806 lines · plain
1; NOTE: Assertions have been autogenerated by utils/update_test_checks.py UTC_ARGS: --prefix-filecheck-ir-name unnamed --version 52; RUN: opt -S -passes=loop-vectorize -force-vector-interleave=1 -force-vector-width=2 < %s 2>&1 | FileCheck %s3 4target datalayout = "e-p:32:32:32-i1:8:8-i8:8:8-i16:16:16-i32:32:32-i64:32:64-f32:32:32-f64:32:64-v64:64:64-v128:128:128-a0:0:64-f80:32:32-n8:16:32-S128"5 6@f = common global i32 0, align 47@.str = private unnamed_addr constant [4 x i8] c"%d\0A\00", align 18@c = common global i32 0, align 49@a = common global i32 0, align 410@b = common global i32 0, align 411@e = common global i32 0, align 412 13; It has a value that is used outside of the loop14; and is not a recognized reduction variable "tmp17".15; However, tmp17 is a non-header phi which is an allowed exit.16 17 18define i32 @test1() {19; CHECK-LABEL: define i32 @test1() {20; CHECK-NEXT: [[BB:.*]]:21; CHECK-NEXT: [[B_PROMOTED:%.*]] = load i32, ptr @b, align 422; CHECK-NEXT: [[TMP0:%.*]] = add i32 [[B_PROMOTED]], 123; CHECK-NEXT: [[SMAX:%.*]] = call i32 @llvm.smax.i32(i32 [[TMP0]], i32 4)24; CHECK-NEXT: [[TMP1:%.*]] = sub i32 [[SMAX]], [[B_PROMOTED]]25; CHECK-NEXT: [[MIN_ITERS_CHECK:%.*]] = icmp ult i32 [[TMP1]], 226; CHECK-NEXT: br i1 [[MIN_ITERS_CHECK]], label %[[_LR_PH_I1:.*]], label %[[VECTOR_PH:.*]]27; CHECK: [[VECTOR_PH]]:28; CHECK-NEXT: [[N_MOD_VF:%.*]] = urem i32 [[TMP1]], 229; CHECK-NEXT: [[N_VEC:%.*]] = sub i32 [[TMP1]], [[N_MOD_VF]]30; CHECK-NEXT: [[TMP2:%.*]] = add i32 [[B_PROMOTED]], [[N_VEC]]31; CHECK-NEXT: [[BROADCAST_SPLATINSERT:%.*]] = insertelement <2 x i32> poison, i32 [[B_PROMOTED]], i64 032; CHECK-NEXT: [[BROADCAST_SPLAT:%.*]] = shufflevector <2 x i32> [[BROADCAST_SPLATINSERT]], <2 x i32> poison, <2 x i32> zeroinitializer33; CHECK-NEXT: [[INDUCTION:%.*]] = add nsw <2 x i32> [[BROADCAST_SPLAT]], <i32 0, i32 1>34; CHECK-NEXT: br label %[[VECTOR_BODY:.*]]35; CHECK: [[VECTOR_BODY]]:36; CHECK-NEXT: [[INDEX:%.*]] = phi i32 [ 0, %[[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], %[[VECTOR_BODY]] ]37; CHECK-NEXT: [[VEC_IND:%.*]] = phi <2 x i32> [ [[INDUCTION]], %[[VECTOR_PH]] ], [ [[VEC_IND_NEXT:%.*]], %[[VECTOR_BODY]] ]38; CHECK-NEXT: [[TMP3:%.*]] = icmp sgt <2 x i32> [[VEC_IND]], splat (i32 10)39; CHECK-NEXT: [[PREDPHI:%.*]] = select <2 x i1> [[TMP3]], <2 x i32> splat (i32 1), <2 x i32> zeroinitializer40; CHECK-NEXT: [[INDEX_NEXT]] = add nuw i32 [[INDEX]], 241; CHECK-NEXT: [[VEC_IND_NEXT]] = add nsw <2 x i32> [[VEC_IND]], splat (i32 2)42; CHECK-NEXT: [[TMP4:%.*]] = icmp eq i32 [[INDEX_NEXT]], [[N_VEC]]43; CHECK-NEXT: br i1 [[TMP4]], label %[[MIDDLE_BLOCK:.*]], label %[[VECTOR_BODY]], !llvm.loop [[LOOP0:![0-9]+]]44; CHECK: [[MIDDLE_BLOCK]]:45; CHECK-NEXT: [[TMP5:%.*]] = extractelement <2 x i32> [[PREDPHI]], i32 146; CHECK-NEXT: [[CMP_N:%.*]] = icmp eq i32 [[TMP1]], [[N_VEC]]47; CHECK-NEXT: br i1 [[CMP_N]], label %[[F1_EXIT_LOOPEXIT:.*]], label %[[_LR_PH_I1]]48; CHECK: [[_LR_PH_I1]]:49; CHECK-NEXT: [[BC_RESUME_VAL:%.*]] = phi i32 [ [[TMP2]], %[[MIDDLE_BLOCK]] ], [ [[B_PROMOTED]], %[[BB]] ]50; CHECK-NEXT: br label %[[DOTLR_PH_I:.*]]51; CHECK: [[_LR_PH_I:.*:]]52; CHECK-NEXT: [[UNNAMEDTMP8:%.*]] = phi i32 [ [[UNNAMEDTMP18:%.*]], %[[BB16:.*]] ], [ [[BC_RESUME_VAL]], %[[_LR_PH_I1]] ]53; CHECK-NEXT: [[UNNAMEDTMP2:%.*]] = icmp sgt i32 [[UNNAMEDTMP8]], 1054; CHECK-NEXT: br i1 [[UNNAMEDTMP2]], label %[[BB16]], label %[[UNNAMEDBB10:.*]]55; CHECK: [[UNNAMEDBB10]]:56; CHECK-NEXT: br label %[[BB16]]57; CHECK: [[BB16]]:58; CHECK-NEXT: [[UNNAMEDTMP17:%.*]] = phi i32 [ 0, %[[UNNAMEDBB10]] ], [ 1, %[[DOTLR_PH_I]] ]59; CHECK-NEXT: [[UNNAMEDTMP18]] = add nsw i32 [[UNNAMEDTMP8]], 160; CHECK-NEXT: [[UNNAMEDTMP19:%.*]] = icmp slt i32 [[UNNAMEDTMP18]], 461; CHECK-NEXT: br i1 [[UNNAMEDTMP19]], label %[[DOTLR_PH_I]], label %[[F1_EXIT_LOOPEXIT]], !llvm.loop [[LOOP3:![0-9]+]]62; CHECK: [[F1_EXIT_LOOPEXIT]]:63; CHECK-NEXT: [[DOTLCSSA:%.*]] = phi i32 [ [[UNNAMEDTMP17]], %[[BB16]] ], [ [[TMP5]], %[[MIDDLE_BLOCK]] ]64; CHECK-NEXT: ret i32 [[DOTLCSSA]]65;66bb:67 %b.promoted = load i32, ptr @b, align 468 br label %.lr.ph.i69 70.lr.ph.i:71 %tmp8 = phi i32 [ %tmp18, %bb16 ], [ %b.promoted, %bb ]72 %tmp2 = icmp sgt i32 %tmp8, 1073 br i1 %tmp2, label %bb16, label %bb1074 75bb10:76 br label %bb1677 78bb16:79 %tmp17 = phi i32 [ 0, %bb10 ], [ 1, %.lr.ph.i ]80 %tmp18 = add nsw i32 %tmp8, 181 %tmp19 = icmp slt i32 %tmp18, 482 br i1 %tmp19, label %.lr.ph.i, label %f1.exit.loopexit83 84f1.exit.loopexit:85 %.lcssa = phi i32 [ %tmp17, %bb16 ]86 ret i32 %.lcssa87}88 89; non-hdr phi depends on header phi.90 91define i32 @test2() {92; CHECK-LABEL: define i32 @test2() {93; CHECK-NEXT: [[BB:.*]]:94; CHECK-NEXT: [[B_PROMOTED:%.*]] = load i32, ptr @b, align 495; CHECK-NEXT: [[TMP0:%.*]] = add i32 [[B_PROMOTED]], 196; CHECK-NEXT: [[SMAX:%.*]] = call i32 @llvm.smax.i32(i32 [[TMP0]], i32 4)97; CHECK-NEXT: [[TMP1:%.*]] = sub i32 [[SMAX]], [[B_PROMOTED]]98; CHECK-NEXT: [[MIN_ITERS_CHECK:%.*]] = icmp ult i32 [[TMP1]], 299; CHECK-NEXT: br i1 [[MIN_ITERS_CHECK]], label %[[_LR_PH_I1:.*]], label %[[VECTOR_PH:.*]]100; CHECK: [[VECTOR_PH]]:101; CHECK-NEXT: [[N_MOD_VF:%.*]] = urem i32 [[TMP1]], 2102; CHECK-NEXT: [[N_VEC:%.*]] = sub i32 [[TMP1]], [[N_MOD_VF]]103; CHECK-NEXT: [[TMP2:%.