// Tests if this needs to be scanned. It marks it if it has not been marked
// unless it has to be scanned.
bool MTGCProcessMarkPointers::TestForScan(PolyWord *pt)
{
if ((*pt).IsTagged())
return false;
// This could contain a forwarding pointer if it points into an
// allocation area and has been moved by the minor GC.
// We have to be a little careful. Another thread could also
// be following any forwarding pointers here. However it's safe
// because they will update it with the same value.
PolyObject *obj = (*pt).AsObjPtr();
if (obj->ContainsForwardingPtr())
{
obj = FollowForwarding(obj);
*pt = obj;
}
if (gMem.LocalSpaceForAddress(obj) == 0)
return false; // Ignore it if it points to a permanent area
POLYUNSIGNED L = obj->LengthWord();
if (L & _OBJ_GC_MARK)
return false; // Already marked
if (debugOptions & DEBUG_GC_DETAIL)
Log("GC: Mark: %p %" POLYUFMT " %u\n", obj, OBJ_OBJECT_LENGTH(L), GetTypeBits(L));
if (OBJ_IS_BYTE_OBJECT(L))
{
obj->SetLengthWord(L | _OBJ_GC_MARK); // Mark it
return false; // We've done as much as we need
}
return true;
}
// The initial entry to process the roots. These may be RTS addresses or addresses in
// a thread stack. Also called recursively to process the addresses of constants in
// code segments. This is used in situations where a scanner may return the
// updated address of an object.
PolyObject *MTGCProcessMarkPointers::ScanObjectAddress(PolyObject *obj)
{
PolyWord val = obj;
LocalMemSpace *space = gMem.LocalSpaceForAddress(val.AsAddress());
if (space == 0)
return obj; // Ignore it if it points to a permanent area
// We may have a forwarding pointer if this has been moved by the
// minor GC.
if (obj->ContainsForwardingPtr())
{
obj = FollowForwarding(obj);
val = obj;
space = gMem.LocalSpaceForAddress(val.AsAddress());
}
ASSERT(obj->ContainsNormalLengthWord());
POLYUNSIGNED L = obj->LengthWord();
if (L & _OBJ_GC_MARK)
return obj; // Already marked
obj->SetLengthWord(L | _OBJ_GC_MARK); // Mark it
if (profileMode == kProfileLiveData || (profileMode == kProfileLiveMutables && obj->IsMutable()))
AddObjectProfile(obj);
POLYUNSIGNED n = OBJ_OBJECT_LENGTH(L);
if (debugOptions & DEBUG_GC_DETAIL)
Log("GC: Mark: %p %" POLYUFMT " %u\n", obj, n, GetTypeBits(L));
if (OBJ_IS_BYTE_OBJECT(L))
return obj;
// If we already have something on the stack we must being called
// recursively to process a constant in a code segment. Just push
// it on the stack and let the caller deal with it.
if (msp != 0)
PushToStack(obj); // Can't check this because it may have forwarding ptrs.
else
{
MTGCProcessMarkPointers::ScanAddressesInObject(obj, L);
// We can only check after we've processed it because if we
// have addresses left over from an incomplete partial GC they
// may need to forwarded.
CheckObject (obj);
}
return obj;
}
// This is called via ScanAddressesInRegion to process the permanent mutables. It is
// also called from ScanObjectAddress to process root addresses.
// It processes all the addresses reachable from the object.
void RecursiveScan::ScanAddressesInObject(PolyObject *obj, POLYUNSIGNED lengthWord)
{
if (OBJ_IS_BYTE_OBJECT(lengthWord))
{
Completed(obj);
return;
}
while (true)
{
ASSERT (OBJ_IS_LENGTH(lengthWord));
// Get the length and base address. N.B. If this is a code segment
// these will be side-effected by GetConstSegmentForCode.
POLYUNSIGNED length = OBJ_OBJECT_LENGTH(lengthWord);
PolyWord *baseAddr = (PolyWord*)obj;
if (OBJ_IS_CODE_OBJECT(lengthWord))
{
// It's better to process the whole code object in one go.
ScanAddress::ScanAddressesInObject(obj, lengthWord);
length = 0; // Finished
}
ASSERT(! OBJ_IS_BYTE_OBJECT(lengthWord)); // Check - remove this later
// else it's a normal object,
// If there are only two addresses in this cell that need to be
// followed we follow them immediately and treat this cell as done.
// If there are more than two we push the address of this cell on
// the stack, follow the first address and then rescan it. That way
// list cells are processed once only but we don't overflow the
// stack by pushing all the addresses in a very large vector.
PolyWord *endWord = baseAddr + length;
PolyObject *firstWord = 0;
PolyObject *secondWord = 0;
while (baseAddr != endWord)
{
PolyWord wordAt = *baseAddr;
if (wordAt.IsDataPtr() && wordAt != PolyWord::FromUnsigned(0))
{
// Normal address. We can have words of all zeros at least in the
// situation where we have a partially constructed code segment where
// the constants at the end of the code have not yet been filled in.
if (TestForScan(baseAddr)) // Test value at baseAddr (may side-effect it)
{
PolyObject *wObj = (*baseAddr).AsObjPtr();
if (wObj->IsByteObject())
{
// Can do this now - don't need to push it
MarkAsScanning(wObj);
Completed(wObj);
}
else if (firstWord == 0)
{
firstWord = wObj;
// We mark the word immediately. We can have
// two words in an object that are the same
// and we don't want to process it again.
