Extend the enum check to pointer and union reads

This change extends the previously added enum discriminant check to
enums read through a union or pointer. At the moment we only insert
the check when transmuting to an enum. Although I hoped for it, this
check isn't yet inserted for calls to `MaybeUninit::assume_init`,
because the pass is running on polymorphic MIR and thus doesn't have
the information yet to know whether the type that is read is an
enum.

Author: 1c3t3a

compiler/rustc_mir_transform/src/check_enums.rs

@@ -2,10 +2,10 @@ use rustc_abi::{Scalar, Size, TagEncoding, Variants, WrappingRange};
 use rustc_hir::LangItem;
 use rustc_index::IndexVec;
 use rustc_middle::bug;
-use rustc_middle::mir::visit::Visitor;
+use rustc_middle::mir::visit::{NonMutatingUseContext, PlaceContext, Visitor};
 use rustc_middle::mir::*;
 use rustc_middle::ty::layout::PrimitiveExt;
-use rustc_middle::ty::{self, Ty, TyCtxt, TypingEnv};
+use rustc_middle::ty::{self, AdtDef, Ty, TyCtxt, TypingEnv};
 use rustc_session::Session;
 use tracing::debug;
 
@@ -148,79 +148,197 @@ impl<'a, 'tcx> EnumFinder<'a, 'tcx> {
     fn into_found_enums(self) -> Vec<EnumCheckType<'tcx>> {
         self.enums
     }
+
+    /// Registers a new enum check in the finder.
+    fn register_new_check(
+        &mut self,
+        enum_ty: Ty<'tcx>,
+        enum_def: AdtDef<'tcx>,
+        source_op: Operand<'tcx>,
+    ) {
+        let Ok(enum_layout) = self.tcx.layout_of(self.typing_env.as_query_input(enum_ty)) else {
+            return;
+        };
+        // If the operand is a pointer, we want to pass on the size of the operand to the check,
+        // as we will dereference the pointer and look at the value directly.
+        let Ok(op_layout) = (if let ty::RawPtr(pointee_ty, _) =
+            source_op.ty(self.local_decls, self.tcx).kind()
+        {
+            self.tcx.layout_of(self.typing_env.as_query_input(*pointee_ty))
+        } else {
+            self.tcx
+                .layout_of(self.typing_env.as_query_input(source_op.ty(self.local_decls, self.tcx)))
+        }) else {
+            return;
+        };
+
+        match enum_layout.variants {
+            Variants::Empty if op_layout.is_uninhabited() => return,
+            // An empty enum that tries to be constructed from an inhabited value, this
+            // is never correct.
+            Variants::Empty => {
+                // The enum layout is uninhabited but we construct it from sth inhabited.
+                // This is always UB.
+                self.enums.push(EnumCheckType::Uninhabited);
+            }
+            // Construction of Single value enums is always fine.
+            Variants::Single { .. } => {}
+            // Construction of an enum with multiple variants but no niche optimizations.
+            Variants::Multiple {
+                tag_encoding: TagEncoding::Direct,
+                tag: Scalar::Initialized { value, .. },
+                ..
+            } => {
+                let valid_discrs =
+                    enum_def.discriminants(self.tcx).map(|(_, discr)| discr.val).collect();
+
+                let discr =
+                    TyAndSize { ty: value.to_int_ty(self.tcx), size: value.size(&self.tcx) };
+                self.enums.push(EnumCheckType::Direct {
+                    source_op: source_op.to_copy(),
+                    discr,
+                    op_size: op_layout.size,
+                    valid_discrs,
+                });
+            }
+            // Construction of an enum with multiple variants and niche optimizations.
+            Variants::Multiple {
+                tag_encoding: TagEncoding::Niche { .. },
+                tag: Scalar::Initialized { value, valid_range, .. },
+                tag_field,
+                ..
+            } => {
+                let discr =
+                    TyAndSize { ty: value.to_int_ty(self.tcx), size: value.size(&self.tcx) };
+                self.enums.push(EnumCheckType::WithNiche {
+                    source_op: source_op.to_copy(),
+                    discr,
+                    op_size: op_layout.size,
+                    offset: enum_layout.fields.offset(tag_field.as_usize()),
+                    valid_range,
+                });
+            }
+            _ => return,
+        }
+    }
 }
 
