rustc_const_eval/interpret/
validity.rs

1//! Check the validity invariant of a given value, and tell the user
2//! where in the value it got violated.
3//! In const context, this goes even further and tries to approximate const safety.
4//! That's useful because it means other passes (e.g. promotion) can rely on `const`s
5//! to be const-safe.
6
7use std::borrow::Cow;
8use std::fmt::Write;
9use std::hash::Hash;
10use std::num::NonZero;
11
12use either::{Left, Right};
13use hir::def::DefKind;
14use rustc_abi::{
15    BackendRepr, FieldIdx, FieldsShape, Scalar as ScalarAbi, Size, VariantIdx, Variants,
16    WrappingRange,
17};
18use rustc_ast::Mutability;
19use rustc_data_structures::fx::FxHashSet;
20use rustc_hir as hir;
21use rustc_middle::bug;
22use rustc_middle::mir::interpret::ValidationErrorKind::{self, *};
23use rustc_middle::mir::interpret::{
24    ExpectedKind, InterpErrorKind, InvalidMetaKind, Misalignment, PointerKind, Provenance,
25    UnsupportedOpInfo, ValidationErrorInfo, alloc_range, interp_ok,
26};
27use rustc_middle::ty::layout::{LayoutCx, TyAndLayout};
28use rustc_middle::ty::{self, Ty};
29use rustc_span::{Symbol, sym};
30use tracing::trace;
31
32use super::machine::AllocMap;
33use super::{
34    AllocId, CheckInAllocMsg, GlobalAlloc, ImmTy, Immediate, InterpCx, InterpResult, MPlaceTy,
35    Machine, MemPlaceMeta, PlaceTy, Pointer, Projectable, Scalar, ValueVisitor, err_ub,
36    format_interp_error,
37};
38use crate::enter_trace_span;
39
40// for the validation errors
41#[rustfmt::skip]
42use super::InterpErrorKind::UndefinedBehavior as Ub;
43use super::InterpErrorKind::Unsupported as Unsup;
44use super::UndefinedBehaviorInfo::*;
45use super::UnsupportedOpInfo::*;
46
47macro_rules! err_validation_failure {
48    ($where:expr, $kind: expr) => {{
49        let where_ = &$where;
50        let path = if !where_.is_empty() {
51            let mut path = String::new();
52            write_path(&mut path, where_);
53            Some(path)
54        } else {
55            None
56        };
57
58        err_ub!(ValidationError(ValidationErrorInfo { path, kind: $kind }))
59    }};
60}
61
62macro_rules! throw_validation_failure {
63    ($where:expr, $kind: expr) => {
64        do yeet err_validation_failure!($where, $kind)
65    };
66}
67
68/// If $e throws an error matching the pattern, throw a validation failure.
69/// Other errors are passed back to the caller, unchanged -- and if they reach the root of
70/// the visitor, we make sure only validation errors and `InvalidProgram` errors are left.
71/// This lets you use the patterns as a kind of validation list, asserting which errors
72/// can possibly happen:
73///
74/// ```ignore(illustrative)
75/// let v = try_validation!(some_fn(), some_path, {
76///     Foo | Bar | Baz => { "some failure" },
77/// });
78/// ```
79///
80/// The patterns must be of type `UndefinedBehaviorInfo`.
81/// An additional expected parameter can also be added to the failure message:
82///
83/// ```ignore(illustrative)
84/// let v = try_validation!(some_fn(), some_path, {
85///     Foo | Bar | Baz => { "some failure" } expected { "something that wasn't a failure" },
86/// });
87/// ```
88///
89/// An additional nicety is that both parameters actually take format args, so you can just write
90/// the format string in directly:
91///
92/// ```ignore(illustrative)
93/// let v = try_validation!(some_fn(), some_path, {
94///     Foo | Bar | Baz => { "{:?}", some_failure } expected { "{}", expected_value },
95/// });
96/// ```
97///
98macro_rules! try_validation {
99    ($e:expr, $where:expr,
100    $( $( $p:pat_param )|+ => $kind: expr ),+ $(,)?
101    ) => {{
102        $e.map_err_kind(|e| {
103            // We catch the error and turn it into a validation failure. We are okay with
104            // allocation here as this can only slow down builds that fail anyway.
105            match e {
106                $(
107                    $($p)|+ => {
108                        err_validation_failure!(
109                            $where,
110                            $kind
111                        )
112                    }
113                ),+,
114                e => e,
115            }
116        })?
117    }};
118}
119
120/// We want to show a nice path to the invalid field for diagnostics,
121/// but avoid string operations in the happy case where no error happens.
122/// So we track a `Vec<PathElem>` where `PathElem` contains all the data we
123/// need to later print something for the user.
124#[derive(Copy, Clone, Debug)]
125pub enum PathElem {
126    Field(Symbol),
127    Variant(Symbol),
128    CoroutineState(VariantIdx),
129    CapturedVar(Symbol),
130    ArrayElem(usize),
131    TupleElem(usize),
132    Deref,
133    EnumTag,
134    CoroutineTag,
135    DynDowncast,
136    Vtable,
137}
138
139/// Extra things to check for during validation of CTFE results.
140#[derive(Copy, Clone)]
141pub enum CtfeValidationMode {
142    /// Validation of a `static`
143    Static { mutbl: Mutability },
144    /// Validation of a promoted.
145    Promoted,
146    /// Validation of a `const`.
147    /// `allow_immutable_unsafe_cell` says whether we allow `UnsafeCell` in immutable memory (which is the
148    /// case for the top-level allocation of a `const`, where this is fine because the allocation will be
149    /// copied at each use site).
150    Const { allow_immutable_unsafe_cell: bool },
151}
152
153impl CtfeValidationMode {
154    fn allow_immutable_unsafe_cell(self) -> bool {
155        match self {
156            CtfeValidationMode::Static { .. } => false,
157            CtfeValidationMode::Promoted { .. } => false,
158            CtfeValidationMode::Const { allow_immutable_unsafe_cell, .. } => {
159                allow_immutable_unsafe_cell
160            }
161        }
162    }
163}
164
165/// State for tracking recursive validation of references
166pub struct RefTracking<T, PATH = ()> {
167    seen: FxHashSet<T>,
168    todo: Vec<(T, PATH)>,
169}
170
171impl<T: Clone + Eq + Hash + std::fmt::Debug, PATH: Default> RefTracking<T, PATH> {
172    pub fn empty() -> Self {
173        RefTracking { seen: FxHashSet::default(), todo: vec![] }
174    }
175    pub fn new(val: T) -> Self {
176        let mut ref_tracking_for_consts =
177            RefTracking { seen: FxHashSet::default(), todo: vec![(val.clone(), PATH::default())] };
178        ref_tracking_for_consts.seen.insert(val);
179        ref_tracking_for_consts
180    }
181    pub fn next(&mut self) -> Option<(T, PATH)> {
182        self.todo.pop()
183    }
184
185    fn track(&mut self, val: T, path: impl FnOnce() -> PATH) {
186        if self.seen.insert(val.clone()) {
187            trace!("Recursing below ptr {:#?}", val);
188            let path = path();
189            // Remember to come back to this later.
190            self.todo.push((val, path));
191        }
192    }
193}
194
195// FIXME make this translatable as well?
196/// Format a path
197fn write_path(out: &mut String, path: &[PathElem]) {
198    use self::PathElem::*;
199
200    for elem in path.iter() {
201        match elem {
202            Field(name) => write!(out, ".{name}"),
203            EnumTag => write!(out, ".<enum-tag>"),
204            Variant(name) => write!(out, ".<enum-variant({name})>"),
205            CoroutineTag => write!(out, ".<coroutine-tag>"),
206            CoroutineState(idx) => write!(out, ".<coroutine-state({})>", idx.index()),
207            CapturedVar(name) => write!(out, ".<captured-var({name})>"),
208            TupleElem(idx) => write!(out, ".{idx}"),
209            ArrayElem(idx) => write!(out, "[{idx}]"),
210            // `.<deref>` does not match Rust syntax, but it is more readable for long paths -- and
211            // some of the other items here also are not Rust syntax. Actually we can't
212            // even use the usual syntax because we are just showing the projections,
213            // not the root.