MarkAsScanning(firstWord);
}
else if (secondWord == 0)
secondWord = wObj;
else break; // More than two words.
}
}
else if (wordAt.IsCodePtr())
{
// If we're processing the constant area of a code segment this could
// be a code address.
PolyObject *oldObject = ObjCodePtrToPtr(wordAt.AsCodePtr());
// Calculate the byte offset of this value within the code object.
POLYUNSIGNED offset = wordAt.AsCodePtr() - (byte*)oldObject;
wordAt = oldObject;
bool test = TestForScan(&wordAt);
// TestForScan may side-effect the word.
PolyObject *newObject = wordAt.AsObjPtr();
wordAt = PolyWord::FromCodePtr((byte*)newObject + offset);
if (wordAt != *baseAddr)
*baseAddr = wordAt;
if (test)
{
if (firstWord == 0)
{
firstWord = newObject;
MarkAsScanning(firstWord);
}
else if (secondWord == 0)
secondWord = newObject;
else break;
}
}
baseAddr++;
}
if (baseAddr == endWord)
{
// We have done everything except possibly firstWord and secondWord.
Completed(obj);
if (secondWord != 0)
{
MarkAsScanning(secondWord);
// Put this on the stack. If this is a list node we will be
// pushing the tail.
PushToStack(secondWord);
}
}
else // Put this back on the stack while we process the first word
PushToStack(obj);
if (firstWord != 0)
// Process it immediately.
obj = firstWord;
else if (StackIsEmpty())
return;
else
obj = PopFromStack();
lengthWord = obj->LengthWord();
}
}
// General purpose object processor, Processes all the addresses in an object.
// Handles the various kinds of object that may contain addresses.
void ScanAddress::ScanAddressesInObject(PolyObject *obj, POLYUNSIGNED lengthWord)
{
do
{
ASSERT (OBJ_IS_LENGTH(lengthWord));
if (OBJ_IS_BYTE_OBJECT(lengthWord))
return; /* Nothing more to do */
POLYUNSIGNED length = OBJ_OBJECT_LENGTH(lengthWord);
PolyWord *baseAddr = (PolyWord*)obj;
if (OBJ_IS_CODE_OBJECT(lengthWord))
{
// Scan constants within the code.
machineDependent->ScanConstantsWithinCode(obj, obj, length, this);
// Skip to the constants and get ready to scan them.
obj->GetConstSegmentForCode(length, baseAddr, length);
} // else it's a normal object,
PolyWord *endWord = baseAddr + length;
// We want to minimise the actual recursion we perform so we try to
// use tail recursion if we can. We first scan from the end and
// remove any words that don't need recursion.
POLYUNSIGNED lastLengthWord = 0;
while (endWord != baseAddr)
{
PolyWord wordAt = endWord[-1];
if (IS_INT(wordAt) || wordAt == PolyWord::FromUnsigned(0))
endWord--; // Don't need to look at this.
else if ((lastLengthWord = ScanAddressAt(endWord-1)) != 0)
// We need to process this one
break;
else endWord--; // We're not interested in this.
}
if (endWord == baseAddr)
return; // We've done everything.
// There is at least one word that needs to be processed, the
// one at endWord-1.
// Now process from the beginning forward to see if there are
// any words before this that need to be handled. This way we are more
// likely to handle the head of a list by recursion and the
// tail by looping (tail recursion).
while (baseAddr < endWord-1)
{
PolyWord wordAt = *baseAddr;
if (IS_INT(wordAt) || wordAt == PolyWord::FromUnsigned(0))
baseAddr++; // Don't need to look at this.
else
{
POLYUNSIGNED lengthWord = ScanAddressAt(baseAddr);
if (lengthWord != 0)
{
wordAt = *baseAddr; // Reload because it may have been side-effected
// We really have to process this recursively.
if (wordAt.IsCodePtr())
ScanAddressesInObject(ObjCodePtrToPtr(wordAt.AsCodePtr()), lengthWord);
else
ScanAddressesInObject(wordAt.AsObjPtr(), lengthWord);
baseAddr++;
}
else baseAddr++;
}
}
// Finally process the last word we found that has to be processed.
// Do this by looping rather than recursion.
PolyWord wordAt = *baseAddr; // Last word to do.
// This must be an address
if (wordAt.IsCodePtr())
obj = ObjCodePtrToPtr(wordAt.AsCodePtr());
else
obj = wordAt.AsObjPtr();
lengthWord = lastLengthWord;
} while(1);
}
// This is called via ScanAddressesInRegion to process the permanent mutables. It is
// also called from ScanObjectAddress to process root addresses.
