remove Core._apply Builtin (#39115)

We can emulate this deprecated function, until we delete it in v2.
This in turn revealed some missing error checks, which we also add.

Fixes #39113

Author: vtjnash

base/boot.jl

@@ -803,4 +803,7 @@ Integer(x::Union{Float16, Float32, Float64}) = Int(x)
 # The internal jl_parse which will call into Core._parse if not `nothing`.
 _parse = nothing
 
+# support for deprecated uses of internal _apply function
+_apply(x...) = Core._apply_iterate(Main.Base.iterate, x...)
+
 ccall(:jl_set_istopmod, Cvoid, (Any, Bool), Core, true)

base/compiler/abstractinterpretation.jl

@@ -577,13 +577,7 @@ end
 
 # simulate iteration protocol on container type up to fixpoint
 function abstract_iteration(interp::AbstractInterpreter, @nospecialize(itft), @nospecialize(itertype), sv::InferenceState)
-    if !isdefined(Main, :Base) || !isdefined(Main.Base, :iterate) || !isconst(Main.Base, :iterate)
-        return Any[Vararg{Any}], nothing
-    end
-    if itft === nothing
-        iteratef = getfield(Main.Base, :iterate)
-        itft = Const(iteratef)
-    elseif isa(itft, Const)
+    if isa(itft, Const)
         iteratef = itft.val
     else
         return Any[Vararg{Any}], nothing
@@ -595,6 +589,7 @@ function abstract_iteration(interp::AbstractInterpreter, @nospecialize(itft), @n
     # Return Bottom if this is not an iterator.
     # WARNING: Changes to the iteration protocol must be reflected here,
     # this is not just an optimization.
+    # TODO: this doesn't realize that Array, SimpleVector, Tuple, and NamedTuple do not use the iterate protocol
     stateordonet === Bottom && return Any[Bottom], AbstractIterationInfo(CallMeta[CallMeta(Bottom, info)])
     valtype = statetype = Bottom
     ret = Any[]
@@ -658,7 +653,7 @@ function abstract_apply(interp::AbstractInterpreter, @nospecialize(itft), @nospe
     aftw = widenconst(aft)
     if !isa(aft, Const) && (!isType(aftw) || has_free_typevars(aftw))
         if !isconcretetype(aftw) || (aftw <: Builtin)
-            add_remark!(interp, sv, "Core._apply called on a function of a non-concrete type")
+            add_remark!(interp, sv, "Core._apply_iterate called on a function of a non-concrete type")
             # bail now, since it seems unlikely that abstract_call will be able to do any better after splitting
             # this also ensures we don't call abstract_call_gf_by_type below on an IntrinsicFunction or Builtin
             return CallMeta(Any, false)
@@ -805,7 +800,8 @@ function abstract_call_builtin(interp::AbstractInterpreter, f::Builtin, fargs::U
     end
     rt = builtin_tfunction(interp, f, argtypes[2:end], sv)
     if f === getfield && isa(fargs, Vector{Any}) && la == 3 && isa(argtypes[3], Const) && isa(argtypes[3].val, Int) && argtypes[2] ⊑ Tuple
-        cti, _ = precise_container_type(interp, nothing, argtypes[2], sv)
+        # TODO: why doesn't this use the getfield_tfunc?
+        cti, _ = precise_container_type(interp, iterate, argtypes[2], sv)
         idx = argtypes[3].val
         if 1 <= idx <= length(cti)
             rt = unwrapva(cti[idx])
@@ -923,11 +919,7 @@ function abstract_call_known(interp::AbstractInterpreter, @nospecialize(f),
     la = length(argtypes)
 
     if isa(f, Builtin)
-        if f === _apply
-            ft = argtype_by_index(argtypes, 2)
-            ft === Bottom && return CallMeta(Bottom, false)
-            return abstract_apply(interp, nothing, ft, argtype_tail(argtypes, 3), sv, max_methods)
-        elseif f === _apply_iterate
+        if f === _apply_iterate
             itft = argtype_by_index(argtypes, 2)
             ft = argtype_by_index(argtypes, 3)
             (itft === Bottom || ft === Bottom) && return CallMeta(Bottom, false)

base/compiler/compiler.jl

@@ -1,11 +1,11 @@
 # This file is a part of Julia. License is MIT: https://julialang.org/license
 