 impl<'a, 'tcx> Visitor<'tcx> for EnumFinder<'a, 'tcx> {
-    fn visit_rvalue(&mut self, rvalue: &Rvalue<'tcx>, ___location: Location) {
-        if let Rvalue::Cast(CastKind::Transmute, op, ty) = rvalue {
-            let ty::Adt(adt_def, _) = ty.kind() else {
-                return;
-            };
-            if !adt_def.is_enum() {
-                return;
-            }
+    fn visit_place(&mut self, place: &Place<'tcx>, context: PlaceContext, ___location: Location) {
+        self.super_place(place, context, ___location);
+        // We only want to emit this check on pointer reads.
+        match context {
+            PlaceContext::NonMutatingUse(
+                NonMutatingUseContext::Copy
+                | NonMutatingUseContext::Move
+                | NonMutatingUseContext::SharedBorrow,
+            ) => (),
+            _ => return,
+        }
+        // Get the place and type we visit.
+        let pointer = Place::from(place.local);
+        let pointer_ty = pointer.ty(self.local_decls, self.tcx).ty;
 
-            let Ok(enum_layout) = self.tcx.layout_of(self.typing_env.as_query_input(*ty)) else {
+        // We only want to check places based on raw pointers to enums or ManuallyDrop<Enum>.
+        let &ty::RawPtr(pointee_ty, _) = pointer_ty.kind() else {
+            return;
+        };
+        let ty::Adt(enum_adt_def, _) = pointee_ty.kind() else {
+            return;
+        };
+
+        let (enum_ty, enum_adt_def) = if enum_adt_def.is_enum() {
+            (pointee_ty, enum_adt_def)
+        } else if enum_adt_def.is_manually_drop() {
+            // Find the type contained in the ManuallyDrop and check whether it is an enum.
+            let Some((manual_drop_arg, adt_def)) =
+                pointee_ty.walk().skip(1).next().map_or(None, |arg| {
+                    if let Some(ty) = arg.as_type()
+                        && let ty::Adt(adt_def, _) = ty.kind()
+                    {
+                        Some((ty, adt_def))
+                    } else {
+                        None
+                    }
+                })
+            else {
                 return;
             };
-            let Ok(op_layout) = self
-                .tcx
-                .layout_of(self.typing_env.as_query_input(op.ty(self.local_decls, self.tcx)))
+
+            (manual_drop_arg, adt_def)
+        } else {
+            return;
+        };
+        // Exclude c_void.
+        if enum_ty.is_c_void(self.tcx) {
+            return;
+        }
+        self.register_new_check(enum_ty, *enum_adt_def, Operand::Copy(*place));
+    }
+
+    fn visit_projection_elem(
+        &mut self,
+        place_ref: PlaceRef<'tcx>,
+        elem: PlaceElem<'tcx>,
+        context: visit::PlaceContext,
+        ___location: Location,
+    ) {
+        self.super_projection_elem(place_ref, elem, context, ___location);
+        // Check whether we are reading an enum or a ManuallyDrop<Enum> from a union.
+        let ty::Adt(union_adt_def, _) = place_ref.ty(self.local_decls, self.tcx).ty.kind() else {
+            return;
+        };
+        if !union_adt_def.is_union() {
+            return;
+        }
+        let PlaceElem::Field(_, extracted_ty) = elem else {
+            return;
+        };
+        let ty::Adt(enum_adt_def, _) = extracted_ty.kind() else {
+            return;
+        };
+        let (enum_ty, enum_adt_def) = if enum_adt_def.is_enum() {
+            (extracted_ty, enum_adt_def)
+        } else if enum_adt_def.is_manually_drop() {
+            // Find the type contained in the ManuallyDrop and check whether it is an enum.
+            let Some((manual_drop_arg, adt_def)) =
+                extracted_ty.walk().skip(1).next().map_or(None, |arg| {
+                    if let Some(ty) = arg.as_type()
+                        && let ty::Adt(adt_def, _) = ty.kind()
+                    {
+                        Some((ty, adt_def))
+                    } else {
+                        None
+                    }
+                })
             else {
                 return;
             };
 