*]] = add i32 [[B_PROMOTED]], [[N_VEC]]104; CHECK-NEXT: [[DOTSPLATINSERT:%.*]] = insertelement <2 x i32> poison, i32 [[B_PROMOTED]], i64 0105; CHECK-NEXT: [[DOTSPLAT:%.*]] = shufflevector <2 x i32> [[DOTSPLATINSERT]], <2 x i32> poison, <2 x i32> zeroinitializer106; CHECK-NEXT: [[INDUCTION:%.*]] = add nsw <2 x i32> [[DOTSPLAT]], <i32 0, i32 1>107; CHECK-NEXT: br label %[[VECTOR_BODY:.*]]108; CHECK: [[VECTOR_BODY]]:109; CHECK-NEXT: [[INDEX:%.*]] = phi i32 [ 0, %[[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], %[[VECTOR_BODY]] ]110; CHECK-NEXT: [[VEC_IND:%.*]] = phi <2 x i32> [ [[INDUCTION]], %[[VECTOR_PH]] ], [ [[VEC_IND_NEXT:%.*]], %[[VECTOR_BODY]] ]111; CHECK-NEXT: [[TMP3:%.*]] = icmp sgt <2 x i32> [[VEC_IND]], splat (i32 10)112; CHECK-NEXT: [[PREDPHI:%.*]] = select <2 x i1> [[TMP3]], <2 x i32> splat (i32 1), <2 x i32> [[VEC_IND]]113; CHECK-NEXT: [[INDEX_NEXT]] = add nuw i32 [[INDEX]], 2114; CHECK-NEXT: [[VEC_IND_NEXT]] = add nsw <2 x i32> [[VEC_IND]], splat (i32 2)115; CHECK-NEXT: [[TMP4:%.*]] = icmp eq i32 [[INDEX_NEXT]], [[N_VEC]]116; CHECK-NEXT: br i1 [[TMP4]], label %[[MIDDLE_BLOCK:.*]], label %[[VECTOR_BODY]], !llvm.loop [[LOOP4:![0-9]+]]117; CHECK: [[MIDDLE_BLOCK]]:118; CHECK-NEXT: [[TMP5:%.*]] = extractelement <2 x i32> [[PREDPHI]], i32 1119; CHECK-NEXT: [[CMP_N:%.*]] = icmp eq i32 [[TMP1]], [[N_VEC]]120; CHECK-NEXT: br i1 [[CMP_N]], label %[[F1_EXIT_LOOPEXIT:.*]], label %[[_LR_PH_I1]]121; CHECK: [[_LR_PH_I1]]:122; CHECK-NEXT: [[BC_RESUME_VAL:%.*]] = phi i32 [ [[TMP2]], %[[MIDDLE_BLOCK]] ], [ [[B_PROMOTED]], %[[BB]] ]123; CHECK-NEXT: br label %[[DOTLR_PH_I:.*]]124; CHECK: [[_LR_PH_I:.*:]]125; CHECK-NEXT: [[UNNAMEDTMP8:%.*]] = phi i32 [ [[UNNAMEDTMP18:%.*]], %[[BB16:.*]] ], [ [[BC_RESUME_VAL]], %[[_LR_PH_I1]] ]126; CHECK-NEXT: [[UNNAMEDTMP2:%.*]] = icmp sgt i32 [[UNNAMEDTMP8]], 10127; CHECK-NEXT: br i1 [[UNNAMEDTMP2]], label %[[BB16]], label %[[UNNAMEDBB10:.*]]128; CHECK: [[UNNAMEDBB10]]:129; CHECK-NEXT: br label %[[BB16]]130; CHECK: [[BB16]]:131; CHECK-NEXT: [[UNNAMEDTMP17:%.*]] = phi i32 [ [[UNNAMEDTMP8]], %[[UNNAMEDBB10]] ], [ 1, %[[DOTLR_PH_I]] ]132; CHECK-NEXT: [[UNNAMEDTMP18]] = add nsw i32 [[UNNAMEDTMP8]], 1133; CHECK-NEXT: [[UNNAMEDTMP19:%.*]] = icmp slt i32 [[UNNAMEDTMP18]], 4134; CHECK-NEXT: br i1 [[UNNAMEDTMP19]], label %[[DOTLR_PH_I]], label %[[F1_EXIT_LOOPEXIT]], !llvm.loop [[LOOP5:![0-9]+]]135; CHECK: [[F1_EXIT_LOOPEXIT]]:136; CHECK-NEXT: [[DOTLCSSA:%.*]] = phi i32 [ [[UNNAMEDTMP17]], %[[BB16]] ], [ [[TMP5]], %[[MIDDLE_BLOCK]] ]137; CHECK-NEXT: ret i32 [[DOTLCSSA]]138;139bb:140 %b.promoted = load i32, ptr @b, align 4141 br label %.lr.ph.i142 143.lr.ph.i:144 %tmp8 = phi i32 [ %tmp18, %bb16 ], [ %b.promoted, %bb ]145 %tmp2 = icmp sgt i32 %tmp8, 10146 br i1 %tmp2, label %bb16, label %bb10147 148bb10:149 br label %bb16150 151bb16:152 %tmp17 = phi i32 [ %tmp8, %bb10 ], [ 1, %.lr.ph.i ]153 %tmp18 = add nsw i32 %tmp8, 1154 %tmp19 = icmp slt i32 %tmp18, 4155 br i1 %tmp19, label %.lr.ph.i, label %f1.exit.loopexit156 157f1.exit.loopexit:158 %.lcssa = phi i32 [ %tmp17, %bb16 ]159 ret i32 %.lcssa160}161 162; more than 2 incoming values for tmp17 phi that is used outside loop.163define i32 @test3(i32 %N) {164; CHECK-LABEL: define i32 @test3(165; CHECK-SAME: i32 [[N:%.*]]) {166; CHECK-NEXT: [[BB:.*]]:167; CHECK-NEXT: [[B_PROMOTED:%.*]] = load i32, ptr @b, align 4168; CHECK-NEXT: [[TMP0:%.*]] = add i32 [[B_PROMOTED]], 1169; CHECK-NEXT: [[SMAX:%.*]] = call i32 @llvm.smax.i32(i32 [[TMP0]], i32 4)170; CHECK-NEXT: [[TMP1:%.*]] = sub i32 [[SMAX]], [[B_PROMOTED]]171; CHECK-NEXT: [[MIN_ITERS_CHECK:%.*]] = icmp ult i32 [[TMP1]], 2172; CHECK-NEXT: br i1 [[MIN_ITERS_CHECK]], label %[[_LR_PH_I:.*]], label %[[VECTOR_PH:.*]]173; CHECK: [[VECTOR_PH]]:174; CHECK-NEXT: [[N_MOD_VF:%.*]] = urem i32 [[TMP1]], 2175; CHECK-NEXT: [[N_VEC:%.*]] = sub i32 [[TMP1]], [[N_MOD_VF]]176; CHECK-NEXT: [[TMP2:%.*]] = add i32 [[B_PROMOTED]], [[N_VEC]]177; CHECK-NEXT: [[BROADCAST_SPLATINSERT:%.*]] = insertelement <2 x i32> poison, i32 [[N]], i64 0178; CHECK-NEXT: [[BROADCAST_SPLAT:%.*]] = shufflevector <2 x i32> [[BROADCAST_SPLATINSERT]], <2 x i32> poison, <2 x i32> zeroinitializer179; CHECK-NEXT: [[DOTSPLATINSERT:%.*]] = insertelement <2 x i32> poison, i32 [[B_PROMOTED]], i64 0180; CHECK-NEXT: [[DOTSPLAT:%.*]] = shufflevector <2 x i32> [[DOTSPLATINSERT]], <2 x i32> poison, <2 x i32> zeroinitializer181; CHECK-NEXT: [[INDUCTION:%.*]] = add nsw <2 x i32> [[DOTSPLAT]], <i32 0, i32 1>182; CHECK-NEXT: br label %[[VECTOR_BODY:.*]]183; CHECK: [[VECTOR_BODY]]:184; CHECK-NEXT: [[INDEX:%.*]] = phi i32 [ 0, %[[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], %[[VECTOR_BODY]] ]185; CHECK-NEXT: [[VEC_IND:%.*]] = phi <2 x i32> [ [[INDUCTION]], %[[VECTOR_PH]] ], [ [[VEC_IND_NEXT:%.*]], %[[VECTOR_BODY]] ]186; CHECK-NEXT: [[TMP4:%.*]] = icmp sle <2 x i32> [[VEC_IND]], splat (i32 10)187; CHECK-NEXT: [[TMP5:%.*]] = icmp sle <2 x i32> [[VEC_IND]], [[BROADCAST_SPLAT]]188; CHECK-NEXT: [[TMP6:%.*]] = select <2 x i1> [[TMP4]], <2 x i1> [[TMP5]], <2 x i1> zeroinitializer189; CHECK-NEXT: [[PREDPHI:%.*]] = select <2 x i1> [[TMP6]], <2 x i32> zeroinitializer, <2 x i32> splat (i32 2)190; CHECK-NEXT: [[PREDPHI1:%.