vm: sampling profiler timer for windows

Spawn a thread and have it spin on the performance counter, triggering safepoints on the main thread every sample time

vm/errors.cpp

@@ -170,18 +170,18 @@ void factor_vm::enqueue_safepoint_fep()
 	code->guard_safepoint();
 }
 
-void factor_vm::enqueue_safepoint_sample(cell pc, bool foreign_thread_p)
+void factor_vm::enqueue_safepoint_sample(cell samples, cell pc, bool foreign_thread_p)
 {
-	if (sampling_profiler_p)
+	if (FACTOR_MEMORY_BARRIER(), sampling_profiler_p)
 	{
-		FACTOR_ATOMIC_ADD(&safepoint_sample_count, 1);
+		FACTOR_ATOMIC_ADD(&safepoint_sample_count, samples);
 		if (foreign_thread_p)
-			FACTOR_ATOMIC_ADD(&safepoint_foreign_thread_sample_count, 1);
+			FACTOR_ATOMIC_ADD(&safepoint_foreign_thread_sample_count, samples);
 		else {
-			if (current_gc)
-				FACTOR_ATOMIC_ADD(&safepoint_gc_sample_count, 1);
-			if (!code->seg->in_segment_p(pc))
-				FACTOR_ATOMIC_ADD(&safepoint_foreign_sample_count, 1);
+			if (FACTOR_MEMORY_BARRIER(), current_gc)
+				FACTOR_ATOMIC_ADD(&safepoint_gc_sample_count, samples);
+			if (pc != 0 && !code->seg->in_segment_p(pc))
+				FACTOR_ATOMIC_ADD(&safepoint_foreign_sample_count, samples);
 		}
 		code->guard_safepoint();
 	}

vm/os-unix.cpp

@@ -209,10 +209,10 @@ void sample_signal_handler(int signal, siginfo_t *siginfo, void *uap)
 {
 	factor_vm *vm = current_vm_p();
 	if (vm)
-		vm->enqueue_safepoint_sample((cell)UAP_PROGRAM_COUNTER(uap), false);
+		vm->enqueue_safepoint_sample(1, (cell)UAP_PROGRAM_COUNTER(uap), false);
 	else if (thread_vms.size() == 1) {
 		factor_vm *the_only_vm = thread_vms.begin()->second;
-		the_only_vm->enqueue_safepoint_sample((cell)UAP_PROGRAM_COUNTER(uap), true);
+		the_only_vm->enqueue_safepoint_sample(1, (cell)UAP_PROGRAM_COUNTER(uap), true);
 	}
 }
 

vm/os-windows.cpp

@@ -301,14 +301,54 @@ void factor_vm::open_console()
 	SetConsoleCtrlHandler(factor::ctrl_handler, TRUE);
 }
 
+void factor_vm::sampler_thread_loop()
+{
+	LARGE_INTEGER counter, new_counter, units_per_second;
+
+	assert(QueryPerformanceFrequency(&units_per_second));
+	long long units_per_sample =
+		units_per_second.QuadPart / FACTOR_PROFILE_SAMPLES_PER_SECOND;
+
+	assert(QueryPerformanceCounter(&counter));
+	while (FACTOR_MEMORY_BARRIER(), sampling_profiler_p)
+	{
+		Sleep(0);
+		assert(QueryPerformanceCounter(&new_counter));
+		cell samples = 0;
+		while (new_counter.QuadPart - counter.QuadPart > units_per_sample) {
+			// We would have to suspend the thread to sample the PC
+			++samples;
+			counter.QuadPart += units_per_sample;
+		}
+		enqueue_safepoint_sample(samples, 0, false);
+	}
+}
+
+static DWORD WINAPI sampler_thread_entry(LPVOID parent_vm)
+{
+	static_cast<factor_vm*>(parent_vm)->sampler_thread_loop();
+	return 0;
+}
+
 void factor_vm::start_sampling_profiler_timer()
 {
-	general_error(ERROR_NOT_IMPLEMENTED, false_object, false_object);
+	sampler_thread = CreateThread(
+		NULL,
+		0,
+		&sampler_thread_entry,
+		static_cast<LPVOID>(this),
+		0,
+		NULL
+	);
 }
 
 void factor_vm::end_sampling_profiler_timer()
 {
-	general_error(ERROR_NOT_IMPLEMENTED, false_object, false_object);
+	DWORD wait_result = WaitForSingleObject(sampler_thread,
+		3000/FACTOR_PROFILE_SAMPLES_PER_SECOND);
+	if (wait_result != WAIT_OBJECT_0)
+		TerminateThread(sampler_thread);
+	sampler_thread = NULL;
 }
 
 }

vm/sampling_profiler.cpp

@@ -101,9 +101,10 @@ void factor_vm::start_sampling_profiler()
 
 void factor_vm::end_sampling_profiler()
 {
+	sampling_profiler_p = false;
+	FACTOR_MEMORY_BARRIER();
 	end_sampling_profiler_timer();
 	record_sample();
-	sampling_profiler_p = false;
 }
 
 void factor_vm::primitive_sampling_profiler()

vm/vm.hpp

@@ -62,7 +62,7 @@ struct factor_vm
 
 	/* Is profiling enabled? */
 	bool counting_profiler_p;
-	bool sampling_profiler_p;
+	volatile bool sampling_profiler_p;
 
 	/* Global variables used to pass fault handler state from signal handler
 	to VM */
@@ -93,7 +93,7 @@ struct factor_vm
 	callback_heap *callbacks;
 
 	/* Only set if we're performing a GC */
-	gc_state *current_gc;
+	volatile gc_state *current_gc;
 
 	/* Mark stack */
 	std::vector<cell> mark_stack;
@@ -216,7 +216,7 @@ struct factor_vm
 	void synchronous_signal_handler_impl();
 	void fp_signal_handler_impl();
 	void enqueue_safepoint_fep();
-	void enqueue_safepoint_sample(cell pc, bool foreign_thread_p);
+	void enqueue_safepoint_sample(cell samples, cell pc, bool foreign_thread_p);
 	void handle_safepoint();
 
 	// bignum
@@ -727,6 +727,9 @@ struct factor_vm
 
 	// os-windows
   #if defined(WINDOWS)
+	HANDLE sampler_thread;
+	void sampler_thread_loop();
+
 	const vm_char *vm_executable_path();
 	const vm_char *default_image_path();
 	void windows_image_path(vm_char *full_path, vm_char *temp_path, unsigned int length);