interpret: reset provenance on typed copies

Author: RalfJung

compiler/rustc_const_eval/src/const_eval/eval_queries.rs

@@ -94,7 +94,7 @@ fn eval_body_using_ecx<'tcx, R: InterpretationResult<'tcx>>(
     let intern_result = intern_const_alloc_recursive(ecx, intern_kind, &ret);
 
     // Since evaluation had no errors, validate the resulting constant.
-    const_validate_mplace(&ecx, &ret, cid)?;
+    const_validate_mplace(ecx, &ret, cid)?;
 
     // Only report this after validation, as validaiton produces much better diagnostics.
     // FIXME: ensure validation always reports this and stop making interning care about it.
@@ -391,7 +391,7 @@ fn eval_in_interpreter<'tcx, R: InterpretationResult<'tcx>>(
 
 #[inline(always)]
 fn const_validate_mplace<'tcx>(
-    ecx: &InterpCx<'tcx, CompileTimeMachine<'tcx>>,
+    ecx: &mut InterpCx<'tcx, CompileTimeMachine<'tcx>>,
     mplace: &MPlaceTy<'tcx>,
     cid: GlobalId<'tcx>,
 ) -> Result<(), ErrorHandled> {

compiler/rustc_const_eval/src/interpret/memory.rs

@@ -8,9 +8,8 @@
 
 use std::assert_matches::assert_matches;
 use std::borrow::Cow;
-use std::cell::Cell;
 use std::collections::VecDeque;
-use std::{fmt, ptr};
+use std::{fmt, mem, ptr};
 
 use rustc_ast::Mutability;
 use rustc_data_structures::fx::{FxHashSet, FxIndexMap};
@@ -118,7 +117,7 @@ pub struct Memory<'tcx, M: Machine<'tcx>> {
     /// This stores whether we are currently doing reads purely for the purpose of validation.
     /// Those reads do not trigger the machine's hooks for memory reads.
     /// Needless to say, this must only be set with great care!
-    validation_in_progress: Cell<bool>,
+    validation_in_progress: bool,
 }
 
 /// A reference to some allocation that was already bounds-checked for the given region
@@ -145,7 +144,7 @@ impl<'tcx, M: Machine<'tcx>> Memory<'tcx, M> {
             alloc_map: M::MemoryMap::default(),
             extra_fn_ptr_map: FxIndexMap::default(),
             dead_alloc_map: FxIndexMap::default(),
-            validation_in_progress: Cell::new(false),
+            validation_in_progress: false,
         }
     }
 
@@ -682,15 +681,15 @@ impl<'tcx, M: Machine<'tcx>> InterpCx<'tcx, M> {
         // We want to call the hook on *all* accesses that involve an AllocId, including zero-sized
         // accesses. That means we cannot rely on the closure above or the `Some` branch below. We
         // do this after `check_and_deref_ptr` to ensure some basic sanity has already been checked.
-        if !self.memory.validation_in_progress.get() {
+        if !self.memory.validation_in_progress {
             if let Ok((alloc_id, ..)) = self.ptr_try_get_alloc_id(ptr, size_i64) {
                 M::before_alloc_read(self, alloc_id)?;
             }
         }
 
