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 traits
362 ty::Dynamic(..) => {
363 assert_eq!(field, 0);
364 PathElem::DynDowncast
365 }
366
367 // nothing else has an aggregate layout
368 _ => bug!("aggregate_field_path_elem: got non-aggregate type {:?}", layout.ty),
369 }
370 }
371
372 fn with_elem<R>(
373 &mut self,
374 elem: PathElem,
375 f: impl FnOnce(&mut Self) -> InterpResult<'tcx, R>,
376 ) -> InterpResult<'tcx, R> {
377 // Remember the old state
378 let path_len = self.path.len();
379 // Record new element
380 self.path.push(elem);
381 // Perform operation
382 let r = f(self)?;
383 // Undo changes
384 self.path.truncate(path_len);
385 // Done
386 interp_ok(r)
387 }
388
389 fn read_immediate(
390 &self,
391 val: &PlaceTy<'tcx, M::Provenance>,
392 expected: ExpectedKind,
393 ) -> InterpResult<'tcx, ImmTy<'tcx, M::Provenance>> {
394 interp_ok(try_validation!(
395 self.ecx.read_immediate(val),
396 self.path,
397 Ub(InvalidUninitBytes(None)) =>
398 Uninit { expected },
399 // The `Unsup` cases can only occur during CTFE
400 Unsup(ReadPointerAsInt(_)) =>
401 PointerAsInt { expected },
402 Unsup(ReadPartialPointer(_)) =>
403 PartialPointer,
404 ))
405 }
406
407 fn read_scalar(
408 &self,
409 val: &PlaceTy<'tcx, M::Provenance>,
410 expected: ExpectedKind,
411 ) -> InterpResult<'tcx, Scalar<M::Provenance>> {
412 interp_ok(self.read_immediate(val, expected)?.to_scalar())
413 }
414
415 fn deref_pointer(
416 &mut self,
417 val: &PlaceTy<'tcx, M::Provenance>,
418 expected: ExpectedKind,
419 ) -> InterpResult<'tcx, MPlaceTy<'tcx, M::Provenance>> {
420 // Not using `ecx.deref_pointer` since we want to use our `read_immediate` wrapper.
421 let imm = self.read_immediate(val, expected)?;
422 // Reset provenance: ensure slice tail metadata does not preserve provenance,
423 // and ensure all pointers do not preserve partial provenance.
424 if self.reset_provenance_and_padding {
425 if matches!(imm.layout.backend_repr, BackendRepr::Scalar(..)) {
426 // A thin pointer. If it has provenance, we don't have to do anything.
427 // If it does not, ensure we clear the provenance in memory.
428 if matches!(imm.to_scalar(), Scalar::Int(..)) {
429 self.ecx.clear_provenance(val)?;
430 }
431 } else {
432 // A wide pointer. This means we have to worry both about the pointer itself and the
433 // metadata. We do the lazy thing and just write back the value we got. Just
434 // clearing provenance in a targeted manner would be more efficient, but unless this
435 // is a perf hotspot it's just not worth the effort.
436 self.ecx.write_immediate_no_validate(*imm, val)?;
437 }
438 // The entire thing is data, not padding.
439 self.add_data_range_place(val);
440 }
441 // Now turn it into a place.
442 self.ecx.ref_to_mplace(&imm)
443 }
444
445 fn check_wide_ptr_meta(
446 &mut self,
447 meta: MemPlaceMeta<M::Provenance>,
448 pointee: TyAndLayout<'tcx>,
449 ) -> InterpResult<'tcx> {
450 let tail = self.ecx.tcx.struct_tail_for_codegen(pointee.ty, self.ecx.typing_env);
451 match tail.kind() {
452 ty::Dynamic(data, _, ty::Dyn) => {
453 let vtable = meta.unwrap_meta().to_pointer(self.ecx)?;
454 // Make sure it is a genuine vtable pointer for the right trait.
455 try_validation!(
456 self.ecx.get_ptr_vtable_ty(vtable, Some(data)),
457 self.path,
458 Ub(DanglingIntPointer{ .. } | InvalidVTablePointer(..)) =>
459 InvalidVTablePtr { value: format!("{vtable}") },
460 Ub(InvalidVTableTrait { vtable_dyn_type, expected_dyn_type }) => {
461 InvalidMetaWrongTrait { vtable_dyn_type, expected_dyn_type }
462 },
463 );
464 }
465 ty::Slice(..) | ty::Str => {
466 let _len = meta.unwrap_meta().to_target_usize(self.ecx)?;
467 // We do not check that `len * elem_size <= isize::MAX`:
468 // that is only required for references, and there it falls out of the
469 // "dereferenceable" check performed by Stacked Borrows.
470 }
471 ty::Foreign(..) => {
472 // Unsized, but not wide.
473 }
474 _ => bug!("Unexpected unsized type tail: {:?}", tail),
475 }
476
477 interp_ok(())
478 }
479
480 /// Check a reference or `Box`.
481 fn check_safe_pointer(
482 &mut self,
483 value: &PlaceTy<'tcx, M::Provenance>,
484 ptr_kind: PointerKind,
485 ) -> InterpResult<'tcx> {
486 let place = self.deref_pointer(value, ptr_kind.into())?;
487 // Handle wide pointers.
488 // Check metadata early, for better diagnostics
489 if place.layout.is_unsized() {
490 self.check_wide_ptr_meta(place.meta(), place.layout)?;
491 }
492 // Make sure this is dereferenceable and all.
493 let size_and_align = try_validation!(
494 self.ecx.size_and_align_of_val(&place),
495 self.path,
496 Ub(InvalidMeta(msg)) => match msg {
497 InvalidMetaKind::SliceTooBig => InvalidMetaSliceTooLarge { ptr_kind },
498 InvalidMetaKind::TooBig => InvalidMetaTooLarge { ptr_kind },
499 }
500 );
501 let (size, align) = size_and_align
502 // for the purpose of validity, consider foreign types to have
503 // alignment and size determined by the layout (size will be 0,
504 // alignment should take attributes into account).
