Julia: armv7l: ptrtoint not supported for non-integral pointers

[staticfloat]

Armv7l bootstrap is failing right at the end during the link step with this:

ptrtoint not supported for non-integral pointers
inttoptr not supported for non-integral pointers
ptrtoint not supported for non-integral pointers
inttoptr not supported for non-integral pointers
in function julia__unsafe_copytoNOT._53532
LLVM ERROR: Broken function found, compilation aborted!

Exemplary build: https://build.julialang.org/#/builders/57/builds/231

[ianshmean]

Looks very similar to this LLVM issue https://github.com/JuliaLang/julia/issues/31156#issuecomment-586449043

ptrtoint not supported for non-integral pointers
inttoptr not supported for non-integral pointers
ptrtoint not supported for non-integral pointers
inttoptr not supported for non-integral pointers
in function julia_copyto!_18518
LLVM ERROR: Broken function found, compilation aborted!

which was patched by @Keno and ultimately merged in https://github.com/JuliaLang/julia/pull/35313

Seems like the patch didn't fix julia__unsafe_copytoNOT.?

[giordano]

@terasakisatoshi has been able to build julia v1.4.2 for armv7l: https://discourse.julialang.org/t/have-a-try-julia-v1-4-2-for-arm32bit/40284

[Keno]

Inttoptr issues like this indicate an un-soundness in some LLVM pass. They need to be fixed one by one.

[ViralBShah]

Will this hurt us on those new arm macs? I imagine they are aarch64, and so this may not be an issue.

[ViralBShah]

@Keno Is this something we can communicate to our various ARM contacts?

[staticfloat]

Correct; armv7l is separate from aarch64 (otherwise known as armv8)

[Keno]

It's not an ARM issue, it's an LLVM issue.

[ViralBShah]

Right - but is it only an LLVM issue on arm - and hence the ARM llvm team should look at it, or is it a general LLVM issue that needs to be fixed by the LLVM community?

[Keno]

It's specific to how we are using LLVM. It's unlikely the ARM LLVM team would care enough to understand how we're using LLVM.

[yuyichao]

Bugpoint reduced repro:

; ModuleID = 'bugpoint-reduced-simplified.bc'
source_filename = "text"
target datalayout = "e-m:e-p:32:32-Fi8-i64:64-v128:64:128-a:0:32-n32-S64-ni:10:11:12:13"
target triple = "armv7l-unknown-linux-gnueabihf"

define void @julia_show_tuple_as_call_23179() local_unnamed_addr {
top:
  br i1 undef, label %L16.lr.ph, label %L51

L16.lr.ph:                                        ; preds = %top
  unreachable

L51:                                              ; preds = %top
  %0 = addrspacecast [2 x {} addrspace(10)*]* undef to [2 x {} addrspace(10)*] addrspace(11)*
  call fastcc void @julia_show_signature_function_32381([2 x {} addrspace(10)*] addrspace(11)* nocapture readonly %0)
  ret void
}

define dso_local fastcc void @julia_show_signature_function_32381([2 x {} addrspace(10)*] addrspace(11)* nocapture nonnull readonly align 4 dereferenceable(8) %0) unnamed_addr {
top:
  %1 = bitcast [2 x {} addrspace(10)*] addrspace(11)* %0 to i64 addrspace(11)*
  ret void
}

!llvm.module.flags = !{!0}

!0 = !{i32 1, !"Debug Info Version", i32 3}

reproducible with opt bugpoint-reduced-simplified.ll -inline.

Not 100% sure if this is the original one but this is what I got by running opt -O2 on the dumpped unoptimized IR. I had to strip other unrelated functions first before feeding it to bugpoint since it otherwise go OOM....

[yuyichao]

Reported as https://bugs.llvm.org/show_bug.cgi?id=46737

[yuyichao]

That's actually not related and is fixed on LLVM master already. The actual one seems to be caused by Loop Strength Reduction.

What I currently have is.

