make ccallable work with system images and precompilation (#35574)

fixes #35014

Author: JeffBezanson

base/c.jl

@@ -482,8 +482,8 @@ function exit_on_sigint(on::Bool)
     ccall(:jl_exit_on_sigint, Cvoid, (Cint,), on)
 end
 
-function ccallable(f::Function, rt::Type, argt::Type, name::Union{AbstractString,Symbol}=string(f))
-    ccall(:jl_extern_c, Cvoid, (Any, Any, Any, Cstring), f, rt, argt, name)
+function _ccallable(rt::Type, sigt::Type)
+    ccall(:jl_extern_c, Cvoid, (Any, Any), rt, sigt)
 end
 
 function expand_ccallable(rt, def)
@@ -499,17 +499,22 @@ function expand_ccallable(rt, def)
             error("@ccallable requires a return type")
         end
         if sig.head === :call
-            name = sig.args[1]
+            f = sig.args[1]
+            if isa(f,Expr) && f.head === :(::)
+                f = f.args[end]
+            else
+                f = :(typeof($f))
+            end
             at = map(sig.args[2:end]) do a
                 if isa(a,Expr) && a.head === :(::)
-                    a.args[2]
+                    a.args[end]
                 else
                     :Any
                 end
             end
             return quote
                 $(esc(def))
-                ccallable($(esc(name)), $(esc(rt)), $(Expr(:curly, :Tuple, map(esc, at)...)), $(string(name)))
+                _ccallable($(esc(rt)), $(Expr(:curly, :Tuple, esc(f), map(esc, at)...)))
             end
         end
     end

src/anticodegen.c

@@ -66,5 +66,3 @@ JL_DLLEXPORT uint32_t jl_get_LLVM_VERSION(void)
 {
     return 0;
 }
-
-jl_array_t *jl_cfunction_list;

src/aotcompile.cpp

@@ -271,7 +271,14 @@ void *jl_create_native(jl_array_t *methods, const jl_cgparams_t cgparams)
             continue;
         size_t i, l;
         for (i = 0, l = jl_array_len(methods); i < l; i++) {
-            mi = (jl_method_instance_t*)jl_array_ptr_ref(methods, i);
+            // each item in this list is either a MethodInstance indicating something
+            // to compile, or an svec(rettype, sig) describing a C-callable alias to create.
+            jl_value_t *item = jl_array_ptr_ref(methods, i);
+            if (jl_is_simplevector(item) && worlds == 1) {
+                jl_compile_extern_c(shadow_output, &params, NULL, jl_svecref(item, 0), jl_svecref(item, 1));
+                continue;
+            }
+            mi = (jl_method_instance_t*)item;
             src = NULL;
             // if this method is generally visible to the current compilation world,
             // and this is either the primary world, or not applicable in the primary world

src/cgutils.cpp

@@ -774,6 +774,13 @@ static Type *julia_struct_to_llvm(jl_codectx_t &ctx, jl_value_t *jt, jl_unionall
     return _julia_struct_to_llvm(&ctx.emission_context, jt, ua, isboxed);
 }
 
+bool jl_type_mappable_to_c(jl_value_t *ty)
+{
+    jl_codegen_params_t params;
+    bool toboxed;
+    return _julia_struct_to_llvm(&params, ty, NULL, &toboxed) != NULL;
+}
+
 static bool is_datatype_all_pointers(jl_datatype_t *dt)
 {
     size_t i, l = jl_datatype_nfields(dt);

