Fork-only changes: what is proposed, what is held, and why

[0pcom]

State of this fork's delta against upstream/dev, so nothing is forgotten.

Proposed upstream (open PRs)

• runtime: deliver signals under the threads scheduler when blocked on I/O tinygo-org/tinygo#5530 — signal delivery under the threads scheduler (review addressed)
• interp: make the timeout recoverable, per package initializer tinygo-org/tinygo#5587 — interp: recoverable timeout per package initializer
• reflect: implement MakeChan tinygo-org/tinygo#5588 — reflect.MakeChan
• os: add File.Chown on the unix path tinygo-org/tinygo#5589 — os.File.Chown
• runtime: add MemStats.NumGC for API compatibility tinygo-org/tinygo#5590 — runtime.MemStats.NumGC
• reflect: provide the linkname targets modern-go/reflect2 expects tinygo-org/tinygo#5591 — reflect2 linkname stubs (trade-off flagged for maintainers)
• runtime/debug: populate BuildInfo so ReadBuildInfo works tinygo-org/tinygo#5592 — runtime/debug ReadBuildInfo

Held, with reasons

src/net submodule bumps — the pointer commits can't be proposed to tinygo; the substance lives in 0magnet/net. See 0magnet/net#1 for what is blocked there on tinygo-org/net#59.

crypto/tls host-target override (loader/goroot.go + src/crypto/tls) — makes the native target use Go's software TLS instead of the netdev-offload wrapper. Held because it changes target behaviour rather than filling a gap, so it needs a maintainer conversation first, not a drive-by PR.

runtime/pprof label stubs — same shape as the reflect2 stubs (link, then panic). Worth proposing only if tinygo-org#5591 is received well; if maintainers dislike that pattern, this one shouldn't be filed either.

OOM heap-cap commit — superseded upstream by runtimeFatal("out of memory"), resolved during the sync. Nothing to propose.

Notes

• Build this fork with -tags=llvm22; upstream's go-llvm now defaults to LLVM 20.
• The LLVM21+ golden-IR normalization was dropped (5b575a5) — upstream handles it, verified by tests passing without it.

[0pcom]

Status sweep, 2026-08-14. Nothing has merged and nothing has been rejected — all seven PRs are still open, and none has a review decision yet (tinygo-org#5530's earlier review feedback remains addressed).

PR	state
tinygo-org#5530 — signals under the threads scheduler	open
tinygo-org#5587 — interp: recoverable timeout per package initializer	open
tinygo-org#5588 — reflect.MakeChan	open
tinygo-org#5589 — os.File.Chown	open
tinygo-org#5590 — runtime.MemStats.NumGC	open
tinygo-org#5591 — reflect2 linkname stubs	open
tinygo-org#5592 — runtime/debug ReadBuildInfo	open

Resolved — the LLVM 21/22 work is upstream now

The Notes section told people to build with -tags=llvm22 because upstream defaulted to LLVM 20. That is out of date. Upstream dev now carries cgo/libclang_config_llvm21.go, cgo/libclang_config_llvm22.go, compileopts/features_llvm22.go and compileopts/triple_llvm22.go, and the interp switch-instruction fix is byte-identical to ours.

Verified rather than assumed: git diff upstream/dev HEAD -- cgo/ compiler/ interp/compiler.go transform/allocs.go is empty. The only remaining difference across those paths is one go.mod line, where golang.org/x/mod v0.37.0 is direct here and indirect upstream — tidy noise, not a change to propose.

So 9e64c7c, d642576 and acfbe0a join the OOM heap-cap commit as superseded. They are still listed by git log upstream/dev..HEAD only because they are distinct commit objects; there is no content left to send.

Newly recorded as held

Golden-IR normalization drop (5b575a5) — this was filed under Notes as "upstream handles it", which reads backwards. Upstream still has normalizeIR; this fork removed it and its call sites in transform/transform_test.go (a 51-line delta, the largest remaining fork-only change outside the proposed PRs). It is a fork-local test simplification that is neither proposed nor deliberately held. It should become one or the other.

Unchanged holds

src/net submodule bumps (substance lives in 0magnet/net — see 0magnet/net#1, where one commit turned out never to have been blocked and is now filed as tinygo-org/net#70), the crypto/tls host-target override, and the runtime/pprof label stubs pending how tinygo-org#5591 lands.

[0pcom]

Status sweep, 2026-08-16

All seven PRs are still open; none merged. Maintainer feedback has arrived on five of them and has now been addressed.

