186 lines · plain
1; NOTE: Assertions have been autogenerated by utils/update_test_checks.py2; RUN: opt -passes="loop-idiom" < %s -S | FileCheck %s3 4define void @looper(ptr noalias nocapture readonly %M, ptr noalias nocapture %out) {5; CHECK-LABEL: @looper(6; CHECK-NEXT: entry:7; CHECK-NEXT: call void @llvm.memcpy.p0.p0.i64(ptr align 8 [[OUT:%.*]], ptr align 8 [[M:%.*]], i64 256, i1 false), !tbaa [[TBAA0:![0-9]+]]8; CHECK-NEXT: br label [[FOR_BODY4:%.*]]9; CHECK: for.body4:10; CHECK-NEXT: [[J_020:%.*]] = phi i64 [ 0, [[ENTRY:%.*]] ], [ [[INC:%.*]], [[FOR_BODY4]] ]11; CHECK-NEXT: [[ARRAYIDX:%.*]] = getelementptr inbounds double, ptr [[M]], i64 [[J_020]]12; CHECK-NEXT: [[A0:%.*]] = load double, ptr [[ARRAYIDX]], align 8, !tbaa [[TBAA0]]13; CHECK-NEXT: [[ARRAYIDX8:%.*]] = getelementptr inbounds double, ptr [[OUT]], i64 [[J_020]]14; CHECK-NEXT: [[INC]] = add nuw nsw i64 [[J_020]], 115; CHECK-NEXT: [[CMP2:%.*]] = icmp ult i64 [[J_020]], 3116; CHECK-NEXT: br i1 [[CMP2]], label [[FOR_BODY4]], label [[FOR_COND_CLEANUP:%.*]]17; CHECK: for.cond.cleanup:18; CHECK-NEXT: ret void19;20entry:21 br label %for.body422 23for.body4: ; preds = %for.cond1.preheader, %for.body424 %j.020 = phi i64 [ 0, %entry ], [ %inc, %for.body4 ]25 %arrayidx = getelementptr inbounds double, ptr %M, i64 %j.02026 %a0 = load double, ptr %arrayidx, align 8, !tbaa !527 %arrayidx8 = getelementptr inbounds double, ptr %out, i64 %j.02028 store double %a0, ptr %arrayidx8, align 8, !tbaa !529 %inc = add nuw nsw i64 %j.020, 130 %cmp2 = icmp ult i64 %j.020, 3131 br i1 %cmp2, label %for.body4, label %for.cond.cleanup32 33for.cond.cleanup: ; preds = %for.cond.cleanup334 ret void35}36 37 38define void @looperBadMerge(ptr noalias nocapture readonly %M, ptr noalias nocapture %out) {39; CHECK-LABEL: @looperBadMerge(40; CHECK-NEXT: entry:41; CHECK-NEXT: call void @llvm.memcpy.p0.p0.i64(ptr align 8 [[OUT:%.*]], ptr align 8 [[M:%.*]], i64 256, i1 false), !tbaa [[TBAA4:![0-9]+]]42; CHECK-NEXT: br label [[FOR_BODY4:%.*]]43; CHECK: for.body4:44; CHECK-NEXT: [[J_020:%.*]] = phi i64 [ 0, [[ENTRY:%.*]] ], [ [[INC:%.*]], [[FOR_BODY4]] ]45; CHECK-NEXT: [[ARRAYIDX:%.*]] = getelementptr inbounds double, ptr [[M]], i64 [[J_020]]46; CHECK-NEXT: [[A0:%.*]] = load double, ptr [[ARRAYIDX]], align 8, !tbaa [[TBAA0]]47; CHECK-NEXT: [[ARRAYIDX8:%.*]] = getelementptr inbounds double, ptr [[OUT]], i64 [[J_020]]48; CHECK-NEXT: [[INC]] = add nuw nsw i64 [[J_020]], 149; CHECK-NEXT: [[CMP2:%.*]] = icmp ult i64 [[J_020]], 3150; CHECK-NEXT: br i1 [[CMP2]], label [[FOR_BODY4]], label [[FOR_COND_CLEANUP:%.