Stream: git-wasmtime

Topic: wasmtime / PR #14770 cranelift: Implement 128-bit big-end...

Wasmtime GitHub notifications bot (Oct 15 2026 at 11:50):

Gelbpunkt opened PR #14770 from Gelbpunkt:s390x-128bit-atomic-load-store to bytecodealliance:main:

Since we store I128 in vector registers, we perform loads into vector registers and use vector stores to implement atomic store.

This was suggested by @theotherjimmy in yesterday's Cranelift meeting (thanks!).

I've verified that the atomic load/store part of the atomic-128 runtest passes on s390x in qemu-user and on z15 in a z/VM environment, but since the atomic-128 runtest also requires atomic_cas (which I have implemented and will PR as a follow-up) and atomic_rmw (WIP), it cannot be enabled on s390x yet.

Wasmtime GitHub notifications bot (Oct 15 2026 at 11:50):

Gelbpunkt requested wasmtime-compiler-s390x-reviewers for a review on PR #14770.

Wasmtime GitHub notifications bot (Oct 15 2026 at 11:50):

Gelbpunkt requested alexcrichton for a review on PR #14770.

Wasmtime GitHub notifications bot (Oct 15 2026 at 11:50):

Gelbpunkt requested wasmtime-compiler-reviewers for a review on PR #14770.

Wasmtime GitHub notifications bot (Oct 15 2026 at 13:10):

Gelbpunkt edited PR #14770:

Since we store I128 in vector registers, we perform loads into vector registers and use vector stores to implement atomic store. This is different from the other backends, which use "regular" atomic load and store instructions from GPRs, which also exist on s390x in the form of LPQ and STPQ. However, s390x differs from the other supported ISAs: its ABI mandates that i128 must be passed indirectly rather than in a pair of GPRs. Using the value of an i128 parameter will generate a load of the value from memory via gen_load (and analogous for gen_store when an i128 is returned), which only gets one Reg to generate the load into - and it will only fit in a VR. Therefore, it makes sense to always keep our i128s in VRs. This also makes sense for atomic_rmw later, where vector operations will generally let us perform the modifications more efficiently than on pairs of GPRs. Therefore, we keep our i128s in vector registers and only move them to GPRs when absolutely necessary (e.g. for CDSG).

This was suggested by @theotherjimmy in yesterday's Cranelift meeting (thanks!).

I've verified that the atomic load/store part of the atomic-128 runtest passes on s390x in qemu-user and on z15 in a z/VM environment, but since the atomic-128 runtest also requires atomic_cas (which I have implemented and will PR as a follow-up) and atomic_rmw (WIP), it cannot be enabled on s390x yet.

Wasmtime GitHub notifications bot (Oct 15 2026 at 13:36):

:thumbs_up: theotherjimmy submitted PR review.

Wasmtime GitHub notifications bot (Oct 15 2026 at 14:24):

:thumbs_up: alexcrichton submitted PR review.

Wasmtime GitHub notifications bot (Oct 15 2026 at 14:24):

alexcrichton added PR #14770 cranelift: Implement 128-bit big-endian atomic load/store for s390x to the merge queue.

Wasmtime GitHub notifications bot (Oct 15 2026 at 14:32):

:repeat: uweigand submitted PR review:

Needs a fix to avoid incorrect little-endian stores.

(We could implement little-endian stores easily on z15 and later, but on z14 we'd probably have to use a vector permute.)

Wasmtime GitHub notifications bot (Oct 15 2026 at 14:32):

:speech_balloon: uweigand created PR review comment:

This is only correct for big-endian stores, so it should have the same check as above.

Wasmtime GitHub notifications bot (Oct 15 2026 at 14:48):

:memo: Gelbpunkt submitted PR review.

Wasmtime GitHub notifications bot (Oct 15 2026 at 14:48):

:speech_balloon: Gelbpunkt created PR review comment:

Ah, good point. I overlooked this because istore64_impl takes care of handling it

Wasmtime GitHub notifications bot (Oct 15 2026 at 14:50):

Gelbpunkt updated PR #14770.

Wasmtime GitHub notifications bot (Oct 15 2026 at 14:50):

:memo: Gelbpunkt submitted PR review.

Wasmtime GitHub notifications bot (Oct 15 2026 at 14:50):

:speech_balloon: Gelbpunkt created PR review comment:

Fixed!

Wasmtime GitHub notifications bot (Oct 15 2026 at 14:51):

alexcrichton removed PR #14770 cranelift: Implement 128-bit big-endian atomic load/store for s390x from the merge queue.

Wasmtime GitHub notifications bot (Oct 15 2026 at 14:52):

alexcrichton has enabled auto merge for PR #14770.

Wasmtime GitHub notifications bot (Oct 15 2026 at 15:04):

:thumbs_up: uweigand submitted PR review:

LGTM now, thanks!

Wasmtime GitHub notifications bot (Oct 15 2026 at 15:12):

alexcrichton added PR #14770 cranelift: Implement 128-bit big-endian atomic load/store for s390x to the merge queue.

Wasmtime GitHub notifications bot (Oct 15 2026 at 15:41):

:check: alexcrichton merged PR #14770.

Wasmtime GitHub notifications bot (Oct 15 2026 at 15:41):

alexcrichton removed PR #14770 cranelift: Implement 128-bit big-endian atomic load/store for s390x from the merge queue.

Last updated: Oct 31 2026 at 23:03 UTC