214            Deref => write!(out, ".<deref>"),
215            DynDowncast => write!(out, ".<dyn-downcast>"),
216            Vtable => write!(out, ".<vtable>"),
217        }
218        .unwrap()
219    }
220}
221
222/// Represents a set of `Size` values as a sorted list of ranges.
223// These are (offset, length) pairs, and they are sorted and mutually disjoint,
224// and never adjacent (i.e. there's always a gap between two of them).
225#[derive(Debug, Clone)]
226pub struct RangeSet(Vec<(Size, Size)>);
227
228impl RangeSet {
229    fn add_range(&mut self, offset: Size, size: Size) {
230        if size.bytes() == 0 {
231            // No need to track empty ranges.
232            return;
233        }
234        let v = &mut self.0;
235        // We scan for a partition point where the left partition is all the elements that end
236        // strictly before we start. Those are elements that are too "low" to merge with us.
237        let idx =
238            v.partition_point(|&(other_offset, other_size)| other_offset + other_size < offset);
239        // Now we want to either merge with the first element of the second partition, or insert ourselves before that.
240        if let Some(&(other_offset, other_size)) = v.get(idx)
241            && offset + size >= other_offset
242        {
243            // Their end is >= our start (otherwise it would not be in the 2nd partition) and
244            // our end is >= their start. This means we can merge the ranges.
245            let new_start = other_offset.min(offset);
246            let mut new_end = (other_offset + other_size).max(offset + size);
247            // We grew to the right, so merge with overlapping/adjacent elements.
248            // (We also may have grown to the left, but that can never make us adjacent with
249            // anything there since we selected the first such candidate via `partition_point`.)
250            let mut scan_right = 1;
251            while let Some(&(next_offset, next_size)) = v.get(idx + scan_right)
252                && new_end >= next_offset
253            {
254                // Increase our size to absorb the next element.
255                new_end = new_end.max(next_offset + next_size);
256                // Look at the next element.
257                scan_right += 1;
258            }
259            // Update the element we grew.
260            v[idx] = (new_start, new_end - new_start);
261            // Remove the elements we absorbed (if any).
262            if scan_right > 1 {
263                drop(v.drain((idx + 1)..(idx + scan_right)));
264            }
265        } else {
266            // Insert new element.
267            v.insert(idx, (offset, size));
268        }
269    }
270}
271
272struct ValidityVisitor<'rt, 'tcx, M: Machine<'tcx>> {
273    /// The `path` may be pushed to, but the part that is present when a function
274    /// starts must not be changed!  `visit_fields` and `visit_array` rely on
275    /// this stack discipline.
276    path: Vec<PathElem>,
277    ref_tracking: Option<&'rt mut RefTracking<MPlaceTy<'tcx, M::Provenance>, Vec<PathElem>>>,
278    /// `None` indicates this is not validating for CTFE (but for runtime).
279    ctfe_mode: Option<CtfeValidationMode>,
280    ecx: &'rt mut InterpCx<'tcx, M>,
281    /// Whether provenance should be reset outside of pointers (emulating the effect of a typed
282    /// copy).
283    reset_provenance_and_padding: bool,
284    /// This tracks which byte ranges in this value contain data; the remaining bytes are padding.
285    /// The ideal representation here would be pointer-length pairs, but to keep things more compact
286    /// we only store a (range) set of offsets -- the base pointer is the same throughout the entire
287    /// visit, after all.
288    /// If this is `Some`, then `reset_provenance_and_padding` must be true (but not vice versa:
289    /// we might not track data vs padding bytes if the operand isn't stored in memory anyway).
290    data_bytes: Option<RangeSet>,
291}
292
293impl<'rt, 'tcx, M: Machine<'tcx>> ValidityVisitor<'rt, 'tcx, M> {
294    fn aggregate_field_path_elem(&mut self, layout: TyAndLayout<'tcx>, field: usize) -> PathElem {
295        // First, check if we are projecting to a variant.
296        match layout.variants {
297            Variants::Multiple { tag_field, .. } => {
298                if tag_field.as_usize() == field {
299                    return match layout.ty.kind() {
300                        ty::Adt(def, ..) if def.is_enum() => PathElem::EnumTag,
301                        ty::Coroutine(..) => PathElem::CoroutineTag,
302                        _ => bug!("non-variant type {:?}", layout.ty),
303                    };
304                }
305            }
306            Variants::Single { .. } | Variants::Empty => {}
307        }
308
309        // Now we know we are projecting to a field, so figure out which one.
310        match layout.ty.kind() {
311            // coroutines, closures, and coroutine-closures all have upvars that may be named.
312            ty::Closure(def_id, _) | ty::Coroutine(def_id, _) | ty::CoroutineClosure(def_id, _) => {
313                let mut name = None;
314                // FIXME this should be more descriptive i.e. CapturePlace instead of CapturedVar
315                // https://github.com/rust-lang/project-rfc-2229/issues/46
316                if let Some(local_def_id) = def_id.as_local() {
317                    let captures = self.ecx.tcx.closure_captures(local_def_id);
318                    if let Some(captured_place) = captures.get(field) {
319                        // Sometimes the index is beyond the number of upvars (seen
320                        // for a coroutine).
321                        let var_hir_id = captured_place.get_root_variable();
322                        let node = self.ecx.tcx.hir_node(var_hir_id);
323                        if let hir::Node::Pat(pat) = node {
324                            if let hir::PatKind::Binding(_, _, ident, _) = pat.kind {
325                                name = Some(ident.name);
326                            }
327                        }
328                    }
329                }
330
331                PathElem::CapturedVar(name.unwrap_or_else(|| {
332                    // Fall back to showing the field index.
333                    sym::integer(field)
334                }))
335            }
336
337            // tuples
338            ty::Tuple(_) => PathElem::TupleElem(field),
339
340            // enums
341            ty::Adt(def, ..) if def.is_enum() => {
342                // we might be projecting *to* a variant, or to a field *in* a variant.
343                match layout.variants {
344                    Variants::Single { index } => {
345                        // Inside a variant
346                        PathElem::Field(def.variant(index).fields[FieldIdx::from_usize(field)].name)
347                    }
348                    Variants::Empty => panic!("there is no field in Variants::Empty types"),
349                    Variants::Multiple { .. } => bug!("we handled variants above"),
350                }
351            }
352
353            // other ADTs
354            ty::Adt(def, _) => {
355                PathElem::Field(def.non_enum_variant().fields[FieldIdx::from_usize(field)].name)
356            }
357
358            // arrays/slices
359            ty::Array(..) | ty::Slice(..) => PathElem::ArrayElem(field),
360
361            // dyn* vtables
362            ty::Dynamic(_, _, ty::DynKind::DynStar) if field == 1 => PathElem::Vtable,
363
364            // dyn traits
365            ty::Dynamic(..) => {
366                assert_eq!(field, 0);
367                PathElem::DynDowncast
368            }
369
370            // nothing else has an aggregate layout
371            _ => bug!("aggregate_field_path_elem: got non-aggregate type {:?}", layout.ty),
372        }
373    }
374
375    fn with_elem<R>(
376        &mut self,
377        elem: PathElem,
378        f: impl FnOnce(&mut Self) -> InterpResult<'tcx, R>,
379    ) -> InterpResult<'tcx, R> {
380        // Remember the old state
381        let path_len = self.path.len();
382        // Record new element
383        self.path.push(elem);
384        // Perform operation
385        let r = f(self)?;
386        // Undo changes
387        self.path.truncate(path_len);
388        // Done
389        interp_ok(r)
390    }
391
392    fn read_immediate(
393        &self,
394        val: &PlaceTy<'tcx, M::Provenance>,
395        expected: ExpectedKind,
396    ) -> InterpResult<'tcx, ImmTy<'tcx, M::Provenance>> {
397        interp_ok(try_validation!(
398            self.ecx.read_immediate(val),
399            self.path,
400            Ub(InvalidUninitBytes(None)) =>
401                Uninit { expected },
402            // The `Unsup` cases can only occur during CTFE
403            Unsup(ReadPointerAsInt(_)) =>
404                PointerAsInt { expected },
405            Unsup(ReadPartialPointer(_)) =>
406                PartialPointer,
407        ))
408    }
409
410    fn read_scalar(
411        &self,
412        val: &PlaceTy<'tcx, M::Provenance>,
413        expected: ExpectedKind,
414    ) -> InterpResult<'tcx, Scalar<M::Provenance>> {
415        interp_ok(self.read_immediate(val, expected)?.to_scalar())
416    }
417
418    fn deref_pointer(
419        &mut self,
420        val: &PlaceTy<'tcx, M::Provenance>,
421        expected: ExpectedKind,
422    ) -> InterpResult<'tcx, MPlaceTy<'tcx, M::Provenance>> {
423        // Not using `ecx.deref_pointer` since we want to use our `read_immediate` wrapper.
