Auto merge of #128002 - matthiaskrgr:rollup-21p0cue, r=matthiaskrgr

Rollup of 6 pull requests

Successful merges:

 - #127463 ( use precompiled rustdoc with CI rustc)
 - #127779 (Add a hook for `should_codegen_locally`)
 - #127843 (unix: document unsafety for std `sig{action,altstack}`)
 - #127873 (kmc-solid: `#![forbid(unsafe_op_in_unsafe_fn)]`)
 - #127917 (match lowering: Split `finalize_or_candidate` into more coherent methods)
 - #127964 (run_make_support: skip rustfmt for lib.rs)

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

compiler/rustc_codegen_cranelift/src/abi/mod.rs

@@ -10,12 +10,12 @@ use std::mem;
 use cranelift_codegen::ir::{ArgumentPurpose, SigRef};
 use cranelift_codegen::isa::CallConv;
 use cranelift_module::ModuleError;
+use rustc_codegen_ssa::base::is_call_from_compiler_builtins_to_upstream_monomorphization;
 use rustc_codegen_ssa::errors::CompilerBuiltinsCannotCall;
 use rustc_middle::middle::codegen_fn_attrs::CodegenFnAttrFlags;
 use rustc_middle::ty::layout::FnAbiOf;
 use rustc_middle::ty::print::with_no_trimmed_paths;
 use rustc_middle::ty::TypeVisitableExt;
-use rustc_monomorphize::is_call_from_compiler_builtins_to_upstream_monomorphization;
 use rustc_session::Session;
 use rustc_span::source_map::Spanned;
 use rustc_target::abi::call::{Conv, FnAbi, PassMode};

compiler/rustc_codegen_cranelift/src/base.rs

@@ -5,13 +5,13 @@ use cranelift_codegen::CodegenError;
 use cranelift_frontend::{FunctionBuilder, FunctionBuilderContext};
 use cranelift_module::ModuleError;
 use rustc_ast::InlineAsmOptions;
+use rustc_codegen_ssa::base::is_call_from_compiler_builtins_to_upstream_monomorphization;
 use rustc_index::IndexVec;
 use rustc_middle::middle::codegen_fn_attrs::CodegenFnAttrFlags;
 use rustc_middle::ty::adjustment::PointerCoercion;
 use rustc_middle::ty::layout::FnAbiOf;
 use rustc_middle::ty::print::with_no_trimmed_paths;
 use rustc_middle::ty::TypeVisitableExt;
-use rustc_monomorphize::is_call_from_compiler_builtins_to_upstream_monomorphization;
 
 use crate::constant::ConstantCx;
 use crate::debuginfo::{FunctionDebugContext, TypeDebugContext};

compiler/rustc_codegen_cranelift/src/lib.rs

@@ -24,7 +24,6 @@ extern crate rustc_hir;
 extern crate rustc_incremental;
 extern crate rustc_index;
 extern crate rustc_metadata;
-extern crate rustc_monomorphize;
 extern crate rustc_session;
 extern crate rustc_span;
 extern crate rustc_target;

compiler/rustc_codegen_ssa/src/base.rs

@@ -806,6 +806,34 @@ pub fn codegen_crate<B: ExtraBackendMethods>(
     ongoing_codegen
 }
 
