make Tuple{Union{}} unconstructable

Type intersection assumed it was equal to Union{}, so this makes it
unconstructable so that holds true. This is similar to what the
NamedTuple constructor does.

Secondarily, this fixes an inference bug where it would create
Vararg{Union{}} and then incorrectly handle that fieldtype.

- Fixes #32392
- Addresses part of the concerns discussed in
  #24614 (comment)
- Addresses part of the issues presented in
  #26175
- May allow improving jl_type_equality_is_identity
  (https://github.com/JuliaLang/julia/pull/49017/files#diff-882927c6e612596e22406ae0d06adcee88a9ec05e8b61ad81b48942e2cb266e9R986)
- May allow improving intersection (finish_unionall can be more
  aggressive at computing varval for any typevars that appears in
  covariant position and has lb=Union{} and ub=leaf type)

Author: vtjnash

base/compiler/abstractinterpretation.jl

@@ -546,6 +546,8 @@ function abstract_call_method(interp::AbstractInterpreter, method::Method, @nosp
         add_remark!(interp, sv, "Refusing to infer into `depwarn`")
         return MethodCallResult(Any, false, false, nothing, Effects())
     end
+    sigtuple = unwrap_unionall(sig)
+    sigtuple isa DataType || return MethodCallResult(Any, false, false, nothing, Effects())
 
     # Limit argument type tuple growth of functions:
     # look through the parents list to see if there's a call to the same method
@@ -581,7 +583,6 @@ function abstract_call_method(interp::AbstractInterpreter, method::Method, @nosp
     washardlimit = hardlimit
 
     if topmost !== nothing
-        sigtuple = unwrap_unionall(sig)::DataType
         msig = unwrap_unionall(method.sig)::DataType
         spec_len = length(msig.parameters) + 1
         ls = length(sigtuple.parameters)
@@ -1376,7 +1377,11 @@ function precise_container_type(interp::AbstractInterpreter, @nospecialize(itft)
             va = isvarargtype(last)
             elts = Any[ fieldtype(tti0, i) for i = 1:len ]
             if va
-                elts[len] = Vararg{elts[len]}
+                if elts[len] === Union{}
+                    pop!(elts)
+                else
+                    elts[len] = Vararg{elts[len]}
+                end
             end
             return AbstractIterationResult(elts, nothing)
         end
@@ -1385,6 +1390,9 @@ function precise_container_type(interp::AbstractInterpreter, @nospecialize(itft)
     elseif tti0 === Any
         return AbstractIterationResult(Any[Vararg{Any}], nothing, Effects())
     elseif tti0 <: Array
+        if eltype(tti0) === Union{}
+            return AbstractIterationResult(Any[], nothing)
+        end
         return AbstractIterationResult(Any[Vararg{eltype(tti0)}], nothing)
     else
         return abstract_iteration(interp, itft, typ, sv)
@@ -2098,7 +2106,7 @@ end
 
 function sp_type_rewrap(@nospecialize(T), linfo::MethodInstance, isreturn::Bool)
     isref = false
-    if T === Bottom
+    if unwrapva(T) === Bottom
         return Bottom
     elseif isa(T, Type)
         if isa(T, DataType) && (T::DataType).name === _REF_NAME
@@ -2135,8 +2143,13 @@ end
 function abstract_eval_cfunction(interp::AbstractInterpreter, e::Expr, vtypes::VarTable, sv::InferenceState)
     f = abstract_eval_value(interp, e.args[2], vtypes, sv)
     # rt = sp_type_rewrap(e.args[3], sv.linfo, true)
-    at = Any[ sp_type_rewrap(argt, sv.linfo, false) for argt in e.args[4]::SimpleVector ]
-    pushfirst!(at, f)
+    atv = e.args[4]::SimpleVector
+    at = Vector{Any}(undef, length(atv) + 1)
+    at[1] = f
+    for i = 1:length(atv)
+        at[i + 1] = sp_type_rewrap(at[i], sv.linfo, false)
+        at[i + 1] === Bottom && return
+    end
     # this may be the wrong world for the call,
     # but some of the result is likely to be valid anyways
     # and that may help generate better codegen
@@ -2339,7 +2352,7 @@ function abstract_eval_statement_expr(interp::AbstractInterpreter, e::Expr, vtyp
                 let t = t, at = at; all(i::Int->getfield(at.val::Tuple, i) isa fieldtype(t, i), 1:n); end
                 nothrow = isexact
                 t = Const(ccall(:jl_new_structt, Any, (Any, Any), t, at.val))
-            elseif isa(at, PartialStruct) && at ⊑ᵢ Tuple && n == length(at.fields::Vector{Any}) &&
+            elseif isa(at, PartialStruct) && at ⊑ᵢ Tuple && n > 0 && n == length(at.fields::Vector{Any}) && !isvarargtype(at.fields[end]) &&
                 let t = t, at = at; all(i::Int->(at.fields::Vector{Any})[i] ⊑ᵢ fieldtype(t, i), 1:n); end
                 nothrow = isexact
                 t = PartialStruct(t, at.fields::Vector{Any})

