Merge zizzer.eecs.umich.edu:/bk/newmem/

into  zower.eecs.umich.edu:/home/gblack/m5/newmemmemops

--HG--
extra : convert_revision : c49b760eac758dbde30867cb638fcb3b790f4721

21 files changed, 199 insertions, 69 deletions

diff --git a/src/cpu/base.cc b/src/cpu/base.cc
index 4c243a2e9..7cbbb0b96 100644
--- a/src/cpu/base.cc
+++ b/src/cpu/base.cc
@@ -97,11 +97,13 @@ CPUProgressEvent::description()
 #if FULL_SYSTEM
 BaseCPU::BaseCPU(Params *p)
     : MemObject(p->name), clock(p->clock), checkInterrupts(true),
-      params(p), number_of_threads(p->numberOfThreads), system(p->system)
+      params(p), number_of_threads(p->numberOfThreads), system(p->system),
+      phase(p->phase)
 #else
 BaseCPU::BaseCPU(Params *p)
     : MemObject(p->name), clock(p->clock), params(p),
-      number_of_threads(p->numberOfThreads), system(p->system)
+      number_of_threads(p->numberOfThreads), system(p->system),
+      phase(p->phase)
 #endif
 {
 //    currentTick = curTick;
@@ -257,8 +259,9 @@ BaseCPU::regStats()
 Tick
 BaseCPU::nextCycle()
 {
-    Tick next_tick = curTick + clock - 1;
+    Tick next_tick = curTick - phase + clock - 1;
     next_tick -= (next_tick % clock);
+    next_tick += phase;
     return next_tick;
 }
 
@@ -266,11 +269,12 @@ Tick
 BaseCPU::nextCycle(Tick begin_tick)
 {
     Tick next_tick = begin_tick;
+    next_tick -= (next_tick % clock);
+    next_tick += phase;
 
     while (next_tick < curTick)
         next_tick += clock;
 
-    next_tick -= (next_tick % clock);
     assert(next_tick >= curTick);
     return next_tick;
 }
diff --git a/src/cpu/base.hh b/src/cpu/base.hh
index 788f77e3a..1d9b6a93b 100644
--- a/src/cpu/base.hh
+++ b/src/cpu/base.hh
@@ -153,6 +153,7 @@ class BaseCPU : public MemObject
         Tick functionTraceStart;
         System *system;
         int cpu_id;
+        Tick phase;
 #if FULL_SYSTEM
         Tick profile;
 
@@ -209,6 +210,8 @@ class BaseCPU : public MemObject
 
     System *system;
 
+    Tick phase;
+
 #if FULL_SYSTEM
     /**
      * Serialize this object to the given output stream.
diff --git a/src/cpu/memtest/memtest.cc b/src/cpu/memtest/memtest.cc
index 91e073cf0..180f41541 100644
--- a/src/cpu/memtest/memtest.cc
+++ b/src/cpu/memtest/memtest.cc
@@ -81,8 +81,13 @@ MemTest::CpuPort::recvFunctional(PacketPtr pkt)
 void
 MemTest::CpuPort::recvStatusChange(Status status)
 {
-    if (status == RangeChange)
+    if (status == RangeChange) {
+        if (!snoopRangeSent) {
+            snoopRangeSent = true;
+            sendStatusChange(Port::RangeChange);
+        }
         return;
+    }
 
     panic("MemTest doesn't expect recvStatusChange callback!");
 }
@@ -145,6 +150,9 @@ MemTest::MemTest(const string &name,
     //  thread = new SimpleThread(NULL, 0, NULL, 0, mainMem);
     curTick = 0;
 
+    cachePort.snoopRangeSent = false;
+    funcPort.snoopRangeSent = true;
+
     // Needs to be masked off once we know the block size.
     traceBlockAddr = _traceAddr;
     baseAddr1 = 0x100000;
diff --git a/src/cpu/memtest/memtest.hh b/src/cpu/memtest/memtest.hh
index 2694efd39..7bf34d827 100644
--- a/src/cpu/memtest/memtest.hh
+++ b/src/cpu/memtest/memtest.hh
@@ -100,6 +100,8 @@ class MemTest : public MemObject
             : Port(_name, _memtest), memtest(_memtest)
         { }
 
