Bad codegen in Boolean::TryParse

[MichalStrehovsky]

TryParse(ReadOnlySpan<char> value, out bool result) accesses two string literals and makes a Span out of them - "True" and "False".

RyuJIT with optimizations enabled will reuse the first Span for "True" instead of making a new one for "False".

To repro, simply compile Console.WriteLine(Boolean.Parse("False")); with optimizations enabled. It will throw an exception because we don't match "False" to "False", but try to match "True" twice.

My guess is there's something wrong with this piece of code in RyuJIT:

https://github.com/dotnet/coreclr/blob/0f0320e58fd006a02cdecf7ae45426f54da333e5/src/jit/gentree.cpp#L6426-L6430

[MichalStrehovsky]

Blocking AvaloniaUI/Avalonia#1476.

@sandreenko Is my analysis remotely correct? Could you point me in the right direction?

JitDump:
jitdump.txt

The interesting lines are the ones with (reloc 0x421150) and (reloc 0x421158) - those are the string literal loads. Notice how we touch the second reloc, but then don't do anything with it.

[sandreenko]

Thanks for your analysis, it looks correct, I will try to fix it asap.

[sandreenko]

It is a value numbering problem where we for some reasons have wrong null VNs for both true and false cases. I am working on the fix.