rust-lang/rust
1
0
Fork 0
Code Issues Pull Requests Actions 1 Packages Projects Releases Wiki Activity

Rename rustc_mir to rustc_const_eval.

Browse Source
This commit is contained in:
Camille GILLOT
2021-01-05 20:08:11 +01:00
parent fd9c04fe32
commit c5fc2609f0
64 changed files with 66 additions and 66 deletions
Download Patch File Download Diff File Expand all files Collapse all files

365
compiler/rustc_const_eval/src/interpret/cast.rs Normal file
View File

@@ -0,0 +1,365 @@
use std::convert::TryFrom;
use rustc_apfloat::ieee::{Double, Single};
use rustc_apfloat::{Float, FloatConvert};
use rustc_middle::mir::interpret::{InterpResult, PointerArithmetic, Scalar};
use rustc_middle::mir::CastKind;
use rustc_middle::ty::adjustment::PointerCast;
use rustc_middle::ty::layout::{IntegerExt, LayoutOf, TyAndLayout};
use rustc_middle::ty::{self, FloatTy, Ty, TypeAndMut};
use rustc_target::abi::{Integer, Variants};
use super::{
util::ensure_monomorphic_enough, FnVal, ImmTy, Immediate, InterpCx, Machine, OpTy, PlaceTy,
};
impl<'mir, 'tcx: 'mir, M: Machine<'mir, 'tcx>> InterpCx<'mir, 'tcx, M> {
pub fn cast(
&mut self,
src: &OpTy<'tcx, M::PointerTag>,
cast_kind: CastKind,
cast_ty: Ty<'tcx>,
dest: &PlaceTy<'tcx, M::PointerTag>,
) -> InterpResult<'tcx> {
use rustc_middle::mir::CastKind::*;
// FIXME: In which cases should we trigger UB when the source is uninit?
match cast_kind {
Pointer(PointerCast::Unsize) => {
let cast_ty = self.layout_of(cast_ty)?;
self.unsize_into(src, cast_ty, dest)?;
}
Misc => {
let src = self.read_immediate(src)?;
let res = self.misc_cast(&src, cast_ty)?;
self.write_immediate(res, dest)?;
}
Pointer(PointerCast::MutToConstPointer | PointerCast::ArrayToPointer) => {
// These are NOPs, but can be wide pointers.
let v = self.read_immediate(src)?;
self.write_immediate(*v, dest)?;
}
Pointer(PointerCast::ReifyFnPointer) => {
// The src operand does not matter, just its type
match *src.layout.ty.kind() {
ty::FnDef(def_id, substs) => {
// All reifications must be monomorphic, bail out otherwise.
ensure_monomorphic_enough(*self.tcx, src.layout.ty)?;
let instance = ty::Instance::resolve_for_fn_ptr(
*self.tcx,
self.param_env,
def_id,
substs,
)
.ok_or_else(|| err_inval!(TooGeneric))?;
let fn_ptr = self.memory.create_fn_alloc(FnVal::Instance(instance));
self.write_pointer(fn_ptr, dest)?;
}
_ => span_bug!(self.cur_span(), "reify fn pointer on {:?}", src.layout.ty),
}
}
Pointer(PointerCast::UnsafeFnPointer) => {
let src = self.read_immediate(src)?;
match cast_ty.kind() {
ty::FnPtr(_) => {
// No change to value
self.write_immediate(*src, dest)?;
}
_ => span_bug!(self.cur_span(), "fn to unsafe fn cast on {:?}", cast_ty),
}
}
Pointer(PointerCast::ClosureFnPointer(_)) => {
// The src operand does not matter, just its type
match *src.layout.ty.kind() {
ty::Closure(def_id, substs) => {
// All reifications must be monomorphic, bail out otherwise.
ensure_monomorphic_enough(*self.tcx, src.layout.ty)?;
let instance = ty::Instance::resolve_closure(
*self.tcx,
def_id,
substs,
ty::ClosureKind::FnOnce,
);
let fn_ptr = self.memory.create_fn_alloc(FnVal::Instance(instance));
self.write_pointer(fn_ptr, dest)?;
}
_ => span_bug!(self.cur_span(), "closure fn pointer on {:?}", src.layout.ty),
}
}
}
Ok(())
}
fn misc_cast(
&self,
src: &ImmTy<'tcx, M::PointerTag>,
cast_ty: Ty<'tcx>,
) -> InterpResult<'tcx, Immediate<M::PointerTag>> {
use rustc_middle::ty::TyKind::*;
trace!("Casting {:?}: {:?} to {:?}", *src, src.layout.ty, cast_ty);
match src.layout.ty.kind() {
// Floating point
Float(FloatTy::F32) => {
return Ok(self.cast_from_float(src.to_scalar()?.to_f32()?, cast_ty).into());
}
Float(FloatTy::F64) => {
return Ok(self.cast_from_float(src.to_scalar()?.to_f64()?, cast_ty).into());
}
// The rest is integer/pointer-"like", including fn ptr casts and casts from enums that
// are represented as integers.
_ => assert!(
src.layout.ty.is_bool()
|| src.layout.ty.is_char()
|| src.layout.ty.is_enum()
|| src.layout.ty.is_integral()
|| src.layout.ty.is_any_ptr(),
"Unexpected cast from type {:?}",
src.layout.ty
),
}
// # First handle non-scalar source values.
