Auto merge of #140043 - ChrisDenton:rollup-vwf0s9j, r=ChrisDenton

Rollup of 8 pull requests

Successful merges:

 - #138934 (support config extensions)
 - #139091 (Rewrite on_unimplemented format string parser.)
 - #139753 (Make `#[naked]` an unsafe attribute)
 - #139762 (Don't assemble non-env/bound candidates if projection is rigid)
 - #139834 (Don't canonicalize crate paths)
 - #139868 (Move `pal::env` to `std::sys::env_consts`)
 - #139978 (Add citool command for generating a test dashboard)
 - #139995 (Clean UI tests 4 of n)

r? `@ghost`
`@rustbot` modify labels: rollup

Author: bors

bootstrap.example.toml

@@ -19,6 +19,14 @@
 # Note that this has no default value (x.py uses the defaults in `bootstrap.example.toml`).
 #profile = <none>
 
+# Inherits configuration values from different configuration files (a.k.a. config extensions).
+# Supports absolute paths, and uses the current directory (where the bootstrap was invoked)
+# as the base if the given path is not absolute.
+#
+# The overriding logic follows a right-to-left order. For example, in `include = ["a.toml", "b.toml"]`,
+# extension `b.toml` overrides `a.toml`. Also, parent extensions always overrides the inner ones.
+#include = []
+
 # Keeps track of major changes made to this configuration.
 #
 # This value also represents ID of the PR that caused major changes. Meaning,

compiler/rustc_builtin_macros/messages.ftl

@@ -247,9 +247,9 @@ builtin_macros_multiple_defaults = multiple declared defaults
     .suggestion = make `{$ident}` default
 
 builtin_macros_naked_functions_testing_attribute =
-    cannot use `#[naked]` with testing attributes
+    cannot use `#[unsafe(naked)]` with testing attributes
     .label = function marked with testing attribute here
-    .naked_attribute = `#[naked]` is incompatible with testing attributes
+    .naked_attribute = `#[unsafe(naked)]` is incompatible with testing attributes
 
 builtin_macros_no_default_variant = `#[derive(Default)]` on enum with no `#[default]`
     .label = this enum needs a unit variant marked with `#[default]`

compiler/rustc_codegen_cranelift/example/mini_core_hello_world.rs

@@ -387,11 +387,9 @@ global_asm! {
 }
 
 #[cfg(all(not(jit), target_arch = "x86_64"))]
-#[naked]
+#[unsafe(naked)]
 extern "C" fn naked_test() {
-    unsafe {
-        naked_asm!("ret");
-    }
+    naked_asm!("ret")
 }
 
 #[repr(C)]

compiler/rustc_error_codes/src/error_codes/E0736.md

@@ -11,7 +11,7 @@ Erroneous code example:
 
 ```compile_fail,E0736
 #[inline]
-#[naked]
+#[unsafe(naked)]
 fn foo() {}
 ```
 

