blob: 20f6c96e778b6f6f4603fa49962100e7003d3686 [file] [edit]
; NOTE: Assertions have been autogenerated by utils/update_test_checks.py UTC_ARGS: --version 6
; RUN: opt -mtriple=wasm32 -mattr=+simd128 -passes=loop-vectorize %s -S -o - | FileCheck %s
; RUN: opt -mtriple=wasm32 -mattr=+simd128 -passes=loop-vectorize %s --vectorizer-maximize-bandwidth -S -o - | FileCheck %s --check-prefix=MAXIMIZE
define dso_local void @aim_for_vf8(ptr %0, i32 %1, ptr %2, ptr %3, i1 %4, ptr %5, i32 %6, i32 %7, i32 %8, i32 %9, ptr %.out, ptr %.out1, ptr %.out2, ptr %.out3, ptr %.out4) {
; CHECK-LABEL: define dso_local void @aim_for_vf8(
; CHECK-SAME: ptr [[TMP0:%.*]], i32 [[TMP1:%.*]], ptr [[TMP2:%.*]], ptr [[TMP3:%.*]], i1 [[TMP4:%.*]], ptr [[TMP5:%.*]], i32 [[TMP6:%.*]], i32 [[TMP7:%.*]], i32 [[TMP8:%.*]], i32 [[TMP9:%.*]], ptr [[DOTOUT:%.*]], ptr [[DOTOUT1:%.*]], ptr [[DOTOUT2:%.*]], ptr [[DOTOUT3:%.*]], ptr [[DOTOUT4:%.*]]) #[[ATTR0:[0-9]+]] {
; CHECK-NEXT: [[NEWFUNCROOT:.*]]:
; CHECK-NEXT: br label %[[BB10:.*]]
; CHECK: [[BB10]]:
; CHECK-NEXT: [[TMP11:%.*]] = phi i32 [ 0, %[[NEWFUNCROOT]] ], [ [[TMP168:%.*]], %[[TMP163:.*]] ]
; CHECK-NEXT: [[TMP12:%.*]] = phi i32 [ 0, %[[NEWFUNCROOT]] ], [ [[TMP167:%.*]], %[[TMP163]] ]
; CHECK-NEXT: [[TMP13:%.*]] = phi i32 [ 0, %[[NEWFUNCROOT]] ], [ [[TMP166:%.*]], %[[TMP163]] ]
; CHECK-NEXT: [[TMP14:%.*]] = phi i32 [ 0, %[[NEWFUNCROOT]] ], [ [[TMP165:%.*]], %[[TMP163]] ]
; CHECK-NEXT: [[TMP15:%.*]] = phi ptr [ [[TMP0]], %[[NEWFUNCROOT]] ], [ [[TMP164:%.*]], %[[TMP163]] ]
; CHECK-NEXT: [[TMP16:%.*]] = phi i32 [ 0, %[[NEWFUNCROOT]] ], [ [[TMP169:%.*]], %[[TMP163]] ]
; CHECK-NEXT: [[TMP17:%.*]] = mul i32 [[TMP1]], [[TMP16]]
; CHECK-NEXT: [[TMP18:%.*]] = getelementptr i8, ptr [[TMP2]], i32 [[TMP17]]
; CHECK-NEXT: [[TMP19:%.*]] = getelementptr i8, ptr [[TMP3]], i32 [[TMP17]]
; CHECK-NEXT: br i1 [[TMP4]], label %[[DOTPREHEADER:.*]], label %[[TMP163]]
; CHECK: [[DOTPREHEADER]]:
; CHECK-NEXT: [[MIN_ITERS_CHECK:%.*]] = icmp ult i32 [[TMP7]], 4
; CHECK-NEXT: br i1 [[MIN_ITERS_CHECK]], label %[[SCALAR_PH:.*]], label %[[VECTOR_PH:.*]]
; CHECK: [[VECTOR_PH]]:
; CHECK-NEXT: [[TMP20:%.*]] = and i32 [[TMP7]], 3
; CHECK-NEXT: [[N_VEC:%.*]] = sub i32 [[TMP7]], [[TMP20]]
; CHECK-NEXT: [[TMP21:%.*]] = shl i32 [[N_VEC]], 2
; CHECK-NEXT: [[TMP22:%.*]] = getelementptr i8, ptr [[TMP15]], i32 [[TMP21]]
; CHECK-NEXT: [[TMP23:%.*]] = insertelement <4 x i32> zeroinitializer, i32 [[TMP11]], i64 0
; CHECK-NEXT: [[TMP24:%.*]] = insertelement <4 x i32> zeroinitializer, i32 [[TMP12]], i64 0
; CHECK-NEXT: [[TMP25:%.*]] = insertelement <4 x i32> zeroinitializer, i32 [[TMP13]], i64 0
; CHECK-NEXT: [[TMP26:%.*]] = insertelement <4 x i32> zeroinitializer, i32 [[TMP14]], i64 0
; CHECK-NEXT: br label %[[VECTOR_BODY:.*]]
; CHECK: [[VECTOR_BODY]]:
; CHECK-NEXT: [[INDEX:%.*]] = phi i32 [ 0, %[[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], %[[VECTOR_BODY]] ]
; CHECK-NEXT: [[VEC_PHI:%.*]] = phi <4 x i32> [ [[TMP23]], %[[VECTOR_PH]] ], [ [[TMP71:%.*]], %[[VECTOR_BODY]] ]
; CHECK-NEXT: [[VEC_PHI4:%.*]] = phi <4 x i32> [ [[TMP24]], %[[VECTOR_PH]] ], [ [[TMP102:%.*]], %[[VECTOR_BODY]] ]
; CHECK-NEXT: [[VEC_PHI5:%.*]] = phi <4 x i32> [ [[TMP25]], %[[VECTOR_PH]] ], [ [[TMP74:%.*]], %[[VECTOR_BODY]] ]
; CHECK-NEXT: [[VEC_PHI6:%.*]] = phi <4 x i32> [ [[TMP26]], %[[VECTOR_PH]] ], [ [[TMP104:%.*]], %[[VECTOR_BODY]] ]
; CHECK-NEXT: [[TMP27:%.*]] = shl i32 [[INDEX]], 2
; CHECK-NEXT: [[NEXT_GEP:%.*]] = getelementptr i8, ptr [[TMP15]], i32 [[TMP27]]
; CHECK-NEXT: [[TMP28:%.*]] = getelementptr inbounds [4 x i8], ptr [[TMP5]], i32 [[INDEX]]
; CHECK-NEXT: [[WIDE_LOAD:%.*]] = load <4 x i32>, ptr [[TMP28]], align 4
; CHECK-NEXT: [[TMP29:%.*]] = extractelement <4 x i32> [[WIDE_LOAD]], i64 0
; CHECK-NEXT: [[TMP30:%.*]] = getelementptr inbounds i8, ptr [[TMP18]], i32 [[TMP29]]
; CHECK-NEXT: [[TMP31:%.*]] = extractelement <4 x i32> [[WIDE_LOAD]], i64 1
; CHECK-NEXT: [[TMP32:%.*]] = getelementptr inbounds i8, ptr [[TMP18]], i32 [[TMP31]]
; CHECK-NEXT: [[TMP33:%.*]] = extractelement <4 x i32> [[WIDE_LOAD]], i64 2
; CHECK-NEXT: [[TMP34:%.*]] = getelementptr inbounds i8, ptr [[TMP18]], i32 [[TMP33]]
; CHECK-NEXT: [[TMP35:%.*]] = extractelement <4 x i32> [[WIDE_LOAD]], i64 3
; CHECK-NEXT: [[TMP36:%.*]] = getelementptr inbounds i8, ptr [[TMP18]], i32 [[TMP35]]
; CHECK-NEXT: [[TMP37:%.*]] = getelementptr inbounds i8, ptr [[TMP19]], i32 [[TMP29]]
; CHECK-NEXT: [[TMP38:%.*]] = getelementptr inbounds i8, ptr [[TMP19]], i32 [[TMP31]]
; CHECK-NEXT: [[TMP39:%.*]] = getelementptr inbounds i8, ptr [[TMP19]], i32 [[TMP33]]
; CHECK-NEXT: [[TMP40:%.*]] = getelementptr inbounds i8, ptr [[TMP19]], i32 [[TMP35]]
; CHECK-NEXT: [[TMP41:%.*]] = load i8, ptr [[TMP30]], align 1
; CHECK-NEXT: [[TMP42:%.*]] = load i8, ptr [[TMP32]], align 1
; CHECK-NEXT: [[TMP43:%.*]] = load i8, ptr [[TMP34]], align 1
; CHECK-NEXT: [[TMP44:%.*]] = load i8, ptr [[TMP36]], align 1
; CHECK-NEXT: [[TMP45:%.*]] = insertelement <4 x i8> poison, i8 [[TMP41]], i64 0
; CHECK-NEXT: [[TMP46:%.*]] = insertelement <4 x i8> [[TMP45]], i8 [[TMP42]], i64 1
; CHECK-NEXT: [[TMP47:%.*]] = insertelement <4 x i8> [[TMP46]], i8 [[TMP43]], i64 2
; CHECK-NEXT: [[TMP48:%.*]] = insertelement <4 x i8> [[TMP47]], i8 [[TMP44]], i64 3
; CHECK-NEXT: [[TMP49:%.*]] = sext <4 x i8> [[TMP48]] to <4 x i32>
; CHECK-NEXT: [[WIDE_VEC:%.*]] = load <16 x i8>, ptr [[NEXT_GEP]], align 1
; CHECK-NEXT: [[STRIDED_VEC:%.*]] = shufflevector <16 x i8> [[WIDE_VEC]], <16 x i8> poison, <4 x i32> <i32 0, i32 4, i32 8, i32 12>
; CHECK-NEXT: [[STRIDED_VEC7:%.*]] = shufflevector <16 x i8> [[WIDE_VEC]], <16 x i8> poison, <4 x i32> <i32 1, i32 5, i32 9, i32 13>
; CHECK-NEXT: [[STRIDED_VEC8:%.*]] = shufflevector <16 x i8> [[WIDE_VEC]], <16 x i8> poison, <4 x i32> <i32 2, i32 6, i32 10, i32 14>
; CHECK-NEXT: [[STRIDED_VEC9:%.*]] = shufflevector <16 x i8> [[WIDE_VEC]], <16 x i8> poison, <4 x i32> <i32 3, i32 7, i32 11, i32 15>
; CHECK-NEXT: [[TMP50:%.*]] = sext <4 x i8> [[STRIDED_VEC]] to <4 x i32>
; CHECK-NEXT: [[TMP51:%.*]] = mul nsw <4 x i32> [[TMP50]], [[TMP49]]
; CHECK-NEXT: [[TMP52:%.*]] = add <4 x i32> [[TMP51]], [[VEC_PHI]]
; CHECK-NEXT: [[TMP53:%.*]] = sext <4 x i8> [[STRIDED_VEC7]] to <4 x i32>
; CHECK-NEXT: [[TMP54:%.*]] = mul nsw <4 x i32> [[TMP53]], [[TMP49]]
; CHECK-NEXT: [[TMP55:%.*]] = add <4 x i32> [[TMP54]], [[VEC_PHI5]]
; CHECK-NEXT: [[TMP56:%.*]] = getelementptr inbounds i8, ptr [[TMP30]], i32 [[TMP6]]