424        let imm = self.read_immediate(val, expected)?;
425        // Reset provenance: ensure slice tail metadata does not preserve provenance,
426        // and ensure all pointers do not preserve partial provenance.
427        if self.reset_provenance_and_padding {
428            if matches!(imm.layout.backend_repr, BackendRepr::Scalar(..)) {
429                // A thin pointer. If it has provenance, we don't have to do anything.
430                // If it does not, ensure we clear the provenance in memory.
431                if matches!(imm.to_scalar(), Scalar::Int(..)) {
432                    self.ecx.clear_provenance(val)?;
433                }
434            } else {
435                // A wide pointer. This means we have to worry both about the pointer itself and the
436                // metadata. We do the lazy thing and just write back the value we got. Just
437                // clearing provenance in a targeted manner would be more efficient, but unless this
438                // is a perf hotspot it's just not worth the effort.
439                self.ecx.write_immediate_no_validate(*imm, val)?;
440            }
441            // The entire thing is data, not padding.
442            self.add_data_range_place(val);
443        }
444        // Now turn it into a place.
445        self.ecx.ref_to_mplace(&imm)
446    }
447
448    fn check_wide_ptr_meta(
449        &mut self,
450        meta: MemPlaceMeta<M::Provenance>,
451        pointee: TyAndLayout<'tcx>,
452    ) -> InterpResult<'tcx> {
453        let tail = self.ecx.tcx.struct_tail_for_codegen(pointee.ty, self.ecx.typing_env);
454        match tail.kind() {
455            ty::Dynamic(data, _, ty::Dyn) => {
456                let vtable = meta.unwrap_meta().to_pointer(self.ecx)?;
457                // Make sure it is a genuine vtable pointer for the right trait.
458                try_validation!(
459                    self.ecx.get_ptr_vtable_ty(vtable, Some(data)),
460                    self.path,
461                    Ub(DanglingIntPointer{ .. } | InvalidVTablePointer(..)) =>
462                        InvalidVTablePtr { value: format!("{vtable}") },
463                    Ub(InvalidVTableTrait { vtable_dyn_type, expected_dyn_type }) => {
464                        InvalidMetaWrongTrait { vtable_dyn_type, expected_dyn_type }
465                    },
466                );
467            }
468            ty::Slice(..) | ty::Str => {
469                let _len = meta.unwrap_meta().to_target_usize(self.ecx)?;
470                // We do not check that `len * elem_size <= isize::MAX`:
471                // that is only required for references, and there it falls out of the
472                // "dereferenceable" check performed by Stacked Borrows.
473            }
474            ty::Foreign(..) => {
475                // Unsized, but not wide.
476            }
477            _ => bug!("Unexpected unsized type tail: {:?}", tail),
478        }
479
480        interp_ok(())
481    }
482
483    /// Check a reference or `Box`.
484    fn check_safe_pointer(
485        &mut self,
486        value: &PlaceTy<'tcx, M::Provenance>,
487        ptr_kind: PointerKind,
488    ) -> InterpResult<'tcx> {
489        let place = self.deref_pointer(value, ptr_kind.into())?;
490        // Handle wide pointers.
491        // Check metadata early, for better diagnostics
492        if place.layout.is_unsized() {
493            self.check_wide_ptr_meta(place.meta(), place.layout)?;
494        }
495        // Make sure this is dereferenceable and all.
496        let size_and_align = try_validation!(
497            self.ecx.size_and_align_of_val(&place),
498            self.path,
499            Ub(InvalidMeta(msg)) => match msg {
500                InvalidMetaKind::SliceTooBig => InvalidMetaSliceTooLarge { ptr_kind },
501                InvalidMetaKind::TooBig => InvalidMetaTooLarge { ptr_kind },
502            }
503        );
504        let (size, align) = size_and_align
505            // for the purpose of validity, consider foreign types to have
506            // alignment and size determined by the layout (size will be 0,
507            // alignment should take attributes into account).
508            .unwrap_or_else(|| (place.layout.size, place.layout.align.abi));
509        // Direct call to `check_ptr_access_align` checks alignment even on CTFE machines.
510        try_validation!(
511            self.ecx.check_ptr_access(
512                place.ptr(),
513                size,
514                CheckInAllocMsg::Dereferenceable, // will anyway be replaced by validity message
515            ),
516            self.path,
517            Ub(DanglingIntPointer { addr: 0, .. }) => NullPtr { ptr_kind },
518            Ub(DanglingIntPointer { addr: i, .. }) => DanglingPtrNoProvenance {
519                ptr_kind,
520                // FIXME this says "null pointer" when null but we need translate
521                pointer: format!("{}", Pointer::<Option<AllocId>>::from_addr_invalid(i))
522            },
523            Ub(PointerOutOfBounds { .. }) => DanglingPtrOutOfBounds {
524                ptr_kind
525            },
526            Ub(PointerUseAfterFree(..)) => DanglingPtrUseAfterFree {
527                ptr_kind,
528            },
529        );
530        try_validation!(
531            self.ecx.check_ptr_align(
532                place.ptr(),
533                align,
534            ),
535            self.path,
536            Ub(AlignmentCheckFailed(Misalignment { required, has }, _msg)) => UnalignedPtr {
537                ptr_kind,
538                required_bytes: required.bytes(),
539                found_bytes: has.bytes()
540            },
541        );
542        // Make sure this is non-null. We checked dereferenceability above, but if `size` is zero
543        // that does not imply non-null.
544        if self.ecx.scalar_may_be_null(Scalar::from_maybe_pointer(place.ptr(), self.ecx))? {
545            throw_validation_failure!(self.path, NullPtr { ptr_kind })
546        }
547        // Do not allow references to uninhabited types.
548        if place.layout.is_uninhabited() {
549            let ty = place.layout.ty;
550            throw_validation_failure!(self.path, PtrToUninhabited { ptr_kind, ty })
551        }
552        // Recursive checking
553        if let Some(ref_tracking) = self.ref_tracking.as_deref_mut() {
554            // Proceed recursively even for ZST, no reason to skip them!
555            // `!` is a ZST and we want to validate it.
556            if let Some(ctfe_mode) = self.ctfe_mode {
557                let mut skip_recursive_check = false;
558                // CTFE imposes restrictions on what references can point to.
559                if let Ok((alloc_id, _offset, _prov)) =
560                    self.ecx.ptr_try_get_alloc_id(place.ptr(), 0)
561                {
562                    // Everything should be already interned.
