Skip to content

[Go] Model runtime stack semantics and entry maps - #73

Merged
zhouguangyuan0718 merged 1 commit into
llvm23.1.masterfrom
review/runtime-stack-semantics
Aug 13, 2026
Merged

[Go] Model runtime stack semantics and entry maps#73
zhouguangyuan0718 merged 1 commit into
llvm23.1.masterfrom
review/runtime-stack-semantics

Conversation

@zhouguangyuan0718

@zhouguangyuan0718 zhouguangyuan0718 commented Aug 13, 2026

Copy link
Copy Markdown

Summary

  • use go-nosplit as the sole opt-out from the native Go stack check; split functions call runtime.morestack*<ABI0> with an ordinary target call
  • emit one function-level EntryArgsStackMapID record for every GoObj Go function, independent of split/nosplit behavior
  • serialize the entry map as map 0 and use the Machine-CFG-derived PCSP stream to restore map 0 on entry-stack-depth paths, without identifying runtime.morestack* in GoObj emission
  • serialize ordinary statepoints as additional live maps and let same-PC statepoints override a CFG map-0 transition
  • retain complete ABI register homes for morestack retries while omitting unused, potentially poison incoming pointer values from the entry GC bitmap
  • model go-systemstack with g.stackguard1 and runtime.morestackc

Stack position

This is 2/3 in the runtime LLVM support stack.

Depends on #77.
Next/final layer: #76, X86 ABI0/ABIInternal boundary repair.

#75 has been folded into this PR so the stack semantics and the entry-map contract can be reviewed as one coherent change.

Testing

  • fresh Release build of llc for X86 and AArch64
  • 12 GoObjStackMapUtilsTest unit tests
  • 38 Generic/X86/AArch64 GoObj codegen tests
  • dedicated object test covering map 0 -> ordinary map 1 -> CFG-derived map 0 on both targets

@github-actions

Copy link
Copy Markdown

Hello @zhouguangyuan0718 👋

Thank you for submitting a Pull Request (PR) to the LLVM Project. Since this is your first PR, here are a few useful links covering our main contribution policies and review practices.

  • All contributions to LLVM must follow our LLVM AI Tool Use Policy. In particular, if you used AI while working on this PR, remember to add a note to the PR description.
  • The LLVM Code-Review Policy and Practices document contains practical information about the PR process, including how patches are reviewed and accepted, and who can review a PR.
  • Our LLVM Developer Policy describes our expectations for code quality, commit summaries and contains notes on our CI system.

Please reply to this message to confirm that you have read these policies, especially the LLVM AI Tool Use Policy, and that any AI tool usage has been noted in the PR description.


Frequently asked questions

How do I add reviewers?

This PR will be automatically labeled, and the relevant teams will be notified. For some parts of the project, reviewers may also be added automatically.

You can also add reviewers manually using the Reviewers section on this page. If you cannot use that section, it is probably because you do not have write permissions for the repository. In that case, you can request a review by tagging reviewers in a comment using @ followed by their GitHub username.

What if there are no comments?

If you have not received any comments on your PR after a week, you can request a review by pinging the PR with a comment such as “Ping”. The common courtesy ping rate is once a week. Please remember that you are asking for volunteer time from other developers.

Are any special GitHub settings required to contribute to LLVM?

We only require contributors to have a public email address associated with their GitHub commits, see this section of LLVM Developer Policy for details.


If you have questions, feel free to leave a comment on this PR, or ask on LLVM Discord or LLVM Discourse.

Thank you,
The LLVM Community

@zhouguangyuan0718
zhouguangyuan0718 force-pushed the review/runtime-stack-semantics branch from a42f5a4 to 8c41777 Compare August 13, 2026 07:01
@zhouguangyuan0718
zhouguangyuan0718 changed the base branch from llvm23.1.master to review/runtime-abi0-symbol-resolution August 13, 2026 07:02
@zhouguangyuan0718
zhouguangyuan0718 force-pushed the review/runtime-abi0-symbol-resolution branch from aa55715 to e90dafe Compare August 13, 2026 09:15
@zhouguangyuan0718
zhouguangyuan0718 force-pushed the review/runtime-stack-semantics branch from 8c41777 to b0fe7aa Compare August 13, 2026 09:26
@zhouguangyuan0718
zhouguangyuan0718 force-pushed the review/runtime-abi0-symbol-resolution branch from e90dafe to 0c3e791 Compare August 13, 2026 09:34
@zhouguangyuan0718
zhouguangyuan0718 force-pushed the review/runtime-stack-semantics branch 2 times, most recently from e16ab84 to f931e51 Compare August 13, 2026 10:34
@zhouguangyuan0718
zhouguangyuan0718 changed the base branch from review/runtime-abi0-symbol-resolution to llvm23.1.master August 13, 2026 10:34
Emit native Go stack checks by default for GoObj functions and use go-nosplit as the sole opt-out. Lower the morestack slow path as a raw ABI0 call rather than a statepoint.

