Remove MJIT worker thread (#6006)

[Misc #18830]

8 files changed, 351 insertions, 465 deletions

diff --git a/common.mk b/common.mk
index db4a283602..56139211de 100644
--- a/common.mk
+++ b/common.mk
@@ -9657,6 +9657,8 @@ mjit.$(OBJEXT): {$(VPATH)}mjit_worker.c
 mjit.$(OBJEXT): {$(VPATH)}node.h
 mjit.$(OBJEXT): {$(VPATH)}onigmo.h
 mjit.$(OBJEXT): {$(VPATH)}oniguruma.h
+mjit.$(OBJEXT): {$(VPATH)}ractor.h
+mjit.$(OBJEXT): {$(VPATH)}ractor_core.h
 mjit.$(OBJEXT): {$(VPATH)}ruby_assert.h
 mjit.$(OBJEXT): {$(VPATH)}ruby_atomic.h
 mjit.$(OBJEXT): {$(VPATH)}st.h
@@ -13800,6 +13802,7 @@ signal.$(OBJEXT): $(CCAN_DIR)/check_type/check_type.h
 signal.$(OBJEXT): $(CCAN_DIR)/container_of/container_of.h
 signal.$(OBJEXT): $(CCAN_DIR)/list/list.h
 signal.$(OBJEXT): $(CCAN_DIR)/str/str.h
+signal.$(OBJEXT): $(hdrdir)/ruby.h
 signal.$(OBJEXT): $(hdrdir)/ruby/ruby.h
 signal.$(OBJEXT): $(top_srcdir)/internal/array.h
 signal.$(OBJEXT): $(top_srcdir)/internal/compilers.h
@@ -13985,6 +13988,7 @@ signal.$(OBJEXT): {$(VPATH)}internal/warning_push.h
 signal.$(OBJEXT): {$(VPATH)}internal/xmalloc.h
 signal.$(OBJEXT): {$(VPATH)}method.h
 signal.$(OBJEXT): {$(VPATH)}missing.h
+signal.$(OBJEXT): {$(VPATH)}mjit.h
 signal.$(OBJEXT): {$(VPATH)}node.h
 signal.$(OBJEXT): {$(VPATH)}onigmo.h
 signal.$(OBJEXT): {$(VPATH)}oniguruma.h
@@ -14000,6 +14004,7 @@ signal.$(OBJEXT): {$(VPATH)}thread_native.h
 signal.$(OBJEXT): {$(VPATH)}vm_core.h
 signal.$(OBJEXT): {$(VPATH)}vm_debug.h
 signal.$(OBJEXT): {$(VPATH)}vm_opts.h
+signal.$(OBJEXT): {$(VPATH)}yjit.h
 sprintf.$(OBJEXT): $(hdrdir)/ruby/ruby.h
 sprintf.$(OBJEXT): $(top_srcdir)/internal/bignum.h
 sprintf.$(OBJEXT): $(top_srcdir)/internal/bits.h
diff --git a/gc.c b/gc.c
index b3b44e87a1..106f7bb2b3 100644
--- a/gc.c
+++ b/gc.c
@@ -9497,8 +9497,6 @@ gc_enter(rb_objspace_t *objspace, enum gc_enter_event event, unsigned int *lock_
     if (UNLIKELY(during_gc != 0)) rb_bug("during_gc != 0");
     if (RGENGC_CHECK_MODE >= 3) gc_verify_internal_consistency(objspace);
 
-    mjit_gc_start_hook();
-
     during_gc = TRUE;
     RUBY_DEBUG_LOG("%s (%s)",gc_enter_event_cstr(event), gc_current_status(objspace));
     gc_report(1, objspace, "gc_enter: %s [%s]\n", gc_enter_event_cstr(event), gc_current_status(objspace));
@@ -9517,7 +9515,6 @@ gc_exit(rb_objspace_t *objspace, enum gc_enter_event event, unsigned int *lock_l
     gc_report(1, objspace, "gc_exit: %s [%s]\n", gc_enter_event_cstr(event), gc_current_status(objspace));
     during_gc = FALSE;
 
-    mjit_gc_exit_hook();
     gc_exit_clock(objspace, event);
     RB_VM_LOCK_LEAVE_LEV(lock_lev);
 
diff --git a/mjit.c b/mjit.c
index b36331e72e..93e3d823e4 100644
--- a/mjit.c
+++ b/mjit.c
@@ -25,6 +25,7 @@
 #include "internal/hash.h"
 #include "internal/warnings.h"
 #include "vm_sync.h"
+#include "ractor_core.h"
 
 #include "mjit_worker.c"
 
@@ -50,40 +51,6 @@ get_uniq_filename(unsigned long id, const char *prefix, const char *suffix)
     return str;
 }
 
-// Wait until workers don't compile any iseq.  It is called at the
-// start of GC.
-void
-mjit_gc_start_hook(void)
-{
-    if (!mjit_enabled)
-        return;
-    CRITICAL_SECTION_START(4, "mjit_gc_start_hook");
-    while (in_jit) {
-        verbose(4, "Waiting wakeup from a worker for GC");
-        rb_native_cond_wait(&mjit_client_wakeup, &mjit_engine_mutex);
-        verbose(4, "Getting wakeup from a worker for GC");
-    }
-    in_gc++;
-    CRITICAL_SECTION_FINISH(4, "mjit_gc_start_hook");
-}
-
-// Send a signal to workers to continue iseq compilations.  It is
-// called at the end of GC.
-void
-mjit_gc_exit_hook(void)
-{
-    if (!mjit_enabled)
-        return;
-    CRITICAL_SECTION_START(4, "mjit_gc_exit_hook");
-    in_gc--;
-    RUBY_ASSERT_ALWAYS(in_gc >= 0);
-    if (!in_gc) {
-        verbose(4, "Sending wakeup signal to workers after GC");
-        rb_native_cond_broadcast(&mjit_gc_wakeup);
-    }
-    CRITICAL_SECTION_FINISH(4, "mjit_gc_exit_hook");
-}
-
 // Prohibit calling JIT-ed code and let existing JIT-ed frames exit before the next insn.
 void
 mjit_cancel_all(const char *reason)
@@ -133,9 +100,6 @@ mjit_free_iseq(const rb_iseq_t *iseq)
     if (!mjit_enabled)
         return;
 
-    CRITICAL_SECTION_START(4, "mjit_free_iseq");
-    RUBY_ASSERT_ALWAYS(in_gc);
-    RUBY_ASSERT_ALWAYS(!in_jit);
     if (ISEQ_BODY(iseq)->jit_unit) {
         // jit_unit is not freed here because it may be referred by multiple
         // lists of units. `get_from_list` and `mjit_finish` do the job.
@@ -150,7 +114,6 @@ mjit_free_iseq(const rb_iseq_t *iseq)
             unit->iseq = NULL;
         }
     }
-    CRITICAL_SECTION_FINISH(4, "mjit_free_iseq");
 }
 
 // Free unit list. This should be called only when worker is finished
@@ -245,19 +208,169 @@ finish_conts(void)
     }
 }
 
