generators: expose caller world to GeneratedFunctionStub

Expose the demanded world to the GeneratedFunctionStub caller, for users
such as Cassette. If this argument is used, the uesr must return a
CodeInfo with the min/max world field set correctly.

Make the internal representation a tiny bit more compact also, removing
a little bit of unnecessary metadata.

Remove support for returning `body isa CodeInfo` via this wrapper, since
it is impossible to return a correct object via the
GeneratedFunctionStub since it strips off the world argument, which is
required for it to do so. This also removes support for not inferring
these fully (expand_early=false).

Also answer method lookup queries about the future correctly, by
refusing to answer them. This helps keeps execution correct as methods
get added to the system asynchronously.

This reverts "fix #25678: return matters for generated functions
(#40778)" (commit 92c84bf), since this
is no longer sensible to return here anyways, so it is no longer
permitted or supported by this macro.

Fixes various issues where we failed to specify the correct world.

Author: vtjnash

base/Base.jl

@@ -468,7 +468,7 @@ in_sysimage(pkgid::PkgId) = pkgid in _sysimage_modules
 for match = _methods(+, (Int, Int), -1, get_world_counter())
     m = match.method
     delete!(push!(Set{Method}(), m), m)
-    copy(Core.Compiler.retrieve_code_info(Core.Compiler.specialize_method(match)))
+    copy(Core.Compiler.retrieve_code_info(Core.Compiler.specialize_method(match), typemax(UInt)))
 
     empty!(Set())
     push!(push!(Set{Union{GlobalRef,Symbol}}(), :two), GlobalRef(Base, :two))

base/boot.jl

@@ -590,28 +590,25 @@ println(@nospecialize a...) = println(stdout, a...)
 
 struct GeneratedFunctionStub
     gen
-    argnames::Array{Any,1}
-    spnames::Union{Nothing, Array{Any,1}}
-    line::Int
-    file::Symbol
-    expand_early::Bool
+    argnames::SimpleVector
+    spnames::SimpleVector
 end
 
-# invoke and wrap the results of @generated
-function (g::GeneratedFunctionStub)(@nospecialize args...)
+# invoke and wrap the results of @generated expression
+function (g::GeneratedFunctionStub)(world::UInt, source::LineNumberNode, @nospecialize args...)
+    # args is (spvals..., argtypes...)
     body = g.gen(args...)
-    if body isa CodeInfo
-        return body
-    end
-    lam = Expr(:lambda, g.argnames,
-               Expr(Symbol("scope-block"),
+    file = source.file
+    file isa Symbol || (file = :none)
+    lam = Expr(:lambda, Expr(:argnames, g.argnames...).args,
+               Expr(:var"scope-block",
                     Expr(:block,
-                         LineNumberNode(g.line, g.file),
-                         Expr(:meta, :push_loc, g.file, Symbol("@generated body")),
+                         source,
+                         Expr(:meta, :push_loc, file, :var"@generated body"),
                          Expr(:return, body),
                          Expr(:meta, :pop_loc))))
     spnames = g.spnames
-    if spnames === nothing
+    if spnames === svec()
         return lam
     else
         return Expr(Symbol("with-static-parameters"), lam, spnames...)

base/compiler/abstractinterpretation.jl

@@ -573,7 +573,7 @@ function abstract_call_method(interp::AbstractInterpreter, method::Method, @nosp
                 break
             end
             topmost === nothing || continue
-            if edge_matches_sv(infstate, method, sig, sparams, hardlimit, sv)
+            if edge_matches_sv(interp, infstate, method, sig, sparams, hardlimit, sv)
                 topmost = infstate
                 edgecycle = true
             end
@@ -681,12 +681,13 @@ function abstract_call_method(interp::AbstractInterpreter, method::Method, @nosp
     return MethodCallResult(rt, edgecycle, edgelimited, edge, effects)
 end
 
