; ModuleID = 'square_3D_mat'
source_filename = "square_3D_mat"
target datalayout = "e-m:e-i64:64-f80:128-n8:16:32:64-S128"
target triple = "x86_64-unknown-linux-gnu"

@kernel_0 = private unnamed_addr constant [1 x i8] zeroinitializer
@kernel_0_name = private unnamed_addr constant [9 x i8] c"kernel_0\00"

; Function Attrs: sspstrong
define void @julia_square_3D_mat_64454(i8** dereferenceable(40), i8** dereferenceable(40)) #0 !dbg !5 {
top:
  %polly_launch_0_params = alloca [8 x i8*]
  %polly_launch_0_param_0 = alloca i8*
  %polly_launch_0_param_1 = alloca i8*
  %polly_launch_0_param_2 = alloca i64
  %polly_launch_0_param_3 = alloca i64
  %polly_launch_0_param_4 = alloca i64
  %polly_launch_0_param_5 = alloca i64
  %polly_launch_0_param_6 = alloca i64
  %polly_launch_0_param_7 = alloca void (metadata, i64, metadata, metadata)*
  %polly_launch_0_params_i8ptr = bitcast [8 x i8*]* %polly_launch_0_params to i8*
  br label %top.split, !dbg !24

top.split:                                        ; preds = %top
  call void @llvm.dbg.value(metadata i8** null, i64 0, metadata !15, metadata !25), !dbg !24
  call void @llvm.dbg.value(metadata i8** null, i64 0, metadata !16, metadata !25), !dbg !24
  call void @llvm.dbg.value(metadata i8** %0, i64 0, metadata !15, metadata !25), !dbg !24
  %2 = bitcast i8** %0 to double**
  %3 = load double*, double** %2, align 8, !tbaa !26
  %4 = getelementptr i8*, i8** %0, i64 3
  %5 = bitcast i8** %4 to i64*
  %6 = load i64, i64* %5, align 8, !tbaa !26
  %7 = getelementptr i8*, i8** %0, i64 4
  %8 = bitcast i8** %7 to i64*
  %9 = load i64, i64* %8, align 8, !tbaa !26
  call void @llvm.dbg.value(metadata i8** %1, i64 0, metadata !16, metadata !25), !dbg !24
  %10 = bitcast i8** %1 to double**
  %11 = load double*, double** %10, align 8, !tbaa !26
  %12 = getelementptr i8*, i8** %1, i64 3
  %13 = bitcast i8** %12 to i64*
  %14 = load i64, i64* %13, align 8, !tbaa !26
  %15 = getelementptr i8*, i8** %1, i64 4
  %16 = bitcast i8** %15 to i64*
  %17 = load i64, i64* %16, align 8, !tbaa !26
  %18 = getelementptr i8*, i8** %0, i64 5, !dbg !29
  %19 = bitcast i8** %18 to i64*, !dbg !29
  %20 = load i64, i64* %19, align 8, !dbg !29, !tbaa !26
  call void @llvm.dbg.value(metadata i64 %6, i64 0, metadata !21, metadata !25), !dbg !24
  call void @llvm.dbg.value(metadata i64 %6, i64 0, metadata !21, metadata !25), !dbg !24
  %21 = icmp sgt i64 %6, 0, !dbg !39
  %22 = select i1 %21, i64 %6, i64 0, !dbg !39
  %23 = icmp slt i64 %6, 1, !dbg !39
  br i1 %23, label %L55, label %if.lr.ph, !dbg !39

if.lr.ph:                                         ; preds = %top.split
  br label %polly.split_new_and_old, !dbg !39

L31.L22.loopexit_crit_edge:                       ; preds = %L31.loopexit
  br label %L22.loopexit, !dbg !40