-// Create unit for `iseq`. This function may be called from an MJIT worker.
+static void mjit_wait(struct rb_iseq_constant_body *body);
+
+// Check the unit queue and start mjit_compile if nothing is in progress.
 static void
-create_unit(const rb_iseq_t *iseq)
+check_unit_queue(void)
 {
-    struct rb_mjit_unit *unit;
+    if (worker_stopped) return;
+    if (current_cc_pid != 0) return; // still compiling
+
+    // Run unload_units after it's requested `max_cache_size / 10` (default: 10) times.
+    // This throttles the call to mitigate locking in unload_units. It also throttles JIT compaction.
+    int throttle_threshold = mjit_opts.max_cache_size / 10;
+    if (unload_requests >= throttle_threshold) {
+        unload_units();
+        unload_requests = 0;
+        if (active_units.length == mjit_opts.max_cache_size && mjit_opts.wait) { // Sometimes all methods may be in use
+            mjit_opts.max_cache_size++; // avoid infinite loop on `rb_mjit_wait_call`. Note that --jit-wait is just for testing.
+            verbose(1, "No units can be unloaded -- incremented max-cache-size to %d for --jit-wait", mjit_opts.max_cache_size);
+        }
+    }
+    if (active_units.length >= mjit_opts.max_cache_size) return; // wait until unload_units makes a progress
 
-    unit = calloc(1, sizeof(struct rb_mjit_unit));
+    // Dequeue a unit
+    struct rb_mjit_unit *unit = get_from_list(&unit_queue);
+    if (unit == NULL) return;
+
+#ifdef _WIN32
+    // Synchronously compile methods on Windows.
+    // mswin: No SIGCHLD, MinGW: directly compiling .c to .so doesn't work
+    mjit_func_t func = convert_unit_to_func(unit);
+    if ((uintptr_t)func > (uintptr_t)LAST_JIT_ISEQ_FUNC) {
+        add_to_list(unit, &active_units);
+        MJIT_ATOMIC_SET(ISEQ_BODY(unit->iseq)->jit_func, func);
+    }
+#else
+    current_cc_ms = real_ms_time();
+    current_cc_unit = unit;
+    current_cc_pid = start_mjit_compile(unit);
+    // TODO: handle -1
+    if (mjit_opts.wait) {
+        mjit_wait(unit->iseq->body);
+    }
+#endif
+}
+
+// Create unit for `iseq`. This function may be called from an MJIT worker.
+static struct rb_mjit_unit*
+create_unit(const rb_iseq_t *iseq)
+{
+    // To prevent GC, don't use ZALLOC // TODO: just use ZALLOC
+    struct rb_mjit_unit *unit = calloc(1, sizeof(struct rb_mjit_unit));
     if (unit == NULL)
-        return;
+        return NULL;
 
     unit->id = current_unit_num++;
-    unit->iseq = (rb_iseq_t *)iseq;
-    ISEQ_BODY(iseq)->jit_unit = unit;
+    if (iseq == NULL) { // Compact unit
+        unit->compact_p = true;
+    } else { // Normal unit
+        unit->iseq = (rb_iseq_t *)iseq;
+        ISEQ_BODY(iseq)->jit_unit = unit;
+    }
+    return unit;
+}
+
+// Check if it should compact all JIT code and start it as needed
+static void
+check_compaction(void)
+{
+#if USE_JIT_COMPACTION
+    // Allow only `max_cache_size / 100` times (default: 100) of compaction.
+    // Note: GC of compacted code has not been implemented yet.
+    int max_compact_size = mjit_opts.max_cache_size / 100;
+    if (max_compact_size < 10) max_compact_size = 10;
+
+    // Run unload_units after it's requested `max_cache_size / 10` (default: 10) times.
+    // This throttles the call to mitigate locking in unload_units. It also throttles JIT compaction.
+    int throttle_threshold = mjit_opts.max_cache_size / 10;
+
+    if (compact_units.length < max_compact_size
+        && ((!mjit_opts.wait && unit_queue.length == 0 && active_units.length > 1)
+            || (active_units.length == mjit_opts.max_cache_size && compact_units.length * throttle_threshold <= total_unloads))) { // throttle compaction by total_unloads
+        struct rb_mjit_unit *unit = create_unit(NULL);
+        if (unit != NULL) {
+            // TODO: assert unit is null
+            current_cc_ms = real_ms_time();
+            current_cc_unit = unit;
+            current_cc_pid = start_mjit_compact(unit);
+            // TODO: check -1
+        }
+    }
+#endif
+}
+
+// Check the current CC process if any, and start a next C compiler process as needed.
+void
+mjit_notify_waitpid(int status)
+{
+    // TODO: check current_cc_pid?
+    current_cc_pid = 0;
+
+    // Delete .c file
+    char c_file[MAXPATHLEN];
+    sprint_uniq_filename(c_file, (int)sizeof(c_file), current_cc_unit->id, MJIT_TMP_PREFIX, ".c");
+    if (!mjit_opts.save_temps)
+        remove_file(c_file);
+
+    // Check the result
+    bool success = false;
+    if (WIFEXITED(status)) {
+        success = (WEXITSTATUS(status) == 0);
+    }
+    if (!success) {
+        verbose(2, "Failed to generate so");
+        // TODO: free unit?
+        // TODO: set NOT_COMPILED_JIT_ISEQ_FUNC?
+        return;
+    }
+
+    // Load .so file
+    char so_file[MAXPATHLEN];
+    sprint_uniq_filename(so_file, (int)sizeof(so_file), current_cc_unit->id, MJIT_TMP_PREFIX, DLEXT);
+    if (current_cc_unit->compact_p) { // Compact unit
+#if USE_JIT_COMPACTION
+        load_compact_funcs_from_so(current_cc_unit, c_file, so_file);
+        current_cc_unit = NULL;
+#else
+        RUBY_ASSERT(!current_cc_unit->compact_p);
+#endif
+    }
+    else { // Normal unit
+        // Load the function from so
+        char funcname[MAXPATHLEN];
+        sprint_funcname(funcname, current_cc_unit);
+        void *func = load_func_from_so(so_file, funcname, current_cc_unit);
+
+        // Delete .so file
+        if (!mjit_opts.save_temps)
+            remove_file(so_file);
+
+        // Set the jit_func if successful
+        if ((uintptr_t)func > (uintptr_t)LAST_JIT_ISEQ_FUNC) {
+            rb_iseq_t *iseq = current_cc_unit->iseq;
+            double end_time = real_ms_time();
+            verbose(1, "JIT success (%.1fms): %s@%s:%ld -> %s",
+                    end_time - current_cc_ms, RSTRING_PTR(ISEQ_BODY(iseq)->location.label),
+                    RSTRING_PTR(rb_iseq_path(iseq)), FIX2LONG(ISEQ_BODY(iseq)->location.first_lineno), c_file);
+
+            add_to_list(current_cc_unit, &active_units);
+            MJIT_ATOMIC_SET(ISEQ_BODY(iseq)->jit_func, func);
+        }
+        current_cc_unit = NULL;
+
+        // Run compaction if it should
+        if (!stop_worker_p) {
+            check_compaction();
+        }
+    }
+
+    // Skip further compilation if mjit_finish is trying to stop it
+    if (!stop_worker_p) {
+        // Start the next one as needed
+        check_unit_queue();
+    }
 }
 
 // Return true if given ISeq body should be compiled by MJIT
@@ -273,23 +386,14 @@ mjit_target_iseq_p(struct rb_iseq_constant_body *body)
 static void
 mjit_add_iseq_to_process(const rb_iseq_t *iseq, const struct rb_mjit_compile_info *compile_info, bool recompile_p)
 {
-    if (!mjit_enabled || pch_status == PCH_FAILED)
+    // TODO: Support non-main Ractors
+    if (!mjit_enabled || pch_status == PCH_FAILED || !rb_ractor_main_p())
         return;
     if (!mjit_target_iseq_p(ISEQ_BODY(iseq))) {
         ISEQ_BODY(iseq)->jit_func = (mjit_func_t)NOT_COMPILED_JIT_ISEQ_FUNC; // skip mjit_wait
         return;
     }
 
