134 lines · plain
1; NOTE: Assertions have been autogenerated by utils/update_test_checks.py2; RUN: opt -passes=dse %s -S | FileCheck --check-prefixes=CHECK %s3 4 5%struct.ham = type { [3 x double], [3 x double]}6 7declare void @may_throw()8declare void @llvm.memset.p0.i64(ptr nocapture writeonly, i8, i64, i1 immarg)9 10; We miss this case, because of an aggressive limit of partial overlap analysis.11; With a larger partial store limit, we remove the memset.12define void @overlap1(ptr %arg, i1 %cond) {13; CHECK-LABEL: @overlap1(14; CHECK-NEXT: bb:15; CHECK-NEXT: [[TMP:%.*]] = getelementptr inbounds [[STRUCT_HAM:%.*]], ptr [[ARG:%.*]], i64 0, i32 0, i64 216; CHECK-NEXT: [[TMP1:%.*]] = getelementptr inbounds [[STRUCT_HAM]], ptr [[ARG]], i64 0, i32 0, i64 117; CHECK-NEXT: [[TMP3:%.*]] = getelementptr inbounds [[STRUCT_HAM]], ptr [[ARG]], i64 0, i32 1, i64 218; CHECK-NEXT: [[TMP4:%.*]] = getelementptr inbounds [[STRUCT_HAM]], ptr [[ARG]], i64 0, i32 1, i64 119; CHECK-NEXT: [[TMP5:%.*]] = getelementptr inbounds [[STRUCT_HAM]], ptr [[ARG]], i64 0, i32 1, i32 020; CHECK-NEXT: br i1 [[COND:%.*]], label [[BB7:%.*]], label [[BB8:%.*]]21; CHECK: bb7:22; CHECK-NEXT: br label [[BB9:%.*]]23; CHECK: bb8:24; CHECK-NEXT: br label [[BB9]]25; CHECK: bb9:26; CHECK-NEXT: store double 1.000000e+00, ptr [[ARG]], align 827; CHECK-NEXT: store double 2.000000e+00, ptr [[TMP1]], align 828; CHECK-NEXT: store double 3.000000e+00, ptr [[TMP]], align 829; CHECK-NEXT: store double 4.000000e+00, ptr [[TMP5]], align 830; CHECK-NEXT: store double 5.000000e+00, ptr [[TMP4]], align 831; CHECK-NEXT: store double 6.000000e+00, ptr [[TMP3]], align 832; CHECK-NEXT: ret void33;34bb:35 %tmp = getelementptr inbounds %struct.ham, ptr %arg, i64 0, i32 0, i64 236 %tmp1 = getelementptr inbounds %struct.ham, ptr %arg, i64 0, i32 0, i64 137 %tmp3 = getelementptr inbounds %struct.ham, ptr %arg, i64 0,i32 1, i64 238 %tmp4 = getelementptr inbounds %struct.ham, ptr %arg, i64 0, i32 1, i64 139 %tmp5 = getelementptr inbounds %struct.ham, ptr %arg, i64 0, i32 1, i32 040 call void @llvm.memset.p0.i64(ptr nonnull align 8 dereferenceable(48) %arg, i8 0, i64 48, i1 false)41 br i1 %cond, label %bb7, label %bb842 43bb7: ; preds = %bb44 br label %bb945 46bb8: ; preds = %bb47 br label %bb948 49bb9: ; preds = %bb8, %bb750 store double 1.0, ptr %arg, align 851 store double 2.0, ptr %tmp1, align 852 store double 3.0, ptr %tmp, align 853 store double 4.0, ptr %tmp5, align 854 store double 5.0, ptr %tmp4, align 855 store double 6.0, ptr %tmp3, align 856 ret void57}58 59define void @overlap2(ptr %arg, i1 %cond) {60; CHECK-LABEL: @overlap2(61; CHECK-NEXT: bb:62; CHECK-NEXT: [[TMP:%.*]] = getelementptr inbounds [[STRUCT_HAM:%.*]], ptr [[ARG:%.*]], i64 0, i32 0, i64 263; CHECK-NEXT: [[TMP1:%.