+        bool snoopRangeSent;
+
       protected:
 
         virtual bool recvTiming(PacketPtr pkt);
@@ -120,6 +122,8 @@ class MemTest : public MemObject
     CpuPort cachePort;
     CpuPort funcPort;
 
+    bool snoopRangeSent;
+
     class MemTestSenderState : public Packet::SenderState
     {
       public:
diff --git a/src/cpu/o3/alpha/cpu_builder.cc b/src/cpu/o3/alpha/cpu_builder.cc
index 09ccc7f65..5a375a4b8 100644
--- a/src/cpu/o3/alpha/cpu_builder.cc
+++ b/src/cpu/o3/alpha/cpu_builder.cc
@@ -48,6 +48,7 @@ class DerivO3CPU : public AlphaO3CPU<AlphaSimpleImpl>
 BEGIN_DECLARE_SIM_OBJECT_PARAMS(DerivO3CPU)
 
     Param<int> clock;
+    Param<int> phase;
     Param<int> numThreads;
 Param<int> activity;
 
@@ -158,6 +159,7 @@ END_DECLARE_SIM_OBJECT_PARAMS(DerivO3CPU)
 BEGIN_INIT_SIM_OBJECT_PARAMS(DerivO3CPU)
 
     INIT_PARAM(clock, "clock speed"),
+    INIT_PARAM_DFLT(phase, "clock phase", 0),
     INIT_PARAM(numThreads, "number of HW thread contexts"),
     INIT_PARAM_DFLT(activity, "Initial activity count", 0),
 
diff --git a/src/cpu/o3/fetch.hh b/src/cpu/o3/fetch.hh
index cc9a8abf5..04016347a 100644
--- a/src/cpu/o3/fetch.hh
+++ b/src/cpu/o3/fetch.hh
@@ -83,6 +83,8 @@ class DefaultFetch
             : Port(_fetch->name() + "-iport"), fetch(_fetch)
         { }
 
+        bool snoopRangeSent;
+
       protected:
         /** Atomic version of receive.  Panics. */
         virtual Tick recvAtomic(PacketPtr pkt);
diff --git a/src/cpu/o3/fetch_impl.hh b/src/cpu/o3/fetch_impl.hh
index 25faa407e..63d22b293 100644
--- a/src/cpu/o3/fetch_impl.hh
+++ b/src/cpu/o3/fetch_impl.hh
@@ -70,8 +70,13 @@ template<class Impl>
 void
 DefaultFetch<Impl>::IcachePort::recvStatusChange(Status status)
 {
-    if (status == RangeChange)
+    if (status == RangeChange) {
+        if (!snoopRangeSent) {
+            snoopRangeSent = true;
+            sendStatusChange(Port::RangeChange);
+        }
         return;
+    }
 
     panic("DefaultFetch doesn't expect recvStatusChange callback!");
 }
@@ -287,6 +292,8 @@ DefaultFetch<Impl>::setCPU(O3CPU *cpu_ptr)
     // Name is finally available, so create the port.
     icachePort = new IcachePort(this);
 
+    icachePort->snoopRangeSent = false;
+
 #if USE_CHECKER
     if (cpu->checker) {
         cpu->checker->setIcachePort(icachePort);
diff --git a/src/cpu/o3/lsq.hh b/src/cpu/o3/lsq.hh
index 6b12d75b4..7559a36d5 100644
--- a/src/cpu/o3/lsq.hh
+++ b/src/cpu/o3/lsq.hh
@@ -298,6 +298,8 @@ class LSQ {
             : lsq(_lsq)
         { }
 
+        bool snoopRangeSent;
+
       protected:
         /** Atomic version of receive.  Panics. */
         virtual Tick recvAtomic(PacketPtr pkt);
diff --git a/src/cpu/o3/lsq_impl.hh b/src/cpu/o3/lsq_impl.hh
index 5e7945c1c..6758e51c8 100644
--- a/src/cpu/o3/lsq_impl.hh
+++ b/src/cpu/o3/lsq_impl.hh
@@ -53,9 +53,13 @@ template <class Impl>
 void
 LSQ<Impl>::DcachePort::recvStatusChange(Status status)
 {
-    if (status == RangeChange)
+    if (status == RangeChange) {
+        if (!snoopRangeSent) {
+            snoopRangeSent = true;
+            sendStatusChange(Port::RangeChange);
+        }
         return;
-
+    }
     panic("O3CPU doesn't expect recvStatusChange callback!");
 }
 
