Changes... small fix to jim's bugfix...touching our record after we free it is also...

poolalloc the rentry's...avoid running out of memory...looks like speed gains for micro also...

eliminate parentsemaphore...just use the thread local semaphore...needed to simplify freeing rentry records

diff --git a/Robust/src/IR/Flat/BuildCode.java b/Robust/src/IR/Flat/BuildCode.java
index 5573f7610ef4270e7875bc8636ea57d1a364fb46..b50fd27f7c3959b7c25cfc520f73c42beabfad29 100644 (file)
--- a/Robust/src/IR/Flat/BuildCode.java
+++ b/Robust/src/IR/Flat/BuildCode.java
@@ -3266,12 +3266,13 @@ public class BuildCode {
               for (Iterator iterator = waitingElementSet.iterator(); iterator.hasNext();) {
                 Analysis.OoOJava.WaitingElement waitingElement = (Analysis.OoOJava.WaitingElement) iterator.next();
                if( waitingElement.getStatus() >= ConflictNode.COARSE ){
-                  output.println("     rentry=mlpCreateREntry("+ waitingElement.getStatus()+ ", runningSESE);");
+                  output.println("     rentry=mlpCreateREntry(runningSESE->memoryQueueArray["+ waitingElement.getQueueID()+ "],"+waitingElement.getStatus()+ ", runningSESE);");
                 }else{
-                  output.println("     rentry=mlpCreateFineREntry("+ waitingElement.getStatus()+ ", runningSESE,  (void*)&" +generateTemp(fm,waitingElement.getTempDesc(),lb)+ ");");
+                  output.println("     rentry=mlpCreateFineREntry(runningSESE->memoryQueueArray["+ waitingElement.getQueueID()+ "],"+ waitingElement.getStatus()+ ", runningSESE,  (void*)&" +generateTemp(fm,waitingElement.getTempDesc(),lb)+ ");");
                 }         
-                output.println("     psem_init( &(rentry->parentStallSem) );");
-               output.println("     rentry->tag=rentry->parentStallSem.tag;");
+               output.println("     rentry->parentStallSem=&runningSESEstallSem;");
+                output.println("     psem_reset( &runningSESEstallSem);");
+               output.println("     rentry->tag=runningSESEstallSem.tag;");
                 output.println("     rentry->queue=runningSESE->memoryQueueArray["+ waitingElement.getQueueID()+ "];");
                 output.println("     if(ADDRENTRY(runningSESE->memoryQueueArray["+ waitingElement.getQueueID()+ "],rentry)==NOTREADY){");
                 if( state.COREPROF ) {
@@ -3283,7 +3284,7 @@ public class BuildCode {
                  //no need to enqueue parent effect if coarse grained conflict clears us
                  output.println("       while(stallrecord.common.rcrstatus) ;");
                  output.println("         BARRIER();");
-                 output.println("       stallrecord.common.parentsStallSem=&rentry->parentStallSem;");
+                 output.println("       stallrecord.common.parentsStallSem=&runningSESEstallSem;");
                  output.println("       stallrecord.tag=rentry->tag;");
                  output.println("       stallrecord.___obj___=(struct ___Object___ *)"+generateTemp(fm, waitingElement.getTempDesc(), null)+";");
                  output.println("       stallrecord.common.classID=-"+rcr.getTraverserID(waitingElement.getTempDesc(), fn)+";");
@@ -3291,7 +3292,7 @@ public class BuildCode {
                  output.println("       stallrecord.common.rcrstatus=1;");
                  output.println("       enqueueTR(TRqueue, (void *)&stallrecord);");
                 }
-               output.println("        psem_take( &(rentry->parentStallSem), (struct garbagelist *)&___locals___ );");
+               output.println("        psem_take( &runningSESEstallSem, (struct garbagelist *)&___locals___ );");
                 if( state.COREPROF ) {
                   output.println("#ifdef CP_EVENTID_TASKSTALLMEM");
                   output.println("        CP_LOGEVENT( CP_EVENTID_TASKSTALLMEM, CP_EVENTTYPE_END );");
@@ -3327,12 +3328,13 @@ public class BuildCode {
                                        