// Handle cast from a ZST enum (0 or 1 variants).
match src.layout.variants {
Variants::Single { index } => {
if src.layout.abi.is_uninhabited() {
// This is dead code, because an uninhabited enum is UB to
// instantiate.
throw_ub!(Unreachable);
}
if let Some(discr) = src.layout.ty.discriminant_for_variant(*self.tcx, index) {
assert!(src.layout.is_zst());
let discr_layout = self.layout_of(discr.ty)?;
return Ok(self.cast_from_scalar(discr.val, discr_layout, cast_ty).into());
}
}
Variants::Multiple { .. } => {}
}
// Handle casting any ptr to raw ptr (might be a fat ptr).
if src.layout.ty.is_any_ptr() && cast_ty.is_unsafe_ptr() {
let dest_layout = self.layout_of(cast_ty)?;
if dest_layout.size == src.layout.size {
// Thin or fat pointer that just hast the ptr kind of target type changed.
return Ok(**src);
} else {
// Casting the metadata away from a fat ptr.
assert_eq!(src.layout.size, 2 * self.memory.pointer_size());
assert_eq!(dest_layout.size, self.memory.pointer_size());
assert!(src.layout.ty.is_unsafe_ptr());
return match **src {
Immediate::ScalarPair(data, _) => Ok(data.into()),
Immediate::Scalar(..) => span_bug!(
self.cur_span(),
"{:?} input to a fat-to-thin cast ({:?} -> {:?})",
*src,
src.layout.ty,
cast_ty
),
};
}
}
// # The remaining source values are scalar.
// For all remaining casts, we either
// (a) cast a raw ptr to usize, or
// (b) cast from an integer-like (including bool, char, enums).
// In both cases we want the bits.
let bits = src.to_scalar()?.to_bits(src.layout.size)?;
Ok(self.cast_from_scalar(bits, src.layout, cast_ty).into())
}
pub(super) fn cast_from_scalar(
&self,
v: u128, // raw bits (there is no ScalarTy so we separate data+layout)
src_layout: TyAndLayout<'tcx>,
cast_ty: Ty<'tcx>,
) -> Scalar<M::PointerTag> {
// Let's make sure v is sign-extended *if* it has a signed type.
let signed = src_layout.abi.is_signed(); // Also asserts that abi is `Scalar`.
let v = if signed { self.sign_extend(v, src_layout) } else { v };
trace!("cast_from_scalar: {}, {} -> {}", v, src_layout.ty, cast_ty);
use rustc_middle::ty::TyKind::*;
match *cast_ty.kind() {
Int(_) | Uint(_) | RawPtr(_) => {
let size = match *cast_ty.kind() {
Int(t) => Integer::from_int_ty(self, t).size(),
Uint(t) => Integer::from_uint_ty(self, t).size(),
RawPtr(_) => self.pointer_size(),
_ => bug!(),
};
let v = size.truncate(v);
Scalar::from_uint(v, size)
}
Float(FloatTy::F32) if signed => Scalar::from_f32(Single::from_i128(v as i128).value),
Float(FloatTy::F64) if signed => Scalar::from_f64(Double::from_i128(v as i128).value),
Float(FloatTy::F32) => Scalar::from_f32(Single::from_u128(v).value),
Float(FloatTy::F64) => Scalar::from_f64(Double::from_u128(v).value),
Char => {
// `u8` to `char` cast
Scalar::from_u32(u8::try_from(v).unwrap().into())
}
// Casts to bool are not permitted by rustc, no need to handle them here.
_ => span_bug!(self.cur_span(), "invalid int to {:?} cast", cast_ty),
}
}
fn cast_from_float<F>(&self, f: F, dest_ty: Ty<'tcx>) -> Scalar<M::PointerTag>
where
F: Float + Into<Scalar<M::PointerTag>> + FloatConvert<Single> + FloatConvert<Double>,
{
use rustc_middle::ty::TyKind::*;
match *dest_ty.kind() {
// float -> uint
Uint(t) => {
let size = Integer::from_uint_ty(self, t).size();
// `to_u128` is a saturating cast, which is what we need
// (https://doc.rust-lang.org/nightly/nightly-rustc/rustc_apfloat/trait.Float.html#method.to_i128_r).
let v = f.to_u128(size.bits_usize()).value;
// This should already fit the bit width
Scalar::from_uint(v, size)
}
// float -> int
Int(t) => {
let size = Integer::from_int_ty(self, t).size();
// `to_i128` is a saturating cast, which is what we need
// (https://doc.rust-lang.org/nightly/nightly-rustc/rustc_apfloat/trait.Float.html#method.to_i128_r).
let v = f.to_i128(size.bits_usize()).value;
Scalar::from_int(v, size)
}
// float -> f32
Float(FloatTy::F32) => Scalar::from_f32(f.convert(&mut false).value),
// float -> f64
Float(FloatTy::F64) => Scalar::from_f64(f.convert(&mut false).value),
// That's it.