*]] = getelementptr inbounds [[STRUCT_HAM]], ptr [[ARG]], i64 0, i32 0, i64 164; CHECK-NEXT: [[TMP3:%.*]] = getelementptr inbounds [[STRUCT_HAM]], ptr [[ARG]], i64 0, i32 1, i64 265; CHECK-NEXT: [[TMP4:%.*]] = getelementptr inbounds [[STRUCT_HAM]], ptr [[ARG]], i64 0, i32 1, i64 166; CHECK-NEXT: [[TMP5:%.*]] = getelementptr inbounds [[STRUCT_HAM]], ptr [[ARG]], i64 0, i32 1, i32 067; CHECK-NEXT: call void @llvm.memset.p0.i64(ptr nonnull align 8 dereferenceable(48) [[ARG]], i8 0, i64 48, i1 false)68; CHECK-NEXT: br i1 [[COND:%.*]], label [[BB7:%.*]], label [[BB8:%.*]]69; CHECK: bb7:70; CHECK-NEXT: call void @may_throw()71; CHECK-NEXT: br label [[BB9:%.*]]72; CHECK: bb8:73; CHECK-NEXT: br label [[BB9]]74; CHECK: bb9:75; CHECK-NEXT: store double 1.000000e+00, ptr [[ARG]], align 876; CHECK-NEXT: store double 2.000000e+00, ptr [[TMP1]], align 877; CHECK-NEXT: store double 3.000000e+00, ptr [[TMP]], align 878; CHECK-NEXT: store double 4.000000e+00, ptr [[TMP5]], align 879; CHECK-NEXT: store double 5.000000e+00, ptr [[TMP4]], align 880; CHECK-NEXT: store double 6.000000e+00, ptr [[TMP3]], align 881; CHECK-NEXT: ret void82;83bb:84 %tmp = getelementptr inbounds %struct.ham, ptr %arg, i64 0, i32 0, i64 285 %tmp1 = getelementptr inbounds %struct.ham, ptr %arg, i64 0, i32 0, i64 186 %tmp3 = getelementptr inbounds %struct.ham, ptr %arg, i64 0,i32 1, i64 287 %tmp4 = getelementptr inbounds %struct.ham, ptr %arg, i64 0, i32 1, i64 188 %tmp5 = getelementptr inbounds %struct.ham, ptr %arg, i64 0, i32 1, i32 089 call void @llvm.memset.p0.i64(ptr nonnull align 8 dereferenceable(48) %arg, i8 0, i64 48, i1 false)90 br i1 %cond, label %bb7, label %bb891 92bb7: ; preds = %bb93 call void @may_throw()94 br label %bb995 96bb8: ; preds = %bb97 br label %bb998 99bb9: ; preds = %bb8, %bb7100 store double 1.0, ptr %arg, align 8101 store double 2.0, ptr %tmp1, align 8102 store double 3.0, ptr %tmp, align 8103 store double 4.0, ptr %tmp5, align 8104 store double 5.0, ptr %tmp4, align 8105 store double 6.0, ptr %tmp3, align 8106 ret void107}108 109; Test case from PR46513. Make sure we do not crash.110; TODO: we should be able to shorten store i32 844283136, ptr %cast.i32 to a111; store of i16.112define void @overlap_no_dominance(ptr %arg, i1 %c) {113; CHECK-LABEL: @overlap_no_dominance(114; CHECK-NEXT: bb:115; CHECK-NEXT: br i1 [[C:%.*]], label [[BB13:%.*]], label [[BB9:%.*]]116; CHECK: bb9:117; CHECK-NEXT: store i32 844283136, ptr [[ARG:%.*]], align 4118; CHECK-NEXT: br label [[BB13]]119; CHECK: bb13:120; CHECK-NEXT: store i16 0, ptr [[ARG]], align 4121; CHECK-NEXT: ret void122;123bb:124 br i1 %c, label %bb13, label %bb9125 126bb9: ; preds = %bb127 store i32 844283136, ptr %arg, align 4128 br label %bb13129 130bb13: ; preds = %bb9, %bb131 store i16 0, ptr %arg, align 4132 ret void133}134