Me:

• https://github.com/xen0n
• Member of the Go loong64 porters team (@golang/loong64)
• Tinkers with things for fun
  • Linux/LoongArch port reviewer
  • Contributor to LLVM/QEMU/binutils/glibc/...
  • Gentoo developer, Gentoo/LoongArch

LoongArch:

• https://github.com/loongson/LoongArch-Documentation
• https://blog.xen0n.name/posts/tinkering/loongarch-faq/
• RISC ISA blending traditional principles with practical experiences
  • Influenced by MIPS and RISC-V
  • They say it's gonna become open standards

• Timeline of the Loongson fork (https://bbs.loongarch.org/d/95-loongarch-slackware-current-bootstrap/42)
• Unfortunate bikeshedding caused by lack of communication
  • Originally GOARCH=larch64: https://github.com/loongson/go/commit/40c3a0eae3968c2a9946c4e72036be6592c5b772
  • Then GOARCH=loongarch64: https://github.com/loongson/go/commit/2d68237cb4051ce313dd3254b33a0b65019c977e
  • Opposed by community: https://go.dev/issue/46229
  • Final decision is GOARCH=loong64

Apparently Loongson took their inspirations from the mips64x port:

• Asm syntax (I don't like xxxV)
• Code organization and style (smell)

Unfortunately my take was a false start (lack of personal time & priorities), so these persisted into the eventual merge. Most (if not all) are implementation details so we could still progressively improve, though.

Support among the ecosystem:

• Infra
  • x/sys x/net: any tagged version is OK
  • golangci-lint: binaries offered since v1.51.0 (2023-02-02)
  • dlv: stalled due to lack of CI
• CI/CD
  • goreleaser: supported since v1.11.0 (2022-08-29)
  • gox: stalled on review

Popular apps that work:

• frp
• gitea
• helm
• hugo
• kubectl
• syncthing
• more TBD

Popular apps that don't work yet:

• Docker (? -- needs confirmation)
• Kubernetes (need check)
• more TBD

How does one port to loong64?

• Ensure buildability with go1.19+
• Ensure up-to-date x/sys and related modules
• Watch out for particularly problematic deps
  • bbolt - need ≥ 1.3.7
  • psutil - need ≥ 3.23.2
  • TODO: more of this

Community participation:

go1.20:

• native rotates (@wdvxdr1123)
• intrinsify Add64, Sub64 (@wdvxdr1123)
• enable branch elimination pass (@wdvxdr1123)

go1.21:

• improve logical shifts with MASKEQZ (@wdvxdr1123)
• MULH only when necessary (@wdvxdr1123)
• separate DIV and REM ops (@wdvxdr1123)

Thanks!

Current work of the loong64 team (to show up in go1.21):

• buildmode=pie
  • Security hardening by default: Gentoo, many CNCF projects, others
  • Seems PIE support comes for free (without even requiring ELF psABI v2)
• Transition to LoongArch ELF psABI v2
  • Discussion, implementation
  • Unlocks more build modes with less code
    • c-shared
    • access global data via GOT when dynlinking for shared and plugin
• General cleanups
  • asm backend refactors & low-hanging fruits
    • asm case 11 cleanup: no MIPS-esque beq zero, zero, immediately win 0.24% throughput
    • asm case 6 cleanup for beqz/bnez
    • remove family field, support for 4-operand insns
    • single-instruction sign- and zero-extensions
    • PCALIGN, manual asm loop alignment +2.33%
    • auto loop head alignment another +0.29%
    • more TODO
• Reliability improvements
  • 47627.7 overflows float64 -- runtime: floating-point comparison error on Loong64, fix
  • TBD: choose more Gerrit CLs to showcase here
• Low-hanging fruits
  • assembly
    • hash/crc32 +20%
    • xchg micro-benchmark -28.61%
    • And/Or micro-benchmark -44.96%
  • ssa optimization chances
    • Tracker: https://go.dev/issue/59120
    • TrailingZeros
    • FMA
    • Len
    • ReverseBytes
    • bitrev
    • counting ones TODO
    • more CLs TBD

Upstream work:

• loong64 asm backend overhaul
  • machine-generated insn description for maintainability
  • consistent asm grammar
  • some compatibility preserved (mainly for x/sys)
• regabi support
  • another 10%~20% perf win
  • move g from fp to s8: enable frame pointers at all times for debuggability
• internal link support

Possible ecosystem work:

• Communication to make CNCF inroads smooth
• TBD

See you in the bug trackers!