505 .unwrap_or_else(|| (place.layout.size, place.layout.align.abi));
506 // Direct call to `check_ptr_access_align` checks alignment even on CTFE machines.
507 try_validation!(
508 self.ecx.check_ptr_access(
509 place.ptr(),
510 size,
511 CheckInAllocMsg::Dereferenceable, // will anyway be replaced by validity message
512 ),
513 self.path,
514 Ub(DanglingIntPointer { addr: 0, .. }) => NullPtr { ptr_kind },
515 Ub(DanglingIntPointer { addr: i, .. }) => DanglingPtrNoProvenance {
516 ptr_kind,
517 // FIXME this says "null pointer" when null but we need translate
518 pointer: format!("{}", Pointer::<Option<AllocId>>::without_provenance(i))
519 },
520 Ub(PointerOutOfBounds { .. }) => DanglingPtrOutOfBounds {
521 ptr_kind
522 },
523 Ub(PointerUseAfterFree(..)) => DanglingPtrUseAfterFree {
524 ptr_kind,
525 },
526 );
527 try_validation!(
528 self.ecx.check_ptr_align(
529 place.ptr(),
530 align,
531 ),
532 self.path,
533 Ub(AlignmentCheckFailed(Misalignment { required, has }, _msg)) => UnalignedPtr {
534 ptr_kind,
535 required_bytes: required.bytes(),
536 found_bytes: has.bytes()
537 },
538 );
539 // Make sure this is non-null. We checked dereferenceability above, but if `size` is zero
540 // that does not imply non-null.
541 if self.ecx.scalar_may_be_null(Scalar::from_maybe_pointer(place.ptr(), self.ecx))? {
542 throw_validation_failure!(self.path, NullPtr { ptr_kind })
543 }
544 // Do not allow references to uninhabited types.
545 if place.layout.is_uninhabited() {
546 let ty = place.layout.ty;
547 throw_validation_failure!(self.path, PtrToUninhabited { ptr_kind, ty })
548 }
549 // Recursive checking
550 if let Some(ref_tracking) = self.ref_tracking.as_deref_mut() {
551 // Proceed recursively even for ZST, no reason to skip them!
552 // `!` is a ZST and we want to validate it.
553 if let Some(ctfe_mode) = self.ctfe_mode {
554 let mut skip_recursive_check = false;
555 // CTFE imposes restrictions on what references can point to.
556 if let Ok((alloc_id, _offset, _prov)) =
557 self.ecx.ptr_try_get_alloc_id(place.ptr(), 0)
558 {
559 // Everything should be already interned.
560 let Some(global_alloc) = self.ecx.tcx.try_get_global_alloc(alloc_id) else {
561 assert!(self.ecx.memory.alloc_map.get(alloc_id).is_none());
562 // We can't have *any* references to non-existing allocations in const-eval
563 // as the rest of rustc isn't happy with them... so we throw an error, even
564 // though for zero-sized references this isn't really UB.
565 // A potential future alternative would be to resurrect this as a zero-sized allocation
566 // (which codegen will then compile to an aligned dummy pointer anyway).
567 throw_validation_failure!(self.path, DanglingPtrUseAfterFree { ptr_kind });
568 };
569 let (size, _align) =
570 global_alloc.size_and_align(*self.ecx.tcx, self.ecx.typing_env);
571 let alloc_actual_mutbl =
572 global_alloc.mutability(*self.ecx.tcx, self.ecx.typing_env);
573
574 if let GlobalAlloc::Static(did) = global_alloc {
575 let DefKind::Static { nested, .. } = self.ecx.tcx.def_kind(did) else {
576 bug!()
577 };
578 // Special handling for pointers to statics (irrespective of their type).
579 assert!(!self.ecx.tcx.is_thread_local_static(did));
580 assert!(self.ecx.tcx.is_static(did));
581 // Mode-specific checks
582 match ctfe_mode {
583 CtfeValidationMode::Static { .. }
584 | CtfeValidationMode::Promoted { .. } => {
585 // We skip recursively checking other statics. These statics must be sound by
586 // themselves, and the only way to get broken statics here is by using
587 // unsafe code.
588 // The reasons we don't check other statics is twofold. For one, in all
589 // sound cases, the static was already validated on its own, and second, we
590 // trigger cycle errors if we try to compute the value of the other static
591 // and that static refers back to us (potentially through a promoted).
592 // This could miss some UB, but that's fine.
593 // We still walk nested allocations, as they are fundamentally part of this validation run.
594 // This means we will also recurse into nested statics of *other*
595 // statics, even though we do not recurse into other statics directly.
596 // That's somewhat inconsistent but harmless.
597 skip_recursive_check = !nested;
598 }
599 CtfeValidationMode::Const { .. } => {
600 // If this is mutable memory or an `extern static`, there's no point in checking it -- we'd
601 // just get errors trying to read the value.
602 if alloc_actual_mutbl.is_mut() || self.ecx.tcx.is_foreign_item(did)
603 {
604 skip_recursive_check = true;
605 }
606 }
607 }
608 }
609
610 // If this allocation has size zero, there is no actual mutability here.
611 if size != Size::ZERO {
612 // Determine whether this pointer expects to be pointing to something mutable.
613 let ptr_expected_mutbl = match ptr_kind {
614 PointerKind::Box => Mutability::Mut,
615 PointerKind::Ref(mutbl) => {
616 // We do not take into account interior mutability here since we cannot know if
617 // there really is an `UnsafeCell` inside `Option<UnsafeCell>` -- so we check
618 // that in the recursive descent behind this reference (controlled by
619 // `allow_immutable_unsafe_cell`).
620 mutbl
621 }
622 };
623 // Mutable pointer to immutable memory is no good.