*** IR Dump Before Loop Strength Reduction ***
; Preheader:
L37.preheader:                                    ; preds = %idxend4
  %110 = bitcast {} addrspace(10)* %1 to { i8 addrspace(13)*, i32, i16, i16, i32 } addrspace(10)*, !dbg !8332
  %111 = addrspacecast { i8 addrspace(13)*, i32, i16, i16, i32 } addrspace(10)* %110 to { i8 addrspace(13)*, i32, i16, i16, i32 } addrspace(11)*, !dbg !8332
  %112 = getelementptr inbounds { i8 addrspace(13)*, i32, i16, i16, i32 }, { i8 addrspace(13)*, i32, i16, i16, i32 } addrspace(11)* %111, i32 0, i32 1, !dbg !8332
  %113 = load i32, i32 addrspace(11)* %112, align 4, !dbg !8332, !tbaa !8259
  %114 = bitcast {} addrspace(10)* %1 to i8 addrspace(13)* addrspace(10)*, !dbg !8332
  %115 = addrspacecast i8 addrspace(13)* addrspace(10)* %114 to i8 addrspace(13)* addrspace(11)*, !dbg !8332
  %116 = load i8 addrspace(13)*, i8 addrspace(13)* addrspace(11)* %115, align 4, !dbg !8332
  %117 = bitcast {} addrspace(10)* %1 to { i8 addrspace(13)*, i32, i16, i16, i32 } addrspace(10)*, !dbg !8333
  %118 = addrspacecast { i8 addrspace(13)*, i32, i16, i16, i32 } addrspace(10)* %117 to { i8 addrspace(13)*, i32, i16, i16, i32 } addrspace(11)*, !dbg !8333
  %119 = getelementptr inbounds { i8 addrspace(13)*, i32, i16, i16, i32 }, { i8 addrspace(13)*, i32, i16, i16, i32 } addrspace(11)* %118, i32 0, i32 1, !dbg !8333
  %120 = load i32, i32 addrspace(11)* %119, align 4, !dbg !8333
  %121 = bitcast {} addrspace(10)* %1 to i8 addrspace(13)* addrspace(10)*, !dbg !8333
  %122 = addrspacecast i8 addrspace(13)* addrspace(10)* %121 to i8 addrspace(13)* addrspace(11)*, !dbg !8333
  %123 = load i8 addrspace(13)*, i8 addrspace(13)* addrspace(11)* %122, align 4, !dbg !8333
  %124 = bitcast {} addrspace(10)* %1 to { i8 addrspace(13)*, i32, i16, i16, i32 } addrspace(10)*, !dbg !8334
  %125 = addrspacecast { i8 addrspace(13)*, i32, i16, i16, i32 } addrspace(10)* %124 to { i8 addrspace(13)*, i32, i16, i16, i32 } addrspace(11)*, !dbg !8334
  %126 = getelementptr inbounds { i8 addrspace(13)*, i32, i16, i16, i32 }, { i8 addrspace(13)*, i32, i16, i16, i32 } addrspace(11)* %125, i32 0, i32 1, !dbg !8334
  %127 = load i32, i32 addrspace(11)* %126, align 4, !dbg !8334
  %128 = bitcast {} addrspace(10)* %1 to i8 addrspace(13)* addrspace(10)*, !dbg !8334
  %129 = addrspacecast i8 addrspace(13)* addrspace(10)* %128 to i8 addrspace(13)* addrspace(11)*, !dbg !8334
  %130 = load i8 addrspace(13)*, i8 addrspace(13)* addrspace(11)* %129, align 4, !dbg !8334
  %131 = bitcast {} addrspace(10)* %1 to { i8 addrspace(13)*, i32, i16, i16, i32 } addrspace(10)*, !dbg !8335
  %132 = addrspacecast { i8 addrspace(13)*, i32, i16, i16, i32 } addrspace(10)* %131 to { i8 addrspace(13)*, i32, i16, i16, i32 } addrspace(11)*, !dbg !8335
  %133 = getelementptr inbounds { i8 addrspace(13)*, i32, i16, i16, i32 }, { i8 addrspace(13)*, i32, i16, i16, i32 } addrspace(11)* %132, i32 0, i32 1, !dbg !8335
  %134 = load i32, i32 addrspace(11)* %133, align 4, !dbg !8335
  %135 = bitcast {} addrspace(10)* %1 to i8 addrspace(13)* addrspace(10)*, !dbg !8335
  %136 = addrspacecast i8 addrspace(13)* addrspace(10)* %135 to i8 addrspace(13)* addrspace(11)*, !dbg !8335
  %137 = load i8 addrspace(13)*, i8 addrspace(13)* addrspace(11)* %136, align 4, !dbg !8335
  br label %L37, !dbg !8280

