Changes in libcfa/src/concurrency/ready_queue.cfa [d3ba775:f302d80]

File:

• libcfa/src/concurrency/ready_queue.cfa (modified) (5 diffs)

libcfa/src/concurrency/ready_queue.cfa

@@ -398 +398 @@
 
                 if(proc->rdq.target == -1u) {
-                        proc->rdq.target = __tls_rand() % lanes.count;
-                        unsigned it1  = proc->rdq.itr;
-                        unsigned it2  = proc->rdq.itr + 1;
-                        unsigned idx1 = proc->rdq.id + (it1 % READYQ_SHARD_FACTOR);
-                        unsigned idx2 = proc->rdq.id + (it2 % READYQ_SHARD_FACTOR);
+                        _Static_assert(READYQ_SHARD_FACTOR == 2);
+                        unsigned idx1 = proc->rdq.id + 0;
+                        unsigned idx2 = proc->rdq.id + 1;
                         unsigned long long tsc1 = ts(lanes.data[idx1]);
                         unsigned long long tsc2 = ts(lanes.data[idx2]);
-                        proc->rdq.cutoff = min(tsc1, tsc2);
-                        if(proc->rdq.cutoff == 0) proc->rdq.cutoff = -1ull;
+                        proc->rdq.target = __tls_rand() % lanes.count;
+
+                        // WARNING: std::min is polymorphic and therefore causes 500% slowdown instead of the expected 2%
+                        proc->rdq.cutoff = tsc1 < tsc2 ? tsc1 : tsc2;
                 }
                 else {
@@ -418 +418 @@
 
                 for(READYQ_SHARD_FACTOR) {
-                        unsigned i = proc->rdq.id + (--proc->rdq.itr % READYQ_SHARD_FACTOR);
+                        unsigned i = proc->rdq.id + (proc->rdq.itr++ % READYQ_SHARD_FACTOR);
                         if($thread * t = try_pop(cltr, i __STATS(, __tls_stats()->ready.pop.local))) return t;
                 }
@@ -469 +469 @@
         // Actually pop the list
         struct $thread * thrd;
-        thrd = pop(lane);
+        unsigned long long tsv;
+        [thrd, tsv] = pop(lane);
 
         /* paranoid */ verify(thrd);
@@ -481 +482 @@
 
         #if defined(USE_WORK_STEALING)
-                lanes.tscs[w].tv = thrd->link.ts;
+                lanes.tscs[w].tv = tsv;
         #endif
 
@@ -663 +664 @@
                         while(!is_empty(lanes.data[idx])) {
                                 struct $thread * thrd;
-                                thrd = pop(lanes.data[idx]);
+                                unsigned long long _;
+                                [thrd, _] = pop(lanes.data[idx]);
 
                                 push(cltr, thrd);