// Process a value within the stack.
PolyWord ScanAddress::ScanStackAddress(PolyWord val, StackSpace *stack, bool isCode)
{
PolyWord *base = stack->bottom;
PolyWord *end = stack->top;
// If isCode is set we definitely have a code address. It may have the
// bottom bit set or it may be word aligned.
if (isCode || val.IsCodePtr())
{
/* Find the start of the code segment */
PolyObject *oldObject = ObjCodePtrToPtr(val.AsCodePtr());
// Calculate the byte offset of this value within the code object.
POLYUNSIGNED offset = val.AsCodePtr() - (byte*)oldObject;
PolyObject *newObject = ScanObjectAddress(oldObject);
return PolyWord::FromCodePtr((byte*)newObject + offset);
}
else if (val.IsTagged() || val == PolyWord::FromUnsigned(0) ||
(val.AsAddress() > base && val.AsAddress() <= end))
/* We don't need to process tagged integers (now we've checked it isn't
a code address) and we don't need to process addresses within the
current stack. */
/* N.B. We have "<= end" rather than "< end" because it is possible for
the stack to be completely empty on a terminated thread. */
return val;
else
{
ASSERT(val.IsDataPtr());
return ScanObjectAddress(val.AsObjPtr());
}
}
// Returns the new address if the argument is the address of an object that
// has moved, otherwise returns the original.
PolyWord SaveFixupAddress::GetNewAddress(PolyWord old)
{
if (old.IsTagged() || old == PolyWord::FromUnsigned(0) || gMem.IsIOPointer(old.AsAddress()))
return old; // Nothing to do.
// When we are updating addresses in the stack or in code segments we may have
// code pointers.
if (old.IsCodePtr())
{
// Find the start of the code segment
PolyObject *oldObject = ObjCodePtrToPtr(old.AsCodePtr());
// Calculate the byte offset of this value within the code object.
POLYUNSIGNED offset = old.AsCodePtr() - (byte*)oldObject;
PolyWord newObject = GetNewAddress(oldObject);
return PolyWord::FromCodePtr(newObject.AsCodePtr() + offset);
}
ASSERT(old.IsDataPtr());
PolyObject *obj = old.AsObjPtr();
if (obj->ContainsForwardingPtr()) // tombstone is a pointer to a moved object
{
PolyObject *newp = obj->GetForwardingPtr();
ASSERT (newp->ContainsNormalLengthWord());
return newp;
}
ASSERT (obj->ContainsNormalLengthWord()); // object is not moved
return old;
}
// Process the value at a given location and update it as necessary.
POLYUNSIGNED ScanAddress::ScanAddressAt(PolyWord *pt)
{
PolyWord val = *pt;
PolyWord newVal = val;
if (IS_INT(val) || val == PolyWord::FromUnsigned(0))
{
// We can get zeros in the constant area if we garbage collect
// while compiling some code. */
}
else if (val.IsCodePtr())
{
// We can get code pointers either in the stack as return addresses or
// handler pointers or in constants in code segments as the addresses of
// exception handlers.
// Find the start of the code segment
PolyObject *oldObject = ObjCodePtrToPtr(val.AsCodePtr());
// Calculate the byte offset of this value within the code object.
POLYUNSIGNED offset = val.AsCodePtr() - (byte*)oldObject;
// Mustn't use ScanAddressAt here. That's only valid if the value points
// into the area being updated.
PolyObject *newObject = ScanObjectAddress(oldObject);
newVal = PolyWord::FromCodePtr((byte*)newObject + offset);
}
else
{
ASSERT(OBJ_IS_DATAPTR(val));
// Database pointer, local pointer or IO pointer.
// We need to include IO area pointers when we produce an object module.
newVal = ScanObjectAddress(val.AsObjPtr());
}
if (newVal != val) // Only update if we need to.