; Loop:
L37:                                              ; preds = %L37.preheader, %idxend16
  %value_phi7 = phi i32 [ %162, %idxend16 ], [ %107, %L37.preheader ]
  %44 = add i32 %value_phi7, -1, !dbg !8280
  %45 = icmp ult i32 %44, %113, !dbg !8280
  br i1 %45, label %idxend10, label %oob9, !dbg !8280

idxend10:                                         ; preds = %L37
  %140 = getelementptr inbounds i8, i8 addrspace(13)* %116, i32 %44, !dbg !8280
  %141 = load i8, i8 addrspace(13)* %140, align 1, !dbg !8280, !tbaa !8300
  %142 = icmp eq i8 %141, 58, !dbg !8336
  br i1 %142, label %L49, label %L47, !dbg !8281

L49:                                              ; preds = %idxend10
  %50 = add i32 %value_phi7, -1, !dbg !8282
  %51 = icmp ult i32 %50, %120, !dbg !8282
  br i1 %51, label %idxend12, label %oob11, !dbg !8282

idxend12:                                         ; preds = %L49
  %145 = getelementptr inbounds i8, i8 addrspace(13)* %123, i32 %50, !dbg !8282
  store i8 48, i8 addrspace(13)* %145, align 1, !dbg !8282, !tbaa !8300
  %146 = add i32 %value_phi7, -2, !dbg !8342
  %147 = icmp ult i32 %146, %127, !dbg !8342
  br i1 %147, label %idxend14, label %oob13, !dbg !8342

idxend14:                                         ; preds = %idxend12
  %151 = getelementptr inbounds i8, i8 addrspace(13)* %130, i32 %146, !dbg !8342
  %152 = load i8, i8 addrspace(13)* %151, align 1, !dbg !8342, !tbaa !8300
  %153 = zext i8 %152 to i32, !dbg !8344
  %154 = add nuw nsw i32 %153, 1, !dbg !8284
  %155 = trunc i32 %154 to i8, !dbg !8349
  %156 = icmp eq i8 %152, -1, !dbg !8288
  br i1 %156, label %L59, label %L64, !dbg !8288

L64:                                              ; preds = %idxend14
  %64 = add i32 %value_phi7, -2, !dbg !8292
  %65 = icmp ult i32 %64, %134, !dbg !8292
  br i1 %65, label %idxend16, label %oob15, !dbg !8292

idxend16:                                         ; preds = %L64
  %160 = getelementptr inbounds i8, i8 addrspace(13)* %137, i32 %64, !dbg !8292
  store i8 %155, i8 addrspace(13)* %160, align 1, !dbg !8292, !tbaa !8300
  %161 = icmp eq i32 %value_phi7, %108, !dbg !8350
  %162 = add i32 %value_phi7, -1, !dbg !8351
  br i1 %161, label %L77.loopexit, label %L37, !dbg !8287

; Exit blocks
oob9:                                             ; preds = %L37
  %value_phi7.lcssa = phi i32 [ %value_phi7, %L37 ]
  %138 = alloca i32, align 4, !dbg !8280
  store i32 %value_phi7.lcssa, i32* %138, align 4, !dbg !8280
  %139 = addrspacecast {} addrspace(10)* %1 to {} addrspace(12)*, !dbg !8280
  call void @jl_bounds_error_ints({} addrspace(12)* %139, i32* nonnull %138, i32 1), !dbg !8280
  unreachable, !dbg !8280

L47:                                              ; preds = %idxend10
  %.sroa.027.0..sroa_idx = getelementptr inbounds [2 x i32], [2 x i32]* %0, i32 0, i32 0, !dbg !8281
  store i32 %2, i32* %.sroa.027.0..sroa_idx, align 4, !dbg !8281
  %.sroa.228.0..sroa_idx29 = getelementptr inbounds [2 x i32], [2 x i32]* %0, i32 0, i32 1, !dbg !8281
  store i32 %3, i32* %.sroa.228.0..sroa_idx29, align 4, !dbg !8281
  %46 = getelementptr {} addrspace(10)*, {} addrspace(10)** %gcframe, i32 1
  %47 = load {} addrspace(10)*, {} addrspace(10)** %46, align 4, !tbaa !8249
  %48 = getelementptr {}**, {}*** %ptls, i32 0
  %49 = bitcast {}*** %48 to {} addrspace(10)**
  store {} addrspace(10)* %47, {} addrspace(10)** %49, align 4, !tbaa !8249
  ret void, !dbg !8281

