Threads: Context switches cost what they do on hardware

Timed on a PSP, each way of handing the CPU to another thread (the call
and the switch together):

                                     hardware   before   now
  rotate to an equal thread              7        14       7
  signal, better thread runs            10        17      10
  it waits again, back to caller        10        19      12
  wakeup, better thread runs             8        13       6
  it sleeps again, back to caller        7        12       6
  start a better thread, entry          30        28      30
  thread ends, back to its waiter       21        13      20
  notify, better thread's callback      14        13      14

A switch between two threads now costs 1150 cycles instead of 2700.
Starting a better thread costs 2000 cycles more, ending a thread 3300,
and setting up a callback 1800.

Also splits a wait timeout's ~30us into the deadline being taken 12us
into the call and the timeout going off 18us after it. That only changes
the time left written back, which threads/semaphores/wait and
threads/fpl/cancel pin between them. intr/vblank is re-recorded so it
no longer depends on the phase of the frame.

threads/callbacks/combos now passes.

Co-Authored-By: Claude Opus 5.5 (1M context) <[email protected]>

Core/HLE/sceKernelThread.cpp

@@ -1539,8 +1539,9 @@ static bool __KernelChance(int percent) {
 // wait fails with SCE_KERNEL_ERROR_WAIT_TIMEOUT at once, without giving up the CPU or writing the
 // timeout back. For most calls that happens about 85% of the time for 0us, half the time for 1us,
 // 15% for 2us and never from 3us on. sceKernelAllocateVpl does more first: always up to 1us, then
-// about 17% less per us. Longer ones end max(timeout, 205us) plus about 35us after the call, the
-// last few of which we already spend around the event.
+// about 17% less per us. Longer ones end max(timeout, 205us) plus about 35us after the call. The
+// deadline is taken late in the call, so the time left written back when something else ends the
+// wait counts from there.
 bool __KernelWaitTimesOutAtOnce(u32 timeoutPtr, int basePercent, int stepPercent) {
 	if (!Memory::IsValid4AlignedAddress(timeoutPtr)) {
 		return false;
@@ -1550,7 +1551,7 @@ bool __KernelWaitTimesOutAtOnce(u32 timeoutPtr, int basePercent, int stepPercent
 }
 
 s64 __KernelWaitTimeoutUs(u32 micro) {
-	return (s64)std::max(micro, 205U) + WAIT_TIMEOUT_LATENCY_US;
+	return (s64)std::max(micro, 205U) + WAIT_TIMEOUT_DEADLINE_US + WAIT_TIMEOUT_LATENCY_US;
 }
 
 void hleThreadEndTimeout(u64 userdata, int cyclesLate)
@@ -2199,6 +2200,9 @@ int __KernelStartThread(SceUID threadToStartID, int argSize, u32 argBlockPtr, bo
 		KernelValidateThreadTarget(startThread->context.pc);
 		__KernelChangeReadyState(cur, currentThread, true);
 		if (__InterruptsEnabled()) {
+			// Handing over costs more: about 30us from the call to the new thread's entry on
+			// hardware, where starting a worse one takes 10-25us.
+			hleEatCycles(2000);
 			g_startThreadHandoff = threadToStartID;
 			hleReSchedule("thread started");
 		}
@@ -2278,6 +2282,9 @@ int sceKernelGetThreadStackFreeSize(SceUID threadID) {
 	return hleLogDebug(Log::sceKernel, sz & ~3);
 }
 
+// What a thread ending costs, besides the switch away from it.
+static const int THREAD_EXIT_CYCLES = 3300;
+
 void __KernelReturnFromThread()
 {
 	hleSkipDeadbeef();
@@ -2287,6 +2294,8 @@ void __KernelReturnFromThread()
 	_dbg_assert_msg_(thread != NULL, "Returned from a NULL thread.");
 
 	DEBUG_LOG(Log::sceKernel, "__KernelReturnFromThread: %d", exitStatus);
+	// About 20us from a thread ending to a thread waiting on it running, on hardware.
+	hleEatCycles(THREAD_EXIT_CYCLES);
 	__KernelStopThread(currentThread, exitStatus, "thread returned");
 
 	hleReSchedule("thread returned");
@@ -2308,6 +2317,7 @@ int sceKernelExitThread(int exitStatus) {
 	if (exitStatus < 0) {
 		exitStatus = SCE_KERNEL_ERROR_ILLEGAL_ARGUMENT;
 	}
+	hleEatCycles(THREAD_EXIT_CYCLES);
 	__KernelStopThread(currentThread, exitStatus, "thread exited");
 
 	hleReSchedule("thread exited");
@@ -2343,6 +2353,7 @@ int sceKernelExitDeleteThread(int exitStatus) {
 		INFO_LOG(Log::sceKernel,"sceKernelExitDeleteThread(%d)", exitStatus);
 		uint32_t thread_attr = thread->nt.attr;
 		uint32_t uid = thread->GetUID();
+		hleEatCycles(THREAD_EXIT_CYCLES);
 		__KernelDeleteThread(currentThread, exitStatus, "thread exited with delete");
 
 		hleReSchedule("thread exited with delete");
@@ -3265,13 +3276,15 @@ void __KernelSwitchContext(PSPThread *target, const char *reason) {
 	}
 #endif
 
-	// Switching threads eats some cycles.  This is a low approximation.
+	// Switching threads eats some cycles. Between two threads it's about 5us: rotating the ready
+	// queue to an equal thread, or waking a better one and having it block again, each take
+	// 7-10us on hardware including the calls themselves (pspautotests threads/scheduling/handoff).
 	if (fromIdle && toIdle) {
 		// Don't eat any cycles going between idle.
 	} else if (fromIdle || toIdle) {
 		currentMIPS->downcount -= 1200;
 	} else {
-		currentMIPS->downcount -= 2700;
+		currentMIPS->downcount -= 1150;
 	}
 
 	__KernelUpdateBusySyscalls(target);
@@ -3621,6 +3634,10 @@ static void __KernelRunCallbackOnThread(SceUID cbId, PSPThread *thread, bool res
 	cb->nc.notifyCount = 0;
 	cb->nc.notifyArg = 0;
 
+	// Setting up the call costs about 8us on top of the switch: 14us from a notify to a better
+	// thread's callback running on hardware (pspautotests threads/callbacks/combos).
+	currentMIPS->downcount -= 1800;
+
 	ActionAfterCallback *action = (ActionAfterCallback *) __KernelCreateAction(actionAfterCallback);
 	if (action != NULL)
 		action->setCallback(cbId);

Core/HLE/sceKernelThread.h

@@ -386,7 +386,10 @@ void __KernelWaitCurThread(WaitType type, SceUID waitId, u32 waitValue, u32 time
 bool __KernelWaitTimesOutAtOnce(u32 timeoutPtr, int basePercent = 85, int stepPercent = 35);
 s64 __KernelWaitTimeoutUs(u32 micro);
 // How long after its deadline a wait's timeout goes off. Not part of the time left written back.
-const int WAIT_TIMEOUT_LATENCY_US = 30;
+const int WAIT_TIMEOUT_LATENCY_US = 18;
+// The deadline is taken this far into the call, after what we already charge before scheduling.
+// threads/semaphores/wait and threads/fpl/cancel pin it between about 10 and 15us.
+const int WAIT_TIMEOUT_DEADLINE_US = 12;
 void __KernelWaitCallbacksCurThread(WaitType type, SceUID waitID, u32 waitValue, u32 timeoutPtr);
 void __KernelReSchedule(const char *reason = "no reason");
 void __KernelReSchedule(bool doCallbacks, const char *reason);