; NOTE: Assertions have been autogenerated by utils/update_test_checks.py
; RUN: opt -indvars -loop-unroll -mtriple=thumbv8m.main %s -S -o - | FileCheck %s

define dso_local arm_aapcscc void @test(i32* nocapture %pDest, i16* nocapture readonly %pSrcA, i16* nocapture readonly %pSrcB, i32 %blkCnt) local_unnamed_addr #0 {
; CHECK-LABEL: @test(
; CHECK-NEXT:  entry:
; CHECK-NEXT:    [[CMP88:%.*]] = icmp eq i32 [[BLKCNT:%.*]], 0
; CHECK-NEXT:    br i1 [[CMP88]], label [[FOR_COND_CLEANUP:%.*]], label [[FOR_BODY_PREHEADER:%.*]]
; CHECK:       for.body.preheader:
; CHECK-NEXT:    br label [[FOR_BODY:%.*]]
; CHECK:       for.cond.cleanup.loopexit:
; CHECK-NEXT:    br label [[FOR_COND_CLEANUP]]
; CHECK:       for.cond.cleanup:
; CHECK-NEXT:    ret void
; CHECK:       for.body:
; CHECK-NEXT:    [[I_092:%.*]] = phi i32 [ [[INC42:%.*]], [[FOR_END40:%.*]] ], [ 0, [[FOR_BODY_PREHEADER]] ]
; CHECK-NEXT:    [[PDEST_ADDR_091:%.*]] = phi i32* [ [[PDEST_ADDR_2_LCSSA:%.*]], [[FOR_END40]] ], [ [[PDEST:%.*]], [[FOR_BODY_PREHEADER]] ]
; CHECK-NEXT:    [[PSRCA_ADDR_090:%.*]] = phi i16* [ [[PSRCA_ADDR_2_LCSSA:%.*]], [[FOR_END40]] ], [ [[PSRCA:%.*]], [[FOR_BODY_PREHEADER]] ]
; CHECK-NEXT:    [[PSRCB_ADDR_089:%.*]] = phi i16* [ [[PSRCB_ADDR_2_LCSSA:%.*]], [[FOR_END40]] ], [ [[PSRCB:%.*]], [[FOR_BODY_PREHEADER]] ]
; CHECK-NEXT:    [[TMP0:%.*]] = lshr i32 [[I_092]], 2
; CHECK-NEXT:    [[TMP1:%.*]] = add nuw nsw i32 [[TMP0]], 3
; CHECK-NEXT:    [[TMP2:%.*]] = and i32 [[TMP1]], 2147483644
; CHECK-NEXT:    [[CMP272:%.*]] = icmp eq i32 [[TMP0]], 0
; CHECK-NEXT:    br i1 [[CMP272]], label [[FOR_END:%.*]], label [[FOR_BODY3_PREHEADER:%.*]]
; CHECK:       for.body3.preheader:
; CHECK-NEXT:    br label [[FOR_BODY3:%.*]]
; CHECK:       for.body3:
; CHECK-NEXT:    [[J_076:%.*]] = phi i32 [ [[ADD24:%.*]], [[FOR_BODY3]] ], [ 0, [[FOR_BODY3_PREHEADER]] ]
; CHECK-NEXT:    [[PDEST_ADDR_175:%.*]] = phi i32* [ [[INCDEC_PTR:%.*]], [[FOR_BODY3]] ], [ [[PDEST_ADDR_091]], [[FOR_BODY3_PREHEADER]] ]
; CHECK-NEXT:    [[PSRCA_ADDR_174:%.*]] = phi i16* [ [[ADD_PTR:%.*]], [[FOR_BODY3]] ], [ [[PSRCA_ADDR_090]], [[FOR_BODY3_PREHEADER]] ]
; CHECK-NEXT:    [[PSRCB_ADDR_173:%.*]] = phi i16* [ [[ADD_PTR23:%.*]], [[FOR_BODY3]] ], [ [[PSRCB_ADDR_089]], [[FOR_BODY3_PREHEADER]] ]
; CHECK-NEXT:    [[TMP3:%.*]] = load i16, i16* [[PSRCA_ADDR_174]], align 2
; CHECK-NEXT:    [[CONV:%.*]] = sext i16 [[TMP3]] to i32
