Meta: surpass CRuby on Rails-shaped workloads (perf roadmap A → B → C) #378

Where we are (measured 2026-09-23, Mac Studio M2 Max)

Rails 8.1 hello world through Rack::MockRequest, steady state:

CRuby TracePoint census per request: 1427 Ruby calls, 1654 C calls, 223 block calls, 433 allocations. So rubyrs spends ≈380 ns per event, versus ≈70 ns for CRuby.

The gap is uniform, not a few hot spots

poc/rails-spike/shape-bench.rb measures ns/op after subtracting the empty-loop cost. Selected rows:

Plan

A — targeted fixes. Estimated 10–20% on Rails. One issue each:

• Native Process.clock_gettime with a real monotonic clock — perf: native Process.clock_gettime with a real monotonic clock #379
• Native Symbol#start_with? / #end_with? — perf: native Symbol#start_with? / #end_with? #380
• Fast paths for Thread.current / Fiber.current / Mutex#synchronize — perf: fast paths for Thread.current, Fiber.current, Mutex#synchronize #381
• Interpreter block invocation from native iterators (port the tier-2 path) — perf: interpreter block invocation from native iterators #382
• Cheap raise: lazy backtrace, id-based class matching — perf: cheap raise/rescue (lazy backtrace, id-based class matching) #383

B — structural interpreter work. Needed to reach CRuby-interp parity. Tracked in #384; B2 needs an ADR:

• B1: kwargs passing without a Hash allocation
• B2: 8-byte Copy Value (no Rc in the enum)
• B3: per-call-site cached method + call handler (skip the do_call cascade); SymId compares instead of name-string compares on hot paths
• B4: Proc-free yield/blocks; frame elision (ADR 0037)

C — JIT that beats YJIT. After B. Tracked in #385.

Process

• One PR per item. Each PR includes shape-bench before/after numbers plus the Rails bench delta, and passes cargo test, ruby_spec and diff_cruby.
• Re-run the per-request census after A to re-rank B.