Fix for 126457

[dhartglassMSFT]

For a bool-typed inlinee that contains a tail call, where the return is in a seperate block from the tailcall (for a single-return method, for example), the inliner inserts an int->ubyte->int normalizing cast

*  RETURN  int
\--*  CAST    int <- ubyte <- int
   \--*  LCL_VAR int    V04 tmp3

The tail-call validation in morph doesn't expect a cast there. However, a normalizing cast here may be safely ignored for tailcalls because
1 - If the calling function was not inlined, the normalizing cast is intead inserted later in morph after the tailcall transformation, and
2 - Callees are responsible for this adjustment anyway so the tailcall return value would have already been normalized

fixes #126457

[Copilot]

Pull request overview

Addresses a CoreCLR JIT debug assertion during tailcall validation when inlining introduces small-type-normalizing casts (e.g., bool normalization) between a tailcall and the return, which was triggering failures in tvos-arm64 Debug crossgen2/R2R scenarios (issue #126457).

Changes:

• Update fgValidateIRForTailCall (DEBUG-only validation) to accept/skip normalizing GT_CAST nodes when tracing a tailcall’s return value to the final GT_RETURN.
• Add a new JIT regression IL test (Runtime_126457) to exercise the inlinee+tailcall+separate-return-block pattern implicated in the crash.

Reviewed changes

Copilot reviewed 3 out of 3 changed files in this pull request and generated no comments.

File	Description
src/coreclr/jit/morph.cpp	Extends tailcall IR validation to tolerate small-type normalizing casts introduced by the inliner.
src/tests/JIT/Regression/JitBlue/Runtime_126457/Runtime_126457.ilproj	Adds a new IL-based JIT regression test project configuration.
src/tests/JIT/Regression/JitBlue/Runtime_126457/Runtime_126457.il	Adds IL repro that drives inlining + tailcall + return-flow shape relevant to the bug.

[jakobbotsch]

This has to be the right type of normalizing cast -- e.g. if the inserted cast is zero-extending, then the call should also have a zero-extended return of the right size. I think you can just replace this with tree->AsCast()->CastToType() == m_tailcall->gtReturnType.

[jakobbotsch]

Since we have fgCastNeeded you can also just use that here: fgCastNeeded(m_tailcall, tree->CastToType()))

[jakobbotsch]

Why move this here? It looks inconsistent with how the other cases are handled

[dhartglassMSFT]

Are you talking about this cast stripping logic? We need this here because the GT_RETURN node calls into ValidateUse

or are you referring to the fgCastNeeded check? I can move that back into the assert, thats probably better

[Copilot]

Pull request overview

Copilot reviewed 3 out of 3 changed files in this pull request and generated 2 comments.

[Copilot]

New IL source file is missing the standard .NET Foundation/MIT license header comment that most JIT regression .il sources include. Please add the license header at the top of this file to match repo conventions.

[Copilot]

This regression is specifically about tailcall validation; to make the test deterministic across JIT configs, consider forcing the call to be a tailcall by adding the tail. IL prefix before the callvirt to ISite::tailcall() (many existing tailcall IL regressions do this). Without it, the JIT may decide not to tailcall-optimize and the test may not cover the intended scenario.

Suggested change

	CALL:
	CALL:
	tail.