-            match enum_layout.variants {
-                Variants::Empty if op_layout.is_uninhabited() => return,
-                // An empty enum that tries to be constructed from an inhabited value, this
-                // is never correct.
-                Variants::Empty => {
-                    // The enum layout is uninhabited but we construct it from sth inhabited.
-                    // This is always UB.
-                    self.enums.push(EnumCheckType::Uninhabited);
-                }
-                // Construction of Single value enums is always fine.
-                Variants::Single { .. } => {}
-                // Construction of an enum with multiple variants but no niche optimizations.
-                Variants::Multiple {
-                    tag_encoding: TagEncoding::Direct,
-                    tag: Scalar::Initialized { value, .. },
-                    ..
-                } => {
-                    let valid_discrs =
-                        adt_def.discriminants(self.tcx).map(|(_, discr)| discr.val).collect();
-
-                    let discr =
-                        TyAndSize { ty: value.to_int_ty(self.tcx), size: value.size(&self.tcx) };
-                    self.enums.push(EnumCheckType::Direct {
-                        source_op: op.to_copy(),
-                        discr,
-                        op_size: op_layout.size,
-                        valid_discrs,
-                    });
-                }
-                // Construction of an enum with multiple variants and niche optimizations.
-                Variants::Multiple {
-                    tag_encoding: TagEncoding::Niche { .. },
-                    tag: Scalar::Initialized { value, valid_range, .. },
-                    tag_field,
-                    ..
-                } => {
-                    let discr =
-                        TyAndSize { ty: value.to_int_ty(self.tcx), size: value.size(&self.tcx) };
-                    self.enums.push(EnumCheckType::WithNiche {
-                        source_op: op.to_copy(),
-                        discr,
-                        op_size: op_layout.size,
-                        offset: enum_layout.fields.offset(tag_field.as_usize()),
-                        valid_range,
-                    });
-                }
-                _ => return,
+            (manual_drop_arg, adt_def)
+        } else {
+            return;
+        };
+
+        self.register_new_check(
+            enum_ty,
+            *enum_adt_def,
+            Operand::Copy(place_ref.to_place(self.tcx)),
+        );
+    }
+
+    fn visit_rvalue(&mut self, rvalue: &Rvalue<'tcx>, ___location: Location) {
+        if let Rvalue::Cast(CastKind::Transmute, op, ty) = rvalue {
+            let ty::Adt(adt_def, _) = ty.kind() else {
+                return;
+            };
+            if !adt_def.is_enum() {
+                return;
             }
 
-            self.super_rvalue(rvalue, ___location);
+            self.register_new_check(*ty, *adt_def, op.to_copy());
         }
+        self.super_rvalue(rvalue, ___location);
     }
 }
 
@@ -262,6 +380,22 @@ fn insert_discr_cast_to_u128<'tcx>(
         }
     };
 
+    // If the enum is behind a pointer, cast it to a *[const|mut] MaybeUninit<T> and then extract the discriminant through that.
+    let source_op = if let ty::RawPtr(pointee_ty, mutbl) = source_op.ty(local_decls, tcx).kind() {
+        let mu_ptr_ty = Ty::new_ptr(tcx, Ty::new_maybe_uninit(tcx, *pointee_ty), *mutbl);
+        let mu_ptr_decl =
+            local_decls.push(LocalDecl::with_source_info(mu_ptr_ty, source_info)).into();
+        let rvalue = Rvalue::Cast(CastKind::Transmute, source_op, mu_ptr_ty);
+        block_data.statements.push(Statement::new(
+            source_info,
+            StatementKind::Assign(Box::new((mu_ptr_decl, rvalue))),
+        ));
+
+        Operand::Copy(mu_ptr_decl.project_deeper(&[ProjectionElem::Deref], tcx))
+    } else {
+        source_op
+    };
+
     let (cast_kind, discr_ty_bits) = if discr.size.bytes() < op_size.bytes() {
         // The discriminant is less wide than the operand, cast the operand into
         // [MaybeUninit; N] and then index into it.

