This is a hack-fix to #53843, but I am worried it might break things
because it makes the "inference pollution" problem worse.
Fixes#53843 (but introduces a bug that someone might notice).
This patch changes `opt_normalize_project_type` so it appends
obligations to a given obligations vector, instead of returning a new
obligations vector.
This change avoids lots of allocations. In the most extreme case, for a
clean "Check" build of serde it reduces the total number of allocations
by 20%.
Previously, on a type mismatch (and if this wasn't preëmpted by a
higher-priority suggestion), we would look for argumentless methods
returning the expected type, and list them in a `help` note.
This had two major shortcomings. Firstly, a lot of the suggestions didn't
really make sense (if you used a &str where a String was expected,
`.to_ascii_uppercase()` is probably not the solution you were hoping
for). Secondly, we weren't generating suggestions from the most useful
traits!
We address the first problem with an internal
`#[rustc_conversion_suggestion]` attribute meant to mark methods that keep
the "same value" in the relevant sense, just converting the type. We
address the second problem by making `FnCtxt.probe_for_return_type` pass
the `ProbeScope::AllTraits` to `probe_op`: this would seem to be safe
because grep reveals no other callers of `probe_for_return_type`.
Also, structured suggestions are preferred (because they're pretty, but
also for RLS and friends).
Also also, we make the E0055 autoderef recursion limit error use the
one-time-diagnostics set, because we can potentially hit the limit a lot
during probing. (Without this,
test/ui/did_you_mean/recursion_limit_deref.rs would report "aborting due to
51 errors").
Unfortunately, the trait probing is still not all one would hope for: at a
minimum, we don't know how to rule out `into()` in cases where it wouldn't
actually work, and we don't know how to rule in `.to_owned()` where it
would. Issues #46459 and #46460 have been filed and are ref'd in a FIXME.
This is hoped to resolve#42929, #44672, and #45777.
This commit moves the calculation of the `LanguageItems` structure into a
query rather than being calculated before the `TyCtxt` exists, with the eventual
end goal of removing some `CrateStore` methods.
Decompose Adjustment into smaller steps and remove the method map.
The method map held method callee information for:
* actual method calls (`x.f(...)`)
* overloaded unary, binary, indexing and call operators
* *every overloaded deref adjustment* (many can exist for each expression)
That last one was a historical ~~accident~~ hack, and part of the motivation for this PR, along with:
* a desire to compose adjustments more freely
* containing the autoderef logic better to avoid mutation within an inference snapshot
* not creating `TyFnDef` types which are incompatible with the original one
* i.e. we used to take a`TyFnDef`'s `for<'a> &'a T -> &'a U` signature and instantiate `'a` using a region inference variable, *then* package the resulting `&'b T -> &'b U` signature in another `TyFnDef`, while keeping *the same* `DefId` and `Substs`
* to fix#3548 by explicitly writing autorefs for the RHS of comparison operators
Individual commits tell their own story, of "atomic" changes avoiding breaking semantics.
Future work based on this PR could include:
* removing the signature from `TyFnDef`, now that it's always "canonical"
* some questions of variance remain, as subtyping *still* treats the signature differently
* moving part of the typeck logic for methods, autoderef and coercion into `rustc::traits`
* allowing LUB coercions (joining multiple expressions) to "stack up" many adjustments
* transitive coercions (e.g. reify or unsize after multiple steps of autoderef)
r? @nikomatsakis