+/// Returns whether a call from the current crate to the [`Instance`] would produce a call
+/// from `compiler_builtins` to a symbol the linker must resolve.
+///
+/// Such calls from `compiler_bultins` are effectively impossible for the linker to handle. Some
+/// linkers will optimize such that dead calls to unresolved symbols are not an error, but this is
+/// not guaranteed. So we used this function in codegen backends to ensure we do not generate any
+/// unlinkable calls.
+///
+/// Note that calls to LLVM intrinsics are uniquely okay because they won't make it to the linker.
+pub fn is_call_from_compiler_builtins_to_upstream_monomorphization<'tcx>(
+    tcx: TyCtxt<'tcx>,
+    instance: Instance<'tcx>,
+) -> bool {
+    fn is_llvm_intrinsic(tcx: TyCtxt<'_>, def_id: DefId) -> bool {
+        if let Some(name) = tcx.codegen_fn_attrs(def_id).link_name {
+            name.as_str().starts_with("llvm.")
+        } else {
+            false
+        }
+    }
+
+    let def_id = instance.def_id();
+    !def_id.is_local()
+        && tcx.is_compiler_builtins(LOCAL_CRATE)
+        && !is_llvm_intrinsic(tcx, def_id)
+        && !tcx.should_codegen_locally(instance)
+}
+
 impl CrateInfo {
     pub fn new(tcx: TyCtxt<'_>, target_cpu: String) -> CrateInfo {
         let crate_types = tcx.crate_types().to_vec();

compiler/rustc_codegen_ssa/src/mir/block.rs

@@ -3,7 +3,7 @@ use super::operand::OperandValue::{Immediate, Pair, Ref, ZeroSized};
 use super::place::{PlaceRef, PlaceValue};
 use super::{CachedLlbb, FunctionCx, LocalRef};
 
-use crate::base;
+use crate::base::{self, is_call_from_compiler_builtins_to_upstream_monomorphization};
 use crate::common::{self, IntPredicate};
 use crate::errors::CompilerBuiltinsCannotCall;
 use crate::meth;
@@ -18,7 +18,6 @@ use rustc_middle::ty::layout::{HasTyCtxt, LayoutOf, ValidityRequirement};
 use rustc_middle::ty::print::{with_no_trimmed_paths, with_no_visible_paths};
 use rustc_middle::ty::{self, Instance, Ty};
 use rustc_middle::{bug, span_bug};
-use rustc_monomorphize::is_call_from_compiler_builtins_to_upstream_monomorphization;
 use rustc_session::config::OptLevel;
 use rustc_span::{source_map::Spanned, sym, Span};
 use rustc_target::abi::call::{ArgAbi, FnAbi, PassMode, Reg};

compiler/rustc_middle/src/hooks/mod.rs

@@ -103,6 +103,10 @@ declare_hooks! {
 
     /// Create a list-like THIR representation for debugging.
     hook thir_flat(key: LocalDefId) -> String;
+
+    /// Returns `true` if we should codegen an instance in the local crate, or returns `false` if we
+    /// can just link to the upstream crate and therefore don't need a mono item.
+    hook should_codegen_locally(instance: crate::ty::Instance<'tcx>) -> bool;
 }
 
 #[cold]

compiler/rustc_mir_build/src/build/matches/mod.rs

@@ -1598,6 +1598,9 @@ impl<'a, 'tcx> Builder<'a, 'tcx> {
                 for subcandidate in candidate.subcandidates.iter_mut() {
                     expanded_candidates.push(subcandidate);
                 }
+                // Note that the subcandidates have been added to `expanded_candidates`,
+                // but `candidate` itself has not. If the last candidate has more match pairs,
+                // they are handled separately by `test_remaining_match_pairs_after_or`.
             } else {
                 // A candidate that doesn't start with an or-pattern has nothing to
                 // expand, so it is included in the post-expansion list as-is.
@@ -1613,19 +1616,28 @@ impl<'a, 'tcx> Builder<'a, 'tcx> {
             expanded_candidates.as_mut_slice(),
         );
 
-        // Simplify subcandidates and process any leftover match pairs.
-        for candidate in candidates_to_expand {
+        // Postprocess subcandidates, and process any leftover match pairs.
+        // (Only the last candidate can possibly have more match pairs.)
+        debug_assert!({
+            let mut all_except_last = candidates_to_expand.iter().rev().skip(1);
+            all_except_last.all(|candidate| candidate.match_pairs.is_empty())
+        });
+        for candidate in candidates_to_expand.iter_mut() {
             if !candidate.subcandidates.is_empty() {
-                self.finalize_or_candidate(span, scrutinee_span, candidate);
+                self.merge_trivial_subcandidates(candidate);
+                self.remove_never_subcandidates(candidate);
             }
         }
+        if let Some(last_candidate) = candidates_to_expand.last_mut() {
+            self.test_remaining_match_pairs_after_or(span, scrutinee_span, last_candidate);
+        }
 