_ => span_bug!(self.cur_span(), "invalid float to {:?} cast", dest_ty),
}
}
fn unsize_into_ptr(
&mut self,
src: &OpTy<'tcx, M::PointerTag>,
dest: &PlaceTy<'tcx, M::PointerTag>,
// The pointee types
source_ty: Ty<'tcx>,
cast_ty: Ty<'tcx>,
) -> InterpResult<'tcx> {
// A<Struct> -> A<Trait> conversion
let (src_pointee_ty, dest_pointee_ty) =
self.tcx.struct_lockstep_tails_erasing_lifetimes(source_ty, cast_ty, self.param_env);
match (&src_pointee_ty.kind(), &dest_pointee_ty.kind()) {
(&ty::Array(_, length), &ty::Slice(_)) => {
let ptr = self.read_immediate(src)?.to_scalar()?;
// u64 cast is from usize to u64, which is always good
let val =
Immediate::new_slice(ptr, length.eval_usize(*self.tcx, self.param_env), self);
self.write_immediate(val, dest)
}
(&ty::Dynamic(ref data_a, ..), &ty::Dynamic(ref data_b, ..)) => {
let val = self.read_immediate(src)?;
if data_a.principal_def_id() == data_b.principal_def_id() {
return self.write_immediate(*val, dest);
}
// trait upcasting coercion
let vptr_entry_idx = self.tcx.vtable_trait_upcasting_coercion_new_vptr_slot((
src_pointee_ty,
dest_pointee_ty,
));
if let Some(entry_idx) = vptr_entry_idx {
let entry_idx = u64::try_from(entry_idx).unwrap();
let (old_data, old_vptr) = val.to_scalar_pair()?;
let old_vptr = self.scalar_to_ptr(old_vptr);
let new_vptr = self
.read_new_vtable_after_trait_upcasting_from_vtable(old_vptr, entry_idx)?;
self.write_immediate(Immediate::new_dyn_trait(old_data, new_vptr, self), dest)
} else {
self.write_immediate(*val, dest)
}
}
(_, &ty::Dynamic(ref data, _)) => {
// Initial cast from sized to dyn trait
let vtable = self.get_vtable(src_pointee_ty, data.principal())?;
let ptr = self.read_immediate(src)?.to_scalar()?;
let val = Immediate::new_dyn_trait(ptr, vtable, &*self.tcx);
self.write_immediate(val, dest)
}
_ => {
span_bug!(self.cur_span(), "invalid unsizing {:?} -> {:?}", src.layout.ty, cast_ty)
}
}
}
fn unsize_into(
&mut self,
src: &OpTy<'tcx, M::PointerTag>,
cast_ty: TyAndLayout<'tcx>,
dest: &PlaceTy<'tcx, M::PointerTag>,
) -> InterpResult<'tcx> {
trace!("Unsizing {:?} of type {} into {:?}", *src, src.layout.ty, cast_ty.ty);
match (&src.layout.ty.kind(), &cast_ty.ty.kind()) {
(&ty::Ref(_, s, _), &ty::Ref(_, c, _) | &ty::RawPtr(TypeAndMut { ty: c, .. }))
| (&ty::RawPtr(TypeAndMut { ty: s, .. }), &ty::RawPtr(TypeAndMut { ty: c, .. })) => {
self.unsize_into_ptr(src, dest, s, c)
}
(&ty::Adt(def_a, _), &ty::Adt(def_b, _)) => {
assert_eq!(def_a, def_b);
if def_a.is_box() || def_b.is_box() {
if !def_a.is_box() || !def_b.is_box() {
span_bug!(
self.cur_span(),
"invalid unsizing between {:?} -> {:?}",
src.layout.ty,
cast_ty.ty
);
}
return self.unsize_into_ptr(
src,
dest,
src.layout.ty.boxed_ty(),
cast_ty.ty.boxed_ty(),
);
}
// unsizing of generic struct with pointer fields
// Example: `Arc<T>` -> `Arc<Trait>`
// here we need to increase the size of every &T thin ptr field to a fat ptr
for i in 0..src.layout.fields.count() {
let cast_ty_field = cast_ty.field(self, i);
if cast_ty_field.is_zst() {
continue;
}
let src_field = self.operand_field(src, i)?;
let dst_field = self.place_field(dest, i)?;
if src_field.layout.ty == cast_ty_field.ty {
self.copy_op(&src_field, &dst_field)?;
} else {
self.unsize_into(&src_field, cast_ty_field, &dst_field)?;
}
}
Ok(())
}
_ => span_bug!(
self.cur_span(),
"unsize_into: invalid conversion: {:?} -> {:?}",
src.layout,
dest.layout
),
}
}
}

1049
compiler/rustc_const_eval/src/interpret/eval_context.rs Normal file
View File

File diff suppressed because it is too large Load Diff

437
compiler/rustc_const_eval/src/interpret/intern.rs Normal file
View File

@@ -0,0 +1,437 @@
//! This module specifies the type based interner for constants.
//!
//! After a const evaluation has computed a value, before we destroy the const evaluator's session
//! memory, we need to extract all memory allocations to the global memory pool so they stay around.
//!
//! In principle, this is not very complicated: we recursively walk the final value, follow all the
//! pointers, and move all reachable allocations to the global `tcx` memory. The only complication
//! is picking the right mutability for the allocations in a `static` initializer: we want to make
//! as many allocations as possible immutable so LLVM can put them into read-only memory. At the
//! same time, we need to make memory that could be mutated by the program mutable to avoid
//! incorrect compilations. To achieve this, we do a type-based traversal of the final value,
//! tracking mutable and shared references and `UnsafeCell` to determine the current mutability.
//! (In principle, we could skip this type-based part for `const` and promoteds, as they need to be
//! always immutable. At least for `const` however we use this opportunity to reject any `const`
//! that contains allocations whose mutability we cannot identify.)
use super::validity::RefTracking;
use rustc_data_structures::fx::{FxHashMap, FxHashSet};
use rustc_errors::ErrorReported;
use rustc_hir as hir;
use rustc_middle::mir::interpret::InterpResult;
use rustc_middle::ty::{self, layout::TyAndLayout, Ty};
use rustc_ast::Mutability;
use super::{AllocId, Allocation, InterpCx, MPlaceTy, Machine, MemoryKind, PlaceTy, ValueVisitor};
use crate::const_eval;
pub trait CompileTimeMachine<'mir, 'tcx, T> = Machine<
'mir,
'tcx,
MemoryKind = T,
PointerTag = AllocId,
ExtraFnVal = !,
FrameExtra = (),
AllocExtra = (),
MemoryMap = FxHashMap<AllocId, (MemoryKind<T>, Allocation)>,
>;
struct InternVisitor<'rt, 'mir, 'tcx, M: CompileTimeMachine<'mir, 'tcx, const_eval::MemoryKind>> {
/// The ectx from which we intern.
ecx: &'rt mut InterpCx<'mir, 'tcx, M>,
/// Previously encountered safe references.
ref_tracking: &'rt mut RefTracking<(MPlaceTy<'tcx>, InternMode)>,
/// A list of all encountered allocations. After type-based interning, we traverse this list to
/// also intern allocations that are only referenced by a raw pointer or inside a union.
leftover_allocations: &'rt mut FxHashSet<AllocId>,
/// The root kind of the value that we're looking at. This field is never mutated for a
/// particular allocation. It is primarily used to make as many allocations as possible
/// read-only so LLVM can place them in const memory.
mode: InternMode,
/// This field stores whether we are *currently* inside an `UnsafeCell`. This can affect
/// the intern mode of references we encounter.
inside_unsafe_cell: bool,
}
#[derive(Copy, Clone, Debug, PartialEq, Hash, Eq)]
enum InternMode {
/// A static and its current mutability. Below shared references inside a `static mut`,
/// this is *immutable*, and below mutable references inside an `UnsafeCell`, this
/// is *mutable*.