base/compiler/inferenceresult.jl

@@ -119,9 +119,9 @@ function most_general_argtypes(method::Union{Method, Nothing}, @nospecialize(spe
     # to the appropriate `Tuple` type or `PartialStruct` instance.
     if !toplevel && isva
         if specTypes::Type == Tuple
+            linfo_argtypes = Any[Any for i = 1:nargs]
             if nargs > 1
-                linfo_argtypes = Any[Any for i = 1:nargs]
-                linfo_argtypes[end] = Vararg{Any}
+                linfo_argtypes[end] = Tuple
             end
             vargtype = Tuple
         else

base/compiler/ssair/inlining.jl

@@ -1170,7 +1170,7 @@ function handle_invoke_call!(todo::Vector{Pair{Int,Any}},
         if isa(result, ConstPropResult)
             mi = result.result.linfo
             validate_sparams(mi.sparam_vals) || return nothing
-            if argtypes_to_type(argtypes) <: mi.def.sig
+            if Union{} !== argtypes_to_type(argtypes) <: mi.def.sig
                 item = resolve_todo(mi, result.result, argtypes, info, flag, state; invokesig)
                 handle_single_case!(todo, ir, idx, stmt, item, OptimizationParams(state.interp), true)
                 return nothing

base/compiler/tfuncs.jl

@@ -1872,7 +1872,15 @@ add_tfunc(apply_type, 1, INT_INF, apply_type_tfunc, 10)
 # convert the dispatch tuple type argtype to the real (concrete) type of
 # the tuple of those values
 function tuple_tfunc(𝕃::AbstractLattice, argtypes::Vector{Any})
+    isempty(argtypes) && return Const(())
     argtypes = anymap(widenslotwrapper, argtypes)
+    if isvarargtype(argtypes[end]) && unwrapva(argtypes[end]) === Union{}
+        # Drop the Vararg in Tuple{...,Vararg{Union{}}} since it must be length 0.
+        # If there is a Vararg num also, it must be a TypeVar, and it must be
+        # zero, but that generally shouldn't show up here, since it implies a
+        # UnionAll context is missing around this.
+        pop!(argtypes)
+    end
     all_are_const = true
     for i in 1:length(argtypes)
         if !isa(argtypes[i], Const)
@@ -1915,6 +1923,8 @@ function tuple_tfunc(𝕃::AbstractLattice, argtypes::Vector{Any})
                 params[i] = x
             elseif !isvarargtype(x) && hasintersect(x, Type)
                 params[i] = Union{x, Type}
+            elseif x === Union{}
+                return Bottom # argtypes is malformed, but try not to crash
             else
                 params[i] = x
             end

base/compiler/typeutils.jl

@@ -187,6 +187,7 @@ function typesubtract(@nospecialize(a), @nospecialize(b), max_union_splitting::I
                             bp = b.parameters[i]
                             (isvarargtype(ap) || isvarargtype(bp)) && return a
                             ta[i] = typesubtract(ap, bp, min(2, max_union_splitting))
+                            ta[i] === Union{} && return Union{}
                             return Tuple{ta...}
                         end
                     end

base/iterators.jl

@@ -41,6 +41,12 @@ if Base !== Core.Compiler
 export partition
 end
 
+function TupleOrBottom(tt...)
+    any(p -> p === Union{}, tt) && return Union{}
+    return Tuple{tt...}
+end
+
+
 """
     Iterators.map(f, iterators...)
 
@@ -209,7 +215,7 @@ size(e::Enumerate) = size(e.itr)
 end
 last(e::Enumerate) = (length(e.itr), e.itr[end])
 
-eltype(::Type{Enumerate{I}}) where {I} = Tuple{Int, eltype(I)}
+eltype(::Type{Enumerate{I}}) where {I} = TupleOrBottom(Int, eltype(I))
 