L22.loopexit:                                     ; preds = %L31.L22.loopexit_crit_edge, %if
  %24 = icmp eq i64 %"#temp#2.031", %22, !dbg !39
  br i1 %24, label %L22.L55_crit_edge, label %if, !dbg !39

polly.split_new_and_old:                          ; preds = %if.lr.ph
  %25 = icmp sge i64 %17, %9
  %26 = icmp sge i64 %14, %6
  %27 = and i1 %25, %26
  %28 = icmp sle i64 %6, 0
  %29 = sext i1 %28 to i64
  %30 = icmp eq i64 0, %29
  %31 = and i1 %27, %30
  %32 = call { i64, i1 } @llvm.ssub.with.overflow.i64(i64 %20, i64 1)
  %.obit = extractvalue { i64, i1 } %32, 1
  %polly.overflow.state = or i1 false, %.obit
  %.res = extractvalue { i64, i1 } %32, 0
  %polly.access.mul. = call { i64, i1 } @llvm.smul.with.overflow.i64(i64 %.res, i64 %17)
  %polly.access.mul..obit = extractvalue { i64, i1 } %polly.access.mul., 1
  %polly.overflow.state32 = or i1 %polly.overflow.state, %polly.access.mul..obit
  %polly.access.mul..res = extractvalue { i64, i1 } %polly.access.mul., 0
  %33 = call { i64, i1 } @llvm.ssub.with.overflow.i64(i64 %9, i64 1)
  %.obit33 = extractvalue { i64, i1 } %33, 1
  %polly.overflow.state34 = or i1 %polly.overflow.state32, %.obit33
  %.res35 = extractvalue { i64, i1 } %33, 0
  %polly.access.add. = call { i64, i1 } @llvm.sadd.with.overflow.i64(i64 %polly.access.mul..res, i64 %.res35)
  %polly.access.add..obit = extractvalue { i64, i1 } %polly.access.add., 1
  %polly.overflow.state36 = or i1 %polly.overflow.state34, %polly.access.add..obit
  %polly.access.add..res = extractvalue { i64, i1 } %polly.access.add., 0
  %polly.access.mul.37 = call { i64, i1 } @llvm.smul.with.overflow.i64(i64 %polly.access.add..res, i64 %14)
  %polly.access.mul..obit38 = extractvalue { i64, i1 } %polly.access.mul.37, 1
  %polly.overflow.state39 = or i1 %polly.overflow.state36, %polly.access.mul..obit38
  %polly.access.mul..res40 = extractvalue { i64, i1 } %polly.access.mul.37, 0
  %polly.access.add.41 = call { i64, i1 } @llvm.sadd.with.overflow.i64(i64 %polly.access.mul..res40, i64 %6)
  %polly.access.add..obit42 = extractvalue { i64, i1 } %polly.access.add.41, 1
  %polly.overflow.state43 = or i1 %polly.overflow.state39, %polly.access.add..obit42
  %polly.access.add..res44 = extractvalue { i64, i1 } %polly.access.add.41, 0
  %polly.access. = getelementptr double, double* %11, i64 %polly.access.add..res44
  %polly.access.mul.45 = call { i64, i1 } @llvm.smul.with.overflow.i64(i64 0, i64 %9)
  %polly.access.mul..obit46 = extractvalue { i64, i1 } %polly.access.mul.45, 1
  %polly.overflow.state47 = or i1 %polly.overflow.state43, %polly.access.mul..obit46
  %polly.access.mul..res48 = extractvalue { i64, i1 } %polly.access.mul.45, 0
  %polly.access.add.49 = call { i64, i1 } @llvm.sadd.with.overflow.i64(i64 %polly.access.mul..res48, i64 0)
  %polly.access.add..obit50 = extractvalue { i64, i1 } %polly.access.add.49, 1
  %polly.overflow.state51 = or i1 %polly.overflow.state47, %polly.access.add..obit50
  %polly.access.add..res52 = extractvalue { i64, i1 } %polly.access.add.49, 0
  %polly.access.mul.53 = call { i64, i1 } @llvm.smul.with.overflow.i64(i64 %polly.access.add..res52, i64 %6)
  %polly.access.mul..obit54 = extractvalue { i64, i1 } %polly.access.mul.53, 1
  %polly.overflow.state55 = or i1 %polly.overflow.state51, %polly.access.mul..obit54
  %polly.access.mul..res56 = extractvalue { i64, i1 } %polly.access.mul.53, 0
  %polly.access.add.57 = call { i64, i1 } @llvm.sadd.with.overflow.i64(i64 %polly.access.mul..res56, i64 0)
  %polly.access.add..obit58 = extractvalue { i64, i1 } %polly.access.add.57, 1
  %polly.overflow.state59 = or i1 %polly.overflow.state55, %polly.access.add..obit58
  %polly.access.add..res60 = extractvalue { i64, i1 } %polly.access.add.57, 0
  %polly.access.61 = getelementptr double, double* %3, i64 %polly.access.add..res60
  %34 = ptrtoint double* %polly.access. to i64
  %35 = ptrtoint double* %polly.access.61 to i64
  %36 = icmp ule i64 %34, %35
  %37 = call { i64, i1 } @llvm.ssub.with.overflow.i64(i64 %20, i64 1)
  %.obit62 = extractvalue { i64, i1 } %37, 1
  %polly.overflow.state63 = or i1 %polly.overflow.state59, %.obit62
  %.res64 = extractvalue { i64, i1 } %37, 0
  %polly.access.mul.65 = call { i64, i1 } @llvm.smul.with.overflow.i64(i64 %.res64, i64 %9)
  %polly.access.mul..obit66 = extractvalue { i64, i1 } %polly.access.mul.65, 1
  %polly.overflow.state67 = or i1 %polly.overflow.state63, %polly.access.mul..obit66
  %polly.access.mul..res68 = extractvalue { i64, i1 } %polly.access.mul.65, 0
