am,amtest,mpe: support atomic_add and barrier

am/src/xs/isa/riscv/mpe.c

@@ -5,6 +5,7 @@ int __am_ncpu = 1;  // One core by default
 void _mpe_setncpu(char arg) {
   __am_ncpu = arg ? atoi(&arg) : 1;
   assert(0 < __am_ncpu && __am_ncpu <= MAX_CPU);
+  return;
 }
 
 void _mpe_wakeup(int cpu) {
@@ -47,5 +48,29 @@ intptr_t _atomic_xchg(volatile intptr_t *addr, intptr_t newval) {
 }
 
 intptr_t _atomic_add(volatile intptr_t *addr, intptr_t adder) {
-  return 0;
+  intptr_t result;
+  asm volatile(
+    "amoadd.d %0, %1, (%2);"
+    : "=r"(result)
+    : "r"(adder), "r"(addr)
+  );
+  return result;
+}
+
+void barrier() {
+  static volatile intptr_t sense = 0;
+  static volatile intptr_t count = 0;
+  static __thread intptr_t threadsense;
+
+  asm volatile("fence;");
+
+  threadsense = !threadsense;
+  if (_atomic_add(&count, 1) == _ncpu()-1) {
+    count = 0;
+    sense = threadsense;
+  }
+  else while(sense != threadsense)
+    ;
+
+  asm volatile("fence;");
 }

am/xsextra.h

@@ -17,6 +17,7 @@ extern "C" {
 void _mpe_setncpu(char arg);
 void _mpe_wakeup(int cpu);
 intptr_t _atomic_add(volatile intptr_t *addr, intptr_t adder);
+void barrier();
 
 #ifdef __cplusplus
 }

tests/amtest/src/tests/mp.c

@@ -4,7 +4,8 @@
  * Two build mp test, try `make ARCH=riscv64-xs-dual mainargs='m2'`
  */
 
-static volatile intptr_t sum;
+static volatile intptr_t sum = 0;
+static volatile intptr_t atomic_sum = 0;
 static volatile intptr_t print_lock = 0;
 
 void __am_uartlite_putchar(char ch);
@@ -14,11 +15,16 @@ void mp_print() {
   while (_atomic_xchg(&print_lock, 1) == 1);
   printf("My CPU ID is %d, nrCPU is %d\n", _cpu(), _ncpu());
   _atomic_xchg(&print_lock, 0);
-  _atomic_add(&sum, 1);
+  for (int i = 0; i < 100; i++) {
+    sum++;
+    _atomic_add(&atomic_sum, 1);
+  }
 }
 
 void finalize() {
+  barrier();
   while (_atomic_xchg(&print_lock, 1) == 1);
+  printf("sum = %d atomic_sum = %d\n", sum, atomic_sum);
   printf("Finalize CPU ID: %d\n", _cpu());
   _atomic_xchg(&print_lock, 0);
   while(1);