-    if (!recompile_p) {
-        CRITICAL_SECTION_START(3, "in add_iseq_to_process");
-
-        // This prevents multiple Ractors from enqueueing the same ISeq twice.
-        if (rb_multi_ractor_p() && (uintptr_t)ISEQ_BODY(iseq)->jit_func != NOT_ADDED_JIT_ISEQ_FUNC) {
-            CRITICAL_SECTION_FINISH(3, "in add_iseq_to_process");
-            return;
-        }
-    }
-
     RB_DEBUG_COUNTER_INC(mjit_add_iseq_to_process);
     ISEQ_BODY(iseq)->jit_func = (mjit_func_t)NOT_READY_JIT_ISEQ_FUNC;
     create_unit(iseq);
@@ -302,12 +406,6 @@ mjit_add_iseq_to_process(const rb_iseq_t *iseq, const struct rb_mjit_compile_inf
     if (active_units.length >= mjit_opts.max_cache_size) {
         unload_requests++;
     }
-
-    if (!recompile_p) {
-        verbose(3, "Sending wakeup signal to workers in mjit_add_iseq_to_process");
-        rb_native_cond_broadcast(&mjit_worker_wakeup);
-        CRITICAL_SECTION_FINISH(3, "in add_iseq_to_process");
-    }
 }
 
 // Add ISEQ to be JITed in parallel with the current thread.
@@ -316,6 +414,7 @@ void
 rb_mjit_add_iseq_to_process(const rb_iseq_t *iseq)
 {
     mjit_add_iseq_to_process(iseq, NULL, false);
+    check_unit_queue();
 }
 
 // For this timeout seconds, --jit-wait will wait for JIT compilation finish.
@@ -324,23 +423,21 @@ rb_mjit_add_iseq_to_process(const rb_iseq_t *iseq)
 static void
 mjit_wait(struct rb_iseq_constant_body *body)
 {
+    pid_t initial_pid = current_cc_pid;
     struct timeval tv;
     int tries = 0;
     tv.tv_sec = 0;
     tv.tv_usec = 1000;
-    while (body->jit_func == (mjit_func_t)NOT_READY_JIT_ISEQ_FUNC) {
+    while (body == NULL ? current_cc_pid == initial_pid : body->jit_func == (mjit_func_t)NOT_READY_JIT_ISEQ_FUNC) { // TODO: refactor this
         tries++;
         if (tries / 1000 > MJIT_WAIT_TIMEOUT_SECONDS || pch_status == PCH_FAILED) {
-            CRITICAL_SECTION_START(3, "in rb_mjit_wait_call to set jit_func");
-            body->jit_func = (mjit_func_t)NOT_COMPILED_JIT_ISEQ_FUNC; // JIT worker seems dead. Give up.
-            CRITICAL_SECTION_FINISH(3, "in rb_mjit_wait_call to set jit_func");
+            if (body != NULL) {
+                body->jit_func = (mjit_func_t) NOT_COMPILED_JIT_ISEQ_FUNC; // JIT worker seems dead. Give up.
+            }
             mjit_warning("timed out to wait for JIT finish");
             break;
         }
 
-        CRITICAL_SECTION_START(3, "in rb_mjit_wait_call for a client wakeup");
-        rb_native_cond_broadcast(&mjit_worker_wakeup);
-        CRITICAL_SECTION_FINISH(3, "in rb_mjit_wait_call for a client wakeup");
         rb_thread_wait_for(tv);
     }
 }
@@ -377,24 +474,8 @@ mjit_recompile(const rb_iseq_t *iseq)
             RSTRING_PTR(rb_iseq_path(iseq)), FIX2INT(ISEQ_BODY(iseq)->location.first_lineno));
     assert(ISEQ_BODY(iseq)->jit_unit != NULL);
 
-    if (UNLIKELY(mjit_opts.wait)) {
-        CRITICAL_SECTION_START(3, "in rb_mjit_recompile_iseq");
-        remove_from_list(ISEQ_BODY(iseq)->jit_unit, &active_units);
-        add_to_list(ISEQ_BODY(iseq)->jit_unit, &stale_units);
-        mjit_add_iseq_to_process(iseq, &ISEQ_BODY(iseq)->jit_unit->compile_info, true);
-        CRITICAL_SECTION_FINISH(3, "in rb_mjit_recompile_iseq");
-        mjit_wait(ISEQ_BODY(iseq));
-    }
-    else {
-        // Lazily move active_units to stale_units to avoid race conditions around active_units with compaction.
-        // Also, it's lazily moved to unit_queue as well because otherwise it won't be added to stale_units properly.
-        // It's good to avoid a race condition between mjit_add_iseq_to_process and mjit_compile around jit_unit as well.
-        CRITICAL_SECTION_START(3, "in rb_mjit_recompile_iseq");
-        ISEQ_BODY(iseq)->jit_unit->stale_p = true;
-        ISEQ_BODY(iseq)->jit_func = (mjit_func_t)NOT_READY_JIT_ISEQ_FUNC;
-        pending_stale_p = true;
-        CRITICAL_SECTION_FINISH(3, "in rb_mjit_recompile_iseq");
-    }
+    mjit_add_iseq_to_process(iseq, &ISEQ_BODY(iseq)->jit_unit->compile_info, true);
+    check_unit_queue();
 }
 
 // Recompile iseq, disabling send optimization
@@ -642,17 +723,6 @@ start_worker(void)
 {
     stop_worker_p = false;
     worker_stopped = false;
-
-    if (!rb_thread_create_mjit_thread(mjit_worker)) {
-        mjit_enabled = false;
-        rb_native_mutex_destroy(&mjit_engine_mutex);
-        rb_native_cond_destroy(&mjit_pch_wakeup);
-        rb_native_cond_destroy(&mjit_client_wakeup);
-        rb_native_cond_destroy(&mjit_worker_wakeup);
-        rb_native_cond_destroy(&mjit_gc_wakeup);
-        verbose(1, "Failure in MJIT thread initialization\n");
-        return false;
-    }
     return true;
 }
 
@@ -816,21 +886,21 @@ mjit_init(const struct mjit_options *opts)
 
     // Initialize worker thread
     start_worker();
+
+#ifndef _MSC_VER
+    // TODO: Consider running C compiler asynchronously
+    make_pch();
+#endif
 }
 
 static void
 stop_worker(void)
 {
-    rb_execution_context_t *ec = GET_EC();
-
-    while (!worker_stopped) {
-        verbose(3, "Sending cancel signal to worker");
-        CRITICAL_SECTION_START(3, "in stop_worker");
-        stop_worker_p = true; // Setting this inside loop because RUBY_VM_CHECK_INTS may make this false.
-        rb_native_cond_broadcast(&mjit_worker_wakeup);
-        CRITICAL_SECTION_FINISH(3, "in stop_worker");
-        RUBY_VM_CHECK_INTS(ec);
+    stop_worker_p = true;
+    if (current_cc_unit != NULL) {
+        mjit_wait(current_cc_unit->iseq->body);
     }
+    worker_stopped = true;
 }
 
 // Stop JIT-compiling methods but compiled code is kept available.
@@ -846,15 +916,12 @@ mjit_pause(bool wait_p)
 
