fix memory leak
[c11tester.git] / datarace.cc
index 9ba59399f2a07c52cf62136a897e59cd55717739..cfbe94bff3553761f81069c40ed63739db578e7a 100644 (file)
@@ -9,11 +9,12 @@
 #include "action.h"
 #include "execution.h"
 #include "stl-model.h"
+#include <execinfo.h>
 
 static struct ShadowTable *root;
-static SnapVector<DataRace *> *unrealizedraces;
 static void *memory_base;
 static void *memory_top;
+static RaceSet * raceset;
 
 static const ModelExecution * get_execution()
 {
@@ -26,7 +27,7 @@ void initRaceDetector()
        root = (struct ShadowTable *)snapshot_calloc(sizeof(struct ShadowTable), 1);
        memory_base = snapshot_calloc(sizeof(struct ShadowBaseTable) * SHADOWBASETABLES, 1);
        memory_top = ((char *)memory_base) + sizeof(struct ShadowBaseTable) * SHADOWBASETABLES;
-       unrealizedraces = new SnapVector<DataRace *>();
+       raceset = new RaceSet();
 }
 
 void * table_calloc(size_t size)
@@ -59,6 +60,50 @@ static uint64_t * lookupAddressEntry(const void *address)
        return &basetable->array[((uintptr_t)address) & MASK16BIT];
 }
 
+
+bool hasNonAtomicStore(const void *address) {
+       uint64_t * shadow = lookupAddressEntry(address);
+       uint64_t shadowval = *shadow;
+       if (ISSHORTRECORD(shadowval)) {
+               //Do we have a non atomic write with a non-zero clock
+               return !(ATOMICMASK & shadowval);
+       } else {
+               if (shadowval == 0)
+                       return true;
+               struct RaceRecord *record = (struct RaceRecord *)shadowval;
+               return !record->isAtomic;
+       }
+}
+
+void setAtomicStoreFlag(const void *address) {
+       uint64_t * shadow = lookupAddressEntry(address);
+       uint64_t shadowval = *shadow;
+       if (ISSHORTRECORD(shadowval)) {
+               *shadow = shadowval | ATOMICMASK;
+       } else {
+               if (shadowval == 0) {
+                       *shadow = ATOMICMASK | ENCODEOP(0, 0, 0, 0);
+                       return;
+               }
+               struct RaceRecord *record = (struct RaceRecord *)shadowval;
+               record->isAtomic = 1;
+       }
+}
+
+void getStoreThreadAndClock(const void *address, thread_id_t * thread, modelclock_t * clock) {
+       uint64_t * shadow = lookupAddressEntry(address);
+       uint64_t shadowval = *shadow;
+       if (ISSHORTRECORD(shadowval) || shadowval == 0) {
+               //Do we have a non atomic write with a non-zero clock
+               *thread = WRTHREADID(shadowval);
+               *clock = WRITEVECTOR(shadowval);
+       } else {
+               struct RaceRecord *record = (struct RaceRecord *)shadowval;
+               *thread = record->writeThread;
+               *clock = record->writeClock;
+       }
+}
+
 /**
  * Compares a current clock-vector/thread-ID pair with a clock/thread-ID pair
  * to check the potential for a data race.
@@ -92,59 +137,52 @@ static void expandRecord(uint64_t *shadow)
        record->writeClock = writeClock;
 
        if (readClock != 0) {
-               record->capacity = INITCAPACITY;
-               record->thread = (thread_id_t *)snapshot_malloc(sizeof(thread_id_t) * record->capacity);
-               record->readClock = (modelclock_t *)snapshot_malloc(sizeof(modelclock_t) * record->capacity);
+               record->thread = (thread_id_t *)snapshot_malloc(sizeof(thread_id_t) * INITCAPACITY);
+               record->readClock = (modelclock_t *)snapshot_malloc(sizeof(modelclock_t) * INITCAPACITY);
                record->numReads = 1;
+               ASSERT(readThread >= 0);
                record->thread[0] = readThread;
                record->readClock[0] = readClock;
+       } else {
+               record->thread = NULL;
        }
+       if (shadowval & ATOMICMASK)
+               record->isAtomic = 1;
        *shadow = (uint64_t) record;
 }
 
+#define FIRST_STACK_FRAME 2
+
+unsigned int race_hash(struct DataRace *race) {
+       unsigned int hash = 0;
+       for(int i=FIRST_STACK_FRAME;i < race->numframes;i++) {
+               hash ^= ((uintptr_t)race->backtrace[i]);
+               hash = (hash >> 3) | (hash << 29);
+       }
+       return hash;
+}
+
+bool race_equals(struct DataRace *r1, struct DataRace *r2) {
+       if (r1->numframes != r2->numframes)
+               return false;
+       for(int i=FIRST_STACK_FRAME;i < r1->numframes;i++) {
+               if (r1->backtrace[i] != r2->backtrace[i])
+                       return false;
+       }
+       return true;
+}
+
 /** This function is called when we detect a data race.*/
-static void reportDataRace(thread_id_t oldthread, modelclock_t oldclock, bool isoldwrite, ModelAction *newaction, bool isnewwrite, const void *address)
+static struct DataRace * reportDataRace(thread_id_t oldthread, modelclock_t oldclock, bool isoldwrite, ModelAction *newaction, bool isnewwrite, const void *address)
 {
-       struct DataRace *race = (struct DataRace *)snapshot_malloc(sizeof(struct DataRace));
+       struct DataRace *race = (struct DataRace *)model_malloc(sizeof(struct DataRace));
        race->oldthread = oldthread;
        race->oldclock = oldclock;
        race->isoldwrite = isoldwrite;
        race->newaction = newaction;
        race->isnewwrite = isnewwrite;
        race->address = address;
-       unrealizedraces->push_back(race);
-
-       /* If the race is realized, bail out now. */
-       if (checkDataRaces())
-               model->switch_to_master(NULL);
-}
-
-/**
- * @brief Check and report data races
- *
- * If the trace is feasible (a feasible prefix), clear out the list of
- * unrealized data races, asserting any realized ones as execution bugs so that
- * the model-checker will end the execution.
- *
- * @return True if any data races were realized
- */
-bool checkDataRaces()
-{
-       if (get_execution()->isfeasibleprefix()) {
-               bool race_asserted = false;
-               /* Prune the non-racing unrealized dataraces */
-               for (unsigned i = 0;i < unrealizedraces->size();i++) {
-                       struct DataRace *race = (*unrealizedraces)[i];
-                       if (clock_may_race(race->newaction->get_cv(), race->newaction->get_tid(), race->oldclock, race->oldthread)) {
-                               assert_race(race);
-                               race_asserted = true;
-                       }
-                       snapshot_free(race);
-               }
-               unrealizedraces->clear();
-               return race_asserted;
-       }
-       return false;
+       return race;
 }
 