oob11:                                            ; preds = %L49
  %value_phi7.lcssa39 = phi i32 [ %value_phi7, %L49 ]
  %143 = alloca i32, align 4, !dbg !8282
  store i32 %value_phi7.lcssa39, i32* %143, align 4, !dbg !8282
  %144 = addrspacecast {} addrspace(10)* %1 to {} addrspace(12)*, !dbg !8282
  call void @jl_bounds_error_ints({} addrspace(12)* %144, i32* nonnull %143, i32 1), !dbg !8282
  unreachable, !dbg !8282

oob13:                                            ; preds = %idxend12
  %value_phi7.lcssa51 = phi i32 [ %value_phi7, %idxend12 ]
  %148 = add i32 %value_phi7.lcssa51, -1, !dbg !8343
  %149 = alloca i32, align 4, !dbg !8342
  store i32 %148, i32* %149, align 4, !dbg !8342
  %150 = addrspacecast {} addrspace(10)* %1 to {} addrspace(12)*, !dbg !8342
  call void @jl_bounds_error_ints({} addrspace(12)* %150, i32* nonnull %149, i32 1), !dbg !8342
  unreachable, !dbg !8342

L59:                                              ; preds = %idxend14
  %.lcssa47 = phi i32 [ %154, %idxend14 ], !dbg !8284
  %52 = load {}*, {}** @jl_globalYY.16, align 4, !dbg !8288, !tbaa !8276, !nonnull !4
  %53 = addrspacecast {}* %52 to {} addrspace(10)*, !dbg !8288
  %54 = load {}*, {}** @jl_symYY.trunc61, align 4, !dbg !8288, !tbaa !8276, !nonnull !4
  %55 = addrspacecast {}* %54 to {} addrspace(10)*, !dbg !8288
  %56 = load {}*, {}** @SUM.CoreDOT.UInt8254, align 4, !dbg !8288, !tbaa !8276, !nonnull !4, !dereferenceable !8278, !align !8279
  %57 = addrspacecast {}* %56 to {} addrspace(10)*, !dbg !8288
  %58 = call {} addrspace(10)* @jl_box_int32(i32 signext %.lcssa47), !dbg !8288
  %59 = getelementptr {} addrspace(10)*, {} addrspace(10)** %gcframe, i32 2
  store {} addrspace(10)* %58, {} addrspace(10)** %59
  %60 = getelementptr {} addrspace(10)*, {} addrspace(10)** %4, i32 0, !dbg !8288
  store {} addrspace(10)* %55, {} addrspace(10)** %60, align 4, !dbg !8288
  %61 = getelementptr {} addrspace(10)*, {} addrspace(10)** %4, i32 1, !dbg !8288
  store {} addrspace(10)* %57, {} addrspace(10)** %61, align 4, !dbg !8288
  %62 = getelementptr {} addrspace(10)*, {} addrspace(10)** %4, i32 2, !dbg !8288
  store {} addrspace(10)* %58, {} addrspace(10)** %62, align 4, !dbg !8288
  %63 = call nonnull {} addrspace(10)* @jl_apply_generic({} addrspace(10)* %53, {} addrspace(10)** %4, i32 3), !dbg !8288
  call void @llvm.trap(), !dbg !8288
  unreachable, !dbg !8288

oob15:                                            ; preds = %L64
  %value_phi7.lcssa50 = phi i32 [ %value_phi7, %L64 ]
  %157 = add i32 %value_phi7.lcssa50, -1, !dbg !8343
  %158 = alloca i32, align 4, !dbg !8292
  store i32 %157, i32* %158, align 4, !dbg !8292
  %159 = addrspacecast {} addrspace(10)* %1 to {} addrspace(12)*, !dbg !8292
  call void @jl_bounds_error_ints({} addrspace(12)* %159, i32* nonnull %158, i32 1), !dbg !8292
  unreachable, !dbg !8292