Static(hir::Mutability),
/// A `const`.
Const,
}
/// Signalling data structure to ensure we don't recurse
/// into the memory of other constants or statics
struct IsStaticOrFn;
/// Intern an allocation without looking at its children.
/// `mode` is the mode of the environment where we found this pointer.
/// `mutablity` is the mutability of the place to be interned; even if that says
/// `immutable` things might become mutable if `ty` is not frozen.
/// `ty` can be `None` if there is no potential interior mutability
/// to account for (e.g. for vtables).
fn intern_shallow<'rt, 'mir, 'tcx, M: CompileTimeMachine<'mir, 'tcx, const_eval::MemoryKind>>(
ecx: &'rt mut InterpCx<'mir, 'tcx, M>,
leftover_allocations: &'rt mut FxHashSet<AllocId>,
alloc_id: AllocId,
mode: InternMode,
ty: Option<Ty<'tcx>>,
) -> Option<IsStaticOrFn> {
trace!("intern_shallow {:?} with {:?}", alloc_id, mode);
// remove allocation
let tcx = ecx.tcx;
let (kind, mut alloc) = match ecx.memory.alloc_map.remove(&alloc_id) {
Some(entry) => entry,
None => {
// Pointer not found in local memory map. It is either a pointer to the global
// map, or dangling.
// If the pointer is dangling (neither in local nor global memory), we leave it
// to validation to error -- it has the much better error messages, pointing out where
// in the value the dangling reference lies.
// The `delay_span_bug` ensures that we don't forget such a check in validation.
if tcx.get_global_alloc(alloc_id).is_none() {
tcx.sess.delay_span_bug(ecx.tcx.span, "tried to intern dangling pointer");
}
// treat dangling pointers like other statics
// just to stop trying to recurse into them
return Some(IsStaticOrFn);
}
};
// This match is just a canary for future changes to `MemoryKind`, which most likely need
// changes in this function.
match kind {
MemoryKind::Stack
| MemoryKind::Machine(const_eval::MemoryKind::Heap)
| MemoryKind::CallerLocation => {}
}
// Set allocation mutability as appropriate. This is used by LLVM to put things into
// read-only memory, and also by Miri when evaluating other globals that
// access this one.
if let InternMode::Static(mutability) = mode {
// For this, we need to take into account `UnsafeCell`. When `ty` is `None`, we assume
// no interior mutability.
let frozen = ty.map_or(true, |ty| ty.is_freeze(ecx.tcx, ecx.param_env));
// For statics, allocation mutability is the combination of place mutability and
// type mutability.
// The entire allocation needs to be mutable if it contains an `UnsafeCell` anywhere.
let immutable = mutability == Mutability::Not && frozen;
if immutable {
alloc.mutability = Mutability::Not;
} else {
// Just making sure we are not "upgrading" an immutable allocation to mutable.
assert_eq!(alloc.mutability, Mutability::Mut);
}
} else {
// No matter what, *constants are never mutable*. Mutating them is UB.
// See const_eval::machine::MemoryExtra::can_access_statics for why
// immutability is so important.
// Validation will ensure that there is no `UnsafeCell` on an immutable allocation.
alloc.mutability = Mutability::Not;
};
// link the alloc id to the actual allocation
let alloc = tcx.intern_const_alloc(alloc);
leftover_allocations.extend(alloc.relocations().iter().map(|&(_, alloc_id)| alloc_id));
tcx.set_alloc_id_memory(alloc_id, alloc);
None
}
impl<'rt, 'mir, 'tcx, M: CompileTimeMachine<'mir, 'tcx, const_eval::MemoryKind>>
InternVisitor<'rt, 'mir, 'tcx, M>
{
fn intern_shallow(
&mut self,
alloc_id: AllocId,
mode: InternMode,
ty: Option<Ty<'tcx>>,
) -> Option<IsStaticOrFn> {
intern_shallow(self.ecx, self.leftover_allocations, alloc_id, mode, ty)
}
}
impl<'rt, 'mir, 'tcx: 'mir, M: CompileTimeMachine<'mir, 'tcx, const_eval::MemoryKind>>
ValueVisitor<'mir, 'tcx, M> for InternVisitor<'rt, 'mir, 'tcx, M>
{
type V = MPlaceTy<'tcx>;
#[inline(always)]
fn ecx(&self) -> &InterpCx<'mir, 'tcx, M> {
&self.ecx
}
fn visit_aggregate(
&mut self,
mplace: &MPlaceTy<'tcx>,
fields: impl Iterator<Item = InterpResult<'tcx, Self::V>>,
) -> InterpResult<'tcx> {
// ZSTs cannot contain pointers, so we can skip them.
if mplace.layout.is_zst() {
return Ok(());
}
if let Some(def) = mplace.layout.ty.ty_adt_def() {
if Some(def.did) == self.ecx.tcx.lang_items().unsafe_cell_type() {
// We are crossing over an `UnsafeCell`, we can mutate again. This means that
// References we encounter inside here are interned as pointing to mutable
// allocations.
