 139; Test a comparison of 258 bytes in which the CC result can be used directly.140define void @f9(ptr %src1, ptr %src2, ptr %dest) {141; CHECK-LABEL: f9:142; CHECK: clc 0(256,%r3), 0(%r2)143; CHECK: jlh [[LABEL:\..*]]144; CHECK: clc 256(1,%r3), 256(%r2)145; CHECK: [[LABEL]]:146; CHECK-NEXT: bhr %r14147; CHECK: br %r14148entry:149 %res = call i32 @memcmp(ptr %src1, ptr %src2, i64 257)150 %cmp = icmp slt i32 %res, 0151 br i1 %cmp, label %exit, label %store152 153store:154 store i32 0, ptr %dest155 br label %exit156 157exit:158 ret void159}160 161; Test the largest size that can use two CLCs.162define i32 @f10(ptr %src1, ptr %src2) {163; CHECK-LABEL: f10:164; CHECK: clc 0(256,%r3), 0(%r2)165; CHECK: jlh [[LABEL:\..*]]166; CHECK: clc 256(256,%r3), 256(%r2)167; CHECK: [[LABEL]]:168; CHECK: ipm [[REG:%r[0-5]]]169; CHECK: br %r14170 %res = call i32 @memcmp(ptr %src1, ptr %src2, i64 512)171 ret i32 %res172}173 174; Test the smallest size that needs 3 CLCs.175define i32 @f11(ptr %src1, ptr %src2) {176; CHECK-LABEL: f11:177; CHECK: clc 0(256,%r3), 0(%r2)178; CHECK: jlh [[LABEL:\..*]]179; CHECK: clc 256(256,%r3), 256(%r2)180; CHECK: jlh [[LABEL]]181; CHECK: clc 512(1,%r3), 512(%r2)182; CHECK: [[LABEL]]:183; CHECK: ipm [[REG:%r[0-5]]]184; CHECK: br %r14185 %res = call i32 @memcmp(ptr %src1, ptr %src2, i64 513)186 ret i32 %res187}188 189; Test the largest size than can use 3 CLCs.190define i32 @f12(ptr %src1, ptr %src2) {191; CHECK-LABEL: f12:192; CHECK: clc 0(256,%r3), 0(%r2)193; CHECK: jlh [[LABEL:\..*]]194; CHECK: clc 256(256,%r3), 256(%r2)195; CHECK: jlh [[LABEL]]196; CHECK: clc 512(256,%r3), 512(%r2)197; CHECK: [[LABEL]]:198; CHECK: ipm [[REG:%r[0-5]]]199; CHECK: br %r14200 %res = call i32 @memcmp(ptr %src1, ptr %src2, i64 768)201 ret i32 %res202}203 204; The next size up uses a loop instead. We leave the more complicated205; loop tests to memcpy-01.ll, which shares the same form.206define i32 @f13(ptr %src1, ptr %src2) {207; CHECK-LABEL: f13:208; CHECK: lghi [[COUNT:%r[0-5]]], 3209; CHECK: [[LOOP:.L[^:]*]]:210; CHECK: clc 0(256,%r3), 0(%r2)211; CHECK: jlh [[LABEL:\..*]]212; CHECK-DAG: la %r2, 256(%r2)213; CHECK-DAG: la %r3, 256(%r3)214; CHECK: brctg [[COUNT]], [[LOOP]]215; CHECK: clc 0(1,%r3), 0(%r2)216; CHECK: [[LABEL]]:217; CHECK: ipm [[REG:%r[0-5]]]218; CHECK: br %r14219 %res = call i32 @memcmp(ptr %src1, ptr %src2, i64 769)220 ret i32 %res221}222 223define i32 @f14(ptr %src1, ptr %src2, i64 %Len) {224; CHECK-LABEL: f14:225; CHECK: # %bb.0:226; CHECK-NEXT: aghi %r4, -1227; CHECK-NEXT: cghi %r4, -1228; CHECK-NEXT: je .LBB13_5229; CHECK-NEXT: # %bb.1:230; CHECK-NEXT: srlg %r0, %r4, 8231; CHECK-NEXT: cgije %r0, 0, .LBB13_4232; CHECK-NEXT: .LBB13_2: # =>This Inner Loop Header: Depth=1233; CHECK-NEXT: clc 0(256,%r3), 0(%r2)234; CHECK-NEXT: jlh .LBB13_5235; CHECK-NEXT: # %bb.3: # in Loop: Header=BB13_2 Depth=1236; CHECK-NEXT: la %r3, 256(%r3)237; CHECK-NEXT: la %r2, 256(%r2)238; CHECK-NEXT: brctg %r0, .LBB13_2239; CHECK-NEXT: .LBB13_4:240; CHECK-NEXT: exrl %r4, .Ltmp0241; CHECK-NEXT: .LBB13_5:242; CHECK-NEXT: ipm %r2243; CHECK-NEXT: sll %r2, 2244; CHECK-NEXT: sra %r2, 30245; CHECK-NEXT: br %r14246 %res = call i32 @memcmp(ptr %src1, ptr %src2, i64 %Len)247 ret i32 %res248}249