Auto merge of #135709 - lqd:bring-back-len, r=compiler-errors

Temporarily bring back `Rvalue::Len`

r? `@compiler-errors` as requested in https://github.com/rust-lang/rust/issues/135671#issuecomment-2599580364

> However, in the mean time, I'd rather we not crunch trying to find and more importantly validate the soundness of a solution 🤔

Agreed. To fix the IMO P-critical #135671 for which we somehow didn't have test coverage, this PR temporarily reverts:
- https://github.com/rust-lang/rust/pull/133734
- its bugfix https://github.com/rust-lang/rust/pull/134371
- https://github.com/rust-lang/rust/pull/134330

cc `@scottmcm`

I added the few samples from that issue as a test, but we can add more in the future, in particular it seems `@steffahn` [will work on that](https://github.com/rust-lang/rust/issues/135671#issuecomment-2599714354).

Fixes #135671. And if we want to land this, it should also be nominated for beta backport.

compiler/rustc_mir_transform/src/dataflow_const_prop.rs

@@ -408,6 +408,18 @@ impl<'a, 'tcx> ConstAnalysis<'a, 'tcx> {
         state: &mut State<FlatSet<Scalar>>,
     ) -> ValueOrPlace<FlatSet<Scalar>> {
         let val = match rvalue {
+            Rvalue::Len(place) => {
+                let place_ty = place.ty(self.local_decls, self.tcx);
+                if let ty::Array(_, len) = place_ty.ty.kind() {
+                    Const::Ty(self.tcx.types.usize, *len)
+                        .try_eval_scalar(self.tcx, self.typing_env)
+                        .map_or(FlatSet::Top, FlatSet::Elem)
+                } else if let [ProjectionElem::Deref] = place.projection[..] {
+                    state.get_len(place.local.into(), &self.map)
+                } else {
+                    FlatSet::Top
+                }
+            }
             Rvalue::Cast(CastKind::IntToInt | CastKind::IntToFloat, operand, ty) => {
                 let Ok(layout) = self.tcx.layout_of(self.typing_env.as_query_input(*ty)) else {
                     return ValueOrPlace::Value(FlatSet::Top);

compiler/rustc_mir_transform/src/dest_prop.rs

@@ -575,6 +575,7 @@ impl WriteInfo {
                     | Rvalue::NullaryOp(_, _)
                     | Rvalue::Ref(_, _, _)
                     | Rvalue::RawPtr(_, _)
+                    | Rvalue::Len(_)
                     | Rvalue::Discriminant(_)
                     | Rvalue::CopyForDeref(_) => {}
                 }

compiler/rustc_mir_transform/src/gvn.rs

@@ -223,6 +223,8 @@ enum Value<'tcx> {
     Projection(VnIndex, ProjectionElem<VnIndex, Ty<'tcx>>),
     /// Discriminant of the given value.
     Discriminant(VnIndex),
+    /// Length of an array or slice.
+    Len(VnIndex),
 
     // Operations.
     NullaryOp(NullOp<'tcx>, Ty<'tcx>),
@@ -511,6 +513,13 @@ impl<'body, 'tcx> VnState<'body, 'tcx> {
                     self.ecx.discriminant_for_variant(base.layout.ty, variant).discard_err()?;
                 discr_value.into()
             }
+            Len(slice) => {
+                let slice = self.evaluated[slice].as_ref()?;
+                let usize_layout = self.ecx.layout_of(self.tcx.types.usize).unwrap();
+                let len = slice.len(&self.ecx).discard_err()?;
+                let imm = ImmTy::from_uint(len, usize_layout);
+                imm.into()
+            }
             NullaryOp(null_op, ty) => {
                 let layout = self.ecx.layout_of(ty).ok()?;
                 if let NullOp::SizeOf | NullOp::AlignOf = null_op
@@ -854,6 +863,7 @@ impl<'body, 'tcx> VnState<'body, 'tcx> {
             }
 
             // Operations.
+            Rvalue::Len(ref mut place) => return self.simplify_len(place, location),
             Rvalue::Cast(ref mut kind, ref mut value, to) => {
                 return self.simplify_cast(kind, value, to, location);
             }
@@ -1474,6 +1484,47 @@ impl<'body, 'tcx> VnState<'body, 'tcx> {
         Some(self.insert(Value::Cast { kind: *kind, value, from, to }))
     }
 
+    fn simplify_len(&mut self, place: &mut Place<'tcx>, location: Location) -> Option<VnIndex> {
+        // Trivial case: we are fetching a statically known length.
+        let place_ty = place.ty(self.local_decls, self.tcx).ty;
+        if let ty::Array(_, len) = place_ty.kind() {
+            return self.insert_constant(Const::from_ty_const(
+                *len,
+                self.tcx.types.usize,
+                self.tcx,
+            ));
+        }
+
+        let mut inner = self.simplify_place_value(place, location)?;
+
+        // The length information is stored in the wide pointer.
+        // Reborrowing copies length information from one pointer to the other.
+        while let Value::Address { place: borrowed, .. } = self.get(inner)
+            && let [PlaceElem::Deref] = borrowed.projection[..]
+            && let Some(borrowed) = self.locals[borrowed.local]
+        {
+            inner = borrowed;
+        }
+
+        // We have an unsizing cast, which assigns the length to wide pointer metadata.
+        if let Value::Cast { kind, from, to, .. } = self.get(inner)
+            && let CastKind::PointerCoercion(ty::adjustment::PointerCoercion::Unsize, _) = kind
+            && let Some(from) = from.builtin_deref(true)
+            && let ty::Array(_, len) = from.kind()
+            && let Some(to) = to.builtin_deref(true)
+            && let ty::Slice(..) = to.kind()
+        {
+            return self.insert_constant(Const::from_ty_const(
+                *len,
+                self.tcx.types.usize,
+                self.tcx,
+            ));
+        }
+
+        // Fallback: a symbolic `Len`.
+        Some(self.insert(Value::Len(inner)))
+    }
+
     fn pointers_have_same_metadata(&self, left_ptr_ty: Ty<'tcx>, right_ptr_ty: Ty<'tcx>) -> bool {
         let left_meta_ty = left_ptr_ty.pointee_metadata_ty_or_projection(self.tcx);
         let right_meta_ty = right_ptr_ty.pointee_metadata_ty_or_projection(self.tcx);