 IteratorSize(::Type{Enumerate{I}}) where {I} = IteratorSize(I)
 IteratorEltype(::Type{Enumerate{I}}) where {I} = IteratorEltype(I)
@@ -394,7 +400,7 @@ _promote_tuple_shape((m,)::Tuple{Integer}, (n,)::Tuple{Integer}) = (min(m, n),)
 _promote_tuple_shape(a, b) = promote_shape(a, b)
 _promote_tuple_shape(a, b...) = _promote_tuple_shape(a, _promote_tuple_shape(b...))
 _promote_tuple_shape(a) = a
-eltype(::Type{Zip{Is}}) where {Is<:Tuple} = Tuple{map(eltype, fieldtypes(Is))...}
+eltype(::Type{Zip{Is}}) where {Is<:Tuple} = TupleOrBottom(map(eltype, fieldtypes(Is))...)
 #eltype(::Type{Zip{Tuple{}}}) = Tuple{}
 #eltype(::Type{Zip{Tuple{A}}}) where {A} = Tuple{eltype(A)}
 #eltype(::Type{Zip{Tuple{A, B}}}) where {A, B} = Tuple{eltype(A), eltype(B)}
@@ -1072,8 +1078,7 @@ end
 
 eltype(::Type{ProductIterator{I}}) where {I} = _prod_eltype(I)
 _prod_eltype(::Type{Tuple{}}) = Tuple{}
-_prod_eltype(::Type{I}) where {I<:Tuple} =
-    Tuple{ntuple(n -> eltype(fieldtype(I, n)), _counttuple(I)::Int)...}
+_prod_eltype(::Type{I}) where {I<:Tuple} = TupleOrBottom(ntuple(n -> eltype(fieldtype(I, n)), _counttuple(I)::Int)...)
 
 iterate(::ProductIterator{Tuple{}}) = (), true
 iterate(::ProductIterator{Tuple{}}, state) = nothing
@@ -1442,6 +1447,7 @@ end
 function _approx_iter_type(itrT::Type, vstate::Type)
     vstate <: Union{Nothing, Tuple{Any, Any}} || return Any
     vstate <: Union{} && return Union{}
+    itrT <: Union{} && return Union{}
     nextvstate = Base._return_type(doiterate, Tuple{itrT, vstate})
     return (nextvstate <: vstate ? vstate : Any)
 end

src/builtins.c

@@ -1318,7 +1318,7 @@ JL_CALLABLE(jl_f_apply_type)
                 jl_type_error_rt("Tuple", "parameter", (jl_value_t*)jl_type_type, pi);
             }
         }
-        return (jl_value_t*)jl_apply_tuple_type_v(&args[1], nargs-1);
+        return jl_apply_tuple_type_v(&args[1], nargs-1);
     }
     else if (args[0] == (jl_value_t*)jl_uniontype_type) {
         // Union{} has extra restrictions, so it needs to be checked after

src/codegen.cpp

@@ -5146,7 +5146,7 @@ static std::pair<Function*, Function*> get_oc_function(jl_codectx_t &ctx, jl_met
     for (size_t i = 0; i < jl_svec_len(argt_typ->parameters); ++i) {
         jl_svecset(sig_args, 1+i, jl_svecref(argt_typ->parameters, i));
     }
-    sigtype = (jl_value_t*)jl_apply_tuple_type_v(jl_svec_data(sig_args), nsig);
+    sigtype = jl_apply_tuple_type_v(jl_svec_data(sig_args), nsig);
 
     jl_method_instance_t *mi = jl_specializations_get_linfo(closure_method, sigtype, jl_emptysvec);
     jl_code_instance_t *ci = (jl_code_instance_t*)jl_rettype_inferred(mi, ctx.world, ctx.world);
@@ -5469,7 +5469,7 @@ static jl_cgval_t emit_expr(jl_codectx_t &ctx, jl_value_t *expr, ssize_t ssaidx_
 
         if (can_optimize) {
             jl_value_t *closure_t = NULL;
-            jl_tupletype_t *env_t = NULL;
+            jl_value_t *env_t = NULL;
             JL_GC_PUSH2(&closure_t, &env_t);
 
             jl_value_t **env_component_ts = (jl_value_t**)alloca(sizeof(jl_value_t*) * (nargs-4));
@@ -5479,10 +5479,10 @@ static jl_cgval_t emit_expr(jl_codectx_t &ctx, jl_value_t *expr, ssize_t ssaidx_
 
             env_t = jl_apply_tuple_type_v(env_component_ts, nargs-4);
             // we need to know the full env type to look up the right specialization
-            if (jl_is_concrete_type((jl_value_t*)env_t)) {
+            if (jl_is_concrete_type(env_t)) {
                 jl_tupletype_t *argt_typ = (jl_tupletype_t*)argt.constant;
                 Function *F, *specF;
-                std::tie(F, specF) = get_oc_function(ctx, (jl_method_t*)source.constant, env_t, argt_typ, ub.constant);
+                std::tie(F, specF) = get_oc_function(ctx, (jl_method_t*)source.constant, (jl_datatype_t*)env_t, argt_typ, ub.constant);
                 if (F) {
                     jl_cgval_t jlcall_ptr = mark_julia_type(ctx, F, false, jl_voidpointer_type);
                     jl_aliasinfo_t ai = jl_aliasinfo_t::fromTBAA(ctx, ctx.tbaa().tbaa_gcframe);
@@ -5495,7 +5495,7 @@ static jl_cgval_t emit_expr(jl_codectx_t &ctx, jl_value_t *expr, ssize_t ssaidx_
                         fptr = mark_julia_type(ctx, (llvm::Value*)Constant::getNullValue(ctx.types().T_size), false, jl_voidpointer_type);
 
