Why malloc always does more than I asked for?
ssenthilnathan3.github.io
ssenthilnathan3.github.io
Malloc doesn't know the required alignment (because has no idea what the type is, everything is cast through void). So all malloc implementations have a minimum alignment guarantee. Typically 16 bytes these days on x86, as that means even 128bit SSE values will end up aligned by default.
You couldn't go below the sizeof(void ) anyway, the backpointer needs to aligned too.
The padding only happens when you use memalign or aligned_malloc to specify a much larger alignment.
But most C code out there assumes malloc will always return something that is at least aligned to sizeof(void *), it's very rare to see aligned_alloc. So how is your alloc allocation going to know when it can get away with a smaller alignment?
Even if you are on a cpu that doesn't fault on unaligned memory access, any malloc implementation that doesn't align by default will have serious disadvantages in any benchmarks. IMO, There is no good reason to use an unaligned backpointer.
However, the pointer to metadata is not really necessary.
The associated metadata could be stored in a table, and the index of the metadata could be computed from the offset of the pointer returned by malloc to the start of the heap (possibly using a hash function).
The ancient versions of the Microsoft C/C++ compilers were using a malloc with inline metadata. I have no idea if they replaced this more recently.
[0] https://www.intel.com/content/www/us/en/docs/intrinsics-guid...
Fun story. Back in the early 2000s, my Delphi OpenGL program randomly crashed when I tried to use vertex buffers. I spent a lot of time checking this and that, before I realized it might be an alignment issue.
Sure enough, Delphi memory allocator at the time only provided 4-byte alignment, while NVIDIA's drivers used aligned SSE instructions which require 16-byte alignment.
So, had to manually align the buffers before passing them to OpenGL.
However Microsoft's malloc at the time was indeed returning 16-byte aligned buffers so they probably didn't notice it in internal testing.
if (a->cursor + size > a->limit) return NULL; // out of memory
if (size > a->limit - a->cursor) return NULL; // out of memory if (size > a->limit - a->cursor) return NULL; [ Header ][ ...variable padding... ][ Back Pointer ][ User Memory ]
^ always exactly sizeof(void*)
bytes before User Memory,
no matter how much padding
came before it
[ Header ][ Back Pointer ][ Padding ][ User Memory ]
You might as well always align to 8 bytes and make your header a multiple of 8.
Third party allocators improved on this as well as usually avoiding the heap fragmentation problems that system heaps often suffered from back then.
These days the system heaps have improved to a point where you really only need to think about a third party heap for very specialized circumstances. They do usually come with nifty debugging tools though.
"to get individual free() working, the allocator needs to remember something about every allocation it handed out. and that’s the moment metadata stops being optional."
That's just a very nice distillation of an important concept.
The conventional approach for allocating memory on GPUs for games and other applications is to use a real-time allocator such as TLSF. However, it is not usually discussed that TLSF is real-time because it stores metadata in-band. It is possible to create a variant of TLSF that preserves its real-time properties while storing metadata out-of-band, but this requires careful consideration.
If you can convince the caller to keep track of that metadata themselves you obviously don’t need to. That can be important.
Something I noticed is that _very often_ the code that is calling malloc(n) is keeping track of n somehow for its own reasons (bounds checking, grow/gap pointers, etc) so merging the value halves stack churn and it’s an easy win.
> That's why C23 introduce free_sized
Is it? C23 still has free, so there’s no guarantee callers wil use free_sized, so the allocator still has to be able to obtain a block’s size from a pointer.
Or do I overlook something?
I think the only thing free_size adds is robustness. Allocators can check the passed in size with what they know and abort the program if they do not match. That can thwart some security issues.
likely it'd a be perf. hit in most cases. They'd have to copy to the tail end (likely) of the allocated area. Or the start and offset the pointer, they'd need to know the size of the metadata and account for that, including aligning it.Hence, the tail feels 'nicer'
It's possible to manually use mmap and forgo malloc entirely, rolling your own arena manager.
So, most likely, there are two sizes in reality: the size of your user data that you care about; and the size of the memory chunk in which your user data resides, that free() cares about. So, unless you're willing to go for an API like this, you can't rely on the consumer:
int* ptr_to_dest = NULL;
size_t size = 10 * sizeof(int);
size_t allocated = malloc(&ptr_to_dest, size);
if (allocated <= 0 || !ptr_to_dest) {
//handle allocation error
}
//... use ptr_to_dest, size
free(ptr_to_dest, allocated); //careful not to pass size here!The metadata that malloc always attaches to the allocated memory would normally be useful for the user (i.e. having access to values like the currently allocated size and the total size of the allocation).
Very frequently, the user must duplicate inside the allocated memory the same information that already exists in the metadata, wasting memory. Also time may be wasted with requests for reallocation, instead of just adjusting the currently allocated size, when this is sufficient.
With a better API, the metadata would have been visible for the user. Being hidden from the user is not a protection in a language like C, where using pointer arithmetic can trash any memory location. Being exposed as non-mutable would have been a better protection.
Moreover, a better metadata structure for malloc should have always included a reference count, to be handled automatically by the compiler, and the malloc/free functions should have been invoked only implicitly, never explicitly.
1) If the app knew about and could use the additional underlying capacity, memory checkers wouldn't be able to find small overflows
2) Calling realloc instead of knowing existing capacity doesn't really provide better performance, at least not asymptotically.
3) If you need to know the requested allocation size, usually its for a dynamic array. Constantly requesting the size from the allocator would often have horrible performance if the size wasn't stored in adjacent metadata, but instead required a lookup operation, as is the case with hardened allocators. In practice hardened allocators probably wouldn't be a thing, or alternatively it would be idiomatic to store the size separately anyhow.
4) Reference counts would impose significant space and layout limitations despite most allocations never needing it.
There's another group who would argue the allocator should require passing the size and alignment to the free function, so allocators can be optimized better. Most of the time this is known statically. I think C2y will add APIs like this, though there's also a proposal to standardize an interface to query allocation capacity. Altogether I wouldn't be surprised if someday people appreciated the balance struck by malloc/free/realloc, especially as a minimal interface for overlaying application-specific allocators or injecting instrumentation. It has a certain elegance, in the sense of nothing left to remove.
I bet this is better optimized in Rust.
> You can't even address 4 bits.
This is not true, you can use a pointer plus extra offset data, etc.
Using a fat pointer to address individual bits is a waste of memory. You'd need an offset and length, and you'd have to deal with byte boundaries when accessing data. The only way I see this working is if you had a language that natively supported sub byte types and could deal with alignment of such types natively. Your best bet would be bit fields, which Rust doesn't support natively.
Dynamically allocating tiny amounts of data is a niche issue. You don't want a genetic solution to this anyway, because the efficiency will come from the constraints specific to your problem.