 /**
@@ -157,24 +195,26 @@ bool checkDataRaces()
  */
 void assert_race(struct DataRace *race)
 {
-       model->assert_bug(
-               "Data race detected @ address %p:\n"
-               "    Access 1: %5s in thread %2d @ clock %3u\n"
-               "    Access 2: %5s in thread %2d @ clock %3u",
-               race->address,
-               race->isoldwrite ? "write" : "read",
-               id_to_int(race->oldthread),
-               race->oldclock,
-               race->isnewwrite ? "write" : "read",
-               id_to_int(race->newaction->get_tid()),
-               race->newaction->get_seq_number()
-               );
+       model_print("Race detected at location: \n");
+       backtrace_symbols_fd(race->backtrace, race->numframes, model_out);
+       model_print("\nData race detected @ address %p:\n"
+                                                       "    Access 1: %5s in thread %2d @ clock %3u\n"
+                                                       "    Access 2: %5s in thread %2d @ clock %3u\n\n",
+                                                       race->address,
+                                                       race->isoldwrite ? "write" : "read",
+                                                       id_to_int(race->oldthread),
+                                                       race->oldclock,
+                                                       race->isnewwrite ? "write" : "read",
+                                                       id_to_int(race->newaction->get_tid()),
+                                                       race->newaction->get_seq_number()
+                                                       );
 }
 
 /** This function does race detection for a write on an expanded record. */
-void fullRaceCheckWrite(thread_id_t thread, void *location, uint64_t *shadow, ClockVector *currClock)
+struct DataRace * fullRaceCheckWrite(thread_id_t thread, const void *location, uint64_t *shadow, ClockVector *currClock)
 {
        struct RaceRecord *record = (struct RaceRecord *)(*shadow);
+       struct DataRace * race = NULL;
 
        /* Check for datarace against last read. */
 
@@ -187,36 +227,237 @@ void fullRaceCheckWrite(thread_id_t thread, void *location, uint64_t *shadow, Cl
 
                if (clock_may_race(currClock, thread, readClock, readThread)) {
                        /* We have a datarace */
-                       reportDataRace(readThread, readClock, false, get_execution()->get_parent_action(thread), true, location);
+                       race = reportDataRace(readThread, readClock, false, get_execution()->get_parent_action(thread), true, location);
+                       goto Exit;
                }
        }
 
        /* Check for datarace against last write. */
+       {
+               modelclock_t writeClock = record->writeClock;
+               thread_id_t writeThread = record->writeThread;
 
-       modelclock_t writeClock = record->writeClock;
-       thread_id_t writeThread = record->writeThread;
+               if (clock_may_race(currClock, thread, writeClock, writeThread)) {
+                       /* We have a datarace */
+                       race = reportDataRace(writeThread, writeClock, true, get_execution()->get_parent_action(thread), true, location);
+                       goto Exit;
+               }
+       }
+Exit:
+       record->numReads = 0;
+       record->writeThread = thread;
+       record->isAtomic = 0;
+       modelclock_t ourClock = currClock->getClock(thread);
+       record->writeClock = ourClock;
+       return race;
+}
 
-       if (clock_may_race(currClock, thread, writeClock, writeThread)) {
-               /* We have a datarace */
-               reportDataRace(writeThread, writeClock, true, get_execution()->get_parent_action(thread), true, location);
+/** This function does race detection on a write. */
+void raceCheckWrite(thread_id_t thread, void *location)
+{
+       uint64_t *shadow = lookupAddressEntry(location);
+       uint64_t shadowval = *shadow;
+       ClockVector *currClock = get_execution()->get_cv(thread);
+       if (currClock == NULL)
+               return;
+
+       struct DataRace * race = NULL;
+       /* Do full record */
+       if (shadowval != 0 && !ISSHORTRECORD(shadowval)) {
+               race = fullRaceCheckWrite(thread, location, shadow, currClock);
+               goto Exit;
+       }
+
+       {
+               int threadid = id_to_int(thread);
+               modelclock_t ourClock = currClock->getClock(thread);
+
+               /* Thread ID is too large or clock is too large. */
+               if (threadid > MAXTHREADID || ourClock > MAXWRITEVECTOR) {
+                       expandRecord(shadow);
+                       race = fullRaceCheckWrite(thread, location, shadow, currClock);
+                       goto Exit;
+               }
+
+               {
+                       /* Check for datarace against last read. */
+                       modelclock_t readClock = READVECTOR(shadowval);
+                       thread_id_t readThread = int_to_id(RDTHREADID(shadowval));
+
+                       if (clock_may_race(currClock, thread, readClock, readThread)) {
+                               /* We have a datarace */
+                               race = reportDataRace(readThread, readClock, false, get_execution()->get_parent_action(thread), true, location);
+                               goto ShadowExit;
+                       }
+               }
+
+               {
+                       /* Check for datarace against last write. */
+                       modelclock_t writeClock = WRITEVECTOR(shadowval);
+                       thread_id_t writeThread = int_to_id(WRTHREADID(shadowval));
+
+                       if (clock_may_race(currClock, thread, writeClock, writeThread)) {
+                               /* We have a datarace */
+                               race = reportDataRace(writeThread, writeClock, true, get_execution()->get_parent_action(thread), true, location);
+                               goto ShadowExit;
+                       }
+               }
+
+ShadowExit:
+               *shadow = ENCODEOP(0, 0, threadid, ourClock);
+       }
+
+Exit:
+       if (race) {
+               race->numframes=backtrace(race->backtrace, sizeof(race->backtrace)/sizeof(void*));
+               if (raceset->add(race))
+                       assert_race(race);
+               else model_free(race);
        }
+}
+
+/** This function does race detection for a write on an expanded record. */
+struct DataRace * atomfullRaceCheckWrite(thread_id_t thread, const void *location, uint64_t *shadow, ClockVector *currClock)
+{
+       struct RaceRecord *record = (struct RaceRecord *)(*shadow);
+       struct DataRace * race = NULL;
+
+       if (record->isAtomic)
+               goto Exit;
+
+       /* Check for datarace against last read. */
 
