Performance improvements thanks to code paths tuning using likely/unlikely macros.

git-svn-id: svn://svn.code.sf.net/p/chibios/svn/trunk@14528 27425a3e-05d8-49a3-a47f-9c15f0e5edd8

doc/rt/reports/benchmarks/rt-stm32g4-170mhz-gcc.txt

@@ -1,11 +1,11 @@
 
 *** ChibiOS/RT Test Suite
 ***
-*** Compiled:     Jun  7 2021 - 08:51:33
+*** Compiled:     Jun 11 2021 - 11:12:01
 *** Platform:     STM32G4 Hi-resolution Line
 *** Test Board:   STMicroelectronics STM32 Nucleo64-G474RE
 ***
-*** Text size:    51148 bytes
+*** Text size:    50964 bytes
 *** RO data size: 9904 bytes
 *** Data size:    220 bytes
 *** BSS size:     5400 bytes
@@ -189,7 +189,7 @@
 === Test Sequence 12 (Benchmarks)
 ----------------------------------------------------------------------------
 --- Test Case 12.1 (Messages performance #1)
---- Score : 904248 msgs/S, 1808496 ctxswc/S
+--- Score : 923906 msgs/S, 1847812 ctxswc/S
 --- Result: SUCCESS
 ----------------------------------------------------------------------------
 --- Test Case 12.2 (Messages performance #2)
@@ -201,7 +201,7 @@
 --- Result: SUCCESS
 ----------------------------------------------------------------------------
 --- Test Case 12.4 (Context Switch performance)
---- Score : 2821568 ctxswc/S
+--- Score : 2893608 ctxswc/S
 --- Result: SUCCESS
 ----------------------------------------------------------------------------
 --- Test Case 12.5 (Threads performance, full cycle)
@@ -209,15 +209,15 @@
 --- Result: SUCCESS
 ----------------------------------------------------------------------------
 --- Test Case 12.6 (Threads performance, create/exit only)
---- Score : 634325 threads/S
+--- Score : 639095 threads/S
 --- Result: SUCCESS
 ----------------------------------------------------------------------------
 --- Test Case 12.7 (Mass reschedule performance)
---- Score : 288135 reschedules/S, 1728810 ctxswc/S
+--- Score : 289115 reschedules/S, 1734690 ctxswc/S
 --- Result: SUCCESS
 ----------------------------------------------------------------------------
 --- Test Case 12.8 (Round-Robin voluntary reschedule)
---- Score : 1748060 ctxswc/S
+--- Score : 1752560 ctxswc/S
 --- Result: SUCCESS
 ----------------------------------------------------------------------------
 --- Test Case 12.9 (Virtual Timers set/reset performance)
@@ -249,11 +249,11 @@ Final result: SUCCESS
 
 *** ChibiOS OS Library Test Suite
 ***
-*** Compiled:     Jun  7 2021 - 08:51:33
+*** Compiled:     Jun 11 2021 - 11:12:01
 *** Platform:     STM32G4 Hi-resolution Line
 *** Test Board:   STMicroelectronics STM32 Nucleo64-G474RE
 ***
-*** Text size:    51148 bytes
+*** Text size:    50964 bytes
 *** RO data size: 9904 bytes
 *** Data size:    220 bytes
 *** BSS size:     5400 bytes

os/rt/src/chschd.c

@@ -414,11 +414,13 @@ void chSchWakeupS(thread_t *ntp, msg_t msg) {
   }
 #endif
 
-  /* If the waken thread has a not-greater priority than the current
+  /* If the woken thread has a not-greater priority than the current
      one then it is just inserted in the ready list else it made
      running immediately and the invoking thread goes in the ready
-     list instead.*/
+     list instead.
-  if (ntp->hdr.pqueue.prio <= otp->hdr.pqueue.prio) {
+     Note, we are favoring the path where the woken thread has higher
+     priority.*/
+  if (unlikely(ntp->hdr.pqueue.prio <= otp->hdr.pqueue.prio)) {
     (void) __sch_ready_behind(ntp);
   }
   else {
@@ -455,7 +457,9 @@ void chSchRescheduleS(void) {
 
   chDbgCheckClassS();
 
-  if (firstprio(&oip->rlist.pqueue) > tp->hdr.pqueue.prio) {
+  /* Note, we are favoring the path where the reschedule is necessary
+     because higher priority threads are ready.*/
+  if (likely(firstprio(&oip->rlist.pqueue) > tp->hdr.pqueue.prio)) {
     __sch_reschedule_ahead();
   }
 }
@@ -563,19 +567,21 @@ void chSchPreemption(void) {
   tprio_t p1 = firstprio(&oip->rlist.pqueue);
   tprio_t p2 = tp->hdr.pqueue.prio;
 
+  /* Note, we are favoring the path where preemption is necessary
+     because higher priority threads are ready.*/
 #if CH_CFG_TIME_QUANTUM > 0
   if (tp->ticks > (tslices_t)0) {
-    if (p1 > p2) {
+    if (likely(p1 > p2)) {
       __sch_reschedule_ahead();
     }
   }
   else {
-    if (p1 >= p2) {
+    if (likely(p1 >= p2)) {
       __sch_reschedule_behind();
     }
   }
 #else /* CH_CFG_TIME_QUANTUM == 0 */
-  if (p1 > p2) {
+  if (likely(p1 > p2)) {
     __sch_reschedule_ahead();
   }
 #endif /* CH_CFG_TIME_QUANTUM == 0 */
@@ -595,7 +601,9 @@ void chSchDoYieldS(void) {
 
   chDbgCheckClassS();
 
-  if (firstprio(&oip->rlist.pqueue) >= tp->hdr.pqueue.prio) {
+  /* If this function has been called then it is likely there are threads
+     at same priority level.*/
+  if (likely(firstprio(&oip->rlist.pqueue) >= tp->hdr.pqueue.prio)) {
     __sch_reschedule_behind();
   }
 }

release_note_next.txt

@@ -47,6 +47,9 @@ integrated in our structure: WolfSSL, FatFS and lwIP.
 
 - Support for full-SMP multi-core threading.
 - Support for decoupled multi-core threading.
+- Performance improvements thanks to code paths tuning using likely/unlikely
+  macros. This feature requires compiler support and is currently enabled
+  for GCC.
 - 64 bits monotonic time stamps with the same resolution of system time.
 - Much more efficient and accurate Virtual Timers in tick-less mode.
 - Automatic reload of Virtual Timers, now both one-shot and continuous timers