  %38 = call { i64, i1 } @llvm.ssub.with.overflow.i64(i64 %9, i64 1)
  %.obit69 = extractvalue { i64, i1 } %38, 1
  %polly.overflow.state70 = or i1 %polly.overflow.state67, %.obit69
  %.res71 = extractvalue { i64, i1 } %38, 0
  %polly.access.add.72 = call { i64, i1 } @llvm.sadd.with.overflow.i64(i64 %polly.access.mul..res68, i64 %.res71)
  %polly.access.add..obit73 = extractvalue { i64, i1 } %polly.access.add.72, 1
  %polly.overflow.state74 = or i1 %polly.overflow.state70, %polly.access.add..obit73
  %polly.access.add..res75 = extractvalue { i64, i1 } %polly.access.add.72, 0
  %polly.access.mul.76 = call { i64, i1 } @llvm.smul.with.overflow.i64(i64 %polly.access.add..res75, i64 %6)
  %polly.access.mul..obit77 = extractvalue { i64, i1 } %polly.access.mul.76, 1
  %polly.overflow.state78 = or i1 %polly.overflow.state74, %polly.access.mul..obit77
  %polly.access.mul..res79 = extractvalue { i64, i1 } %polly.access.mul.76, 0
  %polly.access.add.80 = call { i64, i1 } @llvm.sadd.with.overflow.i64(i64 %polly.access.mul..res79, i64 %6)
  %polly.access.add..obit81 = extractvalue { i64, i1 } %polly.access.add.80, 1
  %polly.overflow.state82 = or i1 %polly.overflow.state78, %polly.access.add..obit81
  %polly.access.add..res83 = extractvalue { i64, i1 } %polly.access.add.80, 0
  %polly.access.84 = getelementptr double, double* %3, i64 %polly.access.add..res83
  %polly.access.mul.85 = call { i64, i1 } @llvm.smul.with.overflow.i64(i64 0, i64 %17)
  %polly.access.mul..obit86 = extractvalue { i64, i1 } %polly.access.mul.85, 1
  %polly.overflow.state87 = or i1 %polly.overflow.state82, %polly.access.mul..obit86
  %polly.access.mul..res88 = extractvalue { i64, i1 } %polly.access.mul.85, 0
  %polly.access.add.89 = call { i64, i1 } @llvm.sadd.with.overflow.i64(i64 %polly.access.mul..res88, i64 0)
  %polly.access.add..obit90 = extractvalue { i64, i1 } %polly.access.add.89, 1
  %polly.overflow.state91 = or i1 %polly.overflow.state87, %polly.access.add..obit90
  %polly.access.add..res92 = extractvalue { i64, i1 } %polly.access.add.89, 0
  %polly.access.mul.93 = call { i64, i1 } @llvm.smul.with.overflow.i64(i64 %polly.access.add..res92, i64 %14)
  %polly.access.mul..obit94 = extractvalue { i64, i1 } %polly.access.mul.93, 1
  %polly.overflow.state95 = or i1 %polly.overflow.state91, %polly.access.mul..obit94
  %polly.access.mul..res96 = extractvalue { i64, i1 } %polly.access.mul.93, 0
  %polly.access.add.97 = call { i64, i1 } @llvm.sadd.with.overflow.i64(i64 %polly.access.mul..res96, i64 0)
  %polly.access.add..obit98 = extractvalue { i64, i1 } %polly.access.add.97, 1
  %polly.overflow.state99 = or i1 %polly.overflow.state95, %polly.access.add..obit98
  %polly.access.add..res100 = extractvalue { i64, i1 } %polly.access.add.97, 0
  %polly.access.101 = getelementptr double, double* %11, i64 %polly.access.add..res100
  %39 = ptrtoint double* %polly.access.84 to i64
  %40 = ptrtoint double* %polly.access.101 to i64
  %41 = icmp ule i64 %39, %40
  %42 = or i1 %36, %41
  %43 = and i1 %31, %42
  %44 = call { i64, i1 } @llvm.smul.with.overflow.i64(i64 1, i64 %6)
  %.obit102 = extractvalue { i64, i1 } %44, 1
  %polly.overflow.state103 = or i1 %polly.overflow.state99, %.obit102
  %.res104 = extractvalue { i64, i1 } %44, 0
  %45 = call { i64, i1 } @llvm.smul.with.overflow.i64(i64 %.res104, i64 %9)
  %.obit105 = extractvalue { i64, i1 } %45, 1
  %polly.overflow.state106 = or i1 %polly.overflow.state103, %.obit105
  %.res107 = extractvalue { i64, i1 } %45, 0
  %46 = call { i64, i1 } @llvm.smul.with.overflow.i64(i64 %.res107, i64 %20)
  %.obit108 = extractvalue { i64, i1 } %46, 1
  %polly.overflow.state109 = or i1 %polly.overflow.state106, %.obit108
  %.res110 = extractvalue { i64, i1 } %46, 0
  %47 = call { i64, i1 } @llvm.smul.with.overflow.i64(i64 29, i64 %.res110)
  %.obit111 = extractvalue { i64, i1 } %47, 1
  %polly.overflow.state112 = or i1 %polly.overflow.state109, %.obit111
  %.res113 = extractvalue { i64, i1 } %47, 0
  %48 = call { i64, i1 } @llvm.sadd.with.overflow.i64(i64 0, i64 %.res113)
  %.obit114 = extractvalue { i64, i1 } %48, 1
  %polly.overflow.state115 = or i1 %polly.overflow.state112, %.obit114
  %.res116 = extractvalue { i64, i1 } %48, 0
  %49 = icmp sge i64 %.res116, 2621440
  %50 = and i1 %43, %49
  %polly.rtc.overflown = xor i1 %polly.overflow.state115, true
  %polly.rtc.result = and i1 %50, %polly.rtc.overflown
  br i1 false, label %polly.start, label %if.pre_entry_bb

