I tested this with the MRP from #101494 on an M2 MBP. It improved frame time by about 0.1 ms (10.8 -> 10.7) However, there was enough fluctuation that this observed difference could have been a measurement error.

That being said, I am inclined to believe that there was an improvement since Vector showed up in a few places in my earlier profile. I suspect that the larger cost is the allocations rather than the CoW semantics, so this is only showing a small improvement.

Looks like it is failing checks on a signed/unsigned comparison

The only other change is a few places where we switch from checking if (some_int <0 || some_int > max_value) to if (uint32_t(some_int) > max_value) i.e. relying on integer overflow to avoid a check.

Does this measurably impact performance? IMO that worsens readability / communication of the intention of the code.
It forces the reader (me at least) to think about what's the purpose of the cast and it doesn't immediately bring "[0; max[ index check" to mind.

EDIT : I thought this would be no different either way, but it turned out to not be the case, see below, the behaviour is different depends on the sign, and whether the value is a constant, the compiler will only do the trick under specific circumstances.

https://godbolt.org/z/nTcf86Kos

c++:

#include <stdint.h>

bool func(int32_t input)
{
    if ((input < 0) || (input > 10))
    {
        return true;
    }
    return false;
}

assembly:

func(int):
        cmp     edi, 11 // one comparison
        setae   al // set if carry flag clear
        ret

@lawnjelly Compilers don't know this trick if max_value is an int variable. https://godbolt.org/z/crfKTP333

@lawnjelly Compilers don't know this trick if max_value is an int variable. https://godbolt.org/z/crfKTP333

Good point actually.

Also on further testing, turns out they also can't do the trick if m_size is unsigned and not a constant, because it could be outside the signed range, so they end up doing two branches like the worst case. 😮

So for it to be confident of the trick I think it has to be sure that both the value and the size are unsigned, otherwise all bets are off it seems.

Thanks!