fix(types): make #pointeeProjection confluent with source-first unwrapping strategy

[Stevengre]

Summary

• Fixes stuck execution caused by incorrect projection chains from #pointeeProjection
• Root cause: non-deterministic overlap between struct and array rules — when source is a struct and target is an array (or vice versa), both rules match and the LLVM backend's choice determines correctness
• Commit history shows why [priority(45)] alone doesn't work (it fixes one case but regresses the other)
• Fix: always unwrap the source type first; target-side unwrapping deferred to #pointeeProjectionTarget
• Reproducer from @dkcumming's review comment (minimal Wrapper([u8; 2]) cast)

Test plan

• ptr-cast-wrapper-to-array passes (new reproducer)
• iter_next_2 still passes (regression test for opposite overlap: [Thing;3] → Thing)
• Full integration test suite passes

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: beef218f11

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

[dkcumming]

This feels like it is addressing the symptom of a problem and not the problem itself. The description given does not fill me with much confidence that the problem is being accurately understood or addressed. It feels to me the problem is the projection chain is being built incorrect, and this dismisses a particular bad projection chain. But really we want to construct correct projection chains so everything is fixed for the general case and not this particular instance of a bad chain.

I tried looking the example you provided but it adds a lot of noise, I think something like this is much better to analyse and has the same problem and everything goes wrong within the first basic block:

struct Wrapper([u8; 2]);

fn main() {
    let w = Wrapper([11, 22]);
    let arr: [u8; 2] = unsafe { *((&w) as *const Wrapper as *const [u8; 2]) };
    assert_eq!(arr, [11, 22]);
}

This proof gets stuck on Node 72 (Statement: _3 <- Use(cp((*_4)))) with stuck state:

#traverseProjection ( toLocal ( 1 ) , Aggregate ( variantIdx ( 0 ) , ListItem (Range ( ListItem (Integer ( 11 , 8 , false ))
     ListItem (Integer ( 22 , 8 , false ))
     ))
   ) , projectionElemSingletonArray  projectionElemField ( fieldIdx ( 0 ) , ty ( 80 ) )  projectionElemConstantIndex ( ... offset: 0 , minLength: 0 , fromEnd: false )  .ProjectionElems , .Contexts )

Which shows there is projectionElemSingletonArray then projectionElemField(0, ...) then projectionElemConstantIndex(0, ...). The first thing that comes to mind is the rules @jberthold added that should cancel the projections out types.md:65-67, but they don't apply when the projectionElemField is intersected.

// special cancellation rules with higher priority
rule maybeConcatProj(projectionElemSingletonArray, projectionElemConstantIndex(0, 0, false) REST:ProjectionElems) => REST [priority(40)]
rule maybeConcatProj(projectionElemConstantIndex(0, 0, false), projectionElemSingletonArray REST:ProjectionElems) => REST [priority(40)]

So is that field projection the problem, or are the array projections the problem? The projections are first introduced at Node 64 (Statement: _4 <- Cast-PtrToPtr mv(5)) which is the result of casting ty(81)/*const Wrapper to ty(76)/*const [u8; 2]:

#cast ( PtrLocal ( 0 , place ( ... local: local ( 1 ) , projection: .ProjectionElems ) , mutabilityNot , metadata ( noMetadataSize , 0 , noMetadataSize ) ) , castKindPtrToPtr , ty ( 81 ) , ty ( 76 ) )

giving:

PtrLocal ( 0 , place ( ... local: local ( 1 ) , projection: projectionElemSingletonArray  projectionElemField ( fieldIdx ( 0 ) , ty ( 80 ) )  projectionElemConstantIndex ( ... offset: 0 , minLength: 0 , fromEnd: false )  .ProjectionElems ) , mutabilityNot , metadata ( staticSize ( 2 ) , 0 , noMetadataSize ) )

So this looks wrong, it shouldn't have these array projections, just the field access.

This projection chain is built from data.md:1458,1469 which appends projections by #typeProjection. If you follow the logic through, what you find is that #pointeeProjection for typeInfoArrayType is being matched:

rule #pointeeProjection(typeInfoArrayType(TY1, _), TY2)
    => maybeConcatProj(
          projectionElemConstantIndex(0, 0, false),
          #pointeeProjection(lookupTy(TY1), TY2)
        )
  rule #pointeeProjection(TY1, typeInfoArrayType(TY2, _))
    => maybeConcatProj(
          projectionElemSingletonArray,
          #pointeeProjection(TY1, lookupTy(TY2))
        )

The comment and the PR that introduces the rule #936 would indicate this is for casts of the form T -> [T; N] or [T; N] -> T, however there is no appropriate guard on the rule, or restriction for the rule to only match for casts of that case. The approach to solve this is to choose an appropriate guard (no easy I imagine because the guard will rely on context that the casts is of the form T -> [T; N] or [T; N] -> T. So problem the flow of rules needs to change.

As hacky tests I added [priority(55)] on the rules, and tried a contrived restrictive guard. Both approaches resulted in the minimised test passing, with node 64 only having the projectionFieldElem(0, ...) as expected:

 PtrLocal ( 0 , place ( ... local: local ( 1 ) , projection: projectionElemField ( fieldIdx ( 0 ) , ty ( 80 ) )  .ProjectionElems ) , mutabilityNot , metadata ( staticSize ( 2 ) , 0 , noMetadataSize ) )

I personally would change the test to be the minimised one I provided, but that is not strictly necessary. However I think any rule being suggested to solve this issue should be argued with thorough evidence that they are addressing the problem and not a symptom of the problem.

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: aeefea2a3b

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 0877ae1791

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: c22c881c74

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".