*]] = select <2 x i1> [[TMP4]], <2 x i32> [[PREDPHI]], <2 x i32> splat (i32 1)191; CHECK-NEXT: [[INDEX_NEXT]] = add nuw i32 [[INDEX]], 2192; CHECK-NEXT: [[VEC_IND_NEXT]] = add nsw <2 x i32> [[VEC_IND]], splat (i32 2)193; CHECK-NEXT: [[TMP7:%.*]] = icmp eq i32 [[INDEX_NEXT]], [[N_VEC]]194; CHECK-NEXT: br i1 [[TMP7]], label %[[MIDDLE_BLOCK:.*]], label %[[VECTOR_BODY]], !llvm.loop [[LOOP6:![0-9]+]]195; CHECK: [[MIDDLE_BLOCK]]:196; CHECK-NEXT: [[TMP8:%.*]] = extractelement <2 x i32> [[PREDPHI1]], i32 1197; CHECK-NEXT: [[CMP_N:%.*]] = icmp eq i32 [[TMP1]], [[N_VEC]]198; CHECK-NEXT: br i1 [[CMP_N]], label %[[F1_EXIT_LOOPEXIT:.*]], label %[[_LR_PH_I]]199; CHECK: [[_LR_PH_I]]:200; CHECK-NEXT: [[BC_RESUME_VAL:%.*]] = phi i32 [ [[TMP2]], %[[MIDDLE_BLOCK]] ], [ [[B_PROMOTED]], %[[BB]] ]201; CHECK-NEXT: br label %[[DOTLR_PH_I:.*]]202; CHECK: [[_LR_PH_I1:.*:]]203; CHECK-NEXT: [[UNNAMEDTMP8:%.*]] = phi i32 [ [[UNNAMEDTMP18:%.*]], %[[BB16:.*]] ], [ [[BC_RESUME_VAL]], %[[_LR_PH_I]] ]204; CHECK-NEXT: [[UNNAMEDTMP2:%.*]] = icmp sgt i32 [[UNNAMEDTMP8]], 10205; CHECK-NEXT: br i1 [[UNNAMEDTMP2]], label %[[BB16]], label %[[UNNAMEDBB10:.*]]206; CHECK: [[UNNAMEDBB10]]:207; CHECK-NEXT: [[CMP:%.*]] = icmp sgt i32 [[UNNAMEDTMP8]], [[N]]208; CHECK-NEXT: br i1 [[CMP]], label %[[UNNAMEDBB12:.*]], label %[[BB16]]209; CHECK: [[UNNAMEDBB12]]:210; CHECK-NEXT: br label %[[BB16]]211; CHECK: [[BB16]]:212; CHECK-NEXT: [[UNNAMEDTMP17:%.*]] = phi i32 [ 0, %[[UNNAMEDBB10]] ], [ 1, %[[DOTLR_PH_I]] ], [ 2, %[[UNNAMEDBB12]] ]213; CHECK-NEXT: [[UNNAMEDTMP18]] = add nsw i32 [[UNNAMEDTMP8]], 1214; CHECK-NEXT: [[UNNAMEDTMP19:%.*]] = icmp slt i32 [[UNNAMEDTMP18]], 4215; CHECK-NEXT: br i1 [[UNNAMEDTMP19]], label %[[DOTLR_PH_I]], label %[[F1_EXIT_LOOPEXIT]], !llvm.loop [[LOOP7:![0-9]+]]216; CHECK: [[F1_EXIT_LOOPEXIT]]:217; CHECK-NEXT: [[DOTLCSSA:%.*]] = phi i32 [ [[UNNAMEDTMP17]], %[[BB16]] ], [ [[TMP8]], %[[MIDDLE_BLOCK]] ]218; CHECK-NEXT: ret i32 [[DOTLCSSA]]219;220bb:221 %b.promoted = load i32, ptr @b, align 4222 br label %.lr.ph.i223 224.lr.ph.i:225 %tmp8 = phi i32 [ %tmp18, %bb16 ], [ %b.promoted, %bb ]226 %tmp2 = icmp sgt i32 %tmp8, 10227 br i1 %tmp2, label %bb16, label %bb10228 229bb10:230 %cmp = icmp sgt i32 %tmp8, %N231 br i1 %cmp, label %bb12, label %bb16232 233bb12:234 br label %bb16235 236bb16:237 %tmp17 = phi i32 [ 0, %bb10 ], [ 1, %.lr.ph.i ], [ 2, %bb12 ]238 %tmp18 = add nsw i32 %tmp8, 1239 %tmp19 = icmp slt i32 %tmp18, 4240 br i1 %tmp19, label %.lr.ph.i, label %f1.exit.loopexit241 242f1.exit.loopexit:243 %.lcssa = phi i32 [ %tmp17, %bb16 ]244 ret i32 %.lcssa245}246 247; more than one incoming value for outside user: %.lcssa248define i32 @test4(i32 %N) {249; CHECK-LABEL: define i32 @test4(250; CHECK-SAME: i32 [[N:%.*]]) {251; CHECK-NEXT: [[BB:.*]]:252; CHECK-NEXT: [[B_PROMOTED:%.*]] = load i32, ptr @b, align 4253; CHECK-NEXT: [[ICMP:%.*]] = icmp slt i32 [[B_PROMOTED]], [[N]]254; CHECK-NEXT: br i1 [[ICMP]], label %[[F1_EXIT_LOOPEXIT:.*]], label %[[DOTLR_PH_I_PREHEADER:.*]]255; CHECK: [[_LR_PH_I_PREHEADER:.*:]]256; CHECK-NEXT: [[TMP0:%.*]] = add i32 [[B_PROMOTED]], 1257; CHECK-NEXT: [[SMAX:%.*]] = call i32 @llvm.smax.i32(i32 [[TMP0]], i32 4)258; CHECK-NEXT: [[TMP1:%.*]] = sub i32 [[SMAX]], [[B_PROMOTED]]259; CHECK-NEXT: [[MIN_ITERS_CHECK:%.*]] = icmp ult i32 [[TMP1]], 2260; CHECK-NEXT: br i1 [[MIN_ITERS_CHECK]], label %[[_LR_PH_I1:.*]], label %[[VECTOR_PH:.*]]261; CHECK: [[VECTOR_PH]]:262; CHECK-NEXT: [[N_MOD_VF:%.*]] = urem i32 [[TMP1]], 2263; CHECK-NEXT: [[N_VEC:%.*]] = sub i32 [[TMP1]], [[N_MOD_VF]]264; CHECK-NEXT: [[TMP2:%.*]] = add i32 [[B_PROMOTED]], [[N_VEC]]265; CHECK-NEXT: [[BROADCAST_SPLATINSERT:%.*]] = insertelement <2 x i32> poison, i32 [[B_PROMOTED]], i64 0266; CHECK-NEXT: [[DOTSPLAT:%.*]] = shufflevector <2 x i32> [[BROADCAST_SPLATINSERT]], <2 x i32> poison, <2 x i32> zeroinitializer267; CHECK-NEXT: [[INDUCTION:%.*]] = add nsw <2 x i32> [[DOTSPLAT]], <i32 0, i32 1>268; CHECK-NEXT: br label %[[VECTOR_BODY:.*]]269; CHECK: [[VECTOR_BODY]]:270; CHECK-NEXT: [[INDEX:%.*]] = phi i32 [ 0, %[[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], %[[VECTOR_BODY]] ]271; CHECK-NEXT: [[VEC_IND:%.*]] = phi <2 x i32> [ [[INDUCTION]], %[[VECTOR_PH]] ], [ [[VEC_IND_NEXT:%.*]], %[[VECTOR_BODY]] ]272; CHECK-NEXT: [[TMP3:%.*]] = icmp sgt <2 x i32> [[VEC_IND]], splat (i32 10)273; CHECK-NEXT: [[PREDPHI:%.*]] = select <2 x i1> [[TMP3]], <2 x i32> splat (i32 1), <2 x i32> zeroinitializer274; CHECK-NEXT: [[INDEX_NEXT]] = add nuw i32 [[INDEX]], 2275; CHECK-NEXT: [[VEC_IND_NEXT]] = add nsw <2 x i32> [[VEC_IND]], splat (i32 2)276; CHECK-NEXT: [[TMP4:%.*]] = icmp eq i32 [[INDEX_NEXT]], [[N_VEC]]277; CHECK-NEXT: br i1 [[TMP4]], label %[[MIDDLE_BLOCK:.*]], label %[[VECTOR_BODY]], !llvm.loop [[LOOP8:![0-9]+]]278; CHECK: [[MIDDLE_BLOCK]]:279; CHECK-NEXT: [[TMP5:%.*]] = extractelement <2 x i32> [[PREDPHI]], i32 1280; CHECK-NEXT: [[CMP_N:%.*]] = icmp eq i32 [[TMP1]], [[N_VEC]]281; CHECK-NEXT: br i1 [[CMP_N]], label %[[F1_EXIT_LOOPEXIT_LOOPEXIT:.*]], label %[[_LR_PH_I1]]282; CHECK: [[_LR_PH_I1]]:283; CHECK-NEXT: [[BC_RESUME_VAL:%.*]] = phi i32 [ [[TMP2]], %[[MIDDLE_BLOCK]] ], [ [[B_PROMOTED]], %[[DOTLR_PH_I_PREHEADER]] ]284; CHECK-NEXT: br label %[[DOTLR_PH_I:.*]]285; CHECK: [[_LR_PH_I:.*:]]286; CHECK-NEXT: [[UNNAMEDTMP8:%.*]] = phi i32 [ [[UNNAMEDTMP18:%.