Improve resolveReferences performance by JialinHuang803 · Pull Request #4004 · Azure/autorest.typescript

JialinHuang803 · 2026-05-26T15:30:41Z

Summary

Two complementary changes to Binder that reduce the cost of the resolve references emit phase:

Batch placeholder replacement in resolveAllReferences so each source file gets one bulk text.replace + one replaceWithText, instead of one cycle per placeholder. Each replaceWithText discards ts-morph's cached AST and re-parses the file, so the previous per-placeholder loop grew as O(placeholders per file × file size).
Bulk-add import declarations with a single addImportDeclarations call per file, instead of looping addImportDeclaration once per import. Each individual call triggers a full file re-parse in ts-morph.

What changed

For each source file the binder now:

Scans the text once via collectPlaceholders to collect placeholders actually present in the file.
Iterates the precomputed declarationByPlaceholder / dependencyByPlaceholder maps (preserving the original insertion order, so name-collision aliasing in addImport is unchanged).
Accumulates a single placeholder → local-name map.
Applies all replacements via applyReplacements in one bulk text.replace + one replaceWithText.
Emits all collected imports for the file in a single addImportDeclarations call (sort order unchanged).

Only packages/typespec-ts/src/framework/hooks/binder.ts is touched.

Performance results

Measured by running npx tsp compile client.tsp --emit=@azure-tools/typespec-ts against the real Azure spec files (specification/network/.../Network/client.tsp and specification/compute/.../Compute/client.tsp), with per-phase console.time instrumentation around $onEmit. Single run per side.

Spec	`resolve references` (before)	`resolve references` (after)	Total `onEmit` (before)	Total `onEmit` (after)
Network	22:08.539	4.397s	31:25.022	7:27.367
Compute	3:20.601	2.082s	5:17.883	1:47.209

Network: ~302× speedup on the binder phase, ~24 min saved per SDK generation (31:25 → 7:27 end-to-end).
Compute: ~96× speedup on the binder phase, ~3.5 min saved per SDK generation (5:18 → 1:47 end-to-end).

Incremental gain from the follow-up addImportDeclarations commit, on top of the placeholder-batching commit:

Spec	`resolve references` before	after	Δ
Network	8.664s	4.397s	−49% (−4.27s)
Compute	2.942s	2.082s	−29% (−0.86s)

See the baseline timing breakdown in #3907 — resolve references accounted for 70.5% of total emit time on Network and 63.1% on Compute before this change.

Batch placeholder replacement in resolveAllReferences so each source file gets one text.replace + one replaceWithText call, instead of one cycle per placeholder. Each replaceWithText call discards ts-morph's cached AST and re-parses the file, so the previous per-placeholder loop made the cost grow as O(placeholders per file * file size). For each file we now: - scan the text once to collect the placeholders actually present; - iterate the precomputed declaration/dependency maps (preserving the original insertion order so name-collision aliasing in addImport is unchanged); - accumulate a single placeholder -> local-name map; - apply all replacements in a single bulk text.replace and a single replaceWithText. Output is byte-identical on batch_modular and openai_modular smoke specs, and the binder integration tests (22) all pass unchanged. Real-SDK measurements (Microsoft.Network and Microsoft.Compute management-plane via 'tsp compile client.tsp --emit=@azure-tools/typespec-ts'): | Spec | resolve references (before) | resolve references (after) | Total onEmit (before) | Total onEmit (after) | |---------|----------------------------:|---------------------------:|----------------------:|---------------------:| | Network | 22:08.539 | 8.664s | 31:25.022 | 8:44.110 | | Compute | 3:20.601 | 3.455s | 5:17.883 | 1:49.726 | That's a 153x speedup on the binder phase for Network (22 minutes saved per SDK generation) and 58x for Compute (3.5 minutes saved). Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>

JialinHuang803 · 2026-05-27T09:28:09Z

There is an existing test file for binder https://github.com/Azure/autorest.typescript/blob/main/packages/typespec-ts/test-next/integration/hooks/binder.test.ts and all cases pass.

Copilot

Pull request overview

This PR optimizes the TypeSpec TS emitter’s binder reference resolution by batching placeholder replacements per source file, avoiding repeated ts-morph replaceWithText calls that force costly re-parses and scale poorly with placeholder count.

Changes:

Precomputes placeholder→declaration/dependency lookup maps once per resolveAllReferences call.
Scans each source file once for present placeholders, builds a per-file replacement map, and applies all substitutions in a single bulk pass.
Replaces per-placeholder replacement logic with collectPlaceholders + applyReplacements utilities.

