wasmtime / issue #1124 Use optimal MOV encodings for SIMD... · git-wasmtime

Stream: git-wasmtime

Topic: wasmtime / issue #1124 Use optimal MOV encodings for SIMD...

Wasmtime GitHub notifications bot (May 04 2022 at 23:11):

@abrown can you check whether this is still an issue in the new backends? We should have better alignment at least of stack slots now, so any remaining MOVUPS can probably become MOVAPS, but we should audit...

Wasmtime GitHub notifications bot (May 04 2022 at 23:18):

abrown commented on issue #1124:

So I saw you mention earlier that the constant pool is now aligned--can you point me to that?

Wasmtime GitHub notifications bot (May 04 2022 at 23:18):

abrown edited a comment on issue #1124:

So I saw you mention earlier that the constant pool is now aligned--can you point me to that? [edit: let me tag @cfallin so that it doesn't get lost in all the notification churn]

Wasmtime GitHub notifications bot (May 04 2022 at 23:22):

cfallin commented on issue #1124:

@abrown that happens due to the alignment method here I think: link. Then we pass that alignment to the MachBuffer here.

Wasmtime GitHub notifications bot (Aug 28 2025 at 16:07):

abrown closed issue #1124:

As @bnjbvr notes in https://github.com/CraneStation/cranelift/pull/992#discussion_r325065157, there are several different x86 opcodes for encoding the movement of SIMD values between registers and between registers and memory (e.g. MOVUPS, MOVAPS, MOVDQU, MOVUPD, etc.). For now, many CLIF instructions (e.g. regmove) use MOVUPS due to a need to get SIMD spec tests up and running to verify correctness. In the future, however, we need to determine the optimal x86 opcode and replace less optimal uses of MOVUPS.

Last updated: Dec 13 2025 at 19:03 UTC