624 if ptr_expected_mutbl == Mutability::Mut
625 && alloc_actual_mutbl == Mutability::Not
626 {
627 // This can actually occur with transmutes.
628 throw_validation_failure!(self.path, MutableRefToImmutable);
629 }
630 // In a const, any kind of mutable reference is not good.
631 if matches!(self.ctfe_mode, Some(CtfeValidationMode::Const { .. })) {
632 if ptr_expected_mutbl == Mutability::Mut {
633 throw_validation_failure!(self.path, MutableRefInConst);
634 }
635 }
636 }
637 }
638 // Potentially skip recursive check.
639 if skip_recursive_check {
640 return interp_ok(());
641 }
642 } else {
643 // This is not CTFE, so it's Miri with recursive checking.
644 // FIXME: we do *not* check behind boxes, since creating a new box first creates it uninitialized
645 // and then puts the value in there, so briefly we have a box with uninit contents.
646 // FIXME: should we also skip `UnsafeCell` behind shared references? Currently that is not
647 // needed since validation reads bypass Stacked Borrows and data race checks.
648 if matches!(ptr_kind, PointerKind::Box) {
649 return interp_ok(());
650 }
651 }
652 let path = &self.path;
653 ref_tracking.track(place, || {
654 // We need to clone the path anyway, make sure it gets created
655 // with enough space for the additional `Deref`.
656 let mut new_path = Vec::with_capacity(path.len() + 1);
657 new_path.extend(path);
658 new_path.push(PathElem::Deref);
659 new_path
660 });
661 }
662 interp_ok(())
663 }
664
665 /// Check if this is a value of primitive type, and if yes check the validity of the value
666 /// at that type. Return `true` if the type is indeed primitive.
667 ///
668 /// Note that not all of these have `FieldsShape::Primitive`, e.g. wide references.
669 fn try_visit_primitive(
670 &mut self,
671 value: &PlaceTy<'tcx, M::Provenance>,
672 ) -> InterpResult<'tcx, bool> {
673 // Go over all the primitive types
674 let ty = value.layout.ty;
675 match ty.kind() {
676 ty::Bool => {
677 let scalar = self.read_scalar(value, ExpectedKind::Bool)?;
678 try_validation!(
679 scalar.to_bool(),
680 self.path,
681 Ub(InvalidBool(..)) => ValidationErrorKind::InvalidBool {
682 value: format!("{scalar:x}"),
683 }
684 );
685 if self.reset_provenance_and_padding {
686 self.ecx.clear_provenance(value)?;
687 self.add_data_range_place(value);
688 }
689 interp_ok(true)
690 }
691 ty::Char => {
692 let scalar = self.read_scalar(value, ExpectedKind::Char)?;
693 try_validation!(
694 scalar.to_char(),
695 self.path,
696 Ub(InvalidChar(..)) => ValidationErrorKind::InvalidChar {
697 value: format!("{scalar:x}"),
698 }
699 );
700 if self.reset_provenance_and_padding {
701 self.ecx.clear_provenance(value)?;
702 self.add_data_range_place(value);
703 }
704 interp_ok(true)
705 }
706 ty::Float(_) | ty::Int(_) | ty::Uint(_) => {
707 // NOTE: Keep this in sync with the array optimization for int/float
708 // types below!
709 self.read_scalar(
710 value,
711 if matches!(ty.kind(), ty::Float(..)) {
712 ExpectedKind::Float
713 } else {
714 ExpectedKind::Int
715 },
716 )?;
717 if self.reset_provenance_and_padding {
718 self.ecx.clear_provenance(value)?;
719 self.add_data_range_place(value);
720 }
721 interp_ok(true)
722 }
723 ty::RawPtr(..) => {
724 let place = self.deref_pointer(value, ExpectedKind::RawPtr)?;
725 if place.layout.is_unsized() {
726 self.check_wide_ptr_meta(place.meta(), place.layout)?;
727 }
728 interp_ok(true)
729 }
730 ty::Ref(_, _ty, mutbl) => {
731 self.check_safe_pointer(value, PointerKind::Ref(*mutbl))?;
732 interp_ok(true)
733 }
734 ty::FnPtr(..) => {
735 let scalar = self.read_scalar(value, ExpectedKind::FnPtr)?;
736
737 // If we check references recursively, also check that this points to a function.
738 if let Some(_) = self.ref_tracking {
739 let ptr = scalar.to_pointer(self.ecx)?;
740 let _fn = try_validation!(
741 self.ecx.get_ptr_fn(ptr),
742 self.path,
743 Ub(DanglingIntPointer{ .. } | InvalidFunctionPointer(..)) =>
744 InvalidFnPtr { value: format!("{ptr}") },
745 );
746 // FIXME: Check if the signature matches
747 } else {
748 // Otherwise (for standalone Miri), we have to still check it to be non-null.
749 if self.ecx.scalar_may_be_null(scalar)? {
750 throw_validation_failure!(self.path, NullFnPtr);
751 }
752 }
753 if self.reset_provenance_and_padding {
754 // Make sure we do not preserve partial provenance. This matches the thin
755 // pointer handling in `deref_pointer`.
756 if matches!(scalar, Scalar::Int(..)) {
757 self.ecx.clear_provenance(value)?;
758 }
759 self.add_data_range_place(value);
760 }
761 interp_ok(true)
762 }
763 ty::Never => throw_validation_failure!(self.path, NeverVal),
764 ty::Foreign(..) | ty::FnDef(..) => {
765 // Nothing to check.
766 interp_ok(true)
767 }
768 ty::UnsafeBinder(_) => todo!("FIXME(unsafe_binder)"),
769 // The above should be all the primitive types. The rest is compound, we
770 // check them by visiting their fields/variants.
771 ty::Adt(..)
772 | ty::Tuple(..)
773 | ty::Array(..)
774 | ty::Slice(..)
775 | ty::Str
776 | ty::Dynamic(..)
777 | ty::Closure(..)
778 | ty::Pat(..)