; CHECK-NEXT: [[TMP57:%.*]] = getelementptr inbounds i8, ptr [[TMP32]], i32 [[TMP6]]
; CHECK-NEXT: [[TMP58:%.*]] = getelementptr inbounds i8, ptr [[TMP34]], i32 [[TMP6]]
; CHECK-NEXT: [[TMP59:%.*]] = getelementptr inbounds i8, ptr [[TMP36]], i32 [[TMP6]]
; CHECK-NEXT: [[TMP60:%.*]] = load i8, ptr [[TMP56]], align 1
; CHECK-NEXT: [[TMP61:%.*]] = load i8, ptr [[TMP57]], align 1
; CHECK-NEXT: [[TMP62:%.*]] = load i8, ptr [[TMP58]], align 1
; CHECK-NEXT: [[TMP63:%.*]] = load i8, ptr [[TMP59]], align 1
; CHECK-NEXT: [[TMP64:%.*]] = insertelement <4 x i8> poison, i8 [[TMP60]], i64 0
; CHECK-NEXT: [[TMP65:%.*]] = insertelement <4 x i8> [[TMP64]], i8 [[TMP61]], i64 1
; CHECK-NEXT: [[TMP66:%.*]] = insertelement <4 x i8> [[TMP65]], i8 [[TMP62]], i64 2
; CHECK-NEXT: [[TMP67:%.*]] = insertelement <4 x i8> [[TMP66]], i8 [[TMP63]], i64 3
; CHECK-NEXT: [[TMP68:%.*]] = sext <4 x i8> [[TMP67]] to <4 x i32>
; CHECK-NEXT: [[TMP69:%.*]] = sext <4 x i8> [[STRIDED_VEC8]] to <4 x i32>
; CHECK-NEXT: [[TMP70:%.*]] = mul nsw <4 x i32> [[TMP69]], [[TMP68]]
; CHECK-NEXT: [[TMP71]] = add <4 x i32> [[TMP52]], [[TMP70]]
; CHECK-NEXT: [[TMP72:%.*]] = sext <4 x i8> [[STRIDED_VEC9]] to <4 x i32>
; CHECK-NEXT: [[TMP73:%.*]] = mul nsw <4 x i32> [[TMP72]], [[TMP68]]
; CHECK-NEXT: [[TMP74]] = add <4 x i32> [[TMP55]], [[TMP73]]
; CHECK-NEXT: [[TMP75:%.*]] = load i8, ptr [[TMP37]], align 1
; CHECK-NEXT: [[TMP76:%.*]] = load i8, ptr [[TMP38]], align 1
; CHECK-NEXT: [[TMP77:%.*]] = load i8, ptr [[TMP39]], align 1
; CHECK-NEXT: [[TMP78:%.*]] = load i8, ptr [[TMP40]], align 1
; CHECK-NEXT: [[TMP79:%.*]] = insertelement <4 x i8> poison, i8 [[TMP75]], i64 0
; CHECK-NEXT: [[TMP80:%.*]] = insertelement <4 x i8> [[TMP79]], i8 [[TMP76]], i64 1
; CHECK-NEXT: [[TMP81:%.*]] = insertelement <4 x i8> [[TMP80]], i8 [[TMP77]], i64 2
; CHECK-NEXT: [[TMP82:%.*]] = insertelement <4 x i8> [[TMP81]], i8 [[TMP78]], i64 3
; CHECK-NEXT: [[TMP83:%.*]] = sext <4 x i8> [[TMP82]] to <4 x i32>
; CHECK-NEXT: [[TMP84:%.*]] = mul nsw <4 x i32> [[TMP83]], [[TMP50]]
; CHECK-NEXT: [[TMP85:%.*]] = add <4 x i32> [[TMP84]], [[VEC_PHI4]]
; CHECK-NEXT: [[TMP86:%.*]] = mul nsw <4 x i32> [[TMP83]], [[TMP53]]
; CHECK-NEXT: [[TMP87:%.*]] = add <4 x i32> [[TMP86]], [[VEC_PHI6]]
; CHECK-NEXT: [[TMP88:%.*]] = getelementptr inbounds i8, ptr [[TMP37]], i32 [[TMP6]]
; CHECK-NEXT: [[TMP89:%.*]] = getelementptr inbounds i8, ptr [[TMP38]], i32 [[TMP6]]
; CHECK-NEXT: [[TMP90:%.*]] = getelementptr inbounds i8, ptr [[TMP39]], i32 [[TMP6]]
; CHECK-NEXT: [[TMP91:%.*]] = getelementptr inbounds i8, ptr [[TMP40]], i32 [[TMP6]]
; CHECK-NEXT: [[TMP92:%.*]] = load i8, ptr [[TMP88]], align 1
; CHECK-NEXT: [[TMP93:%.*]] = load i8, ptr [[TMP89]], align 1
; CHECK-NEXT: [[TMP94:%.*]] = load i8, ptr [[TMP90]], align 1
; CHECK-NEXT: [[TMP95:%.*]] = load i8, ptr [[TMP91]], align 1
; CHECK-NEXT: [[TMP96:%.*]] = insertelement <4 x i8> poison, i8 [[TMP92]], i64 0
; CHECK-NEXT: [[TMP97:%.*]] = insertelement <4 x i8> [[TMP96]], i8 [[TMP93]], i64 1
; CHECK-NEXT: [[TMP98:%.*]] = insertelement <4 x i8> [[TMP97]], i8 [[TMP94]], i64 2
; CHECK-NEXT: [[TMP99:%.*]] = insertelement <4 x i8> [[TMP98]], i8 [[TMP95]], i64 3
; CHECK-NEXT: [[TMP100:%.*]] = sext <4 x i8> [[TMP99]] to <4 x i32>
; CHECK-NEXT: [[TMP101:%.*]] = mul nsw <4 x i32> [[TMP100]], [[TMP69]]
; CHECK-NEXT: [[TMP102]] = add <4 x i32> [[TMP85]], [[TMP101]]
; CHECK-NEXT: [[TMP103:%.*]] = mul nsw <4 x i32> [[TMP100]], [[TMP72]]
; CHECK-NEXT: [[TMP104]] = add <4 x i32> [[TMP87]], [[TMP103]]
; CHECK-NEXT: [[INDEX_NEXT]] = add nuw i32 [[INDEX]], 4
; CHECK-NEXT: [[TMP105:%.*]] = icmp eq i32 [[INDEX_NEXT]], [[N_VEC]]
; CHECK-NEXT: br i1 [[TMP105]], label %[[MIDDLE_BLOCK:.*]], label %[[VECTOR_BODY]], !llvm.loop [[LOOP0:![0-9]+]]
; CHECK: [[MIDDLE_BLOCK]]:
; CHECK-NEXT: [[TMP106:%.*]] = call i32 @llvm.vector.reduce.add.v4i32(<4 x i32> [[TMP71]])
; CHECK-NEXT: [[TMP107:%.*]] = call i32 @llvm.vector.reduce.add.v4i32(<4 x i32> [[TMP102]])
; CHECK-NEXT: [[TMP108:%.*]] = call i32 @llvm.vector.reduce.add.v4i32(<4 x i32> [[TMP74]])
; CHECK-NEXT: [[TMP109:%.*]] = call i32 @llvm.vector.reduce.add.v4i32(<4 x i32> [[TMP104]])
; CHECK-NEXT: [[CMP_N:%.*]] = icmp eq i32 [[TMP7]], [[N_VEC]]
; CHECK-NEXT: br i1 [[CMP_N]], label %[[BB161:.*]], label %[[SCALAR_PH]]
; CHECK: [[SCALAR_PH]]:
; CHECK-NEXT: [[BC_MERGE_RDX:%.*]] = phi i32 [ [[TMP106]], %[[MIDDLE_BLOCK]] ], [ [[TMP11]], %[[DOTPREHEADER]] ]
; CHECK-NEXT: [[BC_MERGE_RDX10:%.*]] = phi i32 [ [[TMP107]], %[[MIDDLE_BLOCK]] ], [ [[TMP12]], %[[DOTPREHEADER]] ]
; CHECK-NEXT: [[BC_MERGE_RDX11:%.*]] = phi i32 [ [[TMP108]], %[[MIDDLE_BLOCK]] ], [ [[TMP13]], %[[DOTPREHEADER]] ]
; CHECK-NEXT: [[BC_MERGE_RDX12:%.*]] = phi i32 [ [[TMP109]], %[[MIDDLE_BLOCK]] ], [ [[TMP14]], %[[DOTPREHEADER]] ]
; CHECK-NEXT: [[BC_RESUME_VAL:%.*]] = phi ptr [ [[TMP22]], %[[MIDDLE_BLOCK]] ], [ [[TMP15]], %[[DOTPREHEADER]] ]
; CHECK-NEXT: [[BC_RESUME_VAL13:%.*]] = phi i32 [ [[N_VEC]], %[[MIDDLE_BLOCK]] ], [ 0, %[[DOTPREHEADER]] ]
; CHECK-NEXT: br label %[[BB110:.*]]
; CHECK: [[BB110]]:
; CHECK-NEXT: [[TMP111:%.*]] = phi i32 [ [[TMP139:%.*]], %[[BB110]] ], [ [[BC_MERGE_RDX]], %[[SCALAR_PH]] ]
; CHECK-NEXT: [[TMP112:%.*]] = phi i32 [ [[TMP155:%.*]], %[[BB110]] ], [ [[BC_MERGE_RDX10]], %[[SCALAR_PH]] ]
; CHECK-NEXT: [[TMP113:%.*]] = phi i32 [ [[TMP144:%.*]], %[[BB110]] ], [ [[BC_MERGE_RDX11]], %[[SCALAR_PH]] ]
; CHECK-NEXT: [[TMP114:%.*]] = phi i32 [ [[TMP157:%.*]], %[[BB110]] ], [ [[BC_MERGE_RDX12]], %[[SCALAR_PH]] ]
; CHECK-NEXT: [[TMP115:%.*]] = phi ptr [ [[TMP158:%.*]], %[[BB110]] ], [ [[BC_RESUME_VAL]], %[[SCALAR_PH]] ]
; CHECK-NEXT: [[TMP116:%.*]] = phi i32 [ [[TMP159:%.*]], %[[BB110]] ], [ [[BC_RESUME_VAL13]], %[[SCALAR_PH]] ]
; CHECK-NEXT: [[TMP117:%.*]] = getelementptr inbounds [4 x i8], ptr [[TMP5]], i32 [[TMP116]]
; CHECK-NEXT: [[TMP118:%.*]] = load i32, ptr [[TMP117]], align 4
; CHECK-NEXT: [[TMP119:%.*]] = getelementptr inbounds i8, ptr [[TMP18]], i32 [[TMP118]]
; CHECK-NEXT: [[TMP120:%.*]] = getelementptr inbounds i8, ptr [[TMP19]], i32 [[TMP118]]
; CHECK-NEXT: [[TMP121:%.*]] = load i8, ptr [[TMP119]], align 1
; CHECK-NEXT: [[TMP122:%.*]] = sext i8 [[TMP121]] to i32
; CHECK-NEXT: [[TMP123:%.*]] = load i8, ptr [[TMP115]], align 1
; CHECK-NEXT: [[TMP124:%.*]] = sext i8 [[TMP123]] to i32
; CHECK-NEXT: [[TMP125:%.*]] = mul nsw i32 [[TMP124]], [[TMP122]]
; CHECK-NEXT: [[TMP126:%.*]] = add nsw i32 [[TMP125]], [[TMP111]]