563                    let Some(global_alloc) = self.ecx.tcx.try_get_global_alloc(alloc_id) else {
564                        assert!(self.ecx.memory.alloc_map.get(alloc_id).is_none());
565                        // We can't have *any* references to non-existing allocations in const-eval
566                        // as the rest of rustc isn't happy with them... so we throw an error, even
567                        // though for zero-sized references this isn't really UB.
568                        // A potential future alternative would be to resurrect this as a zero-sized allocation
569                        // (which codegen will then compile to an aligned dummy pointer anyway).
570                        throw_validation_failure!(self.path, DanglingPtrUseAfterFree { ptr_kind });
571                    };
572                    let (size, _align) =
573                        global_alloc.size_and_align(*self.ecx.tcx, self.ecx.typing_env);
574                    let alloc_actual_mutbl =
575                        global_alloc.mutability(*self.ecx.tcx, self.ecx.typing_env);
576
577                    if let GlobalAlloc::Static(did) = global_alloc {
578                        let DefKind::Static { nested, .. } = self.ecx.tcx.def_kind(did) else {
579                            bug!()
580                        };
581                        // Special handling for pointers to statics (irrespective of their type).
582                        assert!(!self.ecx.tcx.is_thread_local_static(did));
583                        assert!(self.ecx.tcx.is_static(did));
584                        // Mode-specific checks
585                        match ctfe_mode {
586                            CtfeValidationMode::Static { .. }
587                            | CtfeValidationMode::Promoted { .. } => {
588                                // We skip recursively checking other statics. These statics must be sound by
589                                // themselves, and the only way to get broken statics here is by using
590                                // unsafe code.
591                                // The reasons we don't check other statics is twofold. For one, in all
592                                // sound cases, the static was already validated on its own, and second, we
593                                // trigger cycle errors if we try to compute the value of the other static
594                                // and that static refers back to us (potentially through a promoted).
595                                // This could miss some UB, but that's fine.
596                                // We still walk nested allocations, as they are fundamentally part of this validation run.
597                                // This means we will also recurse into nested statics of *other*
598                                // statics, even though we do not recurse into other statics directly.
599                                // That's somewhat inconsistent but harmless.
600                                skip_recursive_check = !nested;
601                            }
602                            CtfeValidationMode::Const { .. } => {
603                                // If this is mutable memory or an `extern static`, there's no point in checking it -- we'd
604                                // just get errors trying to read the value.
605                                if alloc_actual_mutbl.is_mut() || self.ecx.tcx.is_foreign_item(did)
606                                {
607                                    skip_recursive_check = true;
608                                }
609                            }
610                        }
611                    }
612
613                    // If this allocation has size zero, there is no actual mutability here.
614                    if size != Size::ZERO {
615                        // Determine whether this pointer expects to be pointing to something mutable.
616                        let ptr_expected_mutbl = match ptr_kind {
617                            PointerKind::Box => Mutability::Mut,
618                            PointerKind::Ref(mutbl) => {
619                                // We do not take into account interior mutability here since we cannot know if
620                                // there really is an `UnsafeCell` inside `Option<UnsafeCell>` -- so we check
621                                // that in the recursive descent behind this reference (controlled by
622                                // `allow_immutable_unsafe_cell`).
623                                mutbl
624                            }
625                        };
626                        // Mutable pointer to immutable memory is no good.
627                        if ptr_expected_mutbl == Mutability::Mut
628                            && alloc_actual_mutbl == Mutability::Not
629                        {
630                            // This can actually occur with transmutes.
631                            throw_validation_failure!(self.path, MutableRefToImmutable);
632                        }
633                        // In a const, any kind of mutable reference is not good.
634                        if matches!(self.ctfe_mode, Some(CtfeValidationMode::Const { .. })) {
635                            if ptr_expected_mutbl == Mutability::Mut {
636                                throw_validation_failure!(self.path, MutableRefInConst);
637                            }
638                        }
639                    }
640                }
641                // Potentially skip recursive check.
642                if skip_recursive_check {
643                    return interp_ok(());
644                }
645            } else {
646                // This is not CTFE, so it's Miri with recursive checking.
647                // FIXME: we do *not* check behind boxes, since creating a new box first creates it uninitialized
648                // and then puts the value in there, so briefly we have a box with uninit contents.
649                // FIXME: should we also skip `UnsafeCell` behind shared references? Currently that is not
650                // needed since validation reads bypass Stacked Borrows and data race checks.
651                if matches!(ptr_kind, PointerKind::Box) {
652                    return interp_ok(());
653                }
654            }
655            let path = &self.path;
656            ref_tracking.track(place, || {
657                // We need to clone the path anyway, make sure it gets created
658                // with enough space for the additional `Deref`.
659                let mut new_path = Vec::with_capacity(path.len() + 1);
660                new_path.extend(path);
661                new_path.push(PathElem::Deref);
662                new_path
663            });
664        }
665        interp_ok(())
666    }
667
668    /// Check if this is a value of primitive type, and if yes check the validity of the value
669    /// at that type. Return `true` if the type is indeed primitive.
670    ///
671    /// Note that not all of these have `FieldsShape::Primitive`, e.g. wide references.
672    fn try_visit_primitive(
673        &mut self,
674        value: &PlaceTy<'tcx, M::Provenance>,
675    ) -> InterpResult<'tcx, bool> {
676        // Go over all the primitive types
677        let ty = value.layout.ty;
678        match ty.kind() {
679            ty::Bool => {
680                let scalar = self.read_scalar(value, ExpectedKind::Bool)?;
681                try_validation!(
682                    scalar.to_bool(),
683                    self.path,
684                    Ub(InvalidBool(..)) => ValidationErrorKind::InvalidBool {
685                        value: format!("{scalar:x}"),
686                    }
687                );
688                if self.reset_provenance_and_padding {
689                    self.ecx.clear_provenance(value)?;
690                    self.add_data_range_place(value);
691                }
692                interp_ok(true)
693            }
694            ty::Char => {
695                let scalar = self.read_scalar(value, ExpectedKind::Char)?;
696                try_validation!(
697                    scalar.to_char(),
698                    self.path,
699                    Ub(InvalidChar(..)) => ValidationErrorKind::InvalidChar {
700                        value: format!("{scalar:x}"),
701                    }
702                );
703                if self.reset_provenance_and_padding {
704                    self.ecx.clear_provenance(value)?;
705                    self.add_data_range_place(value);
706                }
707                interp_ok(true)
708            }
709            ty::Float(_) | ty::Int(_) | ty::Uint(_) => {
710                // NOTE: Keep this in sync with the array optimization for int/float
711                // types below!
712                self.read_scalar(
713                    value,
714                    if matches!(ty.kind(), ty::Float(..)) {
715                        ExpectedKind::Float
716                    } else {
717                        ExpectedKind::Int
718                    },
719                )?;
720                if self.reset_provenance_and_padding {
721                    self.ecx.clear_provenance(value)?;
722                    self.add_data_range_place(value);
723                }
724                interp_ok(true)
725            }
726            ty::RawPtr(..) => {
727                let place = self.deref_pointer(value, ExpectedKind::RawPtr)?;
728                if place.layout.is_unsized() {
729                    self.check_wide_ptr_meta(place.meta(), place.layout)?;
730                }
731                interp_ok(true)
732            }
733            ty::Ref(_, _ty, mutbl) => {
734                self.check_safe_pointer(value, PointerKind::Ref(*mutbl))?;
735                interp_ok(true)
736            }
737            ty::FnPtr(..) => {
738                let scalar = self.read_scalar(value, ExpectedKind::FnPtr)?;
739
740                // If we check references recursively, also check that this points to a function.