+       for (int i = 0;i < record->numReads;i++) {
+               modelclock_t readClock = record->readClock[i];
+               thread_id_t readThread = record->thread[i];
+
+               /* Note that readClock can't actuall be zero here, so it could be
+                        optimized. */
+
+               if (clock_may_race(currClock, thread, readClock, readThread)) {
+                       /* We have a datarace */
+                       race = reportDataRace(readThread, readClock, false, get_execution()->get_parent_action(thread), true, location);
+                       goto Exit;
+               }
+       }
+
+       /* Check for datarace against last write. */
+
+       {
+               modelclock_t writeClock = record->writeClock;
+               thread_id_t writeThread = record->writeThread;
+
+               if (clock_may_race(currClock, thread, writeClock, writeThread)) {
+                       /* We have a datarace */
+                       race = reportDataRace(writeThread, writeClock, true, get_execution()->get_parent_action(thread), true, location);
+                       goto Exit;
+               }
+       }
+Exit:
        record->numReads = 0;
        record->writeThread = thread;
+       record->isAtomic = 1;
        modelclock_t ourClock = currClock->getClock(thread);
        record->writeClock = ourClock;
+       return race;
 }
 
 /** This function does race detection on a write. */
-void raceCheckWrite(thread_id_t thread, void *location)
+void atomraceCheckWrite(thread_id_t thread, void *location)
 {
        uint64_t *shadow = lookupAddressEntry(location);
        uint64_t shadowval = *shadow;
        ClockVector *currClock = get_execution()->get_cv(thread);
+       if (currClock == NULL)
+               return;
 
+       struct DataRace * race = NULL;
        /* Do full record */
        if (shadowval != 0 && !ISSHORTRECORD(shadowval)) {
-               fullRaceCheckWrite(thread, location, shadow, currClock);
+               race = atomfullRaceCheckWrite(thread, location, shadow, currClock);
+               goto Exit;
+       }
+
+       {
+               int threadid = id_to_int(thread);
+               modelclock_t ourClock = currClock->getClock(thread);
+
+               /* Thread ID is too large or clock is too large. */
+               if (threadid > MAXTHREADID || ourClock > MAXWRITEVECTOR) {
+                       expandRecord(shadow);
+                       race = atomfullRaceCheckWrite(thread, location, shadow, currClock);
+                       goto Exit;
+               }
+
+               /* Can't race with atomic */
+               if (shadowval & ATOMICMASK)
+                       goto ShadowExit;
+
+               {
+                       /* Check for datarace against last read. */
+                       modelclock_t readClock = READVECTOR(shadowval);
+                       thread_id_t readThread = int_to_id(RDTHREADID(shadowval));
+
+                       if (clock_may_race(currClock, thread, readClock, readThread)) {
+                               /* We have a datarace */
+                               race = reportDataRace(readThread, readClock, false, get_execution()->get_parent_action(thread), true, location);
+                               goto ShadowExit;
+                       }
+               }
+
+               {
+                       /* Check for datarace against last write. */
+                       modelclock_t writeClock = WRITEVECTOR(shadowval);
+                       thread_id_t writeThread = int_to_id(WRTHREADID(shadowval));
+
+                       if (clock_may_race(currClock, thread, writeClock, writeThread)) {
+                               /* We have a datarace */
+                               race = reportDataRace(writeThread, writeClock, true, get_execution()->get_parent_action(thread), true, location);
+                               goto ShadowExit;
+                       }
+               }
+
+ShadowExit:
+               *shadow = ENCODEOP(0, 0, threadid, ourClock) | ATOMICMASK;
+       }
+
+Exit:
+       if (race) {
+               race->numframes=backtrace(race->backtrace, sizeof(race->backtrace)/sizeof(void*));
+               if (raceset->add(race))
+                       assert_race(race);
+               else model_free(race);
+       }
+}
+
+/** This function does race detection for a write on an expanded record. */
+void fullRecordWrite(thread_id_t thread, void *location, uint64_t *shadow, ClockVector *currClock) {
+       struct RaceRecord *record = (struct RaceRecord *)(*shadow);
+       record->numReads = 0;
+       record->writeThread = thread;
+       modelclock_t ourClock = currClock->getClock(thread);
+       record->writeClock = ourClock;
+       record->isAtomic = 1;
+}
+
+/** This function does race detection for a write on an expanded record. */
+void fullRecordWriteNonAtomic(thread_id_t thread, void *location, uint64_t *shadow, ClockVector *currClock) {
+       struct RaceRecord *record = (struct RaceRecord *)(*shadow);
+       record->numReads = 0;
+       record->writeThread = thread;
+       modelclock_t ourClock = currClock->getClock(thread);
+       record->writeClock = ourClock;
+       record->isAtomic = 0;
+}
+
+/** This function just updates metadata on atomic write. */
+void recordWrite(thread_id_t thread, void *location) {
+       uint64_t *shadow = lookupAddressEntry(location);
+       uint64_t shadowval = *shadow;
+       ClockVector *currClock = get_execution()->get_cv(thread);
+       /* Do full record */
+       if (shadowval != 0 && !ISSHORTRECORD(shadowval)) {
+               fullRecordWrite(thread, location, shadow, currClock);
                return;
        }
 
