CTFE: simplify ConstValue by not checking for alignment

[RalfJung]

I hope the test suite actually covers the problematic cases here?

r? @oli-obk

Fixes #61952

[oli-obk]

What do you think about removing the align field of Mplace in const eval by making it an assoc type of the Engine?

[RalfJung]

What do you think about removing the align field of Mplace in const eval by making it an assoc type of the Engine?

Hm. Not sure how well that code would factor.

With #63075, we actually don't need the align field in places any more. However, my expectation is that for raw pointers we might eventually want to relax the rule that says "when you do *ptr, even if you are just computing a field offset, the ptr must be aligned and dereferencable". So I am reluctant to remove all that infrastructure if we have to bring it back later.

[oli-obk]

I don't want to remove the field. Miri will keep using it. Const eval doesn't need it though. If you do UB in const eval, we don't promise we catch it.

[oli-obk]

We should probably find a way to make miri's const_eval query run with all miri-checks on.

[RalfJung]

Miri will keep using it.

With #63075 the field is effectively dead in Miri, though. For now. I think.

[oli-obk]

Ah, ok. We can revisit once that is resolved then. Can you make sure that it's documented $somewhere in the process, so that we actually do revisit?

r=me with travis green

[RalfJung]

Const eval doesn't need it though.

Agreed. But factoring that through the Machine trait will be somewhat messy: MPlaceTy and PlaceTy would get another generic parameter; and the check_ptr_access method in Memory would end up split across both sides of the interface.

[RalfJung]

@bors r=oli-obk

[bors]

📌 Commit cfb13b0eeb620dee597a99c4100b1464f589b731 has been approved by oli-obk

[bors]

⌛ Testing commit cfb13b0eeb620dee597a99c4100b1464f589b731 with merge 4034b4c3188a41efb8be7f415d0252d232da06b4...

[bors]

💔 Test failed - checks-azure

[RalfJung]

@oli-obk @Centril I made validity check alignment even when the machine does not. I am not very happy with check_ptr_access_align as method name though.

(Also that CI failure looks spurious.)

[RalfJung]

Rebased. @oli-obk what do you think of checking alignment in validity even when the machine otherwise ignores alignment (as is now implemented here)?

[oli-obk]

I think it's very reasonable. We can always weaken validity checks if we need to. Hardening them is not easy to do.

@bors r+

[bors]

📌 Commit 0cf4329 has been approved by oli-obk

[bors]

⌛ Testing commit 0cf4329 with merge 4be0675...

[bors]

☀️ Test successful - checks-azure
Approved by: oli-obk
Pushing 4be0675 to master...