// Remember the `old` value to handle nested `UnsafeCell`.
let old = std::mem::replace(&mut self.inside_unsafe_cell, true);
let walked = self.walk_aggregate(mplace, fields);
self.inside_unsafe_cell = old;
return walked;
}
}
self.walk_aggregate(mplace, fields)
}
fn visit_value(&mut self, mplace: &MPlaceTy<'tcx>) -> InterpResult<'tcx> {
// Handle Reference types, as these are the only relocations supported by const eval.
// Raw pointers (and boxes) are handled by the `leftover_relocations` logic.
let tcx = self.ecx.tcx;
let ty = mplace.layout.ty;
if let ty::Ref(_, referenced_ty, ref_mutability) = *ty.kind() {
let value = self.ecx.read_immediate(&(*mplace).into())?;
let mplace = self.ecx.ref_to_mplace(&value)?;
assert_eq!(mplace.layout.ty, referenced_ty);
// Handle trait object vtables.
if let ty::Dynamic(..) =
tcx.struct_tail_erasing_lifetimes(referenced_ty, self.ecx.param_env).kind()
{
let ptr = self.ecx.scalar_to_ptr(mplace.meta.unwrap_meta());
if let Some(alloc_id) = ptr.provenance {
// Explicitly choose const mode here, since vtables are immutable, even
// if the reference of the fat pointer is mutable.
self.intern_shallow(alloc_id, InternMode::Const, None);
} else {
// Validation will error (with a better message) on an invalid vtable pointer.
// Let validation show the error message, but make sure it *does* error.
tcx.sess
.delay_span_bug(tcx.span, "vtables pointers cannot be integer pointers");
}
}
// Check if we have encountered this pointer+layout combination before.
// Only recurse for allocation-backed pointers.
if let Some(alloc_id) = mplace.ptr.provenance {
// Compute the mode with which we intern this. Our goal here is to make as many
// statics as we can immutable so they can be placed in read-only memory by LLVM.
let ref_mode = match self.mode {
InternMode::Static(mutbl) => {
// In statics, merge outer mutability with reference mutability and
// take into account whether we are in an `UnsafeCell`.
// The only way a mutable reference actually works as a mutable reference is
// by being in a `static mut` directly or behind another mutable reference.
// If there's an immutable reference or we are inside a `static`, then our
// mutable reference is equivalent to an immutable one. As an example:
// `&&mut Foo` is semantically equivalent to `&&Foo`
match ref_mutability {
_ if self.inside_unsafe_cell => {
// Inside an `UnsafeCell` is like inside a `static mut`, the "outer"
// mutability does not matter.
InternMode::Static(ref_mutability)
}
Mutability::Not => {
// A shared reference, things become immutable.
// We do *not* consider `freeze` here: `intern_shallow` considers
// `freeze` for the actual mutability of this allocation; the intern
// mode for references contained in this allocation is tracked more
// precisely when traversing the referenced data (by tracking
// `UnsafeCell`). This makes sure that `&(&i32, &Cell<i32>)` still
// has the left inner reference interned into a read-only
// allocation.
InternMode::Static(Mutability::Not)
}
Mutability::Mut => {
// Mutable reference.
InternMode::Static(mutbl)
}
}
}
InternMode::Const => {
// Ignore `UnsafeCell`, everything is immutable. Validity does some sanity
// checking for mutable references that we encounter -- they must all be
// ZST.
InternMode::Const
}
};
match self.intern_shallow(alloc_id, ref_mode, Some(referenced_ty)) {
// No need to recurse, these are interned already and statics may have
// cycles, so we don't want to recurse there
Some(IsStaticOrFn) => {}
// intern everything referenced by this value. The mutability is taken from the
// reference. It is checked above that mutable references only happen in
// `static mut`
None => self.ref_tracking.track((mplace, ref_mode), || ()),
}
}
Ok(())
} else {
// Not a reference -- proceed recursively.
self.walk_value(mplace)
}
}
}
#[derive(Copy, Clone, Debug, PartialEq, Hash, Eq)]
pub enum InternKind {
/// The `mutability` of the static, ignoring the type which may have interior mutability.
Static(hir::Mutability),
Constant,
Promoted,
}
/// Intern `ret` and everything it references.
///
/// This *cannot raise an interpreter error*. Doing so is left to validation, which
/// tracks where in the value we are and thus can show much better error messages.
/// Any errors here would anyway be turned into `const_err` lints, whereas validation failures
/// are hard errors.
#[tracing::instrument(level = "debug", skip(ecx))]
pub fn intern_const_alloc_recursive<M: CompileTimeMachine<'mir, 'tcx, const_eval::MemoryKind>>(
ecx: &mut InterpCx<'mir, 'tcx, M>,
intern_kind: InternKind,
ret: &MPlaceTy<'tcx>,
) -> Result<(), ErrorReported>
where
'tcx: 'mir,
{
let tcx = ecx.tcx;
let base_intern_mode = match intern_kind {
InternKind::Static(mutbl) => InternMode::Static(mutbl),
// `Constant` includes array lengths.
InternKind::Constant | InternKind::Promoted => InternMode::Const,
};
// Type based interning.
// `ref_tracking` tracks typed references we have already interned and still need to crawl for
// more typed information inside them.
