rts: Use SEQ_CST accesses when touching `wakeup`

These are the two remaining non-atomic accesses to `wakeup` which were
missed by the original TSAN patch.

rts/hooks/LongGCSync.c

@@ -21,7 +21,7 @@ void LongGCSync (uint32_t me USED_IF_THREADS, Time t STG_UNUSED)
     {
         uint32_t i;
         for (i=0; i < n_capabilities; i++) {
-            if (i != me && gc_threads[i]->wakeup != GC_THREAD_STANDING_BY) {
+            if (i != me && SEQ_CST_LOAD(&gc_threads[i]->wakeup) == GC_THREAD_STANDING_BY) {
                 debugBelch("Warning: slow GC sync: still waiting for cap %d\n",
                            i);
             }

rts/sm/GC.c

@@ -1055,7 +1055,7 @@ new_gc_thread (uint32_t n, gc_thread *t)
     initSpinLock(&t->mut_spin);
     ACQUIRE_SPIN_LOCK(&t->gc_spin);
     ACQUIRE_SPIN_LOCK(&t->mut_spin);
-    t->wakeup = GC_THREAD_INACTIVE;  // starts true, so we can wait for the
+    SEQ_CST_STORE(&t->wakeup, GC_THREAD_INACTIVE);  // starts true, so we can wait for the
                           // thread to start up, see wakeup_gc_threads
 #endif
 
@@ -1300,7 +1300,7 @@ gcWorkerThread (Capability *cap)
     //    measurements more accurate on Linux, perhaps because it syncs
     //    the CPU time across the multiple cores.  Without this, CPU time
     //    is heavily skewed towards GC rather than MUT.
-    gct->wakeup = GC_THREAD_STANDING_BY;
+    SEQ_CST_STORE(&gct->wakeup, GC_THREAD_STANDING_BY);
     debugTrace(DEBUG_gc, "GC thread %d standing by...", gct->thread_index);
     ACQUIRE_SPIN_LOCK(&gct->gc_spin);
 

rts/sm/GCThread.h

@@ -127,7 +127,7 @@ typedef struct gc_thread_ {
     OSThreadId id;                 // The OS thread that this struct belongs to
     SpinLock   gc_spin;
     SpinLock   mut_spin;
-    volatile StgWord wakeup;       // NB not StgWord8; only StgWord is guaranteed atomic
+    StgWord wakeup;                // This should only be accessed via atomic accesses
 #endif
     uint32_t thread_index;         // a zero based index identifying the thread