; CHECK-NEXT: [[TMP127:%.*]] = getelementptr inbounds i8, ptr [[TMP115]], i32 1
; CHECK-NEXT: [[TMP128:%.*]] = load i8, ptr [[TMP127]], align 1
; CHECK-NEXT: [[TMP129:%.*]] = sext i8 [[TMP128]] to i32
; CHECK-NEXT: [[TMP130:%.*]] = mul nsw i32 [[TMP129]], [[TMP122]]
; CHECK-NEXT: [[TMP131:%.*]] = add nsw i32 [[TMP130]], [[TMP113]]
; CHECK-NEXT: [[TMP132:%.*]] = getelementptr inbounds i8, ptr [[TMP119]], i32 [[TMP6]]
; CHECK-NEXT: [[TMP133:%.*]] = load i8, ptr [[TMP132]], align 1
; CHECK-NEXT: [[TMP134:%.*]] = sext i8 [[TMP133]] to i32
; CHECK-NEXT: [[TMP135:%.*]] = getelementptr inbounds i8, ptr [[TMP115]], i32 2
; CHECK-NEXT: [[TMP136:%.*]] = load i8, ptr [[TMP135]], align 1
; CHECK-NEXT: [[TMP137:%.*]] = sext i8 [[TMP136]] to i32
; CHECK-NEXT: [[TMP138:%.*]] = mul nsw i32 [[TMP137]], [[TMP134]]
; CHECK-NEXT: [[TMP139]] = add nsw i32 [[TMP126]], [[TMP138]]
; CHECK-NEXT: [[TMP140:%.*]] = getelementptr inbounds i8, ptr [[TMP115]], i32 3
; CHECK-NEXT: [[TMP141:%.*]] = load i8, ptr [[TMP140]], align 1
; CHECK-NEXT: [[TMP142:%.*]] = sext i8 [[TMP141]] to i32
; CHECK-NEXT: [[TMP143:%.*]] = mul nsw i32 [[TMP142]], [[TMP134]]
; CHECK-NEXT: [[TMP144]] = add nsw i32 [[TMP131]], [[TMP143]]
; CHECK-NEXT: [[TMP145:%.*]] = load i8, ptr [[TMP120]], align 1
; CHECK-NEXT: [[TMP146:%.*]] = sext i8 [[TMP145]] to i32
; CHECK-NEXT: [[TMP147:%.*]] = mul nsw i32 [[TMP146]], [[TMP124]]
; CHECK-NEXT: [[TMP148:%.*]] = add nsw i32 [[TMP147]], [[TMP112]]
; CHECK-NEXT: [[TMP149:%.*]] = mul nsw i32 [[TMP146]], [[TMP129]]
; CHECK-NEXT: [[TMP150:%.*]] = add nsw i32 [[TMP149]], [[TMP114]]
; CHECK-NEXT: [[TMP151:%.*]] = getelementptr inbounds i8, ptr [[TMP120]], i32 [[TMP6]]
; CHECK-NEXT: [[TMP152:%.*]] = load i8, ptr [[TMP151]], align 1
; CHECK-NEXT: [[TMP153:%.*]] = sext i8 [[TMP152]] to i32
; CHECK-NEXT: [[TMP154:%.*]] = mul nsw i32 [[TMP153]], [[TMP137]]
; CHECK-NEXT: [[TMP155]] = add nsw i32 [[TMP148]], [[TMP154]]
; CHECK-NEXT: [[TMP156:%.*]] = mul nsw i32 [[TMP153]], [[TMP142]]
; CHECK-NEXT: [[TMP157]] = add nsw i32 [[TMP150]], [[TMP156]]
; CHECK-NEXT: [[TMP158]] = getelementptr inbounds i8, ptr [[TMP115]], i32 4
; CHECK-NEXT: [[TMP159]] = add nuw nsw i32 [[TMP116]], 1
; CHECK-NEXT: [[TMP160:%.*]] = icmp eq i32 [[TMP159]], [[TMP7]]
; CHECK-NEXT: br i1 [[TMP160]], label %[[BB161]], label %[[BB110]], !llvm.loop [[LOOP3:![0-9]+]]
; CHECK: [[BB161]]:
; CHECK-NEXT: [[DOTLCSSA3:%.*]] = phi i32 [ [[TMP139]], %[[BB110]] ], [ [[TMP106]], %[[MIDDLE_BLOCK]] ]
; CHECK-NEXT: [[DOTLCSSA2:%.*]] = phi i32 [ [[TMP144]], %[[BB110]] ], [ [[TMP108]], %[[MIDDLE_BLOCK]] ]
; CHECK-NEXT: [[DOTLCSSA1:%.*]] = phi i32 [ [[TMP155]], %[[BB110]] ], [ [[TMP107]], %[[MIDDLE_BLOCK]] ]
; CHECK-NEXT: [[DOTLCSSA:%.*]] = phi i32 [ [[TMP157]], %[[BB110]] ], [ [[TMP109]], %[[MIDDLE_BLOCK]] ]
; CHECK-NEXT: [[TMP162:%.*]] = getelementptr i8, ptr [[TMP15]], i32 [[TMP8]]
; CHECK-NEXT: br label %[[TMP163]]
; CHECK: [[TMP163]]:
; CHECK-NEXT: [[TMP164]] = phi ptr [ [[TMP15]], %[[BB10]] ], [ [[TMP162]], %[[BB161]] ]
; CHECK-NEXT: [[TMP165]] = phi i32 [ [[TMP14]], %[[BB10]] ], [ [[DOTLCSSA]], %[[BB161]] ]
; CHECK-NEXT: [[TMP166]] = phi i32 [ [[TMP13]], %[[BB10]] ], [ [[DOTLCSSA2]], %[[BB161]] ]
; CHECK-NEXT: [[TMP167]] = phi i32 [ [[TMP12]], %[[BB10]] ], [ [[DOTLCSSA1]], %[[BB161]] ]
; CHECK-NEXT: [[TMP168]] = phi i32 [ [[TMP11]], %[[BB10]] ], [ [[DOTLCSSA3]], %[[BB161]] ]
; CHECK-NEXT: store i32 [[TMP168]], ptr [[DOTOUT4]], align 4
; CHECK-NEXT: store i32 [[TMP167]], ptr [[DOTOUT3]], align 4
; CHECK-NEXT: store i32 [[TMP166]], ptr [[DOTOUT2]], align 4
; CHECK-NEXT: store i32 [[TMP165]], ptr [[DOTOUT1]], align 4
; CHECK-NEXT: store ptr [[TMP164]], ptr [[DOTOUT]], align 4
; CHECK-NEXT: [[TMP169]] = add nuw nsw i32 [[TMP16]], 2
; CHECK-NEXT: [[TMP170:%.*]] = or disjoint i32 [[TMP169]], 1
; CHECK-NEXT: [[TMP171:%.*]] = icmp slt i32 [[TMP170]], [[TMP9]]
; CHECK-NEXT: br i1 [[TMP171]], label %[[BB10]], label %[[DOTEXITSTUB:.*]]
; CHECK: [[DOTEXITSTUB]]:
; CHECK-NEXT: ret void
;
; MAXIMIZE-LABEL: define dso_local void @aim_for_vf8(
; MAXIMIZE-SAME: ptr [[TMP0:%.*]], i32 [[TMP1:%.*]], ptr [[TMP2:%.*]], ptr [[TMP3:%.*]], i1 [[TMP4:%.*]], ptr [[TMP5:%.*]], i32 [[TMP6:%.*]], i32 [[TMP7:%.*]], i32 [[TMP8:%.*]], i32 [[TMP9:%.*]], ptr [[DOTOUT:%.*]], ptr [[DOTOUT1:%.*]], ptr [[DOTOUT2:%.*]], ptr [[DOTOUT3:%.*]], ptr [[DOTOUT4:%.*]]) #[[ATTR0:[0-9]+]] {
; MAXIMIZE-NEXT: [[NEWFUNCROOT:.*]]:
; MAXIMIZE-NEXT: br label %[[BB10:.*]]
; MAXIMIZE: [[BB10]]:
; MAXIMIZE-NEXT: [[TMP11:%.*]] = phi i32 [ 0, %[[NEWFUNCROOT]] ], [ [[TMP406:%.*]], %[[TMP401:.*]] ]
; MAXIMIZE-NEXT: [[TMP12:%.*]] = phi i32 [ 0, %[[NEWFUNCROOT]] ], [ [[TMP405:%.*]], %[[TMP401]] ]
; MAXIMIZE-NEXT: [[TMP13:%.*]] = phi i32 [ 0, %[[NEWFUNCROOT]] ], [ [[TMP404:%.*]], %[[TMP401]] ]
; MAXIMIZE-NEXT: [[TMP14:%.*]] = phi i32 [ 0, %[[NEWFUNCROOT]] ], [ [[TMP403:%.*]], %[[TMP401]] ]
; MAXIMIZE-NEXT: [[TMP15:%.*]] = phi ptr [ [[TMP0]], %[[NEWFUNCROOT]] ], [ [[TMP402:%.*]], %[[TMP401]] ]
; MAXIMIZE-NEXT: [[TMP16:%.*]] = phi i32 [ 0, %[[NEWFUNCROOT]] ], [ [[TMP407:%.*]], %[[TMP401]] ]
; MAXIMIZE-NEXT: [[TMP17:%.*]] = mul i32 [[TMP1]], [[TMP16]]
; MAXIMIZE-NEXT: [[TMP18:%.*]] = getelementptr i8, ptr [[TMP2]], i32 [[TMP17]]
; MAXIMIZE-NEXT: [[TMP19:%.*]] = getelementptr i8, ptr [[TMP3]], i32 [[TMP17]]
; MAXIMIZE-NEXT: br i1 [[TMP4]], label %[[ITER_CHECK:.*]], label %[[TMP401]]
; MAXIMIZE: [[ITER_CHECK]]:
; MAXIMIZE-NEXT: [[MIN_ITERS_CHECK4:%.*]] = icmp ult i32 [[TMP7]], 16
; MAXIMIZE-NEXT: br i1 [[MIN_ITERS_CHECK4]], label %[[VEC_EPILOG_PH:.*]], label %[[VECTOR_PH:.*]]
; MAXIMIZE: [[VECTOR_PH]]:
; MAXIMIZE-NEXT: [[TMP20:%.*]] = and i32 [[TMP7]], 15
; MAXIMIZE-NEXT: [[N_VEC:%.*]] = sub i32 [[TMP7]], [[TMP20]]
; MAXIMIZE-NEXT: [[TMP21:%.*]] = shl i32 [[N_VEC]], 2
; MAXIMIZE-NEXT: [[TMP22:%.*]] = getelementptr i8, ptr [[TMP15]], i32 [[TMP21]]
; MAXIMIZE-NEXT: [[TMP23:%.*]] = insertelement <4 x i32> zeroinitializer, i32 [[TMP11]], i64 0
; MAXIMIZE-NEXT: [[TMP24:%.*]] = insertelement <4 x i32> zeroinitializer, i32 [[TMP12]], i64 0
; MAXIMIZE-NEXT: [[TMP25:%.*]] = insertelement <4 x i32> zeroinitializer, i32 [[TMP13]], i64 0
; MAXIMIZE-NEXT: [[TMP26:%.*]] = insertelement <4 x i32> zeroinitializer, i32 [[TMP14]], i64 0
; MAXIMIZE-NEXT: br label %[[VECTOR_BODY:.*]]
; MAXIMIZE: [[VECTOR_BODY]]:
; MAXIMIZE-NEXT: [[INDEX:%.*]] = phi i32 [ 0, %[[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], %[[VECTOR_BODY]] ]