779 | ty::CoroutineClosure(..)
780 | ty::Coroutine(..) => interp_ok(false),
781 // Some types only occur during typechecking, they have no layout.
782 // We should not see them here and we could not check them anyway.
783 ty::Error(_)
784 | ty::Infer(..)
785 | ty::Placeholder(..)
786 | ty::Bound(..)
787 | ty::Param(..)
788 | ty::Alias(..)
789 | ty::CoroutineWitness(..) => bug!("Encountered invalid type {:?}", ty),
790 }
791 }
792
793 fn visit_scalar(
794 &mut self,
795 scalar: Scalar<M::Provenance>,
796 scalar_layout: ScalarAbi,
797 ) -> InterpResult<'tcx> {
798 let size = scalar_layout.size(self.ecx);
799 let valid_range = scalar_layout.valid_range(self.ecx);
800 let WrappingRange { start, end } = valid_range;
801 let max_value = size.unsigned_int_max();
802 assert!(end <= max_value);
803 let bits = match scalar.try_to_scalar_int() {
804 Ok(int) => int.to_bits(size),
805 Err(_) => {
806 // So this is a pointer then, and casting to an int failed.
807 // Can only happen during CTFE.
808 // We support 2 kinds of ranges here: full range, and excluding zero.
809 if start == 1 && end == max_value {
810 // Only null is the niche. So make sure the ptr is NOT null.
811 if self.ecx.scalar_may_be_null(scalar)? {
812 throw_validation_failure!(
813 self.path,
814 NullablePtrOutOfRange { range: valid_range, max_value }
815 )
816 } else {
817 return interp_ok(());
818 }
819 } else if scalar_layout.is_always_valid(self.ecx) {
820 // Easy. (This is reachable if `enforce_number_validity` is set.)
821 return interp_ok(());
822 } else {
823 // Conservatively, we reject, because the pointer *could* have a bad
824 // value.
825 throw_validation_failure!(
826 self.path,
827 PtrOutOfRange { range: valid_range, max_value }
828 )
829 }
830 }
831 };
832 // Now compare.
833 if valid_range.contains(bits) {
834 interp_ok(())
835 } else {
836 throw_validation_failure!(
837 self.path,
838 OutOfRange { value: format!("{bits}"), range: valid_range, max_value }
839 )
840 }
841 }
842
843 fn in_mutable_memory(&self, val: &PlaceTy<'tcx, M::Provenance>) -> bool {
844 debug_assert!(self.ctfe_mode.is_some());
845 if let Some(mplace) = val.as_mplace_or_local().left() {
846 if let Some(alloc_id) = mplace.ptr().provenance.and_then(|p| p.get_alloc_id()) {
847 let tcx = *self.ecx.tcx;
848 // Everything must be already interned.
849 let mutbl = tcx.global_alloc(alloc_id).mutability(tcx, self.ecx.typing_env);
850 if let Some((_, alloc)) = self.ecx.memory.alloc_map.get(alloc_id) {
851 assert_eq!(alloc.mutability, mutbl);
852 }
853 mutbl.is_mut()
854 } else {
855 // No memory at all.
856 false
857 }
858 } else {
859 // A local variable -- definitely mutable.
860 true
861 }
862 }
863
864 /// Add the given pointer-length pair to the "data" range of this visit.
865 fn add_data_range(&mut self, ptr: Pointer<Option<M::Provenance>>, size: Size) {
866 if let Some(data_bytes) = self.data_bytes.as_mut() {
867 // We only have to store the offset, the rest is the same for all pointers here.
868 // The logic is agnostic to whether the offset is relative or absolute as long as
869 // it is consistent.
870 let (_prov, offset) = ptr.into_raw_parts();
871 // Add this.
872 data_bytes.add_range(offset, size);
873 };
874 }
875
876 /// Add the entire given place to the "data" range of this visit.
877 fn add_data_range_place(&mut self, place: &PlaceTy<'tcx, M::Provenance>) {
878 // Only sized places can be added this way.
879 debug_assert!(place.layout.is_sized());
880 if let Some(data_bytes) = self.data_bytes.as_mut() {
881 let offset = Self::data_range_offset(self.ecx, place);
882 data_bytes.add_range(offset, place.layout.size);
883 }
884 }
885
886 /// Convert a place into the offset it starts at, for the purpose of data_range tracking.
887 /// Must only be called if `data_bytes` is `Some(_)`.
888 fn data_range_offset(ecx: &InterpCx<'tcx, M>, place: &PlaceTy<'tcx, M::Provenance>) -> Size {
889 // The presence of `data_bytes` implies that our place is in memory.
890 let ptr = ecx
891 .place_to_op(place)
892 .expect("place must be in memory")
893 .as_mplace_or_imm()
894 .expect_left("place must be in memory")
895 .ptr();
896 let (_prov, offset) = ptr.into_raw_parts();
897 offset
898 }
899
900 fn reset_padding(&mut self, place: &PlaceTy<'tcx, M::Provenance>) -> InterpResult<'tcx> {
901 let Some(data_bytes) = self.data_bytes.as_mut() else { return interp_ok(()) };
902 // Our value must be in memory, otherwise we would not have set up `data_bytes`.
903 let mplace = self.ecx.force_allocation(place)?;
904 // Determine starting offset and size.
905 let (_prov, start_offset) = mplace.ptr().into_raw_parts();
906 let (size, _align) = self
907 .ecx
908 .size_and_align_of_val(&mplace)?
909 .unwrap_or((mplace.layout.size, mplace.layout.align.abi));
910 // If there is no padding at all, we can skip the rest: check for
911 // a single data range covering the entire value.
912 if data_bytes.0 == &[(start_offset, size)] {
913 return interp_ok(());
914 }
915 // Get a handle for the allocation. Do this only once, to avoid looking up the same
916 // allocation over and over again. (Though to be fair, iterating the value already does
917 // exactly that.)