@@ -97,6 +101,8 @@ LSQ<Impl>::LSQ(Params *params)
 {
     DPRINTF(LSQ, "Creating LSQ object.\n");
 
+    dcachePort.snoopRangeSent = false;
+
     //**********************************************/
     //************ Handle SMT Parameters ***********/
     //**********************************************/
diff --git a/src/cpu/o3/mips/cpu_builder.cc b/src/cpu/o3/mips/cpu_builder.cc
index ee9f2b48d..66741aee9 100644
--- a/src/cpu/o3/mips/cpu_builder.cc
+++ b/src/cpu/o3/mips/cpu_builder.cc
@@ -49,6 +49,7 @@ class DerivO3CPU : public MipsO3CPU<MipsSimpleImpl>
 BEGIN_DECLARE_SIM_OBJECT_PARAMS(DerivO3CPU)
 
 Param<int> clock;
+Param<int> phase;
 Param<int> numThreads;
 Param<int> activity;
 
@@ -146,6 +147,7 @@ END_DECLARE_SIM_OBJECT_PARAMS(DerivO3CPU)
 BEGIN_INIT_SIM_OBJECT_PARAMS(DerivO3CPU)
 
     INIT_PARAM(clock, "clock speed"),
+    INIT_PARAM_DFLT(phase, "clock phase", 0),
     INIT_PARAM(numThreads, "number of HW thread contexts"),
     INIT_PARAM_DFLT(activity, "Initial activity count", 0),
 
diff --git a/src/cpu/simple/atomic.cc b/src/cpu/simple/atomic.cc
index 58dc1fe5f..133b5500b 100644
--- a/src/cpu/simple/atomic.cc
+++ b/src/cpu/simple/atomic.cc
@@ -94,7 +94,7 @@ Tick
 AtomicSimpleCPU::CpuPort::recvAtomic(PacketPtr pkt)
 {
     //Snooping a coherence request, just return
-    return curTick;
+    return 0;
 }
 
 void
@@ -107,8 +107,13 @@ AtomicSimpleCPU::CpuPort::recvFunctional(PacketPtr pkt)
 void
 AtomicSimpleCPU::CpuPort::recvStatusChange(Status status)
 {
-    if (status == RangeChange)
+    if (status == RangeChange) {
+        if (!snoopRangeSent) {
+            snoopRangeSent = true;
+            sendStatusChange(Port::RangeChange);
+        }
         return;
+    }
 
     panic("AtomicSimpleCPU doesn't expect recvStatusChange callback!");
 }
@@ -127,6 +132,9 @@ AtomicSimpleCPU::AtomicSimpleCPU(Params *p)
 {
     _status = Idle;
 
+    icachePort.snoopRangeSent = false;
+    dcachePort.snoopRangeSent = false;
+
     ifetch_req = new Request();
     ifetch_req->setThreadContext(p->cpu_id, 0); // Add thread ID if we add MT
     ifetch_pkt = new Packet(ifetch_req, Packet::ReadReq, Packet::Broadcast);
@@ -512,6 +520,7 @@ BEGIN_DECLARE_SIM_OBJECT_PARAMS(AtomicSimpleCPU)
 #endif // FULL_SYSTEM
 
     Param<int> clock;
+    Param<int> phase;
 
     Param<bool> defer_registration;
     Param<int> width;
@@ -547,6 +556,7 @@ BEGIN_INIT_SIM_OBJECT_PARAMS(AtomicSimpleCPU)
 #endif // FULL_SYSTEM
 
     INIT_PARAM(clock, "clock speed"),
+    INIT_PARAM_DFLT(phase, "clock phase", 0),
     INIT_PARAM(defer_registration, "defer system registration (for sampling)"),
     INIT_PARAM(width, "cpu width"),
     INIT_PARAM(function_trace, "Enable function trace"),
@@ -567,6 +577,7 @@ CREATE_SIM_OBJECT(AtomicSimpleCPU)
     params->max_loads_all_threads = max_loads_all_threads;
     params->progress_interval = progress_interval;
     params->deferRegistration = defer_registration;
+    params->phase = phase;
     params->clock = clock;
     params->functionTrace = function_trace;
     params->functionTraceStart = function_trace_start;
diff --git a/src/cpu/simple/atomic.hh b/src/cpu/simple/atomic.hh
index 166a18127..0df6fe079 100644
--- a/src/cpu/simple/atomic.hh
+++ b/src/cpu/simple/atomic.hh
@@ -90,6 +90,8 @@ class AtomicSimpleCPU : public BaseSimpleCPU
             : Port(_name, _cpu), cpu(_cpu)
         { }
 