-getfield(getfield(Main, :Core), :eval)(getfield(Main, :Core), :(baremodule Compiler
+getfield(Core, :eval)(Core, :(baremodule Compiler
 
 using Core.Intrinsics, Core.IR
 
 import Core: print, println, show, write, unsafe_write, stdout, stderr,
-             _apply, _apply_iterate, svec, apply_type, Builtin, IntrinsicFunction,
+             _apply_iterate, svec, apply_type, Builtin, IntrinsicFunction,
              MethodInstance, CodeInstance, MethodMatch
 
 const getproperty = Core.getfield

base/compiler/optimize.jl

@@ -326,19 +326,19 @@ function statement_cost(ex::Expr, line::Int, src::CodeInfo, sptypes::Vector{Any}
             # The efficiency of operations like a[i] and s.b
             # depend strongly on whether the result can be
             # inferred, so check the type of ex
-            if f === Main.Core.getfield || f === Main.Core.tuple
+            if f === Core.getfield || f === Core.tuple
                 # we might like to penalize non-inferrability, but
                 # tuple iteration/destructuring makes that impossible
                 # return plus_saturate(argcost, isknowntype(extyp) ? 1 : params.inline_nonleaf_penalty)
                 return 0
-            elseif f === Main.Core.isa
+            elseif f === Core.isa
                 # If we're in a union context, we penalize type computations
                 # on union types. In such cases, it is usually better to perform
                 # union splitting on the outside.
                 if union_penalties && isa(argextype(ex.args[2],  src, sptypes, slottypes), Union)
                     return params.inline_nonleaf_penalty
                 end
-            elseif (f === Main.Core.arrayref || f === Main.Core.const_arrayref) && length(ex.args) >= 3
+            elseif (f === Core.arrayref || f === Core.const_arrayref) && length(ex.args) >= 3
                 atyp = argextype(ex.args[3], src, sptypes, slottypes)
                 return isknowntype(atyp) ? 4 : error_path ? params.inline_error_path_cost : params.inline_nonleaf_penalty
             end

base/compiler/ssair/inlining.jl

@@ -573,7 +573,7 @@ function batch_inline!(todo::Vector{Pair{Int, Any}}, ir::IRCode, linetable::Vect
     return ir
 end
 
-# This assumes the caller has verified that all arguments to the _apply call are Tuples.
+# This assumes the caller has verified that all arguments to the _apply_iterate call are Tuples.
 function rewrite_apply_exprargs!(ir::IRCode, todo::Vector{Pair{Int, Any}}, idx::Int,
         argexprs::Vector{Any}, atypes::Vector{Any}, arginfos::Vector{Any},
         arg_start::Int, et::Union{EdgeTracker, Nothing}, caches::Union{InferenceCaches, Nothing},
@@ -909,7 +909,7 @@ end
 function inline_apply!(ir::IRCode, todo::Vector{Pair{Int, Any}}, idx::Int, sig::Signature,
                        et, caches, params::OptimizationParams)
     stmt = ir.stmts[idx][:inst]
-    while sig.f === Core._apply || sig.f === Core._apply_iterate
+    while sig.f === Core._apply_iterate
         info = ir.stmts[idx][:info]
         if isa(info, UnionSplitApplyCallInfo)
             if length(info.infos) != 1
@@ -923,7 +923,7 @@ function inline_apply!(ir::IRCode, todo::Vector{Pair{Int, Any}}, idx::Int, sig::
             @assert info === nothing || info === false
             new_info = info = nothing
         end
-        arg_start = sig.f === Core._apply ? 2 : 3
+        arg_start = 3
         atypes = sig.atypes
         if arg_start > length(atypes)
             return nothing
@@ -1010,7 +1010,7 @@ function process_simple!(ir::IRCode, todo::Vector{Pair{Int, Any}}, idx::Int, sta
     sig = call_sig(ir, stmt)
     sig === nothing && return nothing
 
-    # Handle _apply
+    # Handle _apply_iterate
     sig = inline_apply!(ir, todo, idx, sig, state.et, state.caches, state.params)
     sig === nothing && return nothing
 

base/compiler/types.jl

@@ -102,14 +102,14 @@ struct InferenceParams
     # before computing the set of matching methods
     MAX_UNION_SPLITTING::Int
     # the maximum number of union-tuples to swap / expand
-    # when inferring a call to _apply
+    # when inferring a call to _apply_iterate
     MAX_APPLY_UNION_ENUM::Int
 