L77.loopexit:                                     ; preds = %idxend16
  br label %L77, !dbg !8293
*** IR Dump After Loop Strength Reduction ***
; Preheader:
L37.preheader:                                    ; preds = %idxend4
  %112 = bitcast {} addrspace(10)* %1 to { i8 addrspace(13)*, i32, i16, i16, i32 } addrspace(10)*, !dbg !8332
  %113 = addrspacecast { i8 addrspace(13)*, i32, i16, i16, i32 } addrspace(10)* %112 to { i8 addrspace(13)*, i32, i16, i16, i32 } addrspace(11)*, !dbg !8332
  %114 = getelementptr inbounds { i8 addrspace(13)*, i32, i16, i16, i32 }, { i8 addrspace(13)*, i32, i16, i16, i32 } addrspace(11)* %113, i32 0, i32 1, !dbg !8332
  %115 = load i32, i32 addrspace(11)* %114, align 4, !dbg !8332, !tbaa !8259
  %116 = bitcast {} addrspace(10)* %1 to i8 addrspace(13)* addrspace(10)*, !dbg !8332
  %117 = addrspacecast i8 addrspace(13)* addrspace(10)* %116 to i8 addrspace(13)* addrspace(11)*, !dbg !8332
  %118 = load i8 addrspace(13)*, i8 addrspace(13)* addrspace(11)* %117, align 4, !dbg !8332
  %119 = bitcast {} addrspace(10)* %1 to { i8 addrspace(13)*, i32, i16, i16, i32 } addrspace(10)*, !dbg !8333
  %120 = addrspacecast { i8 addrspace(13)*, i32, i16, i16, i32 } addrspace(10)* %119 to { i8 addrspace(13)*, i32, i16, i16, i32 } addrspace(11)*, !dbg !8333
  %121 = getelementptr inbounds { i8 addrspace(13)*, i32, i16, i16, i32 }, { i8 addrspace(13)*, i32, i16, i16, i32 } addrspace(11)* %120, i32 0, i32 1, !dbg !8333
  %122 = load i32, i32 addrspace(11)* %121, align 4, !dbg !8333
  %123 = bitcast {} addrspace(10)* %1 to i8 addrspace(13)* addrspace(10)*, !dbg !8333
  %124 = addrspacecast i8 addrspace(13)* addrspace(10)* %123 to i8 addrspace(13)* addrspace(11)*, !dbg !8333
  %125 = load i8 addrspace(13)*, i8 addrspace(13)* addrspace(11)* %124, align 4, !dbg !8333
  %126 = bitcast {} addrspace(10)* %1 to { i8 addrspace(13)*, i32, i16, i16, i32 } addrspace(10)*, !dbg !8334
  %127 = addrspacecast { i8 addrspace(13)*, i32, i16, i16, i32 } addrspace(10)* %126 to { i8 addrspace(13)*, i32, i16, i16, i32 } addrspace(11)*, !dbg !8334
  %128 = getelementptr inbounds { i8 addrspace(13)*, i32, i16, i16, i32 }, { i8 addrspace(13)*, i32, i16, i16, i32 } addrspace(11)* %127, i32 0, i32 1, !dbg !8334
  %129 = load i32, i32 addrspace(11)* %128, align 4, !dbg !8334
  %130 = bitcast {} addrspace(10)* %1 to i8 addrspace(13)* addrspace(10)*, !dbg !8334
  %131 = addrspacecast i8 addrspace(13)* addrspace(10)* %130 to i8 addrspace(13)* addrspace(11)*, !dbg !8334
  %132 = load i8 addrspace(13)*, i8 addrspace(13)* addrspace(11)* %131, align 4, !dbg !8334
  %133 = ptrtoint i8 addrspace(13)* %132 to i32
  %134 = bitcast {} addrspace(10)* %1 to { i8 addrspace(13)*, i32, i16, i16, i32 } addrspace(10)*, !dbg !8335
  %135 = addrspacecast { i8 addrspace(13)*, i32, i16, i16, i32 } addrspace(10)* %134 to { i8 addrspace(13)*, i32, i16, i16, i32 } addrspace(11)*, !dbg !8335
  %136 = getelementptr inbounds { i8 addrspace(13)*, i32, i16, i16, i32 }, { i8 addrspace(13)*, i32, i16, i16, i32 } addrspace(11)* %135, i32 0, i32 1, !dbg !8335
  %137 = load i32, i32 addrspace(11)* %136, align 4, !dbg !8335
  %138 = bitcast {} addrspace(10)* %1 to i8 addrspace(13)* addrspace(10)*, !dbg !8335
  %139 = addrspacecast i8 addrspace(13)* addrspace(10)* %138 to i8 addrspace(13)* addrspace(11)*, !dbg !8335
  %140 = load i8 addrspace(13)*, i8 addrspace(13)* addrspace(11)* %139, align 4, !dbg !8335
  %141 = ptrtoint i8 addrspace(13)* %140 to i32
  %142 = add i32 %2, -2, !dbg !8292
  %143 = add i32 %2, -2, !dbg !8336
  %144 = sub i32 0, %141, !dbg !8280
  %145 = inttoptr i32 %144 to i8 addrspace(13)*
  %146 = sub i32 0, %133, !dbg !8280
  %147 = inttoptr i32 %146 to i8 addrspace(13)*
  %148 = add i32 %2, -2, !dbg !8282
  %149 = add i32 %2, -2, !dbg !8280
  %scevgep62 = getelementptr i8, i8 addrspace(13)* %125, i32 -2, !dbg !8280
  %scevgep66 = getelementptr i8, i8 addrspace(13)* %118, i32 -2, !dbg !8280
  %150 = sub i32 0, %110, !dbg !8280
  %151 = add i32 %150, %2, !dbg !8337
  br label %L37, !dbg !8280