*]], %[[BB16:.*]] ], [ [[BC_RESUME_VAL]], %[[_LR_PH_I1]] ]287; CHECK-NEXT: [[UNNAMEDTMP2:%.*]] = icmp sgt i32 [[UNNAMEDTMP8]], 10288; CHECK-NEXT: br i1 [[UNNAMEDTMP2]], label %[[BB16]], label %[[UNNAMEDBB10:.*]]289; CHECK: [[UNNAMEDBB10]]:290; CHECK-NEXT: br label %[[BB16]]291; CHECK: [[BB16]]:292; CHECK-NEXT: [[UNNAMEDTMP17:%.*]] = phi i32 [ 0, %[[UNNAMEDBB10]] ], [ 1, %[[DOTLR_PH_I]] ]293; CHECK-NEXT: [[UNNAMEDTMP18]] = add nsw i32 [[UNNAMEDTMP8]], 1294; CHECK-NEXT: [[UNNAMEDTMP19:%.*]] = icmp slt i32 [[UNNAMEDTMP18]], 4295; CHECK-NEXT: br i1 [[UNNAMEDTMP19]], label %[[DOTLR_PH_I]], label %[[F1_EXIT_LOOPEXIT_LOOPEXIT]], !llvm.loop [[LOOP9:![0-9]+]]296; CHECK: [[F1_EXIT_LOOPEXIT_LOOPEXIT]]:297; CHECK-NEXT: [[TMP17_LCSSA:%.*]] = phi i32 [ [[UNNAMEDTMP17]], %[[BB16]] ], [ [[TMP5]], %[[MIDDLE_BLOCK]] ]298; CHECK-NEXT: br label %[[F1_EXIT_LOOPEXIT]]299; CHECK: [[F1_EXIT_LOOPEXIT]]:300; CHECK-NEXT: [[DOTLCSSA:%.*]] = phi i32 [ 2, %[[BB]] ], [ [[TMP17_LCSSA]], %[[F1_EXIT_LOOPEXIT_LOOPEXIT]] ]301; CHECK-NEXT: ret i32 [[DOTLCSSA]]302;303bb:304 %b.promoted = load i32, ptr @b, align 4305 %icmp = icmp slt i32 %b.promoted, %N306 br i1 %icmp, label %f1.exit.loopexit, label %.lr.ph.i307 308.lr.ph.i:309 %tmp8 = phi i32 [ %tmp18, %bb16 ], [ %b.promoted, %bb ]310 %tmp2 = icmp sgt i32 %tmp8, 10311 br i1 %tmp2, label %bb16, label %bb10312 313bb10:314 br label %bb16315 316bb16:317 %tmp17 = phi i32 [ 0, %bb10 ], [ 1, %.lr.ph.i ]318 %tmp18 = add nsw i32 %tmp8, 1319 %tmp19 = icmp slt i32 %tmp18, 4320 br i1 %tmp19, label %.lr.ph.i, label %f1.exit.loopexit321 322f1.exit.loopexit:323 %.lcssa = phi i32 [ %tmp17, %bb16 ], [ 2, %bb ]324 ret i32 %.lcssa325}326 327; non hdr phi that depends on reduction and is used outside the loop.328; reduction phis are only allowed to have bump or reduction operations as the inside user, so we should329; not vectorize this.330define i32 @reduction_sum(i32 %n, ptr noalias nocapture %A, ptr noalias nocapture %B) nounwind uwtable readonly noinline ssp {331; CHECK-LABEL: define i32 @reduction_sum(332; CHECK-SAME: i32 [[N:%.*]], ptr noalias captures(none) [[A:%.*]], ptr noalias captures(none) [[B:%.*]]) #[[ATTR0:[0-9]+]] {333; CHECK-NEXT: [[ENTRY:.*]]:334; CHECK-NEXT: [[C1:%.*]] = icmp sgt i32 [[N]], 0335; CHECK-NEXT: br i1 [[C1]], label %[[HEADER_PREHEADER:.*]], [[DOT_CRIT_EDGE:label %.*]]336; CHECK: [[HEADER_PREHEADER]]:337; CHECK-NEXT: br label %[[HEADER:.*]]338; CHECK: [[HEADER]]:339; CHECK-NEXT: [[INDVARS_IV:%.*]] = phi i64 [ [[INDVARS_IV_NEXT:%.*]], %[[BB16:.*]] ], [ 0, %[[HEADER_PREHEADER]] ]340; CHECK-NEXT: [[SUM_02:%.*]] = phi i32 [ [[C9:%.*]], %[[BB16]] ], [ 0, %[[HEADER_PREHEADER]] ]341; CHECK-NEXT: [[C2:%.*]] = getelementptr inbounds i32, ptr [[A]], i64 [[INDVARS_IV]]342; CHECK-NEXT: [[C3:%.*]] = load i32, ptr [[C2]], align 4343; CHECK-NEXT: [[C4:%.*]] = getelementptr inbounds i32, ptr [[B]], i64 [[INDVARS_IV]]344; CHECK-NEXT: [[C5:%.*]] = load i32, ptr [[C4]], align 4345; CHECK-NEXT: [[UNNAMEDTMP2:%.*]] = icmp sgt i32 [[SUM_02]], 10346; CHECK-NEXT: br i1 [[UNNAMEDTMP2]], label %[[BB16]], label %[[UNNAMEDBB10:.*]]347; CHECK: [[UNNAMEDBB10]]:348; CHECK-NEXT: br label %[[BB16]]349; CHECK: [[BB16]]:350; CHECK-NEXT: [[UNNAMEDTMP17:%.*]] = phi i32 [ [[SUM_02]], %[[UNNAMEDBB10]] ], [ 1, %[[HEADER]] ]351; CHECK-NEXT: [[C6:%.*]] = trunc i64 [[INDVARS_IV]] to i32352; CHECK-NEXT: [[C7:%.*]] = add i32 [[SUM_02]], [[C6]]353; CHECK-NEXT: [[C8:%.*]] = add i32 [[C7]], [[C3]]354; CHECK-NEXT: [[C9]] = add i32 [[C8]], [[C5]]355; CHECK-NEXT: [[INDVARS_IV_NEXT]] = add i64 [[INDVARS_IV]], 1356; CHECK-NEXT: [[LFTR_WIDEIV:%.*]] = trunc i64 [[INDVARS_IV_NEXT]] to i32357; CHECK-NEXT: [[EXITCOND:%.*]] = icmp eq i32 [[LFTR_WIDEIV]], [[N]]358; CHECK-NEXT: br i1 [[EXITCOND]], label %[[DOT_CRIT_EDGE_LOOPEXIT:.*]], label %[[HEADER]]359; CHECK: [[__CRIT_EDGE_LOOPEXIT:.*:]]360; CHECK-NEXT: [[TMP17_LCSSA:%.*]] = phi i32 [ [[UNNAMEDTMP17]], %[[BB16]] ]361; CHECK-NEXT: [[C9_LCSSA:%.*]] = phi i32 [ [[C9]], %[[BB16]] ]362; CHECK-NEXT: br [[DOT_CRIT_EDGE]]363; CHECK: [[__CRIT_EDGE:.*:]]364; CHECK-NEXT: [[SUM_0_LCSSA:%.*]] = phi i32 [ 0, %[[ENTRY]] ], [ [[C9_LCSSA]], %[[DOT_CRIT_EDGE_LOOPEXIT]] ]365; CHECK-NEXT: [[NONHDR_LCSSA:%.*]] = phi i32 [ 1, %[[ENTRY]] ], [ [[TMP17_LCSSA]], %[[DOT_CRIT_EDGE_LOOPEXIT]] ]366; CHECK-NEXT: ret i32 [[SUM_0_LCSSA]]367;368entry:369 %c1 = icmp sgt i32 %n, 0370 br i1 %c1, label %header, label %._crit_edge371 372header: ; preds = %0, %.lr.ph373 %indvars.iv = phi i64 [ %indvars.iv.next, %bb16 ], [ 0, %entry ]374 %sum.02 = phi i32 [ %c9, %bb16 ], [ 0, %entry ]375 %c2 = getelementptr inbounds i32, ptr %A, i64 %indvars.iv376 %c3 = load i32, ptr %c2, align 4377 %c4 = getelementptr inbounds i32, ptr %B, i64 %indvars.iv378 %c5 = load i32, ptr %c4, align 4379 %tmp2 = icmp sgt i32 %sum.02, 10380 br i1 %tmp2, label %bb16, label %bb10381 382bb10:383 br label %bb16384 385bb16:386 %tmp17 = phi i32 [ %sum.02, %bb10 ], [ 1, %header ]387 %c6 = trunc i64 %indvars.iv to i32388 %c7 = add i32 %sum.02, %c6389 %c8 = add i32 %c7, %c3390 %c9 = add i32 %c8, %c5391 %indvars.iv.next = add i64 %indvars.iv, 1392 %lftr.wideiv = trunc i64 %indvars.iv.next to i32393 %exitcond = icmp eq i32 %lftr.wideiv, %n394 br i1 %exitcond, label %._crit_edge, label %header395 396._crit_edge: ; preds = %.lr.ph, %0397 %sum.0.lcssa = phi i32 [ 0, %entry ], [ %c9, %bb16 ]398 %nonhdr.lcssa = phi i32 [ 1, %entry], [ %tmp17, %bb16 ]399 ret i32 %sum.0.lcssa400}401 402; invalid cyclic dependency with header phi iv, which prevents iv from being403; recognized as induction var.404; cannot vectorize.405define i32 @cyclic_dep_with_indvar() {406; CHECK-LABEL: define i32 @cyclic_dep_with_indvar() {407; CHECK-NEXT: [[BB:.