wasmtime / PR #14373 cranelift: treat regionless loads as... · git-wasmtime

Stream: git-wasmtime

Topic: wasmtime / PR #14373 cranelift: treat regionless loads as...

Wasmtime GitHub notifications bot (Sep 22 2026 at 02:32):

agourakis82 opened PR #14373 from agourakis82:users/agourakis/cranelift-regionless-load-alias-fence to bytecodealliance:main:

Summary

Alias analysis keyed regionless loads only on last_fence. A later named-region store therefore did not invalidate the regionless load's memory version, so store-to-load / load-to-load forwarding and DSE could drop an intervening region store that the regionless load still observes (#14323).

Fix

When computing the last-store version for a regionless access, meet every named-region slot with last_fence (collapses to Unknown if any named store happened since the fence).

When a regionless load observes memory, observe all region stores as well as the last fence (same shape as an all-region observer).

Update alias/no-region.clif so its first case matches the corrected invariant: a regionless load keeps a preceding named-region store observable.

Test plan

New: cranelift/filetests/filetests/runtests/issue-14323-regionless-load.clif (both issue repros)

Updated: cranelift/filetests/filetests/alias/no-region.clif

Local: clif-util test on the issue repros, alias/, runtests/alias.clif, egraph/alias_analysis.clif

Fixes #14323

AI assistance

Developed with AI tool assistance (Claude Code) under human direction and review. I am the author of record, reviewed the generated diffs before opening the PR, and can answer questions about the change during review.

Wasmtime GitHub notifications bot (Sep 22 2026 at 02:32):

agourakis82 requested wasmtime-compiler-reviewers for a review on PR #14373.

Wasmtime GitHub notifications bot (Sep 22 2026 at 02:32):

agourakis82 requested fitzgen for a review on PR #14373.

Wasmtime GitHub notifications bot (Sep 22 2026 at 03:48):

github-actions[bot] added the label cranelift on PR #14373.

Wasmtime GitHub notifications bot (Sep 22 2026 at 14:48):

agourakis82 commented on PR #14373:

Thanks @cfallin — this is the key design question.

I read the issue as Alex framed it (regionless load may still observe a named-region store to the same address), and the PR follows that reading. Your preferred model — “no region” as its own disjoint bucket, with no globally-aliasing ops — matches the existing module docs / no-region.clif comment much more closely, and would make the original #14323 tests frontend-invariant violations rather than mid-end bugs.

I’m happy to treat this PR as provisional pending the Cranelift meeting pin-down:

if the meeting keeps “no region == disjoint other”, I can close/revert this and we can close #14323 as invalid IR;

if the meeting decides regionless may alias named regions, this patch (or a tightened variant) stays useful.

Either outcome is fine with me; I just want the invariant written down one way so alias analysis and frontends don’t diverge.

Wasmtime GitHub notifications bot (Sep 22 2026 at 14:50):

agourakis82 commented on PR #14373:

Follow-up design writeup on the issue, focused on the existing store/load asymmetry:

https://github.com/bytecodealliance/wasmtime/issues/14323#issuecomment-5778631844

(see the newer comment just posted with the G vs D table)

@fitzgen you already own most of this file — a one-line call on G vs D unblocks either keeping this PR or replacing it with the disjoint-other store demotion.

Wasmtime GitHub notifications bot (Sep 22 2026 at 14:51):

agourakis82 commented on PR #14373:

Correction — full G vs D writeup with the store/load asymmetry table is here:

https://github.com/bytecodealliance/wasmtime/issues/14323#issuecomment-5778672877

@fitzgen a one-line call on G (regionless = may-alias-all; keep this PR) vs D (regionless = disjoint other; demote regionless stores from fence and close this PR) unblocks the resolution either way.

Wasmtime GitHub notifications bot (Sep 22 2026 at 15:22):

agourakis82 updated PR #14373.

Wasmtime GitHub notifications bot (Sep 22 2026 at 15:22):

agourakis82 commented on PR #14373:

Rebased on latest main.

Also: I prototyped Model D locally; it breaks alias/issue-13508.clif, whose runtest requires a regionless store to clobber a later named-region load ([-1,-1,42]). Writeup on the issue:

https://github.com/bytecodealliance/wasmtime/issues/14323#issuecomment-new

(see latest comment titled “Model D is falsified by an in-tree runtest”)

So this PR remains the load-side consistency fix for the store semantics already in tree.

Wasmtime GitHub notifications bot (Sep 22 2026 at 15:23):

agourakis82 commented on PR #14373:

Correction — falsification writeup URL:

https://github.com/bytecodealliance/wasmtime/issues/14323#issuecomment-5779157701

issue-13508’s run expectation ([-1, -1, 42]) requires regionless stores to be visible to named-region loads. Model D cannot satisfy that without declaring the test invalid. This PR keeps Model G and only fixes the load side.

Wasmtime GitHub notifications bot (Sep 22 2026 at 15:33):

agourakis82 commented on PR #14373:

Parking this on the meeting outcome for “explicit no-region” vs “no-region as fence”.

I’m fine either rewriting this PR toward the January explicit-no-region store behavior, or keeping the load-side fence consistency — whichever gets pinned down.

Wasmtime GitHub notifications bot (Sep 22 2026 at 15:37):

agourakis82 edited a comment on PR #14373:

Thanks. Parking this on the meeting.

#14373 follows the issue report’s reading (no-region load can see a named-region store). If the call is explicit no-region, I’ll rework/close it.

Wasmtime GitHub notifications bot (Sep 22 2026 at 15:37):

agourakis82 edited a comment on PR #14373:

Wrote up the store/load asymmetry on the issue.

Wasmtime GitHub notifications bot (Sep 22 2026 at 15:37):

agourakis82 edited a comment on PR #14373:

Better summary of the two options is here: https://github.com/bytecodealliance/wasmtime/issues/14323#issuecomment-5778672877

Wasmtime GitHub notifications bot (Sep 22 2026 at 15:37):

agourakis82 edited a comment on PR #14373:

Rebased on main.

Tried the explicit-no-region store change locally; notes on the issue.

Wasmtime GitHub notifications bot (Sep 22 2026 at 15:37):

agourakis82 edited a comment on PR #14373:

https://github.com/bytecodealliance/wasmtime/issues/14323#issuecomment-5779157701

Wasmtime GitHub notifications bot (Sep 22 2026 at 15:37):

agourakis82 edited a comment on PR #14373:

Parking on the meeting: explicit no-region vs no-region as fence. Fine either way.

Wasmtime GitHub notifications bot (Sep 22 2026 at 15:38):

agourakis82 edited PR #14373:

Regionless loads were only keyed/observed via last_fence, while named-region stores update regions[r]. That lets forwarding/DSE miss an intervening named store that a later regionless load can still see (#14323).

This patch makes regionless loads meet/observe the named-region slots as well (matching how regionless stores are already treated as fences).

Also updates alias/no-region.clif and adds runtests/issue-14323-regionless-load.clif.

Parking on the meeting for “explicit no-region” vs “no-region as fence”; happy to rework if the call goes the other way.

Fixes #14323

Assisted-by: Claude

Wasmtime GitHub notifications bot (Sep 22 2026 at 17:11):

agourakis82 updated PR #14373.

Last updated: Oct 11 2026 at 04:10 UTC