tsgo backend: parse in-process, open the project lazily

[johanrd]

Two costs in the TypeScript 7 backend that hit every run, cached or not, on an 838-file app (profile: 11 s of a 22.8 s fully-cached run inside the tsgo client):

• The resolver's syntactic parses of imported component files went through typescript-go as two snapshot round-trips each (~15 ms), outside the extraction cache: 5.6 s. A syntactic parse needs no checker, so the backend now hands the resolver the project's typescript 5/6 library when one is installed (TypeBackend.parserSyntax, in-process createSourceFile), and its own facade otherwise. The resolver's node kinds must match its parser, which is why it is a separate facade from the checker-side one.
• preload opened the whole project before knowing whether any file missed the cache: 1.9 s. It now opens it only when one does.

Same app, @glimmer/syntax patched with emberjs/ember.js#21314 on both sides:

run	before	after
cold	~42 s	16.7 s
warm, all cached	10.8 s	8.2 s
single cached file	2.8 s	0.3 s
--no-glint, whole set	407 s	12.0 s

The --no-glint number is the same parse path (build-maps.ts → syntaxFor) without the extraction cache in front of it.

Both lanes: 291 passed + 1 expected fail.

Cowritten by Claude

[Copilot]

🔵 Needs a closer look

It reworks performance-critical backend AST-parsing infrastructure whose correctness depends on cross-process node-kind/position invariants and the project's runtime-resolved typescript, which warrants final human review despite no defects being found.

Pull request overview

This PR optimizes the TypeScript 7 (tsgo) backend to cut two recurring costs measured on a large (838-file) Ember app. First, the resolver's syntactic parses of imported component files previously went through typescript-go as two IPC snapshot round-trips each; the backend now exposes a new parserSyntax facade that, under tsgo, uses the project's in-process typescript 5/6 library (via createSourceFile) when installed, falling back to the tsgo facade otherwise. Second, preload no longer eagerly opens the whole project — it opens the snapshot lazily only when at least one file misses the cache, matching the existing ts6.ts behavior. The reported effect is a large reduction in cold, warm, single-file, and --no-glint run times.

Changes:

• Add a parserSyntax facade to the TypeBackend interface and wire both backends to it (tsgo resolves the project's typescript library in-process; ts6 reuses its existing syntax facade).
• Switch the resolver-side calls in extractAttrTypeMap and syntaxFor to use parserSyntax, keeping checker-side calls on syntax.
• Make tsgo preload open the project snapshot lazily (if (loaded > 0)), aligning progress reporting with ts6.ts.

File summaries

File	Description
lib/backend/types.ts	Adds the required parserSyntax: TsSyntax field with an explanatory doc comment.
lib/backend/tsgo.ts	Adds libraryTypeScript() to load the project's in-process typescript; sets parserSyntax; makes preload open the snapshot lazily.
lib/backend/ts6.ts	Reuses the single ts6Syntax(ts) facade for both syntax and parserSyntax.
lib/glint.ts	Routes resolver syntactic parses (findTemplateSource, resolveTemplate, etc.) through parserSyntax; checker calls unchanged.
lib/backend/index.ts	syntaxFor now returns the backend's parserSyntax instead of syntax.

Review details

• Files reviewed: 5/5 changed files
• Comments generated: 0
• Review effort level: Balanced

---

💡 Add a code-review agent skill or configure MCP servers for context-aware, tailored reviews. Learn more in the docs.

[github-actions]

🏎️ Benchmark Comparison

	Benchmark	Control (p50)	Experiment (p50)	Δ
⚪	extract small template	834.46 µs	831.63 µs	-0.3%
⚪	extract medium template	3.62 ms	3.78 ms	+4.4%
🟡	extract cross-file resolution	4.27 ms	4.53 ms	+6.0%
🟡	cold run (cache off)	1750.21 ms	1898.92 ms	+8.5%
🟢	warm run (all cached)	1363.45 ms	768.96 ms	-43.6%
⚪	one cached file	494.49 ms	499.71 ms	+1.1%
🟢	no glint	4417.82 ms	1021.91 ms	-76.9%

🟢 faster · 🔴 slower · 🟡 5–10 %, possibly noise · ⚪ within 5 %

Full output

clk: ~3.03 GHz
cpu: AMD EPYC 7763 64-Core Processor
runtime: node 24.19.0 (x64-linux)

benchmark                    avg (min … max) p75 / p99    (min … top 1%)
-------------------------------------------- -------------------------------
extract small template          1.22 ms/iter   1.16 ms  █                   
                       (658.82 µs … 6.87 ms)   4.69 ms ▇█▂                  
                     ( 14.38 kb …   4.25 mb) 312.18 kb ███▅▃▃▂▂▂▁▁▂▂▁▁▁▂▂▁▁▁

extract medium template         4.28 ms/iter   4.30 ms  █                   
                        (3.07 ms … 12.79 ms)  10.06 ms  █▄                  
                     (275.89 kb …   4.33 mb)   1.01 mb ▅██▅▄▂▃▂▂▂▂▂▁▂▁▁▂▁▁▁▁

extract cross-file resolution   4.74 ms/iter   5.03 ms   █                  
                         (3.63 ms … 9.22 ms)   8.15 ms  ▆█▄                 
                     ( 59.28 kb …   2.59 mb)   1.29 mb ▃█████▄▄▂▅▃▃▃▄▁▂▂▂▂▁▂

whole process (min / p50 of 3 runs, 20 files)
  cold run (cache off)           1741 ms /   1750 ms
  warm run (all cached)          1360 ms /   1363 ms
  one cached file                 493 ms /    494 ms
  no glint                       4408 ms /   4418 ms
clk: ~3.10 GHz
cpu: AMD EPYC 7763 64-Core Processor
runtime: node 24.19.0 (x64-linux)

benchmark                    avg (min … max) p75 / p99    (min … top 1%)
-------------------------------------------- -------------------------------
extract small template          1.16 ms/iter   1.15 ms  █                   
                       (649.89 µs … 5.82 ms)   4.21 ms ▃█                   
                     ( 25.17 kb …   3.12 mb) 295.81 kb ██▇▆▃▂▂▂▂▂▁▁▁▂▂▂▁▂▁▁▁

extract medium template         4.57 ms/iter   4.97 ms  █                   
                        (3.16 ms … 12.13 ms)  10.62 ms  █▇                  
                     (315.36 kb …   4.79 mb)   1.04 mb ███▄▅▄▃▄▃▂▄▁▁▁▁▂▁▁▁▁▁

extract cross-file resolution   5.05 ms/iter   5.40 ms  ▅█                  
                        (3.71 ms … 11.86 ms)   9.30 ms ▂██▅  ▂              
                     ( 59.03 kb …   2.74 mb)   1.29 mb █████▇█▆▄▇▂▄▄▁▂▁▂▂▂▃▂

whole process (min / p50 of 3 runs, 20 files)
  cold run (cache off)           1877 ms /   1899 ms
  warm run (all cached)           769 ms /    769 ms
  one cached file                 497 ms /    500 ms
  no glint                       1021 ms /   1022 ms