Fatal error in type inference adding a GC frame

[mbauman]

Version: 0.4.0-dev+6494 e66175b

Minimal example:

@inline function f(A::AbstractArray, I1, I2, I3)
    (size(A,1) >= I1 && size(A,2) >= I2) && size(A,3) >= I3
end
h(A) = f(A, 1, 2, 3)

code_typed(f, (Array{Int, 3}, Int, Int, Int)) # is ok
code_typed(h, (Array{Int, 3},)) # has an unreachable error("fatal error in type inference")
code_llvm(f, (Array{Int, 3}, Int, Int, Int)) # no gc frame
code_llvm(h, (Array{Int, 3},)) # gc frame (with nothing put into it, I think)

Full output: https://gist.github.com/mbauman/a4e9fb36f67e2287a8f6

[mbauman]

It's not that big of a deal to have assertion checks like that, especially when LLVM can verify that things were ok and drop the dead code. But perhaps we could make it a non-allocating error type instead to prevent GC frames?

[yuyichao]

I would still prefer not having to worry about gc frames when an error is thrown. I think my throw local gcframe branch can already handle this and I believe when the codegen rewrite is merged there's even more we can do about it.

Any idea why type inference decide to add an assertion?

[mbauman]

Agreed completely, but this is blocking #11895. No clue as to what's going on, but this is about as simple of an example as it gets, so hopefully one of you inference gurus can figure it out.

[yuyichao]

Hmmm throw(ErrorException("....")) should not create a gc frame and code_llvm agrees. However, there's something wrong with the codegen of error(...). It seems that type inference could not infer that Main.Base is statically known?

[yuyichao]

Maybe you can try to have different definitions of error for base and inference?

There also seems to be a missing Main.Base. for systemerror (although I'm not sure if it is used in inference)...

[yuyichao]

Or replace Expr(:call,:error,"fatal error in type inference") with Expr(:call, :throw, Expr(:new, :ErrorException, "fatal error in type inference")) just for testing.

[mbauman]

Thanks for the quick hack!

Inference can't figure it out because Main isn't a constant binding. Why is error.jl different from all the other tools inference uses (which get copied)?

[yuyichao]

Other tools?

[mbauman]

All the other stuff in coreimg.jl. My knowledge here is rudimentary, but aren't all those functions and types duplicated and frozen into Core.Inference? They don't reach out and try to grab the current methods and types from Main like error.jl is doing.

[yuyichao]

I won't claim I know the bootstrap very well but it does seems a little bit strange that we pull error() but not ErrorException into Core.Inference....

[JeffBezanson]

We should just remove that. Internal compiler errors shouldn't be emitted into generated code.

[carnaval]

yep. maybe emit it in debug mode and only have it be an unreachable llvm op in release.

[mbauman]

The original testcase here now passes… and the fatal error line has changed to use TopNode/GlobalRef in fba99fd which I think solves the root issue.

[tkelman]

Do we have test cases that capture this running as part of the test suite?