; CHECK-NEXT:    [[TMP4:%.*]] = load i16, i16* [[PSRCB_ADDR_173]], align 2
; CHECK-NEXT:    [[CONV5:%.*]] = sext i16 [[TMP4]] to i32
; CHECK-NEXT:    [[MUL:%.*]] = mul nsw i32 [[CONV5]], [[CONV]]
; CHECK-NEXT:    [[ARRAYIDX6:%.*]] = getelementptr inbounds i16, i16* [[PSRCA_ADDR_174]], i32 1
; CHECK-NEXT:    [[TMP5:%.*]] = load i16, i16* [[ARRAYIDX6]], align 2
; CHECK-NEXT:    [[CONV7:%.*]] = sext i16 [[TMP5]] to i32
; CHECK-NEXT:    [[ARRAYIDX8:%.*]] = getelementptr inbounds i16, i16* [[PSRCB_ADDR_173]], i32 1
; CHECK-NEXT:    [[TMP6:%.*]] = load i16, i16* [[ARRAYIDX8]], align 2
; CHECK-NEXT:    [[CONV9:%.*]] = sext i16 [[TMP6]] to i32
; CHECK-NEXT:    [[MUL10:%.*]] = mul nsw i32 [[CONV9]], [[CONV7]]
; CHECK-NEXT:    [[ARRAYIDX11:%.*]] = getelementptr inbounds i16, i16* [[PSRCA_ADDR_174]], i32 2
; CHECK-NEXT:    [[TMP7:%.*]] = load i16, i16* [[ARRAYIDX11]], align 2
; CHECK-NEXT:    [[CONV12:%.*]] = sext i16 [[TMP7]] to i32
; CHECK-NEXT:    [[ARRAYIDX13:%.*]] = getelementptr inbounds i16, i16* [[PSRCB_ADDR_173]], i32 3
; CHECK-NEXT:    [[TMP8:%.*]] = load i16, i16* [[ARRAYIDX13]], align 2
; CHECK-NEXT:    [[CONV14:%.*]] = sext i16 [[TMP8]] to i32
; CHECK-NEXT:    [[MUL15:%.*]] = mul nsw i32 [[CONV14]], [[CONV12]]
; CHECK-NEXT:    [[ARRAYIDX17:%.*]] = getelementptr inbounds i16, i16* [[PSRCA_ADDR_174]], i32 3
; CHECK-NEXT:    [[TMP9:%.*]] = load i16, i16* [[ARRAYIDX17]], align 2
; CHECK-NEXT:    [[CONV18:%.*]] = sext i16 [[TMP9]] to i32
; CHECK-NEXT:    [[ADD21:%.*]] = add i32 [[MUL10]], [[MUL]]
; CHECK-NEXT:    [[ADD:%.*]] = add i32 [[ADD21]], [[CONV14]]
; CHECK-NEXT:    [[ADD16:%.*]] = add i32 [[ADD]], [[MUL15]]
; CHECK-NEXT:    [[ADD22:%.*]] = add i32 [[ADD16]], [[CONV18]]
; CHECK-NEXT:    store i32 [[ADD22]], i32* [[PDEST_ADDR_175]], align 4
; CHECK-NEXT:    [[ADD_PTR]] = getelementptr inbounds i16, i16* [[PSRCA_ADDR_174]], i32 4
; CHECK-NEXT:    [[ADD_PTR23]] = getelementptr inbounds i16, i16* [[PSRCB_ADDR_173]], i32 4
; CHECK-NEXT:    [[INCDEC_PTR]] = getelementptr inbounds i32, i32* [[PDEST_ADDR_175]], i32 1
; CHECK-NEXT:    [[ADD24]] = add nuw nsw i32 [[J_076]], 4
; CHECK-NEXT:    [[CMP2:%.*]] = icmp ult i32 [[ADD24]], [[TMP0]]
; CHECK-NEXT:    br i1 [[CMP2]], label [[FOR_BODY3]], label [[FOR_END_LOOPEXIT:%.*]]
; CHECK:       for.end.loopexit:
; CHECK-NEXT:    [[ADD_PTR_LCSSA:%.*]] = phi i16* [ [[ADD_PTR]], [[FOR_BODY3]] ]