         remainder_start.and(remaining_candidates)
     }
 
     /// Given a match-pair that corresponds to an or-pattern, expand each subpattern into a new
-    /// subcandidate. Any candidate that has been expanded that way should be passed to
-    /// `finalize_or_candidate` after its subcandidates have been processed.
+    /// subcandidate. Any candidate that has been expanded this way should also be postprocessed
+    /// at the end of [`Self::expand_and_match_or_candidates`].
     fn create_or_subcandidates<'pat>(
         &mut self,
         candidate: &mut Candidate<'pat, 'tcx>,
@@ -1642,7 +1654,8 @@ impl<'a, 'tcx> Builder<'a, 'tcx> {
         candidate.subcandidates[0].false_edge_start_block = candidate.false_edge_start_block;
     }
 
-    /// Simplify subcandidates and process any leftover match pairs. The candidate should have been
+    /// Try to merge all of the subcandidates of the given candidate into one. This avoids
+    /// exponentially large CFGs in cases like `(1 | 2, 3 | 4, ...)`. The candidate should have been
     /// expanded with `create_or_subcandidates`.
     ///
     /// Given a pattern `(P | Q, R | S)` we (in principle) generate a CFG like
@@ -1695,103 +1708,128 @@ impl<'a, 'tcx> Builder<'a, 'tcx> {
     ///      |
     ///     ...
     /// ```
-    fn finalize_or_candidate(
-        &mut self,
-        span: Span,
-        scrutinee_span: Span,
-        candidate: &mut Candidate<'_, 'tcx>,
-    ) {
-        if candidate.subcandidates.is_empty() {
+    ///
+    /// Note that this takes place _after_ the subcandidates have participated
+    /// in match tree lowering.
+    fn merge_trivial_subcandidates(&mut self, candidate: &mut Candidate<'_, 'tcx>) {
+        assert!(!candidate.subcandidates.is_empty());
+        if candidate.has_guard {
+            // FIXME(or_patterns; matthewjasper) Don't give up if we have a guard.
             return;
         }
 
-        self.merge_trivial_subcandidates(candidate);
+        // FIXME(or_patterns; matthewjasper) Try to be more aggressive here.
+        let can_merge = candidate.subcandidates.iter().all(|subcandidate| {
+            subcandidate.subcandidates.is_empty() && subcandidate.extra_data.is_empty()
+        });
+        if !can_merge {
+            return;
+        }
 