+        bool snoopRangeSent;
+
       protected:
 
         virtual bool recvTiming(PacketPtr pkt);
diff --git a/src/cpu/simple/timing.cc b/src/cpu/simple/timing.cc
index db2c940c0..3648f7613 100644
--- a/src/cpu/simple/timing.cc
+++ b/src/cpu/simple/timing.cc
@@ -82,8 +82,13 @@ TimingSimpleCPU::CpuPort::recvFunctional(PacketPtr pkt)
 void
 TimingSimpleCPU::CpuPort::recvStatusChange(Status status)
 {
-    if (status == RangeChange)
+    if (status == RangeChange) {
+        if (!snoopRangeSent) {
+            snoopRangeSent = true;
+            sendStatusChange(Port::RangeChange);
+        }
         return;
+    }
 
     panic("TimingSimpleCPU doesn't expect recvStatusChange callback!");
 }
@@ -101,6 +106,10 @@ TimingSimpleCPU::TimingSimpleCPU(Params *p)
       cpu_id(p->cpu_id)
 {
     _status = Idle;
+
+    icachePort.snoopRangeSent = false;
+    dcachePort.snoopRangeSent = false;
+
     ifetch_pkt = dcache_pkt = NULL;
     drainEvent = NULL;
     fetchEvent = NULL;
@@ -160,7 +169,7 @@ TimingSimpleCPU::resume()
 
         fetchEvent =
             new EventWrapper<TimingSimpleCPU, &TimingSimpleCPU::fetch>(this, false);
-        fetchEvent->schedule(curTick);
+        fetchEvent->schedule(nextCycle());
     }
 
     changeState(SimObject::Running);
@@ -232,7 +241,7 @@ TimingSimpleCPU::activateContext(int thread_num, int delay)
     // kick things off by initiating the fetch of the next instruction
     fetchEvent =
         new EventWrapper<TimingSimpleCPU, &TimingSimpleCPU::fetch>(this, false);
-    fetchEvent->schedule(curTick + cycles(delay));
+    fetchEvent->schedule(nextCycle(curTick + cycles(delay)));
 }
 
 
@@ -281,6 +290,8 @@ TimingSimpleCPU::read(Addr addr, T &data, unsigned flags)
             // memory system takes ownership of packet
             dcache_pkt = NULL;
         }
+    } else {
+        delete req;
     }
 
     // This will need a new way to tell if it has a dcache attached.
@@ -366,6 +377,8 @@ TimingSimpleCPU::write(T data, Addr addr, unsigned flags, uint64_t *res)
                 dcache_pkt = NULL;
             }
         }
+    } else {
+        delete req;
     }
 
     // This will need a new way to tell if it's hooked up to a cache or not.
@@ -448,6 +461,8 @@ TimingSimpleCPU::fetch()
             ifetch_pkt = NULL;
         }
     } else {
+        delete ifetch_req;
+        delete ifetch_pkt;
         // fetch fault: advance directly to next instruction (fault handler)
         advanceInst(fault);
     }
@@ -481,13 +496,13 @@ TimingSimpleCPU::completeIfetch(PacketPtr pkt)
 
     _status = Running;
 
-    delete pkt->req;
-    delete pkt;
-
     numCycles += curTick - previousTick;
     previousTick = curTick;
 
     if (getState() == SimObject::Draining) {
+        delete pkt->req;
+        delete pkt;
+
         completeDrain();
         return;
     }
@@ -519,6 +534,9 @@ TimingSimpleCPU::completeIfetch(PacketPtr pkt)
         postExecute();
         advanceInst(fault);
     }
+
+    delete pkt->req;
+    delete pkt;
 }
 
 void
@@ -674,6 +692,7 @@ BEGIN_DECLARE_SIM_OBJECT_PARAMS(TimingSimpleCPU)
 #endif // FULL_SYSTEM
 
     Param<int> clock;
+    Param<int> phase;
 
     Param<bool> defer_registration;
     Param<int> width;
@@ -709,6 +728,7 @@ BEGIN_INIT_SIM_OBJECT_PARAMS(TimingSimpleCPU)
 #endif // FULL_SYSTEM
 