// `leftover_allocations` collects *all* allocations we see, because some might not
// be available in a typed way. They get interned at the end.
let mut ref_tracking = RefTracking::empty();
let leftover_allocations = &mut FxHashSet::default();
// start with the outermost allocation
intern_shallow(
ecx,
leftover_allocations,
// The outermost allocation must exist, because we allocated it with
// `Memory::allocate`.
ret.ptr.provenance.unwrap(),
base_intern_mode,
Some(ret.layout.ty),
);
ref_tracking.track((*ret, base_intern_mode), || ());
while let Some(((mplace, mode), _)) = ref_tracking.todo.pop() {
let res = InternVisitor {
ref_tracking: &mut ref_tracking,
ecx,
mode,
leftover_allocations,
inside_unsafe_cell: false,
}
.visit_value(&mplace);
// We deliberately *ignore* interpreter errors here. When there is a problem, the remaining
// references are "leftover"-interned, and later validation will show a proper error
// and point at the right part of the value causing the problem.
match res {
Ok(()) => {}
Err(error) => {
ecx.tcx.sess.delay_span_bug(
ecx.tcx.span,
&format!(
"error during interning should later cause validation failure: {}",
error
),
);
}
}
}
// Intern the rest of the allocations as mutable. These might be inside unions, padding, raw
// pointers, ... So we can't intern them according to their type rules
let mut todo: Vec<_> = leftover_allocations.iter().cloned().collect();
while let Some(alloc_id) = todo.pop() {
if let Some((_, mut alloc)) = ecx.memory.alloc_map.remove(&alloc_id) {
// We can't call the `intern_shallow` method here, as its logic is tailored to safe
// references and a `leftover_allocations` set (where we only have a todo-list here).
// So we hand-roll the interning logic here again.
match intern_kind {
// Statics may contain mutable allocations even behind relocations.
// Even for immutable statics it would be ok to have mutable allocations behind
// raw pointers, e.g. for `static FOO: *const AtomicUsize = &AtomicUsize::new(42)`.
InternKind::Static(_) => {}
// Raw pointers in promoteds may only point to immutable things so we mark
// everything as immutable.
// It is UB to mutate through a raw pointer obtained via an immutable reference:
// Since all references and pointers inside a promoted must by their very definition
// be created from an immutable reference (and promotion also excludes interior
// mutability), mutating through them would be UB.
// There's no way we can check whether the user is using raw pointers correctly,
// so all we can do is mark this as immutable here.
InternKind::Promoted => {
// See const_eval::machine::MemoryExtra::can_access_statics for why
// immutability is so important.
alloc.mutability = Mutability::Not;
}
InternKind::Constant => {
// If it's a constant, we should not have any "leftovers" as everything
// is tracked by const-checking.
// FIXME: downgrade this to a warning? It rejects some legitimate consts,
// such as `const CONST_RAW: *const Vec<i32> = &Vec::new() as *const _;`.
ecx.tcx
.sess
.span_err(ecx.tcx.span, "untyped pointers are not allowed in constant");
// For better errors later, mark the allocation as immutable.
alloc.mutability = Mutability::Not;
}
}
let alloc = tcx.intern_const_alloc(alloc);
tcx.set_alloc_id_memory(alloc_id, alloc);
for &(_, alloc_id) in alloc.relocations().iter() {
if leftover_allocations.insert(alloc_id) {
todo.push(alloc_id);
}
}
} else if ecx.memory.dead_alloc_map.contains_key(&alloc_id) {
// Codegen does not like dangling pointers, and generally `tcx` assumes that
// all allocations referenced anywhere actually exist. So, make sure we error here.
ecx.tcx.sess.span_err(ecx.tcx.span, "encountered dangling pointer in final constant");
return Err(ErrorReported);
} else if ecx.tcx.get_global_alloc(alloc_id).is_none() {
// We have hit an `AllocId` that is neither in local or global memory and isn't
// marked as dangling by local memory. That should be impossible.
span_bug!(ecx.tcx.span, "encountered unknown alloc id {:?}", alloc_id);
}
}
Ok(())
}
impl<'mir, 'tcx: 'mir, M: super::intern::CompileTimeMachine<'mir, 'tcx, !>>
InterpCx<'mir, 'tcx, M>
{
/// A helper function that allocates memory for the layout given and gives you access to mutate
/// it. Once your own mutation code is done, the backing `Allocation` is removed from the
/// current `Memory` and returned.
pub fn intern_with_temp_alloc(
&mut self,
layout: TyAndLayout<'tcx>,
f: impl FnOnce(
&mut InterpCx<'mir, 'tcx, M>,
&PlaceTy<'tcx, M::PointerTag>,
) -> InterpResult<'tcx, ()>,
) -> InterpResult<'tcx, &'tcx Allocation> {
let dest = self.allocate(layout, MemoryKind::Stack)?;
f(self, &dest.into())?;
let mut alloc = self.memory.alloc_map.remove(&dest.ptr.provenance.unwrap()).unwrap().1;
alloc.mutability = Mutability::Not;
Ok(self.tcx.intern_const_alloc(alloc))
}
}

585
compiler/rustc_const_eval/src/interpret/intrinsics.rs Normal file
View File

@@ -0,0 +1,585 @@
//! Intrinsics and other functions that the miri engine executes without
//! looking at their MIR. Intrinsics/functions supported here are shared by CTFE
//! and miri.
use std::convert::TryFrom;
use rustc_hir::def_id::DefId;
use rustc_middle::mir::{
self,
interpret::{ConstValue, GlobalId, InterpResult, Scalar},
BinOp,
};
use rustc_middle::ty;
use rustc_middle::ty::layout::LayoutOf as _;
use rustc_middle::ty::subst::SubstsRef;
use rustc_middle::ty::{Ty, TyCtxt};
use rustc_span::symbol::{sym, Symbol};
use rustc_target::abi::{Abi, Align, Primitive, Size};
use super::{
util::ensure_monomorphic_enough, CheckInAllocMsg, ImmTy, InterpCx, Machine, OpTy, PlaceTy,
Pointer,
};
mod caller_location;
mod type_name;
fn numeric_intrinsic<Tag>(name: Symbol, bits: u128, kind: Primitive) -> Scalar<Tag> {
let size = match kind {
Primitive::Int(integer, _) => integer.size(),
_ => bug!("invalid `{}` argument: {:?}", name, bits),
};
let extra = 128 - u128::from(size.bits());
let bits_out = match name {
sym::ctpop => u128::from(bits.count_ones()),
sym::ctlz => u128::from(bits.leading_zeros()) - extra,
sym::cttz => u128::from((bits << extra).trailing_zeros()) - extra,
sym::bswap => (bits << extra).swap_bytes(),
sym::bitreverse => (bits << extra).reverse_bits(),
_ => bug!("not a numeric intrinsic: {}", name),
};
Scalar::from_uint(bits_out, size)
}
/// The logic for all nullary intrinsics is implemented here. These intrinsics don't get evaluated
/// inside an `InterpCx` and instead have their value computed directly from rustc internal info.