*]]:408; CHECK-NEXT: [[B_PROMOTED:%.*]] = load i32, ptr @b, align 4409; CHECK-NEXT: br label %[[DOTLR_PH_I:.*]]410; CHECK: [[_LR_PH_I:.*:]]411; CHECK-NEXT: [[IV:%.*]] = phi i32 [ [[IVNEXT:%.*]], %[[BB16:.*]] ], [ [[B_PROMOTED]], %[[BB]] ]412; CHECK-NEXT: [[UNNAMEDTMP2:%.*]] = icmp sgt i32 [[IV]], 10413; CHECK-NEXT: br i1 [[UNNAMEDTMP2]], label %[[BB16]], label %[[UNNAMEDBB10:.*]]414; CHECK: [[UNNAMEDBB10]]:415; CHECK-NEXT: br label %[[BB16]]416; CHECK: [[BB16]]:417; CHECK-NEXT: [[UNNAMEDTMP17:%.*]] = phi i32 [ 0, %[[UNNAMEDBB10]] ], [ [[IV]], %[[DOTLR_PH_I]] ]418; CHECK-NEXT: [[IVNEXT]] = add nsw i32 [[UNNAMEDTMP17]], 1419; CHECK-NEXT: [[UNNAMEDTMP19:%.*]] = icmp slt i32 [[IVNEXT]], 4420; CHECK-NEXT: br i1 [[UNNAMEDTMP19]], label %[[DOTLR_PH_I]], label %[[F1_EXIT_LOOPEXIT:.*]]421; CHECK: [[F1_EXIT_LOOPEXIT]]:422; CHECK-NEXT: [[DOTLCSSA:%.*]] = phi i32 [ [[UNNAMEDTMP17]], %[[BB16]] ]423; CHECK-NEXT: ret i32 [[DOTLCSSA]]424;425bb:426 %b.promoted = load i32, ptr @b, align 4427 br label %.lr.ph.i428 429.lr.ph.i:430 %iv = phi i32 [ %ivnext, %bb16 ], [ %b.promoted, %bb ]431 %tmp2 = icmp sgt i32 %iv, 10432 br i1 %tmp2, label %bb16, label %bb10433 434bb10:435 br label %bb16436 437bb16:438 %tmp17 = phi i32 [ 0, %bb10 ], [ %iv, %.lr.ph.i ]439 %ivnext = add nsw i32 %tmp17, 1440 %tmp19 = icmp slt i32 %ivnext, 4441 br i1 %tmp19, label %.lr.ph.i, label %f1.exit.loopexit442 443f1.exit.loopexit:444 %.lcssa = phi i32 [ %tmp17, %bb16 ]445 ret i32 %.lcssa446}447 448; non-reduction phi 'tmp17' used outside loop has cyclic dependence with %x.05 phi449; cannot vectorize.450define i32 @not_valid_reduction(i32 %n, ptr noalias nocapture %A) nounwind uwtable readonly {451; CHECK-LABEL: define i32 @not_valid_reduction(452; CHECK-SAME: i32 [[N:%.*]], ptr noalias captures(none) [[A:%.*]]) #[[ATTR1:[0-9]+]] {453; CHECK-NEXT: [[ENTRY:.*]]:454; CHECK-NEXT: [[CMP4:%.*]] = icmp sgt i32 [[N]], 0455; CHECK-NEXT: br i1 [[CMP4]], label %[[FOR_BODY_PREHEADER:.*]], label %[[FOR_END:.*]]456; CHECK: [[FOR_BODY_PREHEADER]]:457; CHECK-NEXT: br label %[[FOR_BODY:.*]]458; CHECK: [[FOR_BODY]]:459; CHECK-NEXT: [[INDVARS_IV:%.*]] = phi i64 [ [[INDVARS_IV_NEXT:%.*]], %[[LATCH:.*]] ], [ 0, %[[FOR_BODY_PREHEADER]] ]460; CHECK-NEXT: [[X_05:%.*]] = phi i32 [ [[UNNAMEDTMP17:%.*]], %[[LATCH]] ], [ 0, %[[FOR_BODY_PREHEADER]] ]461; CHECK-NEXT: [[ARRAYIDX:%.*]] = getelementptr inbounds i32, ptr [[A]], i64 [[INDVARS_IV]]462; CHECK-NEXT: [[UNNAMEDTMP0:%.*]] = load i32, ptr [[ARRAYIDX]], align 4463; CHECK-NEXT: [[UNNAMEDTMP2:%.*]] = icmp sgt i64 [[INDVARS_IV]], 10464; CHECK-NEXT: [[SUB:%.*]] = sub nsw i32 [[X_05]], [[UNNAMEDTMP0]]465; CHECK-NEXT: br i1 [[UNNAMEDTMP2]], label %[[UNNAMEDBB16:.*]], label %[[UNNAMEDBB10:.*]]466; CHECK: [[UNNAMEDBB10]]:467; CHECK-NEXT: br label %[[UNNAMEDBB16]]468; CHECK: [[UNNAMEDBB16]]:469; CHECK-NEXT: [[UNNAMEDTMP17]] = phi i32 [ 1, %[[UNNAMEDBB10]] ], [ [[SUB]], %[[FOR_BODY]] ]470; CHECK-NEXT: br label %[[LATCH]]471; CHECK: [[LATCH]]:472; CHECK-NEXT: [[INDVARS_IV_NEXT]] = add i64 [[INDVARS_IV]], 1473; CHECK-NEXT: [[LFTR_WIDEIV:%.*]] = trunc i64 [[INDVARS_IV_NEXT]] to i32474; CHECK-NEXT: [[EXITCOND:%.*]] = icmp eq i32 [[LFTR_WIDEIV]], [[N]]475; CHECK-NEXT: br i1 [[EXITCOND]], label %[[FOR_END_LOOPEXIT:.*]], label %[[FOR_BODY]]476; CHECK: [[FOR_END_LOOPEXIT]]:477; CHECK-NEXT: [[TMP17_LCSSA:%.*]] = phi i32 [ [[UNNAMEDTMP17]], %[[LATCH]] ]478; CHECK-NEXT: br label %[[FOR_END]]479; CHECK: [[FOR_END]]:480; CHECK-NEXT: [[X_0_LCSSA:%.*]] = phi i32 [ 0, %[[ENTRY]] ], [ [[TMP17_LCSSA]], %[[FOR_END_LOOPEXIT]] ]481; CHECK-NEXT: ret i32 [[X_0_LCSSA]]482;483entry:484 %cmp4 = icmp sgt i32 %n, 0485 br i1 %cmp4, label %for.body, label %for.end486 487for.body: ; preds = %entry, %for.body488 %indvars.iv = phi i64 [ %indvars.iv.next, %latch ], [ 0, %entry ]489 %x.05 = phi i32 [ %tmp17, %latch ], [ 0, %entry ]490 %arrayidx = getelementptr inbounds i32, ptr %A, i64 %indvars.iv491 %tmp0 = load i32, ptr %arrayidx, align 4492 %tmp2 = icmp sgt i64 %indvars.iv, 10493 %sub = sub nsw i32 %x.05, %tmp0494 br i1 %tmp2, label %bb16, label %bb10495 496bb10:497 br label %bb16498 499bb16:500 %tmp17 = phi i32 [ 1, %bb10 ], [ %sub, %for.body ]501 br label %latch502 503latch:504 %indvars.iv.next = add i64 %indvars.iv, 1505 %lftr.wideiv = trunc i64 %indvars.iv.next to i32506 %exitcond = icmp eq i32 %lftr.wideiv, %n507 br i1 %exitcond, label %for.end, label %for.body508 509for.end: ; preds = %for.body, %entry510 %x.0.lcssa = phi i32 [ 0, %entry ], [ %tmp17 , %latch ]511 ret i32 %x.0.lcssa512}513 514define i8 @outside_user_non_phi() {515; CHECK-LABEL: define i8 @outside_user_non_phi() {516; CHECK-NEXT: [[BB:.*]]:517; CHECK-NEXT: [[B_PROMOTED:%.*]] = load i32, ptr @b, align 4518; CHECK-NEXT: [[TMP0:%.*]] = add i32 [[B_PROMOTED]], 1519; CHECK-NEXT: [[SMAX:%.