; MAXIMIZE-NEXT: [[VEC_PHI:%.*]] = phi <4 x i32> [ [[TMP23]], %[[VECTOR_PH]] ], [ [[PARTIAL_REDUCE12:%.*]], %[[VECTOR_BODY]] ]
; MAXIMIZE-NEXT: [[VEC_PHI5:%.*]] = phi <4 x i32> [ [[TMP24]], %[[VECTOR_PH]] ], [ [[PARTIAL_REDUCE16:%.*]], %[[VECTOR_BODY]] ]
; MAXIMIZE-NEXT: [[VEC_PHI6:%.*]] = phi <4 x i32> [ [[TMP25]], %[[VECTOR_PH]] ], [ [[PARTIAL_REDUCE13:%.*]], %[[VECTOR_BODY]] ]
; MAXIMIZE-NEXT: [[VEC_PHI7:%.*]] = phi <4 x i32> [ [[TMP26]], %[[VECTOR_PH]] ], [ [[PARTIAL_REDUCE17:%.*]], %[[VECTOR_BODY]] ]
; MAXIMIZE-NEXT: [[TMP27:%.*]] = shl i32 [[INDEX]], 2
; MAXIMIZE-NEXT: [[NEXT_GEP:%.*]] = getelementptr i8, ptr [[TMP15]], i32 [[TMP27]]
; MAXIMIZE-NEXT: [[TMP28:%.*]] = getelementptr inbounds [4 x i8], ptr [[TMP5]], i32 [[INDEX]]
; MAXIMIZE-NEXT: [[WIDE_LOAD:%.*]] = load <16 x i32>, ptr [[TMP28]], align 4
; MAXIMIZE-NEXT: [[TMP29:%.*]] = extractelement <16 x i32> [[WIDE_LOAD]], i64 0
; MAXIMIZE-NEXT: [[TMP30:%.*]] = getelementptr inbounds i8, ptr [[TMP18]], i32 [[TMP29]]
; MAXIMIZE-NEXT: [[TMP31:%.*]] = extractelement <16 x i32> [[WIDE_LOAD]], i64 1
; MAXIMIZE-NEXT: [[TMP32:%.*]] = getelementptr inbounds i8, ptr [[TMP18]], i32 [[TMP31]]
; MAXIMIZE-NEXT: [[TMP33:%.*]] = extractelement <16 x i32> [[WIDE_LOAD]], i64 2
; MAXIMIZE-NEXT: [[TMP34:%.*]] = getelementptr inbounds i8, ptr [[TMP18]], i32 [[TMP33]]
; MAXIMIZE-NEXT: [[TMP35:%.*]] = extractelement <16 x i32> [[WIDE_LOAD]], i64 3
; MAXIMIZE-NEXT: [[TMP36:%.*]] = getelementptr inbounds i8, ptr [[TMP18]], i32 [[TMP35]]
; MAXIMIZE-NEXT: [[TMP37:%.*]] = extractelement <16 x i32> [[WIDE_LOAD]], i64 4
; MAXIMIZE-NEXT: [[TMP38:%.*]] = getelementptr inbounds i8, ptr [[TMP18]], i32 [[TMP37]]
; MAXIMIZE-NEXT: [[TMP39:%.*]] = extractelement <16 x i32> [[WIDE_LOAD]], i64 5
; MAXIMIZE-NEXT: [[TMP40:%.*]] = getelementptr inbounds i8, ptr [[TMP18]], i32 [[TMP39]]
; MAXIMIZE-NEXT: [[TMP41:%.*]] = extractelement <16 x i32> [[WIDE_LOAD]], i64 6
; MAXIMIZE-NEXT: [[TMP42:%.*]] = getelementptr inbounds i8, ptr [[TMP18]], i32 [[TMP41]]
; MAXIMIZE-NEXT: [[TMP43:%.*]] = extractelement <16 x i32> [[WIDE_LOAD]], i64 7
; MAXIMIZE-NEXT: [[TMP44:%.*]] = getelementptr inbounds i8, ptr [[TMP18]], i32 [[TMP43]]
; MAXIMIZE-NEXT: [[TMP45:%.*]] = extractelement <16 x i32> [[WIDE_LOAD]], i64 8
; MAXIMIZE-NEXT: [[TMP46:%.*]] = getelementptr inbounds i8, ptr [[TMP18]], i32 [[TMP45]]
; MAXIMIZE-NEXT: [[TMP47:%.*]] = extractelement <16 x i32> [[WIDE_LOAD]], i64 9
; MAXIMIZE-NEXT: [[TMP48:%.*]] = getelementptr inbounds i8, ptr [[TMP18]], i32 [[TMP47]]
; MAXIMIZE-NEXT: [[TMP49:%.*]] = extractelement <16 x i32> [[WIDE_LOAD]], i64 10
; MAXIMIZE-NEXT: [[TMP50:%.*]] = getelementptr inbounds i8, ptr [[TMP18]], i32 [[TMP49]]
; MAXIMIZE-NEXT: [[TMP51:%.*]] = extractelement <16 x i32> [[WIDE_LOAD]], i64 11
; MAXIMIZE-NEXT: [[TMP52:%.*]] = getelementptr inbounds i8, ptr [[TMP18]], i32 [[TMP51]]
; MAXIMIZE-NEXT: [[TMP53:%.*]] = extractelement <16 x i32> [[WIDE_LOAD]], i64 12
; MAXIMIZE-NEXT: [[TMP54:%.*]] = getelementptr inbounds i8, ptr [[TMP18]], i32 [[TMP53]]
; MAXIMIZE-NEXT: [[TMP55:%.*]] = extractelement <16 x i32> [[WIDE_LOAD]], i64 13
; MAXIMIZE-NEXT: [[TMP56:%.*]] = getelementptr inbounds i8, ptr [[TMP18]], i32 [[TMP55]]
; MAXIMIZE-NEXT: [[TMP57:%.*]] = extractelement <16 x i32> [[WIDE_LOAD]], i64 14
; MAXIMIZE-NEXT: [[TMP58:%.*]] = getelementptr inbounds i8, ptr [[TMP18]], i32 [[TMP57]]
; MAXIMIZE-NEXT: [[TMP59:%.*]] = extractelement <16 x i32> [[WIDE_LOAD]], i64 15
; MAXIMIZE-NEXT: [[TMP60:%.*]] = getelementptr inbounds i8, ptr [[TMP18]], i32 [[TMP59]]
; MAXIMIZE-NEXT: [[TMP61:%.*]] = getelementptr inbounds i8, ptr [[TMP19]], i32 [[TMP29]]
; MAXIMIZE-NEXT: [[TMP62:%.*]] = getelementptr inbounds i8, ptr [[TMP19]], i32 [[TMP31]]
; MAXIMIZE-NEXT: [[TMP63:%.*]] = getelementptr inbounds i8, ptr [[TMP19]], i32 [[TMP33]]
; MAXIMIZE-NEXT: [[TMP64:%.*]] = getelementptr inbounds i8, ptr [[TMP19]], i32 [[TMP35]]
; MAXIMIZE-NEXT: [[TMP65:%.*]] = getelementptr inbounds i8, ptr [[TMP19]], i32 [[TMP37]]
; MAXIMIZE-NEXT: [[TMP66:%.*]] = getelementptr inbounds i8, ptr [[TMP19]], i32 [[TMP39]]
; MAXIMIZE-NEXT: [[TMP67:%.*]] = getelementptr inbounds i8, ptr [[TMP19]], i32 [[TMP41]]
; MAXIMIZE-NEXT: [[TMP68:%.*]] = getelementptr inbounds i8, ptr [[TMP19]], i32 [[TMP43]]
; MAXIMIZE-NEXT: [[TMP69:%.*]] = getelementptr inbounds i8, ptr [[TMP19]], i32 [[TMP45]]
; MAXIMIZE-NEXT: [[TMP70:%.*]] = getelementptr inbounds i8, ptr [[TMP19]], i32 [[TMP47]]
; MAXIMIZE-NEXT: [[TMP71:%.*]] = getelementptr inbounds i8, ptr [[TMP19]], i32 [[TMP49]]
; MAXIMIZE-NEXT: [[TMP72:%.*]] = getelementptr inbounds i8, ptr [[TMP19]], i32 [[TMP51]]
; MAXIMIZE-NEXT: [[TMP73:%.*]] = getelementptr inbounds i8, ptr [[TMP19]], i32 [[TMP53]]
; MAXIMIZE-NEXT: [[TMP74:%.*]] = getelementptr inbounds i8, ptr [[TMP19]], i32 [[TMP55]]
; MAXIMIZE-NEXT: [[TMP75:%.*]] = getelementptr inbounds i8, ptr [[TMP19]], i32 [[TMP57]]
; MAXIMIZE-NEXT: [[TMP76:%.*]] = getelementptr inbounds i8, ptr [[TMP19]], i32 [[TMP59]]
; MAXIMIZE-NEXT: [[TMP77:%.*]] = load i8, ptr [[TMP30]], align 1
; MAXIMIZE-NEXT: [[TMP78:%.*]] = load i8, ptr [[TMP32]], align 1
; MAXIMIZE-NEXT: [[TMP79:%.*]] = load i8, ptr [[TMP34]], align 1
; MAXIMIZE-NEXT: [[TMP80:%.*]] = load i8, ptr [[TMP36]], align 1
; MAXIMIZE-NEXT: [[TMP81:%.*]] = load i8, ptr [[TMP38]], align 1
; MAXIMIZE-NEXT: [[TMP82:%.*]] = load i8, ptr [[TMP40]], align 1
; MAXIMIZE-NEXT: [[TMP83:%.*]] = load i8, ptr [[TMP42]], align 1
; MAXIMIZE-NEXT: [[TMP84:%.*]] = load i8, ptr [[TMP44]], align 1
; MAXIMIZE-NEXT: [[TMP85:%.*]] = load i8, ptr [[TMP46]], align 1
; MAXIMIZE-NEXT: [[TMP86:%.*]] = load i8, ptr [[TMP48]], align 1
; MAXIMIZE-NEXT: [[TMP87:%.*]] = load i8, ptr [[TMP50]], align 1
; MAXIMIZE-NEXT: [[TMP88:%.*]] = load i8, ptr [[TMP52]], align 1
; MAXIMIZE-NEXT: [[TMP89:%.*]] = load i8, ptr [[TMP54]], align 1
; MAXIMIZE-NEXT: [[TMP90:%.*]] = load i8, ptr [[TMP56]], align 1
; MAXIMIZE-NEXT: [[TMP91:%.*]] = load i8, ptr [[TMP58]], align 1
; MAXIMIZE-NEXT: [[TMP92:%.*]] = load i8, ptr [[TMP60]], align 1
; MAXIMIZE-NEXT: [[TMP93:%.*]] = insertelement <16 x i8> poison, i8 [[TMP77]], i64 0
; MAXIMIZE-NEXT: [[TMP94:%.*]] = insertelement <16 x i8> [[TMP93]], i8 [[TMP78]], i64 1
; MAXIMIZE-NEXT: [[TMP95:%.*]] = insertelement <16 x i8> [[TMP94]], i8 [[TMP79]], i64 2
; MAXIMIZE-NEXT: [[TMP96:%.*]] = insertelement <16 x i8> [[TMP95]], i8 [[TMP80]], i64 3
; MAXIMIZE-NEXT: [[TMP97:%.*]] = insertelement <16 x i8> [[TMP96]], i8 [[TMP81]], i64 4
