[Bug libstdc++/104719] Use of `std::move` in libstdc++ leads to worsened debug performance

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

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=104719

Jason Merrill <jason at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
                 CC|                            |jason at gcc dot gnu.org

--- Comment #12 from Jason Merrill <jason at gcc dot gnu.org> ---
(In reply to Vittorio Romeo from comment #9)
> - For the `operator[]` benchmark, when using `-Og` after applying
> `[[gnu::always_inline]]` to all the functions touched by the benchmark, we
> reduce the overhead from 34% to around 11%. 

Quoting from your gist:

-Og, without `[[gnu::always_inline]]` on `operator[]`
carray_squareop_mean          440 ns          439 ns            3
vector_squareop_mean          661 ns          662 ns            3
-Og, with `[[gnu::always_inline]]` on `operator[]`
vector_squareop_mean          494 ns          491 ns            3

Which looks significant...

...but I don't see this when I run your test myself; the vector_squareop
results (and the generated code) are unaffected by adding always_inline.  In
particular there are no calls to operator[].

carray_squareop          547 ns          546 ns      1272023
vector_squareop          591 ns          589 ns      1227215

carray:

        movslq  %edx, %rax
        leaq    (%rbx,%rax,4), %rcx
        movl    (%rcx), %eax
        addl    $1, %eax
        movl    %eax, (%rcx)
        movl    %eax, (%rcx)
        addl    $1, %edx

vector:

        movslq  %ecx, %rax
        salq    $2, %rax
        addq    (%rsp), %rax
        movl    (%rax), %esi
        leal    1(%rsi), %edx
        movl    %edx, (%rax)
        movl    %edx, (%rax)
        addl    $1, %ecx

It seems the main difference between the two is that the vector version needs
to keep loading the base pointer from the stack (%rsp) for some reason, rather
than keep it in a register %rbx.  This doesn't seem like an inlining issue at
all.

The double move in both versions is curious.