; Loop:
L37:                                              ; preds = %L37.preheader, %idxend16
  %lsr.iv67 = phi i8 addrspace(13)* [ %scevgep66, %L37.preheader ], [ %scevgep68, %idxend16 ]
  %lsr.iv63 = phi i8 addrspace(13)* [ %scevgep62, %L37.preheader ], [ %scevgep64, %idxend16 ]
  %lsr.iv58 = phi i8 addrspace(13)* [ %147, %L37.preheader ], [ %scevgep59, %idxend16 ]
  %lsr.iv55 = phi i8 addrspace(13)* [ %145, %L37.preheader ], [ %scevgep, %idxend16 ]
  %lsr.iv53 = phi i32 [ 0, %L37.preheader ], [ %lsr.iv.next54, %idxend16 ]
  %lsr.iv = phi i32 [ -1, %L37.preheader ], [ %lsr.iv.next, %idxend16 ]
  %lsr.iv5860 = ptrtoint i8 addrspace(13)* %lsr.iv58 to i32
  %lsr.iv5556 = ptrtoint i8 addrspace(13)* %lsr.iv55 to i32
  %44 = add i32 %2, %lsr.iv, !dbg !8280
  %45 = add i32 %149, %lsr.iv53, !dbg !8280
  %46 = icmp ult i32 %45, %115, !dbg !8280
  br i1 %46, label %idxend10, label %oob9, !dbg !8280

idxend10:                                         ; preds = %L37
  %scevgep69 = getelementptr i8, i8 addrspace(13)* %lsr.iv67, i32 %2, !dbg !8280
  %154 = load i8, i8 addrspace(13)* %scevgep69, align 1, !dbg !8280, !tbaa !8300
  %155 = icmp eq i8 %154, 58, !dbg !8339
  br i1 %155, label %L49, label %L47, !dbg !8281

L49:                                              ; preds = %idxend10
  %51 = add i32 %148, %lsr.iv53, !dbg !8282
  %52 = icmp ult i32 %51, %122, !dbg !8282
  br i1 %52, label %idxend12, label %oob11, !dbg !8282

idxend12:                                         ; preds = %L49
  %scevgep65 = getelementptr i8, i8 addrspace(13)* %lsr.iv63, i32 %2, !dbg !8282
  store i8 48, i8 addrspace(13)* %scevgep65, align 1, !dbg !8282, !tbaa !8300
  %158 = add i32 %143, %lsr.iv53, !dbg !8336
  %159 = add i32 %158, -1, !dbg !8336
  %160 = icmp ult i32 %159, %129, !dbg !8336
  br i1 %160, label %idxend14, label %oob13, !dbg !8336

idxend14:                                         ; preds = %idxend12
  %164 = sub i32 0, %lsr.iv5860, !dbg !8336
  %165 = add i32 %164, %143, !dbg !8336
  %166 = inttoptr i32 %165 to i8 addrspace(13)*
  %scevgep61 = getelementptr i8, i8 addrspace(13)* %166, i32 -1, !dbg !8336
  %167 = load i8, i8 addrspace(13)* %scevgep61, align 1, !dbg !8336, !tbaa !8300
  %168 = zext i8 %167 to i32, !dbg !8346
  %169 = add nuw nsw i32 %168, 1, !dbg !8284
  %170 = trunc i32 %169 to i8, !dbg !8351
  %171 = icmp eq i8 %167, -1, !dbg !8288
  br i1 %171, label %L59, label %L64, !dbg !8288

