Avoid unnecessary allocation from realloc

[Theodus]

No description provided.

[msftclas]

All CLA requirements met.

[Theodus]

Also, quick question: Should calloc ever result in a call to madvise(... MADV_DONTNEED) on the region being allocated? This seems wrong to me.

[davidchisnall]

I'm not 100% convinced by this change. realloc is often used in a pattern where you allocate something bigger than you'll need, figure out how much you actually need, and then realloc to shrink the allocation. With malloc implementations that subdivide allocations, this lets the allocator reuse the rest of the allocation. It's not clear to me what the correct behaviour should be with snmalloc: we'll waste memory if we don't copy, but we'll also incur extra overhead.

On the calloc / madvise question, it might make sense on Linux. Linux's MADV_DONTNEED is misnamed and actually removes the physical pages and replaces them with CoW mappings of a zero page. Whether this hurts performance more or less than calling memset is not clear. We'll only hit this code path on allocations that are greater than 4KB, so we'll end up blowing away a chunk of cache if we memset. On the other hand, it's a system call that acquires a bunch of VM subsystem locks (probably does some IPIs), then the same thing on page fault a few instructions later when the allocation is touched. We should possibly consider always memseting the first page and only madviseing later ones.

[mjp41]

We could only copy if the size would lead to a different sizeclass?

[davidchisnall]

Why not call Alloc::alloc_size directly here, rather than the wrapper?

[davidchisnall]

Looks good to me. Minor change would be nice, but I'm happy either way.