[nigrosimone]

First part of #1300, only the NoexceptPersistent change.

v8::Global has a move constructor that is not noexcept, so every lambda that captures a UniquePersistent fails the is_small_object_v test in MoveOnlyFunction and is heap allocated, even when the capture is only 16 bytes (the handle plus the isolate pointer). That is one new and one delete per onData, onDataV2 and onWritable registration.

The fix is a small wrapper that restores the noexcept on the move (the underlying move is a pointer swap and cannot throw), used in those three sites. onAborted and collectBody capture more than 16 bytes and are left as they are.

Checked with a static_assert on ofats::any_detail::is_small_object_v for a lambda capturing UniquePersistent (false) and NoexceptPersistent (true), compiled with MSVC against the Node 26 headers.

[nigrosimone]

Measured with the benchmark workflow proposed in #1326: this branch against master, both built on the same hosted runner, the load alternating between them (nigrosimone#2). Every row stays inside the noise band of the run, so at this resolution the change moves nothing in either direction, which is expected: one allocation per registration is well under a percent of a request.

Scenario	Base req/s	Head req/s	Head / base	Rounds	Noise	Busy	CPU per request
http/hello-world	115910	114763	0.991x (-0.9%)	0.98 to 0.99	±2.7%	100% / 100%	8.62 / 8.71 us (1.010x)
http/headers	96551	94921	0.986x (-1.4%)	0.98 to 0.99	±3.5%	100% / 100%	10.36 / 10.54 us (1.014x)
http/json-post-1kb	53351	52818	0.985x (-1.5%)	0.97 to 1.00	±5.3%	100% / 100%	18.76 / 18.94 us (1.016x)
http/cached	129466	128312	0.991x (-0.9%)	0.99 to 0.99	±2.8%	100% / 100%	7.72 / 7.79 us (1.009x)
ws/echo-20b	119060	121176	1.018x (+1.8%)	1.01 to 1.03	±1.1%	100% / 100%	8.39 / 8.25 us (0.982x)
ws/echo-4kb	103676	104278	1.003x (+0.3%)	0.99 to 1.01	±1.8%	100% / 100%	9.64 / 9.58 us (0.998x)

"Noise" is how far base against base and head against head, the same code on both sides, got from 1.0 in the same run; a row is marked only when it moved more than that. Node v26.9.0, AMD EPYC 7763, 4 cores, wrk -t2 -c100.

[uNetworkingAB]

Btw, this seems like a bug in V8.

[nigrosimone]

Yes, I agree. In the same V8 headers TracedReference(TracedReference&&) is noexcept but Global(Global&&) is not, and it can not throw because V8 is built without exceptions.

[nigrosimone]

Close here. PR on v8 https://chromium-review.googlesource.com/c/v8/v8/+/8467989

[uNetworkingAB]

I am following your third party fixes, they are well received but I am fully exhausted with the Node.js core people so I try and stay out of their shit