if.pre_entry_bb:                                  ; preds = %polly.split_new_and_old
  br label %if, !dbg !39

if:                                               ; preds = %if.pre_entry_bb, %L22.loopexit
  %"#temp#2.031" = phi i64 [ %51, %L22.loopexit ], [ 1, %if.pre_entry_bb ]
  %51 = add i64 %"#temp#2.031", 1, !dbg !39
  call void @llvm.dbg.value(metadata i64 %9, i64 0, metadata !22, metadata !25), !dbg !24
  call void @llvm.dbg.value(metadata i64 %9, i64 0, metadata !22, metadata !25), !dbg !24
  %52 = icmp sgt i64 %9, 0, !dbg !40
  %53 = select i1 %52, i64 %9, i64 0, !dbg !40
  %54 = icmp slt i64 %9, 1, !dbg !40
  br i1 %54, label %L22.loopexit, label %if14.lr.ph, !dbg !40

if14.lr.ph:                                       ; preds = %if
  br label %if14, !dbg !40

L22.L55_crit_edge:                                ; preds = %L22.loopexit
  br label %polly.merge_new_and_old, !dbg !39

polly.merge_new_and_old:                          ; preds = %polly.exiting, %L22.L55_crit_edge
  br label %L55, !dbg !41

L55:                                              ; preds = %polly.merge_new_and_old, %top.split
  ret void, !dbg !41

L40.L31.loopexit_crit_edge:                       ; preds = %if15
  br label %L31.loopexit, !dbg !42

