🔒 Locked

Bad access in virtual methods (iOS ARM)

Started by Cordan Sep 28, 2012 at 8:01 PM 9 replies 3.1k views

Original Post

Sep 28, 2012 at 8:01 PM

I getting EXC_BAD_ACCESS in virtual methods when i run program on iPad.
It happens only if method returns registered data type and if method have virtual definition in base class.

My code:

[source lang="cpp"]
virtual my_type Base::method() = 0;

...

my_type Derived::method()
{
...
this->field = 10; // EXC_BAD_ACCESS
...
}
[/source]

If i replace code above with:
[source lang="cpp"]
virtual void Base::method(my_type &res) = 0;

...

void Derived::method(my_type &res)
{
...
this->field = 10; // OK
...
}
[/source]
problem will disappear.

I have this problem only on ios device. i've tested my code on windows, osx and iphone simulator - all is ok. If i replace 'my_type' to 'int' all is ok as well.

It looks like calling of virtual methods with wrong object.

So, what it can be?

ps: as version - 2.25.0

TechRogue

Sep 29, 2012 at 7:42 PM

Can you show the code where you register my_type with the script engine, as well as the registration code for this method?

Cordan

Sep 30, 2012 at 10:51 AM

Of course.

Method:
[source lang="cpp"]int r = engine->RegisterObjectMethod("Derived", "my_type method()",
asSMethodPtr::Convert(&Derived::method),
asCALL_THISCALL);
assert(r >= 0);[/source]

Type:
[source lang="cpp"]int r = engine->RegisterObjectType("my_type", sizeof(my_type), asOBJ_VALUE | asOBJ_APP_CLASS_CDAK);
assert(r >= 0);

r = engine->RegisterObjectBehaviour("my_type", asBEHAVE_CONSTRUCT, "void f()", asFUNCTION(__construct), asCALL_CDECL_OBJLAST);
assert(r >= 0);
r = engine->RegisterObjectBehaviour("my_type", asBEHAVE_CONSTRUCT, "void f(const my_type ∈)", asFUNCTION(__construct_copy),asCALL_CDECL_OBJLAST);
assert(r >= 0);
r = engine->RegisterObjectBehaviour("my_type", asBEHAVE_DESTRUCT, "void f()", asFUNCTION(__destruct), asCALL_CDECL_OBJLAST);
assert(r >= 0);

r = engine->RegisterObjectMethod("my_type", "my_type& opAssign(my_type ∈)",
asSMethodPtr::Convert((my_type& (my_type::*)(const my_type&))(&my_type::operator=)),
asCALL_THISCALL);
assert(r >= 0);[/source]

TechRogue

Sep 30, 2012 at 2:39 PM

I'm not familiar with asSMethodPtr, so I don't know if you have a specific reason for using it, but why not use the asMETHODPR macro?


 r = engine->RegisterObjectMethod("my_type", "my_type& opAssign(my_type ∈)", asMETHODPR(my_type, operator=, (const my_type&), my_type&), asCALL_THISCALL);

It's probably not the problem, but pointers to member functions are tricky and the macros do some extra magic under the hood.

Cordan

Sep 30, 2012 at 7:17 PM

It not helps. asMETHODPR is wrapper around asSMethodPtr.
I'm still getting crashes when use virtual and pure-virtual methods.

I guess it is memory corruption in arm implementation of angel script. After call of any virtual method which returns my type crash occurs. It occurs either in method itself or further in code.

WitchLord

Oct 4, 2012 at 12:50 AM

I haven't done any iOS development on my own so I can't test this, but it looks to be a problem in the AngelScript code when handling the hidden address for returning objects by value.

I'll review the code to see if I can spot the error. I may need some help from you to run some tests for me in order to have this fixed.

Cordan

Oct 4, 2012 at 2:08 PM

I'll review the code to see if I can spot the error. I may need some help from you to run some tests for me in order to have this fixed.

I sent PM to you.

WitchLord

Oct 19, 2012 at 1:26 AM

Alright. I'm sorry for taking so long to start looking into this problem, but I've had a quite chaotic couple of weeks at work since I got back from work. What little time that remained I spent on fixing the logical bugs in AngelScript.

After reviewing the code I believe the problem is caused by an inverted condition on line 204 in as_callfunc_arm.cpp. I believe that line should should be [font=courier new,courier,monospace]#ifdef __GNUC__[/font], i.e:


 case ICC_VIRTUAL_THISCALL_RETURNINMEM:
 // Get virtual function table from the object pointer
 vftable = *(asFUNCTION_t**)obj;
 #ifdef __GNUC__ /// <-- change this to ifdef, so that R0 is populated with retPointer
 retQW = armFuncR0R1(args, (paramSize+1)<<2, vftable[FuncPtrToUInt(func)>>2], (asDWORD)retPointer, (asDWORD)obj);
 #else
 retQW = armFuncR0R1(args, (paramSize+1)<<2, vftable[FuncPtrToUInt(func)>>2], (asDWORD)obj, (asDWORD)retPointer);
 #endif
 break;

Please let me know if this change corrects the problem, and I'll check in the modification.

WitchLord

Oct 19, 2012 at 11:42 PM

Well, I'm so sure the fix is correct that I checked in the change in revision 1440 anyway. Please let me know if I was wrong and I'll undo the change.

Cordan

Oct 20, 2012 at 7:01 PM

Yes, everything works now. Huge thanks!

WitchLord

Oct 21, 2012 at 2:10 AM

Thanks for confirming that it's working properly.

Topic Locked

This topic has been locked by a moderator. New replies are not allowed.