Add inline tags to UninitSlice methods

[danburkert]

This appears to be the primary cause of significant performance
regressions in the prost test suite in the 0.5 to 0.6 transition. See
danburkert/prost#381.

[Darksonn]

Thanks!

[seanmonstar]

I'd expect the bounds check to have been elided, since it's asserted just before... It's not?

[danburkert]

I didn't check the assembly, but empirically I found in #442 that a similar change had a measurable impact on microbenchmarks in the prost repo. It's very possible that the issue there was that the UninitSlice index call was not inlined, and so the similar double check wasn't eligible to be optimized away. Realistically I'm not going to have time to read the ASM here, but I'm happy to back out this part of the change if you'd prefer that.

[seanmonstar]

Using s[i] vs s.as_ptr().add(i) gets me the same thing with the assert in the function:

example::index:
        push    rax
        cmp     rsi, rdx
        jbe     .LBB7_1
        mov     al, byte ptr [rdi + rdx]
        pop     rcx
        ret
.LBB7_1:
        lea     rdi, [rip + .L__unnamed_4]
        call    std::panicking::begin_panic
        ud2

example::ptr:
        push    rax
        cmp     rsi, rdx
        jbe     .LBB8_2
        mov     al, byte ptr [rdi + rdx]
        pop     rcx
        ret
.LBB8_2:
        lea     rdi, [rip + .L__unnamed_5]
        call    std::panicking::begin_panic
        ud2

[danburkert]

OK, would you like me to switch it back?

[seanmonstar]

It seems reasonable to me to keep to safe things if they make no performance difference.

[danburkert]

sounds good, updated.

[taiki-e]

Thanks!