     # parameters limiting large (tuple) types
     TUPLE_COMPLEXITY_LIMIT_DEPTH::Int
 
-    # when attempting to inlining _apply, abort the optimization if the tuple
-    # contains more than this many elements
+    # when attempting to inline _apply_iterate, abort the optimization if the
+    # tuple contains more than this many elements
     MAX_TUPLE_SPLAT::Int
 
     function InferenceParams(;

src/builtin_proto.h

@@ -22,7 +22,7 @@ extern "C" {
 DECLARE_BUILTIN(throw);      DECLARE_BUILTIN(is);
 DECLARE_BUILTIN(typeof);     DECLARE_BUILTIN(sizeof);
 DECLARE_BUILTIN(issubtype);  DECLARE_BUILTIN(isa);
-DECLARE_BUILTIN(_apply);     DECLARE_BUILTIN(_apply_pure);
+DECLARE_BUILTIN(_apply_pure);
 DECLARE_BUILTIN(_call_latest); DECLARE_BUILTIN(_apply_iterate);
 DECLARE_BUILTIN(_call_in_world);
 DECLARE_BUILTIN(isdefined);  DECLARE_BUILTIN(nfields);

src/builtins.c

@@ -525,7 +525,7 @@ STATIC_INLINE void _grow_to(jl_value_t **root, jl_value_t ***oldargs, jl_svec_t
 
 static jl_function_t *jl_iterate_func JL_GLOBALLY_ROOTED;
 
-static jl_value_t *do_apply(jl_value_t *F, jl_value_t **args, uint32_t nargs, jl_value_t *iterate)
+static jl_value_t *do_apply( jl_value_t **args, uint32_t nargs, jl_value_t *iterate)
 {
     jl_function_t *f = args[0];
     if (nargs == 2) {
@@ -567,12 +567,7 @@ static jl_value_t *do_apply(jl_value_t *F, jl_value_t **args, uint32_t nargs, jl
         }
     }
     if (extra && iterate == NULL) {
-        if (jl_iterate_func == NULL) {
-            jl_iterate_func = jl_get_function(jl_top_module, "iterate");
-            if (jl_iterate_func == NULL)
-                jl_undefined_var_error(jl_symbol("iterate"));
-        }
-        iterate = jl_iterate_func;
+        jl_undefined_var_error(jl_symbol("iterate"));
     }
     // allocate space for the argument array and gc roots for it
     // based on our previous estimates
@@ -696,13 +691,7 @@ static jl_value_t *do_apply(jl_value_t *F, jl_value_t **args, uint32_t nargs, jl
 JL_CALLABLE(jl_f__apply_iterate)
 {
     JL_NARGSV(_apply_iterate, 2);
-    return do_apply(F, args+1, nargs-1, args[0]);
-}
-
-JL_CALLABLE(jl_f__apply)
-{
-    JL_NARGSV(_apply, 1);
-    return do_apply(F, args, nargs, NULL);
+    return do_apply(args + 1, nargs - 1, args[0]);
 }
 
 // this is like `_apply`, but with quasi-exact checks to make sure it is pure
@@ -720,7 +709,7 @@ JL_CALLABLE(jl_f__apply_pure)
         // and `promote` works better this way
         size_t last_age = ptls->world_age;
         ptls->world_age = jl_world_counter;
-        ret = jl_f__apply(NULL, args, nargs);
+        ret = do_apply(args, nargs, NULL);
         ptls->world_age = last_age;
         ptls->in_pure_callback = last_in;
     }
@@ -1578,7 +1567,6 @@ void jl_init_primitives(void) JL_GC_DISABLED
 
     // internal functions
     jl_builtin_apply_type = add_builtin_func("apply_type", jl_f_apply_type);
-    jl_builtin__apply = add_builtin_func("_apply", jl_f__apply);
     jl_builtin__apply_iterate = add_builtin_func("_apply_iterate", jl_f__apply_iterate);
     jl_builtin__expr = add_builtin_func("_expr", jl_f__expr);
     jl_builtin_svec = add_builtin_func("svec", jl_f_svec);

src/codegen.cpp

@@ -838,7 +838,6 @@ static const std::map<jl_fptr_args_t, JuliaFunction*> builtin_func_map = {
     { &jl_f_isa,                new JuliaFunction{"jl_f_isa", get_func_sig, get_func_attrs} },
     { &jl_f_typeassert,         new JuliaFunction{"jl_f_typeassert", get_func_sig, get_func_attrs} },
     { &jl_f_ifelse,             new JuliaFunction{"jl_f_ifelse", get_func_sig, get_func_attrs} },
-    { &jl_f__apply,             new JuliaFunction{"jl_f__apply", get_func_sig, get_func_attrs} },
     { &jl_f__apply_iterate,     new JuliaFunction{"jl_f__apply_iterate", get_func_sig, get_func_attrs} },
     { &jl_f__apply_pure,        new JuliaFunction{"jl_f__apply_pure", get_func_sig, get_func_attrs} },
     { &jl_f__call_latest,       new JuliaFunction{"jl_f__call_latest", get_func_sig, get_func_attrs} },
@@ -2684,13 +2683,11 @@ static bool emit_builtin_call(jl_codectx_t &ctx, jl_cgval_t *ret, jl_value_t *f,
         }
     }
 