L31.loopexit:                                     ; preds = %L40.L31.loopexit_crit_edge, %if14
  %55 = icmp eq i64 %"#temp#1.030", %53, !dbg !40
  br i1 %55, label %L31.L22.loopexit_crit_edge, label %if14, !dbg !40

if14:                                             ; preds = %if14.lr.ph, %L31.loopexit
  %"#temp#1.030" = phi i64 [ 1, %if14.lr.ph ], [ %56, %L31.loopexit ]
  %56 = add i64 %"#temp#1.030", 1, !dbg !40
  call void @llvm.dbg.value(metadata i64 %20, i64 0, metadata !23, metadata !25), !dbg !24
  call void @llvm.dbg.value(metadata i64 %20, i64 0, metadata !23, metadata !25), !dbg !24
  %57 = icmp sgt i64 %20, 0, !dbg !42
  %58 = select i1 %57, i64 %20, i64 0, !dbg !42
  %59 = icmp slt i64 %20, 1, !dbg !42
  br i1 %59, label %L31.loopexit, label %if15.lr.ph, !dbg !42

if15.lr.ph:                                       ; preds = %if14
  br label %if15, !dbg !42

if15:                                             ; preds = %if15.lr.ph, %if15
  %"#temp#.029" = phi i64 [ 1, %if15.lr.ph ], [ %60, %if15 ]
  %60 = add i64 %"#temp#.029", 1, !dbg !42
  call void @llvm.dbg.value(metadata i64 %"#temp#2.031", i64 0, metadata !20, metadata !25), !dbg !24
  call void @llvm.dbg.value(metadata i64 %"#temp#1.030", i64 0, metadata !19, metadata !25), !dbg !24
  call void @llvm.dbg.value(metadata i64 %"#temp#.029", i64 0, metadata !17, metadata !25), !dbg !24
  %61 = add i64 %"#temp#1.030", -1, !dbg !41
  %62 = add i64 %"#temp#.029", -1, !dbg !41
  %63 = mul i64 %62, %17, !dbg !41
  %tmp = add i64 %61, %63
  %tmp24 = mul i64 %tmp, %14
  %64 = add i64 %"#temp#2.031", -1, !dbg !41
  %65 = add i64 %64, %tmp24, !dbg !41
  %66 = getelementptr double, double* %11, i64 %65, !dbg !41
  %67 = load double, double* %66, align 8, !dbg !41, !tbaa !43
  call void @llvm.dbg.value(metadata i64 %"#temp#2.031", i64 0, metadata !20, metadata !25), !dbg !24
  call void @llvm.dbg.value(metadata i64 %"#temp#1.030", i64 0, metadata !19, metadata !25), !dbg !24
  call void @llvm.dbg.value(metadata i64 %"#temp#.029", i64 0, metadata !17, metadata !25), !dbg !24
  %68 = fmul double %67, %67, !dbg !41
  call void @llvm.dbg.value(metadata i64 %"#temp#2.031", i64 0, metadata !20, metadata !25), !dbg !24
  call void @llvm.dbg.value(metadata i64 %"#temp#1.030", i64 0, metadata !19, metadata !25), !dbg !24
  call void @llvm.dbg.value(metadata i64 %"#temp#.029", i64 0, metadata !17, metadata !25), !dbg !24
  %69 = mul i64 %62, %9, !dbg !41
  %tmp27 = add i64 %61, %69
  %tmp28 = mul i64 %tmp27, %6
  %70 = add i64 %64, %tmp28, !dbg !41
  %71 = getelementptr double, double* %3, i64 %70, !dbg !41
  store double %68, double* %71, align 8, !dbg !41, !tbaa !43
  %72 = icmp eq i64 %"#temp#.029", %58, !dbg !42
  br i1 %72, label %L40.L31.loopexit_crit_edge, label %if15, !dbg !42

polly.start:                                      ; preds = %polly.split_new_and_old
  br label %polly.acc.initialize

polly.acc.initialize:                             ; preds = %polly.start
  %73 = call i8* @polly_initContext()
  %74 = mul nsw i64 %20, %17
  %75 = mul nsw i64 %74, %14
  %76 = mul i64 8, %75
  %p_dev_array_MemRef0 = call i8* @polly_allocateMemoryForDevice(i64 %76)
  %77 = mul nsw i64 %20, %9
  %78 = mul nsw i64 %77, %6
  %79 = mul i64 8, %78
  %p_dev_array_MemRef1 = call i8* @polly_allocateMemoryForDevice(i64 %79)
  br label %polly.cond