     // Flush all queued units with no option or `wait: true`
     if (wait_p) {
-        struct timeval tv;
-        tv.tv_sec = 0;
-        tv.tv_usec = 1000;
-
-        while (unit_queue.length > 0 && active_units.length < mjit_opts.max_cache_size) { // inverse of condition that waits for mjit_worker_wakeup
-            CRITICAL_SECTION_START(3, "in mjit_pause for a worker wakeup");
-            rb_native_cond_broadcast(&mjit_worker_wakeup);
-            CRITICAL_SECTION_FINISH(3, "in mjit_pause for a worker wakeup");
-            rb_thread_wait_for(tv);
+        while (current_cc_unit != NULL) {
+            if (current_cc_unit->compact_p) {
+                mjit_wait(NULL);
+            } else {
+                mjit_wait(current_cc_unit->iseq->body);
+            }
         }
     }
 
@@ -932,21 +999,8 @@ mjit_finish(bool close_handle_p)
     if (!mjit_enabled)
         return;
 
-    // Wait for pch finish
-    verbose(2, "Stopping worker thread");
-    CRITICAL_SECTION_START(3, "in mjit_finish to wakeup from pch");
-    // As our threads are detached, we could just cancel them.  But it
-    // is a bad idea because OS processes (C compiler) started by
-    // threads can produce temp files.  And even if the temp files are
-    // removed, the used C compiler still complaint about their
-    // absence.  So wait for a clean finish of the threads.
-    while (pch_status == PCH_NOT_READY) {
-        verbose(3, "Waiting wakeup from make_pch");
-        rb_native_cond_wait(&mjit_pch_wakeup, &mjit_engine_mutex);
-    }
-    CRITICAL_SECTION_FINISH(3, "in mjit_finish to wakeup from pch");
-
     // Stop worker
+    verbose(2, "Stopping worker thread");
     stop_worker();
 
     rb_native_mutex_destroy(&mjit_engine_mutex);
@@ -1002,7 +1056,6 @@ mjit_mark(void)
     //
     // Because an MJIT worker may modify active_units anytime, we need to convert
     // the linked list to an array to safely loop its ISeqs without keeping a lock.
-    CRITICAL_SECTION_START(4, "mjit_mark");
     int length = 0;
     if (compiling_iseqs != NULL) {
         while (compiling_iseqs[length]) length++;
@@ -1023,7 +1076,6 @@ mjit_mark(void)
         i++;
     }
     assert(i == length);
-    CRITICAL_SECTION_FINISH(4, "mjit_mark");
 
     for (i = 0; i < length; i++) {
         if (iseqs[i] == NULL) // ISeq is GC-ed
diff --git a/mjit.h b/mjit.h
index fb216da8b3..fad18208fb 100644
--- a/mjit.h
+++ b/mjit.h
@@ -95,14 +95,13 @@ RUBY_SYMBOL_EXPORT_END
 extern void mjit_cancel_all(const char *reason);
 extern bool mjit_compile(FILE *f, const rb_iseq_t *iseq, const char *funcname, int id);
 extern void mjit_init(const struct mjit_options *opts);
-extern void mjit_gc_start_hook(void);
-extern void mjit_gc_exit_hook(void);
 extern void mjit_free_iseq(const rb_iseq_t *iseq);
 extern void mjit_update_references(const rb_iseq_t *iseq);
 extern void mjit_mark(void);
 extern struct mjit_cont *mjit_cont_new(rb_execution_context_t *ec);
 extern void mjit_cont_free(struct mjit_cont *cont);
 extern void mjit_mark_cc_entries(const struct rb_iseq_constant_body *const body);
+extern void mjit_notify_waitpid(int status);
 
 #  ifdef MJIT_HEADER
 NOINLINE(static COLDFUNC VALUE mjit_exec_slowpath(rb_execution_context_t *ec, const rb_iseq_t *iseq, struct rb_iseq_constant_body *body));
@@ -215,8 +214,6 @@ void mjit_finish(bool close_handle_p);
 static inline void mjit_cancel_all(const char *reason){}
 static inline struct mjit_cont *mjit_cont_new(rb_execution_context_t *ec){return NULL;}
 static inline void mjit_cont_free(struct mjit_cont *cont){}
-static inline void mjit_gc_start_hook(void){}
-static inline void mjit_gc_exit_hook(void){}
 static inline void mjit_free_iseq(const rb_iseq_t *iseq){}
 static inline void mjit_mark(void){}
 static inline VALUE mjit_exec(rb_execution_context_t *ec) { return Qundef; /* unreachable */ }
diff --git a/mjit_worker.c b/mjit_worker.c
index 1d0150a2c8..6228031696 100644
--- a/mjit_worker.c
+++ b/mjit_worker.c
@@ -164,8 +164,8 @@ struct rb_mjit_unit {
 #endif
     // Only used by unload_units. Flag to check this unit is currently on stack or not.
     bool used_code_p;
-    // True if this is still in active_units but it's to be lazily removed
-    bool stale_p;
+    // True if it's a unit for JIT compaction
+    bool compact_p;
     // mjit_compile's optimization switches
     struct rb_mjit_compile_info compile_info;
     // captured CC values, they should be marked with iseq.
@@ -191,7 +191,7 @@ extern void rb_native_cond_broadcast(rb_nativethread_cond_t *cond);
 extern void rb_native_cond_wait(rb_nativethread_cond_t *cond, rb_nativethread_lock_t *mutex);
 
 // process.c
-extern rb_pid_t ruby_waitpid_locked(rb_vm_t *, rb_pid_t, int *status, int options, rb_nativethread_cond_t *cond);
+extern void mjit_add_waiting_pid(rb_vm_t *vm, rb_pid_t pid);
 
 // A copy of MJIT portion of MRI options since MJIT initialization.  We
 // need them as MJIT threads still can work when the most MRI data were
@@ -227,12 +227,6 @@ static rb_nativethread_cond_t mjit_client_wakeup;
 static rb_nativethread_cond_t mjit_worker_wakeup;
 // A thread conditional to wake up workers if at the end of GC.
 static rb_nativethread_cond_t mjit_gc_wakeup;
-// Greater than 0 when GC is working.
-static int in_gc = 0;
-// True when JIT is working.
-static bool in_jit = false;
-// True when active_units has at least one stale_p=true unit.
-static bool pending_stale_p = false;
 // The times when unload_units is requested. unload_units is called after some requests.
 static int unload_requests = 0;
 // The total number of unloaded units.
@@ -259,6 +253,13 @@ static rb_pid_t pch_owner_pid;
 // shared by the workers and the pch thread.
 static enum {PCH_NOT_READY, PCH_FAILED, PCH_SUCCESS} pch_status;
 
+// The start timestamp of current compilation
+static double current_cc_ms = 0.0; // TODO: make this part of unit?
+// Currently compiling MJIT unit
+static struct rb_mjit_unit *current_cc_unit = NULL;
+// PID of currently running C compiler process. 0 if nothing is running.
+static pid_t current_cc_pid = 0; // TODO: make this part of unit?
+
 #ifndef _MSC_VER
 // Name of the header file.
 static char *header_file;
@@ -492,12 +493,6 @@ real_ms_time(void)
 static struct rb_mjit_unit *
 get_from_list(struct rb_mjit_unit_list *list)
 {
-    while (in_gc) {
-        verbose(3, "Waiting wakeup from GC");
-        rb_native_cond_wait(&mjit_gc_wakeup, &mjit_engine_mutex);
-    }
-    in_jit = true; // Lock GC
-
     // Find iseq with max total_calls
     struct rb_mjit_unit *unit = NULL, *next, *best = NULL;
     ccan_list_for_each_safe(&list->head, unit, next, unode) {
@@ -512,10 +507,6 @@ get_from_list(struct rb_mjit_unit_list *list)
         }
     }
 