; MAXIMIZE-NEXT: [[TMP98:%.*]] = insertelement <16 x i8> [[TMP97]], i8 [[TMP82]], i64 5
; MAXIMIZE-NEXT: [[TMP99:%.*]] = insertelement <16 x i8> [[TMP98]], i8 [[TMP83]], i64 6
; MAXIMIZE-NEXT: [[TMP100:%.*]] = insertelement <16 x i8> [[TMP99]], i8 [[TMP84]], i64 7
; MAXIMIZE-NEXT: [[TMP101:%.*]] = insertelement <16 x i8> [[TMP100]], i8 [[TMP85]], i64 8
; MAXIMIZE-NEXT: [[TMP102:%.*]] = insertelement <16 x i8> [[TMP101]], i8 [[TMP86]], i64 9
; MAXIMIZE-NEXT: [[TMP103:%.*]] = insertelement <16 x i8> [[TMP102]], i8 [[TMP87]], i64 10
; MAXIMIZE-NEXT: [[TMP104:%.*]] = insertelement <16 x i8> [[TMP103]], i8 [[TMP88]], i64 11
; MAXIMIZE-NEXT: [[TMP105:%.*]] = insertelement <16 x i8> [[TMP104]], i8 [[TMP89]], i64 12
; MAXIMIZE-NEXT: [[TMP106:%.*]] = insertelement <16 x i8> [[TMP105]], i8 [[TMP90]], i64 13
; MAXIMIZE-NEXT: [[TMP107:%.*]] = insertelement <16 x i8> [[TMP106]], i8 [[TMP91]], i64 14
; MAXIMIZE-NEXT: [[TMP108:%.*]] = insertelement <16 x i8> [[TMP107]], i8 [[TMP92]], i64 15
; MAXIMIZE-NEXT: [[WIDE_VEC:%.*]] = load <64 x i8>, ptr [[NEXT_GEP]], align 1
; MAXIMIZE-NEXT: [[STRIDED_VEC:%.*]] = shufflevector <64 x i8> [[WIDE_VEC]], <64 x i8> poison, <16 x i32> <i32 0, i32 4, i32 8, i32 12, i32 16, i32 20, i32 24, i32 28, i32 32, i32 36, i32 40, i32 44, i32 48, i32 52, i32 56, i32 60>
; MAXIMIZE-NEXT: [[STRIDED_VEC8:%.*]] = shufflevector <64 x i8> [[WIDE_VEC]], <64 x i8> poison, <16 x i32> <i32 1, i32 5, i32 9, i32 13, i32 17, i32 21, i32 25, i32 29, i32 33, i32 37, i32 41, i32 45, i32 49, i32 53, i32 57, i32 61>
; MAXIMIZE-NEXT: [[STRIDED_VEC9:%.*]] = shufflevector <64 x i8> [[WIDE_VEC]], <64 x i8> poison, <16 x i32> <i32 2, i32 6, i32 10, i32 14, i32 18, i32 22, i32 26, i32 30, i32 34, i32 38, i32 42, i32 46, i32 50, i32 54, i32 58, i32 62>
; MAXIMIZE-NEXT: [[STRIDED_VEC10:%.*]] = shufflevector <64 x i8> [[WIDE_VEC]], <64 x i8> poison, <16 x i32> <i32 3, i32 7, i32 11, i32 15, i32 19, i32 23, i32 27, i32 31, i32 35, i32 39, i32 43, i32 47, i32 51, i32 55, i32 59, i32 63>
; MAXIMIZE-NEXT: [[TMP109:%.*]] = sext <16 x i8> [[STRIDED_VEC]] to <16 x i32>
; MAXIMIZE-NEXT: [[TMP110:%.*]] = sext <16 x i8> [[TMP108]] to <16 x i32>
; MAXIMIZE-NEXT: [[TMP111:%.*]] = mul nsw <16 x i32> [[TMP109]], [[TMP110]]
; MAXIMIZE-NEXT: [[PARTIAL_REDUCE:%.*]] = call <4 x i32> @llvm.vector.partial.reduce.add.v4i32.v16i32(<4 x i32> [[VEC_PHI]], <16 x i32> [[TMP111]])
; MAXIMIZE-NEXT: [[TMP112:%.*]] = sext <16 x i8> [[STRIDED_VEC8]] to <16 x i32>
; MAXIMIZE-NEXT: [[TMP113:%.*]] = mul nsw <16 x i32> [[TMP112]], [[TMP110]]
; MAXIMIZE-NEXT: [[PARTIAL_REDUCE11:%.*]] = call <4 x i32> @llvm.vector.partial.reduce.add.v4i32.v16i32(<4 x i32> [[VEC_PHI6]], <16 x i32> [[TMP113]])
; MAXIMIZE-NEXT: [[TMP114:%.*]] = getelementptr inbounds i8, ptr [[TMP30]], i32 [[TMP6]]
; MAXIMIZE-NEXT: [[TMP115:%.*]] = getelementptr inbounds i8, ptr [[TMP32]], i32 [[TMP6]]
; MAXIMIZE-NEXT: [[TMP116:%.*]] = getelementptr inbounds i8, ptr [[TMP34]], i32 [[TMP6]]
; MAXIMIZE-NEXT: [[TMP117:%.*]] = getelementptr inbounds i8, ptr [[TMP36]], i32 [[TMP6]]
; MAXIMIZE-NEXT: [[TMP118:%.*]] = getelementptr inbounds i8, ptr [[TMP38]], i32 [[TMP6]]
; MAXIMIZE-NEXT: [[TMP119:%.*]] = getelementptr inbounds i8, ptr [[TMP40]], i32 [[TMP6]]
; MAXIMIZE-NEXT: [[TMP120:%.*]] = getelementptr inbounds i8, ptr [[TMP42]], i32 [[TMP6]]
; MAXIMIZE-NEXT: [[TMP121:%.*]] = getelementptr inbounds i8, ptr [[TMP44]], i32 [[TMP6]]
; MAXIMIZE-NEXT: [[TMP122:%.*]] = getelementptr inbounds i8, ptr [[TMP46]], i32 [[TMP6]]
; MAXIMIZE-NEXT: [[TMP123:%.*]] = getelementptr inbounds i8, ptr [[TMP48]], i32 [[TMP6]]
; MAXIMIZE-NEXT: [[TMP124:%.*]] = getelementptr inbounds i8, ptr [[TMP50]], i32 [[TMP6]]
; MAXIMIZE-NEXT: [[TMP125:%.*]] = getelementptr inbounds i8, ptr [[TMP52]], i32 [[TMP6]]
; MAXIMIZE-NEXT: [[TMP126:%.*]] = getelementptr inbounds i8, ptr [[TMP54]], i32 [[TMP6]]
; MAXIMIZE-NEXT: [[TMP127:%.*]] = getelementptr inbounds i8, ptr [[TMP56]], i32 [[TMP6]]
; MAXIMIZE-NEXT: [[TMP128:%.*]] = getelementptr inbounds i8, ptr [[TMP58]], i32 [[TMP6]]
; MAXIMIZE-NEXT: [[TMP129:%.*]] = getelementptr inbounds i8, ptr [[TMP60]], i32 [[TMP6]]
; MAXIMIZE-NEXT: [[TMP130:%.*]] = load i8, ptr [[TMP114]], align 1
; MAXIMIZE-NEXT: [[TMP131:%.*]] = load i8, ptr [[TMP115]], align 1
; MAXIMIZE-NEXT: [[TMP132:%.*]] = load i8, ptr [[TMP116]], align 1
; MAXIMIZE-NEXT: [[TMP133:%.*]] = load i8, ptr [[TMP117]], align 1
; MAXIMIZE-NEXT: [[TMP134:%.*]] = load i8, ptr [[TMP118]], align 1
; MAXIMIZE-NEXT: [[TMP135:%.*]] = load i8, ptr [[TMP119]], align 1
; MAXIMIZE-NEXT: [[TMP136:%.*]] = load i8, ptr [[TMP120]], align 1
; MAXIMIZE-NEXT: [[TMP137:%.*]] = load i8, ptr [[TMP121]], align 1
; MAXIMIZE-NEXT: [[TMP138:%.*]] = load i8, ptr [[TMP122]], align 1
; MAXIMIZE-NEXT: [[TMP139:%.*]] = load i8, ptr [[TMP123]], align 1
; MAXIMIZE-NEXT: [[TMP140:%.*]] = load i8, ptr [[TMP124]], align 1
; MAXIMIZE-NEXT: [[TMP141:%.*]] = load i8, ptr [[TMP125]], align 1
; MAXIMIZE-NEXT: [[TMP142:%.*]] = load i8, ptr [[TMP126]], align 1
; MAXIMIZE-NEXT: [[TMP143:%.*]] = load i8, ptr [[TMP127]], align 1
; MAXIMIZE-NEXT: [[TMP144:%.*]] = load i8, ptr [[TMP128]], align 1
; MAXIMIZE-NEXT: [[TMP145:%.*]] = load i8, ptr [[TMP129]], align 1
; MAXIMIZE-NEXT: [[TMP146:%.*]] = insertelement <16 x i8> poison, i8 [[TMP130]], i64 0
; MAXIMIZE-NEXT: [[TMP147:%.*]] = insertelement <16 x i8> [[TMP146]], i8 [[TMP131]], i64 1
; MAXIMIZE-NEXT: [[TMP148:%.*]] = insertelement <16 x i8> [[TMP147]], i8 [[TMP132]], i64 2
; MAXIMIZE-NEXT: [[TMP149:%.*]] = insertelement <16 x i8> [[TMP148]], i8 [[TMP133]], i64 3
; MAXIMIZE-NEXT: [[TMP150:%.*]] = insertelement <16 x i8> [[TMP149]], i8 [[TMP134]], i64 4
; MAXIMIZE-NEXT: [[TMP151:%.*]] = insertelement <16 x i8> [[TMP150]], i8 [[TMP135]], i64 5
; MAXIMIZE-NEXT: [[TMP152:%.*]] = insertelement <16 x i8> [[TMP151]], i8 [[TMP136]], i64 6
; MAXIMIZE-NEXT: [[TMP153:%.*]] = insertelement <16 x i8> [[TMP152]], i8 [[TMP137]], i64 7
; MAXIMIZE-NEXT: [[TMP154:%.*]] = insertelement <16 x i8> [[TMP153]], i8 [[TMP138]], i64 8
; MAXIMIZE-NEXT: [[TMP155:%.*]] = insertelement <16 x i8> [[TMP154]], i8 [[TMP139]], i64 9
; MAXIMIZE-NEXT: [[TMP156:%.*]] = insertelement <16 x i8> [[TMP155]], i8 [[TMP140]], i64 10
; MAXIMIZE-NEXT: [[TMP157:%.*]] = insertelement <16 x i8> [[TMP156]], i8 [[TMP141]], i64 11
; MAXIMIZE-NEXT: [[TMP158:%.*]] = insertelement <16 x i8> [[TMP157]], i8 [[TMP142]], i64 12
