280 lines · plain
1; NOTE: Assertions have been autogenerated by utils/update_test_checks.py UTC_ARGS: --version 22; RUN: opt -passes=loop-vectorize -force-vector-width=4 -force-vector-interleave=1 -S %s | FileCheck %s3 4target datalayout = "e-m:o-i64:64-i128:128-n32:64-S128"5 6%struct.Foo = type { [32000 x i32], [32000 x i32] }7 8@foo = global %struct.Foo zeroinitializer, align 49 10define void @accesses_to_struct_dereferenceable(ptr noalias %dst) {11; CHECK-LABEL: define void @accesses_to_struct_dereferenceable12; CHECK-SAME: (ptr noalias [[DST:%.*]]) {13; CHECK-NEXT: entry:14; CHECK-NEXT: br label [[VECTOR_PH:%.*]]15; CHECK: vector.ph:16; CHECK-NEXT: br label [[VECTOR_BODY:%.*]]17; CHECK: vector.body:18; CHECK-NEXT: [[INDEX:%.*]] = phi i64 [ 0, [[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], [[VECTOR_BODY]] ]19; CHECK-NEXT: [[TMP0:%.*]] = getelementptr inbounds i32, ptr [[DST]], i64 [[INDEX]]20; CHECK-NEXT: [[WIDE_LOAD:%.*]] = load <4 x i32>, ptr [[TMP0]], align 421; CHECK-NEXT: [[TMP1:%.*]] = icmp ult <4 x i32> [[WIDE_LOAD]], zeroinitializer22; CHECK-NEXT: [[TMP2:%.*]] = getelementptr [[STRUCT_FOO:%.*]], ptr @foo, i64 0, i32 1, i64 [[INDEX]]23; CHECK-NEXT: [[WIDE_LOAD1:%.*]] = load <4 x i32>, ptr [[TMP2]], align 424; CHECK-NEXT: [[TMP3:%.*]] = getelementptr [[STRUCT_FOO]], ptr @foo, i64 0, i32 0, i64 [[INDEX]]25; CHECK-NEXT: [[WIDE_LOAD2:%.*]] = load <4 x i32>, ptr [[TMP3]], align 426; CHECK-NEXT: [[PREDPHI:%.*]] = select <4 x i1> [[TMP1]], <4 x i32> [[WIDE_LOAD2]], <4 x i32> [[WIDE_LOAD1]]27; CHECK-NEXT: store <4 x i32> [[PREDPHI]], ptr [[TMP0]], align 428; CHECK-NEXT: [[INDEX_NEXT]] = add nuw i64 [[INDEX]], 429; CHECK-NEXT: [[TMP4:%.*]] = icmp eq i64 [[INDEX_NEXT]], 3200030; CHECK-NEXT: br i1 [[TMP4]], label [[MIDDLE_BLOCK:%.*]], label [[VECTOR_BODY]], !llvm.loop [[LOOP0:![0-9]+]]31; CHECK: middle.block:32; CHECK-NEXT: br label [[EXIT:%.*]]33; CHECK: exit:34; CHECK-NEXT: ret void35;36entry:37 br label %loop.header38 39loop.header:40 %iv = phi i64 [ 0, %entry ], [ %iv.next, %loop.latch ]41 %gep.dst = getelementptr inbounds i32, ptr %dst, i64 %iv42 %d = load i32, ptr %gep.dst, align 443 %cmp3 = icmp ult i32 %d, 044 br i1 %cmp3, label %if.then, label %if.else45 46if.then:47 %gep.a = getelementptr inbounds %struct.Foo, ptr @foo, i64 0, i32 0, i64 %iv48 %l.a = load i32, ptr %gep.a, align 449 br label %loop.latch50 51if.else:52 %gep.b = getelementptr inbounds %struct.Foo, ptr @foo, i64 0, i32 1, i64 %iv53 %l.b = load i32, ptr %gep.b, align 454 br label %loop.latch55 56loop.latch:57 %tmp.0 = phi i32 [ %l.a, %if.then ], [ %l.b, %if.else ]58 store i32 %tmp.0, ptr %gep.dst, align 459 %iv.next = add nuw nsw i64 %iv, 160 %exitcond.not = icmp eq i64 %iv.next, 3200061 br i1 %exitcond.not, label %exit, label %loop.header62 63exit:64 ret void65}66 67define void @accesses_to_struct_may_not_be_dereferenceable_due_to_loop_bound(ptr noalias %dst) {68; CHECK-LABEL: define void @accesses_to_struct_may_not_be_dereferenceable_due_to_loop_bound69; CHECK-SAME: (ptr noalias [[DST:%.*]]) {70; CHECK-NEXT: entry:71; CHECK-NEXT: br label [[VECTOR_PH:%.*]]72; CHECK: vector.ph:73; CHECK-NEXT: br label [[VECTOR_BODY:%.*]]74; CHECK: vector.body:75; CHECK-NEXT: [[INDEX:%.*]] = phi i64 [ 0, [[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], [[PRED_LOAD_CONTINUE6:%.*]] ]76; CHECK-NEXT: [[TMP0:%.*]] = getelementptr inbounds i32, ptr [[DST]], i64 [[INDEX]]77; CHECK-NEXT: [[WIDE_LOAD:%.*]] = load <4 x i32>, ptr [[TMP0]], align 478; CHECK-NEXT: [[TMP1:%.*]] = icmp uge <4 x i32> [[WIDE_LOAD]], zeroinitializer79; CHECK-NEXT: [[TMP2:%.*]] = extractelement <4 x i1> [[TMP1]], i32 080; CHECK-NEXT: br i1 [[TMP2]], label [[PRED_LOAD_IF:%.*]], label [[PRED_LOAD_CONTINUE:%.*]]81; CHECK: pred.load.if:82; CHECK-NEXT: [[TMP3:%.*]] = add i64 [[INDEX]], 083; CHECK-NEXT: [[TMP4:%.*]] = getelementptr inbounds [[STRUCT_FOO:%.*]], ptr @foo, i64 0, i32 1, i64 [[TMP3]]84; CHECK-NEXT: [[TMP5:%.*]] = load i32, ptr [[TMP4]], align 485; CHECK-NEXT: [[TMP6:%.*]] = insertelement <4 x i32> poison, i32 [[TMP5]], i32 086; CHECK-NEXT: br label [[PRED_LOAD_CONTINUE]]87; CHECK: pred.load.continue:88; CHECK-NEXT: [[TMP7:%.*]] = phi <4 x i32> [ poison, [[VECTOR_BODY]] ], [ [[TMP6]], [[PRED_LOAD_IF]] ]89; CHECK-NEXT: [[TMP8:%.*]] = extractelement <4 x i1> [[TMP1]], i32 190; CHECK-NEXT: br i1 [[TMP8]], label [[PRED_LOAD_IF1:%.*]], label [[PRED_LOAD_CONTINUE2:%.*]]91; CHECK: pred.load.if1:92; CHECK-NEXT: [[TMP9:%.*]] = add i64 [[INDEX]], 193; CHECK-NEXT: [[TMP10:%.*]] = getelementptr inbounds [[STRUCT_FOO]], ptr @foo, i64 0, i32 1, i64 [[TMP9]]94; CHECK-NEXT: [[TMP11:%.*]] = load i32, ptr [[TMP10]], align 495; CHECK-NEXT: [[TMP12:%.*]] = insertelement <4 x i32> [[TMP7]], i32 [[TMP11]], i32 196; CHECK-NEXT: br label [[PRED_LOAD_CONTINUE2]]97; CHECK: pred.load.continue2:98; CHECK-NEXT: [[TMP13:%.