Disable LLVM Global Isel

[jhass]

It's not enabled by default on x86, but on other targets it is. At least until https://reviews.llvm.org/D80898 is released, it doesn't like us generating a value of a zero sized type

We may need to fix doing that, but until then this workarounds the issue.

See #9297 (comment) for more background.

[jhass]

Oh yeah, fixes #8080 / worksaround (I didn't know we had an issue ticket for this).

[waj]

@jhass is it GlobalIsel used in x86_64? Do you know what are the zero sized values that are currently being generated?

[jhass]

According to the hints I got in IRC it's not by default and barely implemented there.

I don't know for certain why they are generated, but I suspect something around nil.

I found it through bugpoint, so it's a bit hard to correlate back to the original code.

[asterite]

nil is such a type:

crystal/src/compiler/crystal/codegen/llvm_typer.cr

Line 68 in ecea8f9

@nil_type ||= @structs["Nil"] ||= @llvm_context.struct([] of LLVM::Type, "Nil")

[jhass]

Yeah, I forgot, in the comment linked above I actually related it back to (an instance of) the original code. So it has to do with nil return type procs.

[asterite]

Strange, for nil return type procs I would expect that to translate to void, but maybe I'm wrong.

Also, nothing prevents a user from creating an empty struct...

[jhass]

Also LLVM acknowledged and fixed this as a bug, so while I got a few why the hell do you do this in response, it seems legit if we do the easy thing right now and apply this workaround.

[waj]

Ok, but what's the current implication of disabling it? Are we going to get less performance in some platforms or what?

[jhass]

Maybe? But I think it's only slightly and a big point is also an architectural change within LLVM allowing them to do more or different things in the future.

https://releases.llvm.org/9.0.0/docs/GlobalISel.html