Faster IndexOfAny for IgnoreCase Ascii letters

[MihaZupan]

The idea is that if we're searching for both cases of an Ascii letter, we can do it by masking off the bit where they differ and only search for one. E.g. (input | 0x20) is 'a' or 'b' instead of input is 'a' or 'A' or 'b' or 'B'.

This PR adds a variant of packed IndexOfAny(char) and IndexOfAny(char, char) that does the | 0x20 input transformation and teaches SearchValues to use it.
I've also updated the RegexCompiler to use SearchValues for sets of 4/5 values like the source generator does, so that it can make use of this change.

public class SearchValuesPackedIgnoreCase
{
    private static readonly SearchValues<char> s_asciiTwoLetters = SearchValues.Create("Aa");
    private static readonly SearchValues<char> s_asciiFourLetters = SearchValues.Create("AaBb");
    private static readonly string s_text = new string('\n', 20_000);

    [Benchmark] public int IndexOfOneIgnoreCase() => s_text.AsSpan().IndexOfAny(s_asciiTwoLetters);
    [Benchmark] public int IndexOfTwoIgnoreCase() => s_text.AsSpan().IndexOfAny(s_asciiFourLetters);
}

Method	Toolchain	Mean	Error	Ratio
IndexOfOneIgnoreCase	main	583.5 ns	3.56 ns	1.00
IndexOfOneIgnoreCase	pr	499.5 ns	1.71 ns	0.86
IndexOfTwoIgnoreCase	main	742.5 ns	4.02 ns	1.00
IndexOfTwoIgnoreCase	pr	642.2 ns	5.54 ns	0.87

Method	Toolchain	Pattern	Options	Mean	Error	Ratio
Count	main	(?i)Sherlock|Holmes	Compiled	676.61 us	1.858 us	1.00
Count	pr	(?i)Sherlock|Holmes	Compiled	623.75 us	1.762 us	0.92
Count	main	[Zz]	Compiled	32.16 us	0.126 us	1.00
Count	pr	[Zz]	Compiled	30.39 us	0.112 us	0.95
Count	main	[ZzQq]	Compiled	63.99 us	0.289 us	1.00
Count	pr	[ZzQq]	Compiled	46.14 us	0.131 us	0.72

The bigger difference for [ZzQq] is switching from a regular 4-value IndexOfAny to a 2-value packed one.

I tested the same approach for IndexOfAny(char, char, char) and IndexOfAnyInRange, but the throughput there is pretty much on par with or slower than IndexOfAnyAsciiSearcher, so I avoided adding more code for it.

[ghost]

Tagging subscribers to this area: @dotnet/area-system-buffers
See info in area-owners.md if you want to be subscribed.

Issue Details

---

The idea is that if we're searching for both cases of an Ascii letter, we can do it by masking off the bit where they differ and only search for one. E.g. (input | 0x20) is 'a' or 'b' instead of input is 'a' or 'A' or 'b' or 'B'.

This PR adds a variant of packed IndexOfAny(char) and IndexOfAny(char, char) that does the | 0x20 input transformation and teaches SearchValues to use it.
I've also updated the RegexCompiler to use SearchValues for sets of 4/5 values like the source generator does, so that it can make use of this change.

public class SearchValuesPackedIgnoreCase
{
    private static readonly SearchValues<char> s_asciiTwoLetters = SearchValues.Create("Aa");
    private static readonly SearchValues<char> s_asciiFourLetters = SearchValues.Create("AaBb");
    private static readonly string s_text = new string('\n', 20_000);

    [Benchmark] public int IndexOfOneIgnoreCase() => s_text.AsSpan().IndexOfAny(s_asciiTwoLetters);
    [Benchmark] public int IndexOfTwoIgnoreCase() => s_text.AsSpan().IndexOfAny(s_asciiFourLetters);
}

Method	Toolchain	Mean	Error	Ratio
IndexOfOneIgnoreCase	main	583.5 ns	3.56 ns	1.00
IndexOfOneIgnoreCase	pr	499.5 ns	1.71 ns	0.86
IndexOfTwoIgnoreCase	main	742.5 ns	4.02 ns	1.00
IndexOfTwoIgnoreCase	pr	642.2 ns	5.54 ns	0.87

Method	Toolchain	Pattern	Options	Mean	Error	Ratio
Count	main	(?i)Sherlock	Compiled	57.02 us	0.121 us	1.00
Count	pr	(?i)Sherlock	Compiled	57.75 us	0.132 us	1.01
Count	main	(?i)Sherlock|Holmes	Compiled	676.61 us	1.858 us	1.00
Count	pr	(?i)Sherlock|Holmes	Compiled	623.75 us	1.762 us	0.92
Count	main	[Zz]	Compiled	32.16 us	0.126 us	1.00
Count	pr	[Zz]	Compiled	30.39 us	0.112 us	0.95
Count	main	[ZzQq]	Compiled	63.99 us	0.289 us	1.00
Count	pr	[ZzQq]	Compiled	46.14 us	0.131 us	0.72

The bigger difference for [ZzQq] is switching from a regular 4-value IndexOfAny to a 2-value packed one.

Author:	MihaZupan
Assignees:	MihaZupan
Labels:	area-System.Buffers
Milestone:	9.0.0

[tannergooding]

Is this pending anything or can it be merged?

Saw a couple questions and an ask for a comment to be added from Stephen, but its unclear if you want to handle that in a follow up or not.

[MihaZupan]

This one is waiting on me, I still have to follow up on Stephen's feedback

[ghost]

This pull request has been automatically marked no-recent-activity because it has not had any activity for 14 days. It will be closed if no further activity occurs within 14 more days. Any new comment (by anyone, not necessarily the author) will remove no-recent-activity.