*]]51; CHECK: for.cond.cleanup:52; CHECK-NEXT: ret void53;54entry:55 br label %for.body456 57for.body4: ; preds = %for.cond1.preheader, %for.body458 %j.020 = phi i64 [ 0, %entry ], [ %inc, %for.body4 ]59 %arrayidx = getelementptr inbounds double, ptr %M, i64 %j.02060 %a0 = load double, ptr %arrayidx, align 8, !tbaa !561 %arrayidx8 = getelementptr inbounds double, ptr %out, i64 %j.02062 store double %a0, ptr %arrayidx8, align 8, !tbaa !363 %inc = add nuw nsw i64 %j.020, 164 %cmp2 = icmp ult i64 %j.020, 3165 br i1 %cmp2, label %for.body4, label %for.cond.cleanup66 67for.cond.cleanup: ; preds = %for.cond.cleanup368 ret void69}70 71define void @looperGoodMerge(ptr noalias nocapture readonly %M, ptr noalias nocapture %out) {72; CHECK-LABEL: @looperGoodMerge(73; CHECK-NEXT: entry:74; CHECK-NEXT: call void @llvm.memcpy.p0.p0.i64(ptr align 8 [[OUT:%.*]], ptr align 8 [[M:%.*]], i64 256, i1 false)75; CHECK-NEXT: br label [[FOR_BODY4:%.*]]76; CHECK: for.body4:77; CHECK-NEXT: [[J_020:%.*]] = phi i64 [ 0, [[ENTRY:%.*]] ], [ [[INC:%.*]], [[FOR_BODY4]] ]78; CHECK-NEXT: [[ARRAYIDX:%.*]] = getelementptr inbounds double, ptr [[M]], i64 [[J_020]]79; CHECK-NEXT: [[A0:%.*]] = load double, ptr [[ARRAYIDX]], align 8, !tbaa [[TBAA0]]80; CHECK-NEXT: [[ARRAYIDX8:%.*]] = getelementptr inbounds double, ptr [[OUT]], i64 [[J_020]]81; CHECK-NEXT: [[INC]] = add nuw nsw i64 [[J_020]], 182; CHECK-NEXT: [[CMP2:%.*]] = icmp ult i64 [[J_020]], 3183; CHECK-NEXT: br i1 [[CMP2]], label [[FOR_BODY4]], label [[FOR_COND_CLEANUP:%.*]]84; CHECK: for.cond.cleanup:85; CHECK-NEXT: ret void86;87entry:88 br label %for.body489 90for.body4: ; preds = %for.cond1.preheader, %for.body491 %j.020 = phi i64 [ 0, %entry ], [ %inc, %for.body4 ]92 %arrayidx = getelementptr inbounds double, ptr %M, i64 %j.02093 %a0 = load double, ptr %arrayidx, align 8, !tbaa !594 %arrayidx8 = getelementptr inbounds double, ptr %out, i64 %j.02095 store double %a0, ptr %arrayidx8, align 896 %inc = add nuw nsw i64 %j.020, 197 %cmp2 = icmp ult i64 %j.020, 3198 br i1 %cmp2, label %for.body4, label %for.cond.cleanup99 100for.cond.cleanup: ; preds = %for.cond.cleanup3101 ret void102}103 104define void @looperConstantTBAAStruct(ptr nocapture noalias %out, ptr nocapture noalias %in) {105; CHECK-LABEL: @looperConstantTBAAStruct(106; CHECK-NEXT: entry:107; CHECK-NEXT: call void @llvm.memcpy.p0.p0.i64(ptr align 8 [[OUT:%.*]], ptr align 8 [[IN:%.*]], i64 32, i1 false), !tbaa [[TBAA5:![0-9]+]]108; CHECK-NEXT: br label [[FOR_BODY4:%.*]]109; CHECK: for.body4:110; CHECK-NEXT: [[J_020:%.*]] = phi i64 [ 0, [[ENTRY:%.*]] ], [ [[INC:%.*]], [[FOR_BODY4]] ]111; CHECK-NEXT: [[ARRAYIDX:%.*]] = getelementptr inbounds double, ptr [[IN]], i64 [[J_020]]112; CHECK-NEXT: [[A0:%.*]] = load double, ptr [[ARRAYIDX]], align 8, !tbaa [[TBAA9:![0-9]+]]113; CHECK-NEXT: [[ARRAYIDX8:%.*]] = getelementptr inbounds double, ptr [[OUT]], i64 [[J_020]]114; CHECK-NEXT: [[INC]] = add nuw nsw i64 [[J_020]], 1115; CHECK-NEXT: [[CMP2:%.*]] = icmp ult i64 [[J_020]], 3116; CHECK-NEXT: br i1 [[CMP2]], label [[FOR_BODY4]], label [[FOR_COND_CLEANUP:%.