Don’t assume MIN_ALIGN for small sizes

[SimonSapin]

Same as rust-lang/rust#46117

See also discussion of jemalloc’s behavior at jemalloc/jemalloc#1072

[gnzlbg]

@SimonSapin could you check if this use case is covered in #26 ?

If so, it might be good to run the benches on this PR and see if there is a difference, and if so, how big.

But AFAIK this is required for correctness, is that right?

[SimonSapin]

Right, this is correctness fix. The aspect that’s performance-dependent is whether we should go further and remove MIN_ALIGN entirely.

[gnzlbg]

Makes sense.

---

FWIW MIN_ALIGN is also used here to dispatch to calloc instead of mallocx(size, flags | MALLOCX_ZEROED), so it might have performance implications beyond the cost of align_to_flags.

[SimonSapin]

Right, calling calloc is probably similar to the flags == 0 fast path in other jemalloc functions discussed in rust-lang/rust#46117 (comment).

[gnzlbg]

@SimonSapin yes I've added a benches/calloc.rs that benchmarks calloc vs mallocx( , flags | MALLOCX_ZEROED), the differences in most cases is "small", in some cases mallocx is faster and in others calloc is faster, but you will be able to try for yourself once that is merged.

Please feel free to review the benchmarks and give me input. Its a lot of boilerplate that I don't seem to be able to avoid because I cannot concatenate identifiers to form new identifiers in macros.

[alexcrichton]

Thanks!