Instruction does not dominate all uses!, getelementptr inbounds when testing DimensionalData

[KristofferC]

Happens on master and backports-release-1.9.

pkg> add DimensionalData; test DimensionalData
...
Instruction does not dominate all uses!
  %47 = getelementptr inbounds [1 x { [2 x double], [2 x [1 x i64]] }], [1 x { [2 x double], [2 x [1 x i64]] }]* %2, i32 0, i32 0, !dbg !135
  %44 = getelementptr inbounds { [2 x double], [2 x [1 x i64]] }, { [2 x double], [2 x [1 x i64]] }* %47, i32 0, i32 1, !dbg !122
Assertion failed: (!verifyModule(M, &errs())), function operator(), file jitlayers.cpp, line 1155.

[33812] signal (6): Abort trap: 6
in expression starting at /Users/kristoffercarlsson/.julia/packages/DimensionalData/1wSpb/test/selector.jl:1126
__pthread_kill at /usr/lib/system/libsystem_kernel.dylib (unknown line)
Allocations: 57098308 (Pool: 57051008; Big: 47300); GC: 76
ERROR: Package DimensionalData errored during testing (received signal: 6)

Needs asserts builds.

[gbaraldi]

Reduced to

using DimensionalData, Unitful, Extents
using DimensionalData.LookupArrays


a = [1 2  3  4
     5 6  7  8
     9 10 11 12]

da = DimArray(a,(Y(1:3),Ti(Sampled((1:4)u"s"; sampling=Intervals()))))
Extents.extent(da)

[gbaraldi]

5378a31 bad, release-1.8 good

Bad here isn't the first bad commit, just a range.

[gbaraldi]

fe72f81 Is the first bad commit.

[KristofferC]

I guess that's just the place where we actually started verifying this. :/

[gbaraldi]

Yeah.

[gbaraldi]

I'm pretty sure it's #45153 :)

[vtjnash]

Looks probably true, since here is the IR for it, which fails with

Instruction does not dominate all uses!
  %33 = getelementptr inbounds [1 x { [2 x i64], [2 x [1 x i64]] }], [1 x { [2 x i64], [2 x [1 x i64]] }]* %2, i32 0, i32 0, !dbg !102
  %30 = getelementptr inbounds { [2 x i64], [2 x [1 x i64]] }, { [2 x i64], [2 x [1 x i64]] }* %33, i32 0, i32 1, !dbg !89

@code_llvm optimize=false dump_module=true raw=true Extents.extent(da)

48917julia.txt
48917llvm.txt

[gbaraldi]

I'm running creduce on it, because the types are a bit massive on that example so it's a bit harder to see where we go wrong.

[vtjnash]

manually "creduce" yielded:

julia> f(x, w) = (y = (a=1, b=x); z = (; a=(a=(1, w), b=(3, y))))

make a struct early, then do a lot of indexing of wrappers that includes it

[gbaraldi]

Jameson reduce is a lot faster than creduce 😆 . Thanks!