Error: unsigned int convert_struct_offset(llvm::Type*, unsigned int): Assertion `SL->getElementOffset(idx) == byte_offset' failed

[atbug]

I don't have a MWE yet but here it is in my package CI for Julia 1.7:

https://github.com/Hop-developers/Hop.jl/runs/6604926044

This issue doesn't happen on Julia 1.6.

If anyone tries to look into this, to test Hop.jl (not registered yet), one needs to put all the data in https://github.com/Hop-developers/HopTestData into the test/data folder in Hop.jl and then run runtests.jl.

[atbug]

A similar issue (fixed) here: #34288

[vtjnash]

Are you familiar enough with GDB that you could print out the value of the type that is failing there (e.g. print jl_dump_llvm_type(lty))

[atbug]

Are you familiar enough with GDB that you could print out the value of the type that is failing there (e.g. print jl_dump_llvm_type(lty))

Unfortunately, no. But I have reduced the example to something that does not need HopTestData:

using Hop
tm = Hop.SharedTBModel(Hop.Zoo.getBN())
Hop.Optics.get_shift_cond(tm, 2, 2, [3.2], -6.0, [5, 5, 1], ϵ=0.1)

It seems the problem comes from SharedArrays.

I also tried to obtain an rr trace here:
https://s3.amazonaws.com/julialang-dumps/reports/2022-05-26T19-31-02-atbug.tar.zst

But I am not sure I ran rr correctly since it complained: "[FATAL src/PerfCounters.cc:219:start_counter()] Unable to open performance counter with 'perf_event_open'; are hardware perf events available? See https://github.com/rr-debugger/rr/wiki/Will-rr-work-on-my-system". I was running rr on a vultr instance since I don't have sudo access on my cluster.

[vtjnash]

Looks like the rr file is empty.

Thread 1 "julia" received signal SIGABRT, Aborted.
__GI_raise (sig=sig@entry=6) at ../sysdeps/unix/sysv/linux/raise.c:50
50      ../sysdeps/unix/sysv/linux/raise.c: No such file or directory.
(gdb) up
#1  0x00007ffff7da9859 in __GI_abort () at abort.c:79
79      abort.c: No such file or directory.
(gdb) 
#2  0x00007ffff7da9729 in __assert_fail_base (fmt=0x7ffff7f3f588 "%s%s%s:%u: %s%sAssertion `%s' failed.\n%n", assertion=0x7ffff77498a0 "SL->getElementOffset(idx) == byte_offset", 
    file=0x7ffff7748930 "/data/vtjnash/julia17/src/cgutils.cpp", line=480, function=<optimized out>) at assert.c:92
92      assert.c: No such file or directory.
(gdb) 
#3  0x00007ffff7dbb006 in __GI___assert_fail (assertion=assertion@entry=0x7ffff77498a0 "SL->getElementOffset(idx) == byte_offset", 
    file=file@entry=0x7ffff7748930 "/data/vtjnash/julia17/src/cgutils.cpp", line=line@entry=480, 
    function=function@entry=0x7ffff7749860 "unsigned int convert_struct_offset(llvm::Type*, unsigned int)") at assert.c:101
101     in assert.c
(gdb) 
#4  0x00007ffff756bc46 in convert_struct_offset (lty=<optimized out>, byte_offset=<optimized out>) at /data/vtjnash/julia17/src/cgutils.cpp:475
475     static unsigned convert_struct_offset(Type *lty, unsigned byte_offset)
(gdb) 
#5  convert_struct_offset (lty=<optimized out>, byte_offset=<optimized out>) at /data/vtjnash/julia17/src/cgutils.cpp:475
475     static unsigned convert_struct_offset(Type *lty, unsigned byte_offset)
(gdb) 
#6  0x00007ffff759aae8 in convert_struct_offset (ctx=..., byte_offset=208, lty=0x55555767e8e0) at /data/vtjnash/julia17/src/cgutils.cpp:486
486         return convert_struct_offset(lty, byte_offset);
(gdb) p lty
$1 = (llvm::Type *) 0x55555767e8e0
(gdb) p jl_dump_llvm_type(lty)
{ double, { i64, i8, [1 x [9 x double]], [1 x [9 x double]], {} addrspace(10)*, {} addrspace(10)*, {} addrspace(10)*, {} addrspace(10)*, [1 x [3 x {} addrspace(10)*]] }, i64, i64, i64, {} addrspace(10)*, double, {} addrspace(10)* }
$2 = void
(gdb) up
#7  emit_new_struct (ctx=..., ty=0x7fff8e55e720, nargs=8, argv=<optimized out>) at /data/vtjnash/julia17/src/cgutils.cpp:3439
3439                    unsigned llvm_idx = (i > 0 && isa<StructType>(lt)) ? convert_struct_offset(ctx, lt, offs) : i;
(gdb) list -
3434                    Type *fty = julia_type_to_llvm(ctx, jtype);
3435                    if (type_is_ghost(fty))
3436                        continue;
3437                    Value *dest = NULL;
3438                    unsigned offs = jl_field_offset(sty, i);
3439                    unsigned llvm_idx = (i > 0 && isa<StructType>(lt)) ? convert_struct_offset(ctx, lt, offs) : i;
3440                    if (!init_as_value) {
3441                        // avoid unboxing the argument explicitly
3442                        // and use memcpy instead
3443                        dest = ctx.builder.CreateConstInBoundsGEP2_32(lt, strct, 0, llvm_idx);
(gdb) p jl_(sty)
Hop.Optics.var"#11#13"{Float64, Hop.SharedTBModel{Base.Complex{Float64}}, Int64, Int64, Int64, Array{Float64, 1}, Float64, Array{SharedArrays.SharedArray{Float64, 1}, 1}}
$3 = void
(gdb) p jl_(sty->types)
svec(Float64, Hop.SharedTBModel{Base.Complex{Float64}}, Int64, Int64, Int64, Array{Float64, 1}, Float64, Array{SharedArrays.SharedArray{Float64, 1}, 1})
$4 = void
(gdb) p offs
$5 = 208
(gdb) p i
$6 = 2

This appears to be fixed on master.

julia> fieldtypes(Hop.SharedTBModel{Base.Complex{Float64}})
(Int64, Bool, StaticArrays.SMatrix{3, 3, Float64, 9}, StaticArrays.SMatrix{3, 3, Float64, 9}, Matrix{Int16}, Matrix{Float64}, SharedArrays.SharedMatrix{ComplexF64}, Union{Nothing, SharedArrays.SharedMatrix{ComplexF64}}, StaticArrays.SVector{3, SharedArrays.SharedMatrix{ComplexF64}})

julia> map(i -> fieldoffset(Hop.SharedTBModel{Base.Complex{Float64}}, i), 1:9)
9-element Vector{UInt64}:
 0x0000000000000000
 0x0000000000000008
 0x0000000000000010
 0x0000000000000058
 0x00000000000000a0
 0x00000000000000a8
 0x00000000000000b0
 0x00000000000000b8
 0x00000000000000c0

julia> sizeof(Hop.SharedTBModel{Base.Complex{Float64}}) |> UInt
0x00000000000000c8 # v1.7.3
julia> sizeof(Hop.SharedTBModel{Base.Complex{Float64}}) |> UInt
0x00000000000000d8 # master

[vtjnash]

I believe this will not be fixed as there is not a plan for another v1.7 release, given that v1.8.0 is about to be available.

[atbug]

Fair enough. Thank you.