[PATCH v2 0/5] malloc: TCACHE improvement for free and calloc

[Florian Weimer]

* Wilco Dijkstra:

> Hi Wangyang,
>
> So overall this looks a good improvement to me. I'm wondering whether you're
> planning further speedups? libc_malloc/free still do a lot of unnecessary work,
> initialization checks, errno save/restore etc, all of which are redundant in the fast path.
>
> For example, ignoring tagging/checking, libc_free should look like:
>
> void
> __libc_free (void *mem)
> {
>   mstate ar_ptr;
>   mchunkptr p;                          /* chunk corresponding to mem */
>
>  if (mem == 0)                              /* free(0) has no effect */
>     return;
>
>   p = mem2chunk (mem);
>   INTERNAL_SIZE_T size = chunksize (p);
>
>   // no need for init check if we do it in slow path or always init at thread startup
>   // no need for mmap check (blocks never fit in tcache)
>   // no need for errno saving (no system calls in tcache path)
>   if (tcache_free (p, size))   
>     return;

We can skip the mmap check only if we refuse to mmap small chunks, even
on fallback paths.  It's not an unreasonable change, though.

Thanks,
Florian