src/codegen.cpp

@@ -4014,8 +4014,8 @@ static void emit_cfunc_invalidate(
 
 static Function* gen_cfun_wrapper(
     Module *into, jl_codegen_params_t &params,
-    const function_sig_t &sig, jl_value_t *ff,
-    jl_typemap_entry_t *sf, jl_value_t *declrt, jl_method_instance_t *lam,
+    const function_sig_t &sig, jl_value_t *ff, const char *aliasname,
+    jl_value_t *declrt, jl_method_instance_t *lam,
     jl_unionall_t *unionall_env, jl_svec_t *sparam_vals, jl_array_t **closure_types)
 {
     // Generate a c-callable wrapper
@@ -4027,12 +4027,11 @@ static Function* gen_cfun_wrapper(
     jl_value_t *astrt = (jl_value_t*)jl_any_type;
     void *callptr = NULL;
     int calltype = 0;
-    // infer it first, if necessary
-    // FIXME! pretend this is OK
-    if (lam)
+    if (aliasname)
+        name = aliasname;
+    else if (lam)
         name = jl_symbol_name(lam->def.method->name);
     if (lam && params.cache) {
-        // if (!into)
         // TODO: this isn't ideal to be unconditionally calling type inference (and compile) from here
         codeinst = jl_compile_method_internal(lam, world);
         assert(codeinst->invoke);
@@ -4086,20 +4085,6 @@ static Function* gen_cfun_wrapper(
     cw->setAttributes(attributes);
     jl_init_function(cw);
     Function *cw_proto = into ? cw : function_proto(cw);
-    // Save the Function object reference
-    if (sf) {
-        jl_value_t *oldsf = sf->func.value;
-        size_t i, oldlen = jl_svec_len(oldsf);
-        jl_value_t *newsf = (jl_value_t*)jl_alloc_svec(oldlen + 2);
-        JL_GC_PUSH1(&newsf);
-        jl_svecset(newsf, 0, sig.rt);
-        jl_svecset(newsf, 1, jl_box_voidpointer((void*)cw_proto));
-        for (i = 0; i < oldlen; i++)
-            jl_svecset(newsf, i + 2, jl_svecref(oldsf, i));
-        sf->func.value = newsf;
-        jl_gc_wb(sf, sf->func.value);
-        JL_GC_POP();
-    }
 
     jl_codectx_t ctx(jl_LLVMContext, params);
     ctx.f = cw;
@@ -4515,6 +4500,10 @@ static Function* gen_cfun_wrapper(
         cw_proto = into ? cw_make : function_proto(cw_make);
     }
 
+    if (aliasname) {
+        GlobalAlias::create(cw->getType()->getElementType(), cw->getType()->getAddressSpace(),
+                            GlobalValue::ExternalLinkage, aliasname, cw, M);
+    }
     if (!into)
         jl_finalize_module(std::unique_ptr<Module>(M));
 
@@ -4628,8 +4617,8 @@ static jl_cgval_t emit_cfunction(jl_codectx_t &ctx, jl_value_t *output_type, con
     jl_method_instance_t *lam = sigt ? jl_get_specialization1((jl_tupletype_t*)sigt, world, &min_valid, &max_valid, 0) : NULL;
     Value *F = gen_cfun_wrapper(
             jl_Module, ctx.emission_context,
-            sig, fexpr_rt.constant,
-            NULL, declrt, lam,
+            sig, fexpr_rt.constant, NULL,
+            declrt, lam,
             unionall_env, sparam_vals, &closure_types);
     bool outboxed;
     if (nest) {
@@ -4689,116 +4678,54 @@ static jl_cgval_t emit_cfunction(jl_codectx_t &ctx, jl_value_t *output_type, con
     return mark_julia_type(ctx, F, outboxed, output_type);
 }
 