*]] = phi <4 x i32> [ [[TMP7]], [[PRED_LOAD_CONTINUE]] ], [ [[TMP12]], [[PRED_LOAD_IF1]] ]99; CHECK-NEXT: [[TMP14:%.*]] = extractelement <4 x i1> [[TMP1]], i32 2100; CHECK-NEXT: br i1 [[TMP14]], label [[PRED_LOAD_IF3:%.*]], label [[PRED_LOAD_CONTINUE4:%.*]]101; CHECK: pred.load.if3:102; CHECK-NEXT: [[TMP15:%.*]] = add i64 [[INDEX]], 2103; CHECK-NEXT: [[TMP16:%.*]] = getelementptr inbounds [[STRUCT_FOO]], ptr @foo, i64 0, i32 1, i64 [[TMP15]]104; CHECK-NEXT: [[TMP17:%.*]] = load i32, ptr [[TMP16]], align 4105; CHECK-NEXT: [[TMP18:%.*]] = insertelement <4 x i32> [[TMP13]], i32 [[TMP17]], i32 2106; CHECK-NEXT: br label [[PRED_LOAD_CONTINUE4]]107; CHECK: pred.load.continue4:108; CHECK-NEXT: [[TMP19:%.*]] = phi <4 x i32> [ [[TMP13]], [[PRED_LOAD_CONTINUE2]] ], [ [[TMP18]], [[PRED_LOAD_IF3]] ]109; CHECK-NEXT: [[TMP20:%.*]] = extractelement <4 x i1> [[TMP1]], i32 3110; CHECK-NEXT: br i1 [[TMP20]], label [[PRED_LOAD_IF5:%.*]], label [[PRED_LOAD_CONTINUE6]]111; CHECK: pred.load.if5:112; CHECK-NEXT: [[TMP21:%.*]] = add i64 [[INDEX]], 3113; CHECK-NEXT: [[TMP22:%.*]] = getelementptr inbounds [[STRUCT_FOO]], ptr @foo, i64 0, i32 1, i64 [[TMP21]]114; CHECK-NEXT: [[TMP23:%.*]] = load i32, ptr [[TMP22]], align 4115; CHECK-NEXT: [[TMP24:%.*]] = insertelement <4 x i32> [[TMP19]], i32 [[TMP23]], i32 3116; CHECK-NEXT: br label [[PRED_LOAD_CONTINUE6]]117; CHECK: pred.load.continue6:118; CHECK-NEXT: [[TMP25:%.*]] = phi <4 x i32> [ [[TMP19]], [[PRED_LOAD_CONTINUE4]] ], [ [[TMP24]], [[PRED_LOAD_IF5]] ]119; CHECK-NEXT: [[TMP26:%.*]] = getelementptr [[STRUCT_FOO]], ptr @foo, i64 0, i32 0, i64 [[INDEX]]120; CHECK-NEXT: [[WIDE_LOAD7:%.*]] = load <4 x i32>, ptr [[TMP26]], align 4121; CHECK-NEXT: [[PREDPHI:%.*]] = select <4 x i1> [[TMP1]], <4 x i32> [[TMP25]], <4 x i32> [[WIDE_LOAD7]]122; CHECK-NEXT: store <4 x i32> [[PREDPHI]], ptr [[TMP0]], align 4123; CHECK-NEXT: [[INDEX_NEXT]] = add nuw i64 [[INDEX]], 4124; CHECK-NEXT: [[TMP27:%.*]] = icmp eq i64 [[INDEX_NEXT]], 32000125; CHECK-NEXT: br i1 [[TMP27]], label [[MIDDLE_BLOCK:%.*]], label [[VECTOR_BODY]], !llvm.loop [[LOOP3:![0-9]+]]126; CHECK: middle.block:127; CHECK-NEXT: br label [[SCALAR_PH:%.*]]128; CHECK: scalar.ph:129; CHECK-NEXT: br label [[LOOP_HEADER:%.*]]130; CHECK: loop.header:131; CHECK-NEXT: [[IV:%.*]] = phi i64 [ 32000, [[SCALAR_PH]] ], [ [[IV_NEXT:%.*]], [[LOOP_LATCH:%.*]] ]132; CHECK-NEXT: [[GEP_DST:%.*]] = getelementptr inbounds i32, ptr [[DST]], i64 [[IV]]133; CHECK-NEXT: [[D:%.