918 let Some(mut alloc) = self.ecx.get_ptr_alloc_mut(mplace.ptr(), size)? else {
919 // A ZST, no padding to clear.
920 return interp_ok(());
921 };
922 // Add a "finalizer" data range at the end, so that the iteration below finds all gaps
923 // between ranges.
924 data_bytes.0.push((start_offset + size, Size::ZERO));
925 // Iterate, and reset gaps.
926 let mut padding_cleared_until = start_offset;
927 for &(offset, size) in data_bytes.0.iter() {
928 assert!(
929 offset >= padding_cleared_until,
930 "reset_padding on {}: previous field ended at offset {}, next field starts at {} (and has a size of {} bytes)",
931 mplace.layout.ty,
932 (padding_cleared_until - start_offset).bytes(),
933 (offset - start_offset).bytes(),
934 size.bytes(),
935 );
936 if offset > padding_cleared_until {
937 // We found padding. Adjust the range to be relative to `alloc`, and make it uninit.
938 let padding_start = padding_cleared_until - start_offset;
939 let padding_size = offset - padding_cleared_until;
940 let range = alloc_range(padding_start, padding_size);
941 trace!("reset_padding on {}: resetting padding range {range:?}", mplace.layout.ty);
942 alloc.write_uninit(range)?;
943 }
944 padding_cleared_until = offset + size;
945 }
946 assert!(padding_cleared_until == start_offset + size);
947 interp_ok(())
948 }
949
950 /// Computes the data range of this union type:
951 /// which bytes are inside a field (i.e., not padding.)
952 fn union_data_range<'e>(
953 ecx: &'e mut InterpCx<'tcx, M>,
954 layout: TyAndLayout<'tcx>,
955 ) -> Cow<'e, RangeSet> {
956 assert!(layout.ty.is_union());
957 assert!(layout.is_sized(), "there are no unsized unions");
958 let layout_cx = LayoutCx::new(*ecx.tcx, ecx.typing_env);
959 return M::cached_union_data_range(ecx, layout.ty, || {
960 let mut out = RangeSet(Vec::new());
961 union_data_range_uncached(&layout_cx, layout, Size::ZERO, &mut out);
962 out
963 });
964
965 /// Helper for recursive traversal: add data ranges of the given type to `out`.
966 fn union_data_range_uncached<'tcx>(
967 cx: &LayoutCx<'tcx>,
968 layout: TyAndLayout<'tcx>,
969 base_offset: Size,
970 out: &mut RangeSet,
971 ) {
972 // If this is a ZST, we don't contain any data. In particular, this helps us to quickly
973 // skip over huge arrays of ZST.
974 if layout.is_zst() {
975 return;
976 }
977 // Just recursively add all the fields of everything to the output.
978 match &layout.fields {
979 FieldsShape::Primitive => {
980 out.add_range(base_offset, layout.size);
981 }
982 &FieldsShape::Union(fields) => {
983 // Currently, all fields start at offset 0 (relative to `base_offset`).
984 for field in 0..fields.get() {
985 let field = layout.field(cx, field);
986 union_data_range_uncached(cx, field, base_offset, out);
987 }
988 }
989 &FieldsShape::Array { stride, count } => {
990 let elem = layout.field(cx, 0);
991
992 // Fast-path for large arrays of simple types that do not contain any padding.
993 if elem.backend_repr.is_scalar() {
994 out.add_range(base_offset, elem.size * count);
995 } else {
996 for idx in 0..count {
997 // This repeats the same computation for every array element... but the alternative
998 // is to allocate temporary storage for a dedicated `out` set for the array element,
999 // and replicating that N times. Is that better?
1000 union_data_range_uncached(cx, elem, base_offset + idx * stride, out);
1001 }
1002 }
1003 }
1004 FieldsShape::Arbitrary { offsets, .. } => {
1005 for (field, &offset) in offsets.iter_enumerated() {
1006 let field = layout.field(cx, field.as_usize());
1007 union_data_range_uncached(cx, field, base_offset + offset, out);
1008 }
1009 }
1010 }
1011 // Don't forget potential other variants.
1012 match &layout.variants {
1013 Variants::Single { .. } | Variants::Empty => {
1014 // Fully handled above.
1015 }
1016 Variants::Multiple { variants, .. } => {
1017 for variant in variants.indices() {
1018 let variant = layout.for_variant(cx, variant);
1019 union_data_range_uncached(cx, variant, base_offset, out);
1020 }
1021 }
1022 }
1023 }
1024 }
1025}
1026
1027impl<'rt, 'tcx, M: Machine<'tcx>> ValueVisitor<'tcx, M> for ValidityVisitor<'rt, 'tcx, M> {
1028 type V = PlaceTy<'tcx, M::Provenance>;
1029
1030 #[inline(always)]
1031 fn ecx(&self) -> &InterpCx<'tcx, M> {
1032 self.ecx
1033 }
1034
1035 fn read_discriminant(
1036 &mut self,
1037 val: &PlaceTy<'tcx, M::Provenance>,
1038 ) -> InterpResult<'tcx, VariantIdx> {
1039 self.with_elem(PathElem::EnumTag, move |this| {
1040 interp_ok(try_validation!(
1041 this.ecx.read_discriminant(val),
1042 this.path,
1043 Ub(InvalidTag(val)) => InvalidEnumTag {
1044 value: format!("{val:x}"),
1045 },
1046 Ub(UninhabitedEnumVariantRead(_)) => UninhabitedEnumVariant,
1047 // Uninit / bad provenance are not possible since the field was already previously
1048 // checked at its integer type.