     INIT_PARAM(clock, "clock speed"),
+    INIT_PARAM_DFLT(phase, "clock phase", 0),
     INIT_PARAM(defer_registration, "defer system registration (for sampling)"),
     INIT_PARAM(width, "cpu width"),
     INIT_PARAM(function_trace, "Enable function trace"),
@@ -730,6 +750,7 @@ CREATE_SIM_OBJECT(TimingSimpleCPU)
     params->progress_interval = progress_interval;
     params->deferRegistration = defer_registration;
     params->clock = clock;
+    params->phase = phase;
     params->functionTrace = function_trace;
     params->functionTraceStart = function_trace_start;
     params->system = system;
diff --git a/src/cpu/simple/timing.hh b/src/cpu/simple/timing.hh
index 408fa315e..fe5d03666 100644
--- a/src/cpu/simple/timing.hh
+++ b/src/cpu/simple/timing.hh
@@ -82,6 +82,8 @@ class TimingSimpleCPU : public BaseSimpleCPU
             : Port(_name, _cpu), cpu(_cpu), lat(_lat)
         { }
 
+        bool snoopRangeSent;
+
       protected:
 
         virtual Tick recvAtomic(PacketPtr pkt);
@@ -166,8 +168,6 @@ class TimingSimpleCPU : public BaseSimpleCPU
     PacketPtr ifetch_pkt;
     PacketPtr dcache_pkt;
 
-
-
     int cpu_id;
     Tick previousTick;
 
diff --git a/src/mem/bridge.cc b/src/mem/bridge.cc
index 38dcfd2e8..b787f79ca 100644
--- a/src/mem/bridge.cc
+++ b/src/mem/bridge.cc
@@ -91,10 +91,16 @@ Bridge::init()
 bool
 Bridge::BridgePort::recvTiming(PacketPtr pkt)
 {
-    DPRINTF(BusBridge, "recvTiming: src %d dest %d addr 0x%x\n",
-            pkt->getSrc(), pkt->getDest(), pkt->getAddr());
+    if (pkt->flags & SNOOP_COMMIT) {
+        DPRINTF(BusBridge, "recvTiming: src %d dest %d addr 0x%x\n",
+                pkt->getSrc(), pkt->getDest(), pkt->getAddr());
 
-    return otherPort->queueForSendTiming(pkt);
+        return otherPort->queueForSendTiming(pkt);
+    }
+    else {
+        // Else it's just a snoop, properly return if we are blocking
+        return !queueFull();
+    }
 }
 
 
diff --git a/src/mem/bus.cc b/src/mem/bus.cc
index 8ea67a0e4..6b5b63f50 100644
--- a/src/mem/bus.cc
+++ b/src/mem/bus.cc
@@ -160,11 +160,12 @@ Bus::recvTiming(PacketPtr pkt)
 
     short dest = pkt->getDest();
     if (dest == Packet::Broadcast) {
-        if (timingSnoop(pkt)) {
+        port = findPort(pkt->getAddr(), pkt->getSrc());
+        if (timingSnoop(pkt, port ? port : interfaces[pkt->getSrc()])) {
             bool success;
 
             pkt->flags |= SNOOP_COMMIT;
-            success = timingSnoop(pkt);
+            success = timingSnoop(pkt, port ? port : interfaces[pkt->getSrc()]);
             assert(success);
 
             if (pkt->flags & SATISFIED) {
@@ -177,7 +178,6 @@ Bus::recvTiming(PacketPtr pkt)
                 occupyBus(pkt);
                 return true;
             }
-            port = findPort(pkt->getAddr(), pkt->getSrc());
         } else {
             //Snoop didn't succeed
             DPRINTF(Bus, "Adding a retry to RETRY list %i\n", pktPort);
@@ -192,22 +192,28 @@ Bus::recvTiming(PacketPtr pkt)
 
     occupyBus(pkt);
 