-        if !candidate.match_pairs.is_empty() {
-            let or_span = candidate.or_span.unwrap_or(candidate.extra_data.span);
-            let source_info = self.source_info(or_span);
-            // If more match pairs remain, test them after each subcandidate.
-            // We could add them to the or-candidates before the call to `test_or_pattern` but this
-            // would make it impossible to detect simplifiable or-patterns. That would guarantee
-            // exponentially large CFGs for cases like `(1 | 2, 3 | 4, ...)`.
-            let mut last_otherwise = None;
-            candidate.visit_leaves(|leaf_candidate| {
-                last_otherwise = leaf_candidate.otherwise_block;
-            });
-            let remaining_match_pairs = mem::take(&mut candidate.match_pairs);
-            candidate.visit_leaves(|leaf_candidate| {
-                assert!(leaf_candidate.match_pairs.is_empty());
-                leaf_candidate.match_pairs.extend(remaining_match_pairs.iter().cloned());
-                let or_start = leaf_candidate.pre_binding_block.unwrap();
-                let otherwise =
-                    self.match_candidates(span, scrutinee_span, or_start, &mut [leaf_candidate]);
-                // In a case like `(P | Q, R | S)`, if `P` succeeds and `R | S` fails, we know `(Q,
-                // R | S)` will fail too. If there is no guard, we skip testing of `Q` by branching
-                // directly to `last_otherwise`. If there is a guard,
-                // `leaf_candidate.otherwise_block` can be reached by guard failure as well, so we
-                // can't skip `Q`.
-                let or_otherwise = if leaf_candidate.has_guard {
-                    leaf_candidate.otherwise_block.unwrap()
-                } else {
-                    last_otherwise.unwrap()
-                };
-                self.cfg.goto(otherwise, source_info, or_otherwise);
-            });
+        let mut last_otherwise = None;
+        let shared_pre_binding_block = self.cfg.start_new_block();
+        // This candidate is about to become a leaf, so unset `or_span`.
+        let or_span = candidate.or_span.take().unwrap();
+        let source_info = self.source_info(or_span);
+
+        if candidate.false_edge_start_block.is_none() {
+            candidate.false_edge_start_block = candidate.subcandidates[0].false_edge_start_block;
+        }
+
+        // Remove the (known-trivial) subcandidates from the candidate tree,
+        // so that they aren't visible after match tree lowering, and wire them
+        // all to join up at a single shared pre-binding block.
+        // (Note that the subcandidates have already had their part of the match
+        // tree lowered by this point, which is why we can add a goto to them.)
+        for subcandidate in mem::take(&mut candidate.subcandidates) {
+            let subcandidate_block = subcandidate.pre_binding_block.unwrap();
+            self.cfg.goto(subcandidate_block, source_info, shared_pre_binding_block);
+            last_otherwise = subcandidate.otherwise_block;
         }
+        candidate.pre_binding_block = Some(shared_pre_binding_block);
+        assert!(last_otherwise.is_some());
+        candidate.otherwise_block = last_otherwise;
     }
 
-    /// Try to merge all of the subcandidates of the given candidate into one. This avoids
-    /// exponentially large CFGs in cases like `(1 | 2, 3 | 4, ...)`. The candidate should have been
-    /// expanded with `create_or_subcandidates`.
-    fn merge_trivial_subcandidates(&mut self, candidate: &mut Candidate<'_, 'tcx>) {
-        if candidate.subcandidates.is_empty() || candidate.has_guard {
-            // FIXME(or_patterns; matthewjasper) Don't give up if we have a guard.
+    /// Never subcandidates may have a set of bindings inconsistent with their siblings,
+    /// which would break later code. So we filter them out. Note that we can't filter out
+    /// top-level candidates this way.
+    fn remove_never_subcandidates(&mut self, candidate: &mut Candidate<'_, 'tcx>) {
+        if candidate.subcandidates.is_empty() {
             return;
         }
 