*]] = load i32, ptr [[GEP_DST]], align 4134; CHECK-NEXT: [[CMP3:%.*]] = icmp ult i32 [[D]], 0135; CHECK-NEXT: br i1 [[CMP3]], label [[IF_THEN:%.*]], label [[IF_ELSE:%.*]]136; CHECK: if.then:137; CHECK-NEXT: [[GEP_A:%.*]] = getelementptr inbounds [[STRUCT_FOO]], ptr @foo, i64 0, i32 0, i64 [[IV]]138; CHECK-NEXT: [[L_A:%.*]] = load i32, ptr [[GEP_A]], align 4139; CHECK-NEXT: br label [[LOOP_LATCH]]140; CHECK: if.else:141; CHECK-NEXT: [[GEP_B:%.*]] = getelementptr inbounds [[STRUCT_FOO]], ptr @foo, i64 0, i32 1, i64 [[IV]]142; CHECK-NEXT: [[L_B:%.*]] = load i32, ptr [[GEP_B]], align 4143; CHECK-NEXT: br label [[LOOP_LATCH]]144; CHECK: loop.latch:145; CHECK-NEXT: [[TMP_0:%.*]] = phi i32 [ [[L_A]], [[IF_THEN]] ], [ [[L_B]], [[IF_ELSE]] ]146; CHECK-NEXT: store i32 [[TMP_0]], ptr [[GEP_DST]], align 4147; CHECK-NEXT: [[IV_NEXT]] = add nuw nsw i64 [[IV]], 1148; CHECK-NEXT: [[EXITCOND_NOT:%.*]] = icmp eq i64 [[IV_NEXT]], 32001149; CHECK-NEXT: br i1 [[EXITCOND_NOT]], label [[EXIT:%.*]], label [[LOOP_HEADER]], !llvm.loop [[LOOP4:![0-9]+]]150; CHECK: exit:151; CHECK-NEXT: ret void152;153entry:154 br label %loop.header155 156loop.header:157 %iv = phi i64 [ 0, %entry ], [ %iv.next, %loop.latch ]158 %gep.dst = getelementptr inbounds i32, ptr %dst, i64 %iv159 %d = load i32, ptr %gep.dst, align 4160 %cmp3 = icmp ult i32 %d, 0161 br i1 %cmp3, label %if.then, label %if.else162 163if.then:164 %gep.a = getelementptr inbounds %struct.Foo, ptr @foo, i64 0, i32 0, i64 %iv165 %l.a = load i32, ptr %gep.a, align 4166 br label %loop.latch167 168if.else:169 %gep.b = getelementptr inbounds %struct.Foo, ptr @foo, i64 0, i32 1, i64 %iv170 %l.b = load i32, ptr %gep.b, align 4171 br label %loop.latch172 173loop.latch:174 %tmp.0 = phi i32 [ %l.a, %if.then ], [ %l.b, %if.else ]175 store i32 %tmp.0, ptr %gep.dst, align 4176 %iv.next = add nuw nsw i64 %iv, 1177 %exitcond.not = icmp eq i64 %iv.next, 32001178 br i1 %exitcond.not, label %exit, label %loop.header179 180exit:181 ret void182}183 184define void @accesses_to_struct_may_not_be_dereferenceable_access_size(ptr noalias %dst) {185; CHECK-LABEL: define void @accesses_to_struct_may_not_be_dereferenceable_access_size186; CHECK-SAME: (ptr noalias [[DST:%.*]]) {187; CHECK-NEXT: entry:188; CHECK-NEXT: br label [[VECTOR_PH:%.*]]189; CHECK: vector.ph:190; CHECK-NEXT: br label [[VECTOR_BODY:%.*]]191; CHECK: vector.body:192; CHECK-NEXT: [[INDEX:%.