*]]117; CHECK: for.cond.cleanup:118; CHECK-NEXT: ret void119;120entry:121 br label %for.body4122 123for.body4: ; preds = %for.cond1.preheader, %for.body4124 %j.020 = phi i64 [ 0, %entry ], [ %inc, %for.body4 ]125 %arrayidx = getelementptr inbounds double, ptr %in, i64 %j.020126 %a0 = load double, ptr %arrayidx, align 8, !tbaa !10127 %arrayidx8 = getelementptr inbounds double, ptr %out, i64 %j.020128 store double %a0, ptr %arrayidx8, align 8, !tbaa !10129 %inc = add nuw nsw i64 %j.020, 1130 %cmp2 = icmp ult i64 %j.020, 3131 br i1 %cmp2, label %for.body4, label %for.cond.cleanup132 133for.cond.cleanup: ; preds = %for.cond.cleanup3134 ret void135}136 137define void @looperVarTBAAStruct(ptr nocapture noalias %out, ptr nocapture noalias %in, i64 %len) {138; CHECK-LABEL: @looperVarTBAAStruct(139; CHECK-NEXT: entry:140; CHECK-NEXT: [[UMAX:%.*]] = call i64 @llvm.umax.i64(i64 [[LEN:%.*]], i64 1)141; CHECK-NEXT: [[TMP0:%.*]] = shl nuw i64 [[UMAX]], 3142; CHECK-NEXT: call void @llvm.memcpy.p0.p0.i64(ptr align 8 [[OUT:%.*]], ptr align 8 [[IN:%.*]], i64 [[TMP0]], i1 false)143; CHECK-NEXT: br label [[FOR_BODY4:%.*]]144; CHECK: for.body4:145; CHECK-NEXT: [[J_020:%.*]] = phi i64 [ 0, [[ENTRY:%.*]] ], [ [[INC:%.*]], [[FOR_BODY4]] ]146; CHECK-NEXT: [[ARRAYIDX:%.*]] = getelementptr inbounds double, ptr [[IN]], i64 [[J_020]]147; CHECK-NEXT: [[A0:%.*]] = load double, ptr [[ARRAYIDX]], align 8, !tbaa [[TBAA9]]148; CHECK-NEXT: [[ARRAYIDX8:%.*]] = getelementptr inbounds double, ptr [[OUT]], i64 [[J_020]]149; CHECK-NEXT: [[INC]] = add nuw nsw i64 [[J_020]], 1150; CHECK-NEXT: [[CMP2:%.*]] = icmp ult i64 [[INC]], [[LEN]]151; CHECK-NEXT: br i1 [[CMP2]], label [[FOR_BODY4]], label [[FOR_COND_CLEANUP:%.*]]152; CHECK: for.cond.cleanup:153; CHECK-NEXT: ret void154;155entry:156 br label %for.body4157 158for.body4: ; preds = %for.cond1.preheader, %for.body4159 %j.020 = phi i64 [ 0, %entry ], [ %inc, %for.body4 ]160 %arrayidx = getelementptr inbounds double, ptr %in, i64 %j.020161 %a0 = load double, ptr %arrayidx, align 8, !tbaa !10162 %arrayidx8 = getelementptr inbounds double, ptr %out, i64 %j.020163 store double %a0, ptr %arrayidx8, align 8, !tbaa !10164 %inc = add nuw nsw i64 %j.020, 1165 %cmp2 = icmp ult i64 %inc, %len166 br i1 %cmp2, label %for.body4, label %for.cond.cleanup167 168for.cond.cleanup: ; preds = %for.cond.cleanup3169 ret void170}171 172 173 174 175!3 = !{!4, !4, i64 0}176!4 = !{!"float", !7, i64 0}177!5 = !{!6, !6, i64 0}178!6 = !{!"double", !7, i64 0}179!7 = !{!"omnipotent char", !8, i64 0}180!8 = !{!"Simple C++ TBAA"}181 182!15 = !{!8, i64 0, !"omnipotent char"}183!17 = !{!15, i64 8, !"double"}184!9 = !{!15, i64 32, !"_ZTS1A", !17, i64 0, i64 8, !17, i64 8, i64 8, !17, i64 16, i64 8, !17, i64 24, i64 8}185!10 = !{!9, !17, i64 0, i64 8}186