; MAXIMIZE-NEXT: [[TMP159:%.*]] = insertelement <16 x i8> [[TMP158]], i8 [[TMP143]], i64 13
; MAXIMIZE-NEXT: [[TMP160:%.*]] = insertelement <16 x i8> [[TMP159]], i8 [[TMP144]], i64 14
; MAXIMIZE-NEXT: [[TMP161:%.*]] = insertelement <16 x i8> [[TMP160]], i8 [[TMP145]], i64 15
; MAXIMIZE-NEXT: [[TMP162:%.*]] = sext <16 x i8> [[STRIDED_VEC9]] to <16 x i32>
; MAXIMIZE-NEXT: [[TMP163:%.*]] = sext <16 x i8> [[TMP161]] to <16 x i32>
; MAXIMIZE-NEXT: [[TMP164:%.*]] = mul nsw <16 x i32> [[TMP162]], [[TMP163]]
; MAXIMIZE-NEXT: [[PARTIAL_REDUCE12]] = call <4 x i32> @llvm.vector.partial.reduce.add.v4i32.v16i32(<4 x i32> [[PARTIAL_REDUCE]], <16 x i32> [[TMP164]])
; MAXIMIZE-NEXT: [[TMP165:%.*]] = sext <16 x i8> [[STRIDED_VEC10]] to <16 x i32>
; MAXIMIZE-NEXT: [[TMP166:%.*]] = mul nsw <16 x i32> [[TMP165]], [[TMP163]]
; MAXIMIZE-NEXT: [[PARTIAL_REDUCE13]] = call <4 x i32> @llvm.vector.partial.reduce.add.v4i32.v16i32(<4 x i32> [[PARTIAL_REDUCE11]], <16 x i32> [[TMP166]])
; MAXIMIZE-NEXT: [[TMP167:%.*]] = load i8, ptr [[TMP61]], align 1
; MAXIMIZE-NEXT: [[TMP168:%.*]] = load i8, ptr [[TMP62]], align 1
; MAXIMIZE-NEXT: [[TMP169:%.*]] = load i8, ptr [[TMP63]], align 1
; MAXIMIZE-NEXT: [[TMP170:%.*]] = load i8, ptr [[TMP64]], align 1
; MAXIMIZE-NEXT: [[TMP171:%.*]] = load i8, ptr [[TMP65]], align 1
; MAXIMIZE-NEXT: [[TMP172:%.*]] = load i8, ptr [[TMP66]], align 1
; MAXIMIZE-NEXT: [[TMP173:%.*]] = load i8, ptr [[TMP67]], align 1
; MAXIMIZE-NEXT: [[TMP174:%.*]] = load i8, ptr [[TMP68]], align 1
; MAXIMIZE-NEXT: [[TMP175:%.*]] = load i8, ptr [[TMP69]], align 1
; MAXIMIZE-NEXT: [[TMP176:%.*]] = load i8, ptr [[TMP70]], align 1
; MAXIMIZE-NEXT: [[TMP177:%.*]] = load i8, ptr [[TMP71]], align 1
; MAXIMIZE-NEXT: [[TMP178:%.*]] = load i8, ptr [[TMP72]], align 1
; MAXIMIZE-NEXT: [[TMP179:%.*]] = load i8, ptr [[TMP73]], align 1
; MAXIMIZE-NEXT: [[TMP180:%.*]] = load i8, ptr [[TMP74]], align 1
; MAXIMIZE-NEXT: [[TMP181:%.*]] = load i8, ptr [[TMP75]], align 1
; MAXIMIZE-NEXT: [[TMP182:%.*]] = load i8, ptr [[TMP76]], align 1
; MAXIMIZE-NEXT: [[TMP183:%.*]] = insertelement <16 x i8> poison, i8 [[TMP167]], i64 0
; MAXIMIZE-NEXT: [[TMP184:%.*]] = insertelement <16 x i8> [[TMP183]], i8 [[TMP168]], i64 1
; MAXIMIZE-NEXT: [[TMP185:%.*]] = insertelement <16 x i8> [[TMP184]], i8 [[TMP169]], i64 2
; MAXIMIZE-NEXT: [[TMP186:%.*]] = insertelement <16 x i8> [[TMP185]], i8 [[TMP170]], i64 3
; MAXIMIZE-NEXT: [[TMP187:%.*]] = insertelement <16 x i8> [[TMP186]], i8 [[TMP171]], i64 4
; MAXIMIZE-NEXT: [[TMP188:%.*]] = insertelement <16 x i8> [[TMP187]], i8 [[TMP172]], i64 5
; MAXIMIZE-NEXT: [[TMP189:%.*]] = insertelement <16 x i8> [[TMP188]], i8 [[TMP173]], i64 6
; MAXIMIZE-NEXT: [[TMP190:%.*]] = insertelement <16 x i8> [[TMP189]], i8 [[TMP174]], i64 7
; MAXIMIZE-NEXT: [[TMP191:%.*]] = insertelement <16 x i8> [[TMP190]], i8 [[TMP175]], i64 8
; MAXIMIZE-NEXT: [[TMP192:%.*]] = insertelement <16 x i8> [[TMP191]], i8 [[TMP176]], i64 9
; MAXIMIZE-NEXT: [[TMP193:%.*]] = insertelement <16 x i8> [[TMP192]], i8 [[TMP177]], i64 10
; MAXIMIZE-NEXT: [[TMP194:%.*]] = insertelement <16 x i8> [[TMP193]], i8 [[TMP178]], i64 11
; MAXIMIZE-NEXT: [[TMP195:%.*]] = insertelement <16 x i8> [[TMP194]], i8 [[TMP179]], i64 12
; MAXIMIZE-NEXT: [[TMP196:%.*]] = insertelement <16 x i8> [[TMP195]], i8 [[TMP180]], i64 13
; MAXIMIZE-NEXT: [[TMP197:%.*]] = insertelement <16 x i8> [[TMP196]], i8 [[TMP181]], i64 14
; MAXIMIZE-NEXT: [[TMP198:%.*]] = insertelement <16 x i8> [[TMP197]], i8 [[TMP182]], i64 15
; MAXIMIZE-NEXT: [[TMP199:%.*]] = sext <16 x i8> [[TMP198]] to <16 x i32>
; MAXIMIZE-NEXT: [[TMP200:%.*]] = mul nsw <16 x i32> [[TMP199]], [[TMP109]]
; MAXIMIZE-NEXT: [[PARTIAL_REDUCE14:%.*]] = call <4 x i32> @llvm.vector.partial.reduce.add.v4i32.v16i32(<4 x i32> [[VEC_PHI5]], <16 x i32> [[TMP200]])
; MAXIMIZE-NEXT: [[TMP201:%.*]] = mul nsw <16 x i32> [[TMP199]], [[TMP112]]
; MAXIMIZE-NEXT: [[PARTIAL_REDUCE15:%.*]] = call <4 x i32> @llvm.vector.partial.reduce.add.v4i32.v16i32(<4 x i32> [[VEC_PHI7]], <16 x i32> [[TMP201]])
; MAXIMIZE-NEXT: [[TMP202:%.*]] = getelementptr inbounds i8, ptr [[TMP61]], i32 [[TMP6]]
; MAXIMIZE-NEXT: [[TMP203:%.*]] = getelementptr inbounds i8, ptr [[TMP62]], i32 [[TMP6]]
; MAXIMIZE-NEXT: [[TMP204:%.*]] = getelementptr inbounds i8, ptr [[TMP63]], i32 [[TMP6]]
; MAXIMIZE-NEXT: [[TMP205:%.*]] = getelementptr inbounds i8, ptr [[TMP64]], i32 [[TMP6]]
; MAXIMIZE-NEXT: [[TMP206:%.*]] = getelementptr inbounds i8, ptr [[TMP65]], i32 [[TMP6]]
; MAXIMIZE-NEXT: [[TMP207:%.*]] = getelementptr inbounds i8, ptr [[TMP66]], i32 [[TMP6]]
; MAXIMIZE-NEXT: [[TMP208:%.*]] = getelementptr inbounds i8, ptr [[TMP67]], i32 [[TMP6]]
; MAXIMIZE-NEXT: [[TMP209:%.*]] = getelementptr inbounds i8, ptr [[TMP68]], i32 [[TMP6]]
; MAXIMIZE-NEXT: [[TMP210:%.*]] = getelementptr inbounds i8, ptr [[TMP69]], i32 [[TMP6]]
; MAXIMIZE-NEXT: [[TMP211:%.*]] = getelementptr inbounds i8, ptr [[TMP70]], i32 [[TMP6]]
; MAXIMIZE-NEXT: [[TMP212:%.*]] = getelementptr inbounds i8, ptr [[TMP71]], i32 [[TMP6]]
; MAXIMIZE-NEXT: [[TMP213:%.*]] = getelementptr inbounds i8, ptr [[TMP72]], i32 [[TMP6]]
; MAXIMIZE-NEXT: [[TMP214:%.*]] = getelementptr inbounds i8, ptr [[TMP73]], i32 [[TMP6]]
; MAXIMIZE-NEXT: [[TMP215:%.*]] = getelementptr inbounds i8, ptr [[TMP74]], i32 [[TMP6]]
; MAXIMIZE-NEXT: [[TMP216:%.*]] = getelementptr inbounds i8, ptr [[TMP75]], i32 [[TMP6]]
; MAXIMIZE-NEXT: [[TMP217:%.*]] = getelementptr inbounds i8, ptr [[TMP76]], i32 [[TMP6]]
; MAXIMIZE-NEXT: [[TMP218:%.*]] = load i8, ptr [[TMP202]], align 1
; MAXIMIZE-NEXT: [[TMP219:%.*]] = load i8, ptr [[TMP203]], align 1
; MAXIMIZE-NEXT: [[TMP220:%.*]] = load i8, ptr [[TMP204]], align 1
; MAXIMIZE-NEXT: [[TMP221:%.*]] = load i8, ptr [[TMP205]], align 1
; MAXIMIZE-NEXT: [[TMP222:%.*]] = load i8, ptr [[TMP206]], align 1
; MAXIMIZE-NEXT: [[TMP223:%.*]] = load i8, ptr [[TMP207]], align 1
; MAXIMIZE-NEXT: [[TMP224:%.*]] = load i8, ptr [[TMP208]], align 1
; MAXIMIZE-NEXT: [[TMP225:%.*]] = load i8, ptr [[TMP209]], align 1
; MAXIMIZE-NEXT: [[TMP226:%.*]] = load i8, ptr [[TMP210]], align 1
; MAXIMIZE-NEXT: [[TMP227:%.*]] = load i8, ptr [[TMP211]], align 1
; MAXIMIZE-NEXT: [[TMP228:%.*]] = load i8, ptr [[TMP212]], align 1
; MAXIMIZE-NEXT: [[TMP229:%.*]] = load i8, ptr [[TMP213]], align 1
; MAXIMIZE-NEXT: [[TMP230:%.*]] = load i8, ptr [[TMP214]], align 1
; MAXIMIZE-NEXT: [[TMP231:%.*]] = load i8, ptr [[TMP215]], align 1
; MAXIMIZE-NEXT: [[TMP232:%.*]] = load i8, ptr [[TMP216]], align 1