Actioned:

• os: add File.Chown on the unix path tinygo-org/tinygo#5589 os.File.Chown — reviewers asked for the change-history comment to go; reduced to documenting the method, matching the package-level Chown wording. Rebased.
• reflect: implement MakeChan tinygo-org/tinygo#5588 reflect.MakeChan — test added (TestTinyMakeChan), covering buffered/unbuffered creation and the three panic paths. Rebased.
• runtime: add MemStats.NumGC for API compatibility tinygo-org/tinygo#5590 runtime.MemStats.NumGC — dgryski asked for it to be real rather than a constant zero. Now counted in runGC for gc_blocks, taken from bdwgc's gc_no for gc_boehm, and 0 for gc_leaking (which never completes a cycle). One approval already from b0ch3nski.
• runtime/debug: populate BuildInfo so ReadBuildInfo works tinygo-org/tinygo#5592 runtime/debug ReadBuildInfo — was conflicting, and b0ch3nski spotted that it bumped every submodule. Both fixed: all seven submodule pointers dropped (including src/net, which pointed at this fork and had no business in an upstream PR), rebased onto dev, ./builder compiles.

Still needs a decision, not code:

• reflect: provide the linkname targets modern-go/reflect2 expects tinygo-org/tinygo#5591 reflect2 linknames — dgryski does not want a single package named explicitly, and asks whether it can be built on the existing reflect instead. That is a design question. Note the knock-on: the held runtime/pprof label stubs above use the same pattern, so if this one is rejected that decision covers both.
• interp: make the timeout recoverable, per package initializer tinygo-org/tinygo#5587 interp timeout — dgryski linked the earlier attempt in interp: move initializaton to runtime on interp timeout tinygo-org/tinygo#2384; wants reading and answering.
• runtime: deliver signals under the threads scheduler when blocked on I/O tinygo-org/tinygo#5530 signals under the threads scheduler — review already addressed in 7243c59 and e626202; waiting on deadprogram.

Housekeeping: CI has never run on any of these — the repo gates workflow runs on fork PRs behind maintainer approval, so the yellow "unstable" state is not a failure. Also, this fork's default branch is now 64 commits behind upstream/dev, which is what let tinygo-org#5592 drift into conflict; the five PR branches are rebased but the fork itself is not.

[0pcom]

Both design questions resolved, 2026-08-16

tinygo-org#5591 (reflect2 linknames) — closed. dgryski was right not to want a package named explicitly, and the ground moved besides:

• Implementing the six helpers on top of reflect would not have helped. reflect2 reads Go's internal rtype layout directly for field offsets and kind bits, so on TinyGo it would still misbehave — a link error would become a wrong answer.
• The stubs let a program link and then panic when the path is hit, which turns a build-time failure into a runtime one. A missing-symbol link error is better diagnostics than a panic inside a JSON encoder.
• The consumer no longer needs it: skywire now selects encoding/json over jsoniter behind a tinygo build tag (interface_tinygo.go, json_tinygo.go), and neither of its TinyGo targets reaches reflect2. Verified with go list -deps -tags tinygo.

This also retires the held runtime/pprof label stubs. They used the same link-and-panic pattern, so per the note above they will not be filed. That item is closed, not deferred.

tinygo-org#5587 (interp timeout) — branch cleaned, design question put back to maintainers.

The branch was carrying reverts of upstream work: the sync.OnceFunc(program.Build) SSA change, the normalizeIR test refactor, the LLVM 22 golden files, and 59 lines of machine_esp32.go including the RTC pullup/pulldown fix. Same class of defect as tinygo-org#5592's submodule bumps. Rebuilt as the three interp files only, rebased on dev, go test -tags=llvm22 ./interp/ green.

dgryski linked tinygo-org#2384 — his own 2021 attempt, closed unmerged in 2023 after aykevl and niaow objected that timeouts make builds non-reproducible. That objection lands on this PR harder than on tinygo-org#2384: there the clock decided whether the build failed, here it decides what the binary contains, silently. A per-initializer budget on the existing callsExecuted counter would make it deterministic, and that is the direction they converged on — but aykevl also asked for the failing example so the underlying interp bug could be fixed instead, and never got one.

So the reply supplies it: crypto/aes key expansion and secp256k1 table setup (btcec, via skycoin/src/cipher). Either they treat it as an interp bug, in which case this PR should close, or they take the counter version. Not proceeding until they say which — writing the counter version now would risk re-litigating tinygo-org#2384 by accident.

Lesson for this fork: two of seven PR branches carried unrelated upstream reverts. Both were only caught by diffing against upstream/dev rather than trusting the branch. Worth doing that check before filing, not after.