memchr-2.6-mov-regression — What's That AI

To compile and run the benchmark:

$ cargo build --release
$ ./target/release/i139

Profiling the execution (e.g., perf on Linux) should reveal a hot loop where
most of the time is being spent on x86-64:

  0.02 │440:┌─→add        $0x20,%rcx
  0.02 │    │  cmp        %rdx,%rcx
       │    │↑ ja         108
  0.05 │44d:│  mov        0x10(%rsp),%rsi
 32.64 │    │  vpcmpeqb   (%rcx,%rsi,1),%ymm0,%ymm2
       │    │  mov        0x8(%rsp),%rsi
 30.66 │    │  vpcmpeqb   (%rcx,%rsi,1),%ymm1,%ymm3
  0.71 │    │  vpand      %ymm2,%ymm3,%ymm2
  3.69 │    │  vpmovmskb  %ymm2,%ebp
  0.20 │    ├──test       %ebp,%ebp
 28.57 │    └──je         440

This is where the code should be spending most of its time. But if you change
the dependency on memchr to be =2.5.0, then the hot loop should now look
like this:

  0.02 │280:┌─→add          $0x20,%rcx
  0.02 │    │  cmp          %rsi,%rcx
       │    │↑ ja           123
 32.63 │28d:│  vpcmpeqb     (%rcx,%rdx,1),%ymm0,%ymm2
  7.18 │    │  vpcmpeqb     (%rcx,%rdi,1),%ymm1,%ymm3
  1.20 │    │  vpand        %ymm2,%ymm3,%ymm2
 27.80 │    │  vpmovmskb    %ymm2,%r12d
  0.09 │    ├──test         %r12d,%r12d
 28.61 │    └──je           280

Notice that there are two fewer instructions here. Specifically, the two
mov instructions are now gone. This appears to be a result of some code
shuffling done between the 2.5.0 and 2.6.0 release. I can't figure out how
to get back to the code without the extra mov instructions.

Does it matter? this issue seems to suggest it does, but the degree
to which it matters seems to depend on the CPU. In my case, I observe a ~6%
regression on the benchmark in this repository, but the issue reported sees a
much bigger regression of around ~30%. (I have an Intel CPU while the reporter
has an AMD CPU.)

README