compiler/rustc_mir_transform/src/instsimplify.rs

@@ -46,6 +46,7 @@ impl<'tcx> crate::MirPass<'tcx> for InstSimplify {
                         }
                         ctx.simplify_bool_cmp(rvalue);
                         ctx.simplify_ref_deref(rvalue);
+                        ctx.simplify_len(rvalue);
                         ctx.simplify_ptr_aggregate(rvalue);
                         ctx.simplify_cast(rvalue);
                         ctx.simplify_repeated_aggregate(rvalue);
@@ -161,6 +162,18 @@ impl<'tcx> InstSimplifyContext<'_, 'tcx> {
         }
     }
 
+    /// Transform `Len([_; N])` ==> `N`.
+    fn simplify_len(&self, rvalue: &mut Rvalue<'tcx>) {
+        if let Rvalue::Len(ref place) = *rvalue {
+            let place_ty = place.ty(self.local_decls, self.tcx).ty;
+            if let ty::Array(_, len) = *place_ty.kind() {
+                let const_ = Const::from_ty_const(len, self.tcx.types.usize, self.tcx);
+                let constant = ConstOperand { span: DUMMY_SP, const_, user_ty: None };
+                *rvalue = Rvalue::Use(Operand::Constant(Box::new(constant)));
+            }
+        }
+    }
+
     /// Transform `Aggregate(RawPtr, [p, ()])` ==> `Cast(PtrToPtr, p)`.
     fn simplify_ptr_aggregate(&self, rvalue: &mut Rvalue<'tcx>) {
         if let Rvalue::Aggregate(box AggregateKind::RawPtr(pointee_ty, mutability), fields) = rvalue

compiler/rustc_mir_transform/src/known_panics_lint.rs

@@ -440,6 +440,7 @@ impl<'mir, 'tcx> ConstPropagator<'mir, 'tcx> {
             | Rvalue::Use(..)
             | Rvalue::CopyForDeref(..)
             | Rvalue::Repeat(..)
+            | Rvalue::Len(..)
             | Rvalue::Cast(..)
             | Rvalue::ShallowInitBox(..)
             | Rvalue::Discriminant(..)
@@ -599,6 +600,20 @@ impl<'mir, 'tcx> ConstPropagator<'mir, 'tcx> {
                 return None;
             }
 
+            Len(place) => {
+                let len = if let ty::Array(_, n) = place.ty(self.local_decls(), self.tcx).ty.kind()
+                {
+                    n.try_to_target_usize(self.tcx)?
+                } else {
+                    match self.get_const(place)? {
+                        Value::Immediate(src) => src.len(&self.ecx).discard_err()?,
+                        Value::Aggregate { fields, .. } => fields.len() as u64,
+                        Value::Uninit => return None,
+                    }
+                };
+                ImmTy::from_scalar(Scalar::from_target_usize(len, self), layout).into()
+            }
+
             Ref(..) | RawPtr(..) => return None,
 
             NullaryOp(ref null_op, ty) => {

compiler/rustc_mir_transform/src/promote_consts.rs

@@ -430,7 +430,9 @@ impl<'tcx> Validator<'_, 'tcx> {
                 self.validate_operand(op)?
             }
 
-            Rvalue::Discriminant(place) => self.validate_place(place.as_ref())?,
+            Rvalue::Discriminant(place) | Rvalue::Len(place) => {
+                self.validate_place(place.as_ref())?
+            }
 
             Rvalue::ThreadLocalRef(_) => return Err(Unpromotable),
 

compiler/rustc_mir_transform/src/validate.rs

@@ -1018,6 +1018,14 @@ impl<'a, 'tcx> Visitor<'tcx> for TypeChecker<'a, 'tcx> {
                 }
             }
             Rvalue::Ref(..) => {}
+            Rvalue::Len(p) => {
+                let pty = p.ty(&self.body.local_decls, self.tcx).ty;
+                check_kinds!(
+                    pty,
+                    "Cannot compute length of non-array type {:?}",
+                    ty::Array(..) | ty::Slice(..)
+                );
+            }
             Rvalue::BinaryOp(op, vals) => {
                 use BinOp::*;
                 let a = vals.0.ty(&self.body.local_decls, self.tcx);
@@ -1116,6 +1124,14 @@ impl<'a, 'tcx> Visitor<'tcx> for TypeChecker<'a, 'tcx> {
                         );
                     }
                     UnOp::PtrMetadata => {
+                        if !matches!(self.body.phase, MirPhase::Runtime(_)) {
+                            // It would probably be fine to support this in earlier phases, but at
+                            // the time of writing it's only ever introduced from intrinsic
+                            // lowering or other runtime-phase optimization passes, so earlier
+                            // things can just `bug!` on it.
+                            self.fail(location, "PtrMetadata should be in runtime MIR only");
+                        }
+
                         check_kinds!(
                             a,
                             "Cannot PtrMetadata non-pointer non-reference type {:?}",