-    else if (((f == jl_builtin__apply && nargs == 2) ||
-              (f == jl_builtin__apply_iterate && nargs == 3)) && ctx.vaSlot > 0) {
-        int arg_start = f == jl_builtin__apply ? 2 : 3;
-        // turn Core._apply(f, Tuple) ==> f(Tuple...) using the jlcall calling convention if Tuple is the va allocation
-        if (LoadInst *load = dyn_cast_or_null<LoadInst>(argv[arg_start].V)) {
+    else if ((f == jl_builtin__apply_iterate && nargs == 3) && ctx.vaSlot > 0) {
+        // turn Core._apply_iterate(iter, f, Tuple) ==> f(Tuple...) using the jlcall calling convention if Tuple is the va allocation
+        if (LoadInst *load = dyn_cast_or_null<LoadInst>(argv[3].V)) {
             if (load->getPointerOperand() == ctx.slots[ctx.vaSlot].boxroot && ctx.argArray) {
-                Value *theF = boxed(ctx, argv[arg_start-1]);
+                Value *theF = boxed(ctx, argv[2]);
                 Value *nva = emit_n_varargs(ctx);
 #ifdef _P64
                 nva = ctx.builder.CreateTrunc(nva, T_int32);

src/module.c

@@ -623,6 +623,8 @@ JL_DLLEXPORT jl_value_t *jl_get_global(jl_module_t *m, jl_sym_t *var)
 
 JL_DLLEXPORT void jl_set_global(jl_module_t *m JL_ROOTING_ARGUMENT, jl_sym_t *var, jl_value_t *val JL_ROOTED_ARGUMENT)
 {
+    JL_TYPECHK(jl_set_global, module, (jl_value_t*)m);
+    JL_TYPECHK(jl_set_global, symbol, (jl_value_t*)var);
     jl_binding_t *bp = jl_get_binding_wr(m, var, 1);
     JL_GC_PROMISE_ROOTED(bp);
     jl_checked_assignment(bp, val);

src/staticdata.c

@@ -30,7 +30,7 @@ extern "C" {
 // TODO: put WeakRefs on the weak_refs list during deserialization
 // TODO: handle finalizers
 
-#define NUM_TAGS    145
+#define NUM_TAGS    144
 
 // An array of references that need to be restored from the sysimg
 // This is a manually constructed dual of the gvars array, which would be produced by codegen for Julia code, for C.
@@ -175,7 +175,6 @@ jl_value_t **const*const get_tags(void) {
         INSERT_TAG(jl_builtin_issubtype);
         INSERT_TAG(jl_builtin_isa);
         INSERT_TAG(jl_builtin_typeassert);
-        INSERT_TAG(jl_builtin__apply);
         INSERT_TAG(jl_builtin__apply_iterate);
         INSERT_TAG(jl_builtin_isdefined);
         INSERT_TAG(jl_builtin_nfields);
@@ -235,7 +234,7 @@ void *native_functions;
 // This is a manually constructed dual of the fvars array, which would be produced by codegen for Julia code, for C.
 static const jl_fptr_args_t id_to_fptrs[] = {
     &jl_f_throw, &jl_f_is, &jl_f_typeof, &jl_f_issubtype, &jl_f_isa,
-    &jl_f_typeassert, &jl_f__apply, &jl_f__apply_iterate, &jl_f__apply_pure,
+    &jl_f_typeassert, &jl_f__apply_iterate, &jl_f__apply_pure,
     &jl_f__call_latest, &jl_f__call_in_world, &jl_f_isdefined,
     &jl_f_tuple, &jl_f_svec, &jl_f_intrinsic_call, &jl_f_invoke_kwsorter,
     &jl_f_getfield, &jl_f_setfield, &jl_f_fieldtype, &jl_f_nfields,