; MAXIMIZE-NEXT: [[TMP233:%.*]] = load i8, ptr [[TMP217]], align 1
; MAXIMIZE-NEXT: [[TMP234:%.*]] = insertelement <16 x i8> poison, i8 [[TMP218]], i64 0
; MAXIMIZE-NEXT: [[TMP235:%.*]] = insertelement <16 x i8> [[TMP234]], i8 [[TMP219]], i64 1
; MAXIMIZE-NEXT: [[TMP236:%.*]] = insertelement <16 x i8> [[TMP235]], i8 [[TMP220]], i64 2
; MAXIMIZE-NEXT: [[TMP237:%.*]] = insertelement <16 x i8> [[TMP236]], i8 [[TMP221]], i64 3
; MAXIMIZE-NEXT: [[TMP238:%.*]] = insertelement <16 x i8> [[TMP237]], i8 [[TMP222]], i64 4
; MAXIMIZE-NEXT: [[TMP239:%.*]] = insertelement <16 x i8> [[TMP238]], i8 [[TMP223]], i64 5
; MAXIMIZE-NEXT: [[TMP240:%.*]] = insertelement <16 x i8> [[TMP239]], i8 [[TMP224]], i64 6
; MAXIMIZE-NEXT: [[TMP241:%.*]] = insertelement <16 x i8> [[TMP240]], i8 [[TMP225]], i64 7
; MAXIMIZE-NEXT: [[TMP242:%.*]] = insertelement <16 x i8> [[TMP241]], i8 [[TMP226]], i64 8
; MAXIMIZE-NEXT: [[TMP243:%.*]] = insertelement <16 x i8> [[TMP242]], i8 [[TMP227]], i64 9
; MAXIMIZE-NEXT: [[TMP244:%.*]] = insertelement <16 x i8> [[TMP243]], i8 [[TMP228]], i64 10
; MAXIMIZE-NEXT: [[TMP245:%.*]] = insertelement <16 x i8> [[TMP244]], i8 [[TMP229]], i64 11
; MAXIMIZE-NEXT: [[TMP246:%.*]] = insertelement <16 x i8> [[TMP245]], i8 [[TMP230]], i64 12
; MAXIMIZE-NEXT: [[TMP247:%.*]] = insertelement <16 x i8> [[TMP246]], i8 [[TMP231]], i64 13
; MAXIMIZE-NEXT: [[TMP248:%.*]] = insertelement <16 x i8> [[TMP247]], i8 [[TMP232]], i64 14
; MAXIMIZE-NEXT: [[TMP249:%.*]] = insertelement <16 x i8> [[TMP248]], i8 [[TMP233]], i64 15
; MAXIMIZE-NEXT: [[TMP250:%.*]] = sext <16 x i8> [[TMP249]] to <16 x i32>
; MAXIMIZE-NEXT: [[TMP251:%.*]] = mul nsw <16 x i32> [[TMP250]], [[TMP162]]
; MAXIMIZE-NEXT: [[PARTIAL_REDUCE16]] = call <4 x i32> @llvm.vector.partial.reduce.add.v4i32.v16i32(<4 x i32> [[PARTIAL_REDUCE14]], <16 x i32> [[TMP251]])
; MAXIMIZE-NEXT: [[TMP252:%.*]] = mul nsw <16 x i32> [[TMP250]], [[TMP165]]
; MAXIMIZE-NEXT: [[PARTIAL_REDUCE17]] = call <4 x i32> @llvm.vector.partial.reduce.add.v4i32.v16i32(<4 x i32> [[PARTIAL_REDUCE15]], <16 x i32> [[TMP252]])
; MAXIMIZE-NEXT: [[INDEX_NEXT]] = add nuw i32 [[INDEX]], 16
; MAXIMIZE-NEXT: [[TMP253:%.*]] = icmp eq i32 [[INDEX_NEXT]], [[N_VEC]]
; MAXIMIZE-NEXT: br i1 [[TMP253]], label %[[VEC_EPILOG_MIDDLE_BLOCK:.*]], label %[[VECTOR_BODY]], !llvm.loop [[LOOP0:![0-9]+]]
; MAXIMIZE: [[VEC_EPILOG_MIDDLE_BLOCK]]:
; MAXIMIZE-NEXT: [[TMP254:%.*]] = call i32 @llvm.vector.reduce.add.v4i32(<4 x i32> [[PARTIAL_REDUCE12]])
; MAXIMIZE-NEXT: [[TMP255:%.*]] = call i32 @llvm.vector.reduce.add.v4i32(<4 x i32> [[PARTIAL_REDUCE16]])
; MAXIMIZE-NEXT: [[TMP256:%.*]] = call i32 @llvm.vector.reduce.add.v4i32(<4 x i32> [[PARTIAL_REDUCE13]])
; MAXIMIZE-NEXT: [[TMP257:%.*]] = call i32 @llvm.vector.reduce.add.v4i32(<4 x i32> [[PARTIAL_REDUCE17]])
; MAXIMIZE-NEXT: [[CMP_N:%.*]] = icmp eq i32 [[TMP7]], [[N_VEC]]
; MAXIMIZE-NEXT: br i1 [[CMP_N]], label %[[BB399:.*]], label %[[VEC_EPILOG_PH]]
; MAXIMIZE: [[VEC_EPILOG_PH]]:
; MAXIMIZE-NEXT: [[BC_MERGE_RDX:%.*]] = phi i32 [ [[TMP254]], %[[VEC_EPILOG_MIDDLE_BLOCK]] ], [ [[TMP11]], %[[ITER_CHECK]] ]
; MAXIMIZE-NEXT: [[BC_MERGE_RDX17:%.*]] = phi i32 [ [[TMP255]], %[[VEC_EPILOG_MIDDLE_BLOCK]] ], [ [[TMP12]], %[[ITER_CHECK]] ]
; MAXIMIZE-NEXT: [[BC_MERGE_RDX18:%.*]] = phi i32 [ [[TMP256]], %[[VEC_EPILOG_MIDDLE_BLOCK]] ], [ [[TMP13]], %[[ITER_CHECK]] ]
; MAXIMIZE-NEXT: [[BC_MERGE_RDX19:%.*]] = phi i32 [ [[TMP257]], %[[VEC_EPILOG_MIDDLE_BLOCK]] ], [ [[TMP14]], %[[ITER_CHECK]] ]
; MAXIMIZE-NEXT: [[BC_RESUME_VAL:%.*]] = phi ptr [ [[TMP22]], %[[VEC_EPILOG_MIDDLE_BLOCK]] ], [ [[TMP15]], %[[ITER_CHECK]] ]
; MAXIMIZE-NEXT: [[BC_RESUME_VAL20:%.*]] = phi i32 [ [[N_VEC]], %[[VEC_EPILOG_MIDDLE_BLOCK]] ], [ 0, %[[ITER_CHECK]] ]
; MAXIMIZE-NEXT: br label %[[BB348:.*]]
; MAXIMIZE: [[BB348]]:
; MAXIMIZE-NEXT: [[TMP349:%.*]] = phi i32 [ [[TMP377:%.*]], %[[BB348]] ], [ [[BC_MERGE_RDX]], %[[VEC_EPILOG_PH]] ]
; MAXIMIZE-NEXT: [[TMP350:%.*]] = phi i32 [ [[TMP393:%.*]], %[[BB348]] ], [ [[BC_MERGE_RDX17]], %[[VEC_EPILOG_PH]] ]
; MAXIMIZE-NEXT: [[TMP351:%.*]] = phi i32 [ [[TMP382:%.*]], %[[BB348]] ], [ [[BC_MERGE_RDX18]], %[[VEC_EPILOG_PH]] ]
; MAXIMIZE-NEXT: [[TMP352:%.*]] = phi i32 [ [[TMP395:%.*]], %[[BB348]] ], [ [[BC_MERGE_RDX19]], %[[VEC_EPILOG_PH]] ]
; MAXIMIZE-NEXT: [[TMP353:%.*]] = phi ptr [ [[TMP396:%.*]], %[[BB348]] ], [ [[BC_RESUME_VAL]], %[[VEC_EPILOG_PH]] ]
; MAXIMIZE-NEXT: [[TMP354:%.*]] = phi i32 [ [[TMP397:%.*]], %[[BB348]] ], [ [[BC_RESUME_VAL20]], %[[VEC_EPILOG_PH]] ]
; MAXIMIZE-NEXT: [[TMP355:%.*]] = getelementptr inbounds [4 x i8], ptr [[TMP5]], i32 [[TMP354]]
; MAXIMIZE-NEXT: [[TMP356:%.*]] = load i32, ptr [[TMP355]], align 4
; MAXIMIZE-NEXT: [[TMP357:%.*]] = getelementptr inbounds i8, ptr [[TMP18]], i32 [[TMP356]]
; MAXIMIZE-NEXT: [[TMP358:%.*]] = getelementptr inbounds i8, ptr [[TMP19]], i32 [[TMP356]]
; MAXIMIZE-NEXT: [[TMP359:%.*]] = load i8, ptr [[TMP357]], align 1
; MAXIMIZE-NEXT: [[TMP360:%.*]] = sext i8 [[TMP359]] to i32
; MAXIMIZE-NEXT: [[TMP361:%.*]] = load i8, ptr [[TMP353]], align 1
; MAXIMIZE-NEXT: [[TMP362:%.*]] = sext i8 [[TMP361]] to i32
; MAXIMIZE-NEXT: [[TMP363:%.*]] = mul nsw i32 [[TMP362]], [[TMP360]]
; MAXIMIZE-NEXT: [[TMP364:%.*]] = add nsw i32 [[TMP363]], [[TMP349]]
; MAXIMIZE-NEXT: [[TMP365:%.*]] = getelementptr inbounds i8, ptr [[TMP353]], i32 1
; MAXIMIZE-NEXT: [[TMP366:%.*]] = load i8, ptr [[TMP365]], align 1
; MAXIMIZE-NEXT: [[TMP367:%.*]] = sext i8 [[TMP366]] to i32
; MAXIMIZE-NEXT: [[TMP368:%.*]] = mul nsw i32 [[TMP367]], [[TMP360]]
; MAXIMIZE-NEXT: [[TMP369:%.*]] = add nsw i32 [[TMP368]], [[TMP351]]
; MAXIMIZE-NEXT: [[TMP370:%.*]] = getelementptr inbounds i8, ptr [[TMP357]], i32 [[TMP6]]
; MAXIMIZE-NEXT: [[TMP371:%.*]] = load i8, ptr [[TMP370]], align 1
; MAXIMIZE-NEXT: [[TMP372:%.*]] = sext i8 [[TMP371]] to i32
; MAXIMIZE-NEXT: [[TMP373:%.*]] = getelementptr inbounds i8, ptr [[TMP353]], i32 2
; MAXIMIZE-NEXT: [[TMP374:%.*]] = load i8, ptr [[TMP373]], align 1
; MAXIMIZE-NEXT: [[TMP375:%.*]] = sext i8 [[TMP374]] to i32
; MAXIMIZE-NEXT: [[TMP376:%.*]] = mul nsw i32 [[TMP375]], [[TMP372]]
; MAXIMIZE-NEXT: [[TMP377]] = add nsw i32 [[TMP364]], [[TMP376]]
; MAXIMIZE-NEXT: [[TMP378:%.*]] = getelementptr inbounds i8, ptr [[TMP353]], i32 3
; MAXIMIZE-NEXT: [[TMP379:%.*]] = load i8, ptr [[TMP378]], align 1
; MAXIMIZE-NEXT: [[TMP380:%.*]] = sext i8 [[TMP379]] to i32
; MAXIMIZE-NEXT: [[TMP381:%.*]] = mul nsw i32 [[TMP380]], [[TMP372]]
; MAXIMIZE-NEXT: [[TMP382]] = add nsw i32 [[TMP369]], [[TMP381]]
; MAXIMIZE-NEXT: [[TMP383:%.*]] = load i8, ptr [[TMP358]], align 1
; MAXIMIZE-NEXT: [[TMP384:%.*]] = sext i8 [[TMP383]] to i32
; MAXIMIZE-NEXT: [[TMP385:%.*]] = mul nsw i32 [[TMP384]], [[TMP362]]
; MAXIMIZE-NEXT: [[TMP386:%.*]] = add nsw i32 [[TMP385]], [[TMP350]]
; MAXIMIZE-NEXT: [[TMP387:%.*]] = mul nsw i32 [[TMP384]], [[TMP367]]
; MAXIMIZE-NEXT: [[TMP388:%.*]] = add nsw i32 [[TMP387]], [[TMP352]]
; MAXIMIZE-NEXT: [[TMP389:%.*]] = getelementptr inbounds i8, ptr [[TMP358]], i32 [[TMP6]]
; MAXIMIZE-NEXT: [[TMP390:%.*]] = load i8, ptr [[TMP389]], align 1
; MAXIMIZE-NEXT: [[TMP391:%.*]] = sext i8 [[TMP390]] to i32
; MAXIMIZE-NEXT: [[TMP392:%.*]] = mul nsw i32 [[TMP391]], [[TMP375]]
; MAXIMIZE-NEXT: [[TMP393]] = add nsw i32 [[TMP386]], [[TMP392]]
; MAXIMIZE-NEXT: [[TMP394:%.*]] = mul nsw i32 [[TMP391]], [[TMP380]]
; MAXIMIZE-NEXT: [[TMP395]] = add nsw i32 [[TMP388]], [[TMP394]]
; MAXIMIZE-NEXT: [[TMP396]] = getelementptr inbounds i8, ptr [[TMP353]], i32 4
; MAXIMIZE-NEXT: [[TMP397]] = add nuw nsw i32 [[TMP354]], 1
; MAXIMIZE-NEXT: [[TMP398:%.*]] = icmp eq i32 [[TMP397]], [[TMP7]]
; MAXIMIZE-NEXT: br i1 [[TMP398]], label %[[BB399]], label %[[BB348]], !llvm.loop [[LOOP3:![0-9]+]]
; MAXIMIZE: [[BB399]]:
; MAXIMIZE-NEXT: [[DOTLCSSA3:%.*]] = phi i32 [ [[TMP377]], %[[BB348]] ], [ [[TMP254]], %[[VEC_EPILOG_MIDDLE_BLOCK]] ]
; MAXIMIZE-NEXT: [[DOTLCSSA2:%.*]] = phi i32 [ [[TMP382]], %[[BB348]] ], [ [[TMP256]], %[[VEC_EPILOG_MIDDLE_BLOCK]] ]
; MAXIMIZE-NEXT: [[DOTLCSSA1:%.*]] = phi i32 [ [[TMP393]], %[[BB348]] ], [ [[TMP255]], %[[VEC_EPILOG_MIDDLE_BLOCK]] ]
; MAXIMIZE-NEXT: [[DOTLCSSA:%.*]] = phi i32 [ [[TMP395]], %[[BB348]] ], [ [[TMP257]], %[[VEC_EPILOG_MIDDLE_BLOCK]] ]
; MAXIMIZE-NEXT: [[TMP400:%.*]] = getelementptr i8, ptr [[TMP15]], i32 [[TMP8]]
; MAXIMIZE-NEXT: br label %[[TMP401]]
; MAXIMIZE: [[TMP401]]:
; MAXIMIZE-NEXT: [[TMP402]] = phi ptr [ [[TMP15]], %[[BB10]] ], [ [[TMP400]], %[[BB399]] ]
; MAXIMIZE-NEXT: [[TMP403]] = phi i32 [ [[TMP14]], %[[BB10]] ], [ [[DOTLCSSA]], %[[BB399]] ]
; MAXIMIZE-NEXT: [[TMP404]] = phi i32 [ [[TMP13]], %[[BB10]] ], [ [[DOTLCSSA2]], %[[BB399]] ]
; MAXIMIZE-NEXT: [[TMP405]] = phi i32 [ [[TMP12]], %[[BB10]] ], [ [[DOTLCSSA1]], %[[BB399]] ]
; MAXIMIZE-NEXT: [[TMP406]] = phi i32 [ [[TMP11]], %[[BB10]] ], [ [[DOTLCSSA3]], %[[BB399]] ]
; MAXIMIZE-NEXT: store i32 [[TMP406]], ptr [[DOTOUT4]], align 4
; MAXIMIZE-NEXT: store i32 [[TMP405]], ptr [[DOTOUT3]], align 4
; MAXIMIZE-NEXT: store i32 [[TMP404]], ptr [[DOTOUT2]], align 4
; MAXIMIZE-NEXT: store i32 [[TMP403]], ptr [[DOTOUT1]], align 4
; MAXIMIZE-NEXT: store ptr [[TMP402]], ptr [[DOTOUT]], align 4
; MAXIMIZE-NEXT: [[TMP407]] = add nuw nsw i32 [[TMP16]], 2
; MAXIMIZE-NEXT: [[TMP408:%.*]] = or disjoint i32 [[TMP407]], 1
; MAXIMIZE-NEXT: [[TMP409:%.*]] = icmp slt i32 [[TMP408]], [[TMP9]]
; MAXIMIZE-NEXT: br i1 [[TMP409]], label %[[BB10]], label %[[DOTEXITSTUB:.*]]
; MAXIMIZE: [[DOTEXITSTUB]]:
; MAXIMIZE-NEXT: ret void
;
newFuncRoot:
br label %10
10: ; preds = %newFuncRoot, %73
%11 = phi i32 [ 0, %newFuncRoot ], [ %78, %73 ]
%12 = phi i32 [ 0, %newFuncRoot ], [ %77, %73 ]
%13 = phi i32 [ 0, %newFuncRoot ], [ %76, %73 ]
%14 = phi i32 [ 0, %newFuncRoot ], [ %75, %73 ]
%15 = phi ptr [ %0, %newFuncRoot ], [ %74, %73 ]
%16 = phi i32 [ 0, %newFuncRoot ], [ %79, %73 ]
%17 = mul i32 %1, %16
%18 = getelementptr i8, ptr %2, i32 %17
%19 = getelementptr i8, ptr %3, i32 %17
br i1 %4, label %20, label %73
20: ; preds = %20, %10
%21 = phi i32 [ %49, %20 ], [ %11, %10 ]
%22 = phi i32 [ %65, %20 ], [ %12, %10 ]
%23 = phi i32 [ %54, %20 ], [ %13, %10 ]
%24 = phi i32 [ %67, %20 ], [ %14, %10 ]
%25 = phi ptr [ %68, %20 ], [ %15, %10 ]
%26 = phi i32 [ %69, %20 ], [ 0, %10 ]
%27 = getelementptr inbounds [4 x i8], ptr %5, i32 %26
%28 = load i32, ptr %27, align 4
%29 = getelementptr inbounds i8, ptr %18, i32 %28
%30 = getelementptr inbounds i8, ptr %19, i32 %28
%31 = load i8, ptr %29, align 1
%32 = sext i8 %31 to i32
%33 = load i8, ptr %25, align 1
%34 = sext i8 %33 to i32
%35 = mul nsw i32 %34, %32
%36 = add nsw i32 %35, %21
%37 = getelementptr inbounds i8, ptr %25, i32 1
%38 = load i8, ptr %37, align 1
%39 = sext i8 %38 to i32
%40 = mul nsw i32 %39, %32
%41 = add nsw i32 %40, %23
%42 = getelementptr inbounds i8, ptr %29, i32 %6
%43 = load i8, ptr %42, align 1
%44 = sext i8 %43 to i32
%45 = getelementptr inbounds i8, ptr %25, i32 2
%46 = load i8, ptr %45, align 1
%47 = sext i8 %46 to i32
%48 = mul nsw i32 %47, %44
%49 = add nsw i32 %36, %48
%50 = getelementptr inbounds i8, ptr %25, i32 3
%51 = load i8, ptr %50, align 1
%52 = sext i8 %51 to i32
%53 = mul nsw i32 %52, %44
%54 = add nsw i32 %41, %53
%55 = load i8, ptr %30, align 1
%56 = sext i8 %55 to i32
%57 = mul nsw i32 %56, %34
%58 = add nsw i32 %57, %22
%59 = mul nsw i32 %56, %39
%60 = add nsw i32 %59, %24
%61 = getelementptr inbounds i8, ptr %30, i32 %6
%62 = load i8, ptr %61, align 1
%63 = sext i8 %62 to i32
%64 = mul nsw i32 %63, %47
%65 = add nsw i32 %58, %64
%66 = mul nsw i32 %63, %52
%67 = add nsw i32 %60, %66
%68 = getelementptr inbounds i8, ptr %25, i32 4
%69 = add nuw nsw i32 %26, 1
%70 = icmp eq i32 %69, %7
br i1 %70, label %71, label %20
71: ; preds = %20
%72 = getelementptr i8, ptr %15, i32 %8
br label %73
73: ; preds = %71, %10
%74 = phi ptr [ %15, %10 ], [ %72, %71 ]
%75 = phi i32 [ %14, %10 ], [ %67, %71 ]
%76 = phi i32 [ %13, %10 ], [ %54, %71 ]
%77 = phi i32 [ %12, %10 ], [ %65, %71 ]
%78 = phi i32 [ %11, %10 ], [ %49, %71 ]
store i32 %78, ptr %.out4, align 4
store i32 %77, ptr %.out3, align 4
store i32 %76, ptr %.out2, align 4
store i32 %75, ptr %.out1, align 4
store ptr %74, ptr %.out, align 4
%79 = add nuw nsw i32 %16, 2
%80 = or disjoint i32 %79, 1
%81 = icmp slt i32 %80, %9
br i1 %81, label %10, label %.exitStub
.exitStub: ; preds = %73
ret void
}
;.
; CHECK: [[LOOP0]] = distinct !{[[LOOP0]], [[META1:![0-9]+]], [[META2:![0-9]+]]}
; CHECK: [[META1]] = !{!"llvm.loop.isvectorized", i32 1}
; CHECK: [[META2]] = !{!"llvm.loop.unroll.runtime.disable"}
; CHECK: [[LOOP3]] = distinct !{[[LOOP3]], [[META2]], [[META1]]}
;.
; MAXIMIZE: [[LOOP0]] = distinct !{[[LOOP0]], [[META1:![0-9]+]], [[META2:![0-9]+]]}
; MAXIMIZE: [[META1]] = !{!"llvm.loop.isvectorized", i32 1}
; MAXIMIZE: [[META2]] = !{!"llvm.loop.unroll.runtime.disable"}
; MAXIMIZE: [[LOOP3]] = distinct !{[[LOOP3]], [[META2]], [[META1]]}
;.