         if let Some((alloc_id, offset, prov, alloc)) = ptr_and_alloc {
             let range = alloc_range(offset, size);
-            if !self.memory.validation_in_progress.get() {
+            if !self.memory.validation_in_progress {
                 M::before_memory_read(
                     self.tcx,
                     &self.machine,
@@ -766,11 +765,14 @@ impl<'tcx, M: Machine<'tcx>> InterpCx<'tcx, M> {
         let parts = self.get_ptr_access(ptr, size)?;
         if let Some((alloc_id, offset, prov)) = parts {
             let tcx = self.tcx;
+            let validation_in_progress = self.memory.validation_in_progress;
             // FIXME: can we somehow avoid looking up the allocation twice here?
             // We cannot call `get_raw_mut` inside `check_and_deref_ptr` as that would duplicate `&mut self`.
             let (alloc, machine) = self.get_alloc_raw_mut(alloc_id)?;
             let range = alloc_range(offset, size);
-            M::before_memory_write(tcx, machine, &mut alloc.extra, (alloc_id, prov), range)?;
+            if !validation_in_progress {
+                M::before_memory_write(tcx, machine, &mut alloc.extra, (alloc_id, prov), range)?;
+            }
             Ok(Some(AllocRefMut { alloc, range, tcx: *tcx, alloc_id }))
         } else {
             Ok(None)
@@ -1014,16 +1016,16 @@ impl<'tcx, M: Machine<'tcx>> InterpCx<'tcx, M> {
     ///
     /// We do this so Miri's allocation access tracking does not show the validation
     /// reads as spurious accesses.
-    pub fn run_for_validation<R>(&self, f: impl FnOnce() -> R) -> R {
+    pub fn run_for_validation<R>(&mut self, f: impl FnOnce(&mut Self) -> R) -> R {
         // This deliberately uses `==` on `bool` to follow the pattern
         // `assert!(val.replace(new) == old)`.
         assert!(
-            self.memory.validation_in_progress.replace(true) == false,
+            mem::replace(&mut self.memory.validation_in_progress, true) == false,
             "`validation_in_progress` was already set"
         );
-        let res = f();
+        let res = f(self);
         assert!(
-            self.memory.validation_in_progress.replace(false) == true,
+            mem::replace(&mut self.memory.validation_in_progress, false) == true,
             "`validation_in_progress` was unset by someone else"
         );
         res
@@ -1115,6 +1117,10 @@ impl<'a, 'tcx, M: Machine<'tcx>> std::fmt::Debug for DumpAllocs<'a, 'tcx, M> {
 impl<'tcx, 'a, Prov: Provenance, Extra, Bytes: AllocBytes>
     AllocRefMut<'a, 'tcx, Prov, Extra, Bytes>
 {
+    pub fn as_ref<'b>(&'b self) -> AllocRef<'b, 'tcx, Prov, Extra, Bytes> {
+        AllocRef { alloc: self.alloc, range: self.range, tcx: self.tcx, alloc_id: self.alloc_id }
+    }
+
     /// `range` is relative to this allocation reference, not the base of the allocation.
     pub fn write_scalar(&mut self, range: AllocRange, val: Scalar<Prov>) -> InterpResult<'tcx> {
         let range = self.range.subrange(range);
@@ -1137,6 +1143,14 @@ impl<'tcx, 'a, Prov: Provenance, Extra, Bytes: AllocBytes>
             .write_uninit(&self.tcx, self.range)
             .map_err(|e| e.to_interp_error(self.alloc_id))?)
     }
+
+    /// Remove all provenance in the reference range.
+    pub fn clear_provenance(&mut self) -> InterpResult<'tcx> {
+        Ok(self
+            .alloc
+            .clear_provenance(&self.tcx, self.range)
+            .map_err(|e| e.to_interp_error(self.alloc_id))?)
+    }
 }
 
 impl<'tcx, 'a, Prov: Provenance, Extra, Bytes: AllocBytes> AllocRef<'a, 'tcx, Prov, Extra, Bytes> {
@@ -1278,7 +1292,7 @@ impl<'tcx, M: Machine<'tcx>> InterpCx<'tcx, M> {
         };
         let src_alloc = self.get_alloc_raw(src_alloc_id)?;
         let src_range = alloc_range(src_offset, size);
-        assert!(!self.memory.validation_in_progress.get(), "we can't be copying during validation");
+        assert!(!self.memory.validation_in_progress, "we can't be copying during validation");
         M::before_memory_read(
             tcx,
             &self.machine,

compiler/rustc_const_eval/src/interpret/operand.rs

@@ -137,6 +137,20 @@ impl<Prov: Provenance> Immediate<Prov> {
             }
         }
     }
+
+    pub fn clear_provenance<'tcx>(&mut self) -> InterpResult<'tcx> {
+        match self {
+            Immediate::Scalar(s) => {
+                s.clear_provenance()?;
+            }
+            Immediate::ScalarPair(a, b) => {
+                a.clear_provenance()?;
+                b.clear_provenance()?;
+            }
+            Immediate::Uninit => {}
+        }
+        Ok(())
+    }
 }
 
 // ScalarPair needs a type to interpret, so we often have an immediate and a type together

compiler/rustc_const_eval/src/interpret/place.rs

@@ -605,8 +605,9 @@ where
         if M::enforce_validity(self, dest.layout()) {
             // Data got changed, better make sure it matches the type!
             self.validate_operand(
-                &dest.to_op(self)?,
+                &dest.to_place(),
                 M::enforce_validity_recursively(self, dest.layout()),
+                /*reset_provenance*/ true,
             )?;
         }
 