*]] = phi i64 [ 0, [[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], [[PRED_LOAD_CONTINUE6:%.*]] ]193; CHECK-NEXT: [[TMP0:%.*]] = getelementptr inbounds i32, ptr [[DST]], i64 [[INDEX]]194; CHECK-NEXT: [[WIDE_LOAD:%.*]] = load <4 x i32>, ptr [[TMP0]], align 4195; CHECK-NEXT: [[TMP1:%.*]] = icmp uge <4 x i32> [[WIDE_LOAD]], zeroinitializer196; CHECK-NEXT: [[TMP2:%.*]] = extractelement <4 x i1> [[TMP1]], i32 0197; CHECK-NEXT: br i1 [[TMP2]], label [[PRED_LOAD_IF:%.*]], label [[PRED_LOAD_CONTINUE:%.*]]198; CHECK: pred.load.if:199; CHECK-NEXT: [[TMP3:%.*]] = add i64 [[INDEX]], 0200; CHECK-NEXT: [[TMP4:%.*]] = getelementptr inbounds [[STRUCT_FOO:%.*]], ptr @foo, i64 0, i32 1, i64 [[TMP3]]201; CHECK-NEXT: [[TMP5:%.*]] = load i64, ptr [[TMP4]], align 4202; CHECK-NEXT: [[TMP6:%.*]] = insertelement <4 x i64> poison, i64 [[TMP5]], i32 0203; CHECK-NEXT: br label [[PRED_LOAD_CONTINUE]]204; CHECK: pred.load.continue:205; CHECK-NEXT: [[TMP7:%.*]] = phi <4 x i64> [ poison, [[VECTOR_BODY]] ], [ [[TMP6]], [[PRED_LOAD_IF]] ]206; CHECK-NEXT: [[TMP8:%.*]] = extractelement <4 x i1> [[TMP1]], i32 1207; CHECK-NEXT: br i1 [[TMP8]], label [[PRED_LOAD_IF1:%.*]], label [[PRED_LOAD_CONTINUE2:%.*]]208; CHECK: pred.load.if1:209; CHECK-NEXT: [[TMP9:%.*]] = add i64 [[INDEX]], 1210; CHECK-NEXT: [[TMP10:%.*]] = getelementptr inbounds [[STRUCT_FOO]], ptr @foo, i64 0, i32 1, i64 [[TMP9]]211; CHECK-NEXT: [[TMP11:%.*]] = load i64, ptr [[TMP10]], align 4212; CHECK-NEXT: [[TMP12:%.*]] = insertelement <4 x i64> [[TMP7]], i64 [[TMP11]], i32 1213; CHECK-NEXT: br label [[PRED_LOAD_CONTINUE2]]214; CHECK: pred.load.continue2:215; CHECK-NEXT: [[TMP13:%.*]] = phi <4 x i64> [ [[TMP7]], [[PRED_LOAD_CONTINUE]] ], [ [[TMP12]], [[PRED_LOAD_IF1]] ]216; CHECK-NEXT: [[TMP14:%.*]] = extractelement <4 x i1> [[TMP1]], i32 2217; CHECK-NEXT: br i1 [[TMP14]], label [[PRED_LOAD_IF3:%.*]], label [[PRED_LOAD_CONTINUE4:%.*]]218; CHECK: pred.load.if3:219; CHECK-NEXT: [[TMP15:%.*]] = add i64 [[INDEX]], 2220; CHECK-NEXT: [[TMP16:%.*]] = getelementptr inbounds [[STRUCT_FOO]], ptr @foo, i64 0, i32 1, i64 [[TMP15]]221; CHECK-NEXT: [[TMP17:%.*]] = load i64, ptr [[TMP16]], align 4222; CHECK-NEXT: [[TMP18:%.*]] = insertelement <4 x i64> [[TMP13]], i64 [[TMP17]], i32 2223; CHECK-NEXT: br label [[PRED_LOAD_CONTINUE4]]224; CHECK: pred.load.continue4:225; CHECK-NEXT: [[TMP19:%.*]] = phi <4 x i64> [ [[TMP13]], [[PRED_LOAD_CONTINUE2]] ], [ [[TMP18]], [[PRED_LOAD_IF3]] ]226; CHECK-NEXT: [[TMP20:%.