crate fn eval_nullary_intrinsic<'tcx>(
tcx: TyCtxt<'tcx>,
param_env: ty::ParamEnv<'tcx>,
def_id: DefId,
substs: SubstsRef<'tcx>,
) -> InterpResult<'tcx, ConstValue<'tcx>> {
let tp_ty = substs.type_at(0);
let name = tcx.item_name(def_id);
Ok(match name {
sym::type_name => {
ensure_monomorphic_enough(tcx, tp_ty)?;
let alloc = type_name::alloc_type_name(tcx, tp_ty);
ConstValue::Slice { data: alloc, start: 0, end: alloc.len() }
}
sym::needs_drop => {
ensure_monomorphic_enough(tcx, tp_ty)?;
ConstValue::from_bool(tp_ty.needs_drop(tcx, param_env))
}
sym::min_align_of | sym::pref_align_of => {
// Correctly handles non-monomorphic calls, so there is no need for ensure_monomorphic_enough.
let layout = tcx.layout_of(param_env.and(tp_ty)).map_err(|e| err_inval!(Layout(e)))?;
let n = match name {
sym::pref_align_of => layout.align.pref.bytes(),
sym::min_align_of => layout.align.abi.bytes(),
_ => bug!(),
};
ConstValue::from_machine_usize(n, &tcx)
}
sym::type_id => {
ensure_monomorphic_enough(tcx, tp_ty)?;
ConstValue::from_u64(tcx.type_id_hash(tp_ty))
}
sym::variant_count => match tp_ty.kind() {
// Correctly handles non-monomorphic calls, so there is no need for ensure_monomorphic_enough.
ty::Adt(ref adt, _) => ConstValue::from_machine_usize(adt.variants.len() as u64, &tcx),
ty::Projection(_)
| ty::Opaque(_, _)
| ty::Param(_)
| ty::Bound(_, _)
| ty::Placeholder(_)
| ty::Infer(_) => throw_inval!(TooGeneric),
ty::Bool
| ty::Char
| ty::Int(_)
| ty::Uint(_)
| ty::Float(_)
| ty::Foreign(_)
| ty::Str
| ty::Array(_, _)
| ty::Slice(_)
| ty::RawPtr(_)
| ty::Ref(_, _, _)
| ty::FnDef(_, _)
| ty::FnPtr(_)
| ty::Dynamic(_, _)
| ty::Closure(_, _)
| ty::Generator(_, _, _)
| ty::GeneratorWitness(_)
| ty::Never
| ty::Tuple(_)
| ty::Error(_) => ConstValue::from_machine_usize(0u64, &tcx),
},
other => bug!("`{}` is not a zero arg intrinsic", other),
})
}
impl<'mir, 'tcx: 'mir, M: Machine<'mir, 'tcx>> InterpCx<'mir, 'tcx, M> {
/// Returns `true` if emulation happened.
/// Here we implement the intrinsics that are common to all Miri instances; individual machines can add their own
/// intrinsic handling.
pub fn emulate_intrinsic(
&mut self,
instance: ty::Instance<'tcx>,
args: &[OpTy<'tcx, M::PointerTag>],
ret: Option<(&PlaceTy<'tcx, M::PointerTag>, mir::BasicBlock)>,
) -> InterpResult<'tcx, bool> {
let substs = instance.substs;
let intrinsic_name = self.tcx.item_name(instance.def_id());
// First handle intrinsics without return place.
let (dest, ret) = match ret {
None => match intrinsic_name {
sym::transmute => throw_ub_format!("transmuting to uninhabited type"),
sym::abort => M::abort(self, "the program aborted execution".to_owned())?,
// Unsupported diverging intrinsic.
_ => return Ok(false),
},
Some(p) => p,
};
// Keep the patterns in this match ordered the same as the list in
// `src/librustc_middle/ty/constness.rs`
match intrinsic_name {
sym::caller_location => {
let span = self.find_closest_untracked_caller_location();
let location = self.alloc_caller_location_for_span(span);
self.write_immediate(location.to_ref(self), dest)?;
}
sym::min_align_of_val | sym::size_of_val => {
// Avoid `deref_operand` -- this is not a deref, the ptr does not have to be
// dereferencable!
let place = self.ref_to_mplace(&self.read_immediate(&args[0])?)?;
let (size, align) = self
.size_and_align_of_mplace(&place)?