@@ -226,37 +467,46 @@ void raceCheckWrite(thread_id_t thread, void *location)
        /* Thread ID is too large or clock is too large. */
        if (threadid > MAXTHREADID || ourClock > MAXWRITEVECTOR) {
                expandRecord(shadow);
-               fullRaceCheckWrite(thread, location, shadow, currClock);
+               fullRecordWrite(thread, location, shadow, currClock);
                return;
        }
 
-       /* Check for datarace against last read. */
-
-       modelclock_t readClock = READVECTOR(shadowval);
-       thread_id_t readThread = int_to_id(RDTHREADID(shadowval));
+       *shadow = ENCODEOP(0, 0, threadid, ourClock) | ATOMICMASK;
+}
 
-       if (clock_may_race(currClock, thread, readClock, readThread)) {
-               /* We have a datarace */
-               reportDataRace(readThread, readClock, false, get_execution()->get_parent_action(thread), true, location);
-       }
+/** This function just updates metadata on atomic write. */
+void recordCalloc(void *location, size_t size) {
+       thread_id_t thread = thread_current()->get_id();
+       for(;size != 0;size--) {
+               uint64_t *shadow = lookupAddressEntry(location);
+               uint64_t shadowval = *shadow;
+               ClockVector *currClock = get_execution()->get_cv(thread);
+               /* Do full record */
+               if (shadowval != 0 && !ISSHORTRECORD(shadowval)) {
+                       fullRecordWriteNonAtomic(thread, location, shadow, currClock);
+                       return;
+               }
 
-       /* Check for datarace against last write. */
+               int threadid = id_to_int(thread);
+               modelclock_t ourClock = currClock->getClock(thread);
 
-       modelclock_t writeClock = WRITEVECTOR(shadowval);
-       thread_id_t writeThread = int_to_id(WRTHREADID(shadowval));
+               /* Thread ID is too large or clock is too large. */
+               if (threadid > MAXTHREADID || ourClock > MAXWRITEVECTOR) {
+                       expandRecord(shadow);
+                       fullRecordWriteNonAtomic(thread, location, shadow, currClock);
+                       return;
+               }
 
-       if (clock_may_race(currClock, thread, writeClock, writeThread)) {
-               /* We have a datarace */
-               reportDataRace(writeThread, writeClock, true, get_execution()->get_parent_action(thread), true, location);
+               *shadow = ENCODEOP(0, 0, threadid, ourClock);
+               location = (void *)(((char *) location) + 1);
        }
-       *shadow = ENCODEOP(0, 0, threadid, ourClock);
 }
 
 /** This function does race detection on a read for an expanded record. */
-void fullRaceCheckRead(thread_id_t thread, const void *location, uint64_t *shadow, ClockVector *currClock)
+struct DataRace * fullRaceCheckRead(thread_id_t thread, const void *location, uint64_t *shadow, ClockVector *currClock)
 {
        struct RaceRecord *record = (struct RaceRecord *) (*shadow);
-
+       struct DataRace * race = NULL;
        /* Check for datarace against last write. */
 
        modelclock_t writeClock = record->writeClock;
@@ -264,7 +514,7 @@ void fullRaceCheckRead(thread_id_t thread, const void *location, uint64_t *shado
 
        if (clock_may_race(currClock, thread, writeClock, writeThread)) {
                /* We have a datarace */
-               reportDataRace(writeThread, writeClock, true, get_execution()->get_parent_action(thread), false, location);
+               race = reportDataRace(writeThread, writeClock, true, get_execution()->get_parent_action(thread), false, location);
        }
 
        /* Shorten vector when possible */
@@ -275,7 +525,7 @@ void fullRaceCheckRead(thread_id_t thread, const void *location, uint64_t *shado
                modelclock_t readClock = record->readClock[i];
                thread_id_t readThread = record->thread[i];
 
-               /*  Note that is not really a datarace check as reads cannott
+               /*  Note that is not really a datarace check as reads cannot
                                actually race.  It is just determining that this read subsumes
                                another in the sense that either this read races or neither
                                read races. Note that readClock can't actually be zero, so it
@@ -284,6 +534,7 @@ void fullRaceCheckRead(thread_id_t thread, const void *location, uint64_t *shado
                if (clock_may_race(currClock, thread, readClock, readThread)) {
                        /* Still need this read in vector */
                        if (copytoindex != i) {
+                               ASSERT(record->thread[i] >= 0);
                                record->readClock[copytoindex] = record->readClock[i];
                                record->thread[copytoindex] = record->thread[i];
                        }
@@ -291,24 +542,31 @@ void fullRaceCheckRead(thread_id_t thread, const void *location, uint64_t *shado
                }
        }
 
-       if (copytoindex >= record->capacity) {
-               int newCapacity = record->capacity * 2;
-               thread_id_t *newthread = (thread_id_t *)snapshot_malloc(sizeof(thread_id_t) * newCapacity);
-               modelclock_t *newreadClock = (modelclock_t *)snapshot_malloc(sizeof(modelclock_t) * newCapacity);
-               std::memcpy(newthread, record->thread, record->capacity * sizeof(thread_id_t));
-               std::memcpy(newreadClock, record->readClock, record->capacity * sizeof(modelclock_t));
-               snapshot_free(record->readClock);
-               snapshot_free(record->thread);
-               record->readClock = newreadClock;
-               record->thread = newthread;
-               record->capacity = newCapacity;
+       if (__builtin_popcount(copytoindex) <= 1) {
+               if (copytoindex == 0 && record->thread == NULL) {
+                       int newCapacity = INITCAPACITY;
+                       record->thread = (thread_id_t *)snapshot_malloc(sizeof(thread_id_t) * newCapacity);
+                       record->readClock = (modelclock_t *)snapshot_malloc(sizeof(modelclock_t) * newCapacity);
+               } else if (copytoindex>=INITCAPACITY) {
+                       int newCapacity = copytoindex * 2;
+                       thread_id_t *newthread = (thread_id_t *)snapshot_malloc(sizeof(thread_id_t) * newCapacity);
+                       modelclock_t *newreadClock = (modelclock_t *)snapshot_malloc(sizeof(modelclock_t) * newCapacity);
+                       std::memcpy(newthread, record->thread, copytoindex * sizeof(thread_id_t));
+                       std::memcpy(newreadClock, record->readClock, copytoindex * sizeof(modelclock_t));
+                       snapshot_free(record->readClock);
+                       snapshot_free(record->thread);
+                       record->readClock = newreadClock;
+                       record->thread = newthread;
+               }
        }
 
        modelclock_t ourClock = currClock->getClock(thread);
 