polly.cond:                                       ; preds = %polly.acc.initialize
  %80 = icmp sge i64 %9, 1
  %81 = icmp sge i64 %20, 1
  %82 = and i1 %80, %81
  %83 = icmp sge i64 %6, 1
  %84 = and i1 %82, %83
  br i1 %84, label %polly.then, label %polly.else

polly.merge:                                      ; preds = %polly.else, %polly.merge118
  call void @polly_freeDeviceMemory(i8* %p_dev_array_MemRef1)
  call void @polly_freeDeviceMemory(i8* %p_dev_array_MemRef0)
  call void @polly_freeContext(i8* %73)
  br label %polly.exiting

polly.exiting:                                    ; preds = %polly.merge
  br label %polly.merge_new_and_old

polly.then:                                       ; preds = %polly.cond
  br label %polly.cond117

polly.cond117:                                    ; preds = %polly.then
  %85 = icmp sge i64 %17, 1
  %86 = icmp sge i64 %14, 1
  %87 = and i1 %85, %86
  br i1 %87, label %polly.then119, label %polly.else120

polly.merge118:                                   ; preds = %polly.else120, %polly.then119
  %88 = call i8* @polly_getDevicePtr(i8* %p_dev_array_MemRef0)
  %89 = getelementptr [8 x i8*], [8 x i8*]* %polly_launch_0_params, i64 0, i64 0
  store i8* %88, i8** %polly_launch_0_param_0
  %90 = bitcast i8** %polly_launch_0_param_0 to i8*
  store i8* %90, i8** %89
  %91 = call i8* @polly_getDevicePtr(i8* %p_dev_array_MemRef1)
  %92 = getelementptr [8 x i8*], [8 x i8*]* %polly_launch_0_params, i64 0, i64 1
  store i8* %91, i8** %polly_launch_0_param_1
  %93 = bitcast i8** %polly_launch_0_param_1 to i8*
  store i8* %93, i8** %92
  store i64 %9, i64* %polly_launch_0_param_2
  %94 = getelementptr [8 x i8*], [8 x i8*]* %polly_launch_0_params, i64 0, i64 2
  %95 = bitcast i64* %polly_launch_0_param_2 to i8*
  store i8* %95, i8** %94
  store i64 %20, i64* %polly_launch_0_param_3
  %96 = getelementptr [8 x i8*], [8 x i8*]* %polly_launch_0_params, i64 0, i64 3
  %97 = bitcast i64* %polly_launch_0_param_3 to i8*
  store i8* %97, i8** %96
  store i64 %6, i64* %polly_launch_0_param_4
  %98 = getelementptr [8 x i8*], [8 x i8*]* %polly_launch_0_params, i64 0, i64 4
  %99 = bitcast i64* %polly_launch_0_param_4 to i8*
  store i8* %99, i8** %98
  store i64 %17, i64* %polly_launch_0_param_5
  %100 = getelementptr [8 x i8*], [8 x i8*]* %polly_launch_0_params, i64 0, i64 5
  %101 = bitcast i64* %polly_launch_0_param_5 to i8*
  store i8* %101, i8** %100
  store i64 %14, i64* %polly_launch_0_param_6
  %102 = getelementptr [8 x i8*], [8 x i8*]* %polly_launch_0_params, i64 0, i64 6
  %103 = bitcast i64* %polly_launch_0_param_6 to i8*
  store i8* %103, i8** %102
  store void (metadata, i64, metadata, metadata)* @llvm.dbg.value, void (metadata, i64, metadata, metadata)** %polly_launch_0_param_7
  %104 = getelementptr [8 x i8*], [8 x i8*]* %polly_launch_0_params, i64 0, i64 7
  %105 = bitcast void (metadata, i64, metadata, metadata)** %polly_launch_0_param_7 to i8*
  store i8* %105, i8** %104
  %106 = call i8* @polly_getKernel(i8* getelementptr inbounds ([1 x i8], [1 x i8]* @kernel_0, i32 0, i32 0), i8* getelementptr inbounds ([9 x i8], [9 x i8]* @kernel_0_name, i32 0, i32 0))
  %107 = icmp sge i64 %6, 8161
  %108 = add nsw i64 %6, 31
  %polly.fdiv_q.shr = ashr i64 %108, 5
  %109 = select i1 %107, i64 256, i64 %polly.fdiv_q.shr
  %110 = trunc i64 %109 to i32
  %111 = icmp sge i64 %9, 8162
  %112 = add nsw i64 %9, 31
  %polly.fdiv_q.shr121 = ashr i64 %112, 5
  %113 = select i1 %111, i64 256, i64 %polly.fdiv_q.shr121
  %114 = trunc i64 %113 to i32
  call void @polly_launchKernel(i8* %106, i32 %110, i32 %114, i32 32, i32 4, i32 4, i8* %polly_launch_0_params_i8ptr)
  call void @polly_freeKernel(i8* %106)
  %115 = mul nsw i64 %20, %9
  %116 = mul nsw i64 %115, %6
  %117 = mul i64 8, %116
  %118 = bitcast double* %3 to i8*
  call void @polly_copyFromDeviceToHost(i8* %p_dev_array_MemRef1, i8* %118, i64 %117)
  br label %polly.merge