*pt = newVal;
return 0;
}
// Store a constant value. Also used with a patch table when importing a saved heap which has
// been exported using the C exporter.
void ScanAddress::SetConstantValue(byte *addressOfConstant, PolyWord p, ScanRelocationKind code)
{
switch (code)
{
case PROCESS_RELOC_DIRECT: // 32 or 64 bit address of target
{
POLYUNSIGNED valu = p.AsUnsigned();
for (unsigned i = 0; i < sizeof(PolyWord); i++)
{
addressOfConstant[i] = (byte)(valu & 255);
valu >>= 8;
}
}
break;
case PROCESS_RELOC_I386RELATIVE: // 32 bit relative address
{
POLYSIGNED newDisp = p.AsCodePtr() - addressOfConstant - 4;
#if (SIZEOF_VOIDP != 4)
ASSERT(newDisp < 0x80000000 && newDisp >= -(POLYSIGNED)0x80000000);
#endif
for (unsigned i = 0; i < 4; i++) {
addressOfConstant[i] = (byte)(newDisp & 0xff);
newDisp >>= 8;
}
}
break;
}
}
void PExport::printValue(PolyWord q)
{
if (IS_INT(q) || q == PolyWord::FromUnsigned(0))
fprintf(exportFile, "%" POLYSFMT, UNTAGGED(q));
else if (OBJ_IS_CODEPTR(q))
printCodeAddr(q.AsCodePtr());
else
printAddress(q.AsAddress());
}
bool PImport::GetValue(PolyWord *result)
{
int ch = getc(f);
if (ch == '@')
{
/* Address of an object. */
POLYUNSIGNED obj;
fscanf(f, "%" POLYUFMT, &obj);
ASSERT(obj < nObjects);
*result = objMap[obj];
}
else if (ch == '$')
{
/* Code address. */
POLYUNSIGNED obj, offset;
fscanf(f, "%" POLYUFMT "+%" POLYUFMT, &obj, &offset);
ASSERT(obj < nObjects);
PolyObject *q = objMap[obj];
ASSERT(q->IsCodeObject());
*result = PolyWord::FromCodePtr((PolyWord(q)).AsCodePtr() + offset); /* The offset is in bytes. */
}
else if ((ch >= '0' && ch <= '9') || ch == '-')
{
/* Tagged integer. */
POLYSIGNED j;
ungetc(ch, f);
fscanf(f, "%" POLYSFMT, &j);
/* The assertion may be false if we are porting to a machine
with a shorter tagged representation. */
ASSERT(j >= -MAXTAGGED-1 && j <= MAXTAGGED);
*result = TAGGED(j);
}
else if (ch == 'I')
{
/* IO entry number. */
POLYUNSIGNED j;
fscanf(f, "%" POLYUFMT, &j);
ASSERT(j < POLY_SYS_vecsize);
*result = (PolyObject*)&gMem.ioSpace->bottom[j * IO_SPACING];
}
else if (ch == 'J')
{
/* IO entry number with offset. */
POLYUNSIGNED j, offset;
fscanf(f, "%" POLYUFMT "+%" POLYUFMT, &j, &offset);
ASSERT(j < POLY_SYS_vecsize);
PolyWord base = (PolyObject*)&gMem.ioSpace->bottom[j * IO_SPACING];
*result = PolyWord::FromCodePtr(base.AsCodePtr() + offset);
}
else
{
fprintf(stderr, "Unexpected character in stream");
return false;
}
return true;
}
// Returns the new address if the argument is the address of an object that
// has moved, otherwise returns the original.
PolyWord ProcessFixupAddress::GetNewAddress(PolyWord old)
{
if (old.IsTagged() || old == PolyWord::FromUnsigned(0) || gMem.IsIOPointer(old.AsAddress()))
return old; // Nothing to do.
// When we are updating addresses in the stack or in code segments we may have
// code pointers.
if (old.IsCodePtr())
{
// Find the start of the code segment
PolyObject *oldObject = ObjCodePtrToPtr(old.AsCodePtr());
// Calculate the byte offset of this value within the code object.
POLYUNSIGNED offset = old.AsCodePtr() - (byte*)oldObject;
PolyWord newObject = GetNewAddress(oldObject);
return PolyWord::FromCodePtr(newObject.AsCodePtr() + offset);
}
ASSERT(old.IsDataPtr());
PolyObject *obj = old.AsObjPtr();
POLYUNSIGNED L = obj->LengthWord();
// Generally each address will point to an object processed at a lower depth.
// The exception is if we have a cycle and have assigned the rest of the
// structure to a higher depth.
// N.B. We return the original address here but this could actually share
// with something else and not be retained.
if (OBJ_IS_DEPTH(L))
return old;
if (obj->ContainsForwardingPtr()) // tombstone is a pointer to a shared object
{
PolyObject *newp = obj->GetForwardingPtr();
// ASSERT (newp->ContainsNormalLengthWord());
return newp;
}
ASSERT (obj->ContainsNormalLengthWord()); // object is not shared
return old;
}
// 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();
}
}