-const struct jl_typemap_info cfunction_cache = {
-    1, (jl_datatype_t**)&jl_array_any_type
-};
-
-jl_array_t *jl_cfunction_list;
-
-Function *jl_cfunction_object(jl_function_t *ff, jl_value_t *declrt, jl_tupletype_t *argt,
-                              jl_codegen_params_t &params)
+// do codegen to create a C-callable alias/wrapper, or if sysimg_handle is set,
+// restore one from a loaded system image.
+Function *jl_generate_ccallable(void *llvmmod, void *sysimg_handle, jl_value_t *declrt, jl_value_t *sigt, jl_codegen_params_t &params)
 {
-    // Assumes the codegen lock is acquired. The caller is responsible for that.
-    jl_ptls_t ptls = jl_get_ptls_states();
-    if (ptls->in_pure_callback)
-        jl_error("cfunction cannot be used in a generated function");
-
-    // validate and unpack the arguments
-    JL_TYPECHK(cfunction, type, declrt);
-    if (!jl_is_tuple_type(argt)) // the C API requires that argt Tuple type actually be an svec
-        jl_type_error("cfunction", (jl_value_t*)jl_anytuple_type_type, (jl_value_t*)argt);
-    // trampolines are not supported here:
-    // check that f is a guaranteed singleton type
-    jl_value_t *ft = jl_typeof(ff);
-    if (((jl_datatype_t*)ft)->instance != ff)
-        jl_error("cfunction: use `@cfunction` to make closures");
-
-    // check the cache structure
-    // this has three levels (for the 3 parameters above)
-    // first split on `ft` using a simple eqtable
-    // then use the typemap to split on argt
-    // and finally, pick declrt from the pair-list
-    jl_typemap_t *cache_l2 = NULL;
-    jl_typemap_entry_t *cache_l3 = NULL;
-    if (!jl_cfunction_list) {
-        jl_cfunction_list = jl_alloc_vec_any(16);
-    }
-    else {
-        cache_l2 = jl_eqtable_get(jl_cfunction_list, ft, NULL);
-        if (cache_l2) {
-            struct jl_typemap_assoc search = {(jl_value_t*)argt, 1, NULL, 0, ~(size_t)0};
-            cache_l3 = jl_typemap_assoc_by_type(cache_l2, &search, /*offs*/0, /*subtype*/0);
-            if (cache_l3) {
-                jl_svec_t *sf = (jl_svec_t*)cache_l3->func.value;
-                size_t i, l = jl_svec_len(sf);
-                for (i = 0; i < l; i += 2) {
-                    jl_value_t *ti = jl_svecref(sf, i);
-                    if (jl_egal(ti, declrt)) {
-                        return (Function*)jl_unbox_voidpointer(jl_svecref(sf, i + 1));
-                    }
-                }
-            }
-        }
-    }
-
-    if (cache_l3 == NULL) {
-        jl_typemap_t *insert = cache_l2;
-        if (!insert)
-            insert = jl_nothing;
-        cache_l3 = jl_typemap_insert(&insert, (jl_value_t*)insert, (jl_tupletype_t*)argt,
-            NULL, jl_emptysvec, (jl_value_t*)jl_emptysvec, /*offs*/0, &cfunction_cache, 1, ~(size_t)0);
-        if (insert != cache_l2)
-            jl_cfunction_list = jl_eqtable_put(jl_cfunction_list, ft, insert, NULL);
-    }
-
-    // compute / validate return type
+    jl_datatype_t *ft = (jl_datatype_t*)jl_tparam0(sigt);
+    jl_value_t *ff = ft->instance;
+    assert(ff);
+    const char *name = jl_symbol_name(ft->name->mt->name);
     jl_value_t *crt = declrt;
     if (jl_is_abstract_ref_type(declrt)) {
         declrt = jl_tparam0(declrt);
-        if (jl_is_typevar(declrt))
-            jl_error("cfunction: return type Ref should have an element type, not Ref{<:T}");
-        if (declrt == (jl_value_t*)jl_any_type)
-            jl_error("cfunction: return type Ref{Any} is invalid. Use Any or Ptr{Any} instead.");
         crt = (jl_value_t*)jl_any_type;
     }
     bool toboxed;
     Type *lcrt = _julia_struct_to_llvm(&params, crt, NULL, &toboxed);
-    if (lcrt == NULL)
-        jl_error("cfunction: return type doesn't correspond to a C type");
-    else if (toboxed)
+    if (toboxed)
         lcrt = T_prjlvalue;
-
-    // compute / validate method signature
-    jl_value_t *sigt = NULL; // dispatch sig: type signature (argt) with Ref{} annotations removed and ft added
-    JL_GC_PUSH1(&sigt);
-    size_t i, nargs = jl_nparams(argt);
-    sigt = (jl_value_t*)jl_alloc_svec(nargs + 1);
-    jl_svecset(sigt, 0, ft);
-    for (i = 0; i < nargs; i++) {
-        jl_value_t *ati = jl_tparam(argt, i);
-        if (jl_is_abstract_ref_type(ati)) {
-            ati = jl_tparam0(ati);
-            if (jl_is_typevar(ati))
-                jl_error("cfunction: argument type Ref should have an element type, not Ref{<:T}");
-        }
-        if (jl_is_pointer(ati) && jl_is_typevar(jl_tparam0(ati)))
-            jl_error("cfunction: argument type Ptr should have an element type, Ptr{<:T}");
-        jl_svecset(sigt, i + 1, ati);
-    }
-    sigt = (jl_value_t*)jl_apply_tuple_type((jl_svec_t*)sigt);
-
-    // emit cfunction (trampoline)
+    size_t nargs = jl_nparams(sigt)-1;
+    jl_svec_t *argtypes = NULL;
+    JL_GC_PUSH1(&argtypes);
+    argtypes = jl_alloc_svec(nargs);
+    for (size_t i = 0; i < nargs; i++) {
+        jl_svecset(argtypes, i, jl_tparam(sigt, i+1));
+    }
     jl_value_t *err;
     { // scope block for sig
         function_sig_t sig("cfunction", lcrt, crt, toboxed,
-                           argt->parameters, NULL, false, CallingConv::C, false, &params);
+                           argtypes, NULL, false, CallingConv::C, false, &params);
         if (sig.err_msg.empty()) {
             size_t world = jl_world_counter;
             size_t min_valid = 0;
             size_t max_valid = ~(size_t)0;
-            // try to look up this function for direct invoking
-            jl_method_instance_t *lam = jl_get_specialization1((jl_tupletype_t*)sigt, world, &min_valid, &max_valid, 0);
-            Function *F = gen_cfun_wrapper(NULL, params, sig, ff, cache_l3, declrt, lam, NULL, NULL, NULL);
+            Function *F = NULL;
+            if (sysimg_handle) {
+                // restore a ccallable from the system image
+                void *addr;
+                int found = jl_dlsym(sysimg_handle, name, &addr, 0);
+                if (found) {
+                    FunctionType *ftype = sig.functype();
+                    F = Function::Create(ftype, GlobalVariable::ExternalLinkage,
+                                         name, shadow_output);
+                    add_named_global(F, addr);
+                }
+            }
+            else {
+                jl_method_instance_t *lam = jl_get_specialization1((jl_tupletype_t*)sigt, world, &min_valid, &max_valid, 0);
+                F = gen_cfun_wrapper((Module*)llvmmod, params, sig, ff, name, declrt, lam, NULL, NULL, NULL);
+            }
             JL_GC_POP();
             return F;
         }