-    in_jit = false; // Unlock GC
-    verbose(3, "Sending wakeup signal to client in a mjit-worker for GC");
-    rb_native_cond_signal(&mjit_client_wakeup);
-
     if (best) {
         remove_from_list(best, list);
     }
@@ -632,18 +623,9 @@ static int
 exec_process(const char *path, char *const argv[])
 {
     int stat, exit_code = -2;
-    rb_vm_t *vm = WAITPID_USE_SIGCHLD ? GET_VM() : 0;
-    rb_nativethread_cond_t cond;
-
-    if (vm) {
-        rb_native_cond_initialize(&cond);
-        rb_native_mutex_lock(&vm->waitpid_lock);
-    }
-
     pid_t pid = start_process(path, argv);
     for (;pid > 0;) {
-        pid_t r = vm ? ruby_waitpid_locked(vm, pid, &stat, 0, &cond)
-                     : waitpid(pid, &stat, 0);
+        pid_t r = waitpid(pid, &stat, 0);
         if (r == -1) {
             if (errno == EINTR) continue;
             fprintf(stderr, "[%"PRI_PIDT_PREFIX"d] waitpid(%lu): %s (SIGCHLD=%d,%u)\n",
@@ -662,11 +644,6 @@ exec_process(const char *path, char *const argv[])
             }
         }
     }
-
-    if (vm) {
-        rb_native_mutex_unlock(&vm->waitpid_lock);
-        rb_native_cond_destroy(&cond);
-    }
     return exit_code;
 }
 
@@ -881,16 +858,13 @@ make_pch(void)
     char **args = form_args(4, cc_common_args, CC_CODEFLAG_ARGS, cc_added_args, rest_args);
     if (args == NULL) {
         mjit_warning("making precompiled header failed on forming args");
-        CRITICAL_SECTION_START(3, "in make_pch");
         pch_status = PCH_FAILED;
-        CRITICAL_SECTION_FINISH(3, "in make_pch");
         return;
     }
 
     int exit_code = exec_process(cc_path, args);
     free(args);
 
-    CRITICAL_SECTION_START(3, "in make_pch");
     if (exit_code == 0) {
         pch_status = PCH_SUCCESS;
     }
@@ -898,14 +872,10 @@ make_pch(void)
         mjit_warning("Making precompiled header failed on compilation. Stopping MJIT worker...");
         pch_status = PCH_FAILED;
     }
-    /* wakeup `mjit_finish` */
-    rb_native_cond_broadcast(&mjit_pch_wakeup);
-    CRITICAL_SECTION_FINISH(3, "in make_pch");
 }
 
-// Compile .c file to .so file. It returns true if it succeeds. (non-mswin)
-static bool
-compile_c_to_so(const char *c_file, const char *so_file)
+static pid_t
+start_compiling_c_to_so(const char *c_file, const char *so_file)
 {
     const char *so_args[] = {
         "-o", so_file,
@@ -919,21 +889,27 @@ compile_c_to_so(const char *c_file, const char *so_file)
     };
     char **args = form_args(7, CC_LDSHARED_ARGS, CC_CODEFLAG_ARGS, cc_added_args,
                             so_args, CC_LIBS, CC_DLDFLAGS_ARGS, CC_LINKER_ARGS);
-    if (args == NULL) return false;
-    int exit_code = exec_process(cc_path, args);
+    if (args == NULL) return -1;
+
+    rb_vm_t *vm = GET_VM();
+    rb_native_mutex_lock(&vm->waitpid_lock);
+
+    pid_t pid = start_process(cc_path, args);
+    mjit_add_waiting_pid(vm, pid);
+
+    rb_native_mutex_unlock(&vm->waitpid_lock);
+
     free(args);
-    if (exit_code != 0) {
-        verbose(2, "compile_c_to_so: failed to compile .c to .so: %d", exit_code);
-    }
-    return exit_code == 0;
+    return pid;
 }
 #endif // _MSC_VER
 
 #if USE_JIT_COMPACTION
 static void compile_prelude(FILE *f);
 
+// Compile all JIT code into a single .c file
 static bool
-compile_compact_jit_code(char* c_file)
+mjit_compact(char* c_file)
 {
     FILE *f;
     int fd = rb_cloexec_open(c_file, c_file_access_mode, 0600);
@@ -946,32 +922,6 @@ compile_compact_jit_code(char* c_file)
 
     compile_prelude(f);
 
-    // wait until mjit_gc_exit_hook is called
-    CRITICAL_SECTION_START(3, "before mjit_compile to wait GC finish");
-    while (in_gc) {
-        verbose(3, "Waiting wakeup from GC");
-        rb_native_cond_wait(&mjit_gc_wakeup, &mjit_engine_mutex);
-    }
-    // We need to check again here because we could've waited on GC above
-    bool iseq_gced = false;
-    struct rb_mjit_unit *child_unit = 0, *next;
-    ccan_list_for_each_safe(&active_units.head, child_unit, next, unode) {
-        if (child_unit->iseq == NULL) { // ISeq is GC-ed
-            iseq_gced = true;
-            verbose(1, "JIT compaction: A method for JIT code u%d is obsoleted. Compaction will be skipped.", child_unit->id);
-            remove_from_list(child_unit, &active_units);
-            free_unit(child_unit); // unload it without waiting for throttled unload_units to retry compaction quickly
-        }
-    }
-    in_jit = !iseq_gced;
-    CRITICAL_SECTION_FINISH(3, "before mjit_compile to wait GC finish");
-    if (!in_jit) {
-        fclose(f);
-        if (!mjit_opts.save_temps)
-            remove_file(c_file);
-        return false;
-    }
-
     // This entire loop lock GC so that we do not need to consider a case that
     // ISeq is GC-ed in a middle of re-compilation. It takes 3~4ms with 100 methods
     // on my machine. It's not too bad compared to compilation time of C (7200~8000ms),
@@ -980,10 +930,9 @@ compile_compact_jit_code(char* c_file)
     // TODO: Consider using a more granular lock after we implement inlining across
     // compacted functions (not done yet).
     bool success = true;
+    struct rb_mjit_unit *child_unit = 0;
     ccan_list_for_each(&active_units.head, child_unit, unode) {
-        CRITICAL_SECTION_START(3, "before set_compiling_iseqs");
         success &= set_compiling_iseqs(child_unit->iseq);
-        CRITICAL_SECTION_FINISH(3, "after set_compiling_iseqs");
         if (!success) continue;
 
         char funcname[MAXPATHLEN];
@@ -1000,86 +949,68 @@ compile_compact_jit_code(char* c_file)
         fprintf(f, "\n/* %s%s%s:%ld */\n", iseq_label, sep, iseq_path, iseq_lineno);
         success &= mjit_compile(f, child_unit->iseq, funcname, child_unit->id);
 
-        CRITICAL_SECTION_START(3, "before compiling_iseqs free");
         free_compiling_iseqs();
-        CRITICAL_SECTION_FINISH(3, "after compiling_iseqs free");
     }
 
-    // release blocking mjit_gc_start_hook
-    CRITICAL_SECTION_START(3, "after mjit_compile to wakeup client for GC");
-    in_jit = false;
-    verbose(3, "Sending wakeup signal to client in a mjit-worker for GC");
-    rb_native_cond_signal(&mjit_client_wakeup);
-    CRITICAL_SECTION_FINISH(3, "in worker to wakeup client for GC");
-
     fclose(f);
     return success;
 }
 
