Collections crate that uses allocator api

help
1

Is there some crate that implements collections, such as vector or hashmaps but uses the new allocator API solely, that is, doesn't rely on the GlobalAlloc trait?

2

Both std::vec::Vec and hashbrown::HashMap use the allocator API, allowing you to specify an allocator. Is that not what you're referring to?

3

Yes they do. But I found out that in-order to use the alloc crate in a no-std binary, you need to define a #[global_allocator] nevertheless.
Or am I wrong about this part?

4

I havenā€™t tried it, but I assume you could define a global allocator that simply panics whenever an allocation request comes in. Itā€™s not the best solution, as the errors come at runtime instead of compile time, but it might get the job done in practice.

5

Thanks! I'll try it out.

6

I just pulled up the docs for GlobalAlloc, and panicking is UB. You can return a NULL pointer from the allocation methods, though, to indicate an allocation failure. All of the other methods require a previously-allocated chunk to be passed inā€” If you never gave any of those out, the program is already in UB-land anyway.

7

So I guess the solution would be return null from alloc and put unreachable in the rest since they literally should not be called?

8

Thatā€™s essentially the way I read it:

  • alloc always returns null
  • dealloc calls unreachable!
  • Default implementation for everything else
9

With Allocator::deallocate it's clear (nonnull), but for GlobalAlloc::dealloc it's not - is a null pointer a legal input to dealloc? Given how free works, I assume it is and is a no-op. So that's how I'd implement dealloc.

10

No-op would certainly be acceptable here. Iā€™m basing my UB claim on this clause from the dealloc docs:

This function is unsafe because undefined behavior can result if the caller does not ensure all of the following:

  • ptr must denote a block of memory currently allocated via this allocator,
  • layout must be the same layout that was used to allocate that block of memory.

As ā€˜this allocatorā€™ never allocates blocks of memory, itā€™s always UB to call its dealloc method. If youā€™re feeling particularly daring, Iā€™m pretty sure unreachable_unchecked would be valid here, which might optimize out all code paths that involve a successful allocation. Personally, that seems like too risky a move.

Also, the alloc docs say that the method

Returns a pointer to newly-allocated memory, or null to indicate allocation failure.

I donā€™t believe that an explicitly-listed sentinel value indicating failure can reasonably be construed as representing a ā€œblock of memory currently allocatedā€.