polly.else:                                       ; preds = %polly.cond
  br label %polly.merge

polly.then119:                                    ; preds = %polly.cond117
  %119 = mul nsw i64 %20, %17
  %120 = mul nsw i64 %119, %14
  %121 = mul i64 8, %120
  %122 = bitcast double* %11 to i8*
  call void @polly_copyFromHostToDevice(i8* %122, i8* %p_dev_array_MemRef0, i64 %121)
  br label %polly.merge118

polly.else120:                                    ; preds = %polly.cond117
  br label %polly.merge118
}

define i8** @jlcall_square_3D_mat_64453(i8**, i8***, i32) #1 {
top:
  br label %top.split

top.split:                                        ; preds = %top
  %3 = load i8**, i8*** %1, align 8
  %4 = getelementptr i8**, i8*** %1, i64 1
  %5 = load i8**, i8*** %4, align 8
  call void @julia_square_3D_mat_64454(i8** %3, i8** %5)
  ret i8** inttoptr (i64 140082821496848 to i8**)
}

declare i8**** @jl_get_ptls_states()

; Function Attrs: nounwind readnone
declare void @llvm.dbg.declare(metadata, metadata, metadata) #2

; Function Attrs: noreturn
declare void @jl_error(i8*) #3

; Function Attrs: argmemonly nounwind
declare void @llvm.memcpy.p0i8.p0i8.i32(i8* nocapture writeonly, i8* nocapture readonly, i32, i32, i1) #4

; Function Attrs: argmemonly nounwind
declare void @llvm.memcpy.p0i8.p0i8.i64(i8* nocapture writeonly, i8* nocapture readonly, i64, i32, i1) #4

declare void @jl_enter_handler(i8*)

; Function Attrs: returns_twice
declare i32 @__sigsetjmp(i8*, i32) #5

; Function Attrs: nounwind readnone
declare void @llvm.dbg.value(metadata, i64, metadata, metadata) #2

; Function Attrs: nounwind readnone
declare { i64, i1 } @llvm.ssub.with.overflow.i64(i64, i64) #2

; Function Attrs: nounwind readnone
declare { i64, i1 } @llvm.smul.with.overflow.i64(i64, i64) #2

; Function Attrs: nounwind readnone
declare { i64, i1 } @llvm.sadd.with.overflow.i64(i64, i64) #2

declare i8* @polly_initContext()

declare i8* @polly_allocateMemoryForDevice(i64)

declare void @polly_copyFromHostToDevice(i8*, i8*, i64)

declare i8* @polly_getDevicePtr(i8*)

declare i8* @polly_getKernel(i8*, i8*)

declare void @polly_launchKernel(i8*, i32, i32, i32, i32, i32, i8*)

declare void @polly_freeKernel(i8*)

declare void @polly_copyFromDeviceToHost(i8*, i8*, i64)

declare void @polly_freeDeviceMemory(i8*)

declare void @polly_freeContext(i8*)

attributes #0 = { sspstrong "no-frame-pointer-elim"="true" }
attributes #1 = { "no-frame-pointer-elim"="true" }
attributes #2 = { nounwind readnone }
attributes #3 = { noreturn }
attributes #4 = { argmemonly nounwind }
attributes #5 = { returns_twice }

