Skip to content

fix(deps): update module github.com/go-asmgen/asmgen to v0.9.0 - #6

Merged
tannevaled merged 1 commit into
mainfrom
renovate/deps
Sep 30, 2026
Merged

tannevaled merged 1 commit into
mainfrom
renovate/deps

Conversation

@tannevaled

Copy link
Copy Markdown
Contributor

This PR contains the following updates:

Package Change Age Confidence
github.com/go-asmgen/asmgen v0.8.0 → v0.9.0 age confidence

Release Notes

go-asmgen/asmgen (github.com/go-asmgen/asmgen)

v0.9.0: — amd64.FeatureProbe

Compare Source

Adds amd64.FeatureProbe(name, amd64.AVX2), which emits the CPUID/XGETBV gate a dispatched SIMD kernel needs in front of it, as a NOSPLIT func() bool.

The fleet had two answers to one question: go-fft/fft hand-wrote 24 lines of Plan 9 assembly for AVX2, while go-simd/popcount and go-simd/base64 took a runtime dependency on golang.org/x/sys/cpu. A generator that already writes the kernel can now write the gate, so a package that wants no runtime dependency does not have to hand-write CPUID to get one.

⛔ The OS half is the part a hand-rolled probe drops: a CPU can report AVX2 on a kernel that does not save YMM state. The probe checks OSXSAVE, then XGETBV's XMM and YMM bits in XCR0, and only then CPUID leaf 7 — the same predicate x/sys/cpu computes. A test holds go-fft/fft's committed cpu_amd64.s as a constant and compares the emitted instruction sequence line by line; they agree on all 23, differing only in the label, which is derived from the function name so two probes can share a file.

POPCNT is the second Feature and deliberately the last; an unrecognised one panics rather than emitting a gate that says yes without asking.

Also in this release since v0.8.0: the go directive raised to 1.26.4, go test over the wasm examples, and documentation corrections (stale version pin, over-claimed target parity).

See #​6, #​5, #​3, #​2.


Configuration

📅 Schedule: (UTC)

  • Branch creation
    • At any time (no schedule defined)
  • Automerge
    • At any time (no schedule defined)

🚦 Automerge: Enabled.

♻ Rebasing: Whenever PR is behind base branch, or you tick the rebase/retry checkbox.

🔕 Ignore: Close this PR and you won't be reminded about this update again.


  • If you want to rebase/retry this PR, check this box

This PR has been generated by Mend Renovate CLI.

@tannevaled
tannevaled merged commit 73ce783 into main Sep 30, 2026
9 checks passed
@tannevaled
tannevaled deleted the renovate/deps branch September 30, 2026 18:30
tannevaled added a commit that referenced this pull request Oct 4, 2026
…tor (#10)

overflow_test.go wrote its overflowing shapes as 1<<62 and 1<<40 literals,
which do not compile where int is 32 bits, so `go test` failed to build on
linux/386, arm, mips and mipsle. The extents are now derived from
bits.UintSize, so the same products overflow on both word sizes; the suite
passes on real 386 hardware.

Nothing caught this because CI never ran a 32-bit target. Three lanes now
cover it:
- arch 386: GOARCH=386 natively on the amd64 runner (not qemu-i386).
- arch arm (qemu): a 32-bit ARM lane beside the existing qemu ones.
- compile tests on every target: `go test -c` for each `go tool dist list`
  pair (`go vet` for android/ios, which cannot link without cgo). Checked
  with a planted 1<<62 literal, which both modes refuse.

The asm generators carried //go:build ignore, so `go mod tidy` saw no
import of go-asmgen and dropped the requirement. Renovate's "update to
v0.9.0" (#6) did exactly that, leaving CI to resolve whatever version
-mod=mod found. The tag is gone (each generator is the only file of its
own main package), and go-asmgen is pinned at v0.10.0 again, which
regenerates the committed .s byte for byte.

Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant