blob: 0c1f159ac61292f012a0b038fbdb94754e420ae3 [file] [edit]
; NOTE: Assertions have been autogenerated by utils/update_test_checks.py UTC_ARGS: --version 6
; RUN: opt -passes=loop-vectorize -force-vector-width=4 -S %s | FileCheck %s
; RUN: opt -passes=loop-vectorize -force-vector-width=4 -force-vector-interleave=4 -S %s | FileCheck %s --check-prefix=IC4
; RUN: opt -passes=loop-vectorize -force-vector-width=4 -force-vector-interleave=8 -S %s | FileCheck %s --check-prefix=IC8
; A clamped (i % 2^N) load guarded by an if-condition is widened into a
; predicated (masked) consecutive load over a single clamp window.
; TODO: This is the desired codegen. Currently the load is fully scalarized
; into per-lane pred.load.if blocks with urem-indexed GEPs (a gather); no
; llvm.masked.load or wide consecutive load is emitted.
define i32 @clamped_load_predicated(ptr %A, i32 %n) {
; CHECK-LABEL: define i32 @clamped_load_predicated(
; CHECK-SAME: ptr [[A:%.*]], i32 [[N:%.*]]) {
; CHECK-NEXT: [[ENTRY:.*]]:
; CHECK-NEXT: [[MIN_ITERS_CHECK:%.*]] = icmp ult i32 [[N]], 4
; CHECK-NEXT: br i1 [[MIN_ITERS_CHECK]], label %[[SCALAR_PH:.*]], label %[[VECTOR_SCEVCHECK:.*]]
; CHECK: [[VECTOR_SCEVCHECK]]:
; CHECK-NEXT: [[TMP0:%.*]] = add i32 [[N]], -1
; CHECK-NEXT: [[TMP1:%.*]] = icmp ugt i32 [[TMP0]], 127
; CHECK-NEXT: br i1 [[TMP1]], label %[[SCALAR_PH]], label %[[VECTOR_PH:.*]]
; CHECK: [[VECTOR_PH]]:
; CHECK-NEXT: [[N_MOD_VF:%.*]] = urem i32 [[N]], 4
; CHECK-NEXT: [[N_VEC:%.*]] = sub i32 [[N]], [[N_MOD_VF]]
; CHECK-NEXT: br label %[[VECTOR_BODY:.*]]
; CHECK: [[VECTOR_BODY]]:
; CHECK-NEXT: [[INDEX:%.*]] = phi i32 [ 0, %[[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], %[[PRED_LOAD_CONTINUE6:.*]] ]
; CHECK-NEXT: [[VEC_IND:%.*]] = phi <4 x i32> [ <i32 0, i32 1, i32 2, i32 3>, %[[VECTOR_PH]] ], [ [[VEC_IND_NEXT:%.*]], %[[PRED_LOAD_CONTINUE6]] ]
; CHECK-NEXT: [[VEC_PHI:%.*]] = phi <4 x i32> [ zeroinitializer, %[[VECTOR_PH]] ], [ [[TMP29:%.*]], %[[PRED_LOAD_CONTINUE6]] ]
; CHECK-NEXT: [[TMP2:%.*]] = and <4 x i32> [[VEC_IND]], splat (i32 1)
; CHECK-NEXT: [[TMP3:%.*]] = icmp eq <4 x i32> [[TMP2]], zeroinitializer
; CHECK-NEXT: [[TMP4:%.*]] = urem <4 x i32> [[VEC_IND]], splat (i32 128)
; CHECK-NEXT: [[TMP5:%.*]] = extractelement <4 x i1> [[TMP3]], i64 0
; CHECK-NEXT: br i1 [[TMP5]], label %[[PRED_LOAD_IF:.*]], label %[[PRED_LOAD_CONTINUE:.*]]
; CHECK: [[PRED_LOAD_IF]]:
; CHECK-NEXT: [[TMP6:%.*]] = extractelement <4 x i32> [[TMP4]], i64 0
; CHECK-NEXT: [[TMP7:%.*]] = getelementptr inbounds i32, ptr [[A]], i32 [[TMP6]]
; CHECK-NEXT: [[TMP8:%.*]] = load i32, ptr [[TMP7]], align 4
; CHECK-NEXT: [[TMP9:%.*]] = insertelement <4 x i32> poison, i32 [[TMP8]], i64 0
; CHECK-NEXT: br label %[[PRED_LOAD_CONTINUE]]
; CHECK: [[PRED_LOAD_CONTINUE]]:
; CHECK-NEXT: [[TMP10:%.*]] = phi <4 x i32> [ poison, %[[VECTOR_BODY]] ], [ [[TMP9]], %[[PRED_LOAD_IF]] ]
; CHECK-NEXT: [[TMP11:%.*]] = extractelement <4 x i1> [[TMP3]], i64 1
; CHECK-NEXT: br i1 [[TMP11]], label %[[PRED_LOAD_IF1:.*]], label %[[PRED_LOAD_CONTINUE2:.*]]
; CHECK: [[PRED_LOAD_IF1]]:
; CHECK-NEXT: [[TMP12:%.*]] = extractelement <4 x i32> [[TMP4]], i64 1
; CHECK-NEXT: [[TMP13:%.*]] = getelementptr inbounds i32, ptr [[A]], i32 [[TMP12]]
; CHECK-NEXT: [[TMP14:%.*]] = load i32, ptr [[TMP13]], align 4
; CHECK-NEXT: [[TMP15:%.*]] = insertelement <4 x i32> [[TMP10]], i32 [[TMP14]], i64 1
; CHECK-NEXT: br label %[[PRED_LOAD_CONTINUE2]]
; CHECK: [[PRED_LOAD_CONTINUE2]]:
; CHECK-NEXT: [[TMP16:%.*]] = phi <4 x i32> [ [[TMP10]], %[[PRED_LOAD_CONTINUE]] ], [ [[TMP15]], %[[PRED_LOAD_IF1]] ]
; CHECK-NEXT: [[TMP17:%.*]] = extractelement <4 x i1> [[TMP3]], i64 2
; CHECK-NEXT: br i1 [[TMP17]], label %[[PRED_LOAD_IF3:.*]], label %[[PRED_LOAD_CONTINUE4:.*]]
; CHECK: [[PRED_LOAD_IF3]]:
; CHECK-NEXT: [[TMP18:%.*]] = extractelement <4 x i32> [[TMP4]], i64 2
; CHECK-NEXT: [[TMP19:%.*]] = getelementptr inbounds i32, ptr [[A]], i32 [[TMP18]]
; CHECK-NEXT: [[TMP20:%.*]] = load i32, ptr [[TMP19]], align 4
; CHECK-NEXT: [[TMP21:%.*]] = insertelement <4 x i32> [[TMP16]], i32 [[TMP20]], i64 2
; CHECK-NEXT: br label %[[PRED_LOAD_CONTINUE4]]
; CHECK: [[PRED_LOAD_CONTINUE4]]:
; CHECK-NEXT: [[TMP22:%.*]] = phi <4 x i32> [ [[TMP16]], %[[PRED_LOAD_CONTINUE2]] ], [ [[TMP21]], %[[PRED_LOAD_IF3]] ]
; CHECK-NEXT: [[TMP23:%.*]] = extractelement <4 x i1> [[TMP3]], i64 3
; CHECK-NEXT: br i1 [[TMP23]], label %[[PRED_LOAD_IF5:.*]], label %[[PRED_LOAD_CONTINUE6]]
; CHECK: [[PRED_LOAD_IF5]]:
; CHECK-NEXT: [[TMP24:%.*]] = extractelement <4 x i32> [[TMP4]], i64 3
; CHECK-NEXT: [[TMP25:%.*]] = getelementptr inbounds i32, ptr [[A]], i32 [[TMP24]]
; CHECK-NEXT: [[TMP26:%.*]] = load i32, ptr [[TMP25]], align 4
; CHECK-NEXT: [[TMP27:%.*]] = insertelement <4 x i32> [[TMP22]], i32 [[TMP26]], i64 3
; CHECK-NEXT: br label %[[PRED_LOAD_CONTINUE6]]
; CHECK: [[PRED_LOAD_CONTINUE6]]:
; CHECK-NEXT: [[TMP28:%.*]] = phi <4 x i32> [ [[TMP22]], %[[PRED_LOAD_CONTINUE4]] ], [ [[TMP27]], %[[PRED_LOAD_IF5]] ]
; CHECK-NEXT: [[PREDPHI:%.*]] = select <4 x i1> [[TMP3]], <4 x i32> [[TMP28]], <4 x i32> zeroinitializer
; CHECK-NEXT: [[TMP29]] = add <4 x i32> [[VEC_PHI]], [[PREDPHI]]
; CHECK-NEXT: [[INDEX_NEXT]] = add nuw i32 [[INDEX]], 4
; CHECK-NEXT: [[VEC_IND_NEXT]] = add <4 x i32> [[VEC_IND]], splat (i32 4)
; CHECK-NEXT: [[TMP30:%.*]] = icmp eq i32 [[INDEX_NEXT]], [[N_VEC]]
; CHECK-NEXT: br i1 [[TMP30]], label %[[MIDDLE_BLOCK:.*]], label %[[VECTOR_BODY]], !llvm.loop [[LOOP0:![0-9]+]]
; CHECK: [[MIDDLE_BLOCK]]:
; CHECK-NEXT: [[TMP31:%.*]] = call i32 @llvm.vector.reduce.add.v4i32(<4 x i32> [[TMP29]])
; CHECK-NEXT: [[CMP_N:%.*]] = icmp eq i32 [[N]], [[N_VEC]]
; CHECK-NEXT: br i1 [[CMP_N]], label %[[EXIT:.*]], label %[[SCALAR_PH]]
; CHECK: [[SCALAR_PH]]:
; CHECK-NEXT: [[BC_RESUME_VAL:%.*]] = phi i32 [ [[N_VEC]], %[[MIDDLE_BLOCK]] ], [ 0, %[[ENTRY]] ], [ 0, %[[VECTOR_SCEVCHECK]] ]
; CHECK-NEXT: [[BC_MERGE_RDX:%.*]] = phi i32 [ [[TMP31]], %[[MIDDLE_BLOCK]] ], [ 0, %[[ENTRY]] ], [ 0, %[[VECTOR_SCEVCHECK]] ]
; CHECK-NEXT: br label %[[LOOP:.*]]
; CHECK: [[LOOP]]:
; CHECK-NEXT: [[IV:%.*]] = phi i32 [ [[BC_RESUME_VAL]], %[[SCALAR_PH]] ], [ [[IV_NEXT:%.*]], %[[LATCH:.*]] ]
; CHECK-NEXT: [[SUM:%.*]] = phi i32 [ [[BC_MERGE_RDX]], %[[SCALAR_PH]] ], [ [[SUM_NEXT:%.*]], %[[LATCH]] ]
; CHECK-NEXT: [[AND_IV:%.*]] = and i32 [[IV]], 1
; CHECK-NEXT: [[CMP:%.*]] = icmp eq i32 [[AND_IV]], 0
; CHECK-NEXT: br i1 [[CMP]], label %[[THEN:.*]], label %[[LATCH]]
; CHECK: [[THEN]]:
; CHECK-NEXT: [[CLAMPED:%.*]] = urem i32 [[IV]], 128
; CHECK-NEXT: [[GEP_A:%.*]] = getelementptr inbounds i32, ptr [[A]], i32 [[CLAMPED]]
; CHECK-NEXT: [[LV:%.*]] = load i32, ptr [[GEP_A]], align 4
; CHECK-NEXT: br label %[[LATCH]]
; CHECK: [[LATCH]]:
; CHECK-NEXT: [[V:%.*]] = phi i32 [ [[LV]], %[[THEN]] ], [ 0, %[[LOOP]] ]
; CHECK-NEXT: [[SUM_NEXT]] = add i32 [[SUM]], [[V]]
; CHECK-NEXT: [[IV_NEXT]] = add i32 [[IV]], 1
; CHECK-NEXT: [[EC:%.*]] = icmp eq i32 [[IV_NEXT]], [[N]]
; CHECK-NEXT: br i1 [[EC]], label %[[EXIT]], label %[[LOOP]], !llvm.loop [[LOOP3:![0-9]+]]
; CHECK: [[EXIT]]:
; CHECK-NEXT: [[SUM_NEXT_LCSSA:%.*]] = phi i32 [ [[SUM_NEXT]], %[[LATCH]] ], [ [[TMP31]], %[[MIDDLE_BLOCK]] ]
; CHECK-NEXT: ret i32 [[SUM_NEXT_LCSSA]]
;
; IC4-LABEL: define i32 @clamped_load_predicated(
; IC4-SAME: ptr [[A:%.*]], i32 [[N:%.*]]) {
; IC4-NEXT: [[ENTRY:.*]]:
; IC4-NEXT: [[MIN_ITERS_CHECK:%.*]] = icmp ult i32 [[N]], 16
; IC4-NEXT: br i1 [[MIN_ITERS_CHECK]], label %[[SCALAR_PH:.*]], label %[[VECTOR_SCEVCHECK:.*]]
; IC4: [[VECTOR_SCEVCHECK]]:
; IC4-NEXT: [[TMP0:%.*]] = add i32 [[N]], -1
; IC4-NEXT: [[TMP1:%.*]] = icmp ugt i32 [[TMP0]], 127
; IC4-NEXT: br i1 [[TMP1]], label %[[SCALAR_PH]], label %[[VECTOR_PH:.*]]
; IC4: [[VECTOR_PH]]:
; IC4-NEXT: [[N_MOD_VF:%.*]] = urem i32 [[N]], 16
; IC4-NEXT: [[N_VEC:%.*]] = sub i32 [[N]], [[N_MOD_VF]]
; IC4-NEXT: br label %[[VECTOR_BODY:.*]]
; IC4: [[VECTOR_BODY]]:
; IC4-NEXT: [[INDEX:%.*]] = phi i32 [ 0, %[[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], %[[PRED_LOAD_CONTINUE33:.*]] ]
; IC4-NEXT: [[VEC_IND:%.*]] = phi <4 x i32> [ <i32 0, i32 1, i32 2, i32 3>, %[[VECTOR_PH]] ], [ [[VEC_IND_NEXT:%.*]], %[[PRED_LOAD_CONTINUE33]] ]
; IC4-NEXT: [[VEC_PHI:%.*]] = phi <4 x i32> [ zeroinitializer, %[[VECTOR_PH]] ], [ [[TMP110:%.*]], %[[PRED_LOAD_CONTINUE33]] ]
; IC4-NEXT: [[VEC_PHI1:%.*]] = phi <4 x i32> [ zeroinitializer, %[[VECTOR_PH]] ], [ [[TMP111:%.*]], %[[PRED_LOAD_CONTINUE33]] ]
; IC4-NEXT: [[VEC_PHI2:%.*]] = phi <4 x i32> [ zeroinitializer, %[[VECTOR_PH]] ], [ [[TMP112:%.*]], %[[PRED_LOAD_CONTINUE33]] ]
; IC4-NEXT: [[VEC_PHI3:%.*]] = phi <4 x i32> [ zeroinitializer, %[[VECTOR_PH]] ], [ [[TMP113:%.*]], %[[PRED_LOAD_CONTINUE33]] ]
; IC4-NEXT: [[STEP_ADD:%.*]] = add nuw <4 x i32> [[VEC_IND]], splat (i32 4)
; IC4-NEXT: [[STEP_ADD_2:%.*]] = add nuw <4 x i32> [[STEP_ADD]], splat (i32 4)
; IC4-NEXT: [[STEP_ADD_3:%.*]] = add nuw <4 x i32> [[STEP_ADD_2]], splat (i32 4)
; IC4-NEXT: [[TMP2:%.*]] = and <4 x i32> [[VEC_IND]], splat (i32 1)
; IC4-NEXT: [[TMP3:%.*]] = and <4 x i32> [[STEP_ADD]], splat (i32 1)
; IC4-NEXT: [[TMP4:%.*]] = and <4 x i32> [[STEP_ADD_2]], splat (i32 1)
; IC4-NEXT: [[TMP5:%.*]] = and <4 x i32> [[STEP_ADD_3]], splat (i32 1)
; IC4-NEXT: [[TMP6:%.*]] = icmp eq <4 x i32> [[TMP2]], zeroinitializer
; IC4-NEXT: [[TMP7:%.*]] = icmp eq <4 x i32> [[TMP3]], zeroinitializer
; IC4-NEXT: [[TMP8:%.*]] = icmp eq <4 x i32> [[TMP4]], zeroinitializer
; IC4-NEXT: [[TMP9:%.*]] = icmp eq <4 x i32> [[TMP5]], zeroinitializer
; IC4-NEXT: [[TMP10:%.*]] = urem <4 x i32> [[VEC_IND]], splat (i32 128)
; IC4-NEXT: [[TMP11:%.*]] = urem <4 x i32> [[STEP_ADD]], splat (i32 128)
; IC4-NEXT: [[TMP12:%.*]] = urem <4 x i32> [[STEP_ADD_2]], splat (i32 128)
; IC4-NEXT: [[TMP13:%.*]] = urem <4 x i32> [[STEP_ADD_3]], splat (i32 128)
; IC4-NEXT: [[TMP14:%.*]] = extractelement <4 x i1> [[TMP6]], i64 0
; IC4-NEXT: br i1 [[TMP14]], label %[[PRED_LOAD_IF:.*]], label %[[PRED_LOAD_CONTINUE:.*]]
; IC4: [[PRED_LOAD_IF]]:
; IC4-NEXT: [[TMP15:%.*]] = extractelement <4 x i32> [[TMP10]], i64 0
; IC4-NEXT: [[TMP16:%.*]] = getelementptr inbounds i32, ptr [[A]], i32 [[TMP15]]
; IC4-NEXT: [[TMP17:%.*]] = load i32, ptr [[TMP16]], align 4
; IC4-NEXT: [[TMP18:%.*]] = insertelement <4 x i32> poison, i32 [[TMP17]], i64 0
; IC4-NEXT: br label %[[PRED_LOAD_CONTINUE]]
; IC4: [[PRED_LOAD_CONTINUE]]:
; IC4-NEXT: [[TMP19:%.*]] = phi <4 x i32> [ poison, %[[VECTOR_BODY]] ], [ [[TMP18]], %[[PRED_LOAD_IF]] ]
; IC4-NEXT: [[TMP20:%.*]] = extractelement <4 x i1> [[TMP6]], i64 1
; IC4-NEXT: br i1 [[TMP20]], label %[[PRED_LOAD_IF4:.*]], label %[[PRED_LOAD_CONTINUE5:.*]]
; IC4: [[PRED_LOAD_IF4]]:
; IC4-NEXT: [[TMP21:%.*]] = extractelement <4 x i32> [[TMP10]], i64 1
; IC4-NEXT: [[TMP22:%.*]] = getelementptr inbounds i32, ptr [[A]], i32 [[TMP21]]
; IC4-NEXT: [[TMP23:%.*]] = load i32, ptr [[TMP22]], align 4
; IC4-NEXT: [[TMP24:%.*]] = insertelement <4 x i32> [[TMP19]], i32 [[TMP23]], i64 1
; IC4-NEXT: br label %[[PRED_LOAD_CONTINUE5]]
; IC4: [[PRED_LOAD_CONTINUE5]]:
; IC4-NEXT: [[TMP25:%.*]] = phi <4 x i32> [ [[TMP19]], %[[PRED_LOAD_CONTINUE]] ], [ [[TMP24]], %[[PRED_LOAD_IF4]] ]
; IC4-NEXT: [[TMP26:%.*]] = extractelement <4 x i1> [[TMP6]], i64 2
; IC4-NEXT: br i1 [[TMP26]], label %[[PRED_LOAD_IF6:.*]], label %[[PRED_LOAD_CONTINUE7:.*]]
; IC4: [[PRED_LOAD_IF6]]:
; IC4-NEXT: [[TMP27:%.*]] = extractelement <4 x i32> [[TMP10]], i64 2
; IC4-NEXT: [[TMP28:%.*]] = getelementptr inbounds i32, ptr [[A]], i32 [[TMP27]]
; IC4-NEXT: [[TMP29:%.*]] = load i32, ptr [[TMP28]], align 4
; IC4-NEXT: [[TMP30:%.*]] = insertelement <4 x i32> [[TMP25]], i32 [[TMP29]], i64 2
; IC4-NEXT: br label %[[PRED_LOAD_CONTINUE7]]
; IC4: [[PRED_LOAD_CONTINUE7]]:
; IC4-NEXT: [[TMP31:%.*]] = phi <4 x i32> [ [[TMP25]], %[[PRED_LOAD_CONTINUE5]] ], [ [[TMP30]], %[[PRED_LOAD_IF6]] ]
; IC4-NEXT: [[TMP32:%.*]] = extractelement <4 x i1> [[TMP6]], i64 3
; IC4-NEXT: br i1 [[TMP32]], label %[[PRED_LOAD_IF8:.*]], label %[[PRED_LOAD_CONTINUE9:.*]]
; IC4: [[PRED_LOAD_IF8]]:
; IC4-NEXT: [[TMP33:%.*]] = extractelement <4 x i32> [[TMP10]], i64 3
; IC4-NEXT: [[TMP34:%.*]] = getelementptr inbounds i32, ptr [[A]], i32 [[TMP33]]
; IC4-NEXT: [[TMP35:%.*]] = load i32, ptr [[TMP34]], align 4
; IC4-NEXT: [[TMP36:%.*]] = insertelement <4 x i32> [[TMP31]], i32 [[TMP35]], i64 3
; IC4-NEXT: br label %[[PRED_LOAD_CONTINUE9]]
; IC4: [[PRED_LOAD_CONTINUE9]]:
; IC4-NEXT: [[TMP37:%.*]] = phi <4 x i32> [ [[TMP31]], %[[PRED_LOAD_CONTINUE7]] ], [ [[TMP36]], %[[PRED_LOAD_IF8]] ]
; IC4-NEXT: [[TMP38:%.*]] = extractelement <4 x i1> [[TMP7]], i64 0
; IC4-NEXT: br i1 [[TMP38]], label %[[PRED_LOAD_IF10:.*]], label %[[PRED_LOAD_CONTINUE11:.*]]
; IC4: [[PRED_LOAD_IF10]]:
; IC4-NEXT: [[TMP39:%.*]] = extractelement <4 x i32> [[TMP11]], i64 0
; IC4-NEXT: [[TMP40:%.*]] = getelementptr inbounds i32, ptr [[A]], i32 [[TMP39]]
; IC4-NEXT: [[TMP41:%.*]] = load i32, ptr [[TMP40]], align 4
; IC4-NEXT: [[TMP42:%.*]] = insertelement <4 x i32> poison, i32 [[TMP41]], i64 0
; IC4-NEXT: br label %[[PRED_LOAD_CONTINUE11]]
; IC4: [[PRED_LOAD_CONTINUE11]]:
; IC4-NEXT: [[TMP43:%.*]] = phi <4 x i32> [ poison, %[[PRED_LOAD_CONTINUE9]] ], [ [[TMP42]], %[[PRED_LOAD_IF10]] ]
; IC4-NEXT: [[TMP44:%.*]] = extractelement <4 x i1> [[TMP7]], i64 1
; IC4-NEXT: br i1 [[TMP44]], label %[[PRED_LOAD_IF12:.*]], label %[[PRED_LOAD_CONTINUE13:.*]]
; IC4: [[PRED_LOAD_IF12]]:
; IC4-NEXT: [[TMP45:%.*]] = extractelement <4 x i32> [[TMP11]], i64 1
; IC4-NEXT: [[TMP46:%.*]] = getelementptr inbounds i32, ptr [[A]], i32 [[TMP45]]
; IC4-NEXT: [[TMP47:%.*]] = load i32, ptr [[TMP46]], align 4
; IC4-NEXT: [[TMP48:%.*]] = insertelement <4 x i32> [[TMP43]], i32 [[TMP47]], i64 1
; IC4-NEXT: br label %[[PRED_LOAD_CONTINUE13]]
; IC4: [[PRED_LOAD_CONTINUE13]]:
; IC4-NEXT: [[TMP49:%.*]] = phi <4 x i32> [ [[TMP43]], %[[PRED_LOAD_CONTINUE11]] ], [ [[TMP48]], %[[PRED_LOAD_IF12]] ]
; IC4-NEXT: [[TMP50:%.*]] = extractelement <4 x i1> [[TMP7]], i64 2
; IC4-NEXT: br i1 [[TMP50]], label %[[PRED_LOAD_IF14:.*]], label %[[PRED_LOAD_CONTINUE15:.*]]
; IC4: [[PRED_LOAD_IF14]]:
; IC4-NEXT: [[TMP51:%.*]] = extractelement <4 x i32> [[TMP11]], i64 2
; IC4-NEXT: [[TMP52:%.*]] = getelementptr inbounds i32, ptr [[A]], i32 [[TMP51]]
; IC4-NEXT: [[TMP53:%.*]] = load i32, ptr [[TMP52]], align 4
; IC4-NEXT: [[TMP54:%.*]] = insertelement <4 x i32> [[TMP49]], i32 [[TMP53]], i64 2
; IC4-NEXT: br label %[[PRED_LOAD_CONTINUE15]]
; IC4: [[PRED_LOAD_CONTINUE15]]:
; IC4-NEXT: [[TMP55:%.*]] = phi <4 x i32> [ [[TMP49]], %[[PRED_LOAD_CONTINUE13]] ], [ [[TMP54]], %[[PRED_LOAD_IF14]] ]
; IC4-NEXT: [[TMP56:%.*]] = extractelement <4 x i1> [[TMP7]], i64 3
; IC4-NEXT: br i1 [[TMP56]], label %[[PRED_LOAD_IF16:.*]], label %[[PRED_LOAD_CONTINUE17:.*]]
; IC4: [[PRED_LOAD_IF16]]:
; IC4-NEXT: [[TMP57:%.*]] = extractelement <4 x i32> [[TMP11]], i64 3
; IC4-NEXT: [[TMP58:%.*]] = getelementptr inbounds i32, ptr [[A]], i32 [[TMP57]]
; IC4-NEXT: [[TMP59:%.*]] = load i32, ptr [[TMP58]], align 4
; IC4-NEXT: [[TMP60:%.*]] = insertelement <4 x i32> [[TMP55]], i32 [[TMP59]], i64 3
; IC4-NEXT: br label %[[PRED_LOAD_CONTINUE17]]
; IC4: [[PRED_LOAD_CONTINUE17]]:
; IC4-NEXT: [[TMP61:%.*]] = phi <4 x i32> [ [[TMP55]], %[[PRED_LOAD_CONTINUE15]] ], [ [[TMP60]], %[[PRED_LOAD_IF16]] ]
; IC4-NEXT: [[TMP62:%.*]] = extractelement <4 x i1> [[TMP8]], i64 0
; IC4-NEXT: br i1 [[TMP62]], label %[[PRED_LOAD_IF18:.*]], label %[[PRED_LOAD_CONTINUE19:.*]]
; IC4: [[PRED_LOAD_IF18]]:
; IC4-NEXT: [[TMP63:%.*]] = extractelement <4 x i32> [[TMP12]], i64 0
; IC4-NEXT: [[TMP64:%.*]] = getelementptr inbounds i32, ptr [[A]], i32 [[TMP63]]
; IC4-NEXT: [[TMP65:%.*]] = load i32, ptr [[TMP64]], align 4
; IC4-NEXT: [[TMP66:%.*]] = insertelement <4 x i32> poison, i32 [[TMP65]], i64 0
; IC4-NEXT: br label %[[PRED_LOAD_CONTINUE19]]
; IC4: [[PRED_LOAD_CONTINUE19]]:
; IC4-NEXT: [[TMP67:%.*]] = phi <4 x i32> [ poison, %[[PRED_LOAD_CONTINUE17]] ], [ [[TMP66]], %[[PRED_LOAD_IF18]] ]
; IC4-NEXT: [[TMP68:%.*]] = extractelement <4 x i1> [[TMP8]], i64 1
; IC4-NEXT: br i1 [[TMP68]], label %[[PRED_LOAD_IF20:.*]], label %[[PRED_LOAD_CONTINUE21:.*]]
; IC4: [[PRED_LOAD_IF20]]:
; IC4-NEXT: [[TMP69:%.*]] = extractelement <4 x i32> [[TMP12]], i64 1
; IC4-NEXT: [[TMP70:%.*]] = getelementptr inbounds i32, ptr [[A]], i32 [[TMP69]]
; IC4-NEXT: [[TMP71:%.*]] = load i32, ptr [[TMP70]], align 4
; IC4-NEXT: [[TMP72:%.*]] = insertelement <4 x i32> [[TMP67]], i32 [[TMP71]], i64 1
; IC4-NEXT: br label %[[PRED_LOAD_CONTINUE21]]
; IC4: [[PRED_LOAD_CONTINUE21]]:
; IC4-NEXT: [[TMP73:%.*]] = phi <4 x i32> [ [[TMP67]], %[[PRED_LOAD_CONTINUE19]] ], [ [[TMP72]], %[[PRED_LOAD_IF20]] ]
; IC4-NEXT: [[TMP74:%.*]] = extractelement <4 x i1> [[TMP8]], i64 2
; IC4-NEXT: br i1 [[TMP74]], label %[[PRED_LOAD_IF22:.*]], label %[[PRED_LOAD_CONTINUE23:.*]]
; IC4: [[PRED_LOAD_IF22]]:
; IC4-NEXT: [[TMP75:%.*]] = extractelement <4 x i32> [[TMP12]], i64 2
; IC4-NEXT: [[TMP76:%.*]] = getelementptr inbounds i32, ptr [[A]], i32 [[TMP75]]
; IC4-NEXT: [[TMP77:%.*]] = load i32, ptr [[TMP76]], align 4
; IC4-NEXT: [[TMP78:%.*]] = insertelement <4 x i32> [[TMP73]], i32 [[TMP77]], i64 2
; IC4-NEXT: br label %[[PRED_LOAD_CONTINUE23]]
; IC4: [[PRED_LOAD_CONTINUE23]]:
; IC4-NEXT: [[TMP79:%.*]] = phi <4 x i32> [ [[TMP73]], %[[PRED_LOAD_CONTINUE21]] ], [ [[TMP78]], %[[PRED_LOAD_IF22]] ]
; IC4-NEXT: [[TMP80:%.*]] = extractelement <4 x i1> [[TMP8]], i64 3
; IC4-NEXT: br i1 [[TMP80]], label %[[PRED_LOAD_IF24:.*]], label %[[PRED_LOAD_CONTINUE25:.*]]
; IC4: [[PRED_LOAD_IF24]]:
; IC4-NEXT: [[TMP81:%.*]] = extractelement <4 x i32> [[TMP12]], i64 3
; IC4-NEXT: [[TMP82:%.*]] = getelementptr inbounds i32, ptr [[A]], i32 [[TMP81]]
; IC4-NEXT: [[TMP83:%.*]] = load i32, ptr [[TMP82]], align 4
; IC4-NEXT: [[TMP84:%.*]] = insertelement <4 x i32> [[TMP79]], i32 [[TMP83]], i64 3
; IC4-NEXT: br label %[[PRED_LOAD_CONTINUE25]]
; IC4: [[PRED_LOAD_CONTINUE25]]:
; IC4-NEXT: [[TMP85:%.*]] = phi <4 x i32> [ [[TMP79]], %[[PRED_LOAD_CONTINUE23]] ], [ [[TMP84]], %[[PRED_LOAD_IF24]] ]
; IC4-NEXT: [[TMP86:%.*]] = extractelement <4 x i1> [[TMP9]], i64 0
; IC4-NEXT: br i1 [[TMP86]], label %[[PRED_LOAD_IF26:.*]], label %[[PRED_LOAD_CONTINUE27:.*]]
; IC4: [[PRED_LOAD_IF26]]:
; IC4-NEXT: [[TMP87:%.*]] = extractelement <4 x i32> [[TMP13]], i64 0
; IC4-NEXT: [[TMP88:%.*]] = getelementptr inbounds i32, ptr [[A]], i32 [[TMP87]]
; IC4-NEXT: [[TMP89:%.*]] = load i32, ptr [[TMP88]], align 4
; IC4-NEXT: [[TMP90:%.*]] = insertelement <4 x i32> poison, i32 [[TMP89]], i64 0
; IC4-NEXT: br label %[[PRED_LOAD_CONTINUE27]]
; IC4: [[PRED_LOAD_CONTINUE27]]:
; IC4-NEXT: [[TMP91:%.*]] = phi <4 x i32> [ poison, %[[PRED_LOAD_CONTINUE25]] ], [ [[TMP90]], %[[PRED_LOAD_IF26]] ]
; IC4-NEXT: [[TMP92:%.*]] = extractelement <4 x i1> [[TMP9]], i64 1
; IC4-NEXT: br i1 [[TMP92]], label %[[PRED_LOAD_IF28:.*]], label %[[PRED_LOAD_CONTINUE29:.*]]
; IC4: [[PRED_LOAD_IF28]]:
; IC4-NEXT: [[TMP93:%.*]] = extractelement <4 x i32> [[TMP13]], i64 1
; IC4-NEXT: [[TMP94:%.*]] = getelementptr inbounds i32, ptr [[A]], i32 [[TMP93]]
; IC4-NEXT: [[TMP95:%.*]] = load i32, ptr [[TMP94]], align 4
; IC4-NEXT: [[TMP96:%.*]] = insertelement <4 x i32> [[TMP91]], i32 [[TMP95]], i64 1
; IC4-NEXT: br label %[[PRED_LOAD_CONTINUE29]]
; IC4: [[PRED_LOAD_CONTINUE29]]:
; IC4-NEXT: [[TMP97:%.*]] = phi <4 x i32> [ [[TMP91]], %[[PRED_LOAD_CONTINUE27]] ], [ [[TMP96]], %[[PRED_LOAD_IF28]] ]
; IC4-NEXT: [[TMP98:%.*]] = extractelement <4 x i1> [[TMP9]], i64 2
; IC4-NEXT: br i1 [[TMP98]], label %[[PRED_LOAD_IF30:.*]], label %[[PRED_LOAD_CONTINUE31:.*]]
; IC4: [[PRED_LOAD_IF30]]:
; IC4-NEXT: [[TMP99:%.*]] = extractelement <4 x i32> [[TMP13]], i64 2
; IC4-NEXT: [[TMP100:%.*]] = getelementptr inbounds i32, ptr [[A]], i32 [[TMP99]]
; IC4-NEXT: [[TMP101:%.*]] = load i32, ptr [[TMP100]], align 4
; IC4-NEXT: [[TMP102:%.*]] = insertelement <4 x i32> [[TMP97]], i32 [[TMP101]], i64 2
; IC4-NEXT: br label %[[PRED_LOAD_CONTINUE31]]
; IC4: [[PRED_LOAD_CONTINUE31]]:
; IC4-NEXT: [[TMP103:%.*]] = phi <4 x i32> [ [[TMP97]], %[[PRED_LOAD_CONTINUE29]] ], [ [[TMP102]], %[[PRED_LOAD_IF30]] ]
; IC4-NEXT: [[TMP104:%.*]] = extractelement <4 x i1> [[TMP9]], i64 3
; IC4-NEXT: br i1 [[TMP104]], label %[[PRED_LOAD_IF32:.*]], label %[[PRED_LOAD_CONTINUE33]]
; IC4: [[PRED_LOAD_IF32]]:
; IC4-NEXT: [[TMP105:%.*]] = extractelement <4 x i32> [[TMP13]], i64 3
; IC4-NEXT: [[TMP106:%.*]] = getelementptr inbounds i32, ptr [[A]], i32 [[TMP105]]
; IC4-NEXT: [[TMP107:%.*]] = load i32, ptr [[TMP106]], align 4
; IC4-NEXT: [[TMP108:%.*]] = insertelement <4 x i32> [[TMP103]], i32 [[TMP107]], i64 3
; IC4-NEXT: br label %[[PRED_LOAD_CONTINUE33]]
; IC4: [[PRED_LOAD_CONTINUE33]]:
; IC4-NEXT: [[TMP109:%.*]] = phi <4 x i32> [ [[TMP103]], %[[PRED_LOAD_CONTINUE31]] ], [ [[TMP108]], %[[PRED_LOAD_IF32]] ]
; IC4-NEXT: [[PREDPHI:%.*]] = select <4 x i1> [[TMP6]], <4 x i32> [[TMP37]], <4 x i32> zeroinitializer
; IC4-NEXT: [[PREDPHI34:%.*]] = select <4 x i1> [[TMP7]], <4 x i32> [[TMP61]], <4 x i32> zeroinitializer
; IC4-NEXT: [[PREDPHI35:%.*]] = select <4 x i1> [[TMP8]], <4 x i32> [[TMP85]], <4 x i32> zeroinitializer
; IC4-NEXT: [[PREDPHI36:%.*]] = select <4 x i1> [[TMP9]], <4 x i32> [[TMP109]], <4 x i32> zeroinitializer
; IC4-NEXT: [[TMP110]] = add <4 x i32> [[VEC_PHI]], [[PREDPHI]]
; IC4-NEXT: [[TMP111]] = add <4 x i32> [[VEC_PHI1]], [[PREDPHI34]]
; IC4-NEXT: [[TMP112]] = add <4 x i32> [[VEC_PHI2]], [[PREDPHI35]]
; IC4-NEXT: [[TMP113]] = add <4 x i32> [[VEC_PHI3]], [[PREDPHI36]]
; IC4-NEXT: [[INDEX_NEXT]] = add nuw i32 [[INDEX]], 16
; IC4-NEXT: [[VEC_IND_NEXT]] = add <4 x i32> [[STEP_ADD_3]], splat (i32 4)
; IC4-NEXT: [[TMP114:%.*]] = icmp eq i32 [[INDEX_NEXT]], [[N_VEC]]
; IC4-NEXT: br i1 [[TMP114]], label %[[MIDDLE_BLOCK:.*]], label %[[VECTOR_BODY]], !llvm.loop [[LOOP0:![0-9]+]]
; IC4: [[MIDDLE_BLOCK]]:
; IC4-NEXT: [[BIN_RDX:%.*]] = add <4 x i32> [[TMP111]], [[TMP110]]
; IC4-NEXT: [[BIN_RDX37:%.*]] = add <4 x i32> [[TMP112]], [[BIN_RDX]]
; IC4-NEXT: [[BIN_RDX38:%.*]] = add <4 x i32> [[TMP113]], [[BIN_RDX37]]
; IC4-NEXT: [[TMP115:%.*]] = call i32 @llvm.vector.reduce.add.v4i32(<4 x i32> [[BIN_RDX38]])
; IC4-NEXT: [[CMP_N:%.*]] = icmp eq i32 [[N]], [[N_VEC]]
; IC4-NEXT: br i1 [[CMP_N]], label %[[EXIT:.*]], label %[[SCALAR_PH]]
; IC4: [[SCALAR_PH]]:
; IC4-NEXT: [[BC_RESUME_VAL:%.*]] = phi i32 [ [[N_VEC]], %[[MIDDLE_BLOCK]] ], [ 0, %[[ENTRY]] ], [ 0, %[[VECTOR_SCEVCHECK]] ]
; IC4-NEXT: [[BC_MERGE_RDX:%.*]] = phi i32 [ [[TMP115]], %[[MIDDLE_BLOCK]] ], [ 0, %[[ENTRY]] ], [ 0, %[[VECTOR_SCEVCHECK]] ]
; IC4-NEXT: br label %[[LOOP:.*]]
; IC4: [[LOOP]]:
; IC4-NEXT: [[IV:%.*]] = phi i32 [ [[BC_RESUME_VAL]], %[[SCALAR_PH]] ], [ [[IV_NEXT:%.*]], %[[LATCH:.*]] ]
; IC4-NEXT: [[SUM:%.*]] = phi i32 [ [[BC_MERGE_RDX]], %[[SCALAR_PH]] ], [ [[SUM_NEXT:%.*]], %[[LATCH]] ]
; IC4-NEXT: [[AND_IV:%.*]] = and i32 [[IV]], 1
; IC4-NEXT: [[CMP:%.*]] = icmp eq i32 [[AND_IV]], 0
; IC4-NEXT: br i1 [[CMP]], label %[[THEN:.*]], label %[[LATCH]]
; IC4: [[THEN]]:
; IC4-NEXT: [[CLAMPED:%.*]] = urem i32 [[IV]], 128
; IC4-NEXT: [[GEP_A:%.*]] = getelementptr inbounds i32, ptr [[A]], i32 [[CLAMPED]]
; IC4-NEXT: [[LV:%.*]] = load i32, ptr [[GEP_A]], align 4
; IC4-NEXT: br label %[[LATCH]]
; IC4: [[LATCH]]:
; IC4-NEXT: [[V:%.*]] = phi i32 [ [[LV]], %[[THEN]] ], [ 0, %[[LOOP]] ]
; IC4-NEXT: [[SUM_NEXT]] = add i32 [[SUM]], [[V]]
; IC4-NEXT: [[IV_NEXT]] = add i32 [[IV]], 1
; IC4-NEXT: [[EC:%.*]] = icmp eq i32 [[IV_NEXT]], [[N]]
; IC4-NEXT: br i1 [[EC]], label %[[EXIT]], label %[[LOOP]], !llvm.loop [[LOOP3:![0-9]+]]
; IC4: [[EXIT]]:
; IC4-NEXT: [[SUM_NEXT_LCSSA:%.*]] = phi i32 [ [[SUM_NEXT]], %[[LATCH]] ], [ [[TMP115]], %[[MIDDLE_BLOCK]] ]
; IC4-NEXT: ret i32 [[SUM_NEXT_LCSSA]]
;
; IC8-LABEL: define i32 @clamped_load_predicated(
; IC8-SAME: ptr [[A:%.*]], i32 [[N:%.*]]) {
; IC8-NEXT: [[ENTRY:.*]]:
; IC8-NEXT: [[MIN_ITERS_CHECK:%.*]] = icmp ult i32 [[N]], 32
; IC8-NEXT: br i1 [[MIN_ITERS_CHECK]], label %[[SCALAR_PH:.*]], label %[[VECTOR_SCEVCHECK:.*]]
; IC8: [[VECTOR_SCEVCHECK]]:
; IC8-NEXT: [[TMP0:%.*]] = add i32 [[N]], -1
; IC8-NEXT: [[TMP1:%.*]] = icmp ugt i32 [[TMP0]], 127
; IC8-NEXT: br i1 [[TMP1]], label %[[SCALAR_PH]], label %[[VECTOR_PH:.*]]
; IC8: [[VECTOR_PH]]:
; IC8-NEXT: [[N_MOD_VF:%.*]] = urem i32 [[N]], 32
; IC8-NEXT: [[N_VEC:%.*]] = sub i32 [[N]], [[N_MOD_VF]]
; IC8-NEXT: br label %[[VECTOR_BODY:.*]]
; IC8: [[VECTOR_BODY]]:
; IC8-NEXT: [[INDEX:%.*]] = phi i32 [ 0, %[[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], %[[PRED_LOAD_CONTINUE69:.*]] ]
; IC8-NEXT: [[VEC_IND:%.*]] = phi <4 x i32> [ <i32 0, i32 1, i32 2, i32 3>, %[[VECTOR_PH]] ], [ [[VEC_IND_NEXT:%.*]], %[[PRED_LOAD_CONTINUE69]] ]
; IC8-NEXT: [[VEC_PHI:%.*]] = phi <4 x i32> [ zeroinitializer, %[[VECTOR_PH]] ], [ [[TMP218:%.*]], %[[PRED_LOAD_CONTINUE69]] ]
; IC8-NEXT: [[VEC_PHI1:%.*]] = phi <4 x i32> [ zeroinitializer, %[[VECTOR_PH]] ], [ [[TMP219:%.*]], %[[PRED_LOAD_CONTINUE69]] ]
; IC8-NEXT: [[VEC_PHI2:%.*]] = phi <4 x i32> [ zeroinitializer, %[[VECTOR_PH]] ], [ [[TMP220:%.*]], %[[PRED_LOAD_CONTINUE69]] ]
; IC8-NEXT: [[VEC_PHI3:%.*]] = phi <4 x i32> [ zeroinitializer, %[[VECTOR_PH]] ], [ [[TMP221:%.*]], %[[PRED_LOAD_CONTINUE69]] ]
; IC8-NEXT: [[VEC_PHI4:%.*]] = phi <4 x i32> [ zeroinitializer, %[[VECTOR_PH]] ], [ [[TMP222:%.*]], %[[PRED_LOAD_CONTINUE69]] ]
; IC8-NEXT: [[VEC_PHI5:%.*]] = phi <4 x i32> [ zeroinitializer, %[[VECTOR_PH]] ], [ [[TMP223:%.*]], %[[PRED_LOAD_CONTINUE69]] ]
; IC8-NEXT: [[VEC_PHI6:%.*]] = phi <4 x i32> [ zeroinitializer, %[[VECTOR_PH]] ], [ [[TMP224:%.*]], %[[PRED_LOAD_CONTINUE69]] ]
; IC8-NEXT: [[VEC_PHI7:%.*]] = phi <4 x i32> [ zeroinitializer, %[[VECTOR_PH]] ], [ [[TMP225:%.*]], %[[PRED_LOAD_CONTINUE69]] ]
; IC8-NEXT: [[STEP_ADD:%.*]] = add nuw <4 x i32> [[VEC_IND]], splat (i32 4)
; IC8-NEXT: [[STEP_ADD_2:%.*]] = add nuw <4 x i32> [[STEP_ADD]], splat (i32 4)
; IC8-NEXT: [[STEP_ADD_3:%.*]] = add nuw <4 x i32> [[STEP_ADD_2]], splat (i32 4)
; IC8-NEXT: [[STEP_ADD_4:%.*]] = add nuw <4 x i32> [[STEP_ADD_3]], splat (i32 4)
; IC8-NEXT: [[STEP_ADD_5:%.*]] = add nuw <4 x i32> [[STEP_ADD_4]], splat (i32 4)
; IC8-NEXT: [[STEP_ADD_6:%.*]] = add nuw <4 x i32> [[STEP_ADD_5]], splat (i32 4)
; IC8-NEXT: [[STEP_ADD_7:%.*]] = add nuw <4 x i32> [[STEP_ADD_6]], splat (i32 4)
; IC8-NEXT: [[TMP2:%.*]] = and <4 x i32> [[VEC_IND]], splat (i32 1)
; IC8-NEXT: [[TMP3:%.*]] = and <4 x i32> [[STEP_ADD]], splat (i32 1)
; IC8-NEXT: [[TMP4:%.*]] = and <4 x i32> [[STEP_ADD_2]], splat (i32 1)
; IC8-NEXT: [[TMP5:%.*]] = and <4 x i32> [[STEP_ADD_3]], splat (i32 1)
; IC8-NEXT: [[TMP6:%.*]] = and <4 x i32> [[STEP_ADD_4]], splat (i32 1)
; IC8-NEXT: [[TMP7:%.*]] = and <4 x i32> [[STEP_ADD_5]], splat (i32 1)
; IC8-NEXT: [[TMP8:%.*]] = and <4 x i32> [[STEP_ADD_6]], splat (i32 1)
; IC8-NEXT: [[TMP9:%.*]] = and <4 x i32> [[STEP_ADD_7]], splat (i32 1)
; IC8-NEXT: [[TMP10:%.*]] = icmp eq <4 x i32> [[TMP2]], zeroinitializer
; IC8-NEXT: [[TMP11:%.*]] = icmp eq <4 x i32> [[TMP3]], zeroinitializer
; IC8-NEXT: [[TMP12:%.*]] = icmp eq <4 x i32> [[TMP4]], zeroinitializer
; IC8-NEXT: [[TMP13:%.*]] = icmp eq <4 x i32> [[TMP5]], zeroinitializer
; IC8-NEXT: [[TMP14:%.*]] = icmp eq <4 x i32> [[TMP6]], zeroinitializer
; IC8-NEXT: [[TMP15:%.*]] = icmp eq <4 x i32> [[TMP7]], zeroinitializer
; IC8-NEXT: [[TMP16:%.*]] = icmp eq <4 x i32> [[TMP8]], zeroinitializer
; IC8-NEXT: [[TMP17:%.*]] = icmp eq <4 x i32> [[TMP9]], zeroinitializer
; IC8-NEXT: [[TMP18:%.*]] = urem <4 x i32> [[VEC_IND]], splat (i32 128)
; IC8-NEXT: [[TMP19:%.*]] = urem <4 x i32> [[STEP_ADD]], splat (i32 128)
; IC8-NEXT: [[TMP20:%.*]] = urem <4 x i32> [[STEP_ADD_2]], splat (i32 128)
; IC8-NEXT: [[TMP21:%.*]] = urem <4 x i32> [[STEP_ADD_3]], splat (i32 128)
; IC8-NEXT: [[TMP22:%.*]] = urem <4 x i32> [[STEP_ADD_4]], splat (i32 128)
; IC8-NEXT: [[TMP23:%.*]] = urem <4 x i32> [[STEP_ADD_5]], splat (i32 128)
; IC8-NEXT: [[TMP24:%.*]] = urem <4 x i32> [[STEP_ADD_6]], splat (i32 128)
; IC8-NEXT: [[TMP25:%.*]] = urem <4 x i32> [[STEP_ADD_7]], splat (i32 128)
; IC8-NEXT: [[TMP26:%.*]] = extractelement <4 x i1> [[TMP10]], i64 0
; IC8-NEXT: br i1 [[TMP26]], label %[[PRED_LOAD_IF:.*]], label %[[PRED_LOAD_CONTINUE:.*]]
; IC8: [[PRED_LOAD_IF]]:
; IC8-NEXT: [[TMP27:%.*]] = extractelement <4 x i32> [[TMP18]], i64 0
; IC8-NEXT: [[TMP28:%.*]] = getelementptr inbounds i32, ptr [[A]], i32 [[TMP27]]
; IC8-NEXT: [[TMP29:%.*]] = load i32, ptr [[TMP28]], align 4
; IC8-NEXT: [[TMP30:%.*]] = insertelement <4 x i32> poison, i32 [[TMP29]], i64 0
; IC8-NEXT: br label %[[PRED_LOAD_CONTINUE]]
; IC8: [[PRED_LOAD_CONTINUE]]:
; IC8-NEXT: [[TMP31:%.*]] = phi <4 x i32> [ poison, %[[VECTOR_BODY]] ], [ [[TMP30]], %[[PRED_LOAD_IF]] ]
; IC8-NEXT: [[TMP32:%.*]] = extractelement <4 x i1> [[TMP10]], i64 1
; IC8-NEXT: br i1 [[TMP32]], label %[[PRED_LOAD_IF8:.*]], label %[[PRED_LOAD_CONTINUE9:.*]]
; IC8: [[PRED_LOAD_IF8]]:
; IC8-NEXT: [[TMP33:%.*]] = extractelement <4 x i32> [[TMP18]], i64 1
; IC8-NEXT: [[TMP34:%.*]] = getelementptr inbounds i32, ptr [[A]], i32 [[TMP33]]
; IC8-NEXT: [[TMP35:%.*]] = load i32, ptr [[TMP34]], align 4
; IC8-NEXT: [[TMP36:%.*]] = insertelement <4 x i32> [[TMP31]], i32 [[TMP35]], i64 1
; IC8-NEXT: br label %[[PRED_LOAD_CONTINUE9]]
; IC8: [[PRED_LOAD_CONTINUE9]]:
; IC8-NEXT: [[TMP37:%.*]] = phi <4 x i32> [ [[TMP31]], %[[PRED_LOAD_CONTINUE]] ], [ [[TMP36]], %[[PRED_LOAD_IF8]] ]
; IC8-NEXT: [[TMP38:%.*]] = extractelement <4 x i1> [[TMP10]], i64 2
; IC8-NEXT: br i1 [[TMP38]], label %[[PRED_LOAD_IF10:.*]], label %[[PRED_LOAD_CONTINUE11:.*]]
; IC8: [[PRED_LOAD_IF10]]:
; IC8-NEXT: [[TMP39:%.*]] = extractelement <4 x i32> [[TMP18]], i64 2
; IC8-NEXT: [[TMP40:%.*]] = getelementptr inbounds i32, ptr [[A]], i32 [[TMP39]]
; IC8-NEXT: [[TMP41:%.*]] = load i32, ptr [[TMP40]], align 4
; IC8-NEXT: [[TMP42:%.*]] = insertelement <4 x i32> [[TMP37]], i32 [[TMP41]], i64 2
; IC8-NEXT: br label %[[PRED_LOAD_CONTINUE11]]
; IC8: [[PRED_LOAD_CONTINUE11]]:
; IC8-NEXT: [[TMP43:%.*]] = phi <4 x i32> [ [[TMP37]], %[[PRED_LOAD_CONTINUE9]] ], [ [[TMP42]], %[[PRED_LOAD_IF10]] ]
; IC8-NEXT: [[TMP44:%.*]] = extractelement <4 x i1> [[TMP10]], i64 3
; IC8-NEXT: br i1 [[TMP44]], label %[[PRED_LOAD_IF12:.*]], label %[[PRED_LOAD_CONTINUE13:.*]]
; IC8: [[PRED_LOAD_IF12]]:
; IC8-NEXT: [[TMP45:%.*]] = extractelement <4 x i32> [[TMP18]], i64 3
; IC8-NEXT: [[TMP46:%.*]] = getelementptr inbounds i32, ptr [[A]], i32 [[TMP45]]
; IC8-NEXT: [[TMP47:%.*]] = load i32, ptr [[TMP46]], align 4
; IC8-NEXT: [[TMP48:%.*]] = insertelement <4 x i32> [[TMP43]], i32 [[TMP47]], i64 3
; IC8-NEXT: br label %[[PRED_LOAD_CONTINUE13]]
; IC8: [[PRED_LOAD_CONTINUE13]]:
; IC8-NEXT: [[TMP49:%.*]] = phi <4 x i32> [ [[TMP43]], %[[PRED_LOAD_CONTINUE11]] ], [ [[TMP48]], %[[PRED_LOAD_IF12]] ]
; IC8-NEXT: [[TMP50:%.*]] = extractelement <4 x i1> [[TMP11]], i64 0
; IC8-NEXT: br i1 [[TMP50]], label %[[PRED_LOAD_IF14:.*]], label %[[PRED_LOAD_CONTINUE15:.*]]
; IC8: [[PRED_LOAD_IF14]]:
; IC8-NEXT: [[TMP51:%.*]] = extractelement <4 x i32> [[TMP19]], i64 0
; IC8-NEXT: [[TMP52:%.*]] = getelementptr inbounds i32, ptr [[A]], i32 [[TMP51]]
; IC8-NEXT: [[TMP53:%.*]] = load i32, ptr [[TMP52]], align 4
; IC8-NEXT: [[TMP54:%.*]] = insertelement <4 x i32> poison, i32 [[TMP53]], i64 0
; IC8-NEXT: br label %[[PRED_LOAD_CONTINUE15]]
; IC8: [[PRED_LOAD_CONTINUE15]]:
; IC8-NEXT: [[TMP55:%.*]] = phi <4 x i32> [ poison, %[[PRED_LOAD_CONTINUE13]] ], [ [[TMP54]], %[[PRED_LOAD_IF14]] ]
; IC8-NEXT: [[TMP56:%.*]] = extractelement <4 x i1> [[TMP11]], i64 1
; IC8-NEXT: br i1 [[TMP56]], label %[[PRED_LOAD_IF16:.*]], label %[[PRED_LOAD_CONTINUE17:.*]]
; IC8: [[PRED_LOAD_IF16]]:
; IC8-NEXT: [[TMP57:%.*]] = extractelement <4 x i32> [[TMP19]], i64 1
; IC8-NEXT: [[TMP58:%.*]] = getelementptr inbounds i32, ptr [[A]], i32 [[TMP57]]
; IC8-NEXT: [[TMP59:%.*]] = load i32, ptr [[TMP58]], align 4
; IC8-NEXT: [[TMP60:%.*]] = insertelement <4 x i32> [[TMP55]], i32 [[TMP59]], i64 1
; IC8-NEXT: br label %[[PRED_LOAD_CONTINUE17]]
; IC8: [[PRED_LOAD_CONTINUE17]]:
; IC8-NEXT: [[TMP61:%.*]] = phi <4 x i32> [ [[TMP55]], %[[PRED_LOAD_CONTINUE15]] ], [ [[TMP60]], %[[PRED_LOAD_IF16]] ]
; IC8-NEXT: [[TMP62:%.*]] = extractelement <4 x i1> [[TMP11]], i64 2
; IC8-NEXT: br i1 [[TMP62]], label %[[PRED_LOAD_IF18:.*]], label %[[PRED_LOAD_CONTINUE19:.*]]
; IC8: [[PRED_LOAD_IF18]]:
; IC8-NEXT: [[TMP63:%.*]] = extractelement <4 x i32> [[TMP19]], i64 2
; IC8-NEXT: [[TMP64:%.*]] = getelementptr inbounds i32, ptr [[A]], i32 [[TMP63]]
; IC8-NEXT: [[TMP65:%.*]] = load i32, ptr [[TMP64]], align 4
; IC8-NEXT: [[TMP66:%.*]] = insertelement <4 x i32> [[TMP61]], i32 [[TMP65]], i64 2
; IC8-NEXT: br label %[[PRED_LOAD_CONTINUE19]]
; IC8: [[PRED_LOAD_CONTINUE19]]:
; IC8-NEXT: [[TMP67:%.*]] = phi <4 x i32> [ [[TMP61]], %[[PRED_LOAD_CONTINUE17]] ], [ [[TMP66]], %[[PRED_LOAD_IF18]] ]
; IC8-NEXT: [[TMP68:%.*]] = extractelement <4 x i1> [[TMP11]], i64 3
; IC8-NEXT: br i1 [[TMP68]], label %[[PRED_LOAD_IF20:.*]], label %[[PRED_LOAD_CONTINUE21:.*]]
; IC8: [[PRED_LOAD_IF20]]:
; IC8-NEXT: [[TMP69:%.*]] = extractelement <4 x i32> [[TMP19]], i64 3
; IC8-NEXT: [[TMP70:%.*]] = getelementptr inbounds i32, ptr [[A]], i32 [[TMP69]]
; IC8-NEXT: [[TMP71:%.*]] = load i32, ptr [[TMP70]], align 4
; IC8-NEXT: [[TMP72:%.*]] = insertelement <4 x i32> [[TMP67]], i32 [[TMP71]], i64 3
; IC8-NEXT: br label %[[PRED_LOAD_CONTINUE21]]
; IC8: [[PRED_LOAD_CONTINUE21]]:
; IC8-NEXT: [[TMP73:%.*]] = phi <4 x i32> [ [[TMP67]], %[[PRED_LOAD_CONTINUE19]] ], [ [[TMP72]], %[[PRED_LOAD_IF20]] ]
; IC8-NEXT: [[TMP74:%.*]] = extractelement <4 x i1> [[TMP12]], i64 0
; IC8-NEXT: br i1 [[TMP74]], label %[[PRED_LOAD_IF22:.*]], label %[[PRED_LOAD_CONTINUE23:.*]]
; IC8: [[PRED_LOAD_IF22]]:
; IC8-NEXT: [[TMP75:%.*]] = extractelement <4 x i32> [[TMP20]], i64 0
; IC8-NEXT: [[TMP76:%.*]] = getelementptr inbounds i32, ptr [[A]], i32 [[TMP75]]
; IC8-NEXT: [[TMP77:%.*]] = load i32, ptr [[TMP76]], align 4
; IC8-NEXT: [[TMP78:%.*]] = insertelement <4 x i32> poison, i32 [[TMP77]], i64 0
; IC8-NEXT: br label %[[PRED_LOAD_CONTINUE23]]
; IC8: [[PRED_LOAD_CONTINUE23]]:
; IC8-NEXT: [[TMP79:%.*]] = phi <4 x i32> [ poison, %[[PRED_LOAD_CONTINUE21]] ], [ [[TMP78]], %[[PRED_LOAD_IF22]] ]
; IC8-NEXT: [[TMP80:%.*]] = extractelement <4 x i1> [[TMP12]], i64 1
; IC8-NEXT: br i1 [[TMP80]], label %[[PRED_LOAD_IF24:.*]], label %[[PRED_LOAD_CONTINUE25:.*]]
; IC8: [[PRED_LOAD_IF24]]:
; IC8-NEXT: [[TMP81:%.*]] = extractelement <4 x i32> [[TMP20]], i64 1
; IC8-NEXT: [[TMP82:%.*]] = getelementptr inbounds i32, ptr [[A]], i32 [[TMP81]]
; IC8-NEXT: [[TMP83:%.*]] = load i32, ptr [[TMP82]], align 4
; IC8-NEXT: [[TMP84:%.*]] = insertelement <4 x i32> [[TMP79]], i32 [[TMP83]], i64 1
; IC8-NEXT: br label %[[PRED_LOAD_CONTINUE25]]
; IC8: [[PRED_LOAD_CONTINUE25]]:
; IC8-NEXT: [[TMP85:%.*]] = phi <4 x i32> [ [[TMP79]], %[[PRED_LOAD_CONTINUE23]] ], [ [[TMP84]], %[[PRED_LOAD_IF24]] ]
; IC8-NEXT: [[TMP86:%.*]] = extractelement <4 x i1> [[TMP12]], i64 2
; IC8-NEXT: br i1 [[TMP86]], label %[[PRED_LOAD_IF26:.*]], label %[[PRED_LOAD_CONTINUE27:.*]]
; IC8: [[PRED_LOAD_IF26]]:
; IC8-NEXT: [[TMP87:%.*]] = extractelement <4 x i32> [[TMP20]], i64 2
; IC8-NEXT: [[TMP88:%.*]] = getelementptr inbounds i32, ptr [[A]], i32 [[TMP87]]
; IC8-NEXT: [[TMP89:%.*]] = load i32, ptr [[TMP88]], align 4
; IC8-NEXT: [[TMP90:%.*]] = insertelement <4 x i32> [[TMP85]], i32 [[TMP89]], i64 2
; IC8-NEXT: br label %[[PRED_LOAD_CONTINUE27]]
; IC8: [[PRED_LOAD_CONTINUE27]]:
; IC8-NEXT: [[TMP91:%.*]] = phi <4 x i32> [ [[TMP85]], %[[PRED_LOAD_CONTINUE25]] ], [ [[TMP90]], %[[PRED_LOAD_IF26]] ]
; IC8-NEXT: [[TMP92:%.*]] = extractelement <4 x i1> [[TMP12]], i64 3
; IC8-NEXT: br i1 [[TMP92]], label %[[PRED_LOAD_IF28:.*]], label %[[PRED_LOAD_CONTINUE29:.*]]
; IC8: [[PRED_LOAD_IF28]]:
; IC8-NEXT: [[TMP93:%.*]] = extractelement <4 x i32> [[TMP20]], i64 3
; IC8-NEXT: [[TMP94:%.*]] = getelementptr inbounds i32, ptr [[A]], i32 [[TMP93]]
; IC8-NEXT: [[TMP95:%.*]] = load i32, ptr [[TMP94]], align 4
; IC8-NEXT: [[TMP96:%.*]] = insertelement <4 x i32> [[TMP91]], i32 [[TMP95]], i64 3
; IC8-NEXT: br label %[[PRED_LOAD_CONTINUE29]]
; IC8: [[PRED_LOAD_CONTINUE29]]:
; IC8-NEXT: [[TMP97:%.*]] = phi <4 x i32> [ [[TMP91]], %[[PRED_LOAD_CONTINUE27]] ], [ [[TMP96]], %[[PRED_LOAD_IF28]] ]
; IC8-NEXT: [[TMP98:%.*]] = extractelement <4 x i1> [[TMP13]], i64 0
; IC8-NEXT: br i1 [[TMP98]], label %[[PRED_LOAD_IF30:.*]], label %[[PRED_LOAD_CONTINUE31:.*]]
; IC8: [[PRED_LOAD_IF30]]:
; IC8-NEXT: [[TMP99:%.*]] = extractelement <4 x i32> [[TMP21]], i64 0
; IC8-NEXT: [[TMP100:%.*]] = getelementptr inbounds i32, ptr [[A]], i32 [[TMP99]]
; IC8-NEXT: [[TMP101:%.*]] = load i32, ptr [[TMP100]], align 4
; IC8-NEXT: [[TMP102:%.*]] = insertelement <4 x i32> poison, i32 [[TMP101]], i64 0
; IC8-NEXT: br label %[[PRED_LOAD_CONTINUE31]]
; IC8: [[PRED_LOAD_CONTINUE31]]:
; IC8-NEXT: [[TMP103:%.*]] = phi <4 x i32> [ poison, %[[PRED_LOAD_CONTINUE29]] ], [ [[TMP102]], %[[PRED_LOAD_IF30]] ]
; IC8-NEXT: [[TMP104:%.*]] = extractelement <4 x i1> [[TMP13]], i64 1
; IC8-NEXT: br i1 [[TMP104]], label %[[PRED_LOAD_IF32:.*]], label %[[PRED_LOAD_CONTINUE33:.*]]
; IC8: [[PRED_LOAD_IF32]]:
; IC8-NEXT: [[TMP105:%.*]] = extractelement <4 x i32> [[TMP21]], i64 1
; IC8-NEXT: [[TMP106:%.*]] = getelementptr inbounds i32, ptr [[A]], i32 [[TMP105]]
; IC8-NEXT: [[TMP107:%.*]] = load i32, ptr [[TMP106]], align 4
; IC8-NEXT: [[TMP108:%.*]] = insertelement <4 x i32> [[TMP103]], i32 [[TMP107]], i64 1
; IC8-NEXT: br label %[[PRED_LOAD_CONTINUE33]]
; IC8: [[PRED_LOAD_CONTINUE33]]:
; IC8-NEXT: [[TMP109:%.*]] = phi <4 x i32> [ [[TMP103]], %[[PRED_LOAD_CONTINUE31]] ], [ [[TMP108]], %[[PRED_LOAD_IF32]] ]
; IC8-NEXT: [[TMP110:%.*]] = extractelement <4 x i1> [[TMP13]], i64 2
; IC8-NEXT: br i1 [[TMP110]], label %[[PRED_LOAD_IF34:.*]], label %[[PRED_LOAD_CONTINUE35:.*]]
; IC8: [[PRED_LOAD_IF34]]:
; IC8-NEXT: [[TMP111:%.*]] = extractelement <4 x i32> [[TMP21]], i64 2
; IC8-NEXT: [[TMP112:%.*]] = getelementptr inbounds i32, ptr [[A]], i32 [[TMP111]]
; IC8-NEXT: [[TMP113:%.*]] = load i32, ptr [[TMP112]], align 4
; IC8-NEXT: [[TMP114:%.*]] = insertelement <4 x i32> [[TMP109]], i32 [[TMP113]], i64 2
; IC8-NEXT: br label %[[PRED_LOAD_CONTINUE35]]
; IC8: [[PRED_LOAD_CONTINUE35]]:
; IC8-NEXT: [[TMP115:%.*]] = phi <4 x i32> [ [[TMP109]], %[[PRED_LOAD_CONTINUE33]] ], [ [[TMP114]], %[[PRED_LOAD_IF34]] ]
; IC8-NEXT: [[TMP116:%.*]] = extractelement <4 x i1> [[TMP13]], i64 3
; IC8-NEXT: br i1 [[TMP116]], label %[[PRED_LOAD_IF36:.*]], label %[[PRED_LOAD_CONTINUE37:.*]]
; IC8: [[PRED_LOAD_IF36]]:
; IC8-NEXT: [[TMP117:%.*]] = extractelement <4 x i32> [[TMP21]], i64 3
; IC8-NEXT: [[TMP118:%.*]] = getelementptr inbounds i32, ptr [[A]], i32 [[TMP117]]
; IC8-NEXT: [[TMP119:%.*]] = load i32, ptr [[TMP118]], align 4
; IC8-NEXT: [[TMP120:%.*]] = insertelement <4 x i32> [[TMP115]], i32 [[TMP119]], i64 3
; IC8-NEXT: br label %[[PRED_LOAD_CONTINUE37]]
; IC8: [[PRED_LOAD_CONTINUE37]]:
; IC8-NEXT: [[TMP121:%.*]] = phi <4 x i32> [ [[TMP115]], %[[PRED_LOAD_CONTINUE35]] ], [ [[TMP120]], %[[PRED_LOAD_IF36]] ]
; IC8-NEXT: [[TMP122:%.*]] = extractelement <4 x i1> [[TMP14]], i64 0
; IC8-NEXT: br i1 [[TMP122]], label %[[PRED_LOAD_IF38:.*]], label %[[PRED_LOAD_CONTINUE39:.*]]
; IC8: [[PRED_LOAD_IF38]]:
; IC8-NEXT: [[TMP123:%.*]] = extractelement <4 x i32> [[TMP22]], i64 0
; IC8-NEXT: [[TMP124:%.*]] = getelementptr inbounds i32, ptr [[A]], i32 [[TMP123]]
; IC8-NEXT: [[TMP125:%.*]] = load i32, ptr [[TMP124]], align 4
; IC8-NEXT: [[TMP126:%.*]] = insertelement <4 x i32> poison, i32 [[TMP125]], i64 0
; IC8-NEXT: br label %[[PRED_LOAD_CONTINUE39]]
; IC8: [[PRED_LOAD_CONTINUE39]]:
; IC8-NEXT: [[TMP127:%.*]] = phi <4 x i32> [ poison, %[[PRED_LOAD_CONTINUE37]] ], [ [[TMP126]], %[[PRED_LOAD_IF38]] ]
; IC8-NEXT: [[TMP128:%.*]] = extractelement <4 x i1> [[TMP14]], i64 1
; IC8-NEXT: br i1 [[TMP128]], label %[[PRED_LOAD_IF40:.*]], label %[[PRED_LOAD_CONTINUE41:.*]]
; IC8: [[PRED_LOAD_IF40]]:
; IC8-NEXT: [[TMP129:%.*]] = extractelement <4 x i32> [[TMP22]], i64 1
; IC8-NEXT: [[TMP130:%.*]] = getelementptr inbounds i32, ptr [[A]], i32 [[TMP129]]
; IC8-NEXT: [[TMP131:%.*]] = load i32, ptr [[TMP130]], align 4
; IC8-NEXT: [[TMP132:%.*]] = insertelement <4 x i32> [[TMP127]], i32 [[TMP131]], i64 1
; IC8-NEXT: br label %[[PRED_LOAD_CONTINUE41]]
; IC8: [[PRED_LOAD_CONTINUE41]]:
; IC8-NEXT: [[TMP133:%.*]] = phi <4 x i32> [ [[TMP127]], %[[PRED_LOAD_CONTINUE39]] ], [ [[TMP132]], %[[PRED_LOAD_IF40]] ]
; IC8-NEXT: [[TMP134:%.*]] = extractelement <4 x i1> [[TMP14]], i64 2
; IC8-NEXT: br i1 [[TMP134]], label %[[PRED_LOAD_IF42:.*]], label %[[PRED_LOAD_CONTINUE43:.*]]
; IC8: [[PRED_LOAD_IF42]]:
; IC8-NEXT: [[TMP135:%.*]] = extractelement <4 x i32> [[TMP22]], i64 2
; IC8-NEXT: [[TMP136:%.*]] = getelementptr inbounds i32, ptr [[A]], i32 [[TMP135]]
; IC8-NEXT: [[TMP137:%.*]] = load i32, ptr [[TMP136]], align 4
; IC8-NEXT: [[TMP138:%.*]] = insertelement <4 x i32> [[TMP133]], i32 [[TMP137]], i64 2
; IC8-NEXT: br label %[[PRED_LOAD_CONTINUE43]]
; IC8: [[PRED_LOAD_CONTINUE43]]:
; IC8-NEXT: [[TMP139:%.*]] = phi <4 x i32> [ [[TMP133]], %[[PRED_LOAD_CONTINUE41]] ], [ [[TMP138]], %[[PRED_LOAD_IF42]] ]
; IC8-NEXT: [[TMP140:%.*]] = extractelement <4 x i1> [[TMP14]], i64 3
; IC8-NEXT: br i1 [[TMP140]], label %[[PRED_LOAD_IF44:.*]], label %[[PRED_LOAD_CONTINUE45:.*]]
; IC8: [[PRED_LOAD_IF44]]:
; IC8-NEXT: [[TMP141:%.*]] = extractelement <4 x i32> [[TMP22]], i64 3
; IC8-NEXT: [[TMP142:%.*]] = getelementptr inbounds i32, ptr [[A]], i32 [[TMP141]]
; IC8-NEXT: [[TMP143:%.*]] = load i32, ptr [[TMP142]], align 4
; IC8-NEXT: [[TMP144:%.*]] = insertelement <4 x i32> [[TMP139]], i32 [[TMP143]], i64 3
; IC8-NEXT: br label %[[PRED_LOAD_CONTINUE45]]
; IC8: [[PRED_LOAD_CONTINUE45]]:
; IC8-NEXT: [[TMP145:%.*]] = phi <4 x i32> [ [[TMP139]], %[[PRED_LOAD_CONTINUE43]] ], [ [[TMP144]], %[[PRED_LOAD_IF44]] ]
; IC8-NEXT: [[TMP146:%.*]] = extractelement <4 x i1> [[TMP15]], i64 0
; IC8-NEXT: br i1 [[TMP146]], label %[[PRED_LOAD_IF46:.*]], label %[[PRED_LOAD_CONTINUE47:.*]]
; IC8: [[PRED_LOAD_IF46]]:
; IC8-NEXT: [[TMP147:%.*]] = extractelement <4 x i32> [[TMP23]], i64 0
; IC8-NEXT: [[TMP148:%.*]] = getelementptr inbounds i32, ptr [[A]], i32 [[TMP147]]
; IC8-NEXT: [[TMP149:%.*]] = load i32, ptr [[TMP148]], align 4
; IC8-NEXT: [[TMP150:%.*]] = insertelement <4 x i32> poison, i32 [[TMP149]], i64 0
; IC8-NEXT: br label %[[PRED_LOAD_CONTINUE47]]
; IC8: [[PRED_LOAD_CONTINUE47]]:
; IC8-NEXT: [[TMP151:%.*]] = phi <4 x i32> [ poison, %[[PRED_LOAD_CONTINUE45]] ], [ [[TMP150]], %[[PRED_LOAD_IF46]] ]
; IC8-NEXT: [[TMP152:%.*]] = extractelement <4 x i1> [[TMP15]], i64 1
; IC8-NEXT: br i1 [[TMP152]], label %[[PRED_LOAD_IF48:.*]], label %[[PRED_LOAD_CONTINUE49:.*]]
; IC8: [[PRED_LOAD_IF48]]:
; IC8-NEXT: [[TMP153:%.*]] = extractelement <4 x i32> [[TMP23]], i64 1
; IC8-NEXT: [[TMP154:%.*]] = getelementptr inbounds i32, ptr [[A]], i32 [[TMP153]]
; IC8-NEXT: [[TMP155:%.*]] = load i32, ptr [[TMP154]], align 4
; IC8-NEXT: [[TMP156:%.*]] = insertelement <4 x i32> [[TMP151]], i32 [[TMP155]], i64 1
; IC8-NEXT: br label %[[PRED_LOAD_CONTINUE49]]
; IC8: [[PRED_LOAD_CONTINUE49]]:
; IC8-NEXT: [[TMP157:%.*]] = phi <4 x i32> [ [[TMP151]], %[[PRED_LOAD_CONTINUE47]] ], [ [[TMP156]], %[[PRED_LOAD_IF48]] ]
; IC8-NEXT: [[TMP158:%.*]] = extractelement <4 x i1> [[TMP15]], i64 2
; IC8-NEXT: br i1 [[TMP158]], label %[[PRED_LOAD_IF50:.*]], label %[[PRED_LOAD_CONTINUE51:.*]]
; IC8: [[PRED_LOAD_IF50]]:
; IC8-NEXT: [[TMP159:%.*]] = extractelement <4 x i32> [[TMP23]], i64 2
; IC8-NEXT: [[TMP160:%.*]] = getelementptr inbounds i32, ptr [[A]], i32 [[TMP159]]
; IC8-NEXT: [[TMP161:%.*]] = load i32, ptr [[TMP160]], align 4
; IC8-NEXT: [[TMP162:%.*]] = insertelement <4 x i32> [[TMP157]], i32 [[TMP161]], i64 2
; IC8-NEXT: br label %[[PRED_LOAD_CONTINUE51]]
; IC8: [[PRED_LOAD_CONTINUE51]]:
; IC8-NEXT: [[TMP163:%.*]] = phi <4 x i32> [ [[TMP157]], %[[PRED_LOAD_CONTINUE49]] ], [ [[TMP162]], %[[PRED_LOAD_IF50]] ]
; IC8-NEXT: [[TMP164:%.*]] = extractelement <4 x i1> [[TMP15]], i64 3
; IC8-NEXT: br i1 [[TMP164]], label %[[PRED_LOAD_IF52:.*]], label %[[PRED_LOAD_CONTINUE53:.*]]
; IC8: [[PRED_LOAD_IF52]]:
; IC8-NEXT: [[TMP165:%.*]] = extractelement <4 x i32> [[TMP23]], i64 3
; IC8-NEXT: [[TMP166:%.*]] = getelementptr inbounds i32, ptr [[A]], i32 [[TMP165]]
; IC8-NEXT: [[TMP167:%.*]] = load i32, ptr [[TMP166]], align 4
; IC8-NEXT: [[TMP168:%.*]] = insertelement <4 x i32> [[TMP163]], i32 [[TMP167]], i64 3
; IC8-NEXT: br label %[[PRED_LOAD_CONTINUE53]]
; IC8: [[PRED_LOAD_CONTINUE53]]:
; IC8-NEXT: [[TMP169:%.*]] = phi <4 x i32> [ [[TMP163]], %[[PRED_LOAD_CONTINUE51]] ], [ [[TMP168]], %[[PRED_LOAD_IF52]] ]
; IC8-NEXT: [[TMP170:%.*]] = extractelement <4 x i1> [[TMP16]], i64 0
; IC8-NEXT: br i1 [[TMP170]], label %[[PRED_LOAD_IF54:.*]], label %[[PRED_LOAD_CONTINUE55:.*]]
; IC8: [[PRED_LOAD_IF54]]:
; IC8-NEXT: [[TMP171:%.*]] = extractelement <4 x i32> [[TMP24]], i64 0
; IC8-NEXT: [[TMP172:%.*]] = getelementptr inbounds i32, ptr [[A]], i32 [[TMP171]]
; IC8-NEXT: [[TMP173:%.*]] = load i32, ptr [[TMP172]], align 4
; IC8-NEXT: [[TMP174:%.*]] = insertelement <4 x i32> poison, i32 [[TMP173]], i64 0
; IC8-NEXT: br label %[[PRED_LOAD_CONTINUE55]]
; IC8: [[PRED_LOAD_CONTINUE55]]:
; IC8-NEXT: [[TMP175:%.*]] = phi <4 x i32> [ poison, %[[PRED_LOAD_CONTINUE53]] ], [ [[TMP174]], %[[PRED_LOAD_IF54]] ]
; IC8-NEXT: [[TMP176:%.*]] = extractelement <4 x i1> [[TMP16]], i64 1
; IC8-NEXT: br i1 [[TMP176]], label %[[PRED_LOAD_IF56:.*]], label %[[PRED_LOAD_CONTINUE57:.*]]
; IC8: [[PRED_LOAD_IF56]]:
; IC8-NEXT: [[TMP177:%.*]] = extractelement <4 x i32> [[TMP24]], i64 1
; IC8-NEXT: [[TMP178:%.*]] = getelementptr inbounds i32, ptr [[A]], i32 [[TMP177]]
; IC8-NEXT: [[TMP179:%.*]] = load i32, ptr [[TMP178]], align 4
; IC8-NEXT: [[TMP180:%.*]] = insertelement <4 x i32> [[TMP175]], i32 [[TMP179]], i64 1
; IC8-NEXT: br label %[[PRED_LOAD_CONTINUE57]]
; IC8: [[PRED_LOAD_CONTINUE57]]:
; IC8-NEXT: [[TMP181:%.*]] = phi <4 x i32> [ [[TMP175]], %[[PRED_LOAD_CONTINUE55]] ], [ [[TMP180]], %[[PRED_LOAD_IF56]] ]
; IC8-NEXT: [[TMP182:%.*]] = extractelement <4 x i1> [[TMP16]], i64 2
; IC8-NEXT: br i1 [[TMP182]], label %[[PRED_LOAD_IF58:.*]], label %[[PRED_LOAD_CONTINUE59:.*]]
; IC8: [[PRED_LOAD_IF58]]:
; IC8-NEXT: [[TMP183:%.*]] = extractelement <4 x i32> [[TMP24]], i64 2
; IC8-NEXT: [[TMP184:%.*]] = getelementptr inbounds i32, ptr [[A]], i32 [[TMP183]]
; IC8-NEXT: [[TMP185:%.*]] = load i32, ptr [[TMP184]], align 4
; IC8-NEXT: [[TMP186:%.*]] = insertelement <4 x i32> [[TMP181]], i32 [[TMP185]], i64 2
; IC8-NEXT: br label %[[PRED_LOAD_CONTINUE59]]
; IC8: [[PRED_LOAD_CONTINUE59]]:
; IC8-NEXT: [[TMP187:%.*]] = phi <4 x i32> [ [[TMP181]], %[[PRED_LOAD_CONTINUE57]] ], [ [[TMP186]], %[[PRED_LOAD_IF58]] ]
; IC8-NEXT: [[TMP188:%.*]] = extractelement <4 x i1> [[TMP16]], i64 3
; IC8-NEXT: br i1 [[TMP188]], label %[[PRED_LOAD_IF60:.*]], label %[[PRED_LOAD_CONTINUE61:.*]]
; IC8: [[PRED_LOAD_IF60]]:
; IC8-NEXT: [[TMP189:%.*]] = extractelement <4 x i32> [[TMP24]], i64 3
; IC8-NEXT: [[TMP190:%.*]] = getelementptr inbounds i32, ptr [[A]], i32 [[TMP189]]
; IC8-NEXT: [[TMP191:%.*]] = load i32, ptr [[TMP190]], align 4
; IC8-NEXT: [[TMP192:%.*]] = insertelement <4 x i32> [[TMP187]], i32 [[TMP191]], i64 3
; IC8-NEXT: br label %[[PRED_LOAD_CONTINUE61]]
; IC8: [[PRED_LOAD_CONTINUE61]]:
; IC8-NEXT: [[TMP193:%.*]] = phi <4 x i32> [ [[TMP187]], %[[PRED_LOAD_CONTINUE59]] ], [ [[TMP192]], %[[PRED_LOAD_IF60]] ]
; IC8-NEXT: [[TMP194:%.*]] = extractelement <4 x i1> [[TMP17]], i64 0
; IC8-NEXT: br i1 [[TMP194]], label %[[PRED_LOAD_IF62:.*]], label %[[PRED_LOAD_CONTINUE63:.*]]
; IC8: [[PRED_LOAD_IF62]]:
; IC8-NEXT: [[TMP195:%.*]] = extractelement <4 x i32> [[TMP25]], i64 0
; IC8-NEXT: [[TMP196:%.*]] = getelementptr inbounds i32, ptr [[A]], i32 [[TMP195]]
; IC8-NEXT: [[TMP197:%.*]] = load i32, ptr [[TMP196]], align 4
; IC8-NEXT: [[TMP198:%.*]] = insertelement <4 x i32> poison, i32 [[TMP197]], i64 0
; IC8-NEXT: br label %[[PRED_LOAD_CONTINUE63]]
; IC8: [[PRED_LOAD_CONTINUE63]]:
; IC8-NEXT: [[TMP199:%.*]] = phi <4 x i32> [ poison, %[[PRED_LOAD_CONTINUE61]] ], [ [[TMP198]], %[[PRED_LOAD_IF62]] ]
; IC8-NEXT: [[TMP200:%.*]] = extractelement <4 x i1> [[TMP17]], i64 1
; IC8-NEXT: br i1 [[TMP200]], label %[[PRED_LOAD_IF64:.*]], label %[[PRED_LOAD_CONTINUE65:.*]]
; IC8: [[PRED_LOAD_IF64]]:
; IC8-NEXT: [[TMP201:%.*]] = extractelement <4 x i32> [[TMP25]], i64 1
; IC8-NEXT: [[TMP202:%.*]] = getelementptr inbounds i32, ptr [[A]], i32 [[TMP201]]
; IC8-NEXT: [[TMP203:%.*]] = load i32, ptr [[TMP202]], align 4
; IC8-NEXT: [[TMP204:%.*]] = insertelement <4 x i32> [[TMP199]], i32 [[TMP203]], i64 1
; IC8-NEXT: br label %[[PRED_LOAD_CONTINUE65]]
; IC8: [[PRED_LOAD_CONTINUE65]]:
; IC8-NEXT: [[TMP205:%.*]] = phi <4 x i32> [ [[TMP199]], %[[PRED_LOAD_CONTINUE63]] ], [ [[TMP204]], %[[PRED_LOAD_IF64]] ]
; IC8-NEXT: [[TMP206:%.*]] = extractelement <4 x i1> [[TMP17]], i64 2
; IC8-NEXT: br i1 [[TMP206]], label %[[PRED_LOAD_IF66:.*]], label %[[PRED_LOAD_CONTINUE67:.*]]
; IC8: [[PRED_LOAD_IF66]]:
; IC8-NEXT: [[TMP207:%.*]] = extractelement <4 x i32> [[TMP25]], i64 2
; IC8-NEXT: [[TMP208:%.*]] = getelementptr inbounds i32, ptr [[A]], i32 [[TMP207]]
; IC8-NEXT: [[TMP209:%.*]] = load i32, ptr [[TMP208]], align 4
; IC8-NEXT: [[TMP210:%.*]] = insertelement <4 x i32> [[TMP205]], i32 [[TMP209]], i64 2
; IC8-NEXT: br label %[[PRED_LOAD_CONTINUE67]]
; IC8: [[PRED_LOAD_CONTINUE67]]:
; IC8-NEXT: [[TMP211:%.*]] = phi <4 x i32> [ [[TMP205]], %[[PRED_LOAD_CONTINUE65]] ], [ [[TMP210]], %[[PRED_LOAD_IF66]] ]
; IC8-NEXT: [[TMP212:%.*]] = extractelement <4 x i1> [[TMP17]], i64 3
; IC8-NEXT: br i1 [[TMP212]], label %[[PRED_LOAD_IF68:.*]], label %[[PRED_LOAD_CONTINUE69]]
; IC8: [[PRED_LOAD_IF68]]:
; IC8-NEXT: [[TMP213:%.*]] = extractelement <4 x i32> [[TMP25]], i64 3
; IC8-NEXT: [[TMP214:%.*]] = getelementptr inbounds i32, ptr [[A]], i32 [[TMP213]]
; IC8-NEXT: [[TMP215:%.*]] = load i32, ptr [[TMP214]], align 4
; IC8-NEXT: [[TMP216:%.*]] = insertelement <4 x i32> [[TMP211]], i32 [[TMP215]], i64 3
; IC8-NEXT: br label %[[PRED_LOAD_CONTINUE69]]
; IC8: [[PRED_LOAD_CONTINUE69]]:
; IC8-NEXT: [[TMP217:%.*]] = phi <4 x i32> [ [[TMP211]], %[[PRED_LOAD_CONTINUE67]] ], [ [[TMP216]], %[[PRED_LOAD_IF68]] ]
; IC8-NEXT: [[PREDPHI:%.*]] = select <4 x i1> [[TMP10]], <4 x i32> [[TMP49]], <4 x i32> zeroinitializer
; IC8-NEXT: [[PREDPHI70:%.*]] = select <4 x i1> [[TMP11]], <4 x i32> [[TMP73]], <4 x i32> zeroinitializer
; IC8-NEXT: [[PREDPHI71:%.*]] = select <4 x i1> [[TMP12]], <4 x i32> [[TMP97]], <4 x i32> zeroinitializer
; IC8-NEXT: [[PREDPHI72:%.*]] = select <4 x i1> [[TMP13]], <4 x i32> [[TMP121]], <4 x i32> zeroinitializer
; IC8-NEXT: [[PREDPHI73:%.*]] = select <4 x i1> [[TMP14]], <4 x i32> [[TMP145]], <4 x i32> zeroinitializer
; IC8-NEXT: [[PREDPHI74:%.*]] = select <4 x i1> [[TMP15]], <4 x i32> [[TMP169]], <4 x i32> zeroinitializer
; IC8-NEXT: [[PREDPHI75:%.*]] = select <4 x i1> [[TMP16]], <4 x i32> [[TMP193]], <4 x i32> zeroinitializer
; IC8-NEXT: [[PREDPHI76:%.*]] = select <4 x i1> [[TMP17]], <4 x i32> [[TMP217]], <4 x i32> zeroinitializer
; IC8-NEXT: [[TMP218]] = add <4 x i32> [[VEC_PHI]], [[PREDPHI]]
; IC8-NEXT: [[TMP219]] = add <4 x i32> [[VEC_PHI1]], [[PREDPHI70]]
; IC8-NEXT: [[TMP220]] = add <4 x i32> [[VEC_PHI2]], [[PREDPHI71]]
; IC8-NEXT: [[TMP221]] = add <4 x i32> [[VEC_PHI3]], [[PREDPHI72]]
; IC8-NEXT: [[TMP222]] = add <4 x i32> [[VEC_PHI4]], [[PREDPHI73]]
; IC8-NEXT: [[TMP223]] = add <4 x i32> [[VEC_PHI5]], [[PREDPHI74]]
; IC8-NEXT: [[TMP224]] = add <4 x i32> [[VEC_PHI6]], [[PREDPHI75]]
; IC8-NEXT: [[TMP225]] = add <4 x i32> [[VEC_PHI7]], [[PREDPHI76]]
; IC8-NEXT: [[INDEX_NEXT]] = add nuw i32 [[INDEX]], 32
; IC8-NEXT: [[VEC_IND_NEXT]] = add <4 x i32> [[STEP_ADD_7]], splat (i32 4)
; IC8-NEXT: [[TMP226:%.*]] = icmp eq i32 [[INDEX_NEXT]], [[N_VEC]]
; IC8-NEXT: br i1 [[TMP226]], label %[[MIDDLE_BLOCK:.*]], label %[[VECTOR_BODY]], !llvm.loop [[LOOP0:![0-9]+]]
; IC8: [[MIDDLE_BLOCK]]:
; IC8-NEXT: [[BIN_RDX:%.*]] = add <4 x i32> [[TMP219]], [[TMP218]]
; IC8-NEXT: [[BIN_RDX77:%.*]] = add <4 x i32> [[TMP220]], [[BIN_RDX]]
; IC8-NEXT: [[BIN_RDX78:%.*]] = add <4 x i32> [[TMP221]], [[BIN_RDX77]]
; IC8-NEXT: [[BIN_RDX79:%.*]] = add <4 x i32> [[TMP222]], [[BIN_RDX78]]
; IC8-NEXT: [[BIN_RDX80:%.*]] = add <4 x i32> [[TMP223]], [[BIN_RDX79]]
; IC8-NEXT: [[BIN_RDX81:%.*]] = add <4 x i32> [[TMP224]], [[BIN_RDX80]]
; IC8-NEXT: [[BIN_RDX82:%.*]] = add <4 x i32> [[TMP225]], [[BIN_RDX81]]
; IC8-NEXT: [[TMP227:%.*]] = call i32 @llvm.vector.reduce.add.v4i32(<4 x i32> [[BIN_RDX82]])
; IC8-NEXT: [[CMP_N:%.*]] = icmp eq i32 [[N]], [[N_VEC]]
; IC8-NEXT: br i1 [[CMP_N]], label %[[EXIT:.*]], label %[[SCALAR_PH]]
; IC8: [[SCALAR_PH]]:
; IC8-NEXT: [[BC_RESUME_VAL:%.*]] = phi i32 [ [[N_VEC]], %[[MIDDLE_BLOCK]] ], [ 0, %[[ENTRY]] ], [ 0, %[[VECTOR_SCEVCHECK]] ]
; IC8-NEXT: [[BC_MERGE_RDX:%.*]] = phi i32 [ [[TMP227]], %[[MIDDLE_BLOCK]] ], [ 0, %[[ENTRY]] ], [ 0, %[[VECTOR_SCEVCHECK]] ]
; IC8-NEXT: br label %[[LOOP:.*]]
; IC8: [[LOOP]]:
; IC8-NEXT: [[IV:%.*]] = phi i32 [ [[BC_RESUME_VAL]], %[[SCALAR_PH]] ], [ [[IV_NEXT:%.*]], %[[LATCH:.*]] ]
; IC8-NEXT: [[SUM:%.*]] = phi i32 [ [[BC_MERGE_RDX]], %[[SCALAR_PH]] ], [ [[SUM_NEXT:%.*]], %[[LATCH]] ]
; IC8-NEXT: [[AND_IV:%.*]] = and i32 [[IV]], 1
; IC8-NEXT: [[CMP:%.*]] = icmp eq i32 [[AND_IV]], 0
; IC8-NEXT: br i1 [[CMP]], label %[[THEN:.*]], label %[[LATCH]]
; IC8: [[THEN]]:
; IC8-NEXT: [[CLAMPED:%.*]] = urem i32 [[IV]], 128
; IC8-NEXT: [[GEP_A:%.*]] = getelementptr inbounds i32, ptr [[A]], i32 [[CLAMPED]]
; IC8-NEXT: [[LV:%.*]] = load i32, ptr [[GEP_A]], align 4
; IC8-NEXT: br label %[[LATCH]]
; IC8: [[LATCH]]:
; IC8-NEXT: [[V:%.*]] = phi i32 [ [[LV]], %[[THEN]] ], [ 0, %[[LOOP]] ]
; IC8-NEXT: [[SUM_NEXT]] = add i32 [[SUM]], [[V]]
; IC8-NEXT: [[IV_NEXT]] = add i32 [[IV]], 1
; IC8-NEXT: [[EC:%.*]] = icmp eq i32 [[IV_NEXT]], [[N]]
; IC8-NEXT: br i1 [[EC]], label %[[EXIT]], label %[[LOOP]], !llvm.loop [[LOOP3:![0-9]+]]
; IC8: [[EXIT]]:
; IC8-NEXT: [[SUM_NEXT_LCSSA:%.*]] = phi i32 [ [[SUM_NEXT]], %[[LATCH]] ], [ [[TMP227]], %[[MIDDLE_BLOCK]] ]
; IC8-NEXT: ret i32 [[SUM_NEXT_LCSSA]]
;
entry:
br label %loop
loop:
%iv = phi i32 [ 0, %entry ], [ %iv.next, %latch ]
%sum = phi i32 [ 0, %entry ], [ %sum.next, %latch ]
%and.iv = and i32 %iv, 1
%cmp = icmp eq i32 %and.iv, 0
br i1 %cmp, label %then, label %latch
then:
%clamped = urem i32 %iv, 128
%gep.A = getelementptr inbounds i32, ptr %A, i32 %clamped
%lv = load i32, ptr %gep.A
br label %latch
latch:
%v = phi i32 [ %lv, %then ], [ 0, %loop ]
%sum.next = add i32 %sum, %v
%iv.next = add i32 %iv, 1
%ec = icmp eq i32 %iv.next, %n
br i1 %ec, label %exit, label %loop
exit:
ret i32 %sum.next
}
; A forced interleave count that would step past the clamp window must be capped
; down to a power-of-two divisor of Bound / VF. With VF=4 and Bound=16 the cap is
; 16 / 4 = 4.
; TODO: This is the desired behavior. Currently the interleave count is not
; capped: with -force-vector-interleave=8 the full 8-part body is emitted
; (step.add.7, index += 32, 32 pred.load.if blocks). It is only harmless
; because the contradictory min-iters (N >= 32) and SCEV wrap (N <= 16) checks
; make that body unreachable.
define i32 @clamped_load_predicated_ic_capped(ptr %A, i32 %n) {
; CHECK-LABEL: define i32 @clamped_load_predicated_ic_capped(
; CHECK-SAME: ptr [[A:%.*]], i32 [[N:%.*]]) {
; CHECK-NEXT: [[ENTRY:.*]]:
; CHECK-NEXT: [[MIN_ITERS_CHECK:%.*]] = icmp ult i32 [[N]], 4
; CHECK-NEXT: br i1 [[MIN_ITERS_CHECK]], label %[[SCALAR_PH:.*]], label %[[VECTOR_SCEVCHECK:.*]]
; CHECK: [[VECTOR_SCEVCHECK]]:
; CHECK-NEXT: [[TMP0:%.*]] = add i32 [[N]], -1
; CHECK-NEXT: [[TMP1:%.*]] = icmp ugt i32 [[TMP0]], 15
; CHECK-NEXT: br i1 [[TMP1]], label %[[SCALAR_PH]], label %[[VECTOR_PH:.*]]
; CHECK: [[VECTOR_PH]]:
; CHECK-NEXT: [[N_MOD_VF:%.*]] = urem i32 [[N]], 4
; CHECK-NEXT: [[N_VEC:%.*]] = sub i32 [[N]], [[N_MOD_VF]]
; CHECK-NEXT: br label %[[VECTOR_BODY:.*]]
; CHECK: [[VECTOR_BODY]]:
; CHECK-NEXT: [[INDEX:%.*]] = phi i32 [ 0, %[[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], %[[PRED_LOAD_CONTINUE6:.*]] ]
; CHECK-NEXT: [[VEC_IND:%.*]] = phi <4 x i32> [ <i32 0, i32 1, i32 2, i32 3>, %[[VECTOR_PH]] ], [ [[VEC_IND_NEXT:%.*]], %[[PRED_LOAD_CONTINUE6]] ]
; CHECK-NEXT: [[VEC_PHI:%.*]] = phi <4 x i32> [ zeroinitializer, %[[VECTOR_PH]] ], [ [[TMP29:%.*]], %[[PRED_LOAD_CONTINUE6]] ]
; CHECK-NEXT: [[TMP2:%.*]] = and <4 x i32> [[VEC_IND]], splat (i32 1)
; CHECK-NEXT: [[TMP3:%.*]] = icmp eq <4 x i32> [[TMP2]], zeroinitializer
; CHECK-NEXT: [[TMP4:%.*]] = urem <4 x i32> [[VEC_IND]], splat (i32 16)
; CHECK-NEXT: [[TMP5:%.*]] = extractelement <4 x i1> [[TMP3]], i64 0
; CHECK-NEXT: br i1 [[TMP5]], label %[[PRED_LOAD_IF:.*]], label %[[PRED_LOAD_CONTINUE:.*]]
; CHECK: [[PRED_LOAD_IF]]:
; CHECK-NEXT: [[TMP6:%.*]] = extractelement <4 x i32> [[TMP4]], i64 0
; CHECK-NEXT: [[TMP7:%.*]] = getelementptr inbounds i32, ptr [[A]], i32 [[TMP6]]
; CHECK-NEXT: [[TMP8:%.*]] = load i32, ptr [[TMP7]], align 4
; CHECK-NEXT: [[TMP9:%.*]] = insertelement <4 x i32> poison, i32 [[TMP8]], i64 0
; CHECK-NEXT: br label %[[PRED_LOAD_CONTINUE]]
; CHECK: [[PRED_LOAD_CONTINUE]]:
; CHECK-NEXT: [[TMP10:%.*]] = phi <4 x i32> [ poison, %[[VECTOR_BODY]] ], [ [[TMP9]], %[[PRED_LOAD_IF]] ]
; CHECK-NEXT: [[TMP11:%.*]] = extractelement <4 x i1> [[TMP3]], i64 1
; CHECK-NEXT: br i1 [[TMP11]], label %[[PRED_LOAD_IF1:.*]], label %[[PRED_LOAD_CONTINUE2:.*]]
; CHECK: [[PRED_LOAD_IF1]]:
; CHECK-NEXT: [[TMP12:%.*]] = extractelement <4 x i32> [[TMP4]], i64 1
; CHECK-NEXT: [[TMP13:%.*]] = getelementptr inbounds i32, ptr [[A]], i32 [[TMP12]]
; CHECK-NEXT: [[TMP14:%.*]] = load i32, ptr [[TMP13]], align 4
; CHECK-NEXT: [[TMP15:%.*]] = insertelement <4 x i32> [[TMP10]], i32 [[TMP14]], i64 1
; CHECK-NEXT: br label %[[PRED_LOAD_CONTINUE2]]
; CHECK: [[PRED_LOAD_CONTINUE2]]:
; CHECK-NEXT: [[TMP16:%.*]] = phi <4 x i32> [ [[TMP10]], %[[PRED_LOAD_CONTINUE]] ], [ [[TMP15]], %[[PRED_LOAD_IF1]] ]
; CHECK-NEXT: [[TMP17:%.*]] = extractelement <4 x i1> [[TMP3]], i64 2
; CHECK-NEXT: br i1 [[TMP17]], label %[[PRED_LOAD_IF3:.*]], label %[[PRED_LOAD_CONTINUE4:.*]]
; CHECK: [[PRED_LOAD_IF3]]:
; CHECK-NEXT: [[TMP18:%.*]] = extractelement <4 x i32> [[TMP4]], i64 2
; CHECK-NEXT: [[TMP19:%.*]] = getelementptr inbounds i32, ptr [[A]], i32 [[TMP18]]
; CHECK-NEXT: [[TMP20:%.*]] = load i32, ptr [[TMP19]], align 4
; CHECK-NEXT: [[TMP21:%.*]] = insertelement <4 x i32> [[TMP16]], i32 [[TMP20]], i64 2
; CHECK-NEXT: br label %[[PRED_LOAD_CONTINUE4]]
; CHECK: [[PRED_LOAD_CONTINUE4]]:
; CHECK-NEXT: [[TMP22:%.*]] = phi <4 x i32> [ [[TMP16]], %[[PRED_LOAD_CONTINUE2]] ], [ [[TMP21]], %[[PRED_LOAD_IF3]] ]
; CHECK-NEXT: [[TMP23:%.*]] = extractelement <4 x i1> [[TMP3]], i64 3
; CHECK-NEXT: br i1 [[TMP23]], label %[[PRED_LOAD_IF5:.*]], label %[[PRED_LOAD_CONTINUE6]]
; CHECK: [[PRED_LOAD_IF5]]:
; CHECK-NEXT: [[TMP24:%.*]] = extractelement <4 x i32> [[TMP4]], i64 3
; CHECK-NEXT: [[TMP25:%.*]] = getelementptr inbounds i32, ptr [[A]], i32 [[TMP24]]
; CHECK-NEXT: [[TMP26:%.*]] = load i32, ptr [[TMP25]], align 4
; CHECK-NEXT: [[TMP27:%.*]] = insertelement <4 x i32> [[TMP22]], i32 [[TMP26]], i64 3
; CHECK-NEXT: br label %[[PRED_LOAD_CONTINUE6]]
; CHECK: [[PRED_LOAD_CONTINUE6]]:
; CHECK-NEXT: [[TMP28:%.*]] = phi <4 x i32> [ [[TMP22]], %[[PRED_LOAD_CONTINUE4]] ], [ [[TMP27]], %[[PRED_LOAD_IF5]] ]
; CHECK-NEXT: [[PREDPHI:%.*]] = select <4 x i1> [[TMP3]], <4 x i32> [[TMP28]], <4 x i32> zeroinitializer
; CHECK-NEXT: [[TMP29]] = add <4 x i32> [[VEC_PHI]], [[PREDPHI]]
; CHECK-NEXT: [[INDEX_NEXT]] = add nuw i32 [[INDEX]], 4
; CHECK-NEXT: [[VEC_IND_NEXT]] = add <4 x i32> [[VEC_IND]], splat (i32 4)
; CHECK-NEXT: [[TMP30:%.*]] = icmp eq i32 [[INDEX_NEXT]], [[N_VEC]]
; CHECK-NEXT: br i1 [[TMP30]], label %[[MIDDLE_BLOCK:.*]], label %[[VECTOR_BODY]], !llvm.loop [[LOOP4:![0-9]+]]
; CHECK: [[MIDDLE_BLOCK]]:
; CHECK-NEXT: [[TMP31:%.*]] = call i32 @llvm.vector.reduce.add.v4i32(<4 x i32> [[TMP29]])
; CHECK-NEXT: [[CMP_N:%.*]] = icmp eq i32 [[N]], [[N_VEC]]
; CHECK-NEXT: br i1 [[CMP_N]], label %[[EXIT:.*]], label %[[SCALAR_PH]]
; CHECK: [[SCALAR_PH]]:
; CHECK-NEXT: [[BC_RESUME_VAL:%.*]] = phi i32 [ [[N_VEC]], %[[MIDDLE_BLOCK]] ], [ 0, %[[ENTRY]] ], [ 0, %[[VECTOR_SCEVCHECK]] ]
; CHECK-NEXT: [[BC_MERGE_RDX:%.*]] = phi i32 [ [[TMP31]], %[[MIDDLE_BLOCK]] ], [ 0, %[[ENTRY]] ], [ 0, %[[VECTOR_SCEVCHECK]] ]
; CHECK-NEXT: br label %[[LOOP:.*]]
; CHECK: [[LOOP]]:
; CHECK-NEXT: [[IV:%.*]] = phi i32 [ [[BC_RESUME_VAL]], %[[SCALAR_PH]] ], [ [[IV_NEXT:%.*]], %[[LATCH:.*]] ]
; CHECK-NEXT: [[SUM:%.*]] = phi i32 [ [[BC_MERGE_RDX]], %[[SCALAR_PH]] ], [ [[SUM_NEXT:%.*]], %[[LATCH]] ]
; CHECK-NEXT: [[AND_IV:%.*]] = and i32 [[IV]], 1
; CHECK-NEXT: [[CMP:%.*]] = icmp eq i32 [[AND_IV]], 0
; CHECK-NEXT: br i1 [[CMP]], label %[[THEN:.*]], label %[[LATCH]]
; CHECK: [[THEN]]:
; CHECK-NEXT: [[CLAMPED:%.*]] = urem i32 [[IV]], 16
; CHECK-NEXT: [[GEP_A:%.*]] = getelementptr inbounds i32, ptr [[A]], i32 [[CLAMPED]]
; CHECK-NEXT: [[LV:%.*]] = load i32, ptr [[GEP_A]], align 4
; CHECK-NEXT: br label %[[LATCH]]
; CHECK: [[LATCH]]:
; CHECK-NEXT: [[V:%.*]] = phi i32 [ [[LV]], %[[THEN]] ], [ 0, %[[LOOP]] ]
; CHECK-NEXT: [[SUM_NEXT]] = add i32 [[SUM]], [[V]]
; CHECK-NEXT: [[IV_NEXT]] = add i32 [[IV]], 1
; CHECK-NEXT: [[EC:%.*]] = icmp eq i32 [[IV_NEXT]], [[N]]
; CHECK-NEXT: br i1 [[EC]], label %[[EXIT]], label %[[LOOP]], !llvm.loop [[LOOP5:![0-9]+]]
; CHECK: [[EXIT]]:
; CHECK-NEXT: [[SUM_NEXT_LCSSA:%.*]] = phi i32 [ [[SUM_NEXT]], %[[LATCH]] ], [ [[TMP31]], %[[MIDDLE_BLOCK]] ]
; CHECK-NEXT: ret i32 [[SUM_NEXT_LCSSA]]
;
; IC4-LABEL: define i32 @clamped_load_predicated_ic_capped(
; IC4-SAME: ptr [[A:%.*]], i32 [[N:%.*]]) {
; IC4-NEXT: [[ENTRY:.*]]:
; IC4-NEXT: [[MIN_ITERS_CHECK:%.*]] = icmp ult i32 [[N]], 16
; IC4-NEXT: br i1 [[MIN_ITERS_CHECK]], label %[[SCALAR_PH:.*]], label %[[VECTOR_SCEVCHECK:.*]]
; IC4: [[VECTOR_SCEVCHECK]]:
; IC4-NEXT: [[TMP0:%.*]] = add i32 [[N]], -1
; IC4-NEXT: [[TMP1:%.*]] = icmp ugt i32 [[TMP0]], 15
; IC4-NEXT: br i1 [[TMP1]], label %[[SCALAR_PH]], label %[[VECTOR_PH:.*]]
; IC4: [[VECTOR_PH]]:
; IC4-NEXT: [[N_MOD_VF:%.*]] = urem i32 [[N]], 16
; IC4-NEXT: [[N_VEC:%.*]] = sub i32 [[N]], [[N_MOD_VF]]
; IC4-NEXT: br label %[[VECTOR_BODY:.*]]
; IC4: [[VECTOR_BODY]]:
; IC4-NEXT: [[INDEX:%.*]] = phi i32 [ 0, %[[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], %[[PRED_LOAD_CONTINUE33:.*]] ]
; IC4-NEXT: [[VEC_IND:%.*]] = phi <4 x i32> [ <i32 0, i32 1, i32 2, i32 3>, %[[VECTOR_PH]] ], [ [[VEC_IND_NEXT:%.*]], %[[PRED_LOAD_CONTINUE33]] ]
; IC4-NEXT: [[VEC_PHI:%.*]] = phi <4 x i32> [ zeroinitializer, %[[VECTOR_PH]] ], [ [[TMP110:%.*]], %[[PRED_LOAD_CONTINUE33]] ]
; IC4-NEXT: [[VEC_PHI1:%.*]] = phi <4 x i32> [ zeroinitializer, %[[VECTOR_PH]] ], [ [[TMP111:%.*]], %[[PRED_LOAD_CONTINUE33]] ]
; IC4-NEXT: [[VEC_PHI2:%.*]] = phi <4 x i32> [ zeroinitializer, %[[VECTOR_PH]] ], [ [[TMP112:%.*]], %[[PRED_LOAD_CONTINUE33]] ]
; IC4-NEXT: [[VEC_PHI3:%.*]] = phi <4 x i32> [ zeroinitializer, %[[VECTOR_PH]] ], [ [[TMP113:%.*]], %[[PRED_LOAD_CONTINUE33]] ]
; IC4-NEXT: [[STEP_ADD:%.*]] = add nuw <4 x i32> [[VEC_IND]], splat (i32 4)
; IC4-NEXT: [[STEP_ADD_2:%.*]] = add nuw <4 x i32> [[STEP_ADD]], splat (i32 4)
; IC4-NEXT: [[STEP_ADD_3:%.*]] = add nuw <4 x i32> [[STEP_ADD_2]], splat (i32 4)
; IC4-NEXT: [[TMP2:%.*]] = and <4 x i32> [[VEC_IND]], splat (i32 1)
; IC4-NEXT: [[TMP3:%.*]] = and <4 x i32> [[STEP_ADD]], splat (i32 1)
; IC4-NEXT: [[TMP4:%.*]] = and <4 x i32> [[STEP_ADD_2]], splat (i32 1)
; IC4-NEXT: [[TMP5:%.*]] = and <4 x i32> [[STEP_ADD_3]], splat (i32 1)
; IC4-NEXT: [[TMP6:%.*]] = icmp eq <4 x i32> [[TMP2]], zeroinitializer
; IC4-NEXT: [[TMP7:%.*]] = icmp eq <4 x i32> [[TMP3]], zeroinitializer
; IC4-NEXT: [[TMP8:%.*]] = icmp eq <4 x i32> [[TMP4]], zeroinitializer
; IC4-NEXT: [[TMP9:%.*]] = icmp eq <4 x i32> [[TMP5]], zeroinitializer
; IC4-NEXT: [[TMP10:%.*]] = urem <4 x i32> [[VEC_IND]], splat (i32 16)
; IC4-NEXT: [[TMP11:%.*]] = urem <4 x i32> [[STEP_ADD]], splat (i32 16)
; IC4-NEXT: [[TMP12:%.*]] = urem <4 x i32> [[STEP_ADD_2]], splat (i32 16)
; IC4-NEXT: [[TMP13:%.*]] = urem <4 x i32> [[STEP_ADD_3]], splat (i32 16)
; IC4-NEXT: [[TMP14:%.*]] = extractelement <4 x i1> [[TMP6]], i64 0
; IC4-NEXT: br i1 [[TMP14]], label %[[PRED_LOAD_IF:.*]], label %[[PRED_LOAD_CONTINUE:.*]]
; IC4: [[PRED_LOAD_IF]]:
; IC4-NEXT: [[TMP15:%.*]] = extractelement <4 x i32> [[TMP10]], i64 0
; IC4-NEXT: [[TMP16:%.*]] = getelementptr inbounds i32, ptr [[A]], i32 [[TMP15]]
; IC4-NEXT: [[TMP17:%.*]] = load i32, ptr [[TMP16]], align 4
; IC4-NEXT: [[TMP18:%.*]] = insertelement <4 x i32> poison, i32 [[TMP17]], i64 0
; IC4-NEXT: br label %[[PRED_LOAD_CONTINUE]]
; IC4: [[PRED_LOAD_CONTINUE]]:
; IC4-NEXT: [[TMP19:%.*]] = phi <4 x i32> [ poison, %[[VECTOR_BODY]] ], [ [[TMP18]], %[[PRED_LOAD_IF]] ]
; IC4-NEXT: [[TMP20:%.*]] = extractelement <4 x i1> [[TMP6]], i64 1
; IC4-NEXT: br i1 [[TMP20]], label %[[PRED_LOAD_IF4:.*]], label %[[PRED_LOAD_CONTINUE5:.*]]
; IC4: [[PRED_LOAD_IF4]]:
; IC4-NEXT: [[TMP21:%.*]] = extractelement <4 x i32> [[TMP10]], i64 1
; IC4-NEXT: [[TMP22:%.*]] = getelementptr inbounds i32, ptr [[A]], i32 [[TMP21]]
; IC4-NEXT: [[TMP23:%.*]] = load i32, ptr [[TMP22]], align 4
; IC4-NEXT: [[TMP24:%.*]] = insertelement <4 x i32> [[TMP19]], i32 [[TMP23]], i64 1
; IC4-NEXT: br label %[[PRED_LOAD_CONTINUE5]]
; IC4: [[PRED_LOAD_CONTINUE5]]:
; IC4-NEXT: [[TMP25:%.*]] = phi <4 x i32> [ [[TMP19]], %[[PRED_LOAD_CONTINUE]] ], [ [[TMP24]], %[[PRED_LOAD_IF4]] ]
; IC4-NEXT: [[TMP26:%.*]] = extractelement <4 x i1> [[TMP6]], i64 2
; IC4-NEXT: br i1 [[TMP26]], label %[[PRED_LOAD_IF6:.*]], label %[[PRED_LOAD_CONTINUE7:.*]]
; IC4: [[PRED_LOAD_IF6]]:
; IC4-NEXT: [[TMP27:%.*]] = extractelement <4 x i32> [[TMP10]], i64 2
; IC4-NEXT: [[TMP28:%.*]] = getelementptr inbounds i32, ptr [[A]], i32 [[TMP27]]
; IC4-NEXT: [[TMP29:%.*]] = load i32, ptr [[TMP28]], align 4
; IC4-NEXT: [[TMP30:%.*]] = insertelement <4 x i32> [[TMP25]], i32 [[TMP29]], i64 2
; IC4-NEXT: br label %[[PRED_LOAD_CONTINUE7]]
; IC4: [[PRED_LOAD_CONTINUE7]]:
; IC4-NEXT: [[TMP31:%.*]] = phi <4 x i32> [ [[TMP25]], %[[PRED_LOAD_CONTINUE5]] ], [ [[TMP30]], %[[PRED_LOAD_IF6]] ]
; IC4-NEXT: [[TMP32:%.*]] = extractelement <4 x i1> [[TMP6]], i64 3
; IC4-NEXT: br i1 [[TMP32]], label %[[PRED_LOAD_IF8:.*]], label %[[PRED_LOAD_CONTINUE9:.*]]
; IC4: [[PRED_LOAD_IF8]]:
; IC4-NEXT: [[TMP33:%.*]] = extractelement <4 x i32> [[TMP10]], i64 3
; IC4-NEXT: [[TMP34:%.*]] = getelementptr inbounds i32, ptr [[A]], i32 [[TMP33]]
; IC4-NEXT: [[TMP35:%.*]] = load i32, ptr [[TMP34]], align 4
; IC4-NEXT: [[TMP36:%.*]] = insertelement <4 x i32> [[TMP31]], i32 [[TMP35]], i64 3
; IC4-NEXT: br label %[[PRED_LOAD_CONTINUE9]]
; IC4: [[PRED_LOAD_CONTINUE9]]:
; IC4-NEXT: [[TMP37:%.*]] = phi <4 x i32> [ [[TMP31]], %[[PRED_LOAD_CONTINUE7]] ], [ [[TMP36]], %[[PRED_LOAD_IF8]] ]
; IC4-NEXT: [[TMP38:%.*]] = extractelement <4 x i1> [[TMP7]], i64 0
; IC4-NEXT: br i1 [[TMP38]], label %[[PRED_LOAD_IF10:.*]], label %[[PRED_LOAD_CONTINUE11:.*]]
; IC4: [[PRED_LOAD_IF10]]:
; IC4-NEXT: [[TMP39:%.*]] = extractelement <4 x i32> [[TMP11]], i64 0
; IC4-NEXT: [[TMP40:%.*]] = getelementptr inbounds i32, ptr [[A]], i32 [[TMP39]]
; IC4-NEXT: [[TMP41:%.*]] = load i32, ptr [[TMP40]], align 4
; IC4-NEXT: [[TMP42:%.*]] = insertelement <4 x i32> poison, i32 [[TMP41]], i64 0
; IC4-NEXT: br label %[[PRED_LOAD_CONTINUE11]]
; IC4: [[PRED_LOAD_CONTINUE11]]:
; IC4-NEXT: [[TMP43:%.*]] = phi <4 x i32> [ poison, %[[PRED_LOAD_CONTINUE9]] ], [ [[TMP42]], %[[PRED_LOAD_IF10]] ]
; IC4-NEXT: [[TMP44:%.*]] = extractelement <4 x i1> [[TMP7]], i64 1
; IC4-NEXT: br i1 [[TMP44]], label %[[PRED_LOAD_IF12:.*]], label %[[PRED_LOAD_CONTINUE13:.*]]
; IC4: [[PRED_LOAD_IF12]]:
; IC4-NEXT: [[TMP45:%.*]] = extractelement <4 x i32> [[TMP11]], i64 1
; IC4-NEXT: [[TMP46:%.*]] = getelementptr inbounds i32, ptr [[A]], i32 [[TMP45]]
; IC4-NEXT: [[TMP47:%.*]] = load i32, ptr [[TMP46]], align 4
; IC4-NEXT: [[TMP48:%.*]] = insertelement <4 x i32> [[TMP43]], i32 [[TMP47]], i64 1
; IC4-NEXT: br label %[[PRED_LOAD_CONTINUE13]]
; IC4: [[PRED_LOAD_CONTINUE13]]:
; IC4-NEXT: [[TMP49:%.*]] = phi <4 x i32> [ [[TMP43]], %[[PRED_LOAD_CONTINUE11]] ], [ [[TMP48]], %[[PRED_LOAD_IF12]] ]
; IC4-NEXT: [[TMP50:%.*]] = extractelement <4 x i1> [[TMP7]], i64 2
; IC4-NEXT: br i1 [[TMP50]], label %[[PRED_LOAD_IF14:.*]], label %[[PRED_LOAD_CONTINUE15:.*]]
; IC4: [[PRED_LOAD_IF14]]:
; IC4-NEXT: [[TMP51:%.*]] = extractelement <4 x i32> [[TMP11]], i64 2
; IC4-NEXT: [[TMP52:%.*]] = getelementptr inbounds i32, ptr [[A]], i32 [[TMP51]]
; IC4-NEXT: [[TMP53:%.*]] = load i32, ptr [[TMP52]], align 4
; IC4-NEXT: [[TMP54:%.*]] = insertelement <4 x i32> [[TMP49]], i32 [[TMP53]], i64 2
; IC4-NEXT: br label %[[PRED_LOAD_CONTINUE15]]
; IC4: [[PRED_LOAD_CONTINUE15]]:
; IC4-NEXT: [[TMP55:%.*]] = phi <4 x i32> [ [[TMP49]], %[[PRED_LOAD_CONTINUE13]] ], [ [[TMP54]], %[[PRED_LOAD_IF14]] ]
; IC4-NEXT: [[TMP56:%.*]] = extractelement <4 x i1> [[TMP7]], i64 3
; IC4-NEXT: br i1 [[TMP56]], label %[[PRED_LOAD_IF16:.*]], label %[[PRED_LOAD_CONTINUE17:.*]]
; IC4: [[PRED_LOAD_IF16]]:
; IC4-NEXT: [[TMP57:%.*]] = extractelement <4 x i32> [[TMP11]], i64 3
; IC4-NEXT: [[TMP58:%.*]] = getelementptr inbounds i32, ptr [[A]], i32 [[TMP57]]
; IC4-NEXT: [[TMP59:%.*]] = load i32, ptr [[TMP58]], align 4
; IC4-NEXT: [[TMP60:%.*]] = insertelement <4 x i32> [[TMP55]], i32 [[TMP59]], i64 3
; IC4-NEXT: br label %[[PRED_LOAD_CONTINUE17]]
; IC4: [[PRED_LOAD_CONTINUE17]]:
; IC4-NEXT: [[TMP61:%.*]] = phi <4 x i32> [ [[TMP55]], %[[PRED_LOAD_CONTINUE15]] ], [ [[TMP60]], %[[PRED_LOAD_IF16]] ]
; IC4-NEXT: [[TMP62:%.*]] = extractelement <4 x i1> [[TMP8]], i64 0
; IC4-NEXT: br i1 [[TMP62]], label %[[PRED_LOAD_IF18:.*]], label %[[PRED_LOAD_CONTINUE19:.*]]
; IC4: [[PRED_LOAD_IF18]]:
; IC4-NEXT: [[TMP63:%.*]] = extractelement <4 x i32> [[TMP12]], i64 0
; IC4-NEXT: [[TMP64:%.*]] = getelementptr inbounds i32, ptr [[A]], i32 [[TMP63]]
; IC4-NEXT: [[TMP65:%.*]] = load i32, ptr [[TMP64]], align 4
; IC4-NEXT: [[TMP66:%.*]] = insertelement <4 x i32> poison, i32 [[TMP65]], i64 0
; IC4-NEXT: br label %[[PRED_LOAD_CONTINUE19]]
; IC4: [[PRED_LOAD_CONTINUE19]]:
; IC4-NEXT: [[TMP67:%.*]] = phi <4 x i32> [ poison, %[[PRED_LOAD_CONTINUE17]] ], [ [[TMP66]], %[[PRED_LOAD_IF18]] ]
; IC4-NEXT: [[TMP68:%.*]] = extractelement <4 x i1> [[TMP8]], i64 1
; IC4-NEXT: br i1 [[TMP68]], label %[[PRED_LOAD_IF20:.*]], label %[[PRED_LOAD_CONTINUE21:.*]]
; IC4: [[PRED_LOAD_IF20]]:
; IC4-NEXT: [[TMP69:%.*]] = extractelement <4 x i32> [[TMP12]], i64 1
; IC4-NEXT: [[TMP70:%.*]] = getelementptr inbounds i32, ptr [[A]], i32 [[TMP69]]
; IC4-NEXT: [[TMP71:%.*]] = load i32, ptr [[TMP70]], align 4
; IC4-NEXT: [[TMP72:%.*]] = insertelement <4 x i32> [[TMP67]], i32 [[TMP71]], i64 1
; IC4-NEXT: br label %[[PRED_LOAD_CONTINUE21]]
; IC4: [[PRED_LOAD_CONTINUE21]]:
; IC4-NEXT: [[TMP73:%.*]] = phi <4 x i32> [ [[TMP67]], %[[PRED_LOAD_CONTINUE19]] ], [ [[TMP72]], %[[PRED_LOAD_IF20]] ]
; IC4-NEXT: [[TMP74:%.*]] = extractelement <4 x i1> [[TMP8]], i64 2
; IC4-NEXT: br i1 [[TMP74]], label %[[PRED_LOAD_IF22:.*]], label %[[PRED_LOAD_CONTINUE23:.*]]
; IC4: [[PRED_LOAD_IF22]]:
; IC4-NEXT: [[TMP75:%.*]] = extractelement <4 x i32> [[TMP12]], i64 2
; IC4-NEXT: [[TMP76:%.*]] = getelementptr inbounds i32, ptr [[A]], i32 [[TMP75]]
; IC4-NEXT: [[TMP77:%.*]] = load i32, ptr [[TMP76]], align 4
; IC4-NEXT: [[TMP78:%.*]] = insertelement <4 x i32> [[TMP73]], i32 [[TMP77]], i64 2
; IC4-NEXT: br label %[[PRED_LOAD_CONTINUE23]]
; IC4: [[PRED_LOAD_CONTINUE23]]:
; IC4-NEXT: [[TMP79:%.*]] = phi <4 x i32> [ [[TMP73]], %[[PRED_LOAD_CONTINUE21]] ], [ [[TMP78]], %[[PRED_LOAD_IF22]] ]
; IC4-NEXT: [[TMP80:%.*]] = extractelement <4 x i1> [[TMP8]], i64 3
; IC4-NEXT: br i1 [[TMP80]], label %[[PRED_LOAD_IF24:.*]], label %[[PRED_LOAD_CONTINUE25:.*]]
; IC4: [[PRED_LOAD_IF24]]:
; IC4-NEXT: [[TMP81:%.*]] = extractelement <4 x i32> [[TMP12]], i64 3
; IC4-NEXT: [[TMP82:%.*]] = getelementptr inbounds i32, ptr [[A]], i32 [[TMP81]]
; IC4-NEXT: [[TMP83:%.*]] = load i32, ptr [[TMP82]], align 4
; IC4-NEXT: [[TMP84:%.*]] = insertelement <4 x i32> [[TMP79]], i32 [[TMP83]], i64 3
; IC4-NEXT: br label %[[PRED_LOAD_CONTINUE25]]
; IC4: [[PRED_LOAD_CONTINUE25]]:
; IC4-NEXT: [[TMP85:%.*]] = phi <4 x i32> [ [[TMP79]], %[[PRED_LOAD_CONTINUE23]] ], [ [[TMP84]], %[[PRED_LOAD_IF24]] ]
; IC4-NEXT: [[TMP86:%.*]] = extractelement <4 x i1> [[TMP9]], i64 0
; IC4-NEXT: br i1 [[TMP86]], label %[[PRED_LOAD_IF26:.*]], label %[[PRED_LOAD_CONTINUE27:.*]]
; IC4: [[PRED_LOAD_IF26]]:
; IC4-NEXT: [[TMP87:%.*]] = extractelement <4 x i32> [[TMP13]], i64 0
; IC4-NEXT: [[TMP88:%.*]] = getelementptr inbounds i32, ptr [[A]], i32 [[TMP87]]
; IC4-NEXT: [[TMP89:%.*]] = load i32, ptr [[TMP88]], align 4
; IC4-NEXT: [[TMP90:%.*]] = insertelement <4 x i32> poison, i32 [[TMP89]], i64 0
; IC4-NEXT: br label %[[PRED_LOAD_CONTINUE27]]
; IC4: [[PRED_LOAD_CONTINUE27]]:
; IC4-NEXT: [[TMP91:%.*]] = phi <4 x i32> [ poison, %[[PRED_LOAD_CONTINUE25]] ], [ [[TMP90]], %[[PRED_LOAD_IF26]] ]
; IC4-NEXT: [[TMP92:%.*]] = extractelement <4 x i1> [[TMP9]], i64 1
; IC4-NEXT: br i1 [[TMP92]], label %[[PRED_LOAD_IF28:.*]], label %[[PRED_LOAD_CONTINUE29:.*]]
; IC4: [[PRED_LOAD_IF28]]:
; IC4-NEXT: [[TMP93:%.*]] = extractelement <4 x i32> [[TMP13]], i64 1
; IC4-NEXT: [[TMP94:%.*]] = getelementptr inbounds i32, ptr [[A]], i32 [[TMP93]]
; IC4-NEXT: [[TMP95:%.*]] = load i32, ptr [[TMP94]], align 4
; IC4-NEXT: [[TMP96:%.*]] = insertelement <4 x i32> [[TMP91]], i32 [[TMP95]], i64 1
; IC4-NEXT: br label %[[PRED_LOAD_CONTINUE29]]
; IC4: [[PRED_LOAD_CONTINUE29]]:
; IC4-NEXT: [[TMP97:%.*]] = phi <4 x i32> [ [[TMP91]], %[[PRED_LOAD_CONTINUE27]] ], [ [[TMP96]], %[[PRED_LOAD_IF28]] ]
; IC4-NEXT: [[TMP98:%.*]] = extractelement <4 x i1> [[TMP9]], i64 2
; IC4-NEXT: br i1 [[TMP98]], label %[[PRED_LOAD_IF30:.*]], label %[[PRED_LOAD_CONTINUE31:.*]]
; IC4: [[PRED_LOAD_IF30]]:
; IC4-NEXT: [[TMP99:%.*]] = extractelement <4 x i32> [[TMP13]], i64 2
; IC4-NEXT: [[TMP100:%.*]] = getelementptr inbounds i32, ptr [[A]], i32 [[TMP99]]
; IC4-NEXT: [[TMP101:%.*]] = load i32, ptr [[TMP100]], align 4
; IC4-NEXT: [[TMP102:%.*]] = insertelement <4 x i32> [[TMP97]], i32 [[TMP101]], i64 2
; IC4-NEXT: br label %[[PRED_LOAD_CONTINUE31]]
; IC4: [[PRED_LOAD_CONTINUE31]]:
; IC4-NEXT: [[TMP103:%.*]] = phi <4 x i32> [ [[TMP97]], %[[PRED_LOAD_CONTINUE29]] ], [ [[TMP102]], %[[PRED_LOAD_IF30]] ]
; IC4-NEXT: [[TMP104:%.*]] = extractelement <4 x i1> [[TMP9]], i64 3
; IC4-NEXT: br i1 [[TMP104]], label %[[PRED_LOAD_IF32:.*]], label %[[PRED_LOAD_CONTINUE33]]
; IC4: [[PRED_LOAD_IF32]]:
; IC4-NEXT: [[TMP105:%.*]] = extractelement <4 x i32> [[TMP13]], i64 3
; IC4-NEXT: [[TMP106:%.*]] = getelementptr inbounds i32, ptr [[A]], i32 [[TMP105]]
; IC4-NEXT: [[TMP107:%.*]] = load i32, ptr [[TMP106]], align 4
; IC4-NEXT: [[TMP108:%.*]] = insertelement <4 x i32> [[TMP103]], i32 [[TMP107]], i64 3
; IC4-NEXT: br label %[[PRED_LOAD_CONTINUE33]]
; IC4: [[PRED_LOAD_CONTINUE33]]:
; IC4-NEXT: [[TMP109:%.*]] = phi <4 x i32> [ [[TMP103]], %[[PRED_LOAD_CONTINUE31]] ], [ [[TMP108]], %[[PRED_LOAD_IF32]] ]
; IC4-NEXT: [[PREDPHI:%.*]] = select <4 x i1> [[TMP6]], <4 x i32> [[TMP37]], <4 x i32> zeroinitializer
; IC4-NEXT: [[PREDPHI34:%.*]] = select <4 x i1> [[TMP7]], <4 x i32> [[TMP61]], <4 x i32> zeroinitializer
; IC4-NEXT: [[PREDPHI35:%.*]] = select <4 x i1> [[TMP8]], <4 x i32> [[TMP85]], <4 x i32> zeroinitializer
; IC4-NEXT: [[PREDPHI36:%.*]] = select <4 x i1> [[TMP9]], <4 x i32> [[TMP109]], <4 x i32> zeroinitializer
; IC4-NEXT: [[TMP110]] = add <4 x i32> [[VEC_PHI]], [[PREDPHI]]
; IC4-NEXT: [[TMP111]] = add <4 x i32> [[VEC_PHI1]], [[PREDPHI34]]
; IC4-NEXT: [[TMP112]] = add <4 x i32> [[VEC_PHI2]], [[PREDPHI35]]
; IC4-NEXT: [[TMP113]] = add <4 x i32> [[VEC_PHI3]], [[PREDPHI36]]
; IC4-NEXT: [[INDEX_NEXT]] = add nuw i32 [[INDEX]], 16
; IC4-NEXT: [[VEC_IND_NEXT]] = add <4 x i32> [[STEP_ADD_3]], splat (i32 4)
; IC4-NEXT: [[TMP114:%.*]] = icmp eq i32 [[INDEX_NEXT]], [[N_VEC]]
; IC4-NEXT: br i1 [[TMP114]], label %[[MIDDLE_BLOCK:.*]], label %[[VECTOR_BODY]], !llvm.loop [[LOOP4:![0-9]+]]
; IC4: [[MIDDLE_BLOCK]]:
; IC4-NEXT: [[BIN_RDX:%.*]] = add <4 x i32> [[TMP111]], [[TMP110]]
; IC4-NEXT: [[BIN_RDX37:%.*]] = add <4 x i32> [[TMP112]], [[BIN_RDX]]
; IC4-NEXT: [[BIN_RDX38:%.*]] = add <4 x i32> [[TMP113]], [[BIN_RDX37]]
; IC4-NEXT: [[TMP115:%.*]] = call i32 @llvm.vector.reduce.add.v4i32(<4 x i32> [[BIN_RDX38]])
; IC4-NEXT: [[CMP_N:%.*]] = icmp eq i32 [[N]], [[N_VEC]]
; IC4-NEXT: br i1 [[CMP_N]], label %[[EXIT:.*]], label %[[SCALAR_PH]]
; IC4: [[SCALAR_PH]]:
; IC4-NEXT: [[BC_RESUME_VAL:%.*]] = phi i32 [ [[N_VEC]], %[[MIDDLE_BLOCK]] ], [ 0, %[[ENTRY]] ], [ 0, %[[VECTOR_SCEVCHECK]] ]
; IC4-NEXT: [[BC_MERGE_RDX:%.*]] = phi i32 [ [[TMP115]], %[[MIDDLE_BLOCK]] ], [ 0, %[[ENTRY]] ], [ 0, %[[VECTOR_SCEVCHECK]] ]
; IC4-NEXT: br label %[[LOOP:.*]]
; IC4: [[LOOP]]:
; IC4-NEXT: [[IV:%.*]] = phi i32 [ [[BC_RESUME_VAL]], %[[SCALAR_PH]] ], [ [[IV_NEXT:%.*]], %[[LATCH:.*]] ]
; IC4-NEXT: [[SUM:%.*]] = phi i32 [ [[BC_MERGE_RDX]], %[[SCALAR_PH]] ], [ [[SUM_NEXT:%.*]], %[[LATCH]] ]
; IC4-NEXT: [[AND_IV:%.*]] = and i32 [[IV]], 1
; IC4-NEXT: [[CMP:%.*]] = icmp eq i32 [[AND_IV]], 0
; IC4-NEXT: br i1 [[CMP]], label %[[THEN:.*]], label %[[LATCH]]
; IC4: [[THEN]]:
; IC4-NEXT: [[CLAMPED:%.*]] = urem i32 [[IV]], 16
; IC4-NEXT: [[GEP_A:%.*]] = getelementptr inbounds i32, ptr [[A]], i32 [[CLAMPED]]
; IC4-NEXT: [[LV:%.*]] = load i32, ptr [[GEP_A]], align 4
; IC4-NEXT: br label %[[LATCH]]
; IC4: [[LATCH]]:
; IC4-NEXT: [[V:%.*]] = phi i32 [ [[LV]], %[[THEN]] ], [ 0, %[[LOOP]] ]
; IC4-NEXT: [[SUM_NEXT]] = add i32 [[SUM]], [[V]]
; IC4-NEXT: [[IV_NEXT]] = add i32 [[IV]], 1
; IC4-NEXT: [[EC:%.*]] = icmp eq i32 [[IV_NEXT]], [[N]]
; IC4-NEXT: br i1 [[EC]], label %[[EXIT]], label %[[LOOP]], !llvm.loop [[LOOP5:![0-9]+]]
; IC4: [[EXIT]]:
; IC4-NEXT: [[SUM_NEXT_LCSSA:%.*]] = phi i32 [ [[SUM_NEXT]], %[[LATCH]] ], [ [[TMP115]], %[[MIDDLE_BLOCK]] ]
; IC4-NEXT: ret i32 [[SUM_NEXT_LCSSA]]
;
; IC8-LABEL: define i32 @clamped_load_predicated_ic_capped(
; IC8-SAME: ptr [[A:%.*]], i32 [[N:%.*]]) {
; IC8-NEXT: [[ENTRY:.*]]:
; IC8-NEXT: [[MIN_ITERS_CHECK:%.*]] = icmp ult i32 [[N]], 32
; IC8-NEXT: br i1 [[MIN_ITERS_CHECK]], label %[[SCALAR_PH:.*]], label %[[VECTOR_SCEVCHECK:.*]]
; IC8: [[VECTOR_SCEVCHECK]]:
; IC8-NEXT: [[TMP0:%.*]] = add i32 [[N]], -1
; IC8-NEXT: [[TMP1:%.*]] = icmp ugt i32 [[TMP0]], 15
; IC8-NEXT: br i1 [[TMP1]], label %[[SCALAR_PH]], label %[[VECTOR_PH:.*]]
; IC8: [[VECTOR_PH]]:
; IC8-NEXT: [[N_MOD_VF:%.*]] = urem i32 [[N]], 32
; IC8-NEXT: [[N_VEC:%.*]] = sub i32 [[N]], [[N_MOD_VF]]
; IC8-NEXT: br label %[[VECTOR_BODY:.*]]
; IC8: [[VECTOR_BODY]]:
; IC8-NEXT: [[INDEX:%.*]] = phi i32 [ 0, %[[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], %[[PRED_LOAD_CONTINUE69:.*]] ]
; IC8-NEXT: [[VEC_IND:%.*]] = phi <4 x i32> [ <i32 0, i32 1, i32 2, i32 3>, %[[VECTOR_PH]] ], [ [[VEC_IND_NEXT:%.*]], %[[PRED_LOAD_CONTINUE69]] ]
; IC8-NEXT: [[VEC_PHI:%.*]] = phi <4 x i32> [ zeroinitializer, %[[VECTOR_PH]] ], [ [[TMP218:%.*]], %[[PRED_LOAD_CONTINUE69]] ]
; IC8-NEXT: [[VEC_PHI1:%.*]] = phi <4 x i32> [ zeroinitializer, %[[VECTOR_PH]] ], [ [[TMP219:%.*]], %[[PRED_LOAD_CONTINUE69]] ]
; IC8-NEXT: [[VEC_PHI2:%.*]] = phi <4 x i32> [ zeroinitializer, %[[VECTOR_PH]] ], [ [[TMP220:%.*]], %[[PRED_LOAD_CONTINUE69]] ]
; IC8-NEXT: [[VEC_PHI3:%.*]] = phi <4 x i32> [ zeroinitializer, %[[VECTOR_PH]] ], [ [[TMP221:%.*]], %[[PRED_LOAD_CONTINUE69]] ]
; IC8-NEXT: [[VEC_PHI4:%.*]] = phi <4 x i32> [ zeroinitializer, %[[VECTOR_PH]] ], [ [[TMP222:%.*]], %[[PRED_LOAD_CONTINUE69]] ]
; IC8-NEXT: [[VEC_PHI5:%.*]] = phi <4 x i32> [ zeroinitializer, %[[VECTOR_PH]] ], [ [[TMP223:%.*]], %[[PRED_LOAD_CONTINUE69]] ]
; IC8-NEXT: [[VEC_PHI6:%.*]] = phi <4 x i32> [ zeroinitializer, %[[VECTOR_PH]] ], [ [[TMP224:%.*]], %[[PRED_LOAD_CONTINUE69]] ]
; IC8-NEXT: [[VEC_PHI7:%.*]] = phi <4 x i32> [ zeroinitializer, %[[VECTOR_PH]] ], [ [[TMP225:%.*]], %[[PRED_LOAD_CONTINUE69]] ]
; IC8-NEXT: [[STEP_ADD:%.*]] = add nuw <4 x i32> [[VEC_IND]], splat (i32 4)
; IC8-NEXT: [[STEP_ADD_2:%.*]] = add nuw <4 x i32> [[STEP_ADD]], splat (i32 4)
; IC8-NEXT: [[STEP_ADD_3:%.*]] = add nuw <4 x i32> [[STEP_ADD_2]], splat (i32 4)
; IC8-NEXT: [[STEP_ADD_4:%.*]] = add nuw <4 x i32> [[STEP_ADD_3]], splat (i32 4)
; IC8-NEXT: [[STEP_ADD_5:%.*]] = add nuw <4 x i32> [[STEP_ADD_4]], splat (i32 4)
; IC8-NEXT: [[STEP_ADD_6:%.*]] = add nuw <4 x i32> [[STEP_ADD_5]], splat (i32 4)
; IC8-NEXT: [[STEP_ADD_7:%.*]] = add nuw <4 x i32> [[STEP_ADD_6]], splat (i32 4)
; IC8-NEXT: [[TMP2:%.*]] = and <4 x i32> [[VEC_IND]], splat (i32 1)
; IC8-NEXT: [[TMP3:%.*]] = and <4 x i32> [[STEP_ADD]], splat (i32 1)
; IC8-NEXT: [[TMP4:%.*]] = and <4 x i32> [[STEP_ADD_2]], splat (i32 1)
; IC8-NEXT: [[TMP5:%.*]] = and <4 x i32> [[STEP_ADD_3]], splat (i32 1)
; IC8-NEXT: [[TMP6:%.*]] = and <4 x i32> [[STEP_ADD_4]], splat (i32 1)
; IC8-NEXT: [[TMP7:%.*]] = and <4 x i32> [[STEP_ADD_5]], splat (i32 1)
; IC8-NEXT: [[TMP8:%.*]] = and <4 x i32> [[STEP_ADD_6]], splat (i32 1)
; IC8-NEXT: [[TMP9:%.*]] = and <4 x i32> [[STEP_ADD_7]], splat (i32 1)
; IC8-NEXT: [[TMP10:%.*]] = icmp eq <4 x i32> [[TMP2]], zeroinitializer
; IC8-NEXT: [[TMP11:%.*]] = icmp eq <4 x i32> [[TMP3]], zeroinitializer
; IC8-NEXT: [[TMP12:%.*]] = icmp eq <4 x i32> [[TMP4]], zeroinitializer
; IC8-NEXT: [[TMP13:%.*]] = icmp eq <4 x i32> [[TMP5]], zeroinitializer
; IC8-NEXT: [[TMP14:%.*]] = icmp eq <4 x i32> [[TMP6]], zeroinitializer
; IC8-NEXT: [[TMP15:%.*]] = icmp eq <4 x i32> [[TMP7]], zeroinitializer
; IC8-NEXT: [[TMP16:%.*]] = icmp eq <4 x i32> [[TMP8]], zeroinitializer
; IC8-NEXT: [[TMP17:%.*]] = icmp eq <4 x i32> [[TMP9]], zeroinitializer
; IC8-NEXT: [[TMP18:%.*]] = urem <4 x i32> [[VEC_IND]], splat (i32 16)
; IC8-NEXT: [[TMP19:%.*]] = urem <4 x i32> [[STEP_ADD]], splat (i32 16)
; IC8-NEXT: [[TMP20:%.*]] = urem <4 x i32> [[STEP_ADD_2]], splat (i32 16)
; IC8-NEXT: [[TMP21:%.*]] = urem <4 x i32> [[STEP_ADD_3]], splat (i32 16)
; IC8-NEXT: [[TMP22:%.*]] = urem <4 x i32> [[STEP_ADD_4]], splat (i32 16)
; IC8-NEXT: [[TMP23:%.*]] = urem <4 x i32> [[STEP_ADD_5]], splat (i32 16)
; IC8-NEXT: [[TMP24:%.*]] = urem <4 x i32> [[STEP_ADD_6]], splat (i32 16)
; IC8-NEXT: [[TMP25:%.*]] = urem <4 x i32> [[STEP_ADD_7]], splat (i32 16)
; IC8-NEXT: [[TMP26:%.*]] = extractelement <4 x i1> [[TMP10]], i64 0
; IC8-NEXT: br i1 [[TMP26]], label %[[PRED_LOAD_IF:.*]], label %[[PRED_LOAD_CONTINUE:.*]]
; IC8: [[PRED_LOAD_IF]]:
; IC8-NEXT: [[TMP27:%.*]] = extractelement <4 x i32> [[TMP18]], i64 0
; IC8-NEXT: [[TMP28:%.*]] = getelementptr inbounds i32, ptr [[A]], i32 [[TMP27]]
; IC8-NEXT: [[TMP29:%.*]] = load i32, ptr [[TMP28]], align 4
; IC8-NEXT: [[TMP30:%.*]] = insertelement <4 x i32> poison, i32 [[TMP29]], i64 0
; IC8-NEXT: br label %[[PRED_LOAD_CONTINUE]]
; IC8: [[PRED_LOAD_CONTINUE]]:
; IC8-NEXT: [[TMP31:%.*]] = phi <4 x i32> [ poison, %[[VECTOR_BODY]] ], [ [[TMP30]], %[[PRED_LOAD_IF]] ]
; IC8-NEXT: [[TMP32:%.*]] = extractelement <4 x i1> [[TMP10]], i64 1
; IC8-NEXT: br i1 [[TMP32]], label %[[PRED_LOAD_IF8:.*]], label %[[PRED_LOAD_CONTINUE9:.*]]
; IC8: [[PRED_LOAD_IF8]]:
; IC8-NEXT: [[TMP33:%.*]] = extractelement <4 x i32> [[TMP18]], i64 1
; IC8-NEXT: [[TMP34:%.*]] = getelementptr inbounds i32, ptr [[A]], i32 [[TMP33]]
; IC8-NEXT: [[TMP35:%.*]] = load i32, ptr [[TMP34]], align 4
; IC8-NEXT: [[TMP36:%.*]] = insertelement <4 x i32> [[TMP31]], i32 [[TMP35]], i64 1
; IC8-NEXT: br label %[[PRED_LOAD_CONTINUE9]]
; IC8: [[PRED_LOAD_CONTINUE9]]:
; IC8-NEXT: [[TMP37:%.*]] = phi <4 x i32> [ [[TMP31]], %[[PRED_LOAD_CONTINUE]] ], [ [[TMP36]], %[[PRED_LOAD_IF8]] ]
; IC8-NEXT: [[TMP38:%.*]] = extractelement <4 x i1> [[TMP10]], i64 2
; IC8-NEXT: br i1 [[TMP38]], label %[[PRED_LOAD_IF10:.*]], label %[[PRED_LOAD_CONTINUE11:.*]]
; IC8: [[PRED_LOAD_IF10]]:
; IC8-NEXT: [[TMP39:%.*]] = extractelement <4 x i32> [[TMP18]], i64 2
; IC8-NEXT: [[TMP40:%.*]] = getelementptr inbounds i32, ptr [[A]], i32 [[TMP39]]
; IC8-NEXT: [[TMP41:%.*]] = load i32, ptr [[TMP40]], align 4
; IC8-NEXT: [[TMP42:%.*]] = insertelement <4 x i32> [[TMP37]], i32 [[TMP41]], i64 2
; IC8-NEXT: br label %[[PRED_LOAD_CONTINUE11]]
; IC8: [[PRED_LOAD_CONTINUE11]]:
; IC8-NEXT: [[TMP43:%.*]] = phi <4 x i32> [ [[TMP37]], %[[PRED_LOAD_CONTINUE9]] ], [ [[TMP42]], %[[PRED_LOAD_IF10]] ]
; IC8-NEXT: [[TMP44:%.*]] = extractelement <4 x i1> [[TMP10]], i64 3
; IC8-NEXT: br i1 [[TMP44]], label %[[PRED_LOAD_IF12:.*]], label %[[PRED_LOAD_CONTINUE13:.*]]
; IC8: [[PRED_LOAD_IF12]]:
; IC8-NEXT: [[TMP45:%.*]] = extractelement <4 x i32> [[TMP18]], i64 3
; IC8-NEXT: [[TMP46:%.*]] = getelementptr inbounds i32, ptr [[A]], i32 [[TMP45]]
; IC8-NEXT: [[TMP47:%.*]] = load i32, ptr [[TMP46]], align 4
; IC8-NEXT: [[TMP48:%.*]] = insertelement <4 x i32> [[TMP43]], i32 [[TMP47]], i64 3
; IC8-NEXT: br label %[[PRED_LOAD_CONTINUE13]]
; IC8: [[PRED_LOAD_CONTINUE13]]:
; IC8-NEXT: [[TMP49:%.*]] = phi <4 x i32> [ [[TMP43]], %[[PRED_LOAD_CONTINUE11]] ], [ [[TMP48]], %[[PRED_LOAD_IF12]] ]
; IC8-NEXT: [[TMP50:%.*]] = extractelement <4 x i1> [[TMP11]], i64 0
; IC8-NEXT: br i1 [[TMP50]], label %[[PRED_LOAD_IF14:.*]], label %[[PRED_LOAD_CONTINUE15:.*]]
; IC8: [[PRED_LOAD_IF14]]:
; IC8-NEXT: [[TMP51:%.*]] = extractelement <4 x i32> [[TMP19]], i64 0
; IC8-NEXT: [[TMP52:%.*]] = getelementptr inbounds i32, ptr [[A]], i32 [[TMP51]]
; IC8-NEXT: [[TMP53:%.*]] = load i32, ptr [[TMP52]], align 4
; IC8-NEXT: [[TMP54:%.*]] = insertelement <4 x i32> poison, i32 [[TMP53]], i64 0
; IC8-NEXT: br label %[[PRED_LOAD_CONTINUE15]]
; IC8: [[PRED_LOAD_CONTINUE15]]:
; IC8-NEXT: [[TMP55:%.*]] = phi <4 x i32> [ poison, %[[PRED_LOAD_CONTINUE13]] ], [ [[TMP54]], %[[PRED_LOAD_IF14]] ]
; IC8-NEXT: [[TMP56:%.*]] = extractelement <4 x i1> [[TMP11]], i64 1
; IC8-NEXT: br i1 [[TMP56]], label %[[PRED_LOAD_IF16:.*]], label %[[PRED_LOAD_CONTINUE17:.*]]
; IC8: [[PRED_LOAD_IF16]]:
; IC8-NEXT: [[TMP57:%.*]] = extractelement <4 x i32> [[TMP19]], i64 1
; IC8-NEXT: [[TMP58:%.*]] = getelementptr inbounds i32, ptr [[A]], i32 [[TMP57]]
; IC8-NEXT: [[TMP59:%.*]] = load i32, ptr [[TMP58]], align 4
; IC8-NEXT: [[TMP60:%.*]] = insertelement <4 x i32> [[TMP55]], i32 [[TMP59]], i64 1
; IC8-NEXT: br label %[[PRED_LOAD_CONTINUE17]]
; IC8: [[PRED_LOAD_CONTINUE17]]:
; IC8-NEXT: [[TMP61:%.*]] = phi <4 x i32> [ [[TMP55]], %[[PRED_LOAD_CONTINUE15]] ], [ [[TMP60]], %[[PRED_LOAD_IF16]] ]
; IC8-NEXT: [[TMP62:%.*]] = extractelement <4 x i1> [[TMP11]], i64 2
; IC8-NEXT: br i1 [[TMP62]], label %[[PRED_LOAD_IF18:.*]], label %[[PRED_LOAD_CONTINUE19:.*]]
; IC8: [[PRED_LOAD_IF18]]:
; IC8-NEXT: [[TMP63:%.*]] = extractelement <4 x i32> [[TMP19]], i64 2
; IC8-NEXT: [[TMP64:%.*]] = getelementptr inbounds i32, ptr [[A]], i32 [[TMP63]]
; IC8-NEXT: [[TMP65:%.*]] = load i32, ptr [[TMP64]], align 4
; IC8-NEXT: [[TMP66:%.*]] = insertelement <4 x i32> [[TMP61]], i32 [[TMP65]], i64 2
; IC8-NEXT: br label %[[PRED_LOAD_CONTINUE19]]
; IC8: [[PRED_LOAD_CONTINUE19]]:
; IC8-NEXT: [[TMP67:%.*]] = phi <4 x i32> [ [[TMP61]], %[[PRED_LOAD_CONTINUE17]] ], [ [[TMP66]], %[[PRED_LOAD_IF18]] ]
; IC8-NEXT: [[TMP68:%.*]] = extractelement <4 x i1> [[TMP11]], i64 3
; IC8-NEXT: br i1 [[TMP68]], label %[[PRED_LOAD_IF20:.*]], label %[[PRED_LOAD_CONTINUE21:.*]]
; IC8: [[PRED_LOAD_IF20]]:
; IC8-NEXT: [[TMP69:%.*]] = extractelement <4 x i32> [[TMP19]], i64 3
; IC8-NEXT: [[TMP70:%.*]] = getelementptr inbounds i32, ptr [[A]], i32 [[TMP69]]
; IC8-NEXT: [[TMP71:%.*]] = load i32, ptr [[TMP70]], align 4
; IC8-NEXT: [[TMP72:%.*]] = insertelement <4 x i32> [[TMP67]], i32 [[TMP71]], i64 3
; IC8-NEXT: br label %[[PRED_LOAD_CONTINUE21]]
; IC8: [[PRED_LOAD_CONTINUE21]]:
; IC8-NEXT: [[TMP73:%.*]] = phi <4 x i32> [ [[TMP67]], %[[PRED_LOAD_CONTINUE19]] ], [ [[TMP72]], %[[PRED_LOAD_IF20]] ]
; IC8-NEXT: [[TMP74:%.*]] = extractelement <4 x i1> [[TMP12]], i64 0
; IC8-NEXT: br i1 [[TMP74]], label %[[PRED_LOAD_IF22:.*]], label %[[PRED_LOAD_CONTINUE23:.*]]
; IC8: [[PRED_LOAD_IF22]]:
; IC8-NEXT: [[TMP75:%.*]] = extractelement <4 x i32> [[TMP20]], i64 0
; IC8-NEXT: [[TMP76:%.*]] = getelementptr inbounds i32, ptr [[A]], i32 [[TMP75]]
; IC8-NEXT: [[TMP77:%.*]] = load i32, ptr [[TMP76]], align 4
; IC8-NEXT: [[TMP78:%.*]] = insertelement <4 x i32> poison, i32 [[TMP77]], i64 0
; IC8-NEXT: br label %[[PRED_LOAD_CONTINUE23]]
; IC8: [[PRED_LOAD_CONTINUE23]]:
; IC8-NEXT: [[TMP79:%.*]] = phi <4 x i32> [ poison, %[[PRED_LOAD_CONTINUE21]] ], [ [[TMP78]], %[[PRED_LOAD_IF22]] ]
; IC8-NEXT: [[TMP80:%.*]] = extractelement <4 x i1> [[TMP12]], i64 1
; IC8-NEXT: br i1 [[TMP80]], label %[[PRED_LOAD_IF24:.*]], label %[[PRED_LOAD_CONTINUE25:.*]]
; IC8: [[PRED_LOAD_IF24]]:
; IC8-NEXT: [[TMP81:%.*]] = extractelement <4 x i32> [[TMP20]], i64 1
; IC8-NEXT: [[TMP82:%.*]] = getelementptr inbounds i32, ptr [[A]], i32 [[TMP81]]
; IC8-NEXT: [[TMP83:%.*]] = load i32, ptr [[TMP82]], align 4
; IC8-NEXT: [[TMP84:%.*]] = insertelement <4 x i32> [[TMP79]], i32 [[TMP83]], i64 1
; IC8-NEXT: br label %[[PRED_LOAD_CONTINUE25]]
; IC8: [[PRED_LOAD_CONTINUE25]]:
; IC8-NEXT: [[TMP85:%.*]] = phi <4 x i32> [ [[TMP79]], %[[PRED_LOAD_CONTINUE23]] ], [ [[TMP84]], %[[PRED_LOAD_IF24]] ]
; IC8-NEXT: [[TMP86:%.*]] = extractelement <4 x i1> [[TMP12]], i64 2
; IC8-NEXT: br i1 [[TMP86]], label %[[PRED_LOAD_IF26:.*]], label %[[PRED_LOAD_CONTINUE27:.*]]
; IC8: [[PRED_LOAD_IF26]]:
; IC8-NEXT: [[TMP87:%.*]] = extractelement <4 x i32> [[TMP20]], i64 2
; IC8-NEXT: [[TMP88:%.*]] = getelementptr inbounds i32, ptr [[A]], i32 [[TMP87]]
; IC8-NEXT: [[TMP89:%.*]] = load i32, ptr [[TMP88]], align 4
; IC8-NEXT: [[TMP90:%.*]] = insertelement <4 x i32> [[TMP85]], i32 [[TMP89]], i64 2
; IC8-NEXT: br label %[[PRED_LOAD_CONTINUE27]]
; IC8: [[PRED_LOAD_CONTINUE27]]:
; IC8-NEXT: [[TMP91:%.*]] = phi <4 x i32> [ [[TMP85]], %[[PRED_LOAD_CONTINUE25]] ], [ [[TMP90]], %[[PRED_LOAD_IF26]] ]
; IC8-NEXT: [[TMP92:%.*]] = extractelement <4 x i1> [[TMP12]], i64 3
; IC8-NEXT: br i1 [[TMP92]], label %[[PRED_LOAD_IF28:.*]], label %[[PRED_LOAD_CONTINUE29:.*]]
; IC8: [[PRED_LOAD_IF28]]:
; IC8-NEXT: [[TMP93:%.*]] = extractelement <4 x i32> [[TMP20]], i64 3
; IC8-NEXT: [[TMP94:%.*]] = getelementptr inbounds i32, ptr [[A]], i32 [[TMP93]]
; IC8-NEXT: [[TMP95:%.*]] = load i32, ptr [[TMP94]], align 4
; IC8-NEXT: [[TMP96:%.*]] = insertelement <4 x i32> [[TMP91]], i32 [[TMP95]], i64 3
; IC8-NEXT: br label %[[PRED_LOAD_CONTINUE29]]
; IC8: [[PRED_LOAD_CONTINUE29]]:
; IC8-NEXT: [[TMP97:%.*]] = phi <4 x i32> [ [[TMP91]], %[[PRED_LOAD_CONTINUE27]] ], [ [[TMP96]], %[[PRED_LOAD_IF28]] ]
; IC8-NEXT: [[TMP98:%.*]] = extractelement <4 x i1> [[TMP13]], i64 0
; IC8-NEXT: br i1 [[TMP98]], label %[[PRED_LOAD_IF30:.*]], label %[[PRED_LOAD_CONTINUE31:.*]]
; IC8: [[PRED_LOAD_IF30]]:
; IC8-NEXT: [[TMP99:%.*]] = extractelement <4 x i32> [[TMP21]], i64 0
; IC8-NEXT: [[TMP100:%.*]] = getelementptr inbounds i32, ptr [[A]], i32 [[TMP99]]
; IC8-NEXT: [[TMP101:%.*]] = load i32, ptr [[TMP100]], align 4
; IC8-NEXT: [[TMP102:%.*]] = insertelement <4 x i32> poison, i32 [[TMP101]], i64 0
; IC8-NEXT: br label %[[PRED_LOAD_CONTINUE31]]
; IC8: [[PRED_LOAD_CONTINUE31]]:
; IC8-NEXT: [[TMP103:%.*]] = phi <4 x i32> [ poison, %[[PRED_LOAD_CONTINUE29]] ], [ [[TMP102]], %[[PRED_LOAD_IF30]] ]
; IC8-NEXT: [[TMP104:%.*]] = extractelement <4 x i1> [[TMP13]], i64 1
; IC8-NEXT: br i1 [[TMP104]], label %[[PRED_LOAD_IF32:.*]], label %[[PRED_LOAD_CONTINUE33:.*]]
; IC8: [[PRED_LOAD_IF32]]:
; IC8-NEXT: [[TMP105:%.*]] = extractelement <4 x i32> [[TMP21]], i64 1
; IC8-NEXT: [[TMP106:%.*]] = getelementptr inbounds i32, ptr [[A]], i32 [[TMP105]]
; IC8-NEXT: [[TMP107:%.*]] = load i32, ptr [[TMP106]], align 4
; IC8-NEXT: [[TMP108:%.*]] = insertelement <4 x i32> [[TMP103]], i32 [[TMP107]], i64 1
; IC8-NEXT: br label %[[PRED_LOAD_CONTINUE33]]
; IC8: [[PRED_LOAD_CONTINUE33]]:
; IC8-NEXT: [[TMP109:%.*]] = phi <4 x i32> [ [[TMP103]], %[[PRED_LOAD_CONTINUE31]] ], [ [[TMP108]], %[[PRED_LOAD_IF32]] ]
; IC8-NEXT: [[TMP110:%.*]] = extractelement <4 x i1> [[TMP13]], i64 2
; IC8-NEXT: br i1 [[TMP110]], label %[[PRED_LOAD_IF34:.*]], label %[[PRED_LOAD_CONTINUE35:.*]]
; IC8: [[PRED_LOAD_IF34]]:
; IC8-NEXT: [[TMP111:%.*]] = extractelement <4 x i32> [[TMP21]], i64 2
; IC8-NEXT: [[TMP112:%.*]] = getelementptr inbounds i32, ptr [[A]], i32 [[TMP111]]
; IC8-NEXT: [[TMP113:%.*]] = load i32, ptr [[TMP112]], align 4
; IC8-NEXT: [[TMP114:%.*]] = insertelement <4 x i32> [[TMP109]], i32 [[TMP113]], i64 2
; IC8-NEXT: br label %[[PRED_LOAD_CONTINUE35]]
; IC8: [[PRED_LOAD_CONTINUE35]]:
; IC8-NEXT: [[TMP115:%.*]] = phi <4 x i32> [ [[TMP109]], %[[PRED_LOAD_CONTINUE33]] ], [ [[TMP114]], %[[PRED_LOAD_IF34]] ]
; IC8-NEXT: [[TMP116:%.*]] = extractelement <4 x i1> [[TMP13]], i64 3
; IC8-NEXT: br i1 [[TMP116]], label %[[PRED_LOAD_IF36:.*]], label %[[PRED_LOAD_CONTINUE37:.*]]
; IC8: [[PRED_LOAD_IF36]]:
; IC8-NEXT: [[TMP117:%.*]] = extractelement <4 x i32> [[TMP21]], i64 3
; IC8-NEXT: [[TMP118:%.*]] = getelementptr inbounds i32, ptr [[A]], i32 [[TMP117]]
; IC8-NEXT: [[TMP119:%.*]] = load i32, ptr [[TMP118]], align 4
; IC8-NEXT: [[TMP120:%.*]] = insertelement <4 x i32> [[TMP115]], i32 [[TMP119]], i64 3
; IC8-NEXT: br label %[[PRED_LOAD_CONTINUE37]]
; IC8: [[PRED_LOAD_CONTINUE37]]:
; IC8-NEXT: [[TMP121:%.*]] = phi <4 x i32> [ [[TMP115]], %[[PRED_LOAD_CONTINUE35]] ], [ [[TMP120]], %[[PRED_LOAD_IF36]] ]
; IC8-NEXT: [[TMP122:%.*]] = extractelement <4 x i1> [[TMP14]], i64 0
; IC8-NEXT: br i1 [[TMP122]], label %[[PRED_LOAD_IF38:.*]], label %[[PRED_LOAD_CONTINUE39:.*]]
; IC8: [[PRED_LOAD_IF38]]:
; IC8-NEXT: [[TMP123:%.*]] = extractelement <4 x i32> [[TMP22]], i64 0
; IC8-NEXT: [[TMP124:%.*]] = getelementptr inbounds i32, ptr [[A]], i32 [[TMP123]]
; IC8-NEXT: [[TMP125:%.*]] = load i32, ptr [[TMP124]], align 4
; IC8-NEXT: [[TMP126:%.*]] = insertelement <4 x i32> poison, i32 [[TMP125]], i64 0
; IC8-NEXT: br label %[[PRED_LOAD_CONTINUE39]]
; IC8: [[PRED_LOAD_CONTINUE39]]:
; IC8-NEXT: [[TMP127:%.*]] = phi <4 x i32> [ poison, %[[PRED_LOAD_CONTINUE37]] ], [ [[TMP126]], %[[PRED_LOAD_IF38]] ]
; IC8-NEXT: [[TMP128:%.*]] = extractelement <4 x i1> [[TMP14]], i64 1
; IC8-NEXT: br i1 [[TMP128]], label %[[PRED_LOAD_IF40:.*]], label %[[PRED_LOAD_CONTINUE41:.*]]
; IC8: [[PRED_LOAD_IF40]]:
; IC8-NEXT: [[TMP129:%.*]] = extractelement <4 x i32> [[TMP22]], i64 1
; IC8-NEXT: [[TMP130:%.*]] = getelementptr inbounds i32, ptr [[A]], i32 [[TMP129]]
; IC8-NEXT: [[TMP131:%.*]] = load i32, ptr [[TMP130]], align 4
; IC8-NEXT: [[TMP132:%.*]] = insertelement <4 x i32> [[TMP127]], i32 [[TMP131]], i64 1
; IC8-NEXT: br label %[[PRED_LOAD_CONTINUE41]]
; IC8: [[PRED_LOAD_CONTINUE41]]:
; IC8-NEXT: [[TMP133:%.*]] = phi <4 x i32> [ [[TMP127]], %[[PRED_LOAD_CONTINUE39]] ], [ [[TMP132]], %[[PRED_LOAD_IF40]] ]
; IC8-NEXT: [[TMP134:%.*]] = extractelement <4 x i1> [[TMP14]], i64 2
; IC8-NEXT: br i1 [[TMP134]], label %[[PRED_LOAD_IF42:.*]], label %[[PRED_LOAD_CONTINUE43:.*]]
; IC8: [[PRED_LOAD_IF42]]:
; IC8-NEXT: [[TMP135:%.*]] = extractelement <4 x i32> [[TMP22]], i64 2
; IC8-NEXT: [[TMP136:%.*]] = getelementptr inbounds i32, ptr [[A]], i32 [[TMP135]]
; IC8-NEXT: [[TMP137:%.*]] = load i32, ptr [[TMP136]], align 4
; IC8-NEXT: [[TMP138:%.*]] = insertelement <4 x i32> [[TMP133]], i32 [[TMP137]], i64 2
; IC8-NEXT: br label %[[PRED_LOAD_CONTINUE43]]
; IC8: [[PRED_LOAD_CONTINUE43]]:
; IC8-NEXT: [[TMP139:%.*]] = phi <4 x i32> [ [[TMP133]], %[[PRED_LOAD_CONTINUE41]] ], [ [[TMP138]], %[[PRED_LOAD_IF42]] ]
; IC8-NEXT: [[TMP140:%.*]] = extractelement <4 x i1> [[TMP14]], i64 3
; IC8-NEXT: br i1 [[TMP140]], label %[[PRED_LOAD_IF44:.*]], label %[[PRED_LOAD_CONTINUE45:.*]]
; IC8: [[PRED_LOAD_IF44]]:
; IC8-NEXT: [[TMP141:%.*]] = extractelement <4 x i32> [[TMP22]], i64 3
; IC8-NEXT: [[TMP142:%.*]] = getelementptr inbounds i32, ptr [[A]], i32 [[TMP141]]
; IC8-NEXT: [[TMP143:%.*]] = load i32, ptr [[TMP142]], align 4
; IC8-NEXT: [[TMP144:%.*]] = insertelement <4 x i32> [[TMP139]], i32 [[TMP143]], i64 3
; IC8-NEXT: br label %[[PRED_LOAD_CONTINUE45]]
; IC8: [[PRED_LOAD_CONTINUE45]]:
; IC8-NEXT: [[TMP145:%.*]] = phi <4 x i32> [ [[TMP139]], %[[PRED_LOAD_CONTINUE43]] ], [ [[TMP144]], %[[PRED_LOAD_IF44]] ]
; IC8-NEXT: [[TMP146:%.*]] = extractelement <4 x i1> [[TMP15]], i64 0
; IC8-NEXT: br i1 [[TMP146]], label %[[PRED_LOAD_IF46:.*]], label %[[PRED_LOAD_CONTINUE47:.*]]
; IC8: [[PRED_LOAD_IF46]]:
; IC8-NEXT: [[TMP147:%.*]] = extractelement <4 x i32> [[TMP23]], i64 0
; IC8-NEXT: [[TMP148:%.*]] = getelementptr inbounds i32, ptr [[A]], i32 [[TMP147]]
; IC8-NEXT: [[TMP149:%.*]] = load i32, ptr [[TMP148]], align 4
; IC8-NEXT: [[TMP150:%.*]] = insertelement <4 x i32> poison, i32 [[TMP149]], i64 0
; IC8-NEXT: br label %[[PRED_LOAD_CONTINUE47]]
; IC8: [[PRED_LOAD_CONTINUE47]]:
; IC8-NEXT: [[TMP151:%.*]] = phi <4 x i32> [ poison, %[[PRED_LOAD_CONTINUE45]] ], [ [[TMP150]], %[[PRED_LOAD_IF46]] ]
; IC8-NEXT: [[TMP152:%.*]] = extractelement <4 x i1> [[TMP15]], i64 1
; IC8-NEXT: br i1 [[TMP152]], label %[[PRED_LOAD_IF48:.*]], label %[[PRED_LOAD_CONTINUE49:.*]]
; IC8: [[PRED_LOAD_IF48]]:
; IC8-NEXT: [[TMP153:%.*]] = extractelement <4 x i32> [[TMP23]], i64 1
; IC8-NEXT: [[TMP154:%.*]] = getelementptr inbounds i32, ptr [[A]], i32 [[TMP153]]
; IC8-NEXT: [[TMP155:%.*]] = load i32, ptr [[TMP154]], align 4
; IC8-NEXT: [[TMP156:%.*]] = insertelement <4 x i32> [[TMP151]], i32 [[TMP155]], i64 1
; IC8-NEXT: br label %[[PRED_LOAD_CONTINUE49]]
; IC8: [[PRED_LOAD_CONTINUE49]]:
; IC8-NEXT: [[TMP157:%.*]] = phi <4 x i32> [ [[TMP151]], %[[PRED_LOAD_CONTINUE47]] ], [ [[TMP156]], %[[PRED_LOAD_IF48]] ]
; IC8-NEXT: [[TMP158:%.*]] = extractelement <4 x i1> [[TMP15]], i64 2
; IC8-NEXT: br i1 [[TMP158]], label %[[PRED_LOAD_IF50:.*]], label %[[PRED_LOAD_CONTINUE51:.*]]
; IC8: [[PRED_LOAD_IF50]]:
; IC8-NEXT: [[TMP159:%.*]] = extractelement <4 x i32> [[TMP23]], i64 2
; IC8-NEXT: [[TMP160:%.*]] = getelementptr inbounds i32, ptr [[A]], i32 [[TMP159]]
; IC8-NEXT: [[TMP161:%.*]] = load i32, ptr [[TMP160]], align 4
; IC8-NEXT: [[TMP162:%.*]] = insertelement <4 x i32> [[TMP157]], i32 [[TMP161]], i64 2
; IC8-NEXT: br label %[[PRED_LOAD_CONTINUE51]]
; IC8: [[PRED_LOAD_CONTINUE51]]:
; IC8-NEXT: [[TMP163:%.*]] = phi <4 x i32> [ [[TMP157]], %[[PRED_LOAD_CONTINUE49]] ], [ [[TMP162]], %[[PRED_LOAD_IF50]] ]
; IC8-NEXT: [[TMP164:%.*]] = extractelement <4 x i1> [[TMP15]], i64 3
; IC8-NEXT: br i1 [[TMP164]], label %[[PRED_LOAD_IF52:.*]], label %[[PRED_LOAD_CONTINUE53:.*]]
; IC8: [[PRED_LOAD_IF52]]:
; IC8-NEXT: [[TMP165:%.*]] = extractelement <4 x i32> [[TMP23]], i64 3
; IC8-NEXT: [[TMP166:%.*]] = getelementptr inbounds i32, ptr [[A]], i32 [[TMP165]]
; IC8-NEXT: [[TMP167:%.*]] = load i32, ptr [[TMP166]], align 4
; IC8-NEXT: [[TMP168:%.*]] = insertelement <4 x i32> [[TMP163]], i32 [[TMP167]], i64 3
; IC8-NEXT: br label %[[PRED_LOAD_CONTINUE53]]
; IC8: [[PRED_LOAD_CONTINUE53]]:
; IC8-NEXT: [[TMP169:%.*]] = phi <4 x i32> [ [[TMP163]], %[[PRED_LOAD_CONTINUE51]] ], [ [[TMP168]], %[[PRED_LOAD_IF52]] ]
; IC8-NEXT: [[TMP170:%.*]] = extractelement <4 x i1> [[TMP16]], i64 0
; IC8-NEXT: br i1 [[TMP170]], label %[[PRED_LOAD_IF54:.*]], label %[[PRED_LOAD_CONTINUE55:.*]]
; IC8: [[PRED_LOAD_IF54]]:
; IC8-NEXT: [[TMP171:%.*]] = extractelement <4 x i32> [[TMP24]], i64 0
; IC8-NEXT: [[TMP172:%.*]] = getelementptr inbounds i32, ptr [[A]], i32 [[TMP171]]
; IC8-NEXT: [[TMP173:%.*]] = load i32, ptr [[TMP172]], align 4
; IC8-NEXT: [[TMP174:%.*]] = insertelement <4 x i32> poison, i32 [[TMP173]], i64 0
; IC8-NEXT: br label %[[PRED_LOAD_CONTINUE55]]
; IC8: [[PRED_LOAD_CONTINUE55]]:
; IC8-NEXT: [[TMP175:%.*]] = phi <4 x i32> [ poison, %[[PRED_LOAD_CONTINUE53]] ], [ [[TMP174]], %[[PRED_LOAD_IF54]] ]
; IC8-NEXT: [[TMP176:%.*]] = extractelement <4 x i1> [[TMP16]], i64 1
; IC8-NEXT: br i1 [[TMP176]], label %[[PRED_LOAD_IF56:.*]], label %[[PRED_LOAD_CONTINUE57:.*]]
; IC8: [[PRED_LOAD_IF56]]:
; IC8-NEXT: [[TMP177:%.*]] = extractelement <4 x i32> [[TMP24]], i64 1
; IC8-NEXT: [[TMP178:%.*]] = getelementptr inbounds i32, ptr [[A]], i32 [[TMP177]]
; IC8-NEXT: [[TMP179:%.*]] = load i32, ptr [[TMP178]], align 4
; IC8-NEXT: [[TMP180:%.*]] = insertelement <4 x i32> [[TMP175]], i32 [[TMP179]], i64 1
; IC8-NEXT: br label %[[PRED_LOAD_CONTINUE57]]
; IC8: [[PRED_LOAD_CONTINUE57]]:
; IC8-NEXT: [[TMP181:%.*]] = phi <4 x i32> [ [[TMP175]], %[[PRED_LOAD_CONTINUE55]] ], [ [[TMP180]], %[[PRED_LOAD_IF56]] ]
; IC8-NEXT: [[TMP182:%.*]] = extractelement <4 x i1> [[TMP16]], i64 2
; IC8-NEXT: br i1 [[TMP182]], label %[[PRED_LOAD_IF58:.*]], label %[[PRED_LOAD_CONTINUE59:.*]]
; IC8: [[PRED_LOAD_IF58]]:
; IC8-NEXT: [[TMP183:%.*]] = extractelement <4 x i32> [[TMP24]], i64 2
; IC8-NEXT: [[TMP184:%.*]] = getelementptr inbounds i32, ptr [[A]], i32 [[TMP183]]
; IC8-NEXT: [[TMP185:%.*]] = load i32, ptr [[TMP184]], align 4
; IC8-NEXT: [[TMP186:%.*]] = insertelement <4 x i32> [[TMP181]], i32 [[TMP185]], i64 2
; IC8-NEXT: br label %[[PRED_LOAD_CONTINUE59]]
; IC8: [[PRED_LOAD_CONTINUE59]]:
; IC8-NEXT: [[TMP187:%.*]] = phi <4 x i32> [ [[TMP181]], %[[PRED_LOAD_CONTINUE57]] ], [ [[TMP186]], %[[PRED_LOAD_IF58]] ]
; IC8-NEXT: [[TMP188:%.*]] = extractelement <4 x i1> [[TMP16]], i64 3
; IC8-NEXT: br i1 [[TMP188]], label %[[PRED_LOAD_IF60:.*]], label %[[PRED_LOAD_CONTINUE61:.*]]
; IC8: [[PRED_LOAD_IF60]]:
; IC8-NEXT: [[TMP189:%.*]] = extractelement <4 x i32> [[TMP24]], i64 3
; IC8-NEXT: [[TMP190:%.*]] = getelementptr inbounds i32, ptr [[A]], i32 [[TMP189]]
; IC8-NEXT: [[TMP191:%.*]] = load i32, ptr [[TMP190]], align 4
; IC8-NEXT: [[TMP192:%.*]] = insertelement <4 x i32> [[TMP187]], i32 [[TMP191]], i64 3
; IC8-NEXT: br label %[[PRED_LOAD_CONTINUE61]]
; IC8: [[PRED_LOAD_CONTINUE61]]:
; IC8-NEXT: [[TMP193:%.*]] = phi <4 x i32> [ [[TMP187]], %[[PRED_LOAD_CONTINUE59]] ], [ [[TMP192]], %[[PRED_LOAD_IF60]] ]
; IC8-NEXT: [[TMP194:%.*]] = extractelement <4 x i1> [[TMP17]], i64 0
; IC8-NEXT: br i1 [[TMP194]], label %[[PRED_LOAD_IF62:.*]], label %[[PRED_LOAD_CONTINUE63:.*]]
; IC8: [[PRED_LOAD_IF62]]:
; IC8-NEXT: [[TMP195:%.*]] = extractelement <4 x i32> [[TMP25]], i64 0
; IC8-NEXT: [[TMP196:%.*]] = getelementptr inbounds i32, ptr [[A]], i32 [[TMP195]]
; IC8-NEXT: [[TMP197:%.*]] = load i32, ptr [[TMP196]], align 4
; IC8-NEXT: [[TMP198:%.*]] = insertelement <4 x i32> poison, i32 [[TMP197]], i64 0
; IC8-NEXT: br label %[[PRED_LOAD_CONTINUE63]]
; IC8: [[PRED_LOAD_CONTINUE63]]:
; IC8-NEXT: [[TMP199:%.*]] = phi <4 x i32> [ poison, %[[PRED_LOAD_CONTINUE61]] ], [ [[TMP198]], %[[PRED_LOAD_IF62]] ]
; IC8-NEXT: [[TMP200:%.*]] = extractelement <4 x i1> [[TMP17]], i64 1
; IC8-NEXT: br i1 [[TMP200]], label %[[PRED_LOAD_IF64:.*]], label %[[PRED_LOAD_CONTINUE65:.*]]
; IC8: [[PRED_LOAD_IF64]]:
; IC8-NEXT: [[TMP201:%.*]] = extractelement <4 x i32> [[TMP25]], i64 1
; IC8-NEXT: [[TMP202:%.*]] = getelementptr inbounds i32, ptr [[A]], i32 [[TMP201]]
; IC8-NEXT: [[TMP203:%.*]] = load i32, ptr [[TMP202]], align 4
; IC8-NEXT: [[TMP204:%.*]] = insertelement <4 x i32> [[TMP199]], i32 [[TMP203]], i64 1
; IC8-NEXT: br label %[[PRED_LOAD_CONTINUE65]]
; IC8: [[PRED_LOAD_CONTINUE65]]:
; IC8-NEXT: [[TMP205:%.*]] = phi <4 x i32> [ [[TMP199]], %[[PRED_LOAD_CONTINUE63]] ], [ [[TMP204]], %[[PRED_LOAD_IF64]] ]
; IC8-NEXT: [[TMP206:%.*]] = extractelement <4 x i1> [[TMP17]], i64 2
; IC8-NEXT: br i1 [[TMP206]], label %[[PRED_LOAD_IF66:.*]], label %[[PRED_LOAD_CONTINUE67:.*]]
; IC8: [[PRED_LOAD_IF66]]:
; IC8-NEXT: [[TMP207:%.*]] = extractelement <4 x i32> [[TMP25]], i64 2
; IC8-NEXT: [[TMP208:%.*]] = getelementptr inbounds i32, ptr [[A]], i32 [[TMP207]]
; IC8-NEXT: [[TMP209:%.*]] = load i32, ptr [[TMP208]], align 4
; IC8-NEXT: [[TMP210:%.*]] = insertelement <4 x i32> [[TMP205]], i32 [[TMP209]], i64 2
; IC8-NEXT: br label %[[PRED_LOAD_CONTINUE67]]
; IC8: [[PRED_LOAD_CONTINUE67]]:
; IC8-NEXT: [[TMP211:%.*]] = phi <4 x i32> [ [[TMP205]], %[[PRED_LOAD_CONTINUE65]] ], [ [[TMP210]], %[[PRED_LOAD_IF66]] ]
; IC8-NEXT: [[TMP212:%.*]] = extractelement <4 x i1> [[TMP17]], i64 3
; IC8-NEXT: br i1 [[TMP212]], label %[[PRED_LOAD_IF68:.*]], label %[[PRED_LOAD_CONTINUE69]]
; IC8: [[PRED_LOAD_IF68]]:
; IC8-NEXT: [[TMP213:%.*]] = extractelement <4 x i32> [[TMP25]], i64 3
; IC8-NEXT: [[TMP214:%.*]] = getelementptr inbounds i32, ptr [[A]], i32 [[TMP213]]
; IC8-NEXT: [[TMP215:%.*]] = load i32, ptr [[TMP214]], align 4
; IC8-NEXT: [[TMP216:%.*]] = insertelement <4 x i32> [[TMP211]], i32 [[TMP215]], i64 3
; IC8-NEXT: br label %[[PRED_LOAD_CONTINUE69]]
; IC8: [[PRED_LOAD_CONTINUE69]]:
; IC8-NEXT: [[TMP217:%.*]] = phi <4 x i32> [ [[TMP211]], %[[PRED_LOAD_CONTINUE67]] ], [ [[TMP216]], %[[PRED_LOAD_IF68]] ]
; IC8-NEXT: [[PREDPHI:%.*]] = select <4 x i1> [[TMP10]], <4 x i32> [[TMP49]], <4 x i32> zeroinitializer
; IC8-NEXT: [[PREDPHI70:%.*]] = select <4 x i1> [[TMP11]], <4 x i32> [[TMP73]], <4 x i32> zeroinitializer
; IC8-NEXT: [[PREDPHI71:%.*]] = select <4 x i1> [[TMP12]], <4 x i32> [[TMP97]], <4 x i32> zeroinitializer
; IC8-NEXT: [[PREDPHI72:%.*]] = select <4 x i1> [[TMP13]], <4 x i32> [[TMP121]], <4 x i32> zeroinitializer
; IC8-NEXT: [[PREDPHI73:%.*]] = select <4 x i1> [[TMP14]], <4 x i32> [[TMP145]], <4 x i32> zeroinitializer
; IC8-NEXT: [[PREDPHI74:%.*]] = select <4 x i1> [[TMP15]], <4 x i32> [[TMP169]], <4 x i32> zeroinitializer
; IC8-NEXT: [[PREDPHI75:%.*]] = select <4 x i1> [[TMP16]], <4 x i32> [[TMP193]], <4 x i32> zeroinitializer
; IC8-NEXT: [[PREDPHI76:%.*]] = select <4 x i1> [[TMP17]], <4 x i32> [[TMP217]], <4 x i32> zeroinitializer
; IC8-NEXT: [[TMP218]] = add <4 x i32> [[VEC_PHI]], [[PREDPHI]]
; IC8-NEXT: [[TMP219]] = add <4 x i32> [[VEC_PHI1]], [[PREDPHI70]]
; IC8-NEXT: [[TMP220]] = add <4 x i32> [[VEC_PHI2]], [[PREDPHI71]]
; IC8-NEXT: [[TMP221]] = add <4 x i32> [[VEC_PHI3]], [[PREDPHI72]]
; IC8-NEXT: [[TMP222]] = add <4 x i32> [[VEC_PHI4]], [[PREDPHI73]]
; IC8-NEXT: [[TMP223]] = add <4 x i32> [[VEC_PHI5]], [[PREDPHI74]]
; IC8-NEXT: [[TMP224]] = add <4 x i32> [[VEC_PHI6]], [[PREDPHI75]]
; IC8-NEXT: [[TMP225]] = add <4 x i32> [[VEC_PHI7]], [[PREDPHI76]]
; IC8-NEXT: [[INDEX_NEXT]] = add nuw i32 [[INDEX]], 32
; IC8-NEXT: [[VEC_IND_NEXT]] = add <4 x i32> [[STEP_ADD_7]], splat (i32 4)
; IC8-NEXT: [[TMP226:%.*]] = icmp eq i32 [[INDEX_NEXT]], [[N_VEC]]
; IC8-NEXT: br i1 [[TMP226]], label %[[MIDDLE_BLOCK:.*]], label %[[VECTOR_BODY]], !llvm.loop [[LOOP4:![0-9]+]]
; IC8: [[MIDDLE_BLOCK]]:
; IC8-NEXT: [[BIN_RDX:%.*]] = add <4 x i32> [[TMP219]], [[TMP218]]
; IC8-NEXT: [[BIN_RDX77:%.*]] = add <4 x i32> [[TMP220]], [[BIN_RDX]]
; IC8-NEXT: [[BIN_RDX78:%.*]] = add <4 x i32> [[TMP221]], [[BIN_RDX77]]
; IC8-NEXT: [[BIN_RDX79:%.*]] = add <4 x i32> [[TMP222]], [[BIN_RDX78]]
; IC8-NEXT: [[BIN_RDX80:%.*]] = add <4 x i32> [[TMP223]], [[BIN_RDX79]]
; IC8-NEXT: [[BIN_RDX81:%.*]] = add <4 x i32> [[TMP224]], [[BIN_RDX80]]
; IC8-NEXT: [[BIN_RDX82:%.*]] = add <4 x i32> [[TMP225]], [[BIN_RDX81]]
; IC8-NEXT: [[TMP227:%.*]] = call i32 @llvm.vector.reduce.add.v4i32(<4 x i32> [[BIN_RDX82]])
; IC8-NEXT: [[CMP_N:%.*]] = icmp eq i32 [[N]], [[N_VEC]]
; IC8-NEXT: br i1 [[CMP_N]], label %[[EXIT:.*]], label %[[SCALAR_PH]]
; IC8: [[SCALAR_PH]]:
; IC8-NEXT: [[BC_RESUME_VAL:%.*]] = phi i32 [ [[N_VEC]], %[[MIDDLE_BLOCK]] ], [ 0, %[[ENTRY]] ], [ 0, %[[VECTOR_SCEVCHECK]] ]
; IC8-NEXT: [[BC_MERGE_RDX:%.*]] = phi i32 [ [[TMP227]], %[[MIDDLE_BLOCK]] ], [ 0, %[[ENTRY]] ], [ 0, %[[VECTOR_SCEVCHECK]] ]
; IC8-NEXT: br label %[[LOOP:.*]]
; IC8: [[LOOP]]:
; IC8-NEXT: [[IV:%.*]] = phi i32 [ [[BC_RESUME_VAL]], %[[SCALAR_PH]] ], [ [[IV_NEXT:%.*]], %[[LATCH:.*]] ]
; IC8-NEXT: [[SUM:%.*]] = phi i32 [ [[BC_MERGE_RDX]], %[[SCALAR_PH]] ], [ [[SUM_NEXT:%.*]], %[[LATCH]] ]
; IC8-NEXT: [[AND_IV:%.*]] = and i32 [[IV]], 1
; IC8-NEXT: [[CMP:%.*]] = icmp eq i32 [[AND_IV]], 0
; IC8-NEXT: br i1 [[CMP]], label %[[THEN:.*]], label %[[LATCH]]
; IC8: [[THEN]]:
; IC8-NEXT: [[CLAMPED:%.*]] = urem i32 [[IV]], 16
; IC8-NEXT: [[GEP_A:%.*]] = getelementptr inbounds i32, ptr [[A]], i32 [[CLAMPED]]
; IC8-NEXT: [[LV:%.*]] = load i32, ptr [[GEP_A]], align 4
; IC8-NEXT: br label %[[LATCH]]
; IC8: [[LATCH]]:
; IC8-NEXT: [[V:%.*]] = phi i32 [ [[LV]], %[[THEN]] ], [ 0, %[[LOOP]] ]
; IC8-NEXT: [[SUM_NEXT]] = add i32 [[SUM]], [[V]]
; IC8-NEXT: [[IV_NEXT]] = add i32 [[IV]], 1
; IC8-NEXT: [[EC:%.*]] = icmp eq i32 [[IV_NEXT]], [[N]]
; IC8-NEXT: br i1 [[EC]], label %[[EXIT]], label %[[LOOP]], !llvm.loop [[LOOP5:![0-9]+]]
; IC8: [[EXIT]]:
; IC8-NEXT: [[SUM_NEXT_LCSSA:%.*]] = phi i32 [ [[SUM_NEXT]], %[[LATCH]] ], [ [[TMP227]], %[[MIDDLE_BLOCK]] ]
; IC8-NEXT: ret i32 [[SUM_NEXT_LCSSA]]
;
entry:
br label %loop
loop:
%iv = phi i32 [ 0, %entry ], [ %iv.next, %latch ]
%sum = phi i32 [ 0, %entry ], [ %sum.next, %latch ]
%and.iv = and i32 %iv, 1
%cmp = icmp eq i32 %and.iv, 0
br i1 %cmp, label %then, label %latch
then:
%clamped = urem i32 %iv, 16
%gep.A = getelementptr inbounds i32, ptr %A, i32 %clamped
%lv = load i32, ptr %gep.A
br label %latch
latch:
%v = phi i32 [ %lv, %then ], [ 0, %loop ]
%sum.next = add i32 %sum, %v
%iv.next = add i32 %iv, 1
%ec = icmp eq i32 %iv.next, %n
br i1 %ec, label %exit, label %loop
exit:
ret i32 %sum.next
}
;.
; 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]], [[META1]]}
; CHECK: [[LOOP4]] = distinct !{[[LOOP4]], [[META1]], [[META2]]}
; CHECK: [[LOOP5]] = distinct !{[[LOOP5]], [[META1]]}
;.
; IC4: [[LOOP0]] = distinct !{[[LOOP0]], [[META1:![0-9]+]], [[META2:![0-9]+]]}
; IC4: [[META1]] = !{!"llvm.loop.isvectorized", i32 1}
; IC4: [[META2]] = !{!"llvm.loop.unroll.runtime.disable"}
; IC4: [[LOOP3]] = distinct !{[[LOOP3]], [[META1]]}
; IC4: [[LOOP4]] = distinct !{[[LOOP4]], [[META1]], [[META2]]}
; IC4: [[LOOP5]] = distinct !{[[LOOP5]], [[META1]]}
;.
; IC8: [[LOOP0]] = distinct !{[[LOOP0]], [[META1:![0-9]+]], [[META2:![0-9]+]]}
; IC8: [[META1]] = !{!"llvm.loop.isvectorized", i32 1}
; IC8: [[META2]] = !{!"llvm.loop.unroll.runtime.disable"}
; IC8: [[LOOP3]] = distinct !{[[LOOP3]], [[META1]]}
; IC8: [[LOOP4]] = distinct !{[[LOOP4]], [[META1]], [[META2]]}
; IC8: [[LOOP5]] = distinct !{[[LOOP5]], [[META1]]}
;.