*]] = call i32 @llvm.smax.i32(i32 [[TMP0]], i32 4)520; CHECK-NEXT: [[TMP1:%.*]] = sub i32 [[SMAX]], [[B_PROMOTED]]521; CHECK-NEXT: [[MIN_ITERS_CHECK:%.*]] = icmp ult i32 [[TMP1]], 2522; CHECK-NEXT: br i1 [[MIN_ITERS_CHECK]], label %[[_LR_PH_I1:.*]], label %[[VECTOR_PH:.*]]523; CHECK: [[VECTOR_PH]]:524; CHECK-NEXT: [[N_MOD_VF:%.*]] = urem i32 [[TMP1]], 2525; CHECK-NEXT: [[N_VEC:%.*]] = sub i32 [[TMP1]], [[N_MOD_VF]]526; CHECK-NEXT: [[TMP2:%.*]] = add i32 [[B_PROMOTED]], [[N_VEC]]527; CHECK-NEXT: [[DOTSPLATINSERT:%.*]] = insertelement <2 x i32> poison, i32 [[B_PROMOTED]], i64 0528; CHECK-NEXT: [[DOTSPLAT:%.*]] = shufflevector <2 x i32> [[DOTSPLATINSERT]], <2 x i32> poison, <2 x i32> zeroinitializer529; CHECK-NEXT: [[INDUCTION:%.*]] = add nsw <2 x i32> [[DOTSPLAT]], <i32 0, i32 1>530; CHECK-NEXT: br label %[[VECTOR_BODY:.*]]531; CHECK: [[VECTOR_BODY]]:532; CHECK-NEXT: [[INDEX:%.*]] = phi i32 [ 0, %[[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], %[[VECTOR_BODY]] ]533; CHECK-NEXT: [[VEC_IND:%.*]] = phi <2 x i32> [ [[INDUCTION]], %[[VECTOR_PH]] ], [ [[VEC_IND_NEXT:%.*]], %[[VECTOR_BODY]] ]534; CHECK-NEXT: [[TMP3:%.*]] = icmp sgt <2 x i32> [[VEC_IND]], splat (i32 10)535; CHECK-NEXT: [[PREDPHI:%.*]] = select <2 x i1> [[TMP3]], <2 x i32> splat (i32 1), <2 x i32> zeroinitializer536; CHECK-NEXT: [[TMP4:%.*]] = trunc <2 x i32> [[PREDPHI]] to <2 x i8>537; CHECK-NEXT: [[INDEX_NEXT]] = add nuw i32 [[INDEX]], 2538; CHECK-NEXT: [[VEC_IND_NEXT]] = add nsw <2 x i32> [[VEC_IND]], splat (i32 2)539; CHECK-NEXT: [[TMP5:%.*]] = icmp eq i32 [[INDEX_NEXT]], [[N_VEC]]540; CHECK-NEXT: br i1 [[TMP5]], label %[[MIDDLE_BLOCK:.*]], label %[[VECTOR_BODY]], !llvm.loop [[LOOP10:![0-9]+]]541; CHECK: [[MIDDLE_BLOCK]]:542; CHECK-NEXT: [[TMP6:%.*]] = extractelement <2 x i8> [[TMP4]], i32 1543; CHECK-NEXT: [[CMP_N:%.*]] = icmp eq i32 [[TMP1]], [[N_VEC]]544; CHECK-NEXT: br i1 [[CMP_N]], label %[[F1_EXIT_LOOPEXIT:.*]], label %[[_LR_PH_I1]]545; CHECK: [[_LR_PH_I1]]:546; CHECK-NEXT: [[BC_RESUME_VAL:%.*]] = phi i32 [ [[TMP2]], %[[MIDDLE_BLOCK]] ], [ [[B_PROMOTED]], %[[BB]] ]547; CHECK-NEXT: br label %[[DOTLR_PH_I:.*]]548; CHECK: [[_LR_PH_I:.*:]]549; CHECK-NEXT: [[UNNAMEDTMP8:%.*]] = phi i32 [ [[UNNAMEDTMP18:%.*]], %[[BB16:.*]] ], [ [[BC_RESUME_VAL]], %[[_LR_PH_I1]] ]550; CHECK-NEXT: [[UNNAMEDTMP2:%.*]] = icmp sgt i32 [[UNNAMEDTMP8]], 10551; CHECK-NEXT: br i1 [[UNNAMEDTMP2]], label %[[BB16]], label %[[UNNAMEDBB10:.*]]552; CHECK: [[UNNAMEDBB10]]:553; CHECK-NEXT: br label %[[BB16]]554; CHECK: [[BB16]]:555; CHECK-NEXT: [[UNNAMEDTMP17:%.*]] = phi i32 [ 0, %[[UNNAMEDBB10]] ], [ 1, %[[DOTLR_PH_I]] ]556; CHECK-NEXT: [[TMP17_TRUNC:%.*]] = trunc i32 [[UNNAMEDTMP17]] to i8557; CHECK-NEXT: [[UNNAMEDTMP18]] = add nsw i32 [[UNNAMEDTMP8]], 1558; CHECK-NEXT: [[UNNAMEDTMP19:%.*]] = icmp slt i32 [[UNNAMEDTMP18]], 4559; CHECK-NEXT: br i1 [[UNNAMEDTMP19]], label %[[DOTLR_PH_I]], label %[[F1_EXIT_LOOPEXIT]], !llvm.loop [[LOOP11:![0-9]+]]560; CHECK: [[F1_EXIT_LOOPEXIT]]:561; CHECK-NEXT: [[DOTLCSSA:%.*]] = phi i8 [ [[TMP17_TRUNC]], %[[BB16]] ], [ [[TMP6]], %[[MIDDLE_BLOCK]] ]562; CHECK-NEXT: ret i8 [[DOTLCSSA]]563;564bb:565 %b.promoted = load i32, ptr @b, align 4566 br label %.lr.ph.i567 568.lr.ph.i:569 %tmp8 = phi i32 [ %tmp18, %bb16 ], [ %b.promoted, %bb ]570 %tmp2 = icmp sgt i32 %tmp8, 10571 br i1 %tmp2, label %bb16, label %bb10572 573bb10:574 br label %bb16575 576bb16:577 %tmp17 = phi i32 [ 0, %bb10 ], [ 1, %.lr.ph.i ]578 %tmp17.trunc = trunc i32 %tmp17 to i8579 %tmp18 = add nsw i32 %tmp8, 1580 %tmp19 = icmp slt i32 %tmp18, 4581 br i1 %tmp19, label %.lr.ph.i, label %f1.exit.loopexit582 583f1.exit.loopexit:584 %.lcssa = phi i8 [ %tmp17.trunc, %bb16 ]585 ret i8 %.lcssa586}587 588define i32 @no_vectorize_reduction_with_outside_use(i32 %n, ptr nocapture %A, ptr nocapture %B) nounwind uwtable readonly {589; CHECK-LABEL: define i32 @no_vectorize_reduction_with_outside_use(590; CHECK-SAME: i32 [[N:%.*]], ptr captures(none) [[A:%.*]], ptr captures(none) [[B:%.*]]) #[[ATTR1]] {591; CHECK-NEXT: [[ENTRY:.*]]:592; CHECK-NEXT: [[CMP7:%.*]] = icmp sgt i32 [[N]], 0593; CHECK-NEXT: br i1 [[CMP7]], label %[[FOR_BODY_PREHEADER:.*]], label %[[FOR_END:.*]]594; CHECK: [[FOR_BODY_PREHEADER]]:595; CHECK-NEXT: br label %[[FOR_BODY:.*]]596; CHECK: [[FOR_BODY]]:597; CHECK-NEXT: [[INDVARS_IV:%.*]] = phi i64 [ [[INDVARS_IV_NEXT:%.*]], %[[FOR_BODY]] ], [ 0, %[[FOR_BODY_PREHEADER]] ]598; CHECK-NEXT: [[RESULT_08:%.*]] = phi i32 [ [[OR:%.*]], %[[FOR_BODY]] ], [ 0, %[[FOR_BODY_PREHEADER]] ]599; CHECK-NEXT: [[ARRAYIDX:%.*]] = getelementptr inbounds i32, ptr [[A]], i64 [[INDVARS_IV]]600; CHECK-NEXT: [[TMP0:%.*]] = load i32, ptr [[ARRAYIDX]], align 4601; CHECK-NEXT: [[ARRAYIDX2:%.*]] = getelementptr inbounds i32, ptr [[B]], i64 [[INDVARS_IV]]602; CHECK-NEXT: [[TMP1:%.*]] = load i32, ptr [[ARRAYIDX2]], align 4603; CHECK-NEXT: [[ADD:%.*]] = add nsw i32 [[TMP1]], [[TMP0]]604; CHECK-NEXT: [[OR]] = or i32 [[ADD]], [[RESULT_08]]605; CHECK-NEXT: [[INDVARS_IV_NEXT]] = add i64 [[INDVARS_IV]], 1606; CHECK-NEXT: [[LFTR_WIDEIV:%.*]] = trunc i64 [[INDVARS_IV_NEXT]] to i32607; CHECK-NEXT: [[EXITCOND:%.*]] = icmp eq i32 [[LFTR_WIDEIV]], [[N]]608; CHECK-NEXT: br i1 [[EXITCOND]], label %[[FOR_END_LOOPEXIT:.*]], label %[[FOR_BODY]]609; CHECK: [[FOR_END_LOOPEXIT]]:610; CHECK-NEXT: [[DOTLCSSA:%.