Give every Go function an independent entry-argument STACKMAP. Serialize it as map zero, derive map-zero transitions from the CFG-aware PCSP stream, and retain ordinary statepoint maps at live callsites.

Preserve complete ABI register homes for stack-growth retries while omitting unused, potentially poison incoming pointers from the entry GC bitmap.
@zhouguangyuan0718
zhouguangyuan0718 force-pushed the review/runtime-stack-semantics branch from f931e51 to 94a25ae Compare August 13, 2026 12:46
@zhouguangyuan0718 zhouguangyuan0718 changed the title [Go] Model runtime split-stack semantics [Go] Model runtime stack semantics and entry maps Aug 13, 2026
@zhouguangyuan0718
zhouguangyuan0718 merged commit 864953e into llvm23.1.master Aug 13, 2026
8 checks passed
zhouguangyuan0718 added a commit that referenced this pull request Aug 13, 2026
## Summary

- keep R14 and XMM15 reserved under both Go calling conventions and
retain the shared Go preserved-register model
- repair ABIInternal reserved state in a late X86 machine pass after
register allocation, frame lowering, and scheduling
- identify direct ABI0 callees solely through the `<ABI0>` symbol
suffix; statepoints use their existing calling-convention operand
- leave late special calls outside the ordinary ABI transition path:
morestack resumes through `gogo`, and write-barrier thunks use their
dedicated ABIInternal clobber contract
- keep SP-relative Go locals while preserving ABI-specific fixed
incoming argument homes
- emit native-compatible symbol-free x86 TLS relocations for the R14
reload

## Design

R14 and XMM15 are fixed reserved registers rather than ordinary
caller-saved or callee-saved allocation resources. The transition repair
therefore does not participate in SelectionDAG lowering and does not
require a target pseudo or a distinct ABI0 register mask.

The `x86-go-abi` pre-emit pass scans finalized MachineInstr calls. For
an ordinary Go call, `<ABI0>` is the only ABI0 identity. It inserts
`XORPS XMM15, XMM15` and a TLS reload of R14 before ABI0-to-ABIInternal
calls or after ABIInternal-to-ABI0 calls.

Calls inserted late without the ordinary Go call mask are intentionally
ignored. This matches native Go behavior for split-stack morestack calls
and write-barrier thunks.

## Stack position

This is the X86 ABI-boundary layer on top of merged #73 and merged #77.

#71 remains the temporary cross-repository integration and payload PR
until this PR is merged.

## Testing

- independent Release plus assertions build of `llc` and `FileCheck`
- focused LLVM lit validation: 7 of 7 passed
- verified normal calls, statepoints, stack-return statepoints, GoObj
relocations, stack growth, PCSP CFG handling, and ABI symbol naming
- verified pass timing: no repair at finalize-isel or immediately before
`x86-go-abi`; repair appears only after the late pass
- verified native-special behavior: no repair is inserted after
morestack or write-barrier calls
zhouguangyuan0718 added a commit that referenced this pull request Aug 13, 2026
This follows up the runtime stack-growth support merged through #73.

The PR intentionally keeps the work in two reviewable commits:

1. Elide morestack for final-frame small leaf functions on AArch64 and
X86, while keeping the entry ArgsPointerMaps STACKMAP independent. X86
also matches native Go's leaf-like handling of the zero-argument panic
helpers, including the extra 8-byte CALL return-PC limit. The LLVM
frontend consumes generic SSA, so it does not emit Go DUFFCOPY/DUFFZERO
calls; inline memory intrinsics remain call-free instead.
2. Restore native Go's PCDATA_StackMapIndex entry sentinel of -1.
ArgsPointerMaps/LocalsPointerMaps pair 0 is still serialized; runtime
getStackMap normalizes -1 to bitmap 0. CFG ranges that return to entry
SP depth also return to -1.

This removes the artificial 32-byte AArch64 stack-growth prefix from
zero-frame leaf functions without removing their entry argument maps.

The GoALLC LLVM CI test list now includes the leaf-policy,
entry-sentinel, and entry-CFG fixtures so the payload jobs exercise
these contracts directly.

Validation:
- Focused leaf/stack-check and sentinel suite: 8/8 passed.
- Broader Generic/X86/AArch64 GoObj and Go CodeGen suite: 52/52 passed;
go-memset-inline was excluded because its stale REP expectation already
fails on the #76 baseline llc.
- git clang-format --diff reports no changes.
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