+       ASSERT(thread >= 0);
        record->thread[copytoindex] = thread;
        record->readClock[copytoindex] = ourClock;
        record->numReads = copytoindex + 1;
+       return race;
 }
 
 /** This function does race detection on a read. */
@@ -317,47 +575,577 @@ void raceCheckRead(thread_id_t thread, const void *location)
        uint64_t *shadow = lookupAddressEntry(location);
        uint64_t shadowval = *shadow;
        ClockVector *currClock = get_execution()->get_cv(thread);
+       if (currClock == NULL)
+               return;
+
+       struct DataRace * race = NULL;
 
        /* Do full record */
        if (shadowval != 0 && !ISSHORTRECORD(shadowval)) {
-               fullRaceCheckRead(thread, location, shadow, currClock);
-               return;
+               race = fullRaceCheckRead(thread, location, shadow, currClock);
+               goto Exit;
        }
 
-       int threadid = id_to_int(thread);
-       modelclock_t ourClock = currClock->getClock(thread);
+       {
+               int threadid = id_to_int(thread);
+               modelclock_t ourClock = currClock->getClock(thread);
 
-       /* Thread ID is too large or clock is too large. */
-       if (threadid > MAXTHREADID || ourClock > MAXWRITEVECTOR) {
-               expandRecord(shadow);
-               fullRaceCheckRead(thread, location, shadow, currClock);
-               return;
+               /* Thread ID is too large or clock is too large. */
+               if (threadid > MAXTHREADID || ourClock > MAXWRITEVECTOR) {
+                       expandRecord(shadow);
+                       race = fullRaceCheckRead(thread, location, shadow, currClock);
+                       goto Exit;
+               }
+
+               /* Check for datarace against last write. */
+
+               modelclock_t writeClock = WRITEVECTOR(shadowval);
+               thread_id_t writeThread = int_to_id(WRTHREADID(shadowval));
+
+               if (clock_may_race(currClock, thread, writeClock, writeThread)) {
+                       /* We have a datarace */
+                       race = reportDataRace(writeThread, writeClock, true, get_execution()->get_parent_action(thread), false, location);
+                       goto ShadowExit;
+               }
+
+ShadowExit:
+               {
+                       modelclock_t readClock = READVECTOR(shadowval);
+                       thread_id_t readThread = int_to_id(RDTHREADID(shadowval));
+
+                       if (clock_may_race(currClock, thread, readClock, readThread)) {
+                               /* We don't subsume this read... Have to expand record. */
+                               expandRecord(shadow);
+                               fullRaceCheckRead(thread, location, shadow, currClock);
+                               goto Exit;
+                       }
+               }
+
+               *shadow = ENCODEOP(threadid, ourClock, id_to_int(writeThread), writeClock) | (shadowval & ATOMICMASK);
+       }
+Exit:
+       if (race) {
+               race->numframes=backtrace(race->backtrace, sizeof(race->backtrace)/sizeof(void*));
+               if (raceset->add(race))
+                       assert_race(race);
+               else model_free(race);
        }
+}
+
 
+/** This function does race detection on a read for an expanded record. */
+struct DataRace * atomfullRaceCheckRead(thread_id_t thread, const void *location, uint64_t *shadow, ClockVector *currClock)
+{
+       struct RaceRecord *record = (struct RaceRecord *) (*shadow);
+       struct DataRace * race = NULL;
        /* Check for datarace against last write. */
+       if (record->isAtomic)
+               return NULL;
 
-       modelclock_t writeClock = WRITEVECTOR(shadowval);
-       thread_id_t writeThread = int_to_id(WRTHREADID(shadowval));
+       modelclock_t writeClock = record->writeClock;
+       thread_id_t writeThread = record->writeThread;
 
        if (clock_may_race(currClock, thread, writeClock, writeThread)) {
                /* We have a datarace */
-               reportDataRace(writeThread, writeClock, true, get_execution()->get_parent_action(thread), false, location);
+               race = reportDataRace(writeThread, writeClock, true, get_execution()->get_parent_action(thread), false, location);
        }
+       return race;
+}
 
-       modelclock_t readClock = READVECTOR(shadowval);
-       thread_id_t readThread = int_to_id(RDTHREADID(shadowval));
+/** This function does race detection on a read. */
+void atomraceCheckRead(thread_id_t thread, const void *location)
+{
+       uint64_t *shadow = lookupAddressEntry(location);
+       uint64_t shadowval = *shadow;
+       ClockVector *currClock = get_execution()->get_cv(thread);
+       if (currClock == NULL)
+               return;
 