; CHECK-NEXT:    [[ADD_PTR23_LCSSA:%.*]] = phi i16* [ [[ADD_PTR23]], [[FOR_BODY3]] ]
; CHECK-NEXT:    [[INCDEC_PTR_LCSSA:%.*]] = phi i32* [ [[INCDEC_PTR]], [[FOR_BODY3]] ]
; CHECK-NEXT:    br label [[FOR_END]]
; CHECK:       for.end:
; CHECK-NEXT:    [[PSRCB_ADDR_1_LCSSA:%.*]] = phi i16* [ [[PSRCB_ADDR_089]], [[FOR_BODY]] ], [ [[ADD_PTR23_LCSSA]], [[FOR_END_LOOPEXIT]] ]
; CHECK-NEXT:    [[PSRCA_ADDR_1_LCSSA:%.*]] = phi i16* [ [[PSRCA_ADDR_090]], [[FOR_BODY]] ], [ [[ADD_PTR_LCSSA]], [[FOR_END_LOOPEXIT]] ]
; CHECK-NEXT:    [[PDEST_ADDR_1_LCSSA:%.*]] = phi i32* [ [[PDEST_ADDR_091]], [[FOR_BODY]] ], [ [[INCDEC_PTR_LCSSA]], [[FOR_END_LOOPEXIT]] ]
; CHECK-NEXT:    [[J_0_LCSSA:%.*]] = phi i32 [ 0, [[FOR_BODY]] ], [ [[TMP2]], [[FOR_END_LOOPEXIT]] ]
; CHECK-NEXT:    [[REM:%.*]] = and i32 [[TMP0]], 3
; CHECK-NEXT:    [[ADD25:%.*]] = or i32 [[J_0_LCSSA]], [[REM]]
; CHECK-NEXT:    [[CMP2780:%.*]] = icmp ugt i32 [[ADD25]], [[J_0_LCSSA]]
; CHECK-NEXT:    br i1 [[CMP2780]], label [[FOR_BODY29_PREHEADER:%.*]], label [[FOR_END40]]
; CHECK:       for.body29.preheader:
; CHECK-NEXT:    [[TMP10:%.*]] = sub nsw i32 [[ADD25]], [[J_0_LCSSA]]
; CHECK-NEXT:    br label [[FOR_BODY29:%.*]]
; CHECK:       for.body29:
; CHECK-NEXT:    [[J_184:%.*]] = phi i32 [ [[INC:%.*]], [[FOR_BODY29]] ], [ [[J_0_LCSSA]], [[FOR_BODY29_PREHEADER]] ]
; CHECK-NEXT:    [[PDEST_ADDR_283:%.*]] = phi i32* [ [[INCDEC_PTR38:%.*]], [[FOR_BODY29]] ], [ [[PDEST_ADDR_1_LCSSA]], [[FOR_BODY29_PREHEADER]] ]
; CHECK-NEXT:    [[PSRCA_ADDR_282:%.*]] = phi i16* [ [[INCDEC_PTR36:%.*]], [[FOR_BODY29]] ], [ [[PSRCA_ADDR_1_LCSSA]], [[FOR_BODY29_PREHEADER]] ]
; CHECK-NEXT:    [[PSRCB_ADDR_281:%.*]] = phi i16* [ [[INCDEC_PTR37:%.*]], [[FOR_BODY29]] ], [ [[PSRCB_ADDR_1_LCSSA]], [[FOR_BODY29_PREHEADER]] ]
; CHECK-NEXT:    [[ARRAYIDX30:%.*]] = getelementptr inbounds i16, i16* [[PSRCA_ADDR_282]], i32 [[J_184]]
; CHECK-NEXT:    [[TMP11:%.*]] = load i16, i16* [[ARRAYIDX30]], align 2
; CHECK-NEXT:    [[CONV31:%.*]] = sext i16 [[TMP11]] to i32
; CHECK-NEXT:    [[ARRAYIDX32:%.*]] = getelementptr inbounds i16, i16* [[PSRCB_ADDR_281]], i32 [[J_184]]
; CHECK-NEXT:    [[TMP12:%.*]] = load i16, i16* [[ARRAYIDX32]], align 2
; CHECK-NEXT:    [[CONV33:%.*]] = sext i16 [[TMP12]] to i32
; CHECK-NEXT:    [[MUL34:%.*]] = mul nsw i32 [[CONV33]], [[CONV31]]
; CHECK-NEXT:    [[TMP13:%.*]] = load i32, i32* [[PDEST_ADDR_283]], align 4
; CHECK-NEXT:    [[ADD35:%.*]] = add nsw i32 [[MUL34]], [[TMP13]]