compiler/rustc_error_codes/src/error_codes/E0787.md

@@ -5,7 +5,7 @@ Erroneous code example:
 ```compile_fail,E0787
 #![feature(naked_functions)]
 
-#[naked]
+#[unsafe(naked)]
 pub extern "C" fn f() -> u32 {
     42
 }

compiler/rustc_feature/src/builtin_attrs.rs

@@ -517,7 +517,7 @@ pub static BUILTIN_ATTRIBUTES: &[BuiltinAttribute] = &[
 
     // Linking:
     gated!(
-        naked, Normal, template!(Word), WarnFollowing, EncodeCrossCrate::No,
+        unsafe naked, Normal, template!(Word), WarnFollowing, EncodeCrossCrate::No,
         naked_functions, experimental!(naked)
     ),
 

compiler/rustc_metadata/src/locator.rs

@@ -427,12 +427,21 @@ impl<'a> CrateLocator<'a> {
 
                         let (rlibs, rmetas, dylibs) =
                             candidates.entry(hash.to_string()).or_default();
-                        let path =
-                            try_canonicalize(&spf.path).unwrap_or_else(|_| spf.path.to_path_buf());
-                        if seen_paths.contains(&path) {
-                            continue;
-                        };
-                        seen_paths.insert(path.clone());
+                        {
+                            // As a perforamnce optimisation we canonicalize the path and skip
+                            // ones we've already seeen. This allows us to ignore crates
+                            // we know are exactual equal to ones we've already found.
+                            // Going to the same crate through different symlinks does not change the result.
+                            let path = try_canonicalize(&spf.path)
+                                .unwrap_or_else(|_| spf.path.to_path_buf());
+                            if seen_paths.contains(&path) {
+                                continue;
+                            };
+                            seen_paths.insert(path);
+                        }
+                        // Use the original path (potentially with unresolved symlinks),
+                        // filesystem code should not care, but this is nicer for diagnostics.
+                        let path = spf.path.to_path_buf();
                         match kind {
                             CrateFlavor::Rlib => rlibs.insert(path, search_path.kind),
                             CrateFlavor::Rmeta => rmetas.insert(path, search_path.kind),

compiler/rustc_mir_build/src/check_unsafety.rs

@@ -564,13 +564,17 @@ impl<'a, 'tcx> Visitor<'a, 'tcx> for UnsafetyVisitor<'a, 'tcx> {
                 }
             }
             ExprKind::InlineAsm(box InlineAsmExpr {
-                asm_macro: AsmMacro::Asm | AsmMacro::NakedAsm,
+                asm_macro: asm_macro @ (AsmMacro::Asm | AsmMacro::NakedAsm),
                 ref operands,
                 template: _,
                 options: _,
                 line_spans: _,
             }) => {
-                self.requires_unsafe(expr.span, UseOfInlineAssembly);
+                // The `naked` attribute and the `naked_asm!` block form one atomic unit of
+                // unsafety, and `naked_asm!` does not itself need to be wrapped in an unsafe block.
+                if let AsmMacro::Asm = asm_macro {
+                    self.requires_unsafe(expr.span, UseOfInlineAssembly);
+                }
 
                 // For inline asm, do not use `walk_expr`, since we want to handle the label block
                 // specially.

compiler/rustc_next_trait_solver/src/solve/assembly/mod.rs

@@ -288,6 +288,21 @@ where
     ) -> Vec<Candidate<I>>;
 }
 
+/// Allows callers of `assemble_and_evaluate_candidates` to choose whether to limit
+/// candidate assembly to param-env and alias-bound candidates.
+///
+/// On top of being a micro-optimization, as it avoids doing unnecessary work when
+/// a param-env trait bound candidate shadows impls for normalization, this is also
+/// required to prevent query cycles due to RPITIT inference. See the issue at:
+/// <https://github.com/rust-lang/trait-system-refactor-initiative/issues/173>.
+pub(super) enum AssembleCandidatesFrom {
+    All,
+    /// Only assemble candidates from the environment and alias bounds, ignoring
+    /// user-written and built-in impls. We only expect `ParamEnv` and `AliasBound`
+    /// candidates to be assembled.
+    EnvAndBounds,
+}
+
 impl<D, I> EvalCtxt<'_, D>
 where
     D: SolverDelegate<Interner = I>,
@@ -296,6 +311,7 @@ where
     pub(super) fn assemble_and_evaluate_candidates<G: GoalKind<D>>(
         &mut self,
         goal: Goal<I, G>,
+        assemble_from: AssembleCandidatesFrom,
     ) -> Vec<Candidate<I>> {
         let Ok(normalized_self_ty) =
             self.structurally_normalize_ty(goal.param_env, goal.predicate.self_ty())
@@ -322,16 +338,18 @@ where
             }
         }
 
-        self.assemble_impl_candidates(goal, &mut candidates);
-
-        self.assemble_builtin_impl_candidates(goal, &mut candidates);
-
         self.assemble_alias_bound_candidates(goal, &mut candidates);
-
-        self.assemble_object_bound_candidates(goal, &mut candidates);
-
         self.assemble_param_env_candidates(goal, &mut candidates);
 
+        match assemble_from {
+            AssembleCandidatesFrom::All => {
+                self.assemble_impl_candidates(goal, &mut candidates);
+                self.assemble_builtin_impl_candidates(goal, &mut candidates);
+                self.assemble_object_bound_candidates(goal, &mut candidates);
+            }
+            AssembleCandidatesFrom::EnvAndBounds => {}
+        }
+
