fix --compile=all option, and some improvements for --strip-ir (#…

…46935) (#47357)

(cherry picked from commit 6443a46)

base/compiler/utilities.jl

@@ -127,7 +127,10 @@ function retrieve_code_info(linfo::MethodInstance)
     end
     if c === nothing && isdefined(m, :source)
         src = m.source
-        if isa(src, Array{UInt8,1})
+        if src === nothing
+            # can happen in images built with --strip-ir
+            return nothing
+        elseif isa(src, Array{UInt8,1})
             c = ccall(:jl_uncompress_ir, Any, (Any, Ptr{Cvoid}, Any), m, C_NULL, src)
         else
             c = copy(src::CodeInfo)

src/gf.c

@@ -281,8 +281,6 @@ jl_code_info_t *jl_type_infer(jl_method_instance_t *mi, size_t world, int force)
     JL_TIMING(INFERENCE);
     if (jl_typeinf_func == NULL)
         return NULL;
-    if (jl_is_method(mi->def.method) && mi->def.method->unspecialized == mi)
-        return NULL; // avoid inferring the unspecialized method
     static int in_inference;
     if (in_inference > 2)
         return NULL;
@@ -2062,7 +2060,7 @@ jl_code_instance_t *jl_compile_method_internal(jl_method_instance_t *mi, size_t
     if (compile_option == JL_OPTIONS_COMPILE_OFF ||
         compile_option == JL_OPTIONS_COMPILE_MIN) {
         jl_code_info_t *src = jl_code_for_interpreter(mi);
-        if (!jl_code_requires_compiler(src)) {
+        if (!jl_code_requires_compiler(src, 0)) {
             jl_code_instance_t *codeinst = jl_new_codeinst(mi,
                 (jl_value_t*)jl_any_type, NULL, NULL,
                 0, 1, ~(size_t)0, 0, 0, jl_nothing, 0);

src/julia_internal.h

@@ -530,7 +530,7 @@ jl_method_instance_t *jl_get_unspecialized(jl_method_instance_t *method JL_PROPA
 
 JL_DLLEXPORT int jl_compile_hint(jl_tupletype_t *types);
 jl_code_info_t *jl_code_for_interpreter(jl_method_instance_t *lam JL_PROPAGATES_ROOT);
-int jl_code_requires_compiler(jl_code_info_t *src);
+int jl_code_requires_compiler(jl_code_info_t *src, int include_force_compile);
 jl_code_info_t *jl_new_code_info_from_ir(jl_expr_t *ast);
 JL_DLLEXPORT jl_code_info_t *jl_new_code_info_uninit(void);
 void jl_resolve_globals_in_ir(jl_array_t *stmts, jl_module_t *m, jl_svec_t *sparam_vals,

src/precompile.c

@@ -113,13 +113,6 @@ void jl_write_compiler_output(void)
 // and expanding the Union may give a leaf function
 static void _compile_all_tvar_union(jl_value_t *methsig)
 {
-    if (!jl_is_unionall(methsig) && jl_is_dispatch_tupletype(methsig)) {
-        // usually can create a specialized version of the function,
-        // if the signature is already a dispatch type
-        if (jl_compile_hint((jl_tupletype_t*)methsig))
-            return;
-    }
-
     int tvarslen = jl_subtype_env_size(methsig);
     jl_value_t *sigbody = methsig;
     jl_value_t **roots;
@@ -246,74 +239,49 @@ static void _compile_all_union(jl_value_t *sig)
     JL_GC_POP();
 }
 
-static void _compile_all_deq(jl_array_t *found)
-{
-    int found_i, found_l = jl_array_len(found);
-    jl_printf(JL_STDERR, "found %d uncompiled methods for compile-all\n", (int)found_l);
-    jl_method_instance_t *mi = NULL;
-    jl_value_t *src = NULL;
-    JL_GC_PUSH2(&mi, &src);
-    for (found_i = 0; found_i < found_l; found_i++) {
-        if (found_i % (1 + found_l / 300) == 0 || found_i == found_l - 1) // show 300 progress steps, to show progress without overwhelming log files
-            jl_printf(JL_STDERR, " %d / %d\r", found_i + 1, found_l);
-        jl_typemap_entry_t *ml = (jl_typemap_entry_t*)jl_array_ptr_ref(found, found_i);
-        jl_method_t *m = ml->func.method;
-        if (m->source == NULL) // TODO: generic implementations of generated functions
-            continue;
-        mi = jl_get_unspecialized(mi);
-        assert(mi == m->unspecialized); // make sure we didn't get tricked by a generated function, since we can't handle those
-        jl_code_instance_t *ucache = jl_get_method_inferred(mi, (jl_value_t*)jl_any_type, 1, ~(size_t)0);
-        if (ucache->invoke != NULL)
-            continue;
-        src = m->source;
-        assert(src);
-        // TODO: we could now enable storing inferred function pointers in the `unspecialized` cache
-        //src = jl_type_infer(mi, jl_atomic_load_acquire(&jl_world_counter), 1);
-        //if (ucache->invoke != NULL)
-        //    continue;
-
-        // first try to create leaf signatures from the signature declaration and compile those
-        _compile_all_union((jl_value_t*)ml->sig);
-        // then also compile the generic fallback
-        jl_generate_fptr_for_unspecialized(ucache);
-    }
-    JL_GC_POP();
-    jl_printf(JL_STDERR, "\n");
-}
-
-static int compile_all_enq__(jl_typemap_entry_t *ml, void *env)
+static int compile_all_collect__(jl_typemap_entry_t *ml, void *env)
 {
-    jl_array_t *found = (jl_array_t*)env;
-    // method definition -- compile template field
+    jl_array_t *allmeths = (jl_array_t*)env;
     jl_method_t *m = ml->func.method;
     if (m->source) {
-        // found a method to compile
-        jl_array_ptr_1d_push(found, (jl_value_t*)ml);
+        // method has a non-generated definition; can be compiled generically
+        jl_array_ptr_1d_push(allmeths, (jl_value_t*)m);
     }
     return 1;
 }
 