-        // FIXME(or_patterns; matthewjasper) Try to be more aggressive here.
-        let can_merge = candidate.subcandidates.iter().all(|subcandidate| {
-            subcandidate.subcandidates.is_empty() && subcandidate.extra_data.is_empty()
-        });
-        if can_merge {
-            let mut last_otherwise = None;
-            let any_matches = self.cfg.start_new_block();
-            let or_span = candidate.or_span.take().unwrap();
-            let source_info = self.source_info(or_span);
-            if candidate.false_edge_start_block.is_none() {
-                candidate.false_edge_start_block =
-                    candidate.subcandidates[0].false_edge_start_block;
-            }
-            for subcandidate in mem::take(&mut candidate.subcandidates) {
-                let or_block = subcandidate.pre_binding_block.unwrap();
-                self.cfg.goto(or_block, source_info, any_matches);
-                last_otherwise = subcandidate.otherwise_block;
-            }
-            candidate.pre_binding_block = Some(any_matches);
-            assert!(last_otherwise.is_some());
-            candidate.otherwise_block = last_otherwise;
-        } else {
-            // Never subcandidates may have a set of bindings inconsistent with their siblings,
-            // which would break later code. So we filter them out. Note that we can't filter out
-            // top-level candidates this way.
-            candidate.subcandidates.retain_mut(|candidate| {
-                if candidate.extra_data.is_never {
-                    candidate.visit_leaves(|subcandidate| {
-                        let block = subcandidate.pre_binding_block.unwrap();
-                        // That block is already unreachable but needs a terminator to make the MIR well-formed.
-                        let source_info = self.source_info(subcandidate.extra_data.span);
-                        self.cfg.terminate(block, source_info, TerminatorKind::Unreachable);
-                    });
-                    false
-                } else {
-                    true
-                }
-            });
-            if candidate.subcandidates.is_empty() {
-                // If `candidate` has become a leaf candidate, ensure it has a `pre_binding_block`.
-                candidate.pre_binding_block = Some(self.cfg.start_new_block());
+        candidate.subcandidates.retain_mut(|candidate| {
+            if candidate.extra_data.is_never {
+                candidate.visit_leaves(|subcandidate| {
+                    let block = subcandidate.pre_binding_block.unwrap();
+                    // That block is already unreachable but needs a terminator to make the MIR well-formed.
+                    let source_info = self.source_info(subcandidate.extra_data.span);
+                    self.cfg.terminate(block, source_info, TerminatorKind::Unreachable);
+                });
+                false
+            } else {
+                true
             }
+        });
+        if candidate.subcandidates.is_empty() {
+            // If `candidate` has become a leaf candidate, ensure it has a `pre_binding_block`.
+            candidate.pre_binding_block = Some(self.cfg.start_new_block());
+        }
+    }
+
+    /// If more match pairs remain, test them after each subcandidate.
+    /// We could have added them to the or-candidates during or-pattern expansion, but that
+    /// would make it impossible to detect simplifiable or-patterns. That would guarantee
+    /// exponentially large CFGs for cases like `(1 | 2, 3 | 4, ...)`.
+    fn test_remaining_match_pairs_after_or(
+        &mut self,
+        span: Span,
+        scrutinee_span: Span,
+        candidate: &mut Candidate<'_, 'tcx>,
+    ) {
+        if candidate.match_pairs.is_empty() {
+            return;
         }
+
+        let or_span = candidate.or_span.unwrap_or(candidate.extra_data.span);
+        let source_info = self.source_info(or_span);
+        let mut last_otherwise = None;
+        candidate.visit_leaves(|leaf_candidate| {
+            last_otherwise = leaf_candidate.otherwise_block;
+        });
+
+        let remaining_match_pairs = mem::take(&mut candidate.match_pairs);
+        // We're testing match pairs that remained after an `Or`, so the remaining
+        // pairs should all be `Or` too, due to the sorting invariant.
+        debug_assert!(
+            remaining_match_pairs
+                .iter()
+                .all(|match_pair| matches!(match_pair.test_case, TestCase::Or { .. }))
+        );
+
+        candidate.visit_leaves(|leaf_candidate| {
+            // At this point the leaf's own match pairs have all been lowered
+            // and removed, so `extend` and assignment are equivalent,
+            // but extending can also recycle any existing vector capacity.
+            assert!(leaf_candidate.match_pairs.is_empty());
+            leaf_candidate.match_pairs.extend(remaining_match_pairs.iter().cloned());
+
+            let or_start = leaf_candidate.pre_binding_block.unwrap();
+            let otherwise =
+                self.match_candidates(span, scrutinee_span, or_start, &mut [leaf_candidate]);
+            // In a case like `(P | Q, R | S)`, if `P` succeeds and `R | S` fails, we know `(Q,
+            // R | S)` will fail too. If there is no guard, we skip testing of `Q` by branching
+            // directly to `last_otherwise`. If there is a guard,
+            // `leaf_candidate.otherwise_block` can be reached by guard failure as well, so we
+            // can't skip `Q`.
+            let or_otherwise = if leaf_candidate.has_guard {
+                leaf_candidate.otherwise_block.unwrap()
+            } else {
+                last_otherwise.unwrap()
+            };
+            self.cfg.goto(otherwise, source_info, or_otherwise);
+        });
     }
 
     /// Pick a test to run. Which test doesn't matter as long as it is guaranteed to fully match at