fix #9770 (type inference error)

[vtjnash]

here's the segfault-enhanced test case for this:

x9770 = false
function f9770(x)
    if x9770
        g9770(:a, :foo)
    else
        x
    end
end
function g9770(x,y)
   if isa(y,Symbol)
       f9770(x)
   else
       g9770(:a, :foo)
   end
end
@test g9770(:a, "c") === :a
@test g9770(:b, :c) === :b

@JeffBezanson i don't like adding all of the typeinf function to the "don't-infer" list, but it seemed to suffice to get through type inference

[vtjnash]

(also, with this change, we never cache g9770(::Symbol,::Symbol), which also seems incorrect)

edit: the corrected patch needs:

diff --git a/base/inference.jl b/base/inference.jl
index 3e4f091..c8c0e2d 100644
--- a/base/inference.jl
+++ b/base/inference.jl
@@ -1311,7 +1311,7 @@ function typeinf(linfo::LambdaStaticData,atypes::Tuple,sparams::Tuple, def, cop)
             # return best guess so far
             f = f::CallStack
             f.recurred = true
-            f.toprec = !f.rec
+            f.toprec = f.toprec || !f.rec
             f.rec = true
             r = inference_stack
             while !is(r, f)

but I still think that this algorithm may not be correct for handling recurred functions and may record the wrong return value. it's hard to create a failing case however.

[vtjnash]

here's my newest attempt at breaking it:

julia> begin
                              @noinline _true() = true
                              @noinline _false() = false
                              function f(x)
                                 if _false()
                                   g(:a, :foo)
                                 else
                                   x
                                 end
                              end
                              function g(x,y)
                                if isa(y,Symbol)
                                  f(x)
                                elseif _true()
                                  h(g(x,y))
                                end
                              end
                              h(x::Symbol) = Int
                              h(x::ASCIIString) = Float32
                              h(x::Any) = Function

       end
h (generic function with 3 methods)

julia> @code_typed g(:a, "c")
1-element Array{Any,1}:
 :($(Expr(:lambda, Any[:x,:y], Any[Any[],Any[Any[:x,Symbol,0],Any[:y,ASCIIString,0]],Any[]], :(begin  # none, line 12:
        unless $(Expr(:call1, :isa, :(y::ASCIIString), :Symbol)) goto 0 # line 13:
        return $(Expr(:call1, :f, :(x::Symbol)))
        goto 2
        0:  # line 14:
        unless $(Expr(:call1, :_true)) goto 1 # line 15:
        return h($(Expr(:call1, :g, :(x::Symbol), :(y::ASCIIString))))::Union(Type{Int64},Type{Float32},Type{Function})
        goto 2
        1: 
        return
        2: 
    end::Any))))

julia> @code_typed g(:a, :c)
1-element Array{Any,1}:
 :($(Expr(:lambda, Any[:x,:y], Any[Any[],Any[Any[:x,Symbol,0],Any[:y,Symbol,0]],Any[]], :(begin  # none, line 12:
        unless isa(y::Symbol,Symbol)::Bool goto 0 # line 13:
        return f(x::Symbol)::Any
        goto 2
        0:  # line 14:
        unless _true()::Bool goto 1 # line 15:
        return h(g(x::Symbol,y::Symbol)::Any)::Union(Type{Int64},Type{Float32},Type{Function})
        goto 2
        1: 
        return
        2: 
    end::Any))))

julia> @code_typed g(:a, "c")
1-element Array{Any,1}:
 :($(Expr(:lambda, Any[:x,:y], Any[Any[],Any[Any[:x,Symbol,0],Any[:y,ASCIIString,0]],Any[]], :(begin  # none, line 12:
        unless isa(y::ASCIIString,Symbol)::Bool goto 0 # line 13:
        return f(x::Symbol)::Any
        goto 2
        0:  # line 14:
        unless _true()::Bool goto 1 # line 15:
        return h(g(x::Symbol,y::ASCIIString)::Any)::Union(Type{Int64},Type{Float32},Type{Function})
        goto 2
        1: 
        return
        2: 
    end::Any))))

julia> @code_typed f(:a)
1-element Array{Any,1}:
 :($(Expr(:lambda, Any[:x], Any[Any[],Any[Any[:x,Symbol,0]],Any[]], :(begin  # none, line 5:
        unless _false()::Bool goto 0 # line 6:
        return g(:a,:foo)::Any
        goto 1
        0:  # line 8:
        return x::Symbol
        1: 
    end::Any))))

[tkelman]

Is this causing appveyor to time out with a higher-than-usual probability? OSX Travis too?

[vtjnash]

yes, even with the above patch now included, it significantly increases build times

[JeffBezanson]

Rebased. But yes, for me this took the sysimg build time (starting with inference0) from 3 minutes to 4 minutes.

[vtjnash]

that's actually not a bad as I feared. initially i think this was at least doubling the build time (appveyor went from 20minutes to > 40 minute). is it worthwhile to switch inference to a queue instead of recursion to do the much faster fix for this while we wait for the green-fairy (which i've been told already handles this case)?

[vtjnash]

also fixes #9222

[tkelman]

test? and more descriptive commit message?