src/dump.c

@@ -69,6 +69,9 @@ static jl_array_t *serializer_worklist JL_GLOBALLY_ROOTED;
 // (only used by the incremental serializer in MODE_MODULE)
 htable_t edges_map;
 
+// list of requested ccallable signatures
+static arraylist_t ccallable_list;
+
 #define TAG_SYMBOL              2
 #define TAG_SSAVALUE            3
 #define TAG_DATATYPE            4
@@ -873,6 +876,7 @@ static void jl_serialize_value_(jl_serializer_state *s, jl_value_t *v, int as_li
         write_int8(s->s, m->pure);
         jl_serialize_value(s, (jl_value_t*)m->slot_syms);
         jl_serialize_value(s, (jl_value_t*)m->roots);
+        jl_serialize_value(s, (jl_value_t*)m->ccallable);
         jl_serialize_value(s, (jl_value_t*)m->source);
         jl_serialize_value(s, (jl_value_t*)m->unspecialized);
         jl_serialize_value(s, (jl_value_t*)m->generator);
@@ -1778,6 +1782,11 @@ static jl_value_t *jl_deserialize_value_method(jl_serializer_state *s, jl_value_
     m->roots = (jl_array_t*)jl_deserialize_value(s, (jl_value_t**)&m->roots);
     if (m->roots)
         jl_gc_wb(m, m->roots);
+    m->ccallable = (jl_svec_t*)jl_deserialize_value(s, (jl_value_t**)&m->ccallable);
+    if (m->ccallable) {
+        jl_gc_wb(m, m->ccallable);
+        arraylist_push(&ccallable_list, m->ccallable);
+    }
     m->source = jl_deserialize_value(s, &m->source);
     if (m->source)
         jl_gc_wb(m, m->source);
@@ -3217,6 +3226,7 @@ static jl_value_t *_jl_restore_incremental(ios_t *f, jl_array_t *mod_array)
     arraylist_new(&backref_list, 4000);
     arraylist_push(&backref_list, jl_main_module);
     arraylist_new(&flagref_list, 0);
+    arraylist_new(&ccallable_list, 0);
     htable_new(&uniquing_table, 0);
 
     jl_serializer_state s = {
@@ -3267,6 +3277,12 @@ static jl_value_t *_jl_restore_incremental(ios_t *f, jl_array_t *mod_array)
         arraylist_free(tracee_list);
         free(tracee_list);
     }
+    for (int i = 0; i < ccallable_list.len; i++) {
+        jl_svec_t *item = (jl_svec_t*)ccallable_list.items[i];
+        JL_GC_PROMISE_ROOTED(item);
+        jl_compile_extern_c(NULL, NULL, NULL, jl_svecref(item, 0), jl_svecref(item, 1));
+    }
+    arraylist_free(&ccallable_list);
     jl_value_t *ret = (jl_value_t*)jl_svec(2, restored, init_order);
     JL_GC_POP();
 

src/gc.c

@@ -2672,8 +2672,6 @@ static void mark_roots(jl_gc_mark_cache_t *gc_cache, jl_gc_mark_sp_t *sp)
             gc_mark_queue_obj(gc_cache, sp, jl_current_modules.table[i]);
         }
     }
-    if (jl_cfunction_list != NULL)
-        gc_mark_queue_obj(gc_cache, sp, jl_cfunction_list);
     gc_mark_queue_obj(gc_cache, sp, jl_anytuple_type_type);
     for (size_t i = 0; i < N_CALL_CACHE; i++)
         if (call_cache[i])