                   if( waitingElement.getStatus() >= ConflictNode.COARSE ){
                     // HERE! a parent might conflict with a child
-                    output.println("     rentry=mlpCreateREntry("+ waitingElement.getStatus()+ ", runningSESE);");
+                    output.println("     rentry=mlpCreateREntry(runningSESE->memoryQueueArray["+ waitingElement.getQueueID()+ "],"+ waitingElement.getStatus()+ ", runningSESE);");
                   } else {
-                    output.println("     rentry=mlpCreateFineREntry("+ waitingElement.getStatus()+ ", runningSESE,  (void*)&___locals___."+ waitingElement.getDynID() + ");");
+                    output.println("     rentry=mlpCreateFineREntry(runningSESE->memoryQueueArray["+ waitingElement.getQueueID()+ "],"+ waitingElement.getStatus()+ ", runningSESE,  (void*)&___locals___."+ waitingElement.getDynID() + ");");
                   }                                    
-                  output.println("     psem_init(&(rentry->parentStallSem));");
-                 output.println("     rentry->tag=rentry->parentStallSem->tag;");
+                 output.println("     rentry->parentStallSem=&runningSESEstallSem;");
+                 output.println("     psem_reset( &runningSESEstallSem);");
+                 output.println("     rentry->tag=runningSESEstallSem.tag;");
                   output.println("     rentry->queue=runningSESE->memoryQueueArray["+ waitingElement.getQueueID()+ "];");
                   output.println("     if(ADDRENTRY(runningSESE->memoryQueueArray["+ waitingElement.getQueueID()+"],rentry)==NOTREADY) {");
                   if( state.COREPROF ) {
@@ -3340,7 +3342,7 @@ public class BuildCode {
                     output.println("        CP_LOGEVENT( CP_EVENTID_TASKSTALLMEM, CP_EVENTTYPE_BEGIN );");
                     output.println("#endif");
                   }
-                  output.println("        psem_take( &(rentry->parentStallSem), (struct garbagelist *)&___locals___ );");
+                  output.println("        psem_take( &runningSESEstallSem, (struct garbagelist *)&___locals___ );");
                   if( state.COREPROF ) {
                     output.println("#ifdef CP_EVENTID_TASKSTALLMEM");
                     output.println("        CP_LOGEVENT( CP_EVENTID_TASKSTALLMEM, CP_EVENTTYPE_END );");
@@ -4176,7 +4178,7 @@ public class BuildCode {
                 Analysis.OoOJava.WaitingElement waitingElement 
                   = (Analysis.OoOJava.WaitingElement) iterator2.next();
                 if (waitingElement.getStatus() >= ConflictNode.COARSE) {
-                  output.println("       rentry=mlpCreateREntry("
+                  output.println("       rentry=mlpCreateREntry(runningSESE->memoryQueueArray["+ queueID+ "],"
                                  + waitingElement.getStatus()
                                  + ", &(seseToIssue->common));");
                 } else {
@@ -4189,7 +4191,7 @@ public class BuildCode {
                                    + "_srcSESE+seseToIssue->"
                                    + waitingElement.getDynID()
                                    + "_srcOffset;");
-                    output.println("       rentry=mlpCreateFineREntry("
+                    output.println("       rentry=mlpCreateFineREntry(runningSESE->memoryQueueArray["+ queueID+ "],"
                                    + waitingElement.getStatus()
                                    + ", &(seseToIssue->common),  pointer );");
                   } else if (fsen.getStaticInVarSet().contains(td)) {
@@ -4281,7 +4283,7 @@ public class BuildCode {
                 WaitingElement waitingElement = (WaitingElement) iterator2
                   .next();
                 if (waitingElement.getStatus() >= ConflictNode.COARSE) {
-                  output.println("     rentry=mlpCreateREntry("
+                  output.println("     rentry=mlpCreateREntry(runningSESE->memoryQueueArray["+ queueID+ "],"
                                  + waitingElement.getStatus()
                                  + ", &(seseToIssue->common));");
                 } else {
@@ -4297,7 +4299,7 @@ public class BuildCode {
                                    + waitingElement.getDynID()
                                    + "_srcOffset;");
                     output
-                      .println("     rentry=mlpCreateFineREntry("
+                      .println("     rentry=mlpCreateFineREntry(runningSESE->memoryQueueArray["+ queueID+ "],"
                                + waitingElement
                                .getStatus()
                                + ", &(seseToIssue->common),  pointer );");
@@ -4424,7 +4426,7 @@ public class BuildCode {
            Analysis.OoOJava.WaitingElement waitingElement=wtit.next();
            int queueID=waitingElement.getQueueID();
            assert(waitingElement.getStatus()>=ConflictNode.COARSE);
-           output.println("       rentry=mlpCreateREntry(" + waitingElement.getStatus() + ", &(seseToIssue->common));");
+           output.println("       rentry=mlpCreateREntry(runningSESE->memoryQueueArray["+ waitingElement.getQueueID()+ "]," + waitingElement.getStatus() + ", &(seseToIssue->common));");
            output.println("       seseToIssue->common.rentryArray[seseToIssue->common.rentryIdx++]=rentry;");
            output.println("       rentry->queue=runningSESE->memoryQueueArray[" + waitingElement.getQueueID()+"];");
            output.println("       if(ADDRENTRY(runningSESE->memoryQueueArray["+ waitingElement.getQueueID()+ "],rentry)==READY) {");
@@ -4642,6 +4644,9 @@ public class BuildCode {
     output.println("#endif // OOO_DISABLE_TASKMEMPOOL" );
 