-    if (port->sendTiming(pkt))  {
-        // Packet was successfully sent. Return true.
-        // Also take care of retries
-        if (inRetry) {
-            DPRINTF(Bus, "Remove retry from list %i\n", retryList.front());
-            retryList.front()->onRetryList(false);
-            retryList.pop_front();
-            inRetry = false;
+    if (port) {
+        if (port->sendTiming(pkt))  {
+            // Packet was successfully sent. Return true.
+            // Also take care of retries
+            if (inRetry) {
+                DPRINTF(Bus, "Remove retry from list %i\n", retryList.front());
+                retryList.front()->onRetryList(false);
+                retryList.pop_front();
+                inRetry = false;
+            }
+            return true;
         }
+
+        // Packet not successfully sent. Leave or put it on the retry list.
+        DPRINTF(Bus, "Adding a retry to RETRY list %i\n", pktPort);
+        addToRetryList(pktPort);
+        return false;
+    }
+    else {
+        //Forwarding up from responder, just return true;
         return true;
     }
-
-    // Packet not successfully sent. Leave or put it on the retry list.
-    DPRINTF(Bus, "Adding a retry to RETRY list %i\n", pktPort);
-    addToRetryList(pktPort);
-    return false;
 }
 
 void
@@ -290,7 +296,10 @@ Bus::findPort(Addr addr, int id)
 
 
     // we shouldn't be sending this back to where it came from
-    assert(dest_id != id);
+    // only on a functional access and then we should terminate
+    // the cyclical call.
+    if (dest_id == id)
+        return 0;
 
     return interfaces[dest_id];
 }
@@ -307,7 +316,18 @@ Bus::findSnoopPorts(Addr addr, int id)
         if (portSnoopList[i].range == addr && portSnoopList[i].portId != id) {
             //Careful  to not overlap ranges
             //or snoop will be called more than once on the port
-            ports.push_back(portSnoopList[i].portId);
+
+            //@todo Fix this hack because ranges are overlapping
+            //need to make sure we dont't create overlapping ranges
+            bool hack_overlap = false;
+            int size = ports.size();
+            for (int j=0; j < size; j++) {
+                if (ports[j] == portSnoopList[i].portId)
+                    hack_overlap = true;
+            }
+
+            if (!hack_overlap)
+                ports.push_back(portSnoopList[i].portId);
 //            DPRINTF(Bus, "  found snoop addr %#llx on device%d\n", addr,
 //                    portSnoopList[i].portId);
         }
@@ -317,17 +337,19 @@ Bus::findSnoopPorts(Addr addr, int id)
 }
 
 Tick
-Bus::atomicSnoop(PacketPtr pkt)
+Bus::atomicSnoop(PacketPtr pkt, Port *responder)
 {
     std::vector<int> ports = findSnoopPorts(pkt->getAddr(), pkt->getSrc());
     Tick response_time = 0;
 
     while (!ports.empty())
     {
-        Tick response = interfaces[ports.back()]->sendAtomic(pkt);
-        if (response) {
-            assert(!response_time);  //Multiple responders
-            response_time = response;
+        if (interfaces[ports.back()] != responder) {
+            Tick response = interfaces[ports.back()]->sendAtomic(pkt);
+            if (response) {
+                assert(!response_time);  //Multiple responders
+                response_time = response;
+            }
         }
         ports.pop_back();
     }
@@ -335,26 +357,31 @@ Bus::atomicSnoop(PacketPtr pkt)
 }
 
 void
-Bus::functionalSnoop(PacketPtr pkt)
+Bus::functionalSnoop(PacketPtr pkt, Port *responder)
 {
     std::vector<int> ports = findSnoopPorts(pkt->getAddr(), pkt->getSrc());
 
+    //The packet may be changed by another bus on snoops, restore the id after each
+    int id = pkt->getSrc();
     while (!ports.empty() && pkt->result != Packet::Success)
     {
-        interfaces[ports.back()]->sendFunctional(pkt);
+        if (interfaces[ports.back()] != responder)
+            interfaces[ports.back()]->sendFunctional(pkt);
         ports.pop_back();
+        pkt->setSrc(id);
     }
 }
 
 bool
-Bus::timingSnoop(PacketPtr pkt)
+Bus::timingSnoop(PacketPtr pkt, Port* responder)
 {
     std::vector<int> ports = findSnoopPorts(pkt->getAddr(), pkt->getSrc());
     bool success = true;
 
     while (!ports.empty() && success)
     {
-        success = interfaces[ports.back()]->sendTiming(pkt);
+        if (interfaces[ports.back()] != responder) //Don't call if responder also, once will do
+            success = interfaces[ports.back()]->sendTiming(pkt);
         ports.pop_back();
     }
 