; CHECK-NEXT:    store i32 [[ADD35]], i32* [[PDEST_ADDR_283]], align 4
; CHECK-NEXT:    [[INCDEC_PTR36]] = getelementptr inbounds i16, i16* [[PSRCA_ADDR_282]], i32 1
; CHECK-NEXT:    [[INCDEC_PTR37]] = getelementptr inbounds i16, i16* [[PSRCB_ADDR_281]], i32 1
; CHECK-NEXT:    [[INCDEC_PTR38]] = getelementptr inbounds i32, i32* [[PDEST_ADDR_283]], i32 1
; CHECK-NEXT:    [[INC]] = add nuw i32 [[J_184]], 1
; CHECK-NEXT:    [[EXITCOND:%.*]] = icmp eq i32 [[INC]], [[ADD25]]
; CHECK-NEXT:    br i1 [[EXITCOND]], label [[FOR_END40_LOOPEXIT:%.*]], label [[FOR_BODY29]]
; CHECK:       for.end40.loopexit:
; CHECK-NEXT:    [[SCEVGEP93:%.*]] = getelementptr i16, i16* [[PSRCB_ADDR_1_LCSSA]], i32 [[TMP10]]
; CHECK-NEXT:    [[SCEVGEP:%.*]] = getelementptr i16, i16* [[PSRCA_ADDR_1_LCSSA]], i32 [[TMP10]]
; CHECK-NEXT:    [[SCEVGEP94:%.*]] = getelementptr i32, i32* [[PDEST_ADDR_1_LCSSA]], i32 [[TMP10]]
; CHECK-NEXT:    br label [[FOR_END40]]
; CHECK:       for.end40:
; CHECK-NEXT:    [[PSRCB_ADDR_2_LCSSA]] = phi i16* [ [[PSRCB_ADDR_1_LCSSA]], [[FOR_END]] ], [ [[SCEVGEP93]], [[FOR_END40_LOOPEXIT]] ]
; CHECK-NEXT:    [[PSRCA_ADDR_2_LCSSA]] = phi i16* [ [[PSRCA_ADDR_1_LCSSA]], [[FOR_END]] ], [ [[SCEVGEP]], [[FOR_END40_LOOPEXIT]] ]
; CHECK-NEXT:    [[PDEST_ADDR_2_LCSSA]] = phi i32* [ [[PDEST_ADDR_1_LCSSA]], [[FOR_END]] ], [ [[SCEVGEP94]], [[FOR_END40_LOOPEXIT]] ]
; CHECK-NEXT:    [[INC42]] = add nuw i32 [[I_092]], 1
; CHECK-NEXT:    [[EXITCOND95:%.*]] = icmp eq i32 [[INC42]], [[BLKCNT]]
; CHECK-NEXT:    br i1 [[EXITCOND95]], label [[FOR_COND_CLEANUP_LOOPEXIT:%.*]], label [[FOR_BODY]]
;
entry:
  %cmp88 = icmp eq i32 %blkCnt, 0
  br i1 %cmp88, label %for.cond.cleanup, label %for.body

for.cond.cleanup:                                 ; preds = %for.end40, %entry
  ret void

for.body:                                         ; preds = %for.end40, %entry
  %i.092 = phi i32 [ %inc42, %for.end40 ], [ 0, %entry ]
  %pDest.addr.091 = phi i32* [ %pDest.addr.2.lcssa, %for.end40 ], [ %pDest, %entry ]
  %pSrcA.addr.090 = phi i16* [ %pSrcA.addr.2.lcssa, %for.end40 ], [ %pSrcA, %entry ]
  %pSrcB.addr.089 = phi i16* [ %pSrcB.addr.2.lcssa, %for.end40 ], [ %pSrcB, %entry ]
  %0 = lshr i32 %i.092, 2
  %1 = add nuw nsw i32 %0, 3
  %2 = and i32 %1, 2147483644
  %cmp272 = icmp eq i32 %0, 0
  br i1 %cmp272, label %for.end, label %for.body3

for.body3:                                        ; preds = %for.body3, %for.body
