Scalable allocator vs virtual methods - a bug ?

[Bartlomiej]

Hello,

I just started playing with memory allocators - up to now I only used new/delete.
And possibly I observeda bug - the following code crashes on my box (for both scalable and cache aligned allocators):

<sup>#include

#include "tbb/scalable_allocator.h"
#include "tbb/cache_aligned_allocator.h"

struct sth {</sup>

^{double x;}
int i;

^{sth (double ax, int ai) : x(ax), i(ai) {};}

^{virtual void show (void) {}

^{std::cout << "x = " << x << std::endl;}

^{std::cout << "i = " << i << std::endl;}

^};

<sup>};

tbb::scalable_allocator alloc;
//tbb::cache_aligned_allocator alloc;

int main () {</sup>

^{sth *ex = alloc.allocate (1);}

^{ex->x = 3.1415926;}

^{ex->i = 13;}

^{ex->show ();}

^{alloc.deallocate (ex, 1);}

^{return 0;}

^}

As a matter of fact, switching the method to be non-virtual makes the program work correctly.
Is it a bug ? Or possibly I'm doing something wrong ?

Best regards
Bartlomiej

[Dmitry_Vyukov]

Besides allocation of a memory for the object, you have to construct the object itself:
void* mem = alloc.allocate(1);
sth* ex = new (mem) sth (3.14, 13);
Otherwise you just consider memory block filled with trash as an C++ object which is dead wrong.
See the ISO C++ Standard section 3.8. Object Lifetime and 20.8.2 (may differ in your revision) Allocators.

[Bartlomiej]

Quoting - Dmitriy Vyukov

Oops, yes. I didn't know this. Sorry.

Also, I hope, using the new() operator that way does not require any synchronization with other threads?

[Dmitry_Vyukov · Accepted solution]

Quoting - bartlomiej

Also, I hope, using the new() operator that way does not require any synchronization with other threads?

Yes, it's no more that a call of a class constructor. If constructor is empty and class does not have virtual functions, then this call will be completely optimized away in release build.