## Significant changes and quirks

While some of the items below are already mentioned separately, this page tracks the
main changes from the old trait system implementation. This also mentions some ways
in which the solver significantly diverges from an idealized implementation. This
document simplifies and ignores edge cases. It is recommended to add an implicit
"mostly" to each statement.

### Canonicalization

The new solver uses [canonicalization] when evaluating nested goals. In case there
are possibly multiple candidates, each candidate is eagerly canonicalized. We then
attempt to merge their canonical responses. This differs from the old implementation
which does not use canonicalization inside of the trait system.

This has a some major impacts on the design of both solvers. Without using
canonicalization to stash the constraints of candidates, candidate selection has
to discard the constraints of each candidate, only applying the constraints by
reevaluating the candidate after it has been selected: [source][evaluate_stack].
Without canonicalization it is also not possible to cache the inference constraints
from evaluating a goal. This causes the old implementation to have two systems:
*evaluate* and *fulfill*. *Evaluation* is cached, does not apply inference constraints
and is used when selecting candidates. *Fulfillment* applies inference and region
constraints is not cached and applies inference constraints.

By using canonicalization, the new implementation is able to merge *evaluation* and
*fulfillment*, avoiding complexity and subtle differences in behavior. It greatly
simplifies caching and prevents accidentally relying on untracked information.
It allows us to avoid reevaluating candidates after selection and enables us to merge
the responses of multiple candidates. However, canonicalizing goals during evaluation
forces the new implementation to use a fixpoint algorithm when encountering cycles
during trait solving: [source][cycle-fixpoint].


### Deferred alias equality

The new implementation emits `AliasRelate` goals when relating aliases while the
old implementation structurally relates the aliases instead. This enables the
new solver to stall equality until it is able to normalize the related aliases.

The behavior of the old solver is incomplete and relies on eager normalization
which replaces ambiguous aliases with inference variables. As this is
not possible for aliases containing bound variables, the old implementation does