[Alexis Delrieu (Amatewasu)]

Fixes #64423

Problem

TypeScript 7 runs out of memory (unbounded growth, zero diagnostics) on programs that TypeScript 6 checks fine. Root cause: the combined-constraint comparison in structuredTypeRelatedTo hoists the constraints of the instantiable constituents of an intersection source into a combined constraint and compares it against the target. While measuring variances, the hoisted constraint may itself be an intersection with instantiable constituents (e.g. deferred conditional types produced by heavily overloaded generic builder types), so a failed comparison re-enters the same logic with a structurally new and ever larger source that neither the relation cache nor the deeply-nested-type guards recognize as recursive — the chase grows without bound until the process dies.

Fix

Skip the combined-constraint check while the variance stack is non-empty; it never contributes to the measured relation between the marker types.

Verification (both repros from the issue)

• JSX repro (@react-three/fiber IntrinsicElements augmentation + <SvgIcon />, exact versions from the issue): before — killed at a 3 GiB / 6 GiB address-space cap with zero output; after — ~4.4 s, ~700 MB RSS, exit 0.
• three/tsl repro (uv().div(vec2(1, 1)), three@0.185.1 + @types/three@0.185.1): before — killed at the cap with zero output; after — ~3.1 s, ~280 MB RSS, exit 0.
• Full go test ./... passes (only pre-existing environmental failure in internal/astnav, which requires a locally built JS typescript package in root node_modules; fails identically on unpatched main).

Note on the approach

An earlier variant of this change additionally capped the chase depth (a counter on Relater limiting nested combined-constraint comparisons to 3). That cap regressed tests/cases/compiler/indexedAccessAndNullableNarrowing.ts (new false TS2345 errors on the syncStoreProp pattern from #57693, which both TS 6 and unpatched TS 7 accept), and the depth needed by legitimate code is not bounded by any small constant, so the cap was dropped. The variance-stack guard alone fixes both repros.

Disclosure

This change was generated by an LLM (as was the investigation in the issue thread).

Checklist notes

• Could not run npx hereby lint / check:format (no root node_modules in this checkout); ran gofmt + go vet (clean) and the full Go test suite instead.
• No standalone regression test added: both known triggers require the external dependency graphs (three/tsl, R3F/MUI) whose overloaded builder types produce the degenerate hoisted constraints; a self-contained test case could not be extracted.
• There is an associated issue (TypeScript 7.0.2 silently runs out of memory while typechecking a project that TypeScript 6.0.3 checks successfully #64423)
• Code is up-to-date with the main branch
• go test ./... passes
• npx hereby lint
• npx hereby check:format
• New or updated tests validating the change

[Copilot]

Copilot review overview

🟡 Changes recommended

The compiler fix lacks a checked-in regression test for the unbounded-chase scenario.

Review effort: Balanced
Findings: 1

Open (1)

• Add regression test for variance-check termination · New

What changed in this PR

Prevents unbounded type-relation expansion during variance measurement.

Changes:

• Skips combined-constraint checks while the variance stack is active.
• Documents the recursive growth scenario.

File	Description
tsc/​internal/​checker/​relater.go	Guards combined-constraint evaluation during variance measurement.

[Alexis Delrieu (Amatewasu)]

LLM-generated test added

---

Regression test checked in as bd456af6e3 (tests/cases/compiler/effectiveConstraintChaseDuringVarianceMeasurement.ts plus .types/.symbols baselines), in response to the review comment asking for a checked-in test for the variance-stack guard.

1. Minimized from the real repro. It is hand-distilled from the three.js TSL node types that triggered TypeScript 7.0.2 silently runs out of memory while typechecking a project that TypeScript 6.0.3 checks successfully #64423. Only the constructs needed to drive the combined-constraint comparison in structuredTypeRelatedTo during variance measurement were kept — resolving uv().div(vec2()) measures the variances of the generic extension interfaces, whose hoisted constraints form deferred conditional intersections.

2. It fails without the fix. On the pre-fix revision (673a5f17d7) this file enters the unbounded relation chase and the process dies with runtime: out of memory. Removing or weakening the variance-stack guard reintroduces this immediately, so CI catches it instead of the regression going unnoticed.

3. It passes with the fix and pins the exact output. With the guard, the file checks to completion in under a second with no diagnostics, and the baselines record the precise inferred types (e.g. x : Node<"vec2">). This also guards the opposite failure mode: a future change that "fixes" a hang by making the checker give up and infer different types would fail the baseline comparison.

Reproducing the pre-fix failure locally:

git show 673a5f17d7:tsc/internal/checker/relater.go > tsc/internal/checker/relater.go
cd tsc && GOMAXPROCS=2 timeout 120 go test ./internal/testrunner \
  -run 'TestLocal/effectiveConstraintChaseDuringVarianceMeasurement'
# pre-fix: never terminates (chase grows until OOM) / fixed: "ok" in ~1s

[Alexis Delrieu (Amatewasu)]

What do you think Jake Bailey (@jakebailey)? :)

[Alexis Delrieu (Amatewasu)]

Verified against the private project from the issue

I ran the full typecheck of the actual project behind #64423 (a private React 19 + R3F + TSX app, ~855 files) with binaries built from this PR branch vs. unpatched main, using the same methodology as the issue (tsc --noEmit --pretty false --incremental false --composite false under a systemd-run --user -p MemoryMax=… cgroup cap):

Binary	Memory cap	Result
TS 7 unpatched main	6 GiB	OOM-killed after ~3 min, zero output
TS 7 with this PR	6 GiB	25.0 s, ~4.9 GiB peak RSS, exit 0, 0 errors
TS 6.0.3 (8 GB node heap)	12 GiB	146.9 s, ~3.2 GiB peak, exit 0, 0 errors

Diagnostics parity holds: the patched TS 7 reports exactly the same result as TS 6 (zero errors) on the full project, so the guard is not observable in output — it only stops the runaway.

Caveats, for completeness:

• Peak RSS with the fix (~4.9 GiB) is still higher than TS 6 (~3.2 GiB) on this project; the check is now bounded and completes, but it needs memory headroom.
• TS 6 with the default 4 GB node heap aborts on this project (exit 134) — consistent with what was reported in the issue thread.

[Copilot]

🔵 Needs a closer look

The core type-relation change needs maintainer confirmation of semantic safety beyond the termination regression test.

1 open issue

• Add regression test for variance-check termination

_{Review effort: Balanced}