wasmtime / PR #13717 cranelift(aarch64): fold an i32x4 ho... · git-wasmtime

Stream: git-wasmtime

Topic: wasmtime / PR #13717 cranelift(aarch64): fold an i32x4 ho...

Wasmtime GitHub notifications bot (Jun 23 2026 at 14:35):

darmie opened PR #13717 from darmie:aarch64-addv-hsum to bytecodealliance:main:

A complete i32x4 horizontal-add reduction has no CLIF opcode, so it is
expressed as two iadd_pairwisees (which GVN shares) followed by a lane-0
extractlane. Currently that lowers to:
addp v.4s, v.4s, v.4s
addp v.4s, v.4s, v.4s
umov w0, v.s[0]
NEON addv reduces the whole vector in a single instruction, so this contracts
the tree to:
addv s0, v.4s
umov w0, v.s[0]
— one fewer instruction and, more importantly, no serial addp dependency chain.

The mid-end has no egraph rules touching iadd_pairwise, so the structured
reduction reaches isel intact, which keeps the match reliable (a plain iadd
tree would be reassociated).

Follow-up

The cleaner long-term fix is a dedicated horizontal-reduce CLIF op (e.g.
vreduce_iadd) that completes the vall_true / vany_true / vhigh_bits
vector→scalar family and lowers to addv on aarch64 with portable fallbacks
elsewhere. Happy to do that as a follow-up if you'd prefer the IR-level
primitive; this is a minimal aarch64-only implementation for now.

Tests

filetests/isa/aarch64/simd-addv-reduce.clif — precise-output, asserts addv; umov.

filetests/runtests/simd-addv-reduce.clif — interpret + run semantics on aarch64/x86_64/s390x.

The existing isa/aarch64 and runtests suites pass unchanged.

Wasmtime GitHub notifications bot (Jun 23 2026 at 14:35):

darmie requested cfallin for a review on PR #13717.

Wasmtime GitHub notifications bot (Jun 23 2026 at 14:35):

darmie requested wasmtime-compiler-reviewers for a review on PR #13717.

Wasmtime GitHub notifications bot (Jun 23 2026 at 15:34):

:thumbs_up: cfallin submitted PR review:

Thanks -- this looks good!

FWIW, we've talked about the tradeoff of putting more domain-specific opcodes into CLIF vs. using simpler ops and pattern-matching, as we do here; in general the direction has been to keep CLIF simpler ("RISC-like") now that we have a fairly robust pattern-matching language in the backend. We can of course always reconsider that if something specific is really difficult but for now this seems fine.

Wasmtime GitHub notifications bot (Jun 23 2026 at 15:34):

cfallin added PR #13717 cranelift(aarch64): fold an i32x4 horizontal add into a single addv to the merge queue.

Wasmtime GitHub notifications bot (Jun 23 2026 at 15:58):

:check: cfallin merged PR #13717.

Wasmtime GitHub notifications bot (Jun 23 2026 at 15:58):

cfallin removed PR #13717 cranelift(aarch64): fold an i32x4 horizontal add into a single addv from the merge queue.

Last updated: Jul 29 2026 at 05:03 UTC