PR12579 (OS-based Synchronisation for Stop-the-World Sections): upstream 51491dd

(cherry picked from commit 20bd57a)

Author: mshinwell

ocaml/configure.ac

@@ -1146,6 +1146,10 @@ AC_CHECK_HEADER([stdatomic.h], [AC_DEFINE([HAS_STDATOMIC_H])])
 
 AC_CHECK_HEADER([sys/mman.h], [AC_DEFINE([HAS_SYS_MMAN_H])])
 
+AS_CASE([$host],
+  [*-*-linux*],
+    [AC_CHECK_HEADER([linux/futex.h], [AC_DEFINE([HAS_LINUX_FUTEX_H])])])
+
 # Checks for types
 
 ## off_t
@@ -2766,11 +2770,12 @@ ocamlc_cppflags="$common_cppflags $CPPFLAGS"
 
 AS_CASE([$host],
   [*-*-mingw32*],
-    [cclibs="$cclibs -lole32 -luuid -lversion"],
+    [cclibs="$cclibs -lole32 -luuid -lversion -lshlwapi -lsynchronization"],
   [*-pc-windows],
-    [# For whatever reason, flexlink includes -ladvapi32 for mingw-w64, but
-    # doesn't include advapi32.lib for MSVC
-    cclibs="$cclibs ole32.lib uuid.lib advapi32.lib version.lib"])
+    [# For whatever reason, flexlink includes -ladvapi32 and -lshell32 for
+    # mingw-w64, but doesn't include advapi32.lib and shell32.lib for MSVC
+    cclibs="$cclibs ole32.lib uuid.lib advapi32.lib shell32.lib version.lib \
+shlwapi.lib synchronization.lib"])
 
 AC_CONFIG_COMMANDS_PRE([cclibs="$cclibs $mathlib $DLLIBS $PTHREAD_LIBS"])
 

ocaml/runtime/caml/domain.h

@@ -101,12 +101,16 @@ Caml_inline intnat caml_domain_alone(void)
 int caml_domain_is_in_stw(void);
 #endif
 
+int caml_domain_terminating(caml_domain_state *);
+int caml_domain_is_terminating(void);
+
 int caml_try_run_on_all_domains_with_spin_work(
   int sync,
   void (*handler)(caml_domain_state*, void*, int, caml_domain_state**),
   void* data,
   void (*leader_setup)(caml_domain_state*),
-  void (*enter_spin_callback)(caml_domain_state*, void*),
+  /* return nonzero if there may still be useful work to do while spinning */
+  int (*enter_spin_callback)(caml_domain_state*, void*),
   void* enter_spin_data);
 int caml_try_run_on_all_domains(
   void (*handler)(caml_domain_state*, void*, int, caml_domain_state**),
@@ -167,16 +171,54 @@ int caml_try_run_on_all_domains(
 */
 
 
-/* barriers */
-typedef uintnat barrier_status;
-void caml_global_barrier(void);
-barrier_status caml_global_barrier_begin(void);
-int caml_global_barrier_is_final(barrier_status);
-void caml_global_barrier_end(barrier_status);
-int caml_global_barrier_num_domains(void);
+/* Barriers */
 
-int caml_domain_terminating(caml_domain_state *);
-int caml_domain_is_terminating(void);
+/* Get the number of parties expected to arrive into the barrier, i.e. the
+   number of domains participating in the STW section. In most cases the barrier
+   is used directly from an STW callback that already has the number of
+   participating domains at hand, which should be used instead. */
+int caml_global_barrier_num_participating(void);
+
+/* Unconditionally arrive at the barrier and wait for all parties,
+   [caml_global_barrier] below should be used instead. */
+void caml_enter_global_barrier(int num_participating);
+/* Arrive at the barrier and wait iff there is more than one party */
+Caml_inline void caml_global_barrier(int num_participating) {
+  if (num_participating != 1) caml_enter_global_barrier(num_participating);
+}
+
+typedef uintnat barrier_status;
+/* Arrive at the barrier; if we are the final party, immediately returns a
+   nonzero value to be passed to [caml_global_barrier_release_as_final]
+   later, otherwise blocks and returns zero. */
+barrier_status caml_global_barrier_and_check_final(int num_participating);
+/* Release the barrier with the given status */
+void caml_global_barrier_release_as_final(barrier_status status);
+/* Arrive at the global barrier and run the body if we are the final party.
+   Other threads will not be released from the barrier until the final party
+   finishes executing the body.
+
+   Example usage:
+
+   Caml_global_barrier_if_final(num_participating) {
+     do_something_in_final_domain();
+   }
+
+   Note: this expands to an [if] and [for] header, do not exit the body using
+   jumps or returns, and do not put an [else] immediately after.
+ */
+#define Caml_global_barrier_if_final(num_participating)                 \
+  /* fast path when alone */                                            \
+  int CAML_GENSYM(alone) = (num_participating) == 1;                    \
+  barrier_status CAML_GENSYM(b) = 0;                                    \
+  if (CAML_GENSYM(alone) ||                                             \
+      (CAML_GENSYM(b)                                                   \
+       = caml_global_barrier_and_check_final(num_participating)))       \
+    for (int CAML_GENSYM(continue) = 1; CAML_GENSYM(continue);          \
+         /* release the barrier after the body has executed once */     \
+         ((CAML_GENSYM(alone) ? (void)0 :                               \
+           caml_global_barrier_release_as_final(CAML_GENSYM(b))),       \
+          CAML_GENSYM(continue) = 0))
 
 #endif /* CAML_INTERNALS */
 

ocaml/runtime/caml/major_gc.h

@@ -31,7 +31,7 @@ Caml_inline int caml_marking_started(void) {
   return caml_gc_phase != Phase_sweep_main;
 }
 
-intnat caml_opportunistic_major_work_available (void);
+intnat caml_opportunistic_major_work_available (caml_domain_state*);
 void caml_opportunistic_major_collection_slice (intnat);
 /* auto-triggered slice from within the GC */
 #define AUTO_TRIGGERED_MAJOR_SLICE -1

ocaml/runtime/caml/misc.h

@@ -595,6 +595,11 @@ CAMLextern int caml_snwprintf(wchar_t * buf,
 #  endif
 #endif
 
+/* Generate a named symbol that is unique within the current macro expansion */
+#define CAML_GENSYM_3(name, l) caml__##name##_##l
+#define CAML_GENSYM_2(name, l) CAML_GENSYM_3(name, l)
+#define CAML_GENSYM(name) CAML_GENSYM_2(name, __LINE__)
+
 #endif /* CAML_INTERNALS */
 
 /* The [backtrace_slot] type represents values stored in