proposal for unaligned loads and stores

[glessard]

Implementation PR: swiftlang/swift#41033

[rjmccall]

@glessard Is this still awaiting an implementation?

[glessard]

The proposal and implementation are ready for review.

[glessard]

I could remove excessive references to drafts, I guess.

[rjmccall]

Ah, I see the link in the PR summary. Could you link it from the document, please? And yeah, if there's any draft cleanup you want to do, please do so.

[glessard]

@rjmccall Ready to go on my end; let me know if you need any other changes.

[rjmccall]

Thanks. We'll try to talk about it in the Core Team ASAP and figure out how it's going to be scheduled, but I wanted to make sure you felt it was ready before we started that process.

[rjmccall]

We'd like to go ahead and run this. I'll be review manager. I think two weeks, running from March 30th until April 12th, is probably appropriate; is that okay with you as the author?

[glessard]

Sounds great. Thanks!

[rjmccall]

Review manager feedback:

• The title of the proposal just talks about unaligned loads, but the proposal includes both load and store operations.
• Please address Karl's idea about asserting POD-ness of as a precondition.
• Personally, I have to say that keeping the alignment requirement for load but removing it from storeBytes feels like an odd decision. You don't have to respond to this; I'm just putting it out there.

With that, I'll put this in review.

[glessard]

loadUnaligned has a debugPrecondition for POD-ness. I'll clarify this in the document and in the doc-comment.

[rjmccall]

loadUnaligned has a debugPrecondition for POD-ness. I'll clarify this in the document and in the doc-comment.

Thank you.

[glessard]

Regarding storeBytes and storeBytesUnaligned: one issue here is that the destination is always, by definition, byte-aligned. At the builtins level, there is also not a different way to spell the two. If we had both functions, they would be identical in release mode, and differ only by their set of debugPrecondition calls.

[rjmccall]

Regarding storeBytes and storeBytesUnaligned: one issue here is that the destination is always, by definition, byte-aligned. At the builtins level, there is also not a different way to spell the two. If we had both functions, they would be identical in release mode, and differ only by their set of debugPrecondition calls.

I mean, we could add a builtin that does an aligned store. But in fact we don't need to: since you're requiring the type to be POD, you can simply use the existing initialize or assign builtin.

[rjmccall]

Do you want me to put this into review? I can just make these same comments in the review thread, and we can see what the community thinks. Worst case, the Core Team asks you to revise the proposal.

[glessard]

Yes, go ahead. Thanks!