-       if (clock_may_race(currClock, thread, readClock, readThread)) {
-               /* We don't subsume this read... Have to expand record. */
-               expandRecord(shadow);
-               fullRaceCheckRead(thread, location, shadow, currClock);
+       struct DataRace * race = NULL;
+
+       /* Do full record */
+       if (shadowval != 0 && !ISSHORTRECORD(shadowval)) {
+               race = atomfullRaceCheckRead(thread, location, shadow, currClock);
+               goto Exit;
+       }
+
+       if (shadowval & ATOMICMASK)
+               return;
+
+       {
+               /* Check for datarace against last write. */
+               modelclock_t writeClock = WRITEVECTOR(shadowval);
+               thread_id_t writeThread = int_to_id(WRTHREADID(shadowval));
+
+               if (clock_may_race(currClock, thread, writeClock, writeThread)) {
+                       /* We have a datarace */
+                       race = reportDataRace(writeThread, writeClock, true, get_execution()->get_parent_action(thread), false, location);
+                       goto Exit;
+               }
+
+
+       }
+Exit:
+       if (race) {
+               race->numframes=backtrace(race->backtrace, sizeof(race->backtrace)/sizeof(void*));
+               if (raceset->add(race))
+                       assert_race(race);
+               else model_free(race);
+       }
+}
+
+static inline uint64_t * raceCheckRead_firstIt(thread_id_t thread, const void * location, uint64_t *old_val, uint64_t *new_val)
+{
+       uint64_t *shadow = lookupAddressEntry(location);
+       uint64_t shadowval = *shadow;
+
+       ClockVector *currClock = get_execution()->get_cv(thread);
+       if (currClock == NULL)
+               return shadow;
+
+       struct DataRace * race = NULL;
+
+       /* Do full record */
+       if (shadowval != 0 && !ISSHORTRECORD(shadowval)) {
+               race = fullRaceCheckRead(thread, location, shadow, currClock);
+               goto Exit;
+       }
+
+       {
+               int threadid = id_to_int(thread);
+               modelclock_t ourClock = currClock->getClock(thread);
+
+               /* Thread ID is too large or clock is too large. */
+               if (threadid > MAXTHREADID || ourClock > MAXWRITEVECTOR) {
+                       expandRecord(shadow);
+                       race = fullRaceCheckRead(thread, location, shadow, currClock);
+                       goto Exit;
+               }
+
+               /* Check for datarace against last write. */
+               modelclock_t writeClock = WRITEVECTOR(shadowval);
+               thread_id_t writeThread = int_to_id(WRTHREADID(shadowval));
+
+               if (clock_may_race(currClock, thread, writeClock, writeThread)) {
+                       /* We have a datarace */
+                       race = reportDataRace(writeThread, writeClock, true, get_execution()->get_parent_action(thread), false, location);
+               }
+
+               modelclock_t readClock = READVECTOR(shadowval);
+               thread_id_t readThread = int_to_id(RDTHREADID(shadowval));
+
+               if (clock_may_race(currClock, thread, readClock, readThread)) {
+                       /* We don't subsume this read... Have to expand record. */
+                       expandRecord(shadow);
+                       struct RaceRecord *record = (struct RaceRecord *) (*shadow);
+                       record->thread[1] = thread;
+                       record->readClock[1] = ourClock;
+                       record->numReads++;
+
+                       goto Exit;
+               }
+
+               *shadow = ENCODEOP(threadid, ourClock, id_to_int(writeThread), writeClock) | (shadowval & ATOMICMASK);
+
+               *old_val = shadowval;
+               *new_val = *shadow;
+       }
+Exit:
+       if (race) {
+               race->numframes=backtrace(race->backtrace, sizeof(race->backtrace)/sizeof(void*));
+               if (raceset->add(race))
+                       assert_race(race);
+               else model_free(race);
+       }
+
+       return shadow;
+}
+
+static inline void raceCheckRead_otherIt(thread_id_t thread, const void * location) {
+       uint64_t *shadow = lookupAddressEntry(location);
+
+       uint64_t shadowval = *shadow;
+
+       ClockVector *currClock = get_execution()->get_cv(thread);
+       if (currClock == NULL)
+               return;
+
+       struct DataRace * race = NULL;
+
+       /* Do full record */
+       if (shadowval != 0 && !ISSHORTRECORD(shadowval)) {
+               race = fullRaceCheckRead(thread, location, shadow, currClock);
+               goto Exit;
+       }
+
+       {
+               int threadid = id_to_int(thread);
+               modelclock_t ourClock = currClock->getClock(thread);
+
+               /* Thread ID is too large or clock is too large. */
+               if (threadid > MAXTHREADID || ourClock > MAXWRITEVECTOR) {
+                       expandRecord(shadow);
+                       race = fullRaceCheckRead(thread, location, shadow, currClock);
+                       goto Exit;
+               }
+
+               /* Check for datarace against last write. */
+               modelclock_t writeClock = WRITEVECTOR(shadowval);
+               thread_id_t writeThread = int_to_id(WRTHREADID(shadowval));
+
+               if (clock_may_race(currClock, thread, writeClock, writeThread)) {
+                       /* We have a datarace */
+                       race = reportDataRace(writeThread, writeClock, true, get_execution()->get_parent_action(thread), false, location);
+               }
+
+               modelclock_t readClock = READVECTOR(shadowval);
+               thread_id_t readThread = int_to_id(RDTHREADID(shadowval));
+
+               if (clock_may_race(currClock, thread, readClock, readThread)) {
+                       /* We don't subsume this read... Have to expand record. */
+                       expandRecord(shadow);
+                       struct RaceRecord *record = (struct RaceRecord *) (*shadow);
+                       record->thread[1] = thread;
+                       record->readClock[1] = ourClock;
+                       record->numReads++;
+
+                       goto Exit;
+               }
+
+               *shadow = ENCODEOP(threadid, ourClock, id_to_int(writeThread), writeClock) | (shadowval & ATOMICMASK);
+       }
+Exit:
+       if (race) {
+               race->numframes=backtrace(race->backtrace, sizeof(race->backtrace)/sizeof(void*));
+               if (raceset->add(race))
+                       assert_race(race);
+               else model_free(race);
+       }
+}
+
+void raceCheckRead64(thread_id_t thread, const void *location)
+{
+       uint64_t old_shadowval, new_shadowval;
+       old_shadowval = new_shadowval = INVALIDSHADOWVAL;
+
+       uint64_t * shadow = raceCheckRead_firstIt(thread, location, &old_shadowval, &new_shadowval);
+       if (CHECKBOUNDARY(location, 7)) {
+               if (shadow[1]==old_shadowval)
+                       shadow[1] = new_shadowval;
+               else goto L1;
+               if (shadow[2]==old_shadowval)
+                       shadow[2] = new_shadowval;
+               else goto L2;
+               if (shadow[3]==old_shadowval)
+                       shadow[3] = new_shadowval;
+               else goto L3;
+               if (shadow[4]==old_shadowval)
+                       shadow[4] = new_shadowval;
+               else goto L4;
+               if (shadow[5]==old_shadowval)
+                       shadow[5] = new_shadowval;
+               else goto L5;
+               if (shadow[6]==old_shadowval)
+                       shadow[6] = new_shadowval;
+               else goto L6;
+               if (shadow[7]==old_shadowval)
+                       shadow[7] = new_shadowval;
+               else goto L7;
+               return;
+       }
+
+L1:
+       raceCheckRead_otherIt(thread, (const void *)(((uintptr_t)location) + 1));
+L2:
+       raceCheckRead_otherIt(thread, (const void *)(((uintptr_t)location) + 2));
+L3:
+       raceCheckRead_otherIt(thread, (const void *)(((uintptr_t)location) + 3));
+L4:
+       raceCheckRead_otherIt(thread, (const void *)(((uintptr_t)location) + 4));
+L5:
+       raceCheckRead_otherIt(thread, (const void *)(((uintptr_t)location) + 5));
+L6:
+       raceCheckRead_otherIt(thread, (const void *)(((uintptr_t)location) + 6));
+L7:
+       raceCheckRead_otherIt(thread, (const void *)(((uintptr_t)location) + 7));
+}
+
+void raceCheckRead32(thread_id_t thread, const void *location)
+{
+       uint64_t old_shadowval, new_shadowval;
+       old_shadowval = new_shadowval = INVALIDSHADOWVAL;
+
+       uint64_t * shadow = raceCheckRead_firstIt(thread, location, &old_shadowval, &new_shadowval);
+       if (CHECKBOUNDARY(location, 3)) {
+               if (shadow[1]==old_shadowval)
+                       shadow[1] = new_shadowval;
+               else goto L1;
+               if (shadow[2]==old_shadowval)
+                       shadow[2] = new_shadowval;
+               else goto L2;
+               if (shadow[3]==old_shadowval)
+                       shadow[3] = new_shadowval;
+               else goto L3;
                return;
        }
 