741                if let Some(_) = self.ref_tracking {
742                    let ptr = scalar.to_pointer(self.ecx)?;
743                    let _fn = try_validation!(
744                        self.ecx.get_ptr_fn(ptr),
745                        self.path,
746                        Ub(DanglingIntPointer{ .. } | InvalidFunctionPointer(..)) =>
747                            InvalidFnPtr { value: format!("{ptr}") },
748                    );
749                    // FIXME: Check if the signature matches
750                } else {
751                    // Otherwise (for standalone Miri), we have to still check it to be non-null.
752                    if self.ecx.scalar_may_be_null(scalar)? {
753                        throw_validation_failure!(self.path, NullFnPtr);
754                    }
755                }
756                if self.reset_provenance_and_padding {
757                    // Make sure we do not preserve partial provenance. This matches the thin
758                    // pointer handling in `deref_pointer`.
759                    if matches!(scalar, Scalar::Int(..)) {
760                        self.ecx.clear_provenance(value)?;
761                    }
762                    self.add_data_range_place(value);
763                }
764                interp_ok(true)
765            }
766            ty::Never => throw_validation_failure!(self.path, NeverVal),
767            ty::Foreign(..) | ty::FnDef(..) => {
768                // Nothing to check.
769                interp_ok(true)
770            }
771            ty::UnsafeBinder(_) => todo!("FIXME(unsafe_binder)"),
772            // The above should be all the primitive types. The rest is compound, we
773            // check them by visiting their fields/variants.
774            ty::Adt(..)
775            | ty::Tuple(..)
776            | ty::Array(..)
777            | ty::Slice(..)
778            | ty::Str
779            | ty::Dynamic(..)
780            | ty::Closure(..)
781            | ty::Pat(..)
782            | ty::CoroutineClosure(..)
783            | ty::Coroutine(..) => interp_ok(false),
784            // Some types only occur during typechecking, they have no layout.
785            // We should not see them here and we could not check them anyway.
786            ty::Error(_)
787            | ty::Infer(..)
788            | ty::Placeholder(..)
789            | ty::Bound(..)
790            | ty::Param(..)
791            | ty::Alias(..)
792            | ty::CoroutineWitness(..) => bug!("Encountered invalid type {:?}", ty),
793        }
794    }
795
796    fn visit_scalar(
797        &mut self,
798        scalar: Scalar<M::Provenance>,
799        scalar_layout: ScalarAbi,
800    ) -> InterpResult<'tcx> {
801        let size = scalar_layout.size(self.ecx);
802        let valid_range = scalar_layout.valid_range(self.ecx);
803        let WrappingRange { start, end } = valid_range;
804        let max_value = size.unsigned_int_max();
805        assert!(end <= max_value);
806        let bits = match scalar.try_to_scalar_int() {
807            Ok(int) => int.to_bits(size),
808            Err(_) => {
809                // So this is a pointer then, and casting to an int failed.
810                // Can only happen during CTFE.
811                // We support 2 kinds of ranges here: full range, and excluding zero.
812                if start == 1 && end == max_value {
813                    // Only null is the niche. So make sure the ptr is NOT null.
814                    if self.ecx.scalar_may_be_null(scalar)? {
815                        throw_validation_failure!(
816                            self.path,
817                            NullablePtrOutOfRange { range: valid_range, max_value }
818                        )
819                    } else {
820                        return interp_ok(());
821                    }
822                } else if scalar_layout.is_always_valid(self.ecx) {
823                    // Easy. (This is reachable if `enforce_number_validity` is set.)
824                    return interp_ok(());
825                } else {
826                    // Conservatively, we reject, because the pointer *could* have a bad
827                    // value.
828                    throw_validation_failure!(
829                        self.path,
830                        PtrOutOfRange { range: valid_range, max_value }
831                    )
832                }
833            }
834        };
835        // Now compare.
836        if valid_range.contains(bits) {
837            interp_ok(())
838        } else {
839            throw_validation_failure!(
840                self.path,
841                OutOfRange { value: format!("{bits}"), range: valid_range, max_value }
842            )
843        }
844    }
845
846    fn in_mutable_memory(&self, val: &PlaceTy<'tcx, M::Provenance>) -> bool {
847        debug_assert!(self.ctfe_mode.is_some());
848        if let Some(mplace) = val.as_mplace_or_local().left() {
849            if let Some(alloc_id) = mplace.ptr().provenance.and_then(|p| p.get_alloc_id()) {
850                let tcx = *self.ecx.tcx;
851                // Everything must be already interned.
852                let mutbl = tcx.global_alloc(alloc_id).mutability(tcx, self.ecx.typing_env);
853                if let Some((_, alloc)) = self.ecx.memory.alloc_map.get(alloc_id) {
854                    assert_eq!(alloc.mutability, mutbl);
855                }
856                mutbl.is_mut()
857            } else {
858                // No memory at all.
859                false
860            }
861        } else {
862            // A local variable -- definitely mutable.
863            true
864        }
865    }
866
867    /// Add the given pointer-length pair to the "data" range of this visit.
868    fn add_data_range(&mut self, ptr: Pointer<Option<M::Provenance>>, size: Size) {
869        if let Some(data_bytes) = self.data_bytes.as_mut() {
870            // We only have to store the offset, the rest is the same for all pointers here.
871            let (_prov, offset) = ptr.into_parts();
872            // Add this.
873            data_bytes.add_range(offset, size);
874        };
875    }
876
877    /// Add the entire given place to the "data" range of this visit.
878    fn add_data_range_place(&mut self, place: &PlaceTy<'tcx, M::Provenance>) {
879        // Only sized places can be added this way.
880        debug_assert!(place.layout.is_sized());
881        if let Some(data_bytes) = self.data_bytes.as_mut() {
882            let offset = Self::data_range_offset(self.ecx, place);
883            data_bytes.add_range(offset, place.layout.size);
884        }
885    }
886
887    /// Convert a place into the offset it starts at, for the purpose of data_range tracking.
888    /// Must only be called if `data_bytes` is `Some(_)`.
889    fn data_range_offset(ecx: &InterpCx<'tcx, M>, place: &PlaceTy<'tcx, M::Provenance>) -> Size {
890        // The presence of `data_bytes` implies that our place is in memory.
891        let ptr = ecx
892            .place_to_op(place)
893            .expect("place must be in memory")
894            .as_mplace_or_imm()
895            .expect_left("place must be in memory")
896            .ptr();
897        let (_prov, offset) = ptr.into_parts();
898        offset
899    }
900
901    fn reset_padding(&mut self, place: &PlaceTy<'tcx, M::Provenance>) -> InterpResult<'tcx> {
902        let Some(data_bytes) = self.data_bytes.as_mut() else { return interp_ok(()) };
903        // Our value must be in memory, otherwise we would not have set up `data_bytes`.
904        let mplace = self.ecx.force_allocation(place)?;
905        // Determine starting offset and size.
906        let (_prov, start_offset) = mplace.ptr().into_parts();
907        let (size, _align) = self
908            .ecx
909            .size_and_align_of_val(&mplace)?
910            .unwrap_or((mplace.layout.size, mplace.layout.align.abi));
911        // If there is no padding at all, we can skip the rest: check for
912        // a single data range covering the entire value.
913        if data_bytes.0 == &[(start_offset, size)] {
914            return interp_ok(());
915        }
916        // Get a handle for the allocation. Do this only once, to avoid looking up the same
917        // allocation over and over again. (Though to be fair, iterating the value already does
918        // exactly that.)
919        let Some(mut alloc) = self.ecx.get_ptr_alloc_mut(mplace.ptr(), size)? else {
920            // A ZST, no padding to clear.
921            return interp_ok(());
922        };
923        // Add a "finalizer" data range at the end, so that the iteration below finds all gaps
924        // between ranges.
925        data_bytes.0.push((start_offset + size, Size::ZERO));
926        // Iterate, and reset gaps.