-
-static int compile_all_enq_(jl_methtable_t *mt, void *env)
+static int compile_all_collect_(jl_methtable_t *mt, void *env)
 {
-    jl_typemap_visitor(mt->defs, compile_all_enq__, env);
+    jl_typemap_visitor(mt->defs, compile_all_collect__, env);
     return 1;
 }
 
-static void jl_compile_all_defs(void)
+static void jl_compile_all_defs(jl_array_t *mis)
 {
-    // this "found" array will contain
-    // TypeMapEntries for Methods and MethodInstances that need to be compiled
-    jl_array_t *m = jl_alloc_vec_any(0);
-    JL_GC_PUSH1(&m);
-    while (1) {
-        jl_foreach_reachable_mtable(compile_all_enq_, m);
-        size_t changes = jl_array_len(m);
-        if (!changes)
-            break;
-        _compile_all_deq(m);
-        jl_array_del_end(m, changes);
+    jl_array_t *allmeths = jl_alloc_vec_any(0);
+    JL_GC_PUSH1(&allmeths);
+
+    jl_foreach_reachable_mtable(compile_all_collect_, allmeths);
+
+    size_t i, l = jl_array_len(allmeths);
+    for (i = 0; i < l; i++) {
+        jl_method_t *m = (jl_method_t*)jl_array_ptr_ref(allmeths, i);
+        if (jl_isa_compileable_sig((jl_tupletype_t*)m->sig, m)) {
+            // method has a single compileable specialization, e.g. its definition
+            // signature is concrete. in this case we can just hint it.
+            jl_compile_hint((jl_tupletype_t*)m->sig);
+        }
+        else {
+            // first try to create leaf signatures from the signature declaration and compile those
+            _compile_all_union(m->sig);
+
+            // finally, compile a fully generic fallback that can work for all arguments
+            jl_method_instance_t *unspec = jl_get_unspecialized(m);
+            if (unspec)
+                jl_array_ptr_1d_push(mis, (jl_value_t*)unspec);
+        }
     }
+
     JL_GC_POP();
 }
 
@@ -371,14 +339,13 @@ static int precompile_enq_all_specializations_(jl_methtable_t *mt, void *env)
 
 static void *jl_precompile(int all)
 {
-    if (all)
-        jl_compile_all_defs();
-    // this "found" array will contain function
-    // type signatures that were inferred but haven't been compiled
+    // array of MethodInstances and ccallable aliases to include in the output
     jl_array_t *m = jl_alloc_vec_any(0);
     jl_array_t *m2 = NULL;
     jl_method_instance_t *mi = NULL;
     JL_GC_PUSH3(&m, &m2, &mi);
+    if (all)
+        jl_compile_all_defs(m);
     jl_foreach_reachable_mtable(precompile_enq_all_specializations_, m);
     m2 = jl_alloc_vec_any(0);
     for (size_t i = 0; i < jl_array_len(m); i++) {
@@ -387,7 +354,7 @@ static void *jl_precompile(int all)
             mi = (jl_method_instance_t*)item;
             size_t min_world = 0;
             size_t max_world = ~(size_t)0;
-            if (!jl_isa_compileable_sig((jl_tupletype_t*)mi->specTypes, mi->def.method))
+            if (mi != jl_atomic_load_relaxed(&mi->def.method->unspecialized) && !jl_isa_compileable_sig((jl_tupletype_t*)mi->specTypes, mi->def.method))
                 mi = jl_get_specialization1((jl_tupletype_t*)mi->specTypes, jl_atomic_load_acquire(&jl_world_counter), &min_world, &max_world, 0);
             if (mi)
                 jl_array_ptr_1d_push(m2, (jl_value_t*)mi);

src/toplevel.c

@@ -337,7 +337,7 @@ JL_DLLEXPORT jl_module_t *jl_base_relative_to(jl_module_t *m)
     return jl_top_module;
 }
 
