cfallin commented on issue #1140:
Is this still a relevant issue with the new backends @abrown?
abrown commented on issue #1140:
Nope, closing.
abrown closed issue #1140:
As noted in https://github.com/CraneStation/cranelift/pull/1120#discussion_r335156013, when moving a scalar value (
I32
orI64
) to a vector register in x86, the destination must be a aI64x2
due to typing restrictions that ensure vectors have more than one lane. This may be rather confusing when looking at CLIF IR because the intent of that move (using thebitcast
instruction) is to zero out the upper bits and only load the lower bits using the scalar value--it's more aI128x1
than anything else.There are several options here, some of which are:
- leave as-is with
I64x2
and remove the TODO comment- change the typing restrictions to allow
I128x1
- use a different CLIF instruction altogether
- etc.
Last updated: Nov 22 2024 at 16:03 UTC