Revert "Reduce unsafe usage in TextInfo.cs"

[jkotas]

Reverts #122116

Investigating #123667

[jkotas]

rt-sz report MichalStrehovsky/rt-sz#203

[MihaZupan]

NativeAOT failures were not all timeouts

Sorry about that.
Would the acceptable alternative be to introduce a dedicated non-generic struct just for this call? Seems unfortunate we'd have to avoid tuple syntax like this.

[jkotas]

Would the acceptable alternative be to introduce a dedicated non-generic struct just for this call?

It is one of the strategies that we have used in other similar situations to avoid the code bloat.

For this specific case, we have 30+ places where we do string allocation using unsafe FastAllocateString in core globalization and formatting routines to avoid the code bloat and other overheads from the safe variant. I do not think we need to go out of our way to avoid the unsafe code for this one. We should aim to have a solution that scales without introducing regressions for majority of places that use FastAllocateString. The solution likely includes #85014 . cc @EgorBo

[Copilot]

Pull request overview

This PR reverts PR #122116 which reduced unsafe usage in TextInfo.cs. The revert is being done to investigate a NativeAOT binary size regression on osx-x64 (issue #123667), where the size increased from ~1.5MB to 1.551MB, causing size tests to fail.

Changes:

• Reverts ToLowerAsciiInvariant(string) method from safe span-based implementation using string.Create back to unsafe pointer-based implementation using fixed statements and manual pointer arithmetic