JIT: Cold code inlining decisions are incorrectly used later

[pentp]

Description

It seems that inlining decisions made in cold code blocks are being remembered by JIT and blindly used later in completely unrelated methods.

Reproduction Steps

https://godbolt.org/z/9zas8rc6h

Expected behavior

C:InnocentMethod implicitly converts the provided Span into a ReadOnlySpan without using a op_Implicit call.

Actual behavior

Trivially inline-able methods (even intrinsic methods) are not inlined in normal methods.

Regression?

.NET 7 seems to not have this issue.

Known Workarounds

No reliable workarounds, can reduce chances by avoiding any potential inlining in cold code blocks or re-ordering method JIT compilation order.

[dotnet-policy-service]

Tagging subscribers to this area: @JulieLeeMSFT, @jakobbotsch
See info in area-owners.md if you want to be subscribed.

[pentp]

There's a second related issue that often masks this bug - using throw helpers to throw exceptions causes the throwing code blocks to have non-zero weights (0.5 seems to be common) which seems to allow inlining to work more normally and thus not poison it for later method compilations.
However this seems like an arbitrary and useless distinction between throw helpers and direct throws, which affects not only inlining, but also PerfScore calculation for example and IMHO cold blocks that throw an exception using a throw helper should also get a zero weight / zero PerfScore.

Should I open a separate issue for that?

[AndyAyersMS]

Are you seeing this same codegen outside of godbolt/crossgen2?

@EgorBo FYI

[pentp]

It's not crossgen2 actually, regular JIT, I forgot to update the window titles. I first noticed it when locally generating JIT dumps, with tiered compilation disabled (maybe that has some role in this also).