WIP: threading support

compileopts/options.go

@@ -10,7 +10,7 @@ import (
 var (
 	validBuildModeOptions     = []string{"default", "c-shared"}
 	validGCOptions            = []string{"none", "leaking", "conservative", "custom", "precise"}
-	validSchedulerOptions     = []string{"none", "tasks", "asyncify"}
+	validSchedulerOptions     = []string{"none", "tasks", "asyncify", "threads"}
 	validSerialOptions        = []string{"none", "uart", "usb", "rtt"}
 	validPrintSizeOptions     = []string{"none", "short", "full"}
 	validPanicStrategyOptions = []string{"print", "trap"}

compileopts/options_test.go

@@ -10,7 +10,7 @@ import (
 func TestVerifyOptions(t *testing.T) {
 
 	expectedGCError := errors.New(`invalid gc option 'incorrect': valid values are none, leaking, conservative, custom, precise`)
-	expectedSchedulerError := errors.New(`invalid scheduler option 'incorrect': valid values are none, tasks, asyncify`)
+	expectedSchedulerError := errors.New(`invalid scheduler option 'incorrect': valid values are none, tasks, asyncify, threads`)
 	expectedPrintSizeError := errors.New(`invalid size option 'incorrect': valid values are none, short, full`)
 	expectedPanicStrategyError := errors.New(`invalid panic option 'incorrect': valid values are print, trap`)
 

compileopts/target.go

@@ -247,7 +247,6 @@ func defaultTarget(options *Options) (*TargetSpec, error) {
 		GOARCH:           options.GOARCH,
 		BuildTags:        []string{options.GOOS, options.GOARCH},
 		GC:               "precise",
-		Scheduler:        "tasks",
 		Linker:           "cc",
 		DefaultStackSize: 1024 * 64, // 64kB
 		GDB:              []string{"gdb"},
@@ -378,6 +377,7 @@ func defaultTarget(options *Options) (*TargetSpec, error) {
 			platformVersion = "11.0.0" // first macosx platform with arm64 support
 		}
 		llvmvendor = "apple"
+		spec.Scheduler = "tasks"
 		spec.Linker = "ld.lld"
 		spec.Libc = "darwin-libSystem"
 		// Use macosx* instead of darwin, otherwise darwin/arm64 will refer to
@@ -395,6 +395,7 @@ func defaultTarget(options *Options) (*TargetSpec, error) {
 			"src/runtime/runtime_unix.c",
 			"src/runtime/signal.c")
 	case "linux":
+		spec.Scheduler = "threads"
 		spec.Linker = "ld.lld"
 		spec.RTLib = "compiler-rt"
 		spec.Libc = "musl"
@@ -415,9 +416,11 @@ func defaultTarget(options *Options) (*TargetSpec, error) {
 		}
 		spec.ExtraFiles = append(spec.ExtraFiles,
 			"src/internal/futex/futex_linux.c",
+			"src/internal/task/task_threads.c",
 			"src/runtime/runtime_unix.c",
 			"src/runtime/signal.c")
 	case "windows":
+		spec.Scheduler = "tasks"
 		spec.Linker = "ld.lld"
 		spec.Libc = "mingw-w64"
 		// Note: using a medium code model, low image base and no ASLR

src/internal/task/atomic-cooperative.go

@@ -1,3 +1,5 @@
+//go:build !scheduler.threads
+
 package task
 
 // Atomics implementation for cooperative systems. The atomic types here aren't

src/internal/task/atomic-preemptive.go

@@ -0,0 +1,14 @@
+//go:build scheduler.threads
+
+package task
+
+// Atomics implementation for non-cooperative systems (multithreaded, etc).
+// These atomic types use real atomic instructions.
+
+import "sync/atomic"
+
+type (
+	Uintptr = atomic.Uintptr
+	Uint32  = atomic.Uint32
+	Uint64  = atomic.Uint64
+)

src/internal/task/futex-cooperative.go

@@ -1,3 +1,5 @@
+//go:build !scheduler.threads
+
 package task
 
 // A futex is a way for userspace to wait with the pointer as the key, and for

src/internal/task/futex-preemptive.go

@@ -0,0 +1,7 @@
+//go:build scheduler.threads
+
+package task
+
+import "internal/futex"
+
+type Futex = futex.Futex

src/internal/task/linux.go

@@ -0,0 +1,9 @@
+//go:build linux && !baremetal
+
+package task
+
+import "unsafe"
+
+// Musl uses a pointer (or unsigned long for C++) so unsafe.Pointer should be
+// fine.
+type threadID unsafe.Pointer

