⚙️ The optimization

remove_copy and unique_copy are different from their non-_copy counterparts in that they don't have room they are allowed to overwrite. This means we can't directly store results from vector registers.

The previous attempt #5062 tried to use masked stores to bypass that limitation. Unfortunately, this doesn't perform well for some CPUs. Also the minimum granularity of AVX2 masked store is 32 bits, so it would not work for smaller elements.

This time, temporary storage comes to rescue. The algorithms already use some additional memory (the tables), so why wouldn't it use a bit more. I arbitrarily picked 512 bytes, should be not too much. Each time the temporary buffer is full, it can be copied to the destination with memcpy, it should be fast enough for this buffer size.

🚫 No find before remove_copy

In #4987, it was explained that doing find before remove is good for both correctness and performance. Originally it was in vectorization code, but during the review @StephanTLavavej observed that it is done in the headers already (#4987 (comment)).

For remove_copy it is not necessary for correctness, and may be harmful for performance. find would need copy in addition, this will be double pass on the input, which can make the performance worse for large input and memry-bound situation.

We may have special handling of the range before the first match in vectorization code, this is another story, and it would not be harmful, but I'm not doing this in the current PR. Maybe later.

So, as we have not called find, and so have not checked if value type can even match iterator value type, we need this _Could_compare_equal_to_value_type check here.

✅ Test coverage

Shared with non-_copy counterparts to save total tests run time and some lines of code, at the expense with otherwise unnecessary coupling.

We check both modified and unmodified destination parts, to make sure unmodified indeed didn't modify.

⏱️ Benchmark results

🥇 Results interpretation

Good improvement!

Not as good as for non-_copy counterparts though, as memcpy takes some noticeable time.

The usual codegen gremlins that cause results variation are observed for non-vectorized tight loops. I've marked the most notorious one with clown. I can't explain that anomality.