Nightly codegen hang

[christiangnrd]

PNMWE (Probably-Not-Minimal Working Example 🙂). Taken from

GPUCompiler.jl/test/native.jl

Line 307 in 3084db6

Native.code_llvm(devnull, Base.print_to_string, Tuple{Int,Int}; optimize=false)

julia> using GPUCompiler; include("test/helpers/runtime.jl"); include("test/helpers/native.jl")
Main.Native

julia> #=Works=# Native.code_llvm(devnull, Base.print_to_string, Tuple{Int,Int}; optimize=false)

julia> #=Hangs=# Native.code_llvm(devnull, Base.print_to_string, Tuple{Int,Int}; optimize=true)

#771 has slightly more discussion.

[maleadt]

Workaround:

232          add!(mpm, FinalLowerGCPass())
233      end
234
235 +    # NOTE: LateLowerGCPass does not correctly invalidate LLVM analyses in the
236 +    # new pass manager, causing stale MemoryDependenceResults/AAResults that make
237 +    # GVN hang. Work around this by explicitly invalidating all analyses.
239 +    add!(mpm, InvalidateAllAnalysesPass())
240 -
235      if opt_level >= 2
236          add!(mpm, NewPMFunctionPassManager()) do fpm
237              add!(fpm, GVNPass()

Trying to figure out where this goes wrong. I don't think it's LateLowerGCPass, but maybe another pass.

[maleadt]

JuliaLang/julia#61422 should fix this