[RFC]: disable floating-point contraction when building native add-ons for local tests

未关闭
#16,596 4 条评论 0 个 reaction 已指派 0 人 在 GitHub 查看

维护者通常 1 天内回复

还没有人认领这个 Issue。

评估

难度
5/5
预计耗时
一周以上
新手友好度
35/100
Issue 类型
功能
描述清晰度
基本清楚
活跃度
活跃
技术栈
cmake, javascript, node.js

调研方向

Start with tools/make/lib/install/addons.mk, tools/make/lib/test/README.md, and the contributing FAQ entry on compiling native add-ons; compare the add-on detection in .github/workflows/scripts/run_affected_tests/run. The RFC leaves several policy questions open, including default behavior, CI flags, stale builds, and C examples. Done means an agreed approach updates test builds and documents the behavior while preserving release flags.

由索引模型根据 Issue 内容生成。

描述

Feature Needs Discussion RFC
Description

This RFC proposes updating make test so that it rebuilds the native add-ons of the packages under test with -ffp-contract=off, so local native test results match CI.

The approach was agreed on in Office Hours (2026-09-29). The open item from that meeting was how to handle packages which don't have a native add-on.

Problem

On Apple Silicon, clang fuses a*b + c into a single FMA instruction by default. The result can differ by 1 ulp from the JavaScript implementation, which rounds after each operation. Native tests using t.strictEqual then fail locally even though they pass in CI.

Reproduction on current develop (macOS arm64, Apple clang 21.0.0, node-gyp 12.4.0):

$ make install-node-addons NODE_ADDONS_PATTERN="stats/base/ndarray/dvariancech"
$ node lib/node_modules/@stdlib/stats/base/ndarray/dvariancech/test/test.native.js
not ok 11 returns expected value
    expected: 6.25
    actual:   6.249999999999999
# tests 12
# pass  11
# fail  1

The same package with contraction disabled:

$ CFLAGS="-ffp-contract=off" make install-node-addons NODE_ADDONS_PATTERN="stats/base/ndarray/dvariancech"
$ node lib/node_modules/@stdlib/stats/base/ndarray/dvariancech/test/test.native.js
# tests 12
# pass  12
Findings
  • No binding.gyp changes are needed (there are 1,719 of them). gyp's Makefile generator reads CFLAGS from the environment on both Linux and macOS (CFLAGS.target ?= $(CPPFLAGS) $(CFLAGS), appended in cmd_cc in gyp/pylib/gyp/generator/make.py). install-node-addons in tools/make/lib/install/addons.mk already passes the environment through to node-gyp rebuild, which is why the command above works today.
  • CI already solves "packages without add-ons". .github/workflows/scripts/run_affected_tests/run only builds an add-on when lib/node_modules/${pkg}/binding.gyp exists. I suggest make test reuse that check rather than swallow a non-zero exit code from the rebuild step. Swallowing the exit code would also hide real compile errors.
  • Release builds should keep their current flags. FMA is faster and is usually at least as accurate. Only the test build needs to match the JavaScript reference.
Proposed changes
  1. Add a variable to tools/make/lib/install/addons.mk (e.g., NODE_ADDONS_CFLAGS) which, when set, is passed as CFLAGS and CXXFLAGS to node-gyp rebuild.
  2. Before make test runs tests, resolve the packages matched by TESTS_FILTER. For each one that has a binding.gyp, run install-node-addons with NODE_ADDONS_CFLAGS="-ffp-contract=off".
  3. Document the behavior in tools/make/lib/test/README.md and in the contributing FAQ entry on compiling native add-ons.
Related Issues
  • stdlib-js/meetings #162
Questions
  1. Should the rebuild be on by default for make test, or opt-in (e.g., TEST_NATIVE=1)? Rebuilding adds compile time to every run, so a filter-less make test may need to skip it.
  2. Should CI also pass -ffp-contract=off explicitly? CI passes today because it builds on ubuntu-latest (x86-64). As far as I can tell, that's only because baseline x86-64 has no FMA instructions to fuse into, not because contraction is turned off. If CI ever moves to arm64 runners, the same 1-ulp failures would show up there.
  3. Should the flag be recorded (e.g., in a stamp file next to the built add-on) so that a test run doesn't use a stale add-on built with different flags?
  4. Do C examples and C benchmarks (make examples-c, make benchmark-c) need the same treatment, or is this only about tests?
Other

The 1-ulp cases have been showing up in the ULP-assertion migration PRs. Several native tests have moved to isAlmostSameValue( actual, expected, 1 ) to tolerate this difference. With contraction disabled in test builds, strictEqual would work again for kernels where the JavaScript and C implementations are meant to match exactly.

Checklist
  • I have read and understood the Code of Conduct.
  • Searched for existing issues and pull requests.
  • The issue name begins with RFC:.
主要语言
JavaScript
星标
6k
派生
1.3k
平均合并
1 天 21 小时
30 天内合并 PR
224

环境准备

在 Codespaces 中打开

在浏览器里用你自己的 GitHub 账号启动这个项目的开发容器。

从这里开始

  1. 先读完整个 Issue,再读项目的贡献指南。
  2. 在 Issue 下留言说明你要接手 —— 这能避免两个人做同样的事。
  3. Fork 仓库,在一个分支上完成修改。
  4. 提交 Pull Request,并在描述里引用这个 Issue 编号。

stdlib-js/stdlib 的其他 Issue

查看 stdlib-js/stdlib 的全部 Issue

相似的 Issue

更多 JavaScript Issue

把新 issue 发到你的邮箱

精选适合新手参与的 GitHub issue 摘要。