[webkit-dev] Question about javascriptcore garbage collection

[Kurt Kohler]

Shouldn't the destructor in ValueImp be virtual? I'm not totally clear
on this but I understand there can be problems if destructors aren't
declared virtual in virtual base classes. ValueImp isn't actually
virtual, but the derived classes are. Maybe that's how platform
dependencies are arising.

HTH

Krzysztof Kowalczyk wrote:

>To answer my own question: the problem is that ValueImp has no virtual
>functions (=> no virtual table) and AllocatedValueImp does. That means
>that cast between ValueImp* and AllocatedValueImp* involve
>adding/substracting 4 bytes to the pointer of the object, since, at
>least in visual c, pointer to vtable is the first member of the
>object. AllocatedValueImp * is allocated at CollectorCell boundary,
>but when casted to  ValueImp * it is no longer recognized by GC.
>
>Here's a small test program that shows this. See the last 3 addresses
>in the output for how the address changes depending on the type of
>pointer.
>
>So I devised another way of fixing the problem: simply marking
>ValueImp::~ValueImp as virtual also fixes crashes for me.
>
>So the remaining question is: what is the behaviour on Mac/gcc ? I
>would aprecciate if a mac person would run this and posted the
>results. On my win machine it says:
>
>C:\kjk\src\ext\libs\webcore-apple\JavaScriptCore>inh-test.exe
>justVal         : 0x322bc0
>num             : 0x322bd0
>numCasted       : 0x322bd4
>numDowncasted   : 0x322bd0
>jsNumber num    : 0x320f60
>jsNumberAsVI    : 0x320f64
>jsNumberAsAVI   : 0x320f60
>
>On a separate note: I don't see why ValueImp couldn't be eliminated
>alltogether and its functionality folded in  AllocatedValueImp (which
>would also solve the problem).
>
>Test program:
>
>#include <stdio.h>
>
>class ValueImp {
>    friend class AllocatedValueImp; // so it can derive from this class
>public:
>    ValueImp() {};
>    ~ValueImp() {};
>    void mark();
>    bool marked() const;
>
>    AllocatedValueImp *downcast();
>    const AllocatedValueImp *downcast() const;
>};
>
>class SimpleNumber {
>public:
>  static inline bool is(const ValueImp *imp) { return false; }
>};
>
>class AllocatedValueImp : public ValueImp {
>    friend class NumberImp;
>private:
>    AllocatedValueImp() {};
>    virtual ~AllocatedValueImp() {};
>public:
>    virtual void mark();
>    bool marked() const;
>private:
>    bool m_marked;
>};
>
>class NumberImp : public AllocatedValueImp {
>  public:
>    NumberImp(double v) : val(v) { }
>  private:
>    virtual ~NumberImp() {};
>    NumberImp() { }
>    double val;
>};
>
>inline void ValueImp::mark()
>{
>    if (!SimpleNumber::is(this))
>        downcast()->mark();
>}
>
>inline bool ValueImp::marked() const
>{
>    return SimpleNumber::is(this) || downcast()->marked();
>}
>
>inline AllocatedValueImp *ValueImp::downcast()
>{
>    return static_cast<AllocatedValueImp *>(this);
>}
>
>inline const AllocatedValueImp *ValueImp::downcast() const
>{
>    return static_cast<const AllocatedValueImp *>(this);
>}
>
>inline bool AllocatedValueImp::marked() const
>{
>    return m_marked;
>}
>
>inline void AllocatedValueImp::mark()
>{
>    m_marked = true;
>}
>
>ValueImp *jsNumber(double d)
>{
>    NumberImp *num = new NumberImp(d);
>    printf("jsNumber num    : 0x%x\n", num);
>    return num;
>}
>
>int main(int argc, char **argv)
>{
>    ValueImp  *justVal = new ValueImp();
>    AllocatedValueImp *num = new NumberImp(15.0);
>    ValueImp *numCasted = static_cast<ValueImp*>(num);
>    AllocatedValueImp *numDowncasted = numCasted->downcast();
>
>    printf("justVal         : 0x%x\n", justVal);
>    printf("num             : 0x%x\n", num);
>    printf("numCasted       : 0x%x\n", numCasted);
>    printf("numDowncasted   : 0x%x\n", numDowncasted);
>
>    ValueImp *jsNumberAsValueImp = jsNumber(5.0);
>    printf("jsNumberAsVI    : 0x%x\n", jsNumberAsValueImp);
>    AllocatedValueImp *asAllocatedValueImp = jsNumberAsValueImp->downcast();
>    printf("jsNumberAsAVI   : 0x%x\n", asAllocatedValueImp);
>    return 0;
>}
>_______________________________________________
>webkit-dev mailing list
>webkit-dev at opendarwin.org
>http://www.opendarwin.org/mailman/listinfo/webkit-dev
>  
>