-function edge_matches_sv(frame::InferenceState, method::Method, @nospecialize(sig), sparams::SimpleVector, hardlimit::Bool, sv::InferenceState)
+function edge_matches_sv(interp::AbstractInterpreter, frame::InferenceState, method::Method, @nospecialize(sig), sparams::SimpleVector, hardlimit::Bool, sv::InferenceState)
     # The `method_for_inference_heuristics` will expand the given method's generator if
     # necessary in order to retrieve this field from the generated `CodeInfo`, if it exists.
     # The other `CodeInfo`s we inspect will already have this field inflated, so we just
     # access it directly instead (to avoid regeneration).
-    callee_method2 = method_for_inference_heuristics(method, sig, sparams) # Union{Method, Nothing}
+    world = get_world_counter(interp)
+    callee_method2 = method_for_inference_heuristics(method, sig, sparams, world) # Union{Method, Nothing}
 
     inf_method2 = frame.src.method_for_inference_limit_heuristics # limit only if user token match
     inf_method2 isa Method || (inf_method2 = nothing)
@@ -723,11 +724,11 @@ function edge_matches_sv(frame::InferenceState, method::Method, @nospecialize(si
 end
 
 # This function is used for computing alternate limit heuristics
-function method_for_inference_heuristics(method::Method, @nospecialize(sig), sparams::SimpleVector)
-    if isdefined(method, :generator) && method.generator.expand_early && may_invoke_generator(method, sig, sparams)
+function method_for_inference_heuristics(method::Method, @nospecialize(sig), sparams::SimpleVector, world::UInt)
+    if isdefined(method, :generator) && !(method.generator isa Core.GeneratedFunctionStub) && may_invoke_generator(method, sig, sparams)
         method_instance = specialize_method(method, sig, sparams)
         if isa(method_instance, MethodInstance)
-            cinfo = get_staged(method_instance)
+            cinfo = get_staged(method_instance, world)
             if isa(cinfo, CodeInfo)
                 method2 = cinfo.method_for_inference_limit_heuristics
                 if method2 isa Method
@@ -791,24 +792,25 @@ end
 function pure_eval_eligible(interp::AbstractInterpreter,
     @nospecialize(f), applicable::Vector{Any}, arginfo::ArgInfo)
     # XXX we need to check that this pure function doesn't call any overlayed method
+    world = get_world_counter(interp)
     return f !== nothing &&
            length(applicable) == 1 &&
-           is_method_pure(applicable[1]::MethodMatch) &&
+           is_method_pure(applicable[1]::MethodMatch, world) &&
            is_all_const_arg(arginfo, #=start=#2)
 end
 
-function is_method_pure(method::Method, @nospecialize(sig), sparams::SimpleVector)
+function is_method_pure(method::Method, @nospecialize(sig), sparams::SimpleVector, world::UInt)
     if isdefined(method, :generator)
-        method.generator.expand_early || return false
+        method.generator isa Core.GeneratedFunctionStub && return false
         mi = specialize_method(method, sig, sparams)
         isa(mi, MethodInstance) || return false
-        staged = get_staged(mi)
+        staged = get_staged(mi, world)
         (staged isa CodeInfo && (staged::CodeInfo).pure) || return false
         return true
     end
     return method.pure
 end
-is_method_pure(match::MethodMatch) = is_method_pure(match.method, match.spec_types, match.sparams)
+is_method_pure(match::MethodMatch, world::UInt) = is_method_pure(match.method, match.spec_types, match.sparams, world)
 
 function pure_eval_call(interp::AbstractInterpreter,
     @nospecialize(f), applicable::Vector{Any}, arginfo::ArgInfo)

base/compiler/bootstrap.jl

@@ -36,7 +36,7 @@ let interp = NativeInterpreter()
         else
             tt = Tuple{typeof(f), Vararg{Any}}
         end
-        for m in _methods_by_ftype(tt, 10, typemax(UInt))::Vector
+        for m in _methods_by_ftype(tt, 10, get_world_counter())::Vector
             # remove any TypeVars from the intersection
             m = m::MethodMatch
             typ = Any[m.spec_types.parameters...]