src/internal/task/mutex.go

@@ -0,0 +1,69 @@
+package task
+
+// Futex-based mutex.
+// This is largely based on the paper "Futexes are Tricky" by Ulrich Drepper.
+// It describes a few ways to implement mutexes using a futex, and how some
+// seemingly-obvious implementations don't exactly work as intended.
+// Unfortunately, Go atomic operations work slightly differently so we can't
+// copy the algorithm verbatim.
+//
+// The implementation works like this. The futex can have 3 different values,
+// depending on the state:
+//
+//   - 0: the futex is currently unlocked.
+//   - 1: the futex is locked, but is uncontended. There is one special case: if
+//     a contended futex is unlocked, it is set to 0. It is possible for another
+//     thread to lock the futex before the next waiter is woken. But because a
+//     waiter will be woken (if there is one), it will always change to 2
+//     regardless. So this is not a problem.
+//   - 2: the futex is locked, and is contended. At least one thread is trying
+//     to obtain the lock (and is in the contended loop, see below).
+//
+// For the paper, see:
+// https://dept-info.labri.fr/~denis/Enseignement/2008-IR/Articles/01-futex.pdf)
+
+type Mutex struct {
+	futex Futex
+}
+
+func (m *Mutex) Lock() {
+	// Fast path: try to take an uncontended lock.
+	if m.futex.CompareAndSwap(0, 1) {
+		// We obtained the mutex.
+		return
+	}
+
+	// The futex is contended, so we enter the contended loop.
+	// If we manage to change the futex from 0 to 2, we managed to take the
+	// look. Else, we have to wait until a call to Unlock unlocks this mutex.
+	// (Unlock will wake one waiter when it finds the futex is set to 2 when
+	// unlocking).
+	for m.futex.Swap(2) != 0 {
+		// Wait until we get resumed in Unlock.
+		m.futex.Wait(2)
+	}
+}
+
+func (m *Mutex) Unlock() {
+	if old := m.futex.Swap(0); old == 0 {
+		// Mutex wasn't locked before.
+		panic("sync: unlock of unlocked Mutex")
+	} else if old == 2 {
+		// Mutex was a contended lock, so we need to wake the next waiter.
+		m.futex.Wake()
+	}
+}
+
+// TryLock tries to lock m and reports whether it succeeded.
+//
+// Note that while correct uses of TryLock do exist, they are rare,
+// and use of TryLock is often a sign of a deeper problem
+// in a particular use of mutexes.
+func (m *Mutex) TryLock() bool {
+	// Fast path: try to take an uncontended lock.
+	if m.futex.CompareAndSwap(0, 1) {
+		// We obtained the mutex.
+		return true
+	}
+	return false
+}

src/internal/task/pmutex-cooperative.go

@@ -1,3 +1,5 @@
+//go:build !scheduler.threads
+
 package task
 
 // PMutex is a real mutex on systems that can be either preemptive or threaded,

src/internal/task/pmutex-preemptive.go

@@ -0,0 +1,11 @@
+//go:build scheduler.threads
+
+package task
+
+// PMutex is a real mutex on systems that can be either preemptive or threaded,
+// and a dummy lock on other (purely cooperative) systems.
+//
+// It is mainly useful for short operations that need a lock when threading may
+// be involved, but which do not need a lock with a purely cooperative
+// scheduler.
+type PMutex = Mutex

src/internal/task/semaphore.go

@@ -0,0 +1,32 @@
+package task
+
+// Barebones semaphore implementation.
+// The main limitation is that if there are multiple waiters, a single Post()
+// call won't do anything. Only when Post() has been called to awaken all
+// waiters will the waiters proceed.
+// This limitation is not a problem when there will only be a single waiter.
+type Semaphore struct {
+	futex Futex
+}
+
+// Post (unlock) the semaphore, incrementing the value in the semaphore.
+func (s *Semaphore) Post() {
+	newValue := s.futex.Add(1)
+	if newValue == 0 {
+		s.futex.WakeAll()
+	}
+}
+
+// Wait (lock) the semaphore, decrementing the value in the semaphore.
+func (s *Semaphore) Wait() {
+	delta := int32(-1)
+	value := s.futex.Add(uint32(delta))
+	for {
+		if int32(value) >= 0 {
+			// Semaphore unlocked!
+			return
+		}
+		s.futex.Wait(value)
+		value = s.futex.Load()
+	}
+}

src/internal/task/task.go

@@ -27,7 +27,7 @@ type Task struct {
 }
 
 // DataUint32 returns the Data field as a uint32. The value is only valid after
-// setting it through SetDataUint32.
+// setting it through SetDataUint32 or by storing to it using DataAtomicUint32.
 func (t *Task) DataUint32() uint32 {
 	return *(*uint32)(unsafe.Pointer(&t.Data))
 }
@@ -38,6 +38,11 @@ func (t *Task) SetDataUint32(val uint32) {
 	*(*uint32)(unsafe.Pointer(&t.Data)) = val
 }
 