*]] = phi i32 [ [[TMP1]], %[[FOR_BODY]] ]611; CHECK-NEXT: br label %[[FOR_END]]612; CHECK: [[FOR_END]]:613; CHECK-NEXT: [[RESULT_0_LCSSA:%.*]] = phi i32 [ 0, %[[ENTRY]] ], [ [[DOTLCSSA]], %[[FOR_END_LOOPEXIT]] ]614; CHECK-NEXT: ret i32 [[RESULT_0_LCSSA]]615;616entry:617 %cmp7 = icmp sgt i32 %n, 0618 br i1 %cmp7, label %for.body, label %for.end619 620for.body: ; preds = %entry, %for.body621 %indvars.iv = phi i64 [ %indvars.iv.next, %for.body ], [ 0, %entry ]622 %result.08 = phi i32 [ %or, %for.body ], [ 0, %entry ]623 %arrayidx = getelementptr inbounds i32, ptr %A, i64 %indvars.iv624 %0 = load i32, ptr %arrayidx, align 4625 %arrayidx2 = getelementptr inbounds i32, ptr %B, i64 %indvars.iv626 %1 = load i32, ptr %arrayidx2, align 4627 %add = add nsw i32 %1, %0628 %or = or i32 %add, %result.08629 %indvars.iv.next = add i64 %indvars.iv, 1630 %lftr.wideiv = trunc i64 %indvars.iv.next to i32631 %exitcond = icmp eq i32 %lftr.wideiv, %n632 br i1 %exitcond, label %for.end, label %for.body633 634for.end: ; preds = %for.body, %entry635 %result.0.lcssa = phi i32 [ 0, %entry ], [ %1, %for.body ]636 ret i32 %result.0.lcssa637}638 639; vectorize c[i] = a[i] + b[i] loop where result of c[i] is used outside the640; loop641define i32 @sum_arrays_outside_use(ptr %B, ptr %A, ptr %C, i32 %N) {642; CHECK-LABEL: define i32 @sum_arrays_outside_use(643; CHECK-SAME: ptr [[B:%.*]], ptr [[A:%.*]], ptr [[C:%.*]], i32 [[N:%.*]]) {644; CHECK-NEXT: [[BB:.*]]:645; CHECK-NEXT: [[A3:%.*]] = ptrtoint ptr [[A]] to i32646; CHECK-NEXT: [[B2:%.*]] = ptrtoint ptr [[B]] to i32647; CHECK-NEXT: [[C1:%.*]] = ptrtoint ptr [[C]] to i32648; CHECK-NEXT: [[B_PROMOTED:%.*]] = load i32, ptr @b, align 4649; CHECK-NEXT: [[TMP0:%.*]] = add nsw i32 [[B_PROMOTED]], 1650; CHECK-NEXT: [[SMAX:%.*]] = call i32 @llvm.smax.i32(i32 [[N]], i32 [[TMP0]])651; CHECK-NEXT: [[TMP1:%.*]] = sub i32 [[SMAX]], [[B_PROMOTED]]652; CHECK-NEXT: [[MIN_ITERS_CHECK:%.*]] = icmp ult i32 [[TMP1]], 2653; CHECK-NEXT: br i1 [[MIN_ITERS_CHECK]], label %[[_LR_PH_I1:.*]], label %[[VECTOR_MEMCHECK:.*]]654; CHECK: [[VECTOR_MEMCHECK]]:655; CHECK-NEXT: [[TMP2:%.*]] = sub i32 [[C1]], [[B2]]656; CHECK-NEXT: [[DIFF_CHECK:%.*]] = icmp ult i32 [[TMP2]], 8657; CHECK-NEXT: [[TMP3:%.*]] = sub i32 [[C1]], [[A3]]658; CHECK-NEXT: [[DIFF_CHECK4:%.*]] = icmp ult i32 [[TMP3]], 8659; CHECK-NEXT: [[CONFLICT_RDX:%.*]] = or i1 [[DIFF_CHECK]], [[DIFF_CHECK4]]660; CHECK-NEXT: br i1 [[CONFLICT_RDX]], label %[[_LR_PH_I1]], label %[[VECTOR_PH:.*]]661; CHECK: [[VECTOR_PH]]:662; CHECK-NEXT: [[N_MOD_VF:%.*]] = urem i32 [[TMP1]], 2663; CHECK-NEXT: [[N_VEC:%.*]] = sub i32 [[TMP1]], [[N_MOD_VF]]664; CHECK-NEXT: [[TMP4:%.*]] = add i32 [[B_PROMOTED]], [[N_VEC]]665; CHECK-NEXT: br label %[[VECTOR_BODY:.*]]666; CHECK: [[VECTOR_BODY]]:667; CHECK-NEXT: [[INDEX:%.*]] = phi i32 [ 0, %[[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], %[[VECTOR_BODY]] ]668; CHECK-NEXT: [[OFFSET_IDX:%.*]] = add i32 [[B_PROMOTED]], [[INDEX]]669; CHECK-NEXT: [[TMP6:%.*]] = sext i32 [[OFFSET_IDX]] to i64670; CHECK-NEXT: [[TMP7:%.*]] = getelementptr inbounds i32, ptr [[B]], i64 [[TMP6]]671; CHECK-NEXT: [[WIDE_LOAD:%.*]] = load <2 x i32>, ptr [[TMP7]], align 4672; CHECK-NEXT: [[TMP9:%.*]] = getelementptr inbounds i32, ptr [[A]], i64 [[TMP6]]673; CHECK-NEXT: [[WIDE_LOAD5:%.*]] = load <2 x i32>, ptr [[TMP9]], align 4674; CHECK-NEXT: [[TMP11:%.*]] = add nsw <2 x i32> [[WIDE_LOAD]], [[WIDE_LOAD5]]675; CHECK-NEXT: [[TMP12:%.*]] = getelementptr inbounds i32, ptr [[C]], i64 [[TMP6]]676; CHECK-NEXT: store <2 x i32> [[TMP11]], ptr [[TMP12]], align 4677; CHECK-NEXT: [[INDEX_NEXT]] = add nuw i32 [[INDEX]], 2678; CHECK-NEXT: [[TMP14:%.*]] = icmp eq i32 [[INDEX_NEXT]], [[N_VEC]]679; CHECK-NEXT: br i1 [[TMP14]], label %[[MIDDLE_BLOCK:.*]], label %[[VECTOR_BODY]], !llvm.loop [[LOOP12:![0-9]+]]680; CHECK: [[MIDDLE_BLOCK]]:681; CHECK-NEXT: [[TMP15:%.*]] = extractelement <2 x i32> [[TMP11]], i32 1682; CHECK-NEXT: [[CMP_N:%.*]] = icmp eq i32 [[TMP1]], [[N_VEC]]683; CHECK-NEXT: br i1 [[CMP_N]], label %[[F1_EXIT_LOOPEXIT:.*]], label %[[_LR_PH_I1]]684; CHECK: [[_LR_PH_I1]]:685; CHECK-NEXT: [[BC_RESUME_VAL:%.*]] = phi i32 [ [[TMP4]], %[[MIDDLE_BLOCK]] ], [ [[B_PROMOTED]], %[[BB]] ], [ [[B_PROMOTED]], %[[VECTOR_MEMCHECK]] ]686; CHECK-NEXT: br label %[[DOTLR_PH_I:.*]]687; CHECK: [[_LR_PH_I:.*:]]688; CHECK-NEXT: [[IV:%.*]] = phi i32 [ [[IVNEXT:%.*]], %[[DOTLR_PH_I]] ], [ [[BC_RESUME_VAL]], %[[_LR_PH_I1]] ]689; CHECK-NEXT: [[INDVARS_IV:%.*]] = sext i32 [[IV]] to i64690; CHECK-NEXT: [[ARRAYIDX2:%.*]] = getelementptr inbounds i32, ptr [[B]], i64 [[INDVARS_IV]]691; CHECK-NEXT: [[BLOAD:%.*]] = load i32, ptr [[ARRAYIDX2]], align 4692; CHECK-NEXT: [[ARRAYIDX:%.*]] = getelementptr inbounds i32, ptr [[A]], i64 [[INDVARS_IV]]693; CHECK-NEXT: [[ALOAD:%.*]] = load i32, ptr [[ARRAYIDX]], align 4694; CHECK-NEXT: [[SUM:%.*]] = add nsw i32 [[BLOAD]], [[ALOAD]]695; CHECK-NEXT: [[ARRAYIDX3:%.*]] = getelementptr inbounds i32, ptr [[C]], i64 [[INDVARS_IV]]696; CHECK-NEXT: store i32 [[SUM]], ptr [[ARRAYIDX3]], align 4697; CHECK-NEXT: [[IVNEXT]] = add nsw i32 [[IV]], 1698; CHECK-NEXT: [[UNNAMEDTMP19:%.