base/compiler/inferencestate.jl

@@ -342,7 +342,8 @@ end
 
 function InferenceState(result::InferenceResult, cache::Symbol, interp::AbstractInterpreter)
     # prepare an InferenceState object for inferring lambda
-    src = retrieve_code_info(result.linfo)
+    world = get_world_counter(interp)
+    src = retrieve_code_info(result.linfo, world)
     src === nothing && return nothing
     validate_code_in_debug_mode(result.linfo, src, "lowered")
     return InferenceState(result, src, cache, interp)

base/compiler/optimize.jl

@@ -182,7 +182,8 @@ function OptimizationState(linfo::MethodInstance, src::CodeInfo, params::Optimiz
     return OptimizationState(linfo, src, nothing, stmt_info, mod, sptypes, slottypes, inlining, nothing)
 end
 function OptimizationState(linfo::MethodInstance, params::OptimizationParams, interp::AbstractInterpreter)
-    src = retrieve_code_info(linfo)
+    world = get_world_counter(interp)
+    src = retrieve_code_info(linfo, world)
     src === nothing && return nothing
     return OptimizationState(linfo, src, params, interp)
 end

base/compiler/typeinfer.jl

@@ -1041,7 +1041,7 @@ function typeinf_ext(interp::AbstractInterpreter, mi::MethodInstance)
         end
     end
     if ccall(:jl_get_module_infer, Cint, (Any,), method.module) == 0 && !generating_sysimg()
-        return retrieve_code_info(mi)
+        return retrieve_code_info(mi, get_world_counter(interp))
     end
     lock_mi_inference(interp, mi)
     result = InferenceResult(mi)

base/compiler/types.jl

@@ -317,7 +317,7 @@ struct NativeInterpreter <: AbstractInterpreter
         cache = Vector{InferenceResult}() # Initially empty cache
 
         # Sometimes the caller is lazy and passes typemax(UInt).
-        # we cap it to the current world age
+        # we cap it to the current world age for correctness
         if world == typemax(UInt)
             world = get_world_counter()
         end

base/compiler/utilities.jl

@@ -114,23 +114,23 @@ end
 invoke_api(li::CodeInstance) = ccall(:jl_invoke_api, Cint, (Any,), li)
 use_const_api(li::CodeInstance) = invoke_api(li) == 2
 
-function get_staged(mi::MethodInstance)
+function get_staged(mi::MethodInstance, world::UInt)
     may_invoke_generator(mi) || return nothing
     try
         # user code might throw errors – ignore them
-        ci = ccall(:jl_code_for_staged, Any, (Any,), mi)::CodeInfo
+        ci = ccall(:jl_code_for_staged, Any, (Any, UInt), mi, world)::CodeInfo
         return ci
     catch
         return nothing
     end
 end
 
-function retrieve_code_info(linfo::MethodInstance)
+function retrieve_code_info(linfo::MethodInstance, world::UInt)
     m = linfo.def::Method
     c = nothing
     if isdefined(m, :generator)
         # user code might throw errors – ignore them
-        c = get_staged(linfo)
+        c = get_staged(linfo, world)
     end
     if c === nothing && isdefined(m, :source)
         src = m.source

base/compiler/validation.jl

@@ -200,15 +200,14 @@ end
 
 """
     validate_code!(errors::Vector{>:InvalidCodeError}, mi::MethodInstance,
-                   c::Union{Nothing,CodeInfo} = Core.Compiler.retrieve_code_info(mi))
+                   c::Union{Nothing,CodeInfo})
 
 Validate `mi`, logging any violation by pushing an `InvalidCodeError` into `errors`.
 
 If `isa(c, CodeInfo)`, also call `validate_code!(errors, c)`. It is assumed that `c` is
-the `CodeInfo` instance associated with `mi`.
+a `CodeInfo` instance associated with `mi`.
 """
-function validate_code!(errors::Vector{>:InvalidCodeError}, mi::Core.MethodInstance,
-                        c::Union{Nothing,CodeInfo} = Core.Compiler.retrieve_code_info(mi))
+function validate_code!(errors::Vector{>:InvalidCodeError}, mi::Core.MethodInstance, c::Union{Nothing,CodeInfo})
     is_top_level = mi.def isa Module
     if is_top_level
         mnargs = 0