ffi: accept safe integer numbers for 64-bit arguments

[HoonDongKang]

Summary

Accept safe integer numbers for int64/uint64 arguments alongside bigint values, allowing buffer lengths to be passed without explicit BigInt() conversions.

-functions.sum_buffer(bytes, BigInt(bytes.byteLength));
+functions.sum_buffer(bytes, bytes.byteLength);

Changes

• Accept safe integer numbers for int64 and non-negative safe integer numbers for uint64.
• Validate and convert arguments across Fast API, SharedBuffer, and generic C++ conversion paths.
• Preserve existing bigint range checks and bigint return values.
• Document the accepted number ranges.

Tests

The tests check that number and bigint arguments produce the same results, including when used together. A new single-argument test covers Fast API conversion before and after requesting V8 optimization.

Separate nine-argument tests exercise SharedBuffer conversion and generic fallback. Both check Number and BigInt boundaries, invalid inputs, and error messages.

// SharedBuffer: validates and converts the integer in JavaScript.
functions.passthrough_i64_9(0n, value, ...zeros);

// Generic fallback: validates and converts the integer in C++.
functions.passthrough_i64_9(buffer, value, ...zeros);

Validation

54 tests passed across the following files. Local lint and whitespace checks also passed.

./node --test \
  test/ffi/test-ffi-calls.js \
  test/ffi/test-ffi-fast-integer-validation.js \
  test/ffi/test-ffi-shared-buffer.js


ℹ tests 54
ℹ suites 0
ℹ pass 54
ℹ fail 0
ℹ cancelled 0
ℹ skipped 0
ℹ todo 0
ℹ duration_ms 934.82625

Refs: #66198
Assisted-by: Codex:Astra-medium

[nodejs-github-bot]

Review requested:

• @nodejs/ffi

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 90.37%. Comparing base (75e4bbe) to head (b0a3f6d).
⚠️ Report is 3 commits behind head on main.

Additional details and impacted files

@@           Coverage Diff           @@
##             main   #66197   +/-   ##
=======================================
  Coverage   90.36%   90.37%           
=======================================
  Files         792      792           
  Lines      275398   275461   +63     
  Branches    52776    52802   +26     
=======================================
+ Hits       248877   248955   +78     
- Misses      16937    16938    +1     
+ Partials     9584     9568   -16     

Files with missing lines	Coverage Δ
lib/internal/ffi-shared-buffer.js	60.11% <100.00%> (+4.36%)	⬆️
lib/internal/ffi/fast-api.js	94.67% <100.00%> (+0.17%)	⬆️
src/ffi/types.cc	63.19% <100.00%> (+6.92%)	⬆️

... and 31 files with indirect coverage changes

🚀 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.

[ShogunPanda]

Do you mind running benchmarks?
While I like the idea, I want to make sure it's not too expensive.

[HoonDongKang]

I updated fast-api.js in b0a3f6d to combine integer validation and conversion into validateAndConvertFastIntegerArg().

In the previous revision of this PR, validation and conversion were separate steps, each looking up the argument’s type metadata. The combined helper performs that lookup once, validates and converts safe integer "number" inputs to "bigint", and preserves the existing type and range checks for "bigint" inputs.

I measured two aspects of performance, with 30 runs per configuration:

1. Existing BigInt inputs: baseline vs. updated implementation

                            confidence   improvement   accuracy (*)    (**)   (***)
ffi/add-i64.js n=10000000  ***            +29.99 %   ±4.82%  ±6.41%  ±8.36%
ffi/add-u64.js n=10000000  ***            +35.96 %   ±5.13%  ±6.83%  ±8.89%

                              -41.1%             0%             +41.1%
ffi/add-i64.js n=10000000                      |██████████████░░     +29.99% ***
ffi/add-u64.js n=10000000                      |█████████████████░░  +35.96% ***

2. BigInt vs. Number inputs in the updated implementation

I added an input: ['bigint', 'number'] configuration to both benchmarks and used scatter.js to measure the cost of passing numbers instead of pre-created bigints.

Benchmark	BigInt throughput (ops/s)	Number throughput (ops/s)	Difference
add-i64.js	29,288,918.5	19,523,825.1	−33.34%
add-u64.js	28,502,908.3	19,819,141.9	−30.47%

For the first comparison, both versions perform the same type and range checks for BigInt inputs, and the Number conversion branch is not taken. The improvement may therefore reflect differences in V8’s generated code rather than less validation work.

The second comparison shows the cost of accepting Number inputs: throughput was approximately 30–33% lower than with pre-created BigInt inputs in these benchmarks. This measures the additional cost of internal validation and conversion when callers use the convenience provided by this PR.

In absolute terms, this corresponds to approximately 15–17 ns of additional time per call in these benchmarks, which pass two 64-bit integer arguments. Do you think this overhead is acceptable given the convenience of passing Number values directly?

==
Update: I compared the same binaries with and without TurboFan inlining to investigate the unexpected BigInt improvement.

With the --no-turbo-inlining flag, which disables TurboFan inlining, I got the following results:

                            confidence   improvement   accuracy (*)    (**)   (***)
ffi/add-i64.js n=10000000                  +0.49 %   ±1.57%  ±2.08%  ±2.71%
ffi/add-u64.js n=10000000                  -0.10 %   ±1.59%  ±2.12%  ±2.76%

                              -2.1%              0%              +2.1%
ffi/add-i64.js n=10000000            ░░░░░░░░░░|▓▓▓▓░░░░░░░░░░░░░░░  +0.49% 
ffi/add-u64.js n=10000000      ░░░░░░░░░░░░░░░▓|░░░░░░░░░░░░░░       -0.10% 

With TurboFan inlining disabled, both performance differences were below 1%. This suggests that the previously observed BigInt improvement is related to TurboFan inlining and its effects on the generated code, rather than reduced validation work.