I previously mentioned that I’m a stickler for good platform code. Of all the design decisions I’ve ever made, there is one that I regret above all others. To this day, it haunts me mercilessly. It has appeared over and over again in countless lines of code for Half-Life 1, and silently passed to a few Half-Life 2 plugins. What is it?

The [g|s]_et_pdata functions in AMX Mod X allow you to reach into a CBaseEntity’s private data (i.e., its member variables) and pull out values given an offset from the this pointer. What’s the problem? Let’s take a look at how each of the retrieval mechanisms work:

• 4-byte retrievals are given by a 4-byte aligned offset ((int *)base + offset)
• Float retrievals are offset by a float aligned offset ((float *)base + offset)

At the time, I probably didn’t understand how addressing or memory really worked. I was also basing the API off years of code that already made these mistakes — learning from an incorrect source doesn’t help. So, what’s the actual problem?

In C/C++, given pointer X and offset Y, the final address is computed as X + Y * sizeof(X). That means when we do (int)base, the actual offset becomes offset * sizeof(int), or 4*offset. Not only have we lost byte addressability, but to get unaligned values we have to do various bitwise operations. This type of deep-rooted mistake is easily visible in the following code in the cstrike module:

if ((int)*((int *)pPlayer->pvPrivateData + OFFSET_SHIELD) & HAS_SHIELD)
	return 1;

It turns out HAS_SHIELD is a bitmask to clear unwanted data, when in reality, the offset is wrong and (int *) should be (bool *).

Alas, backwards compatibility prevents us from ever making these API changes, and changing internal usage would then confuse users further.

In summary, the practice of casting to random types is ridiculous and SourceMod’s API summarily kills it off. The entire practice can now be expressed in a function like this:

template <typename T>
void get_data(void *base, size_t offset, T * buffer)
{
   *buffer = *(T *)((char *)base + offset);
}

One of the most tricky things about C++, as opposed to a garbage-collected language, is the lifetime of a pointer (or the memory it points to). The standard cases are the ones beginners learn pretty quickly:

• “Forever”: Static storage means global lifetime
• Short: Local storage means until the end of the current scope block
• Very short: Objects can be temporarily created and destroyed from passing, return values, or operator overloads.
• Undefined: Manually allocated and destroyed memory

A more subtle type of pointer invalidation is when storing objects in a parent structure. For example, observe the following code:

    std::vector<int> v;
    v.push_back(1);
 
    int &x0 = v[0];
 
    printf("%d\n", x0);
 
    v.push_back(2);
 
    printf("%d\n", x0);

Running this little code block on gcc-4.1, I get:

[dvander@game020 ~]$ ./test
1
0

What happened here? I took an address (references, of course, are just addresses) to a privately managed and stored object (in this case, the object is an integer). When I changed the structure by adding a new number, it allocated new memory internally and the old pointer suddenly became invalid. It still points to readable memory, but the contents are wrong; in a luckier situation it would have crashed.

What’s not always obvious about this type of mistake is that the vector is not being explicitly reorganized, as would be the case with a call like clear(). Instead, it’s important to recognize that if you’re going to cache a local address to something, you need to fully understand how long that pointer is guaranteed to last (if at all).

This type of error can occur when aggressively optimizing. Imagine if the previous example used string instead of int — suddenly, storing a local copy of the string on the stack involves an internal new[] call, since string needs dynamic memory. To avoid unnecessary allocation, a reference or pointer can be used instead. But if you’re not careful, a simple change will render your cached object unusable — you must update the cached pointers after changes.

There was a was very typical error in SourceMod that demonstrated this mistake. SourceMod has a caching system for lump memory allocations. You allocate memory in the cache and the cache returns an index. The index can give you back a temporary pointer. Observe the following pseudo-code:

int index = cache->allocate(sizeof(X));
X *x = (X *)cache->get_pointer(index);
//...code...
x->y_index = cache->allocate(sizeof(Y));

Can you spot the bug? By the time allocate() returns, the pointer to x might already be invalidated. The code must be:

int index = cache->allocate(sizeof(X));
X *x = (X *)cache->get_pointer(index);
//...code...
int y_index = cache->allocate(sizeof(Y));
x = (X *)cache->get_pointer(index);
x->y_index = y_index

There are many ways to create subtle crash bugs by not updating cached pointers, and they often go undetected if the underlying allocation doesn’t trigger address changes very often. Worse yet, whether that happens or not is often very dependent on the hardware or OS configuration, and serious corruption or crash bugs may live happily undiscovered for long periods of time.

The moral of this story is: Be careful when keeping pointers to where you don’t directly control the memory. Even if you’ve written the underlying data structure, make sure you remember exactly what the pointer lifetime is guaranteed to be.