Add --serialize-machine-code-only capability (#253)

As developed by Cody Tapscott of JuliaHub, delivered via Slack at:
https://relationalai.slack.com/archives/C01BC96H69W/p1759440236522629?thread_ts=1750270043.364009&cid=C01BC96H69W

Author: kpamnany

base/options.jl

@@ -62,6 +62,7 @@ struct JLOptions
     safe_crash_log_file::Ptr{UInt8}
     task_metrics::Int8
     timeout_for_safepoint_straggler_s::Int16
+    serialize_machine_code_only::Int8
 end
 
 # This runs early in the sysimage != is not defined yet

src/gf.c

@@ -2882,6 +2882,10 @@ static void _generate_from_hint(jl_method_instance_t *mi, size_t world)
         if (jl_atomic_load_relaxed(&((jl_code_instance_t*)codeinst)->invoke) == jl_fptr_const_return)
             return; // probably not a good idea to generate code
         jl_atomic_store_relaxed(&((jl_code_instance_t*)codeinst)->precompile, 1);
+        if (jl_options.serialize_machine_code_only) {
+            // also trigger compilation so that JIT-based compilation heuristics can succeed
+            (void)jl_compile_method_internal(mi, world);
+        }
     }
 }
 

src/jloptions.c

@@ -95,6 +95,7 @@ JL_DLLEXPORT void jl_init_options(void)
                         NULL, // safe_crash_log_file
                         0, // task_metrics
                         25, // timeout_for_safepoint_straggler_s
+                        0, // serialize_machine_code_only
     };
     jl_options_initialized = 1;
 }
@@ -245,6 +246,7 @@ JL_DLLEXPORT void jl_parse_opts(int *argcp, char ***argvp)
            opt_warn_scope,
            opt_inline,
            opt_polly,
+           opt_serialize_machine_code_only,
            opt_timeout_for_safepoint_straggler,
            opt_trace_compile,
            opt_trace_compile_timing,
@@ -325,6 +327,7 @@ JL_DLLEXPORT void jl_parse_opts(int *argcp, char ***argvp)
         { "warn-scope",      required_argument, 0, opt_warn_scope },
         { "inline",          required_argument, 0, opt_inline },
         { "polly",           required_argument, 0, opt_polly },
+        { "serialize-machine-code-only", no_argument, 0, opt_serialize_machine_code_only },
         { "timeout-for-safepoint-straggler", required_argument, 0, opt_timeout_for_safepoint_straggler },
         { "trace-compile",   required_argument, 0, opt_trace_compile },
         { "trace-compile-timing",  no_argument, 0, opt_trace_compile_timing },
@@ -893,6 +896,9 @@ JL_DLLEXPORT void jl_parse_opts(int *argcp, char ***argvp)
                 jl_errorf("julia: --timeout-for-safepoint-straggler=<seconds>; seconds must be an integer between 1 and %d", INT16_MAX);
             jl_options.timeout_for_safepoint_straggler_s = (int16_t)timeout;
             break;
+        case opt_serialize_machine_code_only:
+            jl_options.serialize_machine_code_only = 1;
+            break;
         case opt_task_metrics:
             if (!strcmp(optarg, "no"))
                 jl_options.task_metrics = JL_OPTIONS_TASK_METRICS_OFF;

src/jloptions.h

@@ -66,6 +66,7 @@ typedef struct {
     const char *safe_crash_log_file;
     int8_t task_metrics;
     int16_t timeout_for_safepoint_straggler_s;
+    int8_t serialize_machine_code_only;
 } jl_options_t;
 
 #endif

src/precompile_utils.c

@@ -184,14 +184,20 @@ static int precompile_enq_specialization_(jl_method_instance_t *mi, void *closur
         int do_compile = 0;
         if (jl_atomic_load_relaxed(&codeinst->invoke) != jl_fptr_const_return) {
             jl_value_t *inferred = jl_atomic_load_relaxed(&codeinst->inferred);
-            if (inferred &&
-                inferred != jl_nothing &&
-                jl_ir_flag_inferred(inferred) &&
-                (jl_ir_inlining_cost(inferred) == UINT16_MAX)) {
-                do_compile = 1;
-            }
-            else if (jl_atomic_load_relaxed(&codeinst->invoke) != NULL || jl_atomic_load_relaxed(&codeinst->precompile)) {
-                do_compile = 1;
+            if (jl_options.serialize_machine_code_only) {
+                if (jl_atomic_load_relaxed(&codeinst->invoke) != NULL) {
+                    do_compile = 1;
+                }
+            } else {
+                if (inferred &&
+                    inferred != jl_nothing &&
+                    jl_ir_flag_inferred(inferred) &&
+                    (jl_ir_inlining_cost(inferred) == UINT16_MAX)) {
+                    do_compile = 1;
+                }
+                else if (jl_atomic_load_relaxed(&codeinst->invoke) != NULL || jl_atomic_load_relaxed(&codeinst->precompile)) {
+                    do_compile = 1;
+                }
             }
         }
         if (do_compile) {

src/staticdata_utils.c

@@ -234,7 +234,10 @@ static jl_array_t *queue_external_cis(jl_array_t *list)
             continue;
         jl_method_instance_t *mi = ci->def;
         jl_method_t *m = mi->def.method;
-        if (ci->inferred && jl_is_method(m) && jl_object_in_image((jl_value_t*)m->module)) {
+        void *code = (jl_options.serialize_machine_code_only) ? (void *)ci->invoke : (void *)ci->inferred;
+        int for_serialized_method = jl_is_method(m) && jl_object_in_image((jl_value_t*)m->module);
+        int has_code = code != NULL;
+        if (has_code && for_serialized_method) {
             int found = has_backedge_to_worklist(mi, &visited, &stack);
             assert(found == 0 || found == 1 || found == 2);
             assert(stack.len == 0);

test/cmdlineargs.jl

@@ -1132,3 +1132,8 @@ end
     timeout = 120
     @test parse(Int,read(`$exename --timeout-for-safepoint-straggler=$timeout -E "Base.JLOptions().timeout_for_safepoint_straggler_s"`, String)) == timeout
 end
+
+@testset "--serialize-machine-code-only" begin
+    exename = `$(Base.julia_cmd())`
+    @test parse(Int,read(`$exename --serialize-machine-code-only -E "Base.JLOptions().serialize_machine_code_only"`, String)) == 1
+end