perf: build call frames in place and inline apply_int

[linyiru]

Closes #437. Part of the per-call fixed cost work (#378, #385).

What

• push_frame_in_place (vm.rs) writes a Frame straight into frames' spare capacity. Five cached method-call push sites in dispatch.rs use it: serve_explicit_recv_method, try_invoke_explicit_recv_block_cached, try_invoke_class_singleton_cached, try_invoke_fixed_method_from_stack, and try_invoke_norecv_block_cached.
• BinOpKind::apply_int is #[inline(always)].

Why

On starship (x86), while i < n; o.n; i += 1; end retires ~1386 instructions per iteration vs ~328 under CRuby, at a similar IPC. Almost all of the gap is instruction count, which the follow-up ADR will address. Two store-forwarding stalls were the exception, and both are cheap to remove:

• Frame push. Vec::push takes the 192-byte Frame by value. The frame was assembled on the stack with narrow field stores, then copied with a libc memcpy whose wide loads could not be store-forwarded. That memcpy was ~14% of cycles.
• apply_int. It was not inlined into BinOpLocalLocal and returned its Option<Value> through memory.

Numbers

starship x86. taskset -c 5 perf stat -e cycles:u, 30M-iteration o.n loop. Master is 69845c0; both binaries were built with line-tables-only.

binary	cycles (×1e9)	median
master	11.70, 11.88, 12.05, 11.81 (earlier session: 11.72–12.14)	~11.8
this PR	10.70, 10.73, 10.71, 11.87 (earlier session: 10.70–10.82, one 12.17)	~10.7 (−9%)

This PR's results are bimodal: about 1 run in 5 lands at master's level. I have not confirmed why; code or stack alignment is a guess.

On the same machine, the whole shape-bench.rb run took 30.2 → 29.4 e9 cycles (−2.5%, 3 rounds each). Wall-clock call0 on starship stays within noise.

Mac Studio (M2 Max), 5 interleaved rounds, load ~10.

	master	this PR
call0 o.n	73.0–75.8 ns	70.4–73.3 ns
call0 o.r, o.n1, o.n2, loop	—	within noise
shape-bench (0-arg / 2-arg / kwargs / **opts / opt+rest)	—	within noise
Rails hello, median req/s	1011–1065	1008–1101

The Rails numbers are within noise; this change is too small to move Rails measurably. Apple silicon gaining less than x86 is consistent with its cheaper store-forwarding.

Verification (Mac)

• diff_cruby 1103/1103 in each of four modes: default, STRESS_GC=1, RUBYRS_JIT_NATIVE=1, and tier 2 (THRESHOLD=1)
• jit unit, lib unit (153), --no-default-features check
• clippy 1.95 (-D warnings), both the rubyrs everything,jit-native leg and the workspace leg
• --test embed 421/421
• perf/check 9/9 within budget
• lint-gc-rooting and lint-doc-orientation

Not in this PR

• The block-frame push sites (push_block_frame, invoke_block1/2, …) and the remaining method push sites. They can follow once this lands.
• The structural per-op and per-call costs: the dispatch loop, the Return path, and the Frame size. An ADR draft will plan these next.

🤖 Generated with Claude Code

[github-actions]

gapscan PR diff

This PR doesn't touch any file that affects gapscan classification — SUPPORTED_PRISM_NODES / RIDES_ALONG_PRISM_NODES are unchanged. Skipped the base build and the 10 per-target scans.

See docs/gap-reports/ for the dataset and methodology.

[Copilot]

Copilot encountered an error and was unable to review this pull request. You can try again by re-requesting a review.

[Copilot]

Copilot review overview

🟢 Approval recommended

No unresolved review issues remain.

Review effort: Lite
Findings: None