-       *shadow = ENCODEOP(threadid, ourClock, id_to_int(writeThread), writeClock);
+L1:
+       raceCheckRead_otherIt(thread, (const void *)(((uintptr_t)location) + 1));
+L2:
+       raceCheckRead_otherIt(thread, (const void *)(((uintptr_t)location) + 2));
+L3:
+       raceCheckRead_otherIt(thread, (const void *)(((uintptr_t)location) + 3));
+}
+
+void raceCheckRead16(thread_id_t thread, const void *location)
+{
+       uint64_t old_shadowval, new_shadowval;
+       old_shadowval = new_shadowval = INVALIDSHADOWVAL;
+
+
+       uint64_t * shadow = raceCheckRead_firstIt(thread, location, &old_shadowval, &new_shadowval);
+       if (CHECKBOUNDARY(location, 1)) {
+               if (shadow[1]==old_shadowval) {
+                       shadow[1] = new_shadowval;
+                       return;
+               }
+       }
+       raceCheckRead_otherIt(thread, (const void *)(((uintptr_t)location) + 1));
 }
 
-bool haveUnrealizedRaces()
+void raceCheckRead8(thread_id_t thread, const void *location)
 {
-       return !unrealizedraces->empty();
+       uint64_t old_shadowval, new_shadowval;
+       old_shadowval = new_shadowval = INVALIDSHADOWVAL;
+
+       raceCheckRead_firstIt(thread, location, &old_shadowval, &new_shadowval);
+}
+
+static inline uint64_t * raceCheckWrite_firstIt(thread_id_t thread, const void * location, uint64_t *old_val, uint64_t *new_val)
+{
+       uint64_t *shadow = lookupAddressEntry(location);
+       uint64_t shadowval = *shadow;
+       ClockVector *currClock = get_execution()->get_cv(thread);
+       if (currClock == NULL)
+               return shadow;
+
+       struct DataRace * race = NULL;
+       /* Do full record */
+       if (shadowval != 0 && !ISSHORTRECORD(shadowval)) {
+               race = fullRaceCheckWrite(thread, location, shadow, currClock);
+               goto Exit;
+       }
+
+       {
+               int threadid = id_to_int(thread);
+               modelclock_t ourClock = currClock->getClock(thread);
+
+               /* Thread ID is too large or clock is too large. */
+               if (threadid > MAXTHREADID || ourClock > MAXWRITEVECTOR) {
+                       expandRecord(shadow);
+                       race = fullRaceCheckWrite(thread, location, shadow, currClock);
+                       goto Exit;
+               }
+
+               {
+                       /* Check for datarace against last read. */
+                       modelclock_t readClock = READVECTOR(shadowval);
+                       thread_id_t readThread = int_to_id(RDTHREADID(shadowval));
+
+                       if (clock_may_race(currClock, thread, readClock, readThread)) {
+                               /* We have a datarace */
+                               race = reportDataRace(readThread, readClock, false, get_execution()->get_parent_action(thread), true, location);
+                               goto ShadowExit;
+                       }
+               }
+
+               {
+                       /* Check for datarace against last write. */
+                       modelclock_t writeClock = WRITEVECTOR(shadowval);
+                       thread_id_t writeThread = int_to_id(WRTHREADID(shadowval));
+
+                       if (clock_may_race(currClock, thread, writeClock, writeThread)) {
+                               /* We have a datarace */
+                               race = reportDataRace(writeThread, writeClock, true, get_execution()->get_parent_action(thread), true, location);
+                               goto ShadowExit;
+                       }
+               }
+
+ShadowExit:
+               *shadow = ENCODEOP(0, 0, threadid, ourClock);
+
+               *old_val = shadowval;
+               *new_val = *shadow;
+       }
+
+Exit:
+       if (race) {
+               race->numframes=backtrace(race->backtrace, sizeof(race->backtrace)/sizeof(void*));
+               if (raceset->add(race))
+                       assert_race(race);
+               else model_free(race);
+       }
+
+       return shadow;
+}
+
+static inline void raceCheckWrite_otherIt(thread_id_t thread, const void * location) {
+       uint64_t *shadow = lookupAddressEntry(location);
+
+       uint64_t shadowval = *shadow;
+
+       ClockVector *currClock = get_execution()->get_cv(thread);
+       if (currClock == NULL)
+               return;
+
+       struct DataRace * race = NULL;
+       /* Do full record */
+       if (shadowval != 0 && !ISSHORTRECORD(shadowval)) {
+               race = fullRaceCheckWrite(thread, location, shadow, currClock);
+               goto Exit;
+       }
+
+       {
+               int threadid = id_to_int(thread);
+               modelclock_t ourClock = currClock->getClock(thread);
+
+               /* Thread ID is too large or clock is too large. */
+               if (threadid > MAXTHREADID || ourClock > MAXWRITEVECTOR) {
+                       expandRecord(shadow);
+                       race = fullRaceCheckWrite(thread, location, shadow, currClock);
+                       goto Exit;
+               }
+
+               {
+                       /* Check for datarace against last read. */
+                       modelclock_t readClock = READVECTOR(shadowval);