927        let mut padding_cleared_until = start_offset;
928        for &(offset, size) in data_bytes.0.iter() {
929            assert!(
930                offset >= padding_cleared_until,
931                "reset_padding on {}: previous field ended at offset {}, next field starts at {} (and has a size of {} bytes)",
932                mplace.layout.ty,
933                (padding_cleared_until - start_offset).bytes(),
934                (offset - start_offset).bytes(),
935                size.bytes(),
936            );
937            if offset > padding_cleared_until {
938                // We found padding. Adjust the range to be relative to `alloc`, and make it uninit.
939                let padding_start = padding_cleared_until - start_offset;
940                let padding_size = offset - padding_cleared_until;
941                let range = alloc_range(padding_start, padding_size);
942                trace!("reset_padding on {}: resetting padding range {range:?}", mplace.layout.ty);
943                alloc.write_uninit(range)?;
944            }
945            padding_cleared_until = offset + size;
946        }
947        assert!(padding_cleared_until == start_offset + size);
948        interp_ok(())
949    }
950
951    /// Computes the data range of this union type:
952    /// which bytes are inside a field (i.e., not padding.)
953    fn union_data_range<'e>(
954        ecx: &'e mut InterpCx<'tcx, M>,
955        layout: TyAndLayout<'tcx>,
956    ) -> Cow<'e, RangeSet> {
957        assert!(layout.ty.is_union());
958        assert!(layout.is_sized(), "there are no unsized unions");
959        let layout_cx = LayoutCx::new(*ecx.tcx, ecx.typing_env);
960        return M::cached_union_data_range(ecx, layout.ty, || {
961            let mut out = RangeSet(Vec::new());
962            union_data_range_uncached(&layout_cx, layout, Size::ZERO, &mut out);
963            out
964        });
965
966        /// Helper for recursive traversal: add data ranges of the given type to `out`.
967        fn union_data_range_uncached<'tcx>(
968            cx: &LayoutCx<'tcx>,
969            layout: TyAndLayout<'tcx>,
970            base_offset: Size,
971            out: &mut RangeSet,
972        ) {
973            // If this is a ZST, we don't contain any data. In particular, this helps us to quickly
974            // skip over huge arrays of ZST.
975            if layout.is_zst() {
976                return;
977            }
978            // Just recursively add all the fields of everything to the output.
979            match &layout.fields {
980                FieldsShape::Primitive => {
981                    out.add_range(base_offset, layout.size);
982                }
983                &FieldsShape::Union(fields) => {
984                    // Currently, all fields start at offset 0 (relative to `base_offset`).
985                    for field in 0..fields.get() {
986                        let field = layout.field(cx, field);
987                        union_data_range_uncached(cx, field, base_offset, out);
988                    }
989                }
990                &FieldsShape::Array { stride, count } => {
991                    let elem = layout.field(cx, 0);
992
993                    // Fast-path for large arrays of simple types that do not contain any padding.
994                    if elem.backend_repr.is_scalar() {
995                        out.add_range(base_offset, elem.size * count);
996                    } else {
997                        for idx in 0..count {
998                            // This repeats the same computation for every array element... but the alternative
999                            // is to allocate temporary storage for a dedicated `out` set for the array element,
1000                            // and replicating that N times. Is that better?
1001                            union_data_range_uncached(cx, elem, base_offset + idx * stride, out);
1002                        }
1003                    }
1004                }
1005                FieldsShape::Arbitrary { offsets, .. } => {
1006                    for (field, &offset) in offsets.iter_enumerated() {
1007                        let field = layout.field(cx, field.as_usize());
1008                        union_data_range_uncached(cx, field, base_offset + offset, out);
1009                    }
1010                }
1011            }
1012            // Don't forget potential other variants.
1013            match &layout.variants {
1014                Variants::Single { .. } | Variants::Empty => {
1015                    // Fully handled above.
1016                }
1017                Variants::Multiple { variants, .. } => {
1018                    for variant in variants.indices() {
1019                        let variant = layout.for_variant(cx, variant);
1020                        union_data_range_uncached(cx, variant, base_offset, out);
1021                    }
1022                }
1023            }
1024        }
1025    }
1026}
1027
1028impl<'rt, 'tcx, M: Machine<'tcx>> ValueVisitor<'tcx, M> for ValidityVisitor<'rt, 'tcx, M> {
1029    type V = PlaceTy<'tcx, M::Provenance>;
1030
1031    #[inline(always)]
1032    fn ecx(&self) -> &InterpCx<'tcx, M> {
1033        self.ecx
1034    }
1035
1036    fn read_discriminant(
1037        &mut self,
1038        val: &PlaceTy<'tcx, M::Provenance>,
1039    ) -> InterpResult<'tcx, VariantIdx> {
1040        self.with_elem(PathElem::EnumTag, move |this| {
1041            interp_ok(try_validation!(
1042                this.ecx.read_discriminant(val),
1043                this.path,
1044                Ub(InvalidTag(val)) => InvalidEnumTag {
1045                    value: format!("{val:x}"),
1046                },
1047                Ub(UninhabitedEnumVariantRead(_)) => UninhabitedEnumVariant,
1048                // Uninit / bad provenance are not possible since the field was already previously
1049                // checked at its integer type.
1050            ))
1051        })
1052    }
1053
1054    #[inline]
1055    fn visit_field(
1056        &mut self,
1057        old_val: &PlaceTy<'tcx, M::Provenance>,
1058        field: usize,
1059        new_val: &PlaceTy<'tcx, M::Provenance>,
1060    ) -> InterpResult<'tcx> {
1061        let elem = self.aggregate_field_path_elem(old_val.layout, field);
1062        self.with_elem(elem, move |this| this.visit_value(new_val))
1063    }
1064
1065    #[inline]
1066    fn visit_variant(
1067        &mut self,
1068        old_val: &PlaceTy<'tcx, M::Provenance>,
1069        variant_id: VariantIdx,
1070        new_val: &PlaceTy<'tcx, M::Provenance>,
1071    ) -> InterpResult<'tcx> {
1072        let name = match old_val.layout.ty.kind() {
1073            ty::Adt(adt, _) => PathElem::Variant(adt.variant(variant_id).name),
1074            // Coroutines also have variants
1075            ty::Coroutine(..) => PathElem::CoroutineState(variant_id),
1076            _ => bug!("Unexpected type with variant: {:?}", old_val.layout.ty),
1077        };
1078        self.with_elem(name, move |this| this.visit_value(new_val))
1079    }
1080
1081    #[inline(always)]
1082    fn visit_union(
1083        &mut self,
1084        val: &PlaceTy<'tcx, M::Provenance>,
1085        _fields: NonZero<usize>,
1086    ) -> InterpResult<'tcx> {
1087        // Special check for CTFE validation, preventing `UnsafeCell` inside unions in immutable memory.
1088        if self.ctfe_mode.is_some_and(|c| !c.allow_immutable_unsafe_cell()) {
1089            // Unsized unions are currently not a thing, but let's keep this code consistent with
1090            // the check in `visit_value`.
1091            let zst = self.ecx.size_and_align_of_val(val)?.is_some_and(|(s, _a)| s.bytes() == 0);
1092            if !zst && !val.layout.ty.is_freeze(*self.ecx.tcx, self.ecx.typing_env) {
1093                if !self.in_mutable_memory(val) {
1094                    throw_validation_failure!(self.path, UnsafeCellInImmutable);
1095                }
1096            }
1097        }
1098        if self.reset_provenance_and_padding
1099            && let Some(data_bytes) = self.data_bytes.as_mut()
1100        {
1101            let base_offset = Self::data_range_offset(self.ecx, val);
1102            // Determine and add data range for this union.