  %j.076 = phi i32 [ %add24, %for.body3 ], [ 0, %for.body ]
  %pDest.addr.175 = phi i32* [ %incdec.ptr, %for.body3 ], [ %pDest.addr.091, %for.body ]
  %pSrcA.addr.174 = phi i16* [ %add.ptr, %for.body3 ], [ %pSrcA.addr.090, %for.body ]
  %pSrcB.addr.173 = phi i16* [ %add.ptr23, %for.body3 ], [ %pSrcB.addr.089, %for.body ]
  %3 = load i16, i16* %pSrcA.addr.174, align 2
  %conv = sext i16 %3 to i32
  %4 = load i16, i16* %pSrcB.addr.173, align 2
  %conv5 = sext i16 %4 to i32
  %mul = mul nsw i32 %conv5, %conv
  %arrayidx6 = getelementptr inbounds i16, i16* %pSrcA.addr.174, i32 1
  %5 = load i16, i16* %arrayidx6, align 2
  %conv7 = sext i16 %5 to i32
  %arrayidx8 = getelementptr inbounds i16, i16* %pSrcB.addr.173, i32 1
  %6 = load i16, i16* %arrayidx8, align 2
  %conv9 = sext i16 %6 to i32
  %mul10 = mul nsw i32 %conv9, %conv7
  %arrayidx11 = getelementptr inbounds i16, i16* %pSrcA.addr.174, i32 2
  %7 = load i16, i16* %arrayidx11, align 2
  %conv12 = sext i16 %7 to i32
  %arrayidx13 = getelementptr inbounds i16, i16* %pSrcB.addr.173, i32 3
  %8 = load i16, i16* %arrayidx13, align 2
  %conv14 = sext i16 %8 to i32
  %mul15 = mul nsw i32 %conv14, %conv12
  %arrayidx17 = getelementptr inbounds i16, i16* %pSrcA.addr.174, i32 3
  %9 = load i16, i16* %arrayidx17, align 2
  %conv18 = sext i16 %9 to i32
  %add21 = add i32 %mul10, %mul
  %add = add i32 %add21, %conv14
  %add16 = add i32 %add, %mul15
  %add22 = add i32 %add16, %conv18
  store i32 %add22, i32* %pDest.addr.175, align 4
  %add.ptr = getelementptr inbounds i16, i16* %pSrcA.addr.174, i32 4
  %add.ptr23 = getelementptr inbounds i16, i16* %pSrcB.addr.173, i32 4
  %incdec.ptr = getelementptr inbounds i32, i32* %pDest.addr.175, i32 1
  %add24 = add nuw nsw i32 %j.076, 4
  %cmp2 = icmp ult i32 %add24, %0
  br i1 %cmp2, label %for.body3, label %for.end

for.end:                                          ; preds = %for.body3, %for.body
  %pSrcB.addr.1.lcssa = phi i16* [ %pSrcB.addr.089, %for.body ], [ %add.ptr23, %for.body3 ]
  %pSrcA.addr.1.lcssa = phi i16* [ %pSrcA.addr.090, %for.body ], [ %add.ptr, %for.body3 ]
  %pDest.addr.1.lcssa = phi i32* [ %pDest.addr.091, %for.body ], [ %incdec.ptr, %for.body3 ]
  %j.0.lcssa = phi i32 [ 0, %for.body ], [ %2, %for.body3 ]
  %rem = and i32 %0, 3
  %add25 = or i32 %j.0.lcssa, %rem
  %cmp2780 = icmp ugt i32 %add25, %j.0.lcssa
