Custom CowStr type

[kmaasrud]

I've started implementing the CowStr type, heavily based on pulldown_cmark's similar type. Most necessary traits should be in place, but I have not yet tried introducing it to the codebase (currently, I am just importing it via a custom name JotdownCowStr.)

[hellux]

e.g. readme/all benchmarks seem to suggest no significant benefit currently (4d4d1c1 compared to 16491a4):

block
  Instructions:             1212464 (+0.164150%)
  L1 Accesses:              1607567 (+0.196083%)
  L2 Accesses:                 1810 (-0.330396%)
  RAM Accesses:                2181 (+0.553250%)
  Estimated Cycles:         1692952 (+0.209303%)

block_inline
  Instructions:             5222063 (-0.180426%)
  L1 Accesses:              7517519 (-0.143724%)
  L2 Accesses:                 3109 (-11.39926%)
  RAM Accesses:                2549 (+0.750988%)
  Estimated Cycles:         7622279 (-0.159213%)

full
  Instructions:             5719667 (-0.113706%)
  L1 Accesses:              8098320 (-0.031009%)
  L2 Accesses:                 5213 (-21.37255%)
  RAM Accesses:                3023 (+0.800267%)
  Estimated Cycles:         8230190 (-0.106288%)

full/readme             time:   [226.23 µs 226.28 µs 226.33 µs]
                        thrpt:  [53.096 MiB/s 53.108 MiB/s 53.120 MiB/s]
                 change:
                        time:   [+2.3752% +2.4656% +2.5533%] (p = 0.00 < 0.05)
                        thrpt:  [-2.4897% -2.4063% -2.3201%]
                        Performance has regressed.

having a custom type adds some complexity and inconvenience, so it should have a noticeable performance benefit if we want to go through with it.

might be possible to look at benchmarks where there are more short cowstrings, though.

[hellux]

should be possible to avoid allocation for concatenation when we have an inline? e.g. in Attributes::insert:

            if key == "class" {
                *prev = format!("{} {}", prev, val).into();
            } else {
                *prev = val;
            }

[kmaasrud]

I think only the CowStr itself is needed for it to be functional, but we would expect no change in performance then (unless e.g Box is used for owned variant). But adding the Inlined variant might not make any difference unless the push_str is added. replace shouldn't have any impact as it is not used internally.

That's why I was thinking that a complete CowStr (with the Inlined variant and String contained within the Owned variant) would be good to benchmark against our existing Cow. After that, I could add another commit that replaces Owned(String) with Owned(Box) and assess the performance of that change separately.

[kmaasrud]

@hellux Here is the full/all benchmark comparing the commit just before and after the introduction of the new CowStr. I have not added any new benchmarks with many short strings, so this seems promising, doesn't it?

Commit	Throughput	Change
14f6c41 inline: store attributes in vec (squash)	67.958 MiB/s	-
18363ea feat: add custom CowStr type	73.363 MiB/s	+8.0656%

[hellux]

@hellux Here is the full/all benchmark comparing the commit just before and after the introduction of the new CowStr. I have not added any new benchmarks with many short strings, so this seems promising, doesn't it?
Commit Throughput Change
14f6c41 inline: store attributes in vec (squash) 67.958 MiB/s -
18363ea feat: add custom CowStr type 73.363 MiB/s +8.0656%

I'm not able to reproduce similar results on my machines:

AMD 5950X, Zen 3:

commit                                                      cycles                 throughput
14f6c4188 inline: store attributes in vec (squash)          6127389 (-4.265530%)   55.252
c5814c912 feat: add custom CowStr type                      6251397 (+2.023831%)   54.922
200091463 cowstr: add docstrings                            6251397 (No change)    54.820
f067cbd17 cowstr: add tests for push, push_str and replace  6251397 (No change)    55.039

Intel i5-8250U, Coffee Lake:

commit                                                    cycles                 throughput
14f6c41 inline: store attributes in vec (squash)          6564735 (+0.965820%)   34.945
c5814c9 feat: add custom CowStr type                      6526309 (-0.585340%)   34.929
2000914 cowstr: add docstrings                            6526309 (No change)    34.891
f067cbd cowstr: add tests for push, push_str and replace  6526309 (No change)    34.891

I wouldn't expect an improvement to the full/all benchmark unless the cowstr size has been reduced, behavior should be similar in most cases. Should only notice a difference if there are many small cowstrs, which I think we need to add a benchmark for.

Would be interesting to see benchmarks when the cowstr size has been reduced, though.