JIT: Improve inliner - static readonly fields, pinvokes, Span::Length

[EgorBo]

Contributes to #59002

This PR:

1. Recognizes static readonly fields of primitive types as potential constants => finds more potential foldable expressions/branches.
2. Recognizes "pinvokes" in calls in order to avoid inlining of callees with them unless caller already have one - in this case we give callees with pinvokes additional boost to share the transition machinery.
3. Mark ReadOnlySpan/Span::get_Length as [Intrinsic] - this simple change alone generates -6037 negative diff! 😮
4. Drop too optimistic call(CNS, ...) -> CNS heuristic. Currently, if jit sees a call and suspects its first argument is a const - it assumes it is likely to become a constant too - this produced many false-positive "foldable branches" and a large diff as the result.
5. Increases multiplier for foldable branches and intrinsics (e.g. fixes inlining of IsLatin1 into IsDigit here)

Jit-diff after these changes: Total bytes of delta: -292814 (-0.71 % of base) (jitutils, -f, crossgen2) - will post more detailed diffs soon.

TODO:

• Test impact on throughput (from getMethodAttribs and getStaticFieldCurrentClass)
• Test impact on TE benchmarks
• Save final codegen size (or a least a single bit "codegen is likely small") during R2R to then pick up by tier1.
• ARM64-specific tuning (ABI, more registers, etc).

[ghost]

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

Issue Details

---

Contributes to #59002

This PR:

1. Recognizes static readonly fields of primitive types as potential constants => finds more potential foldable expressions/branches.
2. Recognizes "pinvokes" in calls in order to avoid inlining of callees with them unless caller already have one - in this case we give callees with pinvokes additional boost.
3. Mark [ReadOnly]Span.get_Length as intrinsic - this simple change alone generates a large negative diff!
4. Drop too optimistic "call(CNS, ...) -> CNS" heuristic. Currently if jit sees a call and suspect its first argument is a const - it assumes it is likely to become a consant too - this produced many false-positive "foldable branches".
5. Increase multiplier for foldable branches and intrinsics.

Jit-diff after these changes: Total bytes of delta: -292814 (-0.71 % of base) (jitutils, -f, crossgen2) - will post more detailed diffs soon.

TODO:

• Test impact on throughput (from getMethodAttribs and getStaticFieldCurrentClass)
• Test impact on TE benchmarks
• Save codegen size (or a least a single bit "codegen is likely small") during R2R to then pick up by tier1.
• ARM64-specific tuning (ABI, more registers, etc).

Author:	EgorBo
Assignees:	-
Labels:	area-CodeGen-coreclr
Milestone:	-

[stephentoub]

Mark ReadOnlySpan/Span::get_Length as [Intrinsic] - this simple change alone generates -6037 negative diff!

😉

[marek-safar]

/cc @BrzVlad

[ghost]

Draft Pull Request was automatically closed for inactivity. Please let us know if you'd like to reopen it.