L64:                                              ; preds = %idxend14
  %65 = add i32 %142, %lsr.iv53, !dbg !8292
  %66 = add i32 %65, -1, !dbg !8292
  %67 = icmp ult i32 %66, %137, !dbg !8292
  br i1 %67, label %idxend16, label %oob15, !dbg !8292

idxend16:                                         ; preds = %L64
  %175 = sub i32 0, %lsr.iv5556, !dbg !8292
  %176 = add i32 %175, %143, !dbg !8292
  %177 = inttoptr i32 %176 to i8 addrspace(13)*
  %scevgep57 = getelementptr i8, i8 addrspace(13)* %177, i32 -1, !dbg !8292
  store i8 %170, i8 addrspace(13)* %scevgep57, align 1, !dbg !8292, !tbaa !8300
  %lsr.iv.next = add i32 %lsr.iv, -1, !dbg !8337
  %lsr.iv.next54 = add i32 %lsr.iv53, -1, !dbg !8337
  %scevgep = getelementptr i8, i8 addrspace(13)* %lsr.iv55, i32 1, !dbg !8337
  %scevgep59 = getelementptr i8, i8 addrspace(13)* %lsr.iv58, i32 1, !dbg !8337
  %scevgep64 = getelementptr i8, i8 addrspace(13)* %lsr.iv63, i32 -1, !dbg !8337
  %scevgep68 = getelementptr i8, i8 addrspace(13)* %lsr.iv67, i32 -1, !dbg !8337
  %178 = add i32 %151, %lsr.iv.next54, !dbg !8337
  %179 = icmp eq i32 %178, 0, !dbg !8337
  br i1 %179, label %L77.loopexit, label %L37, !dbg !8287

; Exit blocks
oob9:                                             ; preds = %L37
  %value_phi7.lcssa = phi i32 [ %44, %L37 ]
  %152 = alloca i32, align 4, !dbg !8280
  store i32 %value_phi7.lcssa, i32* %152, align 4, !dbg !8280
  %153 = addrspacecast {} addrspace(10)* %1 to {} addrspace(12)*, !dbg !8280
  call void @jl_bounds_error_ints({} addrspace(12)* %153, i32* nonnull %152, i32 1), !dbg !8280
  unreachable, !dbg !8280

L47:                                              ; preds = %idxend10
  %.sroa.027.0..sroa_idx = getelementptr inbounds [2 x i32], [2 x i32]* %0, i32 0, i32 0, !dbg !8281
  store i32 %2, i32* %.sroa.027.0..sroa_idx, align 4, !dbg !8281
  %.sroa.228.0..sroa_idx29 = getelementptr inbounds [2 x i32], [2 x i32]* %0, i32 0, i32 1, !dbg !8281
  store i32 %3, i32* %.sroa.228.0..sroa_idx29, align 4, !dbg !8281
  %47 = getelementptr {} addrspace(10)*, {} addrspace(10)** %gcframe, i32 1
  %48 = load {} addrspace(10)*, {} addrspace(10)** %47, align 4, !tbaa !8249
  %49 = getelementptr {}**, {}*** %ptls, i32 0
  %50 = bitcast {}*** %49 to {} addrspace(10)**
  store {} addrspace(10)* %48, {} addrspace(10)** %50, align 4, !tbaa !8249
  ret void, !dbg !8281

oob11:                                            ; preds = %L49
  %value_phi7.lcssa39 = phi i32 [ %44, %L49 ]
  %156 = alloca i32, align 4, !dbg !8282
  store i32 %value_phi7.lcssa39, i32* %156, align 4, !dbg !8282
  %157 = addrspacecast {} addrspace(10)* %1 to {} addrspace(12)*, !dbg !8282
  call void @jl_bounds_error_ints({} addrspace(12)* %157, i32* nonnull %156, i32 1), !dbg !8282
  unreachable, !dbg !8282

oob13:                                            ; preds = %idxend12
  %value_phi7.lcssa51 = phi i32 [ %44, %idxend12 ]
  %161 = add i32 %value_phi7.lcssa51, -1, !dbg !8345
  %162 = alloca i32, align 4, !dbg !8336
  store i32 %161, i32* %162, align 4, !dbg !8336
  %163 = addrspacecast {} addrspace(10)* %1 to {} addrspace(12)*, !dbg !8336
  call void @jl_bounds_error_ints({} addrspace(12)* %163, i32* nonnull %162, i32 1), !dbg !8336
  unreachable, !dbg !8336