1103            let union_data_range = Self::union_data_range(self.ecx, val.layout);
1104            for &(offset, size) in union_data_range.0.iter() {
1105                data_bytes.add_range(base_offset + offset, size);
1106            }
1107        }
1108        interp_ok(())
1109    }
1110
1111    #[inline]
1112    fn visit_box(
1113        &mut self,
1114        _box_ty: Ty<'tcx>,
1115        val: &PlaceTy<'tcx, M::Provenance>,
1116    ) -> InterpResult<'tcx> {
1117        self.check_safe_pointer(val, PointerKind::Box)?;
1118        interp_ok(())
1119    }
1120
1121    #[inline]
1122    fn visit_value(&mut self, val: &PlaceTy<'tcx, M::Provenance>) -> InterpResult<'tcx> {
1123        trace!("visit_value: {:?}, {:?}", *val, val.layout);
1124
1125        // Check primitive types -- the leaves of our recursive descent.
1126        // This is called even for enum discriminants (which are "fields" of their enum),
1127        // so for integer-typed discriminants the provenance reset will happen here.
1128        // We assume that the Scalar validity range does not restrict these values
1129        // any further than `try_visit_primitive` does!
1130        if self.try_visit_primitive(val)? {
1131            return interp_ok(());
1132        }
1133
1134        // Special check preventing `UnsafeCell` in the inner part of constants
1135        if self.ctfe_mode.is_some_and(|c| !c.allow_immutable_unsafe_cell()) {
1136            // Exclude ZST values. We need to compute the dynamic size/align to properly
1137            // handle slices and trait objects.
1138            let zst = self.ecx.size_and_align_of_val(val)?.is_some_and(|(s, _a)| s.bytes() == 0);
1139            if !zst
1140                && let Some(def) = val.layout.ty.ty_adt_def()
1141                && def.is_unsafe_cell()
1142            {
1143                if !self.in_mutable_memory(val) {
1144                    throw_validation_failure!(self.path, UnsafeCellInImmutable);
1145                }
1146            }
1147        }
1148
1149        // Recursively walk the value at its type. Apply optimizations for some large types.
1150        match val.layout.ty.kind() {
1151            ty::Str => {
1152                let mplace = val.assert_mem_place(); // strings are unsized and hence never immediate
1153                let len = mplace.len(self.ecx)?;
1154                try_validation!(
1155                    self.ecx.read_bytes_ptr_strip_provenance(mplace.ptr(), Size::from_bytes(len)),
1156                    self.path,
1157                    Ub(InvalidUninitBytes(..)) => Uninit { expected: ExpectedKind::Str },
1158                    Unsup(ReadPointerAsInt(_)) => PointerAsInt { expected: ExpectedKind::Str }
1159                );
1160            }
1161            ty::Array(tys, ..) | ty::Slice(tys)
1162                // This optimization applies for types that can hold arbitrary non-provenance bytes (such as
1163                // integer and floating point types).
1164                // FIXME(wesleywiser) This logic could be extended further to arbitrary structs or
1165                // tuples made up of integer/floating point types or inhabited ZSTs with no padding.
1166                if matches!(tys.kind(), ty::Int(..) | ty::Uint(..) | ty::Float(..))
1167                =>
1168            {
1169                let expected = if tys.is_integral() { ExpectedKind::Int } else { ExpectedKind::Float };
1170                // Optimized handling for arrays of integer/float type.
1171
1172                // This is the length of the array/slice.
1173                let len = val.len(self.ecx)?;
1174                // This is the element type size.
1175                let layout = self.ecx.layout_of(*tys)?;
1176                // This is the size in bytes of the whole array. (This checks for overflow.)
1177                let size = layout.size * len;
1178                // If the size is 0, there is nothing to check.
1179                // (`size` can only be 0 if `len` is 0, and empty arrays are always valid.)
1180                if size == Size::ZERO {
1181                    return interp_ok(());
1182                }
1183                // Now that we definitely have a non-ZST array, we know it lives in memory -- except it may
1184                // be an uninitialized local variable, those are also "immediate".
1185                let mplace = match val.to_op(self.ecx)?.as_mplace_or_imm() {
1186                    Left(mplace) => mplace,
1187                    Right(imm) => match *imm {
1188                        Immediate::Uninit =>
1189                            throw_validation_failure!(self.path, Uninit { expected }),
1190                        Immediate::Scalar(..) | Immediate::ScalarPair(..) =>
1191                            bug!("arrays/slices can never have Scalar/ScalarPair layout"),
1192                    }
1193                };
1194
1195                // Optimization: we just check the entire range at once.
1196                // NOTE: Keep this in sync with the handling of integer and float
1197                // types above, in `visit_primitive`.
1198                // No need for an alignment check here, this is not an actual memory access.
1199                let alloc = self.ecx.get_ptr_alloc(mplace.ptr(), size)?.expect("we already excluded size 0");
1200
1201                alloc.get_bytes_strip_provenance().map_err_kind(|kind| {
1202                    // Some error happened, try to provide a more detailed description.
1203                    // For some errors we might be able to provide extra information.
1204                    // (This custom logic does not fit the `try_validation!` macro.)
1205                    match kind {
1206                        Ub(InvalidUninitBytes(Some((_alloc_id, access)))) | Unsup(ReadPointerAsInt(Some((_alloc_id, access)))) => {
1207                            // Some byte was uninitialized, determine which
1208                            // element that byte belongs to so we can
1209                            // provide an index.
1210                            let i = usize::try_from(
1211                                access.bad.start.bytes() / layout.size.bytes(),
1212                            )
1213                            .unwrap();
1214                            self.path.push(PathElem::ArrayElem(i));
1215
1216                            if matches!(kind, Ub(InvalidUninitBytes(_))) {
1217                                err_validation_failure!(self.path, Uninit { expected })
1218                            } else {
1219                                err_validation_failure!(self.path, PointerAsInt { expected })
1220                            }
1221                        }
1222
1223                        // Propagate upwards (that will also check for unexpected errors).
1224                        err => err,
1225                    }
1226                })?;
1227
1228                // Don't forget that these are all non-pointer types, and thus do not preserve
1229                // provenance.
1230                if self.reset_provenance_and_padding {
1231                    // We can't share this with above as above, we might be looking at read-only memory.
1232                    let mut alloc = self.ecx.get_ptr_alloc_mut(mplace.ptr(), size)?.expect("we already excluded size 0");
1233                    alloc.clear_provenance()?;
1234                    // Also, mark this as containing data, not padding.
1235                    self.add_data_range(mplace.ptr(), size);
1236                }
1237            }
1238            // Fast path for arrays and slices of ZSTs. We only need to check a single ZST element
1239            // of an array and not all of them, because there's only a single value of a specific
1240            // ZST type, so either validation fails for all elements or none.
1241            ty::Array(tys, ..) | ty::Slice(tys) if self.ecx.layout_of(*tys)?.is_zst() => {
1242                // Validate just the first element (if any).
1243                if val.len(self.ecx)? > 0 {
1244                    self.visit_field(val, 0, &self.ecx.project_index(val, 0)?)?;
1245                }
1246            }
1247            ty::Pat(base, pat) => {
1248                // First check that the base type is valid
1249                self.visit_value(&val.transmute(self.ecx.layout_of(*base)?, self.ecx)?)?;
1250                // When you extend this match, make sure to also add tests to
1251                // tests/ui/type/pattern_types/validity.rs((
1252                match **pat {
1253                    // Range patterns are precisely reflected into `valid_range` and thus
1254                    // handled fully by `visit_scalar` (called below).
1255                    ty::PatternKind::Range { .. } => {},
1256
1257                    // FIXME(pattern_types): check that the value is covered by one of the variants.
1258                    // For now, we rely on layout computation setting the scalar's `valid_range` to
1259                    // match the pattern. However, this cannot always work; the layout may
1260                    // pessimistically cover actually illegal ranges and Miri would miss that UB.
