Adam Reichold

> A 1.07x regression is enough to give me pause here, particularly on a huge haystack where noise doesn't tend to be as much as a factor. I would not...

> The Assembly generated might provide a clue. Looking at the relevant basic blocks for `memchr::arch::x86_64::memchr::memchr2_raw::find_avx2` on `master`: ```asm .LBB12_9: vmovdqa ymm3, ymmword ptr [rdx] vmovdqa ymm6, ymmword ptr [rdx...

One could unroll that loop one more time while staying within the same (vector) register budget. Benchmarks are similarly inconclusive though. Some wins, some losses.

There is nothing built-in at the moment. If you want this during development, you could try general purposes malloc instrumentation like [Heaptrack](https://github.com/KDE/heaptrack). If you need this for deployments, you would...

I think this would work better as a compile-time switch so that users could avoid embedding a representation of the models that they will not use.

> because values can't, generally, be mutated in-place because the serialized size would have to match Not wanting to relitigate this whole thing, but I found that this case is...