 
+    output.println("{");
+    output.println("SESEcommon *myparent=runningSESE->parent;");
+
     // if this is not the Main sese (which has no parent) then return
     // THIS task's record to the PARENT'S task record pool, and only if
     // the reference count is now zero
@@ -4658,14 +4663,15 @@ public class BuildCode {
 
 
     // last of all, decrement your parent's number of running children    
-    output.println("   if( runningSESE->parent != NULL ) {");
-    output.println("     if( atomic_sub_and_test( 1, &(runningSESE->parent->numRunningChildren) ) ) {");
-    output.println("       pthread_mutex_lock  ( &(runningSESE->parent->lock) );");
-    output.println("       pthread_cond_signal ( &(runningSESE->parent->runningChildrenCond) );");
-    output.println("       pthread_mutex_unlock( &(runningSESE->parent->lock) );");
+    output.println("   if( myparent != NULL ) {");
+    output.println("     if( atomic_sub_and_test( 1, &(myparent->numRunningChildren) ) ) {");
+    output.println("       pthread_mutex_lock  ( &(myparent->lock) );");
+    output.println("       pthread_cond_signal ( &(myparent->runningChildrenCond) );");
+    output.println("       pthread_mutex_unlock( &(myparent->lock) );");
     output.println("     }");
     output.println("   }");
 
+    output.println("}");
     
     // as this thread is wrapping up the task, make sure the thread-local var
     // for the currently running task record references an invalid task

diff --git a/Robust/src/Runtime/memPool.h b/Robust/src/Runtime/memPool.h
index 17e8cc4212a944d59f5514efb73062711606342b..3087881aee088c4b976aa3cc5fa130e39acf5c08 100644 (file)
--- a/Robust/src/Runtime/memPool.h
+++ b/Robust/src/Runtime/memPool.h
@@ -22,7 +22,7 @@
 #include "runtime.h"
 #include "mem.h"
 #include "mlp_lock.h"
-
+#define CACHELINESIZE 64
 
