Add public methods that remove path redundant segments

[carlossanlop]

Fixes #2162

[MihaZupan]

The purpose of TryRemoveRedundantSegments accepting a Span destination is to avoid allocations. As such, it can't delegate to the string-returning methods and only do a copy.

[carlossanlop]

I'll close this issue while it's in draft mode. I'll reopen when the unit tests are ready.

[carlossanlop]

@jkotas Need your help - I cannot rebase because your change #31991 deleted the file ref/System.Runtime.Extensions.cs.

Where should I put my new method references now?

[jkotas]

Where should I put my new method references now?

src/libraries/System.Runtime/ref/System.Runtime.cs

[jkotas]

ValueStringBuilder is meant to be used with stackallocated buffer for sizes up to certain limit. Getting a pooled buffer has too much overhead for small buffers.

[carlossanlop]

@JeremyKuhne - I'm investigating these tests where I put a comment next to them, because their behavior was slightly different than expected.

I also noticed the resulting path is different in Windows than in Unix, because of the separator char used.

[carlossanlop]

I talked with Jeremy about the expected behavior on each edge case. I am adding tests to verify them.

[carlossanlop]

Currently blocked by a bug I found in Path.GetPathRoot, which is the reason why some of the unit tests are failing in the CI: #35260.
I'm investigating it.

[jkotas]

Unbounded stackallocs are prohibited in dotnet/runtime libraries

[danmoseley]

@stephentoub I don't see record of a discussion about trying to catch this with an analyzer. It's not generally possible to be sure (either way) whether it's unbounded. But an analyzer that forced a pattern of passing a constant value would be doable.

Looking at all our stackallocs, almost all of them already pass a constant directly or via a constant field or local. The few exceptions are not self evidently safe by eyeball, eg.,

        private static unsafe int EncryptDecryptHelper(OP op, ISSPIInterface secModule, SafeDeleteContext context, Span<SecurityBuffer> input, uint sequenceNumber)
        {
            Interop.SspiCli.SecBufferDesc sdcInOut = new Interop.SspiCli.SecBufferDesc(input.Length);
            Span<Interop.SspiCli.SecBuffer> unmanagedBuffer = stackalloc Interop.SspiCli.SecBuffer[input.Length];

and it would be easy for one of them to be wrong. Would it be reasonable to have an analyzer that required them to be rewritten in constant terms, or at least something the analyzer could recognize like

            Span<Interop.SspiCli.SecBuffer> unmanagedBuffer = stackalloc Interop.SspiCli.SecBuffer[Math.Min(input.Length, BufferSize)];

[danmoseley]

Moved to #35382

[carlossanlop]

I'll close this while I address the Unix failures.