Add inline caching for T2C indirect jump

[jserv]

This implements inline cache optimization for T2C-compiled code:

• Add inline_cache struct with (key, entry) pairs
• Fast path: check inline cache before seqlock-based jit_cache lookup
• Slow path: full jit_cache lookup + update inline cache on hit
• Skip ISB on ARM64 for inline cache hits (already executed this target)
• Clear inline cache on SFENCE.VMA, FENCE.I, and block eviction

Benchmark: 44% faster on ARM64 (ISB avoidance), ~4% on x86_64.

---

Summary by cubic

Add inline caching for T2C indirect jumps to speed up hot paths by bypassing jit_cache and skipping ARM64 ISB on hits. This reduces indirect jump overhead and lowers contention.

• New Features

  • Per-instance inline_cache storing (key, entry) pairs for indirect call sites.
  • Fast path: check inline_cache first; direct call on hit.
  • Slow path: seqlock jit_cache lookup; populate inline_cache on success; ISB only on ARM64 misses.
  • Invalidation on SFENCE.VMA (full/page), FENCE.I, and block eviction.
  • Initialization and cleanup wired into rv_create/rv_delete and code_cache_flush.

• Performance

  • 44% faster on ARM64 (ISB avoidance on hits).
  • ~4% faster on x86_64.

^{Written for commit df9727a. Summary will update on new commits.}

[cubic-dev-ai]

No issues found across 8 files

[jserv]

Benchmarks

Details

Benchmark suite	Current: df9727a	Previous: c5a3ee4	Ratio
Dhrystone	1593.333 DMIPS	1631.333 DMIPS	1.02
CoreMark	947.117 iterations/sec	946.285 iterations/sec	1.00

This comment was automatically generated by workflow using github-action-benchmark.