!llvm.module.flags = !{!0, !1}
!llvm.dbg.cu = !{!2}

!0 = !{i32 2, !"Dwarf Version", i32 4}
!1 = !{i32 1, !"Debug Info Version", i32 3}
!2 = distinct !DICompileUnit(language: DW_LANG_C89, file: !3, producer: "julia", isOptimized: true, runtimeVersion: 0, emissionKind: FullDebug, enums: !4)
!3 = !DIFile(filename: "REPL[1]", directory: ".")
!4 = !{}
!5 = distinct !DISubprogram(name: "square_3D_mat", linkageName: "julia_square_3D_mat_64454", scope: null, file: !3, type: !6, isLocal: false, isDefinition: true, isOptimized: true, unit: !2, variables: !12)
!6 = !DISubroutineType(types: !7)
!7 = !{!8, !8}
!8 = !DIDerivedType(tag: DW_TAG_pointer_type, baseType: !9, size: 64, align: 64)
!9 = !DICompositeType(tag: DW_TAG_structure_type, name: "jl_value_t", file: !10, line: 71, align: 64, elements: !11)
!10 = !DIFile(filename: "julia.h", directory: "")
!11 = !{!8}
!12 = !{!13, !15, !16, !17, !19, !20, !21, !22, !23}
!13 = !DILocalVariable(name: "#self#", arg: 1, scope: !5, file: !3, line: 2, type: !14)
!14 = !DICompositeType(tag: DW_TAG_structure_type, name: "#square_3D_mat", align: 8, elements: !4, runtimeLang: DW_LANG_Julia, identifier: "#square_3D_mat_64400")
!15 = !DILocalVariable(name: "a", arg: 2, scope: !5, file: !3, line: 2, type: !8)
!16 = !DILocalVariable(name: "b", arg: 3, scope: !5, file: !3, line: 2, type: !8)
!17 = !DILocalVariable(name: "k", scope: !5, file: !3, line: 2, type: !18)
!18 = !DIBasicType(name: "Int64", size: 64, encoding: DW_ATE_unsigned)
!19 = !DILocalVariable(name: "j", scope: !5, file: !3, line: 2, type: !18)
!20 = !DILocalVariable(name: "i", scope: !5, file: !3, line: 2, type: !18)
!21 = !DILocalVariable(name: "ni", scope: !5, file: !3, line: 2, type: !18)
!22 = !DILocalVariable(name: "nj", scope: !5, file: !3, line: 2, type: !18)
!23 = !DILocalVariable(name: "nk", scope: !5, file: !3, line: 2, type: !18)
!24 = !DILocation(line: 2, scope: !5)
!25 = !DIExpression()
!26 = !{!27, !27, i64 0, i64 1}
!27 = !{!"jtbaa_const", !28, i64 0}
!28 = !{!"jtbaa"}
!29 = !DILocation(line: 80, scope: !30, inlinedAt: !33)
!30 = distinct !DISubprogram(name: "_size;", linkageName: "_size", scope: !31, file: !31, type: !32, isLocal: false, isDefinition: true, isOptimized: true, unit: !2, variables: !4)
!31 = !DIFile(filename: "array.jl", directory: ".")
!32 = !DISubroutineType(types: !4)
!33 = !DILocation(line: 80, scope: !34, inlinedAt: !35)
!34 = distinct !DISubprogram(name: "_size;", linkageName: "_size", scope: !31, file: !31, type: !32, isLocal: false, isDefinition: true, isOptimized: true, unit: !2, variables: !4)
!35 = !DILocation(line: 80, scope: !36, inlinedAt: !37)
!36 = distinct !DISubprogram(name: "_size;", linkageName: "_size", scope: !31, file: !31, type: !32, isLocal: false, isDefinition: true, isOptimized: true, unit: !2, variables: !4)
!37 = !DILocation(line: 76, scope: !38, inlinedAt: !24)
!38 = distinct !DISubprogram(name: "size;", linkageName: "size", scope: !31, file: !31, type: !32, isLocal: false, isDefinition: true, isOptimized: true, unit: !2, variables: !4)
!39 = !DILocation(line: 3, scope: !5)
!40 = !DILocation(line: 4, scope: !5)
!41 = !DILocation(line: 6, scope: !5)
!42 = !DILocation(line: 5, scope: !5)
!43 = !{!44, !44, i64 0}
!44 = !{!"jtbaa_arraybuf", !45, i64 0}
!45 = !{!"jtbaa_data", !28, i64 0}