Placement new to initialize inplace object

Question

I recently came across a rather interesting serialization approach that utilized the transparency (common undefined behavior among compilers?) of uninitialized variables for "efficient" deserialization.

Memory is allocated and assigned a predetermined value. Placement new is then used to instantiate a struct (for example a complex inplace data structure) "initializing" the uninitialized variables to the value of the underlying memory. (see code below)

Besides from being rather risky and possibly not a very agreeable coding convention... I was just wondering whether anyone had come across this method or more importantly -- what is it called?

class SomeClass {
public:
  SomeClass() {}

  int someInt;
};

int main(...) {
  int dummy = 42;

  int *pSomeClass = new (&dummy) SomeClass();
  cout << pSomeClass->someInt << endl;
}

This will print out the number 42... neato!

score 5 · Answer 1 · answered Nov 01 '11 at 08:22

5

It's called "relying on UB" or, in laymen's terms, "foolishness".

answered Nov 01 '11 at 08:22

Lightness Races in Orbit

378,754
76
643
1,055

@David: `pSomeClass::someInt` is not initialised. – Lightness Races in Orbit Nov 01 '11 at 08:27
The layout of a `struct` and a `class` is well-defined though, or is it not? – David Heffernan Nov 01 '11 at 08:28
@DavidHeffernan: Padding, alignment... and I don't see what that has to do with it. – Lightness Races in Orbit Nov 01 '11 at 08:29
The first member is always at offset 0, as I understand things. I must admit I don't see why it wouldn't always be easier to initialize in the constructor. – David Heffernan Nov 01 '11 at 09:01
@David: Regardless, in the context of the new object, that member has not been initialised. Placement new may imbue some more lenient policies here, but if it does then I'm not aware of them [yet]. – Lightness Races in Orbit Nov 01 '11 at 09:06
Is that because it's C++ rather than C? – David Heffernan Nov 01 '11 at 09:23
@David: You'll have to be more specific. They are separate languages with separate rules. – Lightness Races in Orbit Nov 01 '11 at 10:13
The problem is that the ctor is called, and _may_ initialize `someInt`. A reasonable compiler, in ISO-conforming debug mode, may set `someInt` to `0xDEADBEEF`. You can't assume it will be initialized, but you can't assume that it will be not-initialised either. – MSalters Nov 01 '11 at 11:56
@MSalters: In the OP's example it's explicit that the member is not initialised; he's asking what it's called when we rely on using the uninitialised value. – Lightness Races in Orbit Nov 01 '11 at 21:38

score 2 · Answer 2 · answered Nov 01 '11 at 10:43

I have seen this done in eCos, an RTOS, to initialize some of their kernel objects.

As pointed out by Tomalak one of the draw-backs is no virtual functions allowed. They try to ensure that by testing for equal size sizeof(kernel object) == sizeof(variable used for initialization).

Their code although was way more complex using a C-struct to mimic the C++ class member variables for the c interface instead of using C functions to get/set the variables in the C++ class

Although the behaviour they went for there was the exact opposite, they used values from the C++ class, set in the constructor, to fill the memory from placement new.

I do not advise on doing this tho.

Placement new to initialize inplace object

2 Answers2