+                       thread_id_t readThread = int_to_id(RDTHREADID(shadowval));
+
+                       if (clock_may_race(currClock, thread, readClock, readThread)) {
+                               /* We have a datarace */
+                               race = reportDataRace(readThread, readClock, false, get_execution()->get_parent_action(thread), true, location);
+                               goto ShadowExit;
+                       }
+               }
+
+               {
+                       /* Check for datarace against last write. */
+                       modelclock_t writeClock = WRITEVECTOR(shadowval);
+                       thread_id_t writeThread = int_to_id(WRTHREADID(shadowval));
+
+                       if (clock_may_race(currClock, thread, writeClock, writeThread)) {
+                               /* We have a datarace */
+                               race = reportDataRace(writeThread, writeClock, true, get_execution()->get_parent_action(thread), true, location);
+                               goto ShadowExit;
+                       }
+               }
+
+ShadowExit:
+               *shadow = ENCODEOP(0, 0, threadid, ourClock);
+       }
+
+Exit:
+       if (race) {
+               race->numframes=backtrace(race->backtrace, sizeof(race->backtrace)/sizeof(void*));
+               if (raceset->add(race))
+                       assert_race(race);
+               else model_free(race);
+       }
+}
+
+void raceCheckWrite64(thread_id_t thread, const void *location)
+{
+       uint64_t old_shadowval, new_shadowval;
+       old_shadowval = new_shadowval = INVALIDSHADOWVAL;
+
+       uint64_t * shadow = raceCheckWrite_firstIt(thread, location, &old_shadowval, &new_shadowval);
+       if (CHECKBOUNDARY(location, 7)) {
+               if (shadow[1]==old_shadowval)
+                       shadow[1] = new_shadowval;
+               else goto L1;
+               if (shadow[2]==old_shadowval)
+                       shadow[2] = new_shadowval;
+               else goto L2;
+               if (shadow[3]==old_shadowval)
+                       shadow[3] = new_shadowval;
+               else goto L3;
+               if (shadow[4]==old_shadowval)
+                       shadow[4] = new_shadowval;
+               else goto L4;
+               if (shadow[5]==old_shadowval)
+                       shadow[5] = new_shadowval;
+               else goto L5;
+               if (shadow[6]==old_shadowval)
+                       shadow[6] = new_shadowval;
+               else goto L6;
+               if (shadow[7]==old_shadowval)
+                       shadow[7] = new_shadowval;
+               else goto L7;
+               return;
+       }
+
+L1:
+       raceCheckWrite_otherIt(thread, (const void *)(((uintptr_t)location) + 1));
+L2:
+       raceCheckWrite_otherIt(thread, (const void *)(((uintptr_t)location) + 2));
+L3:
+       raceCheckWrite_otherIt(thread, (const void *)(((uintptr_t)location) + 3));
+L4:
+       raceCheckWrite_otherIt(thread, (const void *)(((uintptr_t)location) + 4));
+L5:
+       raceCheckWrite_otherIt(thread, (const void *)(((uintptr_t)location) + 5));
+L6:
+       raceCheckWrite_otherIt(thread, (const void *)(((uintptr_t)location) + 6));
+L7:
+       raceCheckWrite_otherIt(thread, (const void *)(((uintptr_t)location) + 7));
+}
+
+void raceCheckWrite32(thread_id_t thread, const void *location)
+{
+       uint64_t old_shadowval, new_shadowval;
+       old_shadowval = new_shadowval = INVALIDSHADOWVAL;
+
+       uint64_t * shadow = raceCheckWrite_firstIt(thread, location, &old_shadowval, &new_shadowval);
+       if (CHECKBOUNDARY(location, 3)) {
+               if (shadow[1]==old_shadowval)
+                       shadow[1] = new_shadowval;
+               else goto L1;
+               if (shadow[2]==old_shadowval)
+                       shadow[2] = new_shadowval;
+               else goto L2;
+               if (shadow[3]==old_shadowval)
+                       shadow[3] = new_shadowval;
+               else goto L3;
+               return;
+       }
+
+L1:
+       raceCheckWrite_otherIt(thread, (const void *)(((uintptr_t)location) + 1));
+L2:
+       raceCheckWrite_otherIt(thread, (const void *)(((uintptr_t)location) + 2));
+L3:
+       raceCheckWrite_otherIt(thread, (const void *)(((uintptr_t)location) + 3));
+}
+
+void raceCheckWrite16(thread_id_t thread, const void *location)
+{
+       uint64_t old_shadowval, new_shadowval;
+       old_shadowval = new_shadowval = INVALIDSHADOWVAL;
+
+       uint64_t * shadow = raceCheckWrite_firstIt(thread, location, &old_shadowval, &new_shadowval);
+       if (CHECKBOUNDARY(location, 1)) {
+               if (shadow[1]==old_shadowval) {
+                       shadow[1] = new_shadowval;
+                       return;
+               }
+       }
+       raceCheckWrite_otherIt(thread, (const void *)(((uintptr_t)location) + 1));
+}
+
+void raceCheckWrite8(thread_id_t thread, const void *location)
+{
+       uint64_t old_shadowval, new_shadowval;
+       old_shadowval = new_shadowval = INVALIDSHADOWVAL;
+
+       raceCheckWrite_firstIt(thread, location, &old_shadowval, &new_shadowval);
 }