Assertion failure on 1.12

[maleadt]

julia: /source/src/gf.c:1885: ijl_method_instance_add_backedge: Assertion `(((((jl_taggedvalue_t*)((char*)(callee) - sizeof(jl_taggedvalue_t)))->header) & ~(uintptr_t)15)==(uintptr_t)(jl_method_instance_type))' failed.

[177] signal 6 (-6): Aborted
in expression starting at /home/pkgeval/.julia/packages/Tricks/UWPzc/test/runtests.jl:84

I'm going to assume this is due to evil functionality in this package rather than a bug in Base, and am going to blacklist Tricks.jl in PkgEval.jl. Feel free to ping me when that should be reverted.

[oxinabox]

I am guessing something has changed with when you are allowed to add to ci.edges, or what format they have to be in.
I guess that returning CodeInfo from generated functions has never been part of the public API.

I am guessing this broke Zygote, IRTools and Cassette also?

[maleadt]

Zygote tests never ran in the first place (depend on CUDA), IRTools fails tests but doesn't crash as badly as Tricks, and the same applies to Cassette.

cc @vtjnash @aviatesk, maybe they can shed some light on what has to change here.

[vtjnash]

The format changed, though we could probably write a verifier in Julia that ensures they follow the basics of these rules without a hard assert in C: https://hackmd.io/sjPig55kS4a5XNWC6HmKSg#Edges-Encoding

[chengchingwen]

It seems that jl_method_instance_add_backedge now requires the caller to be a CodeInstance instead of MethodInstance thus we need an AbstractInterpreter to get CodeInstance but that is unavailable inside the generator?

[vtjnash]

It can still be a MethodInstance (representing that you didn't do inference). The constraints that must be met are:

1. Any edge must come from exactly Core.Compiler.findsup or Core.Compiler.findall. Any other source that creates edges is illegal. This isn't exactly a new requirement, however, findsup and findall have only recently been documented and made public.
2. The edges vector must follow exactly the encoding in that hackmd. Most easily done just by calling Core.Compiler.add_edges(edges::Vector{Any}, ::MethodMatchInfo) on the result gotten from 1

[chengchingwen]

I think I have found the part that caused the crash. It seems the way method table is encoded in edges has slightly changed from [method_table, type] to [type, method_table]. Reverse that order in the code would allow the test to run. However, I couldn't pass all Tricks.jl's test because the method deletion somehow doesn't trigger the recompilation.

[chengchingwen]

BTW what VERSION should I use to branch for the new edge format? I'm currently using VERSION > v"1.12.0-DEV"

[chengchingwen]

However, I couldn't pass all Tricks.jl's test because the method deletion somehow doesn't trigger the recompilation.

This seems to be a separate issue:

Tricks.jl/test/runtests.jl

Lines 157 to 166 in ac53de4

	@testset "delete method" begin
	while length(collect(methods(j))) > 0
	Base.delete_method((first ∘ methods)(j))
	T = Tuple{Any, Vararg{Any}}
	@test (length ∘ collect ∘ static_methods)(j, T) == length(collect(methods(j)))
	@test static_method_count(j, T) == length(collect(methods(j)))
	end
	end

This loop couldn't trigger recompilation, but if manually unrolled, the recompilation works.

[aviatesk]

BTW what VERSION should I use to branch for the new edge format? I'm currently using VERSION > v"1.12.0-DEV"

VERSION >= v"1.12.0-DEV.1531" should work.