1049 ))
1050 })
1051 }
1052
1053 #[inline]
1054 fn visit_field(
1055 &mut self,
1056 old_val: &PlaceTy<'tcx, M::Provenance>,
1057 field: usize,
1058 new_val: &PlaceTy<'tcx, M::Provenance>,
1059 ) -> InterpResult<'tcx> {
1060 let elem = self.aggregate_field_path_elem(old_val.layout, field);
1061 self.with_elem(elem, move |this| this.visit_value(new_val))
1062 }
1063
1064 #[inline]
1065 fn visit_variant(
1066 &mut self,
1067 old_val: &PlaceTy<'tcx, M::Provenance>,
1068 variant_id: VariantIdx,
1069 new_val: &PlaceTy<'tcx, M::Provenance>,
1070 ) -> InterpResult<'tcx> {
1071 let name = match old_val.layout.ty.kind() {
1072 ty::Adt(adt, _) => PathElem::Variant(adt.variant(variant_id).name),
1073 // Coroutines also have variants
1074 ty::Coroutine(..) => PathElem::CoroutineState(variant_id),
1075 _ => bug!("Unexpected type with variant: {:?}", old_val.layout.ty),
1076 };
1077 self.with_elem(name, move |this| this.visit_value(new_val))
1078 }
1079
1080 #[inline(always)]
1081 fn visit_union(
1082 &mut self,
1083 val: &PlaceTy<'tcx, M::Provenance>,
1084 _fields: NonZero<usize>,
1085 ) -> InterpResult<'tcx> {
1086 // Special check for CTFE validation, preventing `UnsafeCell` inside unions in immutable memory.
1087 if self.ctfe_mode.is_some_and(|c| !c.allow_immutable_unsafe_cell()) {
1088 // Unsized unions are currently not a thing, but let's keep this code consistent with
1089 // the check in `visit_value`.
1090 let zst = self.ecx.size_and_align_of_val(val)?.is_some_and(|(s, _a)| s.bytes() == 0);
1091 if !zst && !val.layout.ty.is_freeze(*self.ecx.tcx, self.ecx.typing_env) {
1092 if !self.in_mutable_memory(val) {
1093 throw_validation_failure!(self.path, UnsafeCellInImmutable);
1094 }
1095 }
1096 }
1097 if self.reset_provenance_and_padding
1098 && let Some(data_bytes) = self.data_bytes.as_mut()
1099 {
1100 let base_offset = Self::data_range_offset(self.ecx, val);
1101 // Determine and add data range for this union.
1102 let union_data_range = Self::union_data_range(self.ecx, val.layout);
1103 for &(offset, size) in union_data_range.0.iter() {
1104 data_bytes.add_range(base_offset + offset, size);
1105 }
1106 }
1107 interp_ok(())
1108 }
1109
1110 #[inline]
1111 fn visit_box(
1112 &mut self,
1113 _box_ty: Ty<'tcx>,
1114 val: &PlaceTy<'tcx, M::Provenance>,
1115 ) -> InterpResult<'tcx> {
1116 self.check_safe_pointer(val, PointerKind::Box)?;
1117 interp_ok(())
1118 }
1119
1120 #[inline]
1121 fn visit_value(&mut self, val: &PlaceTy<'tcx, M::Provenance>) -> InterpResult<'tcx> {
1122 trace!("visit_value: {:?}, {:?}", *val, val.layout);
1123
1124 // Check primitive types -- the leaves of our recursive descent.
1125 // This is called even for enum discriminants (which are "fields" of their enum),
1126 // so for integer-typed discriminants the provenance reset will happen here.
1127 // We assume that the Scalar validity range does not restrict these values
1128 // any further than `try_visit_primitive` does!
1129 if self.try_visit_primitive(val)? {
1130 return interp_ok(());
1131 }
1132
1133 // Special check preventing `UnsafeCell` in the inner part of constants
1134 if self.ctfe_mode.is_some_and(|c| !c.allow_immutable_unsafe_cell()) {
1135 // Exclude ZST values. We need to compute the dynamic size/align to properly
1136 // handle slices and trait objects.
1137 let zst = self.ecx.size_and_align_of_val(val)?.is_some_and(|(s, _a)| s.bytes() == 0);
1138 if !zst
1139 && let Some(def) = val.layout.ty.ty_adt_def()
1140 && def.is_unsafe_cell()
1141 {
1142 if !self.in_mutable_memory(val) {
1143 throw_validation_failure!(self.path, UnsafeCellInImmutable);
1144 }
1145 }
1146 }
1147
1148 // Recursively walk the value at its type. Apply optimizations for some large types.
1149 match val.layout.ty.kind() {
1150 ty::Str => {
1151 let mplace = val.assert_mem_place(); // strings are unsized and hence never immediate
1152 let len = mplace.len(self.ecx)?;
1153 try_validation!(
1154 self.ecx.read_bytes_ptr_strip_provenance(mplace.ptr(), Size::from_bytes(len)),
1155 self.path,
1156 Ub(InvalidUninitBytes(..)) => Uninit { expected: ExpectedKind::Str },
1157 Unsup(ReadPointerAsInt(_)) => PointerAsInt { expected: ExpectedKind::Str }
1158 );
1159 }
1160 ty::Array(tys, ..) | ty::Slice(tys)
1161 // This optimization applies for types that can hold arbitrary non-provenance bytes (such as
1162 // integer and floating point types).
1163 // FIXME(wesleywiser) This logic could be extended further to arbitrary structs or
1164 // tuples made up of integer/floating point types or inhabited ZSTs with no padding.
1165 if matches!(tys.kind(), ty::Int(..) | ty::Uint(..) | ty::Float(..))
1166 =>
1167 {
1168 let expected = if tys.is_integral() { ExpectedKind::Int } else { ExpectedKind::Float };
1169 // Optimized handling for arrays of integer/float type.
1170
1171 // This is the length of the array/slice.
1172 let len = val.len(self.ecx)?;
1173 // This is the element type size.
1174 let layout = self.ecx.layout_of(*tys)?;
1175 // This is the size in bytes of the whole array. (This checks for overflow.)
1176 let size = layout.size * len;
1177 // If the size is 0, there is nothing to check.
1178 // (`size` can only be 0 if `len` is 0, and empty arrays are always valid.)
