Loss of invalidation-logging on master

[timholy]

SnoopCompile has tests that look like this:

PkgC.jl

 module PkgC

nbits(::Int8) = 8
nbits(::Int16) = 16

end # module PkgC

PkgD.jl:

module PkgD

using PkgC

call_nbits(x::Integer) = PkgC.nbits(x)
# The rest forces precompilation
map_nbits(list) = map(call_nbits, list)
nbits_list() = map_nbits(Integer[Int8(1), Int16(1)])

nbits_list()

end # module PkgD

and after precompilation,

using PkgC
PkgC.nbits(::UInt8) = 8
using PkgD

The method definition for UInt8 prepares the way to invalidate PkgD.call_nbits, but PkgD is not yet loaded at the time of the invalidation.

On 1.9rc1, one gets the following entries in the _jl_debug_method_invalidation log:

   Tuple{typeof(PkgC.nbits), Integer}
   "insert_backedges_callee"
 56
   Any[nbits(::UInt8) @ Main REPL[9]:4]
   MethodInstance for PkgD.call_nbits(::Integer)
   "verify_methods"
 56

whereas on nightly we get

  Tuple{typeof(PkgC.nbits), Integer}
  "insert_backedges_callee"
 7
  nothing
  MethodInstance for PkgD.call_nbits(::Integer)
  "verify_methods"
 7

(The specific integer values don't matter as long as they are internally consistent.)

The matches log went from a list that contains the method that caused the invalidation to being nothing. Consequently it's not possible to determine the cause of the invalidation.

[timholy]

git blame points to #48841