 
 typedef struct MemPoolItem_t {
@@ -106,7 +106,7 @@ static inline void* poolalloc( MemPool* p ) {
   // it doesn't require an atomic op
   MemPoolItem* headCurrent = p->head;
   MemPoolItem* next=headCurrent->next;
-
+  int i;
   if(next == NULL) {
     // only one item, so don't take from pool
     return (void*) RUNMALLOC( p->itemSize );
@@ -127,6 +127,8 @@ static inline void* poolalloc( MemPool* p ) {
   //////////////////////////////////////////////////////////
   //__builtin_prefetch( &(p->head->next) );
   asm volatile( "prefetcht0 (%0)" :: "r" (next));
+  next=(MemPoolItem*)(((char *)next)+CACHELINESIZE);
+  asm volatile( "prefetcht0 (%0)" :: "r" (next));
 
   return (void*)headCurrent;
 }

diff --git a/Robust/src/Runtime/mlp_runtime.c b/Robust/src/Runtime/mlp_runtime.c
index 5d01810769abc528974be3d4adaa220540cdea78..75a296d2563c3c2d4ec83949698158ce032e1189 100644 (file)
--- a/Robust/src/Runtime/mlp_runtime.c
+++ b/Robust/src/Runtime/mlp_runtime.c
@@ -38,21 +38,24 @@ MemoryQueue** mlpCreateMemoryQueueArray(int numMemoryQueue){
   return newMemoryQueue;
 }
 
-REntry* mlpCreateREntryArray(){
-  REntry* newREntryArray=(REntry*)RUNMALLOC(sizeof(REntry)*NUMRENTRY);
-  return newREntryArray;
-}
-
-REntry* mlpCreateFineREntry(int type, SESEcommon* seseToIssue, void* dynID){
+REntry* mlpCreateFineREntry(MemoryQueue* q, int type, SESEcommon* seseToIssue, void* dynID){
+#ifdef OOO_DISABLE_TASKMEMPOOL
   REntry* newREntry=(REntry*)RUNMALLOC(sizeof(REntry));
+#else
+  REntry* newREntry=poolalloc(q->rentrypool);
+#endif
   newREntry->type=type;
   newREntry->seseRec=seseToIssue;
   newREntry->pointer=dynID;
   return newREntry;
 }
 
-REntry* mlpCreateREntry(int type, SESEcommon* seseToIssue){
+REntry* mlpCreateREntry(MemoryQueue* q, int type, SESEcommon* seseToIssue){
+#ifdef OOO_DISABLE_TASKMEMPOOL
   REntry* newREntry=(REntry*)RUNMALLOC(sizeof(REntry));
+#else
+  REntry* newREntry=poolalloc(q->rentrypool);
+#endif
   newREntry->type=type;
   newREntry->seseRec=seseToIssue;
   return newREntry;
@@ -197,6 +200,9 @@ MemoryQueue* createMemoryQueue(){
   dummy->status=READY;
   queue->head = dummy;
   queue->tail = dummy;
+#ifndef OOO_DISABLE_TASKMEMPOOL
+  queue->rentrypool = poolcreate(sizeof(REntry));
+#endif
   return queue;
 }
 
@@ -568,6 +574,9 @@ void RETIRERENTRY(MemoryQueue* Q, REntry * r) {
   } else if (isSCC(r)) {
     RETIRESCC(Q, r);
   }
+#ifndef OOO_DISABLE_TASKMEMPOOL
+  poolfreeinto(Q->rentrypool, r);
+#endif
 }
 
 RETIRESCC(MemoryQueue *Q, REntry *r) {
@@ -786,7 +795,7 @@ resolveDependencies(REntry* rentry){
       workScheduleSubmit(seseCommon);
     }   
   }else if(rentry->type==PARENTREAD || rentry->type==PARENTWRITE ||rentry->type==PARENTCOARSE){
-    psem_give_tag(&(rentry->parentStallSem), rentry->tag);
+    psem_give_tag(rentry->parentStallSem, rentry->tag);
   }
 }
 

