[cknitt]

Part of #9256 (many refs updated in one loop).

The cause was scalar replacement (Lam_pass_sroa): it analyzed and rewrote a block's whole scope once per block, so n refs used in one function cost O(n²). Now one traversal decides every block and one rewrite replaces them all.

The old order replaced blocks from the innermost out, so when an inner block had a never-read field, the pure values written to or initializing that field were already gone when the blocks around it were analyzed. To give the same result, an occurrence inside such a value is held in a pending slot until the inner block is decided, and dropped if the field turns out never to be read. The rewrite then drops exactly those values without rewriting them.

Measured with bsc on linux-arm64, allocated words and wall time, n refs each updated once in one for loop:

Refs	Before	After
250	12.9M, 48 ms	4.6M, 34 ms
500	42.2M, 130 ms	8.8M, 64 ms
1000	151M, 417 ms	17.4M, 121 ms

The generated JavaScript is unchanged for the test suites and about 1,600 generated functions with refs and mutable records stored in, read from and written to each other. A first version without the pending slots differed on 3 of them.

New tests in sroa_test.res: a ref stored only in a write-only ref, a ref read only by a write-only ref (both still become scalars), and several refs updated in one loop.

🤖 Generated with Claude Code

https://claude.ai/code/session_01Kbf3gK1dcjbe97yY1wRMgs

[github-actions]

Changelog

💅 Polish

• Compile functions with many local refs or mutable records in linear time.

Semantic Diff

2 commits · Full page

Replace local mutable blocks in one pass over the function

_{52c1aa6 · cknitt + Claude · Nov 07, 2026, 10:04 UTC}

Why

The pass analyzed and rewrote a block's whole scope once per block, so n refs used in one function cost O(n²): 1,000 refs updated in one loop allocated 151M words. One traversal now decides every block and one rewrite replaces them, so the same input allocates 17M.

A value written to, or initializing, a never-read field is held back until its block is decided, as the old innermost-first order dropped such values before analyzing the blocks around them. The generated JavaScript is the same for the test suites and 1,600 generated functions.

✨ Analyze blocks without allocating a closure per node

_{ba4c395 · cknitt + Claude · Nov 07, 2026, 19:02 UTC · new}

The benchmarks showed the single pass allocating up to 0.9% more for functions without local blocks: each node passed a new closure to the traversal. The innermost slot is now a reference, so the traversal reuses one closure, and other inputs allocate as much as before.

[codecov]

Codecov Report

❌ Patch coverage is 92.47312% with 7 lines in your changes missing coverage. Please review.
✅ Project coverage is 81.74%. Comparing base (bf35764) to head (ba4c395).

Files with missing lines	Patch %	Lines
compiler/core/lam_pass_sroa.ml	92.47%	7 Missing ⚠️

Additional details and impacted files

@@           Coverage Diff           @@
##           master    #9258   +/-   ##
=======================================
  Coverage   81.74%   81.74%           
=======================================
  Files         457      457           
  Lines       65713    65751   +38     
=======================================
+ Hits        53715    53748   +33     
- Misses      11998    12003    +5     

Files with missing lines	Coverage Δ
compiler/core/lam_pass_sroa.ml	86.08% <92.47%> (+0.37%)	⬆️

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.
• 📦 JS Bundle Analysis: Save yourself from yourself by tracking and limiting bundle sizes in JS merges.

[pkg-pr-new]

Open in StackBlitz

rescript

npm i https://pkg.pr.new/rescript@9258

@rescript/belt

npm i https://pkg.pr.new/@rescript/belt@9258

@rescript/darwin-arm64

npm i https://pkg.pr.new/@rescript/darwin-arm64@9258

@rescript/linux-arm64

npm i https://pkg.pr.new/@rescript/linux-arm64@9258

@rescript/linux-x64

npm i https://pkg.pr.new/@rescript/linux-x64@9258

@rescript/runtime

npm i https://pkg.pr.new/@rescript/runtime@9258

@rescript/win32-x64

npm i https://pkg.pr.new/@rescript/win32-x64@9258

commit: ba4c395

[github-actions]

Developer playground preview: https://rescript-lang.github.io/rescript/dev-playground/?version=pr-9258

[github-actions]

🎉 Compiler benchmarks

Allocated words

In total: 456,989,425 → 444,961,995 (-2.63%).

🎉 Improvement: 1 benchmark allocates more than 1% less.

Benchmark	Base	Head	Change
stress/RefsInLoop.res	17,385,182	5,364,070	-69.15%
packages/@rescript/runtime/Primitive_hash.res	1,501,651	1,497,275	-0.29%
packages/@rescript/belt/src/Belt_Array.res	2,842,879	2,839,869	-0.11%

Instructions

In total: 10,226,961,997 → 10,032,277,336 (-1.90%).

🎉 Improvement: 1 benchmark executes more than 1% fewer instructions.

Benchmark	Base	Head	Change
stress/RefsInLoop.res	411,396,517	214,259,235	-47.92%
packages/@rescript/belt/src/Belt_internalAVLtree.res	166,149,467	166,562,149	+0.25%
packages/@rescript/belt/src/Belt_List.res	145,581,009	145,924,298	+0.24%

Pull request ba4c395 compared with base bf35764 in this run.