 // Compile all cached .c files and build a single .so file. Reload all JIT func from it.
 // This improves the code locality for better performance in terms of iTLB and iCache.
-static void
-compact_all_jit_code(void)
+static pid_t
+start_mjit_compact(struct rb_mjit_unit *unit)
 {
-    struct rb_mjit_unit *unit, *cur = 0;
     static const char c_ext[] = ".c";
     static const char so_ext[] = DLEXT;
     char c_file[MAXPATHLEN], so_file[MAXPATHLEN];
 
-    // Abnormal use case of rb_mjit_unit that doesn't have ISeq
-    unit = calloc(1, sizeof(struct rb_mjit_unit)); // To prevent GC, don't use ZALLOC
-    if (unit == NULL) return;
-    unit->id = current_unit_num++;
     sprint_uniq_filename(c_file, (int)sizeof(c_file), unit->id, MJIT_TMP_PREFIX, c_ext);
     sprint_uniq_filename(so_file, (int)sizeof(so_file), unit->id, MJIT_TMP_PREFIX, so_ext);
 
-    bool success = compile_compact_jit_code(c_file);
-    double start_time = real_ms_time();
+    bool success = mjit_compact(c_file);
     if (success) {
-        success = compile_c_to_so(c_file, so_file);
-        if (!mjit_opts.save_temps)
-            remove_file(c_file);
+        return start_compiling_c_to_so(c_file, so_file);
     }
+    return -1;
+}
+
+static void
+load_compact_funcs_from_so(struct rb_mjit_unit *unit, char *c_file, char *so_file)
+{
+    struct rb_mjit_unit *cur = 0;
     double end_time = real_ms_time();
 
-    if (success) {
-        void *handle = dlopen(so_file, RTLD_NOW);
-        if (handle == NULL) {
-            mjit_warning("failure in loading code from compacted '%s': %s", so_file, dlerror());
-            free(unit);
-            return;
-        }
-        unit->handle = handle;
+    void *handle = dlopen(so_file, RTLD_NOW);
+    if (handle == NULL) {
+        mjit_warning("failure in loading code from compacted '%s': %s", so_file, dlerror());
+        free(unit);
+        return;
+    }
+    unit->handle = handle;
 
-        // lazily dlclose handle (and .so file for win32) on `mjit_finish()`.
-        add_to_list(unit, &compact_units);
+    // lazily dlclose handle (and .so file for win32) on `mjit_finish()`.
+    add_to_list(unit, &compact_units);
 
-        if (!mjit_opts.save_temps)
-            remove_so_file(so_file, unit);
+    if (!mjit_opts.save_temps)
+        remove_so_file(so_file, unit);
 
-        CRITICAL_SECTION_START(3, "in compact_all_jit_code to read list");
-        ccan_list_for_each(&active_units.head, cur, unode) {
-            void *func;
-            char funcname[MAXPATHLEN];
-            sprint_funcname(funcname, cur);
+    ccan_list_for_each(&active_units.head, cur, unode) {
+        void *func;
+        char funcname[MAXPATHLEN];
+        sprint_funcname(funcname, cur);
 
-            if ((func = dlsym(handle, funcname)) == NULL) {
-                mjit_warning("skipping to reload '%s' from '%s': %s", funcname, so_file, dlerror());
-                continue;
-            }
+        if ((func = dlsym(handle, funcname)) == NULL) {
+            mjit_warning("skipping to reload '%s' from '%s': %s", funcname, so_file, dlerror());
+            continue;
+        }
 
-            if (cur->iseq) { // Check whether GCed or not
-                // Usage of jit_code might be not in a critical section.
-                MJIT_ATOMIC_SET(ISEQ_BODY(cur->iseq)->jit_func, (mjit_func_t)func);
-            }
+        if (cur->iseq) { // Check whether GCed or not
+            // Usage of jit_code might be not in a critical section.
+            MJIT_ATOMIC_SET(ISEQ_BODY(cur->iseq)->jit_func, (mjit_func_t)func);
         }
-        CRITICAL_SECTION_FINISH(3, "in compact_all_jit_code to read list");
-        verbose(1, "JIT compaction (%.1fms): Compacted %d methods %s -> %s", end_time - start_time, active_units.length, c_file, so_file);
-    }
-    else {
-        free(unit);
-        verbose(1, "JIT compaction failure (%.1fms): Failed to compact methods", end_time - start_time);
     }
+    verbose(1, "JIT compaction (%.1fms): Compacted %d methods %s -> %s", end_time - current_cc_ms, active_units.length, c_file, so_file);
 }
 #endif // USE_JIT_COMPACTION
 
@@ -1144,6 +1075,62 @@ compile_prelude(FILE *f)
 
 // Compile ISeq in UNIT and return function pointer of JIT-ed code.
 // It may return NOT_COMPILED_JIT_ISEQ_FUNC if something went wrong.
+static pid_t
+start_mjit_compile(struct rb_mjit_unit *unit)
+{
+    static const char c_ext[] = ".c";
+    static const char so_ext[] = DLEXT;
+    char c_file[MAXPATHLEN], so_file[MAXPATHLEN], funcname[MAXPATHLEN];
+
+    sprint_uniq_filename(c_file, (int)sizeof(c_file), unit->id, MJIT_TMP_PREFIX, c_ext);
+    sprint_uniq_filename(so_file, (int)sizeof(so_file), unit->id, MJIT_TMP_PREFIX, so_ext);
+    sprint_funcname(funcname, unit);
+
+    FILE *f;
+    int fd = rb_cloexec_open(c_file, c_file_access_mode, 0600);
+    if (fd < 0 || (f = fdopen(fd, "w")) == NULL) {
+        int e = errno;
+        if (fd >= 0) (void)close(fd);
+        verbose(1, "Failed to fopen '%s', giving up JIT for it (%s)", c_file, strerror(e));
+        return -1;
+    }
+
+    // print #include of MJIT header, etc.
+    compile_prelude(f);
+
+    // This is no longer necessary. TODO: Just reference the ISeq directly in the compiler.
+    if (!set_compiling_iseqs(unit->iseq)) return -1;
+
+    // To make MJIT worker thread-safe against GC.compact, copy ISeq values while `in_jit` is true.
+    long iseq_lineno = 0;
+    if (FIXNUM_P(ISEQ_BODY(unit->iseq)->location.first_lineno))
+        // FIX2INT may fallback to rb_num2long(), which is a method call and dangerous in MJIT worker. So using only FIX2LONG.
+        iseq_lineno = FIX2LONG(ISEQ_BODY(unit->iseq)->location.first_lineno);
+    char *iseq_label = alloca(RSTRING_LEN(ISEQ_BODY(unit->iseq)->location.label) + 1);
+    char *iseq_path  = alloca(RSTRING_LEN(rb_iseq_path(unit->iseq)) + 1);
+    strcpy(iseq_label, RSTRING_PTR(ISEQ_BODY(unit->iseq)->location.label));
+    strcpy(iseq_path,  RSTRING_PTR(rb_iseq_path(unit->iseq)));
+
+    verbose(2, "start compilation: %s@%s:%ld -> %s", iseq_label, iseq_path, iseq_lineno, c_file);
+    fprintf(f, "/* %s@%s:%ld */\n\n", iseq_label, iseq_path, iseq_lineno);
+    bool success = mjit_compile(f, unit->iseq, funcname, unit->id);
+
+    free_compiling_iseqs();
+
+    fclose(f);
+    if (!success) {
+        if (!mjit_opts.save_temps)
+            remove_file(c_file);
+        verbose(1, "JIT failure: %s@%s:%ld -> %s", iseq_label, iseq_path, iseq_lineno, c_file);
+        return -1;
+    }
+
+    return start_compiling_c_to_so(c_file, so_file);
+}
+
+#ifdef _WIN32
+// Compile ISeq in UNIT and return function pointer of JIT-ed code.
+// It may return NOT_COMPILED_JIT_ISEQ_FUNC if something went wrong.
 static mjit_func_t
 convert_unit_to_func(struct rb_mjit_unit *unit)
 {
@@ -1167,18 +1154,7 @@ convert_unit_to_func(struct rb_mjit_unit *unit)
     // print #include of MJIT header, etc.
     compile_prelude(f);
 