diff --git a/Robust/src/Runtime/mlp_runtime.h b/Robust/src/Runtime/mlp_runtime.h
index 520a7d21fc8e230db0e4b0bc24a4ff8116c1c74d..b7657e45951701fbb9a1a7e9da6bc4641e0fe002 100644 (file)
--- a/Robust/src/Runtime/mlp_runtime.h
+++ b/Robust/src/Runtime/mlp_runtime.h
@@ -151,7 +151,7 @@ typedef struct REntry_t{
   struct Vector_t* vector;
   struct SCC_t* scc;
   struct MemoryQueue_t* queue;
-  psemaphore parentStallSem;
+  psemaphore * parentStallSem;
   int tag;
   SESEcommon* seseRec;
   INTPTR* pointer;
@@ -185,7 +185,8 @@ typedef struct MemoryQueueItem_t {
   int type; // hashtable:0, vector:1, singleitem:2
   int total;        //total non-retired
   int status;       //NOTREADY, READY
-  struct MemoryQueueItem_t *next; 
+  struct MemoryQueueItem_t *next;
+  
 } MemoryQueueItem;
 
 typedef struct MemoryQueue_t {
@@ -194,6 +195,9 @@ typedef struct MemoryQueue_t {
   REntry * binbuf[NUMBINS];
   REntry * buf[NUMRENTRY];
   int bufcount;
+#ifndef OOO_DISABLE_TASKMEMPOOL
+  MemPool * rentrypool;
+#endif
 } MemoryQueue;
 
 typedef struct BinItem_t {
@@ -256,8 +260,8 @@ void* mlpAllocSESErecord( int size );
 void  mlpFreeSESErecord( SESEcommon* seseRecord );
 
 MemoryQueue** mlpCreateMemoryQueueArray(int numMemoryQueue);
-REntry* mlpCreateFineREntry(int type, SESEcommon* seseToIssue, void* dynID);
-REntry* mlpCreateREntry    (int type, SESEcommon* seseToIssue);
+REntry* mlpCreateFineREntry(MemoryQueue *q, int type, SESEcommon* seseToIssue, void* dynID);
+REntry* mlpCreateREntry(MemoryQueue *q, int type, SESEcommon* seseToIssue);
 MemoryQueue* createMemoryQueue();
 void rehashMemoryQueue(SESEcommon* seseParent);
 

diff --git a/Robust/src/Runtime/oooJava/hashStructure.c b/Robust/src/Runtime/oooJava/hashStructure.c
index 790c55f4342288deb1e29c95defad1d9cfbaa864..2b8222288fca10be0c24c5e1184b47083a81e89a 100644 (file)
--- a/Robust/src/Runtime/oooJava/hashStructure.c
+++ b/Robust/src/Runtime/oooJava/hashStructure.c
@@ -92,13 +92,13 @@ inline int rcr_BWRITEBINCASE(HashStructure *T, int key, SESEcommon *task, struct
     td->task=task;
     td->bitindexrd=td->bitindexwr=1<<index;
     be->tail=b;
-    
+    BARRIER();//do tail before head
     //release lock
     be->head=b;
     enqueuerecord(rcrrec, key, b);
     return READY;
   }
-
+  BARRIER();//read head before tail
   BinItem_rcr *bintail=be->tail;
   bitvt rdmask=0,wrmask=0;
   int status=NOTREADY;

diff --git a/Robust/src/Runtime/oooJava/trqueue.c b/Robust/src/Runtime/oooJava/trqueue.c
index 852c48ae4e5dc648677d56d2b8b203f8e5a51785..55d105a0ac613b257472bb835fe4d9fc14d6fc45 100644 (file)
--- a/Robust/src/Runtime/oooJava/trqueue.c
+++ b/Robust/src/Runtime/oooJava/trqueue.c
@@ -25,10 +25,11 @@ void * dequeueTR(struct trQueue *q) {
   if(q->head==tail)
     return NULL;
 
-  void * ptr = q->elements[tail];
   tail++;
   if(tail & TRSIZE)
     tail =  0;
+
+  void * ptr = q->elements[tail];
   q->tail=tail;
   return ptr;
 }