1261                    // The consolation here is that codegen also will miss that UB, so at least
1262                    // we won't see optimizations actually breaking such programs.
1263                    ty::PatternKind::Or(_patterns) => {}
1264                }
1265            }
1266            _ => {
1267                // default handler
1268                try_validation!(
1269                    self.walk_value(val),
1270                    self.path,
1271                    // It's not great to catch errors here, since we can't give a very good path,
1272                    // but it's better than ICEing.
1273                    Ub(InvalidVTableTrait { vtable_dyn_type, expected_dyn_type }) => {
1274                        InvalidMetaWrongTrait { vtable_dyn_type, expected_dyn_type }
1275                    },
1276                );
1277            }
1278        }
1279
1280        // *After* all of this, check further information stored in the layout. We need to check
1281        // this to handle types like `NonNull` where the `Scalar` info is more restrictive than what
1282        // the fields say (`rustc_layout_scalar_valid_range_start`). But in most cases, this will
1283        // just propagate what the fields say, and then we want the error to point at the field --
1284        // so, we first recurse, then we do this check.
1285        //
1286        // FIXME: We could avoid some redundant checks here. For newtypes wrapping
1287        // scalars, we do the same check on every "level" (e.g., first we check
1288        // MyNewtype and then the scalar in there).
1289        if val.layout.is_uninhabited() {
1290            let ty = val.layout.ty;
1291            throw_validation_failure!(self.path, UninhabitedVal { ty });
1292        }
1293        match val.layout.backend_repr {
1294            BackendRepr::Scalar(scalar_layout) => {
1295                if !scalar_layout.is_uninit_valid() {
1296                    // There is something to check here.
1297                    let scalar = self.read_scalar(val, ExpectedKind::InitScalar)?;
1298                    self.visit_scalar(scalar, scalar_layout)?;
1299                }
1300            }
1301            BackendRepr::ScalarPair(a_layout, b_layout) => {
1302                // We can only proceed if *both* scalars need to be initialized.
1303                // FIXME: find a way to also check ScalarPair when one side can be uninit but
1304                // the other must be init.
1305                if !a_layout.is_uninit_valid() && !b_layout.is_uninit_valid() {
1306                    let (a, b) =
1307                        self.read_immediate(val, ExpectedKind::InitScalar)?.to_scalar_pair();
1308                    self.visit_scalar(a, a_layout)?;
1309                    self.visit_scalar(b, b_layout)?;
1310                }
1311            }
1312            BackendRepr::SimdVector { .. } => {
1313                // No checks here, we assume layout computation gets this right.
1314                // (This is harder to check since Miri does not represent these as `Immediate`. We
1315                // also cannot use field projections since this might be a newtype around a vector.)
1316            }
1317            BackendRepr::Memory { .. } => {
1318                // Nothing to do.
1319            }
1320        }
1321
1322        interp_ok(())
1323    }
1324}
1325
1326impl<'tcx, M: Machine<'tcx>> InterpCx<'tcx, M> {
1327    fn validate_operand_internal(
1328        &mut self,
1329        val: &PlaceTy<'tcx, M::Provenance>,
1330        path: Vec<PathElem>,
1331        ref_tracking: Option<&mut RefTracking<MPlaceTy<'tcx, M::Provenance>, Vec<PathElem>>>,
1332        ctfe_mode: Option<CtfeValidationMode>,
1333        reset_provenance_and_padding: bool,
1334    ) -> InterpResult<'tcx> {
1335        trace!("validate_operand_internal: {:?}, {:?}", *val, val.layout.ty);
1336
1337        // Run the visitor.
1338        self.run_for_validation_mut(|ecx| {
1339            let reset_padding = reset_provenance_and_padding && {
1340                // Check if `val` is actually stored in memory. If not, padding is not even
1341                // represented and we need not reset it.
1342                ecx.place_to_op(val)?.as_mplace_or_imm().is_left()
1343            };
1344            let mut v = ValidityVisitor {
1345                path,
1346                ref_tracking,
1347                ctfe_mode,
1348                ecx,
1349                reset_provenance_and_padding,
1350                data_bytes: reset_padding.then_some(RangeSet(Vec::new())),
1351            };
1352            v.visit_value(val)?;
1353            v.reset_padding(val)?;
1354            interp_ok(())
1355        })
1356        .map_err_info(|err| {
1357            if !matches!(
1358                err.kind(),
1359                err_ub!(ValidationError { .. })
1360                    | InterpErrorKind::InvalidProgram(_)
1361                    | InterpErrorKind::Unsupported(UnsupportedOpInfo::ExternTypeField)
1362            ) {
1363                bug!(
1364                    "Unexpected error during validation: {}",
1365                    format_interp_error(self.tcx.dcx(), err)
1366                );
1367            }
1368            err
1369        })
1370    }
1371
1372    /// This function checks the data at `val` to be const-valid.
1373    /// `val` is assumed to cover valid memory if it is an indirect operand.
1374    /// It will error if the bits at the destination do not match the ones described by the layout.
1375    ///
1376    /// `ref_tracking` is used to record references that we encounter so that they
1377    /// can be checked recursively by an outside driving loop.
1378    ///
1379    /// `constant` controls whether this must satisfy the rules for constants:
1380    /// - no pointers to statics.
1381    /// - no `UnsafeCell` or non-ZST `&mut`.
1382    #[inline(always)]
1383    pub(crate) fn const_validate_operand(
1384        &mut self,
1385        val: &PlaceTy<'tcx, M::Provenance>,
1386        path: Vec<PathElem>,
1387        ref_tracking: &mut RefTracking<MPlaceTy<'tcx, M::Provenance>, Vec<PathElem>>,
1388        ctfe_mode: CtfeValidationMode,
1389    ) -> InterpResult<'tcx> {
1390        self.validate_operand_internal(
1391            val,
1392            path,
1393            Some(ref_tracking),
1394            Some(ctfe_mode),
1395            /*reset_provenance*/ false,
1396        )
1397    }
1398
1399    /// This function checks the data at `val` to be runtime-valid.
1400    /// `val` is assumed to cover valid memory if it is an indirect operand.
1401    /// It will error if the bits at the destination do not match the ones described by the layout.
1402    #[inline(always)]
1403    pub fn validate_operand(
1404        &mut self,
1405        val: &PlaceTy<'tcx, M::Provenance>,
1406        recursive: bool,
1407        reset_provenance_and_padding: bool,
1408    ) -> InterpResult<'tcx> {
1409        let _span = enter_trace_span!(
1410            M,
1411            "validate_operand",
1412            "recursive={recursive}, reset_provenance_and_padding={reset_provenance_and_padding}, val={val:?}"
1413        );
1414
1415        // Note that we *could* actually be in CTFE here with `-Zextra-const-ub-checks`, but it's
1416        // still correct to not use `ctfe_mode`: that mode is for validation of the final constant
1417        // value, it rules out things like `UnsafeCell` in awkward places.
1418        if !recursive {
1419            return self.validate_operand_internal(
1420                val,
1421                vec![],
1422                None,
1423                None,
1424                reset_provenance_and_padding,
1425            );
1426        }
1427        // Do a recursive check.
1428        let mut ref_tracking = RefTracking::empty();
1429        self.validate_operand_internal(
1430            val,
1431            vec![],
1432            Some(&mut ref_tracking),
1433            None,
1434            reset_provenance_and_padding,
1435        )?;
1436        while let Some((mplace, path)) = ref_tracking.todo.pop() {
1437            // Things behind reference do *not* have the provenance reset.
1438            self.validate_operand_internal(
1439                &mplace.into(),
1440                path,
1441                Some(&mut ref_tracking),
1442                None,
1443                /*reset_provenance_and_padding*/ false,
1444            )?;
1445        }
1446        interp_ok(())
1447    }
1448}