Merge branch 'master' into file-tags-without-comments

4 files changed, 18 insertions, 33 deletions

diff --git a/core/thread/thread.odin b/core/thread/thread.odin
index 17ba1a0a2..c1cbceb42 100644
--- a/core/thread/thread.odin
+++ b/core/thread/thread.odin
@@ -272,7 +272,7 @@ create_and_start :: proc(fn: proc(), init_context: Maybe(runtime.Context) = nil,
 	t := create(thread_proc, priority)
 	t.data = rawptr(fn)
 	if self_cleanup {
-		t.flags += {.Self_Cleanup}
+		intrinsics.atomic_or(&t.flags, {.Self_Cleanup})
 	}
 	t.init_context = init_context
 	start(t)
@@ -307,7 +307,7 @@ create_and_start_with_data :: proc(data: rawptr, fn: proc(data: rawptr), init_co
 	t.user_index = 1
 	t.user_args[0] = data
 	if self_cleanup {
-		t.flags += {.Self_Cleanup}
+		intrinsics.atomic_or(&t.flags, {.Self_Cleanup})
 	}
 	t.init_context = init_context
 	start(t)
@@ -347,7 +347,7 @@ create_and_start_with_poly_data :: proc(data: $T, fn: proc(data: T), init_contex
 	mem.copy(&t.user_args[0], &data, size_of(T))
 
 	if self_cleanup {
-		t.flags += {.Self_Cleanup}
+		intrinsics.atomic_or(&t.flags, {.Self_Cleanup})
 	}
 
 	t.init_context = init_context
@@ -394,7 +394,7 @@ create_and_start_with_poly_data2 :: proc(arg1: $T1, arg2: $T2, fn: proc(T1, T2),
 	_  = copy(user_args[n:], mem.ptr_to_bytes(&arg2))
 
 	if self_cleanup {
-		t.flags += {.Self_Cleanup}
+		intrinsics.atomic_or(&t.flags, {.Self_Cleanup})
 	}
 
 	t.init_context = init_context
@@ -443,7 +443,7 @@ create_and_start_with_poly_data3 :: proc(arg1: $T1, arg2: $T2, arg3: $T3, fn: pr
 	_  = copy(user_args[n:], mem.ptr_to_bytes(&arg3))
 
 	if self_cleanup {
-		t.flags += {.Self_Cleanup}
+		intrinsics.atomic_or(&t.flags, {.Self_Cleanup})
 	}
 
 	t.init_context = init_context
@@ -494,7 +494,7 @@ create_and_start_with_poly_data4 :: proc(arg1: $T1, arg2: $T2, arg3: $T3, arg4:
 	_  = copy(user_args[n:], mem.ptr_to_bytes(&arg4))
 
 	if self_cleanup {
-		t.flags += {.Self_Cleanup}
+		intrinsics.atomic_or(&t.flags, {.Self_Cleanup})
 	}
 
 	t.init_context = init_context
diff --git a/core/thread/thread_pool.odin b/core/thread/thread_pool.odin
index 9bcc42968..d9166b450 100644
--- a/core/thread/thread_pool.odin
+++ b/core/thread/thread_pool.odin
@@ -60,6 +60,7 @@ pool_thread_runner :: proc(t: ^Thread) {
 		if task, ok := pool_pop_waiting(pool); ok {
 			data.task = task
 			pool_do_work(pool, task)
+			sync.guard(&pool.mutex)
 			data.task = {}
 		}
 	}
diff --git a/core/thread/thread_unix.odin b/core/thread/thread_unix.odin
index 2e196c3e6..9576a3040 100644
--- a/core/thread/thread_unix.odin
+++ b/core/thread/thread_unix.odin
@@ -5,18 +5,14 @@ package thread
 import "base:runtime"
 import "core:sync"
 import "core:sys/unix"
-import "core:time"
 
 _IS_SUPPORTED :: true
 
-CAS :: sync.atomic_compare_exchange_strong
-
 // NOTE(tetra): Aligned here because of core/unix/pthread_linux.odin/pthread_t.
 // Also see core/sys/darwin/mach_darwin.odin/semaphore_t.
 Thread_Os_Specific :: struct #align(16) {
 	unix_thread: unix.pthread_t, // NOTE: very large on Darwin, small on Linux.
-	cond:        sync.Cond,
-	mutex:       sync.Mutex,
+	start_ok:    sync.Sema,
 }
 //
 // Creates a thread which will run the given procedure.
@@ -29,14 +25,10 @@ _create :: proc(procedure: Thread_Proc, priority: Thread_Priority) -> ^Thread {
 		// We need to give the thread a moment to start up before we enable cancellation.
 		can_set_thread_cancel_state := unix.pthread_setcancelstate(unix.PTHREAD_CANCEL_ENABLE, nil) == 0
 
