perf: Reduce memory allocations using ReadOnlySpan and String interning

[WanjohiSammy]

Issues

This pull request fixes #3352.

Description

Trying to eliminate string allocations where possible using ReadOnlySpan and String interning

Benchmark

1. Before

2. After

Benchmark project

/WanjohiSammy/ODataJsonReaderBenchmarks-3

Input data:
https://github.com/WanjohiSammy/ODataJsonReaderBenchmarks-3/blob/main/ODataJsonReaderBenchmarks/SamplePayloadWithValues.json

Checklist (Uncheck if it is not completed)

• Test cases added
• Build and test with one-click build and test script passed

Additional work necessary

If documentation update is needed, please add "Docs Needed" label to the issue and provide details about the required document change in the issue.

Repository notes

Team members can start a CI build by adding a comment with the text /AzurePipelines run to a PR. A bot may respond indicating that there is no pipeline associated with the pull request. This can be ignored if the build is triggered.

Team members should not trigger a build this way for pull requests coming from forked repositories. They should instead trigger the build manually by setting the "branch" to refs/pull/{prId}/merge where {prId} is the ID of the PR.

[habbes]

Do we care about case sensitivity in this case?

[WanjohiSammy]

Since we are checking the word or string as it comes, we just need to maintain case sensitivity.

[habbes]

I like the idea of this method, to avoid allocating strings from a span for common values (it doesn't seem limited to properties as the name of the methods suggest). I like that we check the length, so it's cheap to find the suitable bucket.

My concern is around the cases where we have length >= 5 and multiple candidates in the same block. This means we may have to do multiple string comparisons. I guess the obvious optimization here is to make sure that the most likely string candidate appears first in the if statement. Of course this can vary from one service to another. How are we sure this is optimal. I'd like to see what the cost of value of length 5+ is when it's not one of the list candidates (meaning that we would have to compare all candidates, i.e 9 * 5 = 45 characters) each time then fall back. I would like to compare the cost of doing that vs creating the string directly.

Another option would be to create static hashset of common words longer than some threshold (e.g. >= 5), the threshold can be picked by experimentation. Then do a lookup in the hashset and compare the cost of doing that to the cost of comparing all items. Not sure which would be faster. Could be worth the experiment, especially if this method is called each time we want to convert a span to a string. We could also try to replace the hashset with a simpler hash-based datastructure.

[WanjohiSammy]

/AzurePipelines run

[azure-pipelines]

No pipelines are associated with this pull request.

[WanjohiSammy]

/AzurePipelines run

[WanjohiSammy]

/AzurePipelines run

[WanjohiSammy]

/AzurePipelines run