L59:                                              ; preds = %idxend14
  %.lcssa47 = phi i32 [ %169, %idxend14 ], !dbg !8284
  %53 = load {}*, {}** @jl_globalYY.16, align 4, !dbg !8288, !tbaa !8276, !nonnull !4
  %54 = addrspacecast {}* %53 to {} addrspace(10)*, !dbg !8288
  %55 = load {}*, {}** @jl_symYY.trunc61, align 4, !dbg !8288, !tbaa !8276, !nonnull !4
  %56 = addrspacecast {}* %55 to {} addrspace(10)*, !dbg !8288
  %57 = load {}*, {}** @SUM.CoreDOT.UInt8254, align 4, !dbg !8288, !tbaa !8276, !nonnull !4, !dereferenceable !8278, !align !8279
  %58 = addrspacecast {}* %57 to {} addrspace(10)*, !dbg !8288
  %59 = call {} addrspace(10)* @jl_box_int32(i32 signext %.lcssa47), !dbg !8288
  %60 = getelementptr {} addrspace(10)*, {} addrspace(10)** %gcframe, i32 2
  store {} addrspace(10)* %59, {} addrspace(10)** %60
  %61 = getelementptr {} addrspace(10)*, {} addrspace(10)** %4, i32 0, !dbg !8288
  store {} addrspace(10)* %56, {} addrspace(10)** %61, align 4, !dbg !8288
  %62 = getelementptr {} addrspace(10)*, {} addrspace(10)** %4, i32 1, !dbg !8288
  store {} addrspace(10)* %58, {} addrspace(10)** %62, align 4, !dbg !8288
  %63 = getelementptr {} addrspace(10)*, {} addrspace(10)** %4, i32 2, !dbg !8288
  store {} addrspace(10)* %59, {} addrspace(10)** %63, align 4, !dbg !8288
  %64 = call nonnull {} addrspace(10)* @jl_apply_generic({} addrspace(10)* %54, {} addrspace(10)** %4, i32 3), !dbg !8288
  call void @llvm.trap(), !dbg !8288
  unreachable, !dbg !8288

oob15:                                            ; preds = %L64
  %value_phi7.lcssa50 = phi i32 [ %44, %L64 ]
  %172 = add i32 %value_phi7.lcssa50, -1, !dbg !8345
  %173 = alloca i32, align 4, !dbg !8292
  store i32 %172, i32* %173, align 4, !dbg !8292
  %174 = addrspacecast {} addrspace(10)* %1 to {} addrspace(12)*, !dbg !8292
  call void @jl_bounds_error_ints({} addrspace(12)* %174, i32* nonnull %173, i32 1), !dbg !8292
  unreachable, !dbg !8292

L77.loopexit:                                     ; preds = %idxend16
  br label %L77, !dbg !8293

[yuyichao]

Bugpoint reduced IR:

target datalayout = "e-m:e-p:32:32-Fi8-i64:64-v128:64:128-a:0:32-n32-S64-ni:10:11:12:13"
target triple = "armv7l-unknown-linux-gnueabihf"

define dso_local void @julia_roundup_19730() {
top:
  br label %L37

L37:                                              ; preds = %L49, %top
  %value_phi7 = phi i32 [ %3, %L49 ], [ 0, %top ]
  br i1 undef, label %L49, label %L47

L47:                                              ; preds = %L37
  unreachable

L49:                                              ; preds = %L37
  %0 = add i32 %value_phi7, -2
  %1 = getelementptr inbounds i8, i8 addrspace(13)* null, i32 %0
  %2 = load i8, i8 addrspace(13)* %1, align 1
  %3 = add i32 %value_phi7, -1
  br label %L37
}

Reported as https://bugs.llvm.org/show_bug.cgi?id=46757