+// DataAtomicUint32 returns the Data field as an atomic-if-needed Uint32 value.
+func (t *Task) DataAtomicUint32() *Uint32 {
+	return (*Uint32)(unsafe.Pointer(&t.Data))
+}
+
 // getGoroutineStackSize is a compiler intrinsic that returns the stack size for
 // the given function and falls back to the default stack size. It is replaced
 // with a load from a special section just before codegen.

src/internal/task/task_threads.c

@@ -0,0 +1,104 @@
+//go:build none
+
+#define _GNU_SOURCE
+#include <pthread.h>
+#include <semaphore.h>
+#include <signal.h>
+#include <stdint.h>
+#include <stdio.h>
+
+// BDWGC also uses SIGRTMIN+6 on Linux, which seems like a reasonable choice.
+#ifdef __linux__
+#define taskPauseSignal (SIGRTMIN + 6)
+#endif
+
+// Pointer to the current task.Task structure.
+// Ideally the entire task.Task structure would be a thread-local variable but
+// this also works.
+static __thread void *current_task;
+
+struct state_pass {
+    void      *(*start)(void*);
+    void      *args;
+    void      *task;
+    uintptr_t *stackTop;
+    sem_t     startlock;
+};
+
+// Handle the GC pause in Go.
+void tinygo_task_gc_pause(int sig);
+
+// Initialize the main thread.
+void tinygo_task_init(void *mainTask, pthread_t *thread, void *context) {
+    // Make sure the current task pointer is set correctly for the main
+    // goroutine as well.
+    current_task = mainTask;
+
+    // Store the thread ID of the main thread.
+    *thread = pthread_self();
+
+    // Register the "GC pause" signal for the entire process.
+    // Using pthread_kill, we can still send the signal to a specific thread.
+    struct sigaction act = { 0 };
+    act.sa_flags = SA_SIGINFO;
+    act.sa_handler = &tinygo_task_gc_pause;
+    sigaction(taskPauseSignal, &act, NULL);
+}
+
+void tinygo_task_exited(void*);
+
+// Helper to start a goroutine while also storing the 'task' structure.
+static void* start_wrapper(void *arg) {
+    struct state_pass *state = arg;
+    void *(*start)(void*) = state->start;
+    void *args = state->args;
+    current_task = state->task;
+
+    // Save the current stack pointer in the goroutine state, for the GC.
+    int stackAddr;
+    *(state->stackTop) = (uintptr_t)(&stackAddr);
+
+    // Notify the caller that the thread has successfully started and
+    // initialized.
+    sem_post(&state->startlock);
+
+    // Run the goroutine function.
+    start(args);
+
+    // Notify the Go side this thread will exit.
+    tinygo_task_exited(current_task);
+
+    return NULL;
+};
+
+// Start a new goroutine in an OS thread.
+int tinygo_task_start(uintptr_t fn, void *args, void *task, pthread_t *thread, uintptr_t *stackTop, void *context) {
+    // Sanity check. Should get optimized away.
+    if (sizeof(pthread_t) != sizeof(void*)) {
+        __builtin_trap();
+    }
+
+    struct state_pass state = {
+        .start     = (void*)fn,
+        .args      = args,
+        .task      = task,
+        .stackTop  = stackTop,
+    };
+    sem_init(&state.startlock, 0, 0);
+    int result = pthread_create(thread, NULL, &start_wrapper, &state);
+
+    // Wait until the thread has been crated and read all state_pass variables.
+    sem_wait(&state.startlock);
+
+    return result;
+}
+
+// Return the current task (for task.Current()).
+void* tinygo_task_current(void) {
+    return current_task;
+}
+
+// Send a signal to cause the task to pause for the GC mark phase.
+void tinygo_task_send_gc_signal(pthread_t thread) {
+    pthread_kill(thread, taskPauseSignal);
+}