                     // TODO: Inline the env at the end of the opaque closure and generate a descriptor for GC
-                    jl_cgval_t env = emit_new_struct(ctx, (jl_value_t*)env_t, nargs-4, &argv.data()[4]);
+                    jl_cgval_t env = emit_new_struct(ctx, env_t, nargs-4, &argv.data()[4]);
 
                     jl_cgval_t closure_fields[5] = {
                         env,
@@ -6441,7 +6441,7 @@ static jl_cgval_t emit_cfunction(jl_codectx_t &ctx, jl_value_t *output_type, con
         sigt = NULL;
     }
     else {
-        sigt = (jl_value_t*)jl_apply_tuple_type((jl_svec_t*)sigt);
+        sigt = jl_apply_tuple_type((jl_svec_t*)sigt);
     }
     if (sigt && !(unionall_env && jl_has_typevar_from_unionall(rt, unionall_env))) {
         unionall_env = NULL;
@@ -6891,9 +6891,9 @@ static jl_datatype_t *compute_va_type(jl_method_instance_t *lam, size_t nreq)
         }
         jl_svecset(tupargs, i-nreq, argType);
     }
-    jl_datatype_t *typ = jl_apply_tuple_type(tupargs);
+    jl_value_t *typ = jl_apply_tuple_type(tupargs);
     JL_GC_POP();
-    return typ;
+    return (jl_datatype_t*)typ;
 }
 
 

src/gf.c

@@ -1227,7 +1227,7 @@ static jl_method_instance_t *cache_method(
     intptr_t nspec = (kwmt == NULL || kwmt == jl_type_type_mt || kwmt == jl_nonfunction_mt || kwmt == jl_kwcall_mt ? definition->nargs + 1 : jl_atomic_load_relaxed(&kwmt->max_args) + 2 + 2 * (mt == jl_kwcall_mt));
     jl_compilation_sig(tt, sparams, definition, nspec, &newparams);
     if (newparams) {
-        temp2 = (jl_value_t*)jl_apply_tuple_type(newparams);
+        temp2 = jl_apply_tuple_type(newparams);
         // Now there may be a problem: the widened signature is more general
         // than just the given arguments, so it might conflict with another
         // definition that does not have cache instances yet. To fix this, we
@@ -1350,7 +1350,7 @@ static jl_method_instance_t *cache_method(
         }
     }
     if (newparams) {
-        simplett = jl_apply_tuple_type(newparams);
+        simplett = (jl_datatype_t*)jl_apply_tuple_type(newparams);
         temp2 = (jl_value_t*)simplett;
     }
 
@@ -2513,7 +2513,7 @@ JL_DLLEXPORT jl_value_t *jl_normalize_to_compilable_sig(jl_methtable_t *mt, jl_t
     jl_compilation_sig(ti, env, m, nspec, &newparams);
     int is_compileable = ((jl_datatype_t*)ti)->isdispatchtuple;
     if (newparams) {
-        tt = jl_apply_tuple_type(newparams);
+        tt = (jl_datatype_t*)jl_apply_tuple_type(newparams);
         if (!is_compileable) {
             // compute new env, if used below
             jl_value_t *ti = jl_type_intersection_env((jl_value_t*)tt, (jl_value_t*)m->sig, &newparams);

src/intrinsics.cpp

@@ -1409,7 +1409,7 @@ static Value *emit_untyped_intrinsic(jl_codectx_t &ctx, intrinsic f, Value **arg
         jl_value_t *params[2];
         params[0] = xtyp;
         params[1] = (jl_value_t*)jl_bool_type;
-        jl_datatype_t *tuptyp = jl_apply_tuple_type_v(params, 2);
+        jl_datatype_t *tuptyp = (jl_datatype_t*)jl_apply_tuple_type_v(params, 2);
         *newtyp = tuptyp;
 
         Value *tupval;

src/jl_exported_funcs.inc

@@ -478,7 +478,6 @@
     XX(jl_try_substrtod) \
     XX(jl_try_substrtof) \
     XX(jl_tty_set_mode) \
-    XX(jl_tupletype_fill) \
     XX(jl_typeassert) \
     XX(jl_typeinf_lock_begin) \
     XX(jl_typeinf_lock_end) \