@@ -370,15 +397,21 @@ Bus::recvAtomic(PacketPtr pkt)
     DPRINTF(Bus, "recvAtomic: packet src %d dest %d addr 0x%x cmd %s\n",
             pkt->getSrc(), pkt->getDest(), pkt->getAddr(), pkt->cmdString());
     assert(pkt->getDest() == Packet::Broadcast);
+    pkt->flags |= SNOOP_COMMIT;
 
     // Assume one bus cycle in order to get through.  This may have
     // some clock skew issues yet again...
     pkt->finishTime = curTick + clock;
-    Tick snoopTime = atomicSnoop(pkt);
+
+    Port *port = findPort(pkt->getAddr(), pkt->getSrc());
+    Tick snoopTime = atomicSnoop(pkt, port ? port : interfaces[pkt->getSrc()]);
+
     if (snoopTime)
         return snoopTime;  //Snoop satisfies it
+    else if (port)
+        return port->sendAtomic(pkt);
     else
-        return findPort(pkt->getAddr(), pkt->getSrc())->sendAtomic(pkt);
+        return 0;
 }
 
 /** Function called by the port when the bus is receiving a Functional
@@ -389,11 +422,15 @@ Bus::recvFunctional(PacketPtr pkt)
     DPRINTF(Bus, "recvFunctional: packet src %d dest %d addr 0x%x cmd %s\n",
             pkt->getSrc(), pkt->getDest(), pkt->getAddr(), pkt->cmdString());
     assert(pkt->getDest() == Packet::Broadcast);
-    functionalSnoop(pkt);
+    pkt->flags |= SNOOP_COMMIT;
+
+    Port* port = findPort(pkt->getAddr(), pkt->getSrc());
+    functionalSnoop(pkt, port ? port : interfaces[pkt->getSrc()]);
 
     // If the snooping found what we were looking for, we're done.
-    if (pkt->result != Packet::Success)
-        findPort(pkt->getAddr(), pkt->getSrc())->sendFunctional(pkt);
+    if (pkt->result != Packet::Success && port) {
+        port->sendFunctional(pkt);
+    }
 }
 
 /** Function called by the port when the bus is receiving a status change.*/
@@ -451,6 +488,7 @@ Bus::recvStatusChange(Port::Status status, int id)
             dm.portId = id;
             dm.range = *iter;
 
+            //@todo, make sure we don't overlap ranges
             DPRINTF(BusAddrRanges, "Adding snoop range %#llx - %#llx for id %d\n",
                     dm.range.start, dm.range.end, id);
             portSnoopList.push_back(dm);
@@ -493,7 +531,7 @@ Bus::addressRanges(AddrRangeList &resp, AddrRangeList &snoop, int id)
     for (dflt_iter = defaultRange.begin(); dflt_iter != defaultRange.end();
             dflt_iter++) {
         resp.push_back(*dflt_iter);
-        DPRINTF(BusAddrRanges, "  -- %#llx : %#llx\n",dflt_iter->start,
+        DPRINTF(BusAddrRanges, "  -- Dflt: %#llx : %#llx\n",dflt_iter->start,
                 dflt_iter->end);
     }
     for (portIter = portList.begin(); portIter != portList.end(); portIter++) {
@@ -519,6 +557,18 @@ Bus::addressRanges(AddrRangeList &resp, AddrRangeList &snoop, int id)
                     portIter->range.start, portIter->range.end);
         }
     }
+
+    for (portIter = portSnoopList.begin();
+         portIter != portSnoopList.end(); portIter++)
+    {
+        if (portIter->portId != id) {
+            snoop.push_back(portIter->range);
+            DPRINTF(BusAddrRanges, "  -- Snoop: %#llx : %#llx\n",
+                    portIter->range.start, portIter->range.end);
+            //@todo We need to properly insert snoop ranges
+            //not overlapping the ranges (multiple)
+        }
+    }
 }
 
 unsigned int
diff --git a/src/mem/bus.hh b/src/mem/bus.hh
index ff1d2545d..c472b6143 100644
--- a/src/mem/bus.hh
+++ b/src/mem/bus.hh
@@ -109,16 +109,16 @@ class Bus : public MemObject
     std::vector<int> findSnoopPorts(Addr addr, int id);
 
     /** Snoop all relevant ports atomicly. */
-    Tick atomicSnoop(PacketPtr pkt);
+    Tick atomicSnoop(PacketPtr pkt, Port* responder);
 