// It processes all the addresses reachable from the object.
void MTGCProcessMarkPointers::ScanAddressesInObject(PolyObject *obj, POLYUNSIGNED lengthWord)
{
if (OBJ_IS_BYTE_OBJECT(lengthWord))
return;
while (true)
{
ASSERT (OBJ_IS_LENGTH(lengthWord));
// Get the length and base address. N.B. If this is a code segment
// these will be side-effected by GetConstSegmentForCode.
POLYUNSIGNED length = OBJ_OBJECT_LENGTH(lengthWord);
if (OBJ_IS_WEAKREF_OBJECT(lengthWord))
{
// Special case.
ASSERT(OBJ_IS_MUTABLE_OBJECT(lengthWord)); // Should be a mutable.
ASSERT(OBJ_IS_WORD_OBJECT(lengthWord)); // Should be a plain object.
// We need to mark the "SOME" values in this object but we don't mark
// the references contained within the "SOME".
PolyWord *baseAddr = (PolyWord*)obj;
// Mark every word but ignore the result.
for (POLYUNSIGNED i = 0; i < length; i++)
(void)MarkAndTestForScan(baseAddr+i);
// We've finished with this.
length = 0;
}
else if (OBJ_IS_CODE_OBJECT(lengthWord))
{
// It's better to process the whole code object in one go.
ScanAddress::ScanAddressesInObject(obj, lengthWord);
length = 0; // Finished
}
// else it's a normal object,
// If there are only two addresses in this cell that need to be
// followed we follow them immediately and treat this cell as done.
// If there are more than two we push the address of this cell on
// the stack, follow the first address and then rescan it. That way
// list cells are processed once only but we don't overflow the
// stack by pushing all the addresses in a very large vector.
PolyWord *baseAddr = (PolyWord*)obj;
PolyWord *endWord = baseAddr + length;
PolyObject *firstWord = 0;
PolyObject *secondWord = 0;
PolyWord *restartAddr = 0;
if (obj == largeObjectCache[locPtr].base)
{
baseAddr = largeObjectCache[locPtr].current;
ASSERT(baseAddr > (PolyWord*)obj && baseAddr < ((PolyWord*)obj)+length);
if (locPtr == 0) locPtr = LARGECACHE_SIZE-1; else locPtr--;
}
while (baseAddr != endWord)
{
PolyWord wordAt = *baseAddr;
if (wordAt.IsDataPtr() && wordAt != PolyWord::FromUnsigned(0))
{
// Normal address. We can have words of all zeros at least in the
// situation where we have a partially constructed code segment where
// the constants at the end of the code have not yet been filled in.
if (TestForScan(baseAddr))
{
if (firstWord == 0)
firstWord = baseAddr->AsObjPtr();
else if (secondWord == 0)
{
// If we need to rescan because there are three or more words to do
// this is the place we need to restart (or the start of the cell if it's
// small).
restartAddr = baseAddr;
secondWord = baseAddr->AsObjPtr();
}
else break; // More than two words.
}
}
else if (wordAt.IsCodePtr())
{
// If we're processing the constant area of a code segment this could
// be a code address.
// Check that this is actually an address. If we have had a bad pointer
// earlier we may treat some length fields as values.
ASSERT(gMem.SpaceForAddress(wordAt.AsCodePtr()) != 0);
PolyObject *oldObject = ObjCodePtrToPtr(wordAt.AsCodePtr());
// Calculate the byte offset of this value within the code object.
POLYUNSIGNED offset = wordAt.AsCodePtr() - (byte*)oldObject;
wordAt = oldObject;
bool test = TestForScan(&wordAt);
// If we've changed it because we had a left-over forwarding pointer
// we need to update the original.
PolyObject *newObject = wordAt.AsObjPtr();
wordAt = PolyWord::FromCodePtr((byte*)newObject + offset);
if (wordAt != *baseAddr)
*baseAddr = wordAt;
if (test)
{
if (firstWord == 0)
firstWord = newObject;
else if (secondWord == 0)
{
restartAddr = baseAddr;
secondWord = newObject;
}
else break;
}
}
baseAddr++;
}
if (baseAddr != endWord)
// Put this back on the stack while we process the first word
PushToStack(obj, length < largeObjectSize ? 0 : restartAddr, length);
else if (secondWord != 0)
{
// Mark it now because we will process it.
secondWord->SetLengthWord(secondWord->LengthWord() | _OBJ_GC_MARK);
// Put this on the stack. If this is a list node we will be
// pushing the tail.
PushToStack(secondWord);
}
if (firstWord != 0)
{
// Mark it and process it immediately.
firstWord->SetLengthWord(firstWord->LengthWord() | _OBJ_GC_MARK);
obj = firstWord;
}
else if (msp == 0)
{
markStack[msp] = 0; // Really finished
return;
}
else
{
// Clear the item above the top. This really is finished.
if (msp < MARK_STACK_SIZE) markStack[msp] = 0;
// Pop the item from the stack but don't overwrite it yet.
// This allows another thread to steal it if there really
// is nothing else to do. This is only really important
// for large objects.
obj = markStack[--msp]; // Pop something.
}
lengthWord = obj->LengthWord();
}
}