  br i1 %cmp2780, label %for.body29.preheader, label %for.end40

for.body29.preheader:                             ; preds = %for.end
  %10 = sub nsw i32 %add25, %j.0.lcssa
  %scevgep93 = getelementptr i16, i16* %pSrcB.addr.1.lcssa, i32 %10
  br label %for.body29

for.body29:                                       ; preds = %for.body29, %for.body29.preheader
  %j.184 = phi i32 [ %inc, %for.body29 ], [ %j.0.lcssa, %for.body29.preheader ]
  %pDest.addr.283 = phi i32* [ %incdec.ptr38, %for.body29 ], [ %pDest.addr.1.lcssa, %for.body29.preheader ]
  %pSrcA.addr.282 = phi i16* [ %incdec.ptr36, %for.body29 ], [ %pSrcA.addr.1.lcssa, %for.body29.preheader ]
  %pSrcB.addr.281 = phi i16* [ %incdec.ptr37, %for.body29 ], [ %pSrcB.addr.1.lcssa, %for.body29.preheader ]
  %arrayidx30 = getelementptr inbounds i16, i16* %pSrcA.addr.282, i32 %j.184
  %11 = load i16, i16* %arrayidx30, align 2
  %conv31 = sext i16 %11 to i32
  %arrayidx32 = getelementptr inbounds i16, i16* %pSrcB.addr.281, i32 %j.184
  %12 = load i16, i16* %arrayidx32, align 2
  %conv33 = sext i16 %12 to i32
  %mul34 = mul nsw i32 %conv33, %conv31
  %13 = load i32, i32* %pDest.addr.283, align 4
  %add35 = add nsw i32 %mul34, %13
  store i32 %add35, i32* %pDest.addr.283, align 4
  %incdec.ptr36 = getelementptr inbounds i16, i16* %pSrcA.addr.282, i32 1
  %incdec.ptr37 = getelementptr inbounds i16, i16* %pSrcB.addr.281, i32 1
  %incdec.ptr38 = getelementptr inbounds i32, i32* %pDest.addr.283, i32 1
  %inc = add nuw i32 %j.184, 1
  %exitcond = icmp eq i32 %inc, %add25
  br i1 %exitcond, label %for.end40.loopexit, label %for.body29

for.end40.loopexit:                               ; preds = %for.body29
  %scevgep = getelementptr i16, i16* %pSrcA.addr.1.lcssa, i32 %10
  %scevgep94 = getelementptr i32, i32* %pDest.addr.1.lcssa, i32 %10
  br label %for.end40

for.end40:                                        ; preds = %for.end40.loopexit, %for.end
  %pSrcB.addr.2.lcssa = phi i16* [ %pSrcB.addr.1.lcssa, %for.end ], [ %scevgep93, %for.end40.loopexit ]
  %pSrcA.addr.2.lcssa = phi i16* [ %pSrcA.addr.1.lcssa, %for.end ], [ %scevgep, %for.end40.loopexit ]
  %pDest.addr.2.lcssa = phi i32* [ %pDest.addr.1.lcssa, %for.end ], [ %scevgep94, %for.end40.loopexit ]
  %inc42 = add nuw i32 %i.092, 1
  %exitcond95 = icmp eq i32 %inc42, %blkCnt
  br i1 %exitcond95, label %for.cond.cleanup, label %for.body
}

attributes #0 = { minsize optsize }