Rollup merge of #141347 - lcnr:lets-make-it-unsound-3, r=compiler-errors

incorrectly prefer builtin `dyn` impls :3
This makes #57893 slightly more exploitable with the new solver. It's still strictly better than the old solver and the underlying unsoundness persists in the new one even without this preference.
Properly fixing #57893 is something we've been looking at more deeply recently and discussed at the [Types Meetup during the All-Hands](https://hackmd.io/rz-4ghMzTb2wXOkdLKHaHw#Dyn-traits). Whatever approach we'll end up deciding on will likely require a fairly long transition period and some significant further design work. This should not block `-Znext-solver`.
fixesrust-lang/trait-system-refactor-initiative#183
r? `@compiler-errors` cc `@rust-lang/types`

Author: matthiaskrgr

compiler/rustc_next_trait_solver/src/solve/trait_goals.rs

@@ -9,15 +9,15 @@ use rustc_type_ir::{
 selfas ty,Interner,Movability,TraitPredicate,TypeVisitableExtas _,TypingMode,
 Upcastas _, elaborate,
 };
-use tracing::{instrument, trace};
+use tracing::{debug,instrument, trace};
 
 usecrate::delegate::SolverDelegate;
 usecrate::solve::assembly::structural_traits::{self,AsyncCallableRelevantTypes};
 usecrate::solve::assembly::{self,AllowInferenceConstraints,AssembleCandidatesFrom,Candidate};
 usecrate::solve::inspect::ProbeKind;
 usecrate::solve::{
 BuiltinImplSource,CandidateSource,Certainty,EvalCtxt,Goal,GoalSource,MaybeCause,
-NoSolution,ParamEnvSource,QueryResult,
+NoSolution,ParamEnvSource,QueryResult, has_only_region_constraints,
 };
 
 impl<D,I> assembly::GoalKind<D>forTraitPredicate<I>
@@ -1253,6 +1253,45 @@ where
 D:SolverDelegate<Interner = I>,
 I:Interner,
 {
+/// FIXME(#57893): For backwards compatability with the old trait solver implementation,
+/// we need to handle overlap between builtin and user-written impls for trait objects.
+///
+/// This overlap is unsound in general and something which we intend to fix separately.
+/// To avoid blocking the stabilization of the trait solver, we add this hack to avoid
+/// breakage in cases which are *mostly fine*™. Importantly, this preference is strictly
+/// weaker than the old behavior.
+///
+/// We only prefer builtin over user-written impls if there are no inference constraints.
+/// Importantly, we also only prefer the builtin impls for trait goals, and not during
+/// normalization. This means the only case where this special-case results in exploitable
+/// unsoundness should be lifetime dependent user-written impls.
+pub(super)fnunsound_prefer_builtin_dyn_impl(&mutself,candidates:&mutVec<Candidate<I>>){
+matchself.typing_mode(){
+TypingMode::Coherence => return,
+TypingMode::Analysis{ .. }
+ | TypingMode::Borrowck{ .. }
+ | TypingMode::PostBorrowckAnalysis{ .. }
+ | TypingMode::PostAnalysis => {}
+}
+
+if candidates
+.iter()
+.find(|c| {
+matches!(c.source,CandidateSource::BuiltinImpl(BuiltinImplSource::Object(_)))
+})
+.is_some_and(|c| has_only_region_constraints(c.result))
+{
+ candidates.retain(|c| {
+ifmatches!(c.source,CandidateSource::Impl(_)){
+debug!(?c,"unsoundly dropping impl in favor of builtin dyn-candidate");
+false
+}else{
+true
+}
+});
+}
+}
+
 #[instrument(level = "debug", skip(self), ret)]
 pub(super)fnmerge_trait_candidates(
 &mutself,
@@ -1313,6 +1352,7 @@ where
 }
 
 self.filter_specialized_impls(AllowInferenceConstraints::No,&mut candidates);
+self.unsound_prefer_builtin_dyn_impl(&mut candidates);
 
 // If there are *only* global where bounds, then make sure to return that this
 // is still reported as being proven-via the param-env so that rigid projections

tests/ui/traits/next-solver/assembly/better_any-backcompat.rs

@@ -0,0 +1,33 @@
+//@ check-pass
+//@ revisions: current next
+//@[next] compile-flags: -Znext-solver
+//@ ignore-compare-mode-next-solver (explicit revisions)
+
+// A regression test for trait-system-refactor-initiative#183. While
+// this concrete instance is likely not practically unsound, the general
+// pattern is, see #57893.
+
+use std::any::TypeId;
+
+unsafetraitTidAble<'a>:Tid<'a>{}
+traitTidExt<'a>:Tid<'a>{
+fndowncast_box(self:Box<Self>){
+loop{}
+}
+}
+
+impl<'a,X: ?Sized + Tid<'a>>TidExt<'a>forX{}
+
+unsafetraitTid<'a>:'a{}
+
+unsafeimpl<'a,T: ?Sized + TidAble<'a>>Tid<'a>forT{}
+
+impl<'a>dynTid<'a> + 'a{
+fndowncast_any_box(self:Box<Self>){
+self.downcast_box();
+}
+}
+
+unsafeimpl<'a>TidAble<'a>fordynTid<'a> + 'a{}
+
+fnmain(){}