-    // wait until mjit_gc_exit_hook is called
-    CRITICAL_SECTION_START(3, "before mjit_compile to wait GC finish");
-    while (in_gc) {
-        verbose(3, "Waiting wakeup from GC");
-        rb_native_cond_wait(&mjit_gc_wakeup, &mjit_engine_mutex);
-    }
-    // We need to check again here because we could've waited on GC above
-    in_jit = (unit->iseq != NULL);
-    if (in_jit)
-        in_jit &= set_compiling_iseqs(unit->iseq);
-    CRITICAL_SECTION_FINISH(3, "before mjit_compile to wait GC finish");
-    if (!in_jit) {
+    if (!set_compiling_iseqs(unit->iseq)) {
         fclose(f);
         if (!mjit_opts.save_temps)
             remove_file(c_file);
@@ -1200,12 +1176,7 @@ convert_unit_to_func(struct rb_mjit_unit *unit)
     bool success = mjit_compile(f, unit->iseq, funcname, unit->id);
 
     // release blocking mjit_gc_start_hook
-    CRITICAL_SECTION_START(3, "after mjit_compile to wakeup client for GC");
     free_compiling_iseqs();
-    in_jit = false;
-    verbose(3, "Sending wakeup signal to client in a mjit-worker for GC");
-    rb_native_cond_signal(&mjit_client_wakeup);
-    CRITICAL_SECTION_FINISH(3, "in worker to wakeup client for GC");
 
     fclose(f);
     if (!success) {
@@ -1236,6 +1207,7 @@ convert_unit_to_func(struct rb_mjit_unit *unit)
     }
     return (mjit_func_t)func;
 }
+#endif
 
 // To see cc_entries using index returned by `mjit_capture_cc_entries` in mjit_compile.c
 const struct rb_callcache **
@@ -1382,119 +1354,3 @@ unload_units(void)
 }
 
 static void mjit_add_iseq_to_process(const rb_iseq_t *iseq, const struct rb_mjit_compile_info *compile_info, bool worker_p);
-
-// The function implementing a worker. It is executed in a separate
-// thread by rb_thread_create_mjit_thread. It compiles precompiled header
-// and then compiles requested ISeqs.
-void
-mjit_worker(void)
-{
-    // Allow only `max_cache_size / 100` times (default: 100) of compaction.
-    // Note: GC of compacted code has not been implemented yet.
-    int max_compact_size = mjit_opts.max_cache_size / 100;
-    if (max_compact_size < 10) max_compact_size = 10;
-
-    // Run unload_units after it's requested `max_cache_size / 10` (default: 10) times.
-    // This throttles the call to mitigate locking in unload_units. It also throttles JIT compaction.
-    int throttle_threshold = mjit_opts.max_cache_size / 10;
-
-#ifndef _MSC_VER
-    if (pch_status == PCH_NOT_READY) {
-        make_pch();
-    }
-#endif
-    if (pch_status == PCH_FAILED) {
-        mjit_enabled = false;
-        CRITICAL_SECTION_START(3, "in worker to update worker_stopped");
-        worker_stopped = true;
-        verbose(3, "Sending wakeup signal to client in a mjit-worker");
-        rb_native_cond_signal(&mjit_client_wakeup);
-        CRITICAL_SECTION_FINISH(3, "in worker to update worker_stopped");
-        return; // TODO: do the same thing in the latter half of mjit_finish
-    }
-
-    // main worker loop
-    while (!stop_worker_p) {
-        struct rb_mjit_unit *unit;
-
-        // Wait until a unit becomes available
-        CRITICAL_SECTION_START(3, "in worker dequeue");
-        while ((ccan_list_empty(&unit_queue.head) || active_units.length >= mjit_opts.max_cache_size) && !stop_worker_p) {
-            rb_native_cond_wait(&mjit_worker_wakeup, &mjit_engine_mutex);
-            verbose(3, "Getting wakeup from client");
-
-            // Lazily move active_units to stale_units to avoid race conditions around active_units with compaction
-            if (pending_stale_p) {
-                pending_stale_p = false;
-                struct rb_mjit_unit *next;
-                ccan_list_for_each_safe(&active_units.head, unit, next, unode) {
-                    if (unit->stale_p) {
-                        unit->stale_p = false;
-                        remove_from_list(unit, &active_units);
-                        add_to_list(unit, &stale_units);
-                        // Lazily put it to unit_queue as well to avoid race conditions on jit_unit with mjit_compile.
-                        mjit_add_iseq_to_process(unit->iseq, &ISEQ_BODY(unit->iseq)->jit_unit->compile_info, true);
-                    }
-                }
-            }
-
-            // Unload some units as needed
-            if (unload_requests >= throttle_threshold) {
-                while (in_gc) {
-                    verbose(3, "Waiting wakeup from GC");
-                    rb_native_cond_wait(&mjit_gc_wakeup, &mjit_engine_mutex);
-                }
-                in_jit = true; // Lock GC
-
-                RB_DEBUG_COUNTER_INC(mjit_unload_units);
-                unload_units();
-                unload_requests = 0;
-
-                in_jit = false; // Unlock GC
-                verbose(3, "Sending wakeup signal to client in a mjit-worker for GC");
-                rb_native_cond_signal(&mjit_client_wakeup);
-            }
-            if (active_units.length == mjit_opts.max_cache_size && mjit_opts.wait) { // Sometimes all methods may be in use
-                mjit_opts.max_cache_size++; // avoid infinite loop on `rb_mjit_wait_call`. Note that --jit-wait is just for testing.
-                verbose(1, "No units can be unloaded -- incremented max-cache-size to %d for --jit-wait", mjit_opts.max_cache_size);
-            }
-        }
-        unit = get_from_list(&unit_queue);
-        CRITICAL_SECTION_FINISH(3, "in worker dequeue");
-
-        if (unit) {
-            // JIT compile
-            mjit_func_t func = convert_unit_to_func(unit);
-            (void)RB_DEBUG_COUNTER_INC_IF(mjit_compile_failures, func == (mjit_func_t)NOT_COMPILED_JIT_ISEQ_FUNC);
-
-            CRITICAL_SECTION_START(3, "in jit func replace");
-            while (in_gc) { // Make sure we're not GC-ing when touching ISeq
-                verbose(3, "Waiting wakeup from GC");
-                rb_native_cond_wait(&mjit_gc_wakeup, &mjit_engine_mutex);
-            }
-            if (unit->iseq) { // Check whether GCed or not
-                if ((uintptr_t)func > (uintptr_t)LAST_JIT_ISEQ_FUNC) {
-                    add_to_list(unit, &active_units);
-                }
-                // Usage of jit_code might be not in a critical section.
-                MJIT_ATOMIC_SET(ISEQ_BODY(unit->iseq)->jit_func, func);
-            }
-            else {
-                free_unit(unit);
-            }
-            CRITICAL_SECTION_FINISH(3, "in jit func replace");
-
-#if USE_JIT_COMPACTION
-            // Combine .o files to one .so and reload all jit_func to improve memory locality.
-            if (compact_units.length < max_compact_size
-                && ((!mjit_opts.wait && unit_queue.length == 0 && active_units.length > 1)
-                    || (active_units.length == mjit_opts.max_cache_size && compact_units.length * throttle_threshold <= total_unloads))) { // throttle compaction by total_unloads
-                compact_all_jit_code();
-            }
-#endif
-        }
-    }
-
-    // To keep mutex unlocked when it is destroyed by mjit_finish, don't wrap CRITICAL_SECTION here.
-    worker_stopped = true;
-}
diff --git a/process.c b/process.c
index 6de5833c4b..9568399f93 100644
--- a/process.c
+++ b/process.c
@@ -1091,6 +1091,15 @@ void rb_sigwait_sleep(const rb_thread_t *, int fd, const rb_hrtime_t *);
 void rb_sigwait_fd_put(const rb_thread_t *, int fd);
 void rb_thread_sleep_interruptible(void);
 