-static void expr_attributes(jl_value_t *v, int *has_intrinsics, int *has_defs, int *has_opaque)
+static void expr_attributes(jl_value_t *v, int *has_ccall, int *has_defs, int *has_opaque)
 {
     if (!jl_is_expr(v))
         return;
@@ -361,11 +361,11 @@ static void expr_attributes(jl_value_t *v, int *has_intrinsics, int *has_defs, i
         *has_defs = 1;
     }
     else if (head == jl_cfunction_sym) {
-        *has_intrinsics = 1;
+        *has_ccall = 1;
         return;
     }
     else if (head == jl_foreigncall_sym) {
-        *has_intrinsics = 1;
+        *has_ccall = 1;
         return;
     }
     else if (head == jl_new_opaque_closure_sym) {
@@ -389,7 +389,7 @@ static void expr_attributes(jl_value_t *v, int *has_intrinsics, int *has_defs, i
         }
         if (called) {
             if (jl_is_intrinsic(called) && jl_unbox_int32(called) == (int)llvmcall) {
-                *has_intrinsics = 1;
+                *has_ccall = 1;
             }
             if (called == jl_builtin__typebody) {
                 *has_defs = 1;
@@ -401,28 +401,28 @@ static void expr_attributes(jl_value_t *v, int *has_intrinsics, int *has_defs, i
     for (i = 0; i < jl_array_len(e->args); i++) {
         jl_value_t *a = jl_exprarg(e, i);
         if (jl_is_expr(a))
-            expr_attributes(a, has_intrinsics, has_defs, has_opaque);
+            expr_attributes(a, has_ccall, has_defs, has_opaque);
     }
 }
 
-int jl_code_requires_compiler(jl_code_info_t *src)
+int jl_code_requires_compiler(jl_code_info_t *src, int include_force_compile)
 {
     jl_array_t *body = src->code;
     assert(jl_typeis(body, jl_array_any_type));
     size_t i;
-    int has_intrinsics = 0, has_defs = 0, has_opaque = 0;
-    if (jl_has_meta(body, jl_force_compile_sym))
+    int has_ccall = 0, has_defs = 0, has_opaque = 0;
+    if (include_force_compile && jl_has_meta(body, jl_force_compile_sym))
         return 1;
     for(i=0; i < jl_array_len(body); i++) {
         jl_value_t *stmt = jl_array_ptr_ref(body,i);
-        expr_attributes(stmt, &has_intrinsics, &has_defs, &has_opaque);
-        if (has_intrinsics)
+        expr_attributes(stmt, &has_ccall, &has_defs, &has_opaque);
+        if (has_ccall)
             return 1;
     }
     return 0;
 }
 
-static void body_attributes(jl_array_t *body, int *has_intrinsics, int *has_defs, int *has_loops, int *has_opaque, int *forced_compile)
+static void body_attributes(jl_array_t *body, int *has_ccall, int *has_defs, int *has_loops, int *has_opaque, int *forced_compile)
 {
     size_t i;
     *has_loops = 0;
@@ -438,7 +438,7 @@ static void body_attributes(jl_array_t *body, int *has_intrinsics, int *has_defs
                     *has_loops = 1;
             }
         }
-        expr_attributes(stmt, has_intrinsics, has_defs, has_opaque);
+        expr_attributes(stmt, has_ccall, has_defs, has_opaque);
     }
     *forced_compile = jl_has_meta(body, jl_force_compile_sym);
 }
@@ -868,16 +868,16 @@ jl_value_t *jl_toplevel_eval_flex(jl_module_t *JL_NONNULL m, jl_value_t *e, int
         return (jl_value_t*)ex;
     }
 
-    int has_intrinsics = 0, has_defs = 0, has_loops = 0, has_opaque = 0, forced_compile = 0;
+    int has_ccall = 0, has_defs = 0, has_loops = 0, has_opaque = 0, forced_compile = 0;
     assert(head == jl_thunk_sym);
     thk = (jl_code_info_t*)jl_exprarg(ex, 0);
     assert(jl_is_code_info(thk));
     assert(jl_typeis(thk->code, jl_array_any_type));
-    body_attributes((jl_array_t*)thk->code, &has_intrinsics, &has_defs, &has_loops, &has_opaque, &forced_compile);
+    body_attributes((jl_array_t*)thk->code, &has_ccall, &has_defs, &has_loops, &has_opaque, &forced_compile);
 
     jl_value_t *result;
-    if (forced_compile || has_intrinsics ||
-            (!has_defs && fast && has_loops &&
+    if (has_ccall ||
+            ((forced_compile || (!has_defs && fast && has_loops)) &&
             jl_options.compile_enabled != JL_OPTIONS_COMPILE_OFF &&
             jl_options.compile_enabled != JL_OPTIONS_COMPILE_MIN &&
             jl_get_module_compile(m) != JL_OPTIONS_COMPILE_OFF &&