         candidates
     }
 
@@ -754,6 +772,9 @@ where
         })
     }
 
+    /// Assemble and merge candidates for goals which are related to an underlying trait
+    /// goal. Right now, this is normalizes-to and host effect goals.
+    ///
     /// We sadly can't simply take all possible candidates for normalization goals
     /// and check whether they result in the same constraints. We want to make sure
     /// that trying to normalize an alias doesn't result in constraints which aren't
@@ -782,47 +803,44 @@ where
     ///
     /// See trait-system-refactor-initiative#124 for more details.
     #[instrument(level = "debug", skip(self, inject_normalize_to_rigid_candidate), ret)]
-    pub(super) fn merge_candidates(
+    pub(super) fn assemble_and_merge_candidates<G: GoalKind<D>>(
         &mut self,
         proven_via: Option<TraitGoalProvenVia>,
-        candidates: Vec<Candidate<I>>,
+        goal: Goal<I, G>,
         inject_normalize_to_rigid_candidate: impl FnOnce(&mut EvalCtxt<'_, D>) -> QueryResult<I>,
     ) -> QueryResult<I> {
         let Some(proven_via) = proven_via else {
             // We don't care about overflow. If proving the trait goal overflowed, then
             // it's enough to report an overflow error for that, we don't also have to
             // overflow during normalization.
-            return Ok(self.make_ambiguous_response_no_constraints(MaybeCause::Ambiguity));
+            //
+            // We use `forced_ambiguity` here over `make_ambiguous_response_no_constraints`
+            // because the former will also record a built-in candidate in the inspector.
+            return self.forced_ambiguity(MaybeCause::Ambiguity).map(|cand| cand.result);
         };
 
         match proven_via {
             TraitGoalProvenVia::ParamEnv | TraitGoalProvenVia::AliasBound => {
-                let mut considered_candidates = Vec::new();
-                considered_candidates.extend(
-                    candidates
-                        .iter()
-                        .filter(|c| matches!(c.source, CandidateSource::ParamEnv(_)))
-                        .map(|c| c.result),
-                );
-
                 // Even when a trait bound has been proven using a where-bound, we
                 // still need to consider alias-bounds for normalization, see
-                // tests/ui/next-solver/alias-bound-shadowed-by-env.rs.
-                //
+                // `tests/ui/next-solver/alias-bound-shadowed-by-env.rs`.
+                let candidates_from_env_and_bounds: Vec<_> = self
+                    .assemble_and_evaluate_candidates(goal, AssembleCandidatesFrom::EnvAndBounds);
+
                 // We still need to prefer where-bounds over alias-bounds however.
-                // See tests/ui/winnowing/norm-where-bound-gt-alias-bound.rs.
-                //
-                // FIXME(const_trait_impl): should this behavior also be used by
-                // constness checking. Doing so is *at least theoretically* breaking,
-                // see github.com/rust-lang/rust/issues/133044#issuecomment-2500709754
-                if considered_candidates.is_empty() {
-                    considered_candidates.extend(
-                        candidates
-                            .iter()
-                            .filter(|c| matches!(c.source, CandidateSource::AliasBound))
-                            .map(|c| c.result),
-                    );
-                }
+                // See `tests/ui/winnowing/norm-where-bound-gt-alias-bound.rs`.
+                let mut considered_candidates: Vec<_> = if candidates_from_env_and_bounds
+                    .iter()
+                    .any(|c| matches!(c.source, CandidateSource::ParamEnv(_)))
+                {
+                    candidates_from_env_and_bounds
+                        .into_iter()
+                        .filter(|c| matches!(c.source, CandidateSource::ParamEnv(_)))
+                        .map(|c| c.result)
+                        .collect()
+                } else {
+                    candidates_from_env_and_bounds.into_iter().map(|c| c.result).collect()
+                };
 
                 // If the trait goal has been proven by using the environment, we want to treat
                 // aliases as rigid if there are no applicable projection bounds in the environment.
@@ -839,6 +857,9 @@ where
                 }
             }
             TraitGoalProvenVia::Misc => {
+                let candidates =
+                    self.assemble_and_evaluate_candidates(goal, AssembleCandidatesFrom::All);
+
                 // Prefer "orphaned" param-env normalization predicates, which are used
                 // (for example, and ideally only) when proving item bounds for an impl.
                 let candidates_from_env: Vec<_> = candidates

compiler/rustc_next_trait_solver/src/solve/effect_goals.rs

@@ -399,12 +399,11 @@ where
         &mut self,
         goal: Goal<I, ty::HostEffectPredicate<I>>,
     ) -> QueryResult<I> {
-        let candidates = self.assemble_and_evaluate_candidates(goal);
         let (_, proven_via) = self.probe(|_| ProbeKind::ShadowedEnvProbing).enter(|ecx| {
             let trait_goal: Goal<I, ty::TraitPredicate<I>> =
                 goal.with(ecx.cx(), goal.predicate.trait_ref);
             ecx.compute_trait_goal(trait_goal)
         })?;
-        self.merge_candidates(proven_via, candidates, |_ecx| Err(NoSolution))
+        self.assemble_and_merge_candidates(proven_via, goal, |_ecx| Err(NoSolution))
     }
 }