*]] = extractelement <4 x i1> [[TMP1]], i32 3227; CHECK-NEXT: br i1 [[TMP20]], label [[PRED_LOAD_IF5:%.*]], label [[PRED_LOAD_CONTINUE6]]228; CHECK: pred.load.if5:229; CHECK-NEXT: [[TMP21:%.*]] = add i64 [[INDEX]], 3230; CHECK-NEXT: [[TMP22:%.*]] = getelementptr inbounds [[STRUCT_FOO]], ptr @foo, i64 0, i32 1, i64 [[TMP21]]231; CHECK-NEXT: [[TMP23:%.*]] = load i64, ptr [[TMP22]], align 4232; CHECK-NEXT: [[TMP24:%.*]] = insertelement <4 x i64> [[TMP19]], i64 [[TMP23]], i32 3233; CHECK-NEXT: br label [[PRED_LOAD_CONTINUE6]]234; CHECK: pred.load.continue6:235; CHECK-NEXT: [[TMP25:%.*]] = phi <4 x i64> [ [[TMP19]], [[PRED_LOAD_CONTINUE4]] ], [ [[TMP24]], [[PRED_LOAD_IF5]] ]236; CHECK-NEXT: [[TMP26:%.*]] = trunc <4 x i64> [[TMP25]] to <4 x i32>237; CHECK-NEXT: [[TMP27:%.*]] = getelementptr [[STRUCT_FOO]], ptr @foo, i64 0, i32 0, i64 [[INDEX]]238; CHECK-NEXT: [[WIDE_LOAD7:%.*]] = load <4 x i32>, ptr [[TMP27]], align 4239; CHECK-NEXT: [[PREDPHI:%.*]] = select <4 x i1> [[TMP1]], <4 x i32> [[TMP26]], <4 x i32> [[WIDE_LOAD7]]240; CHECK-NEXT: store <4 x i32> [[PREDPHI]], ptr [[TMP0]], align 4241; CHECK-NEXT: [[INDEX_NEXT]] = add nuw i64 [[INDEX]], 4242; CHECK-NEXT: [[TMP28:%.*]] = icmp eq i64 [[INDEX_NEXT]], 32000243; CHECK-NEXT: br i1 [[TMP28]], label [[MIDDLE_BLOCK:%.*]], label [[VECTOR_BODY]], !llvm.loop [[LOOP5:![0-9]+]]244; CHECK: middle.block:245; CHECK-NEXT: br label [[EXIT:%.*]]246; CHECK: exit:247; CHECK-NEXT: ret void248;249entry:250 br label %loop.header251 252loop.header:253 %iv = phi i64 [ 0, %entry ], [ %iv.next, %loop.latch ]254 %gep.dst = getelementptr inbounds i32, ptr %dst, i64 %iv255 %d = load i32, ptr %gep.dst, align 4256 %cmp3 = icmp ult i32 %d, 0257 br i1 %cmp3, label %if.then, label %if.else258 259if.then:260 %gep.a = getelementptr inbounds %struct.Foo, ptr @foo, i64 0, i32 0, i64 %iv261 %l.a = load i32, ptr %gep.a, align 4262 br label %loop.latch263 264if.else:265 %gep.b = getelementptr inbounds %struct.Foo, ptr @foo, i64 0, i32 1, i64 %iv266 %l.b = load i64, ptr %gep.b, align 4267 %t = trunc i64 %l.b to i32268 br label %loop.latch269 270loop.latch:271 %tmp.0 = phi i32 [ %l.a, %if.then ], [ %t, %if.else ]272 store i32 %tmp.0, ptr %gep.dst, align 4273 %iv.next = add nuw nsw i64 %iv, 1274 %exitcond.not = icmp eq i64 %iv.next, 32000275 br i1 %exitcond.not, label %exit, label %loop.header276 277exit:278 ret void279}280