src/toplevel.c

@@ -45,15 +45,13 @@ JL_DLLEXPORT void jl_add_standard_imports(jl_module_t *m)
 // create a new top-level module
 void jl_init_main_module(void)
 {
-    if (jl_main_module != NULL)
-        jl_error("Main module already initialized.");
-
+    assert(jl_main_module == NULL);
     jl_main_module = jl_new_module(jl_symbol("Main"));
     jl_main_module->parent = jl_main_module;
     jl_set_const(jl_main_module, jl_symbol("Core"),
                  (jl_value_t*)jl_core_module);
-    jl_set_global(jl_core_module, jl_symbol("Main"),
-                  (jl_value_t*)jl_main_module);
+    jl_set_const(jl_core_module, jl_symbol("Main"),
+                 (jl_value_t*)jl_main_module);
 }
 
 static jl_function_t *jl_module_get_initializer(jl_module_t *m JL_PROPAGATES_ROOT)

stdlib/InteractiveUtils/src/macros.jl

@@ -153,12 +153,12 @@ function gen_call_with_extracted_types(__module__, fcn, ex0, kws=Expr[])
     exret = Expr(:none)
     if ex.head === :call
         if any(e->(isa(e, Expr) && e.head === :(...)), ex0.args) &&
-            (ex.args[1] === GlobalRef(Core,:_apply) ||
-             ex.args[1] === GlobalRef(Base,:_apply))
+            (ex.args[1] === GlobalRef(Core,:_apply_iterate) ||
+             ex.args[1] === GlobalRef(Base,:_apply_iterate))
             # check for splatting
-            exret = Expr(:call, ex.args[1], fcn,
-                        Expr(:tuple, esc(ex.args[2]),
-                            Expr(:call, typesof, map(esc, ex.args[3:end])...)))
+            exret = Expr(:call, ex.args[2], fcn,
+                        Expr(:tuple, esc(ex.args[3]),
+                            Expr(:call, typesof, map(esc, ex.args[4:end])...)))
         else
             exret = Expr(:call, fcn, esc(ex.args[1]),
                          Expr(:call, typesof, map(esc, ex.args[2:end])...), kws...)

test/compiler/inference.jl

@@ -855,7 +855,7 @@ end
 aa20704(x) = x(nothing)
 @test code_typed(aa20704, (typeof(a20704),))[1][1].pure
 
-#issue #21065, elision of _apply when splatted expression is not effect_free
+#issue #21065, elision of _apply_iterate when splatted expression is not effect_free
 function f21065(x,y)
     println("x=$x, y=$y")
     return x, y
@@ -865,7 +865,7 @@ function test_no_apply(expr::Expr)
     return all(test_no_apply, expr.args)
 end
 function test_no_apply(ref::GlobalRef)
-    return ref.mod != Core || ref.name !== :_apply
+    return ref.mod != Core || ref.name !== :_apply_iterate
 end
 test_no_apply(::Any) = true
 @test all(test_no_apply, code_typed(g21065, Tuple{Int,Int})[1].first.code)
@@ -2041,6 +2041,7 @@ T27078 = Vector{Vector{T}} where T
 # issue #28070
 g28070(f, args...) = f(args...)
 @test @inferred g28070(Core._apply, Base.:/, (1.0, 1.0)) == 1.0
+@test @inferred g28070(Core._apply_iterate, Base.iterate, Base.:/, (1.0, 1.0)) == 1.0
 
 # issue #28079
 struct Foo28079 end
@@ -2298,9 +2299,9 @@ end
 
 @test @inferred(g28955((1,), 1.0)) === Bool
 
-# Test that inlining can look through repeated _applys
+# Test that inlining can look through repeated _apply_iterates
 foo_inlining_apply(args...) = ccall(:jl_, Nothing, (Any,), args[1])
-bar_inlining_apply() = Core._apply(Core._apply, (foo_inlining_apply,), ((1,),))
+bar_inlining_apply() = Core._apply_iterate(iterate, Core._apply_iterate, (iterate,), (foo_inlining_apply,), ((1,),))
 let ci = code_typed(bar_inlining_apply, Tuple{})[1].first
     @test length(ci.code) == 2
     @test ci.code[1].head == :foreigncall

test/compiler/inline.jl

@@ -175,7 +175,7 @@ end
 # 2 for now because the compiler leaves a GotoNode around
 @test_broken length(code_typed(f_ifelse, (String,))[1][1].code) <= 2
 
-# Test that inlining of _apply properly hits the inference cache
+# Test that inlining of _apply_iterate properly hits the inference cache
 @noinline cprop_inline_foo1() = (1, 1)
 @noinline cprop_inline_foo2() = (2, 2)
 function cprop_inline_bar(x...)