Missing elements in invalidation log

[timholy]

#57934 had a side effect of excluding some elements from the invalidation logs processed by SnoopCompile. Here's a MWE. The distinguishing characteristic is the presence of both a poorly-specialized and invoked caller with the same signature:

f(::Integer) = 1
callsfrts(x) = f(Base.inferencebarrier(x)::Signed)
invokesfs(x) = invoke(f, Tuple{Signed}, x)

# compilation
invokesfs(1)                        # invoked callee
callsfrts(1)                        # runtime-dispatched callee

# invalidation
logmeths = ccall(:jl_debug_method_invalidation, Any, (Cint,), 1);
f(::Int) = 2
f(::Signed) = 4
ccall(:jl_debug_method_invalidation, Any, (Cint,), 0);

m = which(f, (Signed,))
m ∈ logmeths

You want the final line to be true (and it was before #57934).

Here's the backedge graph before invalidation:

julia> m = only(methods(f))
f(::Integer)
     @ Main REPL[1]:1

julia> mis = collect(Base.specializations(m))
2-element Vector{Any}:
 MethodInstance for f(::Int64)
 MethodInstance for f(::Signed)

julia> mi = mis[1]
MethodInstance for f(::Int64)

julia> mi.backedges
2-element Vector{Any}:
 Tuple{typeof(f), Signed}
 CodeInstance for MethodInstance for invokesfs(::Int64)

julia> mi2 = mis[2]
MethodInstance for f(::Signed)

julia> mi2.backedges
1-element Vector{Any}:
 CodeInstance for MethodInstance for callsfrts(::Int64)

[Keno]

You want the final line to be true (and it was before #57934).

Why? That method is the newest method added, so by definition none of its code instances were invalidated.

[timholy]

The log is broken into sections, and each section terminates with the new method that triggered the preceding set of invalidations. That way SnoopCompile can identify the "culprit."

[timholy]

BTW that demo works on Julia 1.10 too. On 1.10:

julia> logmeths
14-element Vector{Any}:
  MethodInstance for callsfrts(::Int64)
 1
  MethodInstance for f(::Signed)
  "jl_method_table_insert"
  f(::Int64) @ Main REPL[7]:1                                  # method responsible for the first section
  "jl_method_table_insert"
  MethodInstance for invokesfs(::Int64)
 1
  MethodInstance for f(::Int64)
  "jl_method_table_insert"
  MethodInstance for f(::Signed)
  "jl_method_table_insert"
  f(::Signed) @ Main REPL[8]:1                                 # method responsible for the second section
  "jl_method_table_insert"

vs master:

julia> logmeths
10-element Vector{Any}:
  MethodInstance for callsfrts(::Int64)
 1
  MethodInstance for f(::Signed)
  "jl_method_table_insert"
  f(::Int64) @ Main REPL[7]:1                                  # method responsible for the first section
  "jl_method_table_insert"
  MethodInstance for invokesfs(::Int64)                        # a second section starts, but no causal method is logged
 1
  MethodInstance for f(::Int64)
  "jl_method_table_insert"

And here's what this demo looks like on 1.10 with SnoopCompile:

julia> using SnoopCompileCore

julia> invs = @snoop_invalidations begin
       f(::Integer) = 1
       callsfrts(x) = f(Base.inferencebarrier(x)::Signed)
       invokesfs(x) = invoke(f, Tuple{Signed}, x)

       # compilation
       invokesfs(1)                        # invoked callee
       callsfrts(1)

       f(::Int) = 2
       f(::Signed) = 4
       end
14-element Vector{Any}:
  MethodInstance for callsfrts(::Int64)
 1
  MethodInstance for f(::Signed)
  "jl_method_table_insert"
  f(::Int64) @ Main REPL[2]:10
  "jl_method_table_insert"
  MethodInstance for invokesfs(::Int64)
 1
  MethodInstance for f(::Int64)
  "jl_method_table_insert"
  MethodInstance for f(::Signed)
  "jl_method_table_insert"
  f(::Signed) @ Main REPL[2]:11
  "jl_method_table_insert"

julia> using SnoopCompile

julia> trees = invalidation_trees(invs)
2-element Vector{SnoopCompile.MethodInvalidations}:
 inserting f(::Int64) @ Main REPL[2]:10 invalidated:
   backedges: 1: superseding f(::Integer) @ Main REPL[2]:2 with MethodInstance for f(::Signed) (1 children)

 inserting f(::Signed) @ Main REPL[2]:11 invalidated:
   backedges: 1: superseding f(::Integer) @ Main REPL[2]:2 with MethodInstance for f(::Int64) (1 children)


julia> using SnoopCompile.AbstractTrees

julia> print_tree(only(trees[1].backedges))
MethodInstance for f(::Signed) at depth 0 with 1 children
└─ MethodInstance for callsfrts(::Int64) at depth 1 with 0 children

julia> print_tree(only(trees[2].backedges))
MethodInstance for f(::Int64) at depth 0 with 1 children
└─ MethodInstance for invokesfs(::Int64) at depth 1 with 0 children