Consider specializing std::vec::from_elem to calloc when possible

[sfackler]

This will require adding calloc to the global allocator API, but can significantly speed up situations where a large buffer is allocated but not entirely used.

[brson]

Vec::from for [T] probably, not from_slice (I don't see a from_slice).

What's the benefit you are expecting? When would zeroed pages help in Vec?

[sfackler]

This would be targeting things like vec![0; 40960] which uses from_elem.

Calloc can avoid explicitly zeroing pages if it knows they're already zeroed out (e.g. if they just came from an mmap call). If much of the buffer isn't actually used, this can be way faster. See for example https://bitbucket.org/cffi/cffi/issues/295/cffinew-is-way-slower-than-it-should-be-it

[abonander]

As a practical example, this could improve performance of both std::io::BufReader::new() and Read::read_to_end()/read_to_string() (at least for reading into an empty vector). The box operator could also be specialized to use this for the same reasons.

I asked about this before on one of the easy question threads on Reddit: one of the responses I got said that it probably wouldn't improve performance under jemalloc for various esoteric reasons (which I found hard to believe, but whatever); however, it would definitely help under the system allocator.

Edit: consolidated two comments and removed redundant statements

[bluss]

std::vec::from_elem is the function behind the vec![x; n] repetition. source for from_elem