perf(vm): eliminate recurring stack and return allocations

This commit is contained in:
2026-10-14 10:54:11 +01:00
parent e09f1197bc
commit 46bbad19c2
8 changed files with 308 additions and 30 deletions
@@ -67,3 +67,28 @@ String concatenation still creates a new string when the language semantics requ
- The right performance question is often "where is the repeated cost?" rather than "where is any cost at all?"
- Immutable payload sharing is a strong default when public semantics must remain stable and hot-path copies are the real issue.
- Internal telemetry should help engineering decisions without silently becoming a product contract.
## Successful execution can allocate without materializing guest values
The later stack/return correction (DEC-0061) demonstrated another source of
recurring native allocations: eager diagnostic Strings and disposable return
storage. A warmed scalar call/return loop allocated and freed 3,000 times per
1,000 iterations; lazy diagnostics and in-place return transfer reduced both
counts to zero in debug and release. The original string materialization evidence
above remains a separate observation.
Validate before moving owners. Rotating only the returning frame's suffix and
then truncating locals preserves ordered logical groups, Arc identity and heap
aliases without cloning or scratch storage. Remove the frame after transfer so
the existing last-root safepoint sees the results. GC scratch refusal is a later
failure boundary: transferred roots must remain intact and sweeping must not
start after failed marking admission.
Account for storage that actually exists. Eliminating temporary result storage
can make a return succeed under a zero temporary budget without changing any
quota ceiling. A triggered collector still needs its own admitted scratch.
The [execution allocation qualification](../../../docs/vm-arch/execution-allocation-qualification.md)
records instruction regions, before/after counts and limits. The
[memory contract](../../../docs/specs/runtime/03-memory-stack-heap-and-allocation.md)
remains normative. These measurements say nothing about gameplay throughput.