If attachments are not preserved, they get undefined data, which sounds as good as a write.

But when subsequent subpass wants to use that attachment again, which pipeline stage should it use in src?

Also the rules are bit confusing when attachments become undefined.

For example it says the attachment only becomes undefined if there is a dependency. But I mean, if there is a dependency with srcStage=dstStage=NO_COMMANDS, and srcAccess=dstAccess=0, does it even count?

Also because the subpasses can technically overlap, it is weird that an attachment can be at the same time undefined and defined. And it is not sufficiently explained away in the spec:

Above, subpass 1 and 2 can overlap, but in one the attachment is defined, and in the other it is not. The quantum superposition of the state is weird.

In subpass 3, I think the rules in Vulkan break and say the attachment is both defined and undefined, because the subpass depends on both preserving and non-preserving subpass.

Also I think there should be a VU forbidding to reference attachment in preserve attachments, unless it is subsequenly used in another subpass. Otherwisely it would lead to confusion of which subpass last uses an attachment for the purposes of synchronization.