1179 if size == Size::ZERO {
1180 return interp_ok(());
1181 }
1182 // Now that we definitely have a non-ZST array, we know it lives in memory -- except it may
1183 // be an uninitialized local variable, those are also "immediate".
1184 let mplace = match val.to_op(self.ecx)?.as_mplace_or_imm() {
1185 Left(mplace) => mplace,
1186 Right(imm) => match *imm {
1187 Immediate::Uninit =>
1188 throw_validation_failure!(self.path, Uninit { expected }),
1189 Immediate::Scalar(..) | Immediate::ScalarPair(..) =>
1190 bug!("arrays/slices can never have Scalar/ScalarPair layout"),
1191 }
1192 };
1193
1194 // Optimization: we just check the entire range at once.
1195 // NOTE: Keep this in sync with the handling of integer and float
1196 // types above, in `visit_primitive`.
1197 // No need for an alignment check here, this is not an actual memory access.
1198 let alloc = self.ecx.get_ptr_alloc(mplace.ptr(), size)?.expect("we already excluded size 0");
1199
1200 alloc.get_bytes_strip_provenance().map_err_kind(|kind| {
1201 // Some error happened, try to provide a more detailed description.
1202 // For some errors we might be able to provide extra information.
1203 // (This custom logic does not fit the `try_validation!` macro.)
1204 match kind {
1205 Ub(InvalidUninitBytes(Some((_alloc_id, access)))) | Unsup(ReadPointerAsInt(Some((_alloc_id, access)))) => {
1206 // Some byte was uninitialized, determine which
1207 // element that byte belongs to so we can
1208 // provide an index.
1209 let i = usize::try_from(
1210 access.bad.start.bytes() / layout.size.bytes(),
1211 )
1212 .unwrap();
1213 self.path.push(PathElem::ArrayElem(i));
1214
1215 if matches!(kind, Ub(InvalidUninitBytes(_))) {
1216 err_validation_failure!(self.path, Uninit { expected })
1217 } else {
1218 err_validation_failure!(self.path, PointerAsInt { expected })
1219 }
1220 }
1221
1222 // Propagate upwards (that will also check for unexpected errors).
1223 err => err,
1224 }
1225 })?;
1226
1227 // Don't forget that these are all non-pointer types, and thus do not preserve
1228 // provenance.
1229 if self.reset_provenance_and_padding {
1230 // We can't share this with above as above, we might be looking at read-only memory.
1231 let mut alloc = self.ecx.get_ptr_alloc_mut(mplace.ptr(), size)?.expect("we already excluded size 0");
1232 alloc.clear_provenance()?;
1233 // Also, mark this as containing data, not padding.
1234 self.add_data_range(mplace.ptr(), size);
1235 }
1236 }
1237 // Fast path for arrays and slices of ZSTs. We only need to check a single ZST element
1238 // of an array and not all of them, because there's only a single value of a specific
1239 // ZST type, so either validation fails for all elements or none.
1240 ty::Array(tys, ..) | ty::Slice(tys) if self.ecx.layout_of(*tys)?.is_zst() => {
1241 // Validate just the first element (if any).
1242 if val.len(self.ecx)? > 0 {
1243 self.visit_field(val, 0, &self.ecx.project_index(val, 0)?)?;
1244 }
1245 }
1246 ty::Pat(base, pat) => {
1247 // First check that the base type is valid
1248 self.visit_value(&val.transmute(self.ecx.layout_of(*base)?, self.ecx)?)?;
1249 // When you extend this match, make sure to also add tests to
1250 // tests/ui/type/pattern_types/validity.rs((
1251 match **pat {
1252 // Range patterns are precisely reflected into `valid_range` and thus
1253 // handled fully by `visit_scalar` (called below).
1254 ty::PatternKind::Range { .. } => {},
1255
1256 // FIXME(pattern_types): check that the value is covered by one of the variants.
1257 // For now, we rely on layout computation setting the scalar's `valid_range` to
1258 // match the pattern. However, this cannot always work; the layout may
1259 // pessimistically cover actually illegal ranges and Miri would miss that UB.
1260 // The consolation here is that codegen also will miss that UB, so at least
1261 // we won't see optimizations actually breaking such programs.
1262 ty::PatternKind::Or(_patterns) => {}
1263 }
1264 }
1265 _ => {
1266 // default handler
1267 try_validation!(
1268 self.walk_value(val),
1269 self.path,
1270 // It's not great to catch errors here, since we can't give a very good path,
1271 // but it's better than ICEing.
1272 Ub(InvalidVTableTrait { vtable_dyn_type, expected_dyn_type }) => {
1273 InvalidMetaWrongTrait { vtable_dyn_type, expected_dyn_type }
1274 },
1275 );
1276 }
1277 }
1278
1279 // *After* all of this, check further information stored in the layout. We need to check
1280 // this to handle types like `NonNull` where the `Scalar` info is more restrictive than what
1281 // the fields say (`rustc_layout_scalar_valid_range_start`). But in most cases, this will
1282 // just propagate what the fields say, and then we want the error to point at the field --
1283 // so, we first recurse, then we do this check.
1284 //
1285 // FIXME: We could avoid some redundant checks here. For newtypes wrapping
1286 // scalars, we do the same check on every "level" (e.g., first we check
1287 // MyNewtype and then the scalar in there).
1288 if val.layout.is_uninhabited() {
1289 let ty = val.layout.ty;
1290 throw_validation_failure!(self.path, UninhabitedVal { ty });
1291 }
1292 match val.layout.backend_repr {
1293 BackendRepr::Scalar(scalar_layout) => {
1294 if !scalar_layout.is_uninit_valid() {
1295 // There is something to check here.
1296 let scalar = self.read_scalar(val, ExpectedKind::InitScalar)?;
1297 self.visit_scalar(scalar, scalar_layout)?;
1298 }
1299 }
1300 BackendRepr::ScalarPair(a_layout, b_layout) => {
1301 // We can only proceed if *both* scalars need to be initialized.