*]] = icmp slt i32 [[IVNEXT]], [[N]]699; CHECK-NEXT: br i1 [[UNNAMEDTMP19]], label %[[DOTLR_PH_I]], label %[[F1_EXIT_LOOPEXIT]], !llvm.loop [[LOOP13:![0-9]+]]700; CHECK: [[F1_EXIT_LOOPEXIT]]:701; CHECK-NEXT: [[DOTLCSSA:%.*]] = phi i32 [ [[SUM]], %[[DOTLR_PH_I]] ], [ [[TMP15]], %[[MIDDLE_BLOCK]] ]702; CHECK-NEXT: ret i32 [[DOTLCSSA]]703;704bb:705 %b.promoted = load i32, ptr @b, align 4706 br label %.lr.ph.i707 708.lr.ph.i:709 %iv = phi i32 [ %ivnext, %.lr.ph.i ], [ %b.promoted, %bb ]710 %indvars.iv = sext i32 %iv to i64711 %arrayidx2 = getelementptr inbounds i32, ptr %B, i64 %indvars.iv712 %Bload = load i32, ptr %arrayidx2, align 4713 %arrayidx = getelementptr inbounds i32, ptr %A, i64 %indvars.iv714 %Aload = load i32, ptr %arrayidx, align 4715 %sum = add nsw i32 %Bload, %Aload716 %arrayidx3 = getelementptr inbounds i32, ptr %C, i64 %indvars.iv717 store i32 %sum, ptr %arrayidx3, align 4718 %ivnext = add nsw i32 %iv, 1719 %tmp19 = icmp slt i32 %ivnext, %N720 br i1 %tmp19, label %.lr.ph.i, label %f1.exit.loopexit721 722f1.exit.loopexit:723 %.lcssa = phi i32 [ %sum, %.lr.ph.i ]724 ret i32 %.lcssa725}726 727@tab = common global [32 x i8] zeroinitializer, align 1728 729define i32 @non_uniform_live_out() {730; CHECK-LABEL: define i32 @non_uniform_live_out() {731; CHECK-NEXT: [[ENTRY:.*:]]732; CHECK-NEXT: br label %[[VECTOR_PH:.*]]733; CHECK: [[VECTOR_PH]]:734; CHECK-NEXT: br label %[[VECTOR_BODY:.*]]735; CHECK: [[VECTOR_BODY]]:736; CHECK-NEXT: [[INDEX:%.*]] = phi i32 [ 0, %[[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], %[[VECTOR_BODY]] ]737; CHECK-NEXT: [[VEC_IND:%.*]] = phi <2 x i32> [ <i32 0, i32 1>, %[[VECTOR_PH]] ], [ [[VEC_IND_NEXT:%.*]], %[[VECTOR_BODY]] ]738; CHECK-NEXT: [[TMP0:%.*]] = add <2 x i32> [[VEC_IND]], splat (i32 7)739; CHECK-NEXT: [[TMP1:%.*]] = extractelement <2 x i32> [[TMP0]], i32 0740; CHECK-NEXT: [[TMP2:%.*]] = getelementptr inbounds [32 x i8], ptr @tab, i32 0, i32 [[TMP1]]741; CHECK-NEXT: [[WIDE_LOAD:%.*]] = load <2 x i8>, ptr [[TMP2]], align 1742; CHECK-NEXT: [[TMP4:%.*]] = add <2 x i8> [[WIDE_LOAD]], splat (i8 1)743; CHECK-NEXT: store <2 x i8> [[TMP4]], ptr [[TMP2]], align 1744; CHECK-NEXT: [[INDEX_NEXT]] = add nuw i32 [[INDEX]], 2745; CHECK-NEXT: [[VEC_IND_NEXT]] = add nsw <2 x i32> [[VEC_IND]], splat (i32 2)746; CHECK-NEXT: [[TMP5:%.*]] = icmp eq i32 [[INDEX_NEXT]], 20000747; CHECK-NEXT: br i1 [[TMP5]], label %[[MIDDLE_BLOCK:.*]], label %[[VECTOR_BODY]], !llvm.loop [[LOOP14:![0-9]+]]748; CHECK: [[MIDDLE_BLOCK]]:749; CHECK-NEXT: br label %[[SCALAR_PH:.*]]750; CHECK: [[SCALAR_PH]]:751; CHECK-NEXT: br label %[[FOR_BODY:.*]]752; CHECK: [[FOR_BODY]]:753; CHECK-NEXT: [[I_08:%.*]] = phi i32 [ 20000, %[[SCALAR_PH]] ], [ [[INC:%.*]], %[[FOR_BODY]] ]754; CHECK-NEXT: [[I_09:%.*]] = add i32 [[I_08]], 7755; CHECK-NEXT: [[ARRAYIDX:%.*]] = getelementptr inbounds [32 x i8], ptr @tab, i32 0, i32 [[I_09]]756; CHECK-NEXT: [[TMP7:%.*]] = load i8, ptr [[ARRAYIDX]], align 1757; CHECK-NEXT: [[BUMP:%.*]] = add i8 [[TMP7]], 1758; CHECK-NEXT: store i8 [[BUMP]], ptr [[ARRAYIDX]], align 1759; CHECK-NEXT: [[INC]] = add nsw i32 [[I_08]], 1760; CHECK-NEXT: [[EXITCOND:%.*]] = icmp eq i32 [[I_08]], 20000761; CHECK-NEXT: br i1 [[EXITCOND]], label %[[FOR_END:.*]], label %[[FOR_BODY]], !llvm.loop [[LOOP15:![0-9]+]]762; CHECK: [[FOR_END]]:763; CHECK-NEXT: [[LCSSA:%.*]] = phi i32 [ [[I_09]], %[[FOR_BODY]] ]764; CHECK-NEXT: [[ARRAYIDX_OUT:%.*]] = getelementptr inbounds [32 x i8], ptr @tab, i32 0, i32 [[LCSSA]]765; CHECK-NEXT: store i8 42, ptr [[ARRAYIDX_OUT]], align 1766; CHECK-NEXT: ret i32 0767;768entry:769 br label %for.body770 771for.body: ; preds = %for.body, %entry772 %i.08 = phi i32 [ 0, %entry ], [ %inc, %for.body ]773 %i.09 = add i32 %i.08, 7774 %arrayidx = getelementptr inbounds [32 x i8], ptr @tab, i32 0, i32 %i.09775 %0 = load i8, ptr %arrayidx, align 1776 %bump = add i8 %0, 1777 store i8 %bump, ptr %arrayidx, align 1778 %inc = add nsw i32 %i.08, 1779 %exitcond = icmp eq i32 %i.08, 20000780 br i1 %exitcond, label %for.end, label %for.body781 782for.end: ; preds = %for.body783 %lcssa = phi i32 [%i.09, %for.body]784 %arrayidx.out = getelementptr inbounds [32 x i8], ptr @tab, i32 0, i32 %lcssa785 store i8 42, ptr %arrayidx.out, align 1786 ret i32 0787}788;.789; CHECK: [[LOOP0]] = distinct !{[[LOOP0]], [[META1:![0-9]+]], [[META2:![0-9]+]]}790; CHECK: [[META1]] = !{!"llvm.loop.isvectorized", i32 1}791; CHECK: [[META2]] = !{!"llvm.loop.unroll.runtime.disable"}792; CHECK: [[LOOP3]] = distinct !{[[LOOP3]], [[META2]], [[META1]]}793; CHECK: [[LOOP4]] = distinct !{[[LOOP4]], [[META1]], [[META2]]}794; CHECK: [[LOOP5]] = distinct !{[[LOOP5]], [[META2]], [[META1]]}795; CHECK: [[LOOP6]] = distinct !{[[LOOP6]], [[META1]], [[META2]]}796; CHECK: [[LOOP7]] = distinct !{[[LOOP7]], [[META2]], [[META1]]}797; CHECK: [[LOOP8]] = distinct !{[[LOOP8]], [[META1]], [[META2]]}798; CHECK: [[LOOP9]] = distinct !{[[LOOP9]], [[META2]], [[META1]]}799; CHECK: [[LOOP10]] = distinct !{[[LOOP10]], [[META1]], [[META2]]}800; CHECK: [[LOOP11]] = distinct !{[[LOOP11]], [[META2]], [[META1]]}801; CHECK: [[LOOP12]] = distinct !{[[LOOP12]], [[META1]], [[META2]]}802; CHECK: [[LOOP13]] = distinct !{[[LOOP13]], [[META1]]}803; CHECK: [[LOOP14]] = distinct !{[[LOOP14]], [[META1]], [[META2]]}804; CHECK: [[LOOP15]] = distinct !{[[LOOP15]], [[META2]], [[META1]]}805;.806