+  const text = sourceFile.getFullText();
+  const placeholderRegex = new RegExp(
+    `${escapeRegExp(PLACEHOLDER_PREFIX)}.+?__`,
+    "g"
+  );


qiaozha · 2026-05-27T10:46:43Z

+    // Pre-compute placeholder -> declaration/dependency maps
+    const declarationByPlaceholder = new Map<
+      string,
+      [unknown, DeclarationInfo | StaticHelperMetadata]


why the first element of the tuple is unknown? is it the return type of this.serializePlaceholder?

oh wait it should be the input of this.serializePlaceholder.

Yes, the unknown here is consistent with the key type of this.declarations: private declarations = new Map<unknown, DeclarationInfo>();.

maorleger

I'm far from an expert here but the perf numbers are definitely exciting! I linked to the ts-morph performance documentation in a comment and I wonder if there are other places where we can batch operations for quick wins, not blocking of course

maorleger · 2026-05-27T15:07:26Z

not blocking for this PR but I wonder if we can get some more wins by batching operations. This pattern is explicitly called out in https://ts-morph.com/manipulation/performance#performance-tip-batch-operations and may be an easy replacement to make

Have you seen the performance ts-morph docs ? Could be useful to point an agent to them and find low-hanging fruit to increase perf as well

Great call, thanks for the pointer, Maor! I replaced the loop with a single addImportDeclarations bulk call in the latest commit.

Replace the per-import addImportDeclaration loop with a single addImportDeclarations bulk call. Each individual addImportDeclaration triggers a full re-parse of the source file in ts-morph, so emitting N imports per file was O(N) re-parses; bulk-add collapses this to one.

JialinHuang803 requested review from MaryGao, joheredi and qiaozha as code owners May 26, 2026 15:30

JialinHuang803 linked an issue May 26, 2026 that may be closed by this pull request

[JavaScript][Perf] Optimize speed to generate sdk #3907

Open

qiaozha self-assigned this May 27, 2026

qiaozha added HRLC p0 priority 0 labels May 27, 2026

JialinHuang803 changed the title ~~Improve JS emitter performance~~ Improve resolveReferences performance for JS emitter May 27, 2026

JialinHuang803 removed a link to an issue May 27, 2026

[JavaScript][Perf] Optimize speed to generate sdk #3907

Open

JialinHuang803 requested a review from Copilot May 27, 2026 09:31

Copilot started reviewing on behalf of JialinHuang803 May 27, 2026 09:31 View session

JialinHuang803 requested a review from maorleger May 27, 2026 09:32

Copilot AI reviewed May 27, 2026

View reviewed changes

Comment thread packages/typespec-ts/src/framework/hooks/binder.ts

Comment on lines +490 to +494

const text = sourceFile.getFullText();

const placeholderRegex = new RegExp(

`${escapeRegExp(PLACEHOLDER_PREFIX)}.+?__`,

"g"

);

qiaozha reviewed May 27, 2026

View reviewed changes

qiaozha approved these changes May 27, 2026

View reviewed changes

maorleger approved these changes May 27, 2026

View reviewed changes

JialinHuang803 force-pushed the perf-binder-batched-replacement branch from e2ec4ae to 923390f Compare May 28, 2026 06:49

Merge branch 'main' into perf-binder-batched-replacement

494f4b3

JialinHuang803 changed the title ~~Improve resolveReferences performance for JS emitter~~ Improve resolveReferences performance May 28, 2026

JialinHuang803 merged commit 119dc3a into Azure:main May 28, 2026
16 checks passed

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Improve resolveReferences performance#4004

Improve resolveReferences performance#4004
JialinHuang803 merged 3 commits into
Azure:mainfrom
JialinHuang803:perf-binder-batched-replacement

JialinHuang803 commented May 26, 2026 •

edited

Loading

Uh oh!

JialinHuang803 commented May 27, 2026

Uh oh!

Copilot AI left a comment

Uh oh!

qiaozha May 27, 2026

Uh oh!

qiaozha May 27, 2026

Uh oh!

JialinHuang803 May 28, 2026

Uh oh!

maorleger left a comment

Uh oh!

maorleger May 27, 2026

Uh oh!

JialinHuang803 May 28, 2026

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

4 participants

Conversation

JialinHuang803 commented May 26, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Summary

What changed

Performance results

Uh oh!

JialinHuang803 commented May 27, 2026

Uh oh!

Copilot AI left a comment

Choose a reason for hiding this comment

Pull request overview

Uh oh!

qiaozha May 27, 2026

Choose a reason for hiding this comment

Uh oh!

qiaozha May 27, 2026

Choose a reason for hiding this comment

Uh oh!

JialinHuang803 May 28, 2026

Choose a reason for hiding this comment

Uh oh!

maorleger left a comment

Choose a reason for hiding this comment

Uh oh!

maorleger May 27, 2026

Choose a reason for hiding this comment

Uh oh!

JialinHuang803 May 28, 2026

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

4 participants

JialinHuang803 commented May 26, 2026 •

edited

Loading