-		sync.lock(&t.mutex)
-
 		t.id = sync.current_thread_id()
 
-		for (.Started not_in sync.atomic_load(&t.flags)) {
-			// HACK: use a timeout so in the event that the condition is signalled at THIS comment's exact point
-			// (after checking flags, before starting the wait) it gets itself out of that deadlock after a ms.
-			sync.wait_with_timeout(&t.cond, &t.mutex, time.Millisecond)
+		if .Started not_in sync.atomic_load(&t.flags) {
+			sync.wait(&t.start_ok)
 		}
 
 		if .Joined in sync.atomic_load(&t.flags) {
@@ -66,8 +58,6 @@ _create :: proc(procedure: Thread_Proc, priority: Thread_Priority) -> ^Thread {
 
 		sync.atomic_or(&t.flags, { .Done })
 
-		sync.unlock(&t.mutex)
-
 		if .Self_Cleanup in sync.atomic_load(&t.flags) {
 			res := unix.pthread_detach(t.unix_thread)
 			assert_contextless(res == 0)
@@ -132,7 +122,7 @@ _create :: proc(procedure: Thread_Proc, priority: Thread_Priority) -> ^Thread {
 
 _start :: proc(t: ^Thread) {
 	sync.atomic_or(&t.flags, { .Started })
-	sync.signal(&t.cond)
+	sync.post(&t.start_ok)
 }
 
 _is_done :: proc(t: ^Thread) -> bool {
@@ -140,24 +130,18 @@ _is_done :: proc(t: ^Thread) -> bool {
 }
 
 _join :: proc(t: ^Thread) {
-	// sync.guard(&t.mutex)
-
 	if unix.pthread_equal(unix.pthread_self(), t.unix_thread) {
 		return
 	}
 
-	// Preserve other flags besides `.Joined`, like `.Started`.
-	unjoined := sync.atomic_load(&t.flags) - {.Joined}
-	joined   := unjoined + {.Joined}
-
-	// Try to set `t.flags` from unjoined to joined. If it returns joined,
-	// it means the previous value had that flag set and we can return.
-	if res, ok := CAS(&t.flags, unjoined, joined); res == joined && !ok {
+	// If the previous value was already `Joined`, then we can return.
+	if .Joined in sync.atomic_or(&t.flags, {.Joined}) {
 		return
 	}
+
 	// Prevent non-started threads from blocking main thread with initial wait
 	// condition.
-	if .Started not_in unjoined {
+	if .Started not_in sync.atomic_load(&t.flags) {
 		_start(t)
 	}
 	unix.pthread_join(t.unix_thread, nil)
diff --git a/core/thread/thread_windows.odin b/core/thread/thread_windows.odin
index 300673191..cc73a2d6a 100644
--- a/core/thread/thread_windows.odin
+++ b/core/thread/thread_windows.odin
@@ -27,7 +27,7 @@ _create :: proc(procedure: Thread_Proc, priority: Thread_Priority) -> ^Thread {
 	__windows_thread_entry_proc :: proc "system" (t_: rawptr) -> win32.DWORD {
 		t := (^Thread)(t_)
 
-		if .Joined in t.flags {
+		if .Joined in sync.atomic_load(&t.flags) {
 			return 0
 		}
 
@@ -48,9 +48,9 @@ _create :: proc(procedure: Thread_Proc, priority: Thread_Priority) -> ^Thread {
 			t.procedure(t)
 		}
 
-		intrinsics.atomic_store(&t.flags, t.flags + {.Done})
+		intrinsics.atomic_or(&t.flags, {.Done})
 
-		if .Self_Cleanup in t.flags {
+		if .Self_Cleanup in sync.atomic_load(&t.flags) {
 			win32.CloseHandle(t.win32_thread)
 			t.win32_thread = win32.INVALID_HANDLE
 			// NOTE(ftphikari): It doesn't matter which context 'free' received, right?