     /** Snoop all relevant ports functionally. */
-    void functionalSnoop(PacketPtr pkt);
+    void functionalSnoop(PacketPtr pkt, Port *responder);
 
     /** Call snoop on caches, be sure to set SNOOP_COMMIT bit if you want
      * the snoop to happen
      * @return True if succeds.
      */
-    bool timingSnoop(PacketPtr pkt);
+    bool timingSnoop(PacketPtr pkt, Port *responder);
 
     /** Process address range request.
      * @param resp addresses that we can respond to
diff --git a/src/mem/cache/base_cache.cc b/src/mem/cache/base_cache.cc
index c16cb6945..3af61375d 100644
--- a/src/mem/cache/base_cache.cc
+++ b/src/mem/cache/base_cache.cc
@@ -160,11 +160,14 @@ BaseCache::CachePort::recvRetry()
     PacketPtr pkt;
     assert(waitingOnRetry);
     if (!drainList.empty()) {
-        DPRINTF(CachePort, "%s attempting to send a retry for response\n", name());
+        DPRINTF(CachePort, "%s attempting to send a retry for response (%i waiting)\n"
+                , name(), drainList.size());
         //We have some responses to drain first
-        if (sendTiming(drainList.front())) {
-            DPRINTF(CachePort, "%s sucessful in sending a retry for response\n", name());
-            drainList.pop_front();
+        pkt = drainList.front();
+        drainList.pop_front();
+        if (sendTiming(pkt)) {
+            DPRINTF(CachePort, "%s sucessful in sending a retry for"
+                    "response (%i still waiting)\n", name(), drainList.size());
             if (!drainList.empty() ||
                 !isCpuSide && cache->doMasterRequest() ||
                 isCpuSide && cache->doSlaveRequest()) {
@@ -175,6 +178,9 @@ BaseCache::CachePort::recvRetry()
             }
             waitingOnRetry = false;
         }
+        else {
+            drainList.push_front(pkt);
+        }
         // Check if we're done draining once this list is empty
         if (drainList.empty())
             cache->checkDrain();
diff --git a/src/mem/cache/base_cache.hh b/src/mem/cache/base_cache.hh
index 584c2d5df..ef4955432 100644
--- a/src/mem/cache/base_cache.hh
+++ b/src/mem/cache/base_cache.hh
@@ -144,8 +144,6 @@ class BaseCache : public MemObject
   protected:
     CachePort *memSidePort;
 
-    bool snoopRangesSent;
-
   public:
     virtual Port *getPort(const std::string &if_name, int idx = -1);
 
@@ -171,10 +169,6 @@ class BaseCache : public MemObject
         if (status == Port::RangeChange){
             if (!isCpuSide) {
                 cpuSidePort->sendStatusChange(Port::RangeChange);
-                if (!snoopRangesSent) {
-                    snoopRangesSent = true;
-                    memSidePort->sendStatusChange(Port::RangeChange);
-                }
             }
             else {
                 memSidePort->sendStatusChange(Port::RangeChange);
@@ -358,7 +352,6 @@ class BaseCache : public MemObject
         //Start ports at null if more than one is created we should panic
         cpuSidePort = NULL;
         memSidePort = NULL;
-        snoopRangesSent = false;
     }
 
     ~BaseCache()
diff --git a/src/mem/packet.hh b/src/mem/packet.hh
index 2bc51bf12..19251941f 100644
--- a/src/mem/packet.hh
+++ b/src/mem/packet.hh
@@ -301,7 +301,7 @@ class Packet
 
     /** Destructor. */
     ~Packet()
-    { deleteData(); }
+    { if (staticData || dynamicData) deleteData(); }
 
     /** Reinitialize packet address and size from the associated
      *   Request object, and reset other fields that may have been
diff --git a/src/python/m5/objects/BaseCPU.py b/src/python/m5/objects/BaseCPU.py
index 2f702a4bf..8037c90af 100644
--- a/src/python/m5/objects/BaseCPU.py
+++ b/src/python/m5/objects/BaseCPU.py
@@ -47,6 +47,7 @@ class BaseCPU(SimObject):
         "defer registration with system (for sampling)")
 
     clock = Param.Clock(Parent.clock, "clock speed")
+    phase = Param.Latency("0ns", "clock phase")
 
     _mem_ports = []