The crappiness of shrink-wrapping in gcc and clang (but especially clang) annoys me a lot. It feels like there should be a quite decent amount of general performance to be gained from properly pushing more into slow paths (or, not necessarily even slow paths, but generally paths with high register pressure / uninlined function calls), never mind calling conventions in general.

On the push impl in the article - for non-x86 (and perhaps even on x86 for performance, though not size/instruction count) it'd be better to allow the size increment to reuse the size read done by the capacity check; with C++'s lack of suitable aliasing information, the interleaved memcpy/store prevents the compiler from deciding this itself.

So one thing I thought of when reading this is that very often it is known ahead of time (at runtime or even compile-time) how many push_backs will be done. The programmer could make a reserve call but doesn't bother since the efficiency gain is minimal.

The gain is minimal for doing this optimization at one location. But doing it everywhere, that could matter. Pushing back in a loop could maybe be optimized to a single allocation and a memcopy.