+#if USE_MJIT
+static struct waitpid_state mjit_waitpid_state;
+
+// variables shared with thread.c
+// TODO: implement the same thing with postponed_job and obviate these variables
+bool mjit_waitpid_finished = false;
+int mjit_waitpid_status = 0;
+#endif
+
 static int
 waitpid_signal(struct waitpid_state *w)
 {
@@ -1098,12 +1107,13 @@ waitpid_signal(struct waitpid_state *w)
         rb_threadptr_interrupt(rb_ec_thread_ptr(w->ec));
         return TRUE;
     }
-    else { /* ruby_waitpid_locked */
-        if (w->cond) {
-            rb_native_cond_signal(w->cond);
-            return TRUE;
-        }
+#if USE_MJIT
+    else if (w == &mjit_waitpid_state && w->ret) { /* mjit_add_waiting_pid */
+        mjit_waitpid_finished = true;
+        mjit_waitpid_status = w->status;
+        return TRUE;
     }
+#endif
     return FALSE;
 }
 
@@ -1199,68 +1209,18 @@ waitpid_state_init(struct waitpid_state *w, rb_pid_t pid, int options)
     w->status = 0;
 }
 
-static const rb_hrtime_t *
-sigwait_sleep_time(void)
-{
-    if (SIGCHLD_LOSSY) {
-        static const rb_hrtime_t busy_wait = 100 * RB_HRTIME_PER_MSEC;
-
-        return &busy_wait;
-    }
-    return 0;
-}
-
+#if USE_MJIT
 /*
  * must be called with vm->waitpid_lock held, this is not interruptible
  */
-rb_pid_t
-ruby_waitpid_locked(rb_vm_t *vm, rb_pid_t pid, int *status, int options,
-                    rb_nativethread_cond_t *cond)
+void
+mjit_add_waiting_pid(rb_vm_t *vm, rb_pid_t pid)
 {
-    struct waitpid_state w;
-
-    assert(!ruby_thread_has_gvl_p() && "must not have GVL");
-
-    waitpid_state_init(&w, pid, options);
-    if (w.pid > 0 || ccan_list_empty(&vm->waiting_pids))
-        w.ret = do_waitpid(w.pid, &w.status, w.options | WNOHANG);
-    if (w.ret) {
-        if (w.ret == -1) w.errnum = errno;
-    }
-    else {
-        int sigwait_fd = -1;
-
-        w.ec = 0;
-        ccan_list_add(w.pid > 0 ? &vm->waiting_pids : &vm->waiting_grps, &w.wnode);
-        do {
-            if (sigwait_fd < 0)
-                sigwait_fd = rb_sigwait_fd_get(0);
-
-            if (sigwait_fd >= 0) {
-                w.cond = 0;
-                rb_native_mutex_unlock(&vm->waitpid_lock);
-                rb_sigwait_sleep(0, sigwait_fd, sigwait_sleep_time());
-                rb_native_mutex_lock(&vm->waitpid_lock);
-            }
-            else {
-                w.cond = cond;
-                rb_native_cond_wait(w.cond, &vm->waitpid_lock);
-            }
-        } while (!w.ret);
-        ccan_list_del(&w.wnode);
-
-        /* we're done, maybe other waitpid callers are not: */
-        if (sigwait_fd >= 0) {
-            rb_sigwait_fd_put(0, sigwait_fd);
-            sigwait_fd_migrate_sleeper(vm);
-        }
-    }
-    if (status) {
-        *status = w.status;
-    }
-    if (w.ret == -1) errno = w.errnum;
-    return w.ret;
+    waitpid_state_init(&mjit_waitpid_state, pid, 0);
+    mjit_waitpid_state.ec = 0; // switch the behavior of waitpid_signal
+    ccan_list_add(&vm->waiting_pids, &mjit_waitpid_state.wnode);
 }
+#endif
 
 static VALUE
 waitpid_sleep(VALUE x)
diff --git a/test/ruby/test_rubyvm_mjit.rb b/test/ruby/test_rubyvm_mjit.rb
index 92733a2b15..8ca0fb9ef2 100644
--- a/test/ruby/test_rubyvm_mjit.rb
+++ b/test/ruby/test_rubyvm_mjit.rb
@@ -87,7 +87,11 @@ class TestRubyVMMJIT < Test::Unit::TestCase
       print RubyVM::MJIT.pause(wait: false)
     EOS
     assert_equal('truefalse', out)
-    assert_equal(true, err.scan(/#{JITSupport::JIT_SUCCESS_PREFIX}/).size < 10)
+    if RUBY_PLATFORM.match?(/mswin|mingw/) # MJIT synchronously compiles methods on Windows
+      assert_equal(10, err.scan(/#{JITSupport::JIT_SUCCESS_PREFIX}/).size)
+    else
+      assert_equal(true, err.scan(/#{JITSupport::JIT_SUCCESS_PREFIX}/).size < 10)
+    end
   end
 
   def test_resume
diff --git a/thread.c b/thread.c
index 5e659a643f..0b5cbb8197 100644
--- a/thread.c
+++ b/thread.c
@@ -2233,6 +2233,12 @@ threadptr_get_interrupts(rb_thread_t *th)
     return interrupt & (rb_atomic_t)~ec->interrupt_mask;
 }
 
+#if USE_MJIT
+// process.c
+extern bool mjit_waitpid_finished;
+extern int mjit_waitpid_status;
+#endif
+
 MJIT_FUNC_EXPORTED int
 rb_threadptr_execute_interrupts(rb_thread_t *th, int blocking_timing)
 {
@@ -2280,6 +2286,15 @@ rb_threadptr_execute_interrupts(rb_thread_t *th, int blocking_timing)
 		ret |= rb_signal_exec(th, sig);
 	    }
 	    th->status = prev_status;
+
+#if USE_MJIT
+            // Handle waitpid_signal for MJIT issued by ruby_sigchld_handler. This needs to be done
+            // outside ruby_sigchld_handler to avoid recursively relying on the SIGCHLD handler.
+            if (mjit_waitpid_finished) {
+                mjit_waitpid_finished = false;
+                mjit_notify_waitpid(mjit_waitpid_status);
+            }
+#endif
 	}
 
 	/* exception from another thread */