[Bug target/104611] memcmp/strcmp/strncmp can be optimized when the result is tested for [in]equality with 0 on aarch64

["wilco at gcc dot gnu.org" <gcc-bugzilla@gcc.gnu.org>]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=104611

Wilco <wilco at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
     Ever confirmed|0                           |1
             Status|UNCONFIRMED                 |NEW
   Last reconfirmed|                            |2023-09-28

--- Comment #5 from Wilco <wilco at gcc dot gnu.org> ---
(In reply to Mathias Stearn from comment #4)
> clang has already been using the optimized memcmp code since v16, even at
> -O1: https://www.godbolt.org/z/qEd768TKr. Older versions (at least since v9)
> were still branch-free, but via a less optimal sequence of instructions.
> 
> GCC's code gets even more ridiculous at 32 bytes, because it does a branch
> after every 8-byte compare, while the clang code is fully branch-free (not
> that branch-free is always better, but it seems clearly so in this case).
> 
> Judging by the codegen, there seems to be three deficiencies in GCC: 1) an
> inability to take advantage of the load-pair instructions to load 16-bytes
> at a time, and 2) an inability to use ccmp to combine comparisons. 3) using
> branching rather than cset to fill the output register. Ideally these could
> all be done in the general case by the low level instruction optimizer, but
> even getting them special cased for memcmp (and friends) would be an
> improvement.

I think 1, 2 and 3 are all related due to not having a TImode compare pattern,
so GCC splits things into 8-byte chunks using branches. We could add that and
see whether the result is better or add a backend expander for memcmp similar
to memset and memcpy.

Note what LLVM does is terrible, a 64-byte memcmp is ridiculously inefficient
due to long dependency chains, loading and comparing every byte even if there
is a mismatch in byte 0. So it's obviously better to use branches.