1302 // FIXME: find a way to also check ScalarPair when one side can be uninit but
1303 // the other must be init.
1304 if !a_layout.is_uninit_valid() && !b_layout.is_uninit_valid() {
1305 let (a, b) =
1306 self.read_immediate(val, ExpectedKind::InitScalar)?.to_scalar_pair();
1307 self.visit_scalar(a, a_layout)?;
1308 self.visit_scalar(b, b_layout)?;
1309 }
1310 }
1311 BackendRepr::SimdVector { .. } => {
1312 // No checks here, we assume layout computation gets this right.
1313 // (This is harder to check since Miri does not represent these as `Immediate`. We
1314 // also cannot use field projections since this might be a newtype around a vector.)
1315 }
1316 BackendRepr::Memory { .. } => {
1317 // Nothing to do.
1318 }
1319 }
1320
1321 interp_ok(())
1322 }
1323}
1324
1325impl<'tcx, M: Machine<'tcx>> InterpCx<'tcx, M> {
1326 fn validate_operand_internal(
1327 &mut self,
1328 val: &PlaceTy<'tcx, M::Provenance>,
1329 path: Vec<PathElem>,
1330 ref_tracking: Option<&mut RefTracking<MPlaceTy<'tcx, M::Provenance>, Vec<PathElem>>>,
1331 ctfe_mode: Option<CtfeValidationMode>,
1332 reset_provenance_and_padding: bool,
1333 ) -> InterpResult<'tcx> {
1334 trace!("validate_operand_internal: {:?}, {:?}", *val, val.layout.ty);
1335
1336 // Run the visitor.
1337 self.run_for_validation_mut(|ecx| {
1338 let reset_padding = reset_provenance_and_padding && {
1339 // Check if `val` is actually stored in memory. If not, padding is not even
1340 // represented and we need not reset it.
1341 ecx.place_to_op(val)?.as_mplace_or_imm().is_left()
1342 };
1343 let mut v = ValidityVisitor {
1344 path,
1345 ref_tracking,
1346 ctfe_mode,
1347 ecx,
1348 reset_provenance_and_padding,
1349 data_bytes: reset_padding.then_some(RangeSet(Vec::new())),
1350 };
1351 v.visit_value(val)?;
1352 v.reset_padding(val)?;
1353 interp_ok(())
1354 })
1355 .map_err_info(|err| {
1356 if !matches!(
1357 err.kind(),
1358 err_ub!(ValidationError { .. })
1359 | InterpErrorKind::InvalidProgram(_)
1360 | InterpErrorKind::Unsupported(UnsupportedOpInfo::ExternTypeField)
1361 ) {
1362 bug!(
1363 "Unexpected error during validation: {}",
1364 format_interp_error(self.tcx.dcx(), err)
1365 );
1366 }
1367 err
1368 })
1369 }
1370
1371 /// This function checks the data at `val` to be const-valid.
1372 /// `val` is assumed to cover valid memory if it is an indirect operand.
1373 /// It will error if the bits at the destination do not match the ones described by the layout.
1374 ///
1375 /// `ref_tracking` is used to record references that we encounter so that they
1376 /// can be checked recursively by an outside driving loop.
1377 ///
1378 /// `constant` controls whether this must satisfy the rules for constants:
1379 /// - no pointers to statics.
1380 /// - no `UnsafeCell` or non-ZST `&mut`.
1381 #[inline(always)]
1382 pub(crate) fn const_validate_operand(
1383 &mut self,
1384 val: &PlaceTy<'tcx, M::Provenance>,
1385 path: Vec<PathElem>,
1386 ref_tracking: &mut RefTracking<MPlaceTy<'tcx, M::Provenance>, Vec<PathElem>>,
1387 ctfe_mode: CtfeValidationMode,
1388 ) -> InterpResult<'tcx> {
1389 self.validate_operand_internal(
1390 val,
1391 path,
1392 Some(ref_tracking),
1393 Some(ctfe_mode),
1394 /*reset_provenance*/ false,
1395 )
1396 }
1397
1398 /// This function checks the data at `val` to be runtime-valid.
1399 /// `val` is assumed to cover valid memory if it is an indirect operand.
1400 /// It will error if the bits at the destination do not match the ones described by the layout.
1401 #[inline(always)]
1402 pub fn validate_operand(
1403 &mut self,
1404 val: &PlaceTy<'tcx, M::Provenance>,
1405 recursive: bool,
1406 reset_provenance_and_padding: bool,
1407 ) -> InterpResult<'tcx> {
1408 let _span = enter_trace_span!(
1409 M,
1410 "validate_operand",
1411 "recursive={recursive}, reset_provenance_and_padding={reset_provenance_and_padding}, val={val:?}"
1412 );
1413
1414 // Note that we *could* actually be in CTFE here with `-Zextra-const-ub-checks`, but it's
1415 // still correct to not use `ctfe_mode`: that mode is for validation of the final constant
1416 // value, it rules out things like `UnsafeCell` in awkward places.
1417 if !recursive {
1418 return self.validate_operand_internal(
1419 val,
1420 vec![],
1421 None,
1422 None,
1423 reset_provenance_and_padding,
1424 );
1425 }
1426 // Do a recursive check.
1427 let mut ref_tracking = RefTracking::empty();
1428 self.validate_operand_internal(
1429 val,
1430 vec![],
1431 Some(&mut ref_tracking),
1432 None,
1433 reset_provenance_and_padding,
1434 )?;
1435 while let Some((mplace, path)) = ref_tracking.todo.pop() {
1436 // Things behind reference do *not* have the provenance reset.
1437 self.validate_operand_internal(
1438 &mplace.into(),
1439 path,
1440 Some(&mut ref_tracking),
1441 None,
1442 /*reset_provenance_and_padding*/ false,
1443 )?;
1444 }
1445 interp_ok(())
1446 }
1447}