@@ -636,7 +637,7 @@ where
     /// Write an immediate to a place.
     /// If you use this you are responsible for validating that things got copied at the
     /// right type.
-    fn write_immediate_no_validate(
+    pub(super) fn write_immediate_no_validate(
         &mut self,
         src: Immediate<M::Provenance>,
         dest: &impl Writeable<'tcx, M::Provenance>,
@@ -684,15 +685,7 @@ where
 
         match value {
             Immediate::Scalar(scalar) => {
-                let Abi::Scalar(s) = layout.abi else {
-                    span_bug!(
-                        self.cur_span(),
-                        "write_immediate_to_mplace: invalid Scalar layout: {layout:#?}",
-                    )
-                };
-                let size = s.size(&tcx);
-                assert_eq!(size, layout.size, "abi::Scalar size does not match layout size");
-                alloc.write_scalar(alloc_range(Size::ZERO, size), scalar)
+                alloc.write_scalar(alloc_range(Size::ZERO, scalar.size()), scalar)
             }
             Immediate::ScalarPair(a_val, b_val) => {
                 let Abi::ScalarPair(a, b) = layout.abi else {
@@ -702,16 +695,15 @@ where
                         layout
                     )
                 };
-                let (a_size, b_size) = (a.size(&tcx), b.size(&tcx));
-                let b_offset = a_size.align_to(b.align(&tcx).abi);
+                let b_offset = a.size(&tcx).align_to(b.align(&tcx).abi);
                 assert!(b_offset.bytes() > 0); // in `operand_field` we use the offset to tell apart the fields
 
                 // It is tempting to verify `b_offset` against `layout.fields.offset(1)`,
                 // but that does not work: We could be a newtype around a pair, then the
                 // fields do not match the `ScalarPair` components.
 
-                alloc.write_scalar(alloc_range(Size::ZERO, a_size), a_val)?;
-                alloc.write_scalar(alloc_range(b_offset, b_size), b_val)
+                alloc.write_scalar(alloc_range(Size::ZERO, a_val.size()), a_val)?;
+                alloc.write_scalar(alloc_range(b_offset, b_val.size()), b_val)
             }
             Immediate::Uninit => alloc.write_uninit(),
         }
@@ -736,6 +728,26 @@ where
         Ok(())
     }
 
+    /// Remove all provenance in the given place.
+    pub fn clear_provenance(
+        &mut self,
+        dest: &impl Writeable<'tcx, M::Provenance>,
+    ) -> InterpResult<'tcx> {
+        match self.as_mplace_or_mutable_local(&dest.to_place())? {
+            Right((local_val, _local_layout)) => {
+                local_val.clear_provenance()?;
+            }
+            Left(mplace) => {
+                let Some(mut alloc) = self.get_place_alloc_mut(&mplace)? else {
+                    // Zero-sized access
+                    return Ok(());
+                };
+                alloc.clear_provenance()?;
+            }
+        }
+        Ok(())
+    }
+
     /// Copies the data from an operand to a place.
     /// The layouts of the `src` and `dest` may disagree.
     /// Does not perform validation of the destination.
@@ -789,23 +801,30 @@ where
         allow_transmute: bool,
         validate_dest: bool,
     ) -> InterpResult<'tcx> {
-        // Generally for transmutation, data must be valid both at the old and new type.
-        // But if the types are the same, the 2nd validation below suffices.
-        if src.layout().ty != dest.layout().ty && M::enforce_validity(self, src.layout()) {
-            self.validate_operand(
-                &src.to_op(self)?,
-                M::enforce_validity_recursively(self, src.layout()),
-            )?;
-        }
+        // These are technically *two* typed copies: `src` is a not-yet-loaded value,
+        // so we're going a typed copy at `src` type from there to some intermediate storage.
+        // And then we're doing a second typed copy from that intermediate storage to `dest`.
+        // But as an optimization, we only make a single direct copy here.
 
         // Do the actual copy.
         self.copy_op_no_validate(src, dest, allow_transmute)?;
 
         if validate_dest && M::enforce_validity(self, dest.layout()) {
-            // Data got changed, better make sure it matches the type!
+            let dest = dest.to_place();
+            // Given that there were two typed copies, we have to ensure this is valid at both types,
+            // and we have to ensure this loses provenance and padding according to both types.
+            // But if the types are identical, we only do one pass.
+            if src.layout().ty != dest.layout().ty {
+                self.validate_operand(
+                    &dest.transmute(src.layout(), self)?,
+                    M::enforce_validity_recursively(self, src.layout()),
+                    /*reset_provenance*/ true,
+                )?;
+            }
             self.validate_operand(
-                &dest.to_op(self)?,
+                &dest,
                 M::enforce_validity_recursively(self, dest.layout()),
+                /*reset_provenance*/ true,
             )?;
         }