tests/ui/mir/enum/pointer_manually_drop_read_break.rs

@@ -0,0 +1,19 @@
+//@ run-fail
+//@ compile-flags: -C debug-assertions
+//@ error-pattern: trying to construct an enum from an invalid value 0x1
+
+#[allow(dead_code)]
+#[repr(u16)]
+enum Single {
+    A,
+}
+
+fn main() {
+    let illegal_val: u16 = 1;
+    let illegal_val_ptr = &raw const illegal_val;
+    let foo: *const std::mem::ManuallyDrop<Single> =
+        unsafe { std::mem::transmute(illegal_val_ptr) };
+
+    let val: Single = unsafe { foo.cast::<Single>().read() };
+    println!("{}", val as u16);
+}

tests/ui/mir/enum/pointer_manually_drop_read_ok.rs

@@ -0,0 +1,18 @@
+//@ run-pass
+//@ compile-flags: -C debug-assertions
+
+#[allow(dead_code)]
+#[repr(u16)]
+enum Single {
+    A,
+}
+
+fn main() {
+    let illegal_val: u16 = 0;
+    let illegal_val_ptr = &raw const illegal_val;
+    let foo: *const std::mem::ManuallyDrop<Single> =
+        unsafe { std::mem::transmute(illegal_val_ptr) };
+
+    let val: Single = unsafe { foo.cast::<Single>().read() };
+    println!("{}", val as u16);
+}

tests/ui/mir/enum/pointer_read_break.rs

@@ -0,0 +1,18 @@
+//@ run-fail
+//@ compile-flags: -C debug-assertions
+//@ error-pattern: trying to construct an enum from an invalid value 0x1
+
+#[allow(dead_code)]
+#[repr(u16)]
+enum Single {
+    A,
+}
+
+fn main() {
+    let illegal_val: u16 = 1;
+    let illegal_val_ptr = &raw const illegal_val;
+    let foo: *const Single = unsafe { std::mem::transmute(illegal_val_ptr) };
+
+    let val: Single = unsafe { foo.read() };
+    println!("{}", val as u16);
+}

tests/ui/mir/enum/pointer_read_ok.rs

@@ -0,0 +1,17 @@
+//@ run-pass
+//@ compile-flags: -C debug-assertions
+
+#[allow(dead_code)]
+#[repr(u16)]
+enum Single {
+    A,
+}
+
+fn main() {
+    let illegal_val: u16 = 0;
+    let illegal_val_ptr = &raw const illegal_val;
+    let foo: *const Single = unsafe { std::mem::transmute(illegal_val_ptr) };
+
+    let val: Single = unsafe { foo.read() };
+    println!("{}", val as u16);
+}

tests/ui/mir/enum/union_manually_drop_read_break.rs

@@ -0,0 +1,20 @@
+//@ run-fail
+//@ compile-flags: -C debug-assertions
+//@ error-pattern: trying to construct an enum from an invalid value 0x1
+
+#[allow(dead_code)]
+#[repr(u16)]
+enum Single {
+    A,
+}
+
+union Foo {
+    a: std::mem::ManuallyDrop<Single>,
+}
+
+fn main() {
+    let foo = Foo { a: unsafe { std::mem::transmute(1_u16) } };
+
+    let val: Single = unsafe { std::mem::ManuallyDrop::into_inner(foo.a) };
+    println!("{}", val as u16);
+}