Fix recent IndexOf regressions

[MihaZupan]

#78861 introduced new PackedIndexOf implementations for chars that have higher throughput for ASCII values.

It also regressed some scenarios as tracked by #80441. This PR fixes most of those regressions by:

• dcc8763 which brings back the old codegen for IndexOf over non-chars (e.g. Span<byte>.IndexOf) by doing the type check earlier, before we call into CanUsePackedIndexOf.
• 001b500 which avoids the overhead of checking whether a value can use the packed implementation where we wouldn't benefit anyway.

Example improvement for System.Net.Http because of dcc8763

Top method improvements (bytes):
    -22 (-2.46% of base) : System.Net.Http.dasm - System.Net.Http.HttpConnection:ParseHeadersCore(System.Span`1[ubyte],System.Net.Http.HttpResponseMessage,bool):System.ValueTuple`2[bool,int]:this
    -5 (-3.45% of base) : System.Net.Http.dasm - System.Net.Http.HttpConnection:<FillForHeadersAsync>g__TryFindEndOfLine|83_1(System.ReadOnlySpan`1[ubyte],byref):bool
    -5 (-1.99% of base) : System.Net.Http.dasm - System.Net.Http.HttpConnection:ParseStatusLine(System.Net.Http.HttpResponseMessage):bool:this
    -5 (-1.82% of base) : System.Net.Http.dasm - System.Net.Http.HttpConnection:TryReadNextChunkedLine(byref):bool:this

or on IndexOfAny for bytes:

Method	Toolchain	Mean	Error	Ratio
IndexOfAnyThreeValues	before	2.985 ns	0.0166 ns	1.00
IndexOfAnyThreeValues	main	4.085 ns	0.0573 ns	1.37
IndexOfAnyThreeValues	pr	3.051 ns	0.0118 ns	1.02

[ghost]

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

Issue Details

---

#78861 introduced new PackedIndexOf implementations for chars that have higher throughput for ASCII values.

It also regressed some scenarios as tracked by #80441. This PR fixes most of those regressions by:

• dcc8763 which brings back the old codegen for IndexOf over non-chars (e.g. Span<byte>.IndexOf) by doing the type check earlier, before we call into CanUsePackedIndexOf.
• 001b500 which avoids the overhead of checking whether a value can use the packed implementation where we wouldn't benefit anyway.

Example improvement for System.Net.Http because of dcc8763

Top method improvements (bytes):
    -22 (-2.46% of base) : System.Net.Http.dasm - System.Net.Http.HttpConnection:ParseHeadersCore(System.Span`1[ubyte],System.Net.Http.HttpResponseMessage,bool):System.ValueTuple`2[bool,int]:this
    -5 (-3.45% of base) : System.Net.Http.dasm - System.Net.Http.HttpConnection:<FillForHeadersAsync>g__TryFindEndOfLine|83_1(System.ReadOnlySpan`1[ubyte],byref):bool
    -5 (-1.99% of base) : System.Net.Http.dasm - System.Net.Http.HttpConnection:ParseStatusLine(System.Net.Http.HttpResponseMessage):bool:this
    -5 (-1.82% of base) : System.Net.Http.dasm - System.Net.Http.HttpConnection:TryReadNextChunkedLine(byref):bool:this

or on IndexOfAny for bytes:

Method	Toolchain	Mean	Error	Ratio
IndexOfAnyThreeValues	before	2.985 ns	0.0166 ns	1.00
IndexOfAnyThreeValues	main	4.085 ns	0.0573 ns	1.37
IndexOfAnyThreeValues	pr	3.051 ns	0.0118 ns	1.02

Author:	MihaZupan
Assignees:	MihaZupan
Labels:	area-System.Memory
Milestone:	8.0.0

[dakersnar]

LGTM