StackOverflow with no stacktrace

[SimonDanisch]

I've heard this issue might have been raised before, but I couldn't find it ;)

> .\julia.exe --depwarn=no --inline=no -O1
  | | |_| | | | (_| |  |  Version 0.7.0-beta2.12 (2018-07-15 15:57 UTC)
 _/ |\__'_|_|_|\__'_|  |  Commit a878341b2b* (0 days old master)
|__/                   |  x86_64-w64-mingw32
julia> GPUArrays
julia> fill(JLArray{Float32}, Float32, (4,4))
ERROR: StackOverflowError:
julia>

With gdb:

Starting program: C:\Users\sdani\AppData\Local\Julia-0.7.0-beta2\bin\julia.exe "--depwarn=no" "--inline=no" -O1
  | | |_| | | | (_| |  |  Version 0.7.0-beta2.12 (2018-07-15 15:57 UTC)
 _/ |\__'_|_|_|\__'_|  |  Commit a878341b2b* (0 days old master)
|__/                   |  x86_64-w64-mingw32

julia> [New Thread 244.0x4174]
julia> using GPUArrays
julia> fill(JLArray{Float32}, Float32, (4,4))
Program received signal SIGSEGV, Segmentation fault.
0x00007ff8817e521a in ntdll!RtlSizeHeap () from C:\WINDOWS\SYSTEM32\ntdll.dll
(gdb) backtrace
#0  0x00007ff8817e521a in ntdll!RtlSizeHeap () from C:\WINDOWS\SYSTEM32\ntdll.dll
#1  0x00007ff8817e2b55 in ntdll!RtlAllocateHeap () from C:\WINDOWS\SYSTEM32\ntdll.dll
#2  0x00007ff8818cfade in ntdll!RtlpNtSetValueKey () from C:\WINDOWS\SYSTEM32\ntdll.dll
#3  0x00007ff881878f2a in ntdll!memset () from C:\WINDOWS\SYSTEM32\ntdll.dll
Backtrace stopped: previous frame inner to this frame (corrupt stack?)
(gdb)

It might be related to broadcasting + inference, since when I try walking down the whole call tree manually, everything just works.

This is what gets called:

julia> x = similar(JLArray, Float32, (4,4))
4×4 JLArray{Float32,2}:
...
julia> @code_warntype fill!(x, 0f0)
Body::Union{}
13 1 ─ %1 = :(Base.Broadcast.materialize!)::Core.Compiler.Const(Base.Broadcast.materialize!, false)                    │
   │   %2 = :(Base.Broadcast.broadcasted)::Core.Compiler.Const(Base.Broadcast.broadcasted, false)                      │
   │   %3 = :($(Expr(:static_parameter, 1)))(%%val)::Float32                                                           │
   │   %4 = %2(GPUArrays.identity, %3)::Base.Broadcast.Broadcasted{Base.Broadcast.DefaultArrayStyle{0},Nothing,typeof(identity),Tuple{Float32}}
   │        %1(%%A, %4)                                                                                                │
   └──      unreachable                                                                                                │
   2 ─      unreachable                                                                                                │

[KristofferC]

Could #28016 have something to do with the stacktrace output? Worth testing.

[SimonDanisch]

Did you link the wrong issue, or do I just fail to see the connection? ;)

[KristofferC]

Heh, fixed.

[SimonDanisch]

Some more data:

julia> x = similar(JLArray, Float32, (4, 4))
julia> bc = Broadcast.broadcasted(identity, 0f0)
julia> y = Broadcast.instantiate(Broadcast.Broadcasted(identity, (bc,), axes(x)))
julia> copyto!(x, y)# works:
4×4 JLArray{Float32,2}:
 0.0  0.0  0.0  0.0
 0.0  0.0  0.0  0.0
 0.0  0.0  0.0  0.0
 0.0  0.0  0.0  0.0

[SimonDanisch]

Ah I see...
Followed it more closely (basically exactly executing the IR) and get this now:

julia> @which Base.Broadcast.copyto!(x, ssa6)
copyto!(dest::AbstractArray, bc::Base.Broadcast.Broadcasted{#s608,Axes,F,Args} where Args<:Tuple where F where Axes where #s608<:Base.Broadcast.AbstractArrayStyle{0}) in Base.Broadcast at broadcast.jl:779

https://github.com/JuliaLang/julia/blob/master/base/broadcast.jl#L781

Which calls fill :D

[Liozou]

I tried to reproduce but I can't seem to use GPUArrays... Is the package up to date?

[vtjnash]

Fixed by #13099

[KristofferC]

That PR is merged so closing.