Run sparks in batches, instead of creating a new thread for each one

[ghc-hetmet.git] / rts / Sparks.c

diff --git a/rts/Sparks.c b/rts/Sparks.c
index 360ea41..e7273f3 100644 (file)
--- a/rts/Sparks.c
+++ b/rts/Sparks.c
@@ -44,6 +44,7 @@
 #include "RtsUtils.h"
 #include "ParTicky.h"
 #include "Trace.h"
+#include "Prelude.h"
 
 #include "SMP.h" // for cas
 
@@ -227,8 +228,9 @@ steal(SparkPool *deque)
 }
 
 StgClosure *
-tryStealSpark (SparkPool *pool)
+tryStealSpark (Capability *cap)
 {
+  SparkPool *pool = cap->sparks;
   StgClosure *stolen;
 
   do { 
@@ -264,13 +266,13 @@ looksEmpty(SparkPool* deque)
  * -------------------------------------------------------------------------- */
 
 void
-createSparkThread (Capability *cap, StgClosure *p)
+createSparkThread (Capability *cap)
 {
     StgTSO *tso;
 
-    tso = createGenThread (cap, RtsFlags.GcFlags.initialStkSize, p);
+    tso = createIOThread (cap, RtsFlags.GcFlags.initialStkSize, 
+                          &base_GHCziConc_runSparks_closure);
     appendToRunQueue(cap,tso);
-    cap->sparks_converted++;
 }
 
 /* -----------------------------------------------------------------------------
@@ -371,13 +373,14 @@ newSpark (StgRegTable *reg, StgClosure *p)
  * the spark pool only contains sparkable closures.
  * -------------------------------------------------------------------------- */
 
-static void
-pruneSparkQueue (Capability *cap)
+void
+pruneSparkQueue (evac_fn evac, void *user, Capability *cap)
 { 
     SparkPool *pool;
-    StgClosurePtr spark, *elements;
+    StgClosurePtr spark, tmp, *elements;
     nat n, pruned_sparks; // stats only
     StgWord botInd,oldBotInd,currInd; // indices in array (always < size)
+    const StgInfoTable *info;
     
     PAR_TICKY_MARK_SPARK_QUEUE_START();
     
@@ -386,6 +389,13 @@ pruneSparkQueue (Capability *cap)
     
     pool = cap->sparks;
     
+    // Take this opportunity to reset top/bottom modulo the size of
+    // the array, to avoid overflow.  This is only possible because no
+    // stealing is happening during GC.
+    pool->bottom  -= pool->top & ~pool->moduloSize;
+    pool->top     &= pool->moduloSize;
+    pool->topBound = pool->top;
+
     debugTrace(DEBUG_sched,
                "markSparkQueue: current spark queue len=%d; (hd=%ld; tl=%ld)",
                sparkPoolSize(pool), pool->bottom, pool->top);
@@ -440,14 +450,31 @@ pruneSparkQueue (Capability *cap)
         botInd, otherwise move on */
       spark = elements[currInd];
 
-      /* if valuable work: shift inside the pool */
-      if ( closure_SHOULD_SPARK(spark) ) {
-       elements[botInd] = spark; // keep entry (new address)
-       botInd++;
-       n++;
-      } else { 
-       pruned_sparks++; // discard spark
-        cap->sparks_pruned++;
+      // We have to be careful here: in the parallel GC, another
+      // thread might evacuate this closure while we're looking at it,
+      // so grab the info pointer just once.
+      info = spark->header.info;
+      if (IS_FORWARDING_PTR(info)) {
+          tmp = (StgClosure*)UN_FORWARDING_PTR(info);
+          /* if valuable work: shift inside the pool */
+          if (closure_SHOULD_SPARK(tmp)) {
+              elements[botInd] = tmp; // keep entry (new address)
+              botInd++;
+              n++;
+          } else {
+              pruned_sparks++; // discard spark
+              cap->sparks_pruned++;
+          }
+      } else {
+          if (!(closure_flags[INFO_PTR_TO_STRUCT(info)->type] & _NS)) {
+              elements[botInd] = spark; // keep entry (new address)
+              evac (user, &elements[botInd]);
+              botInd++;
+              n++;
+          } else {
+              pruned_sparks++; // discard spark
+              cap->sparks_pruned++;
+          }
       }
       currInd++;
 
@@ -477,15 +504,6 @@ pruneSparkQueue (Capability *cap)
     ASSERT_SPARK_POOL_INVARIANTS(pool);
 }
 
-void
-pruneSparkQueues (void)
-{
-    nat i;
-    for (i = 0; i < n_capabilities; i++) {
-        pruneSparkQueue(&capabilities[i]);
-    }
-}
-
 /* GC for the spark pool, called inside Capability.c for all
    capabilities in turn. Blindly "evac"s complete spark pool. */
 void