.ok_or_else(|| err_unsup_format!("`extern type` does not have known layout"))?;
let result = match intrinsic_name {
sym::min_align_of_val => align.bytes(),
sym::size_of_val => size.bytes(),
_ => bug!(),
};
self.write_scalar(Scalar::from_machine_usize(result, self), dest)?;
}
sym::min_align_of
| sym::pref_align_of
| sym::needs_drop
| sym::type_id
| sym::type_name
| sym::variant_count => {
let gid = GlobalId { instance, promoted: None };
let ty = match intrinsic_name {
sym::min_align_of | sym::pref_align_of | sym::variant_count => {
self.tcx.types.usize
}
sym::needs_drop => self.tcx.types.bool,
sym::type_id => self.tcx.types.u64,
sym::type_name => self.tcx.mk_static_str(),
_ => bug!("already checked for nullary intrinsics"),
};
let val =
self.tcx.const_eval_global_id(self.param_env, gid, Some(self.tcx.span))?;
let val = self.const_val_to_op(val, ty, Some(dest.layout))?;
self.copy_op(&val, dest)?;
}
sym::ctpop
| sym::cttz
| sym::cttz_nonzero
| sym::ctlz
| sym::ctlz_nonzero
| sym::bswap
| sym::bitreverse => {
let ty = substs.type_at(0);
let layout_of = self.layout_of(ty)?;
let val = self.read_scalar(&args[0])?.check_init()?;
let bits = val.to_bits(layout_of.size)?;
let kind = match layout_of.abi {
Abi::Scalar(ref scalar) => scalar.value,
_ => span_bug!(
self.cur_span(),
"{} called on invalid type {:?}",
intrinsic_name,
ty
),
};
let (nonzero, intrinsic_name) = match intrinsic_name {
sym::cttz_nonzero => (true, sym::cttz),
sym::ctlz_nonzero => (true, sym::ctlz),
other => (false, other),
};
if nonzero && bits == 0 {
throw_ub_format!("`{}_nonzero` called on 0", intrinsic_name);
}
let out_val = numeric_intrinsic(intrinsic_name, bits, kind);
self.write_scalar(out_val, dest)?;
}
sym::add_with_overflow | sym::sub_with_overflow | sym::mul_with_overflow => {
let lhs = self.read_immediate(&args[0])?;
let rhs = self.read_immediate(&args[1])?;
let bin_op = match intrinsic_name {
sym::add_with_overflow => BinOp::Add,
sym::sub_with_overflow => BinOp::Sub,
sym::mul_with_overflow => BinOp::Mul,
_ => bug!("Already checked for int ops"),
};
self.binop_with_overflow(bin_op, &lhs, &rhs, dest)?;
}
sym::saturating_add | sym::saturating_sub => {
let l = self.read_immediate(&args[0])?;
let r = self.read_immediate(&args[1])?;
let is_add = intrinsic_name == sym::saturating_add;
let (val, overflowed, _ty) = self.overflowing_binary_op(
if is_add { BinOp::Add } else { BinOp::Sub },
&l,
&r,
)?;
let val = if overflowed {
let num_bits = l.layout.size.bits();
if l.layout.abi.is_signed() {
// For signed ints the saturated value depends on the sign of the first
// term since the sign of the second term can be inferred from this and
// the fact that the operation has overflowed (if either is 0 no
// overflow can occur)
let first_term: u128 = l.to_scalar()?.to_bits(l.layout.size)?;
let first_term_positive = first_term & (1 << (num_bits - 1)) == 0;
if first_term_positive {
// Negative overflow not possible since the positive first term
// can only increase an (in range) negative term for addition
// or corresponding negated positive term for subtraction
Scalar::from_uint(
(1u128 << (num_bits - 1)) - 1, // max positive
Size::from_bits(num_bits),
)
} else {
// Positive overflow not possible for similar reason
// max negative
Scalar::from_uint(1u128 << (num_bits - 1), Size::from_bits(num_bits))
}
} else {
// unsigned
if is_add {
// max unsigned
Scalar::from_uint(
u128::MAX >> (128 - num_bits),
Size::from_bits(num_bits),
)
} else {
// underflow to 0
Scalar::from_uint(0u128, Size::from_bits(num_bits))
}
}
} else {
val
};
self.write_scalar(val, dest)?;
}
sym::discriminant_value => {
let place = self.deref_operand(&args[0])?;
let discr_val = self.read_discriminant(&place.into())?.0;
self.write_scalar(discr_val, dest)?;
}
sym::unchecked_shl
| sym::unchecked_shr
| sym::unchecked_add
| sym::unchecked_sub
| sym::unchecked_mul
| sym::unchecked_div
| sym::unchecked_rem => {
let l = self.read_immediate(&args[0])?;
let r = self.read_immediate(&args[1])?;
let bin_op = match intrinsic_name {
sym::unchecked_shl => BinOp::Shl,
sym::unchecked_shr => BinOp::Shr,
sym::unchecked_add => BinOp::Add,
sym::unchecked_sub => BinOp::Sub,
sym::unchecked_mul => BinOp::Mul,
sym::unchecked_div => BinOp::Div,
sym::unchecked_rem => BinOp::Rem,
_ => bug!("Already checked for int ops"),
};
let (val, overflowed, _ty) = self.overflowing_binary_op(bin_op, &l, &r)?;
if overflowed {
let layout = self.layout_of(substs.type_at(0))?;
let r_val = r.to_scalar()?.to_bits(layout.size)?;
if let sym::unchecked_shl | sym::unchecked_shr = intrinsic_name {
throw_ub_format!("overflowing shift by {} in `{}`", r_val, intrinsic_name);
} else {
throw_ub_format!("overflow executing `{}`", intrinsic_name);
}
}
self.write_scalar(val, dest)?;
}
sym::rotate_left | sym::rotate_right => {
// rotate_left: (X << (S % BW)) | (X >> ((BW - S) % BW))
// rotate_right: (X << ((BW - S) % BW)) | (X >> (S % BW))
let layout = self.layout_of(substs.type_at(0))?;
let val = self.read_scalar(&args[0])?.check_init()?;
let val_bits = val.to_bits(layout.size)?;
let raw_shift = self.read_scalar(&args[1])?.check_init()?;
let raw_shift_bits = raw_shift.to_bits(layout.size)?;
let width_bits = u128::from(layout.size.bits());
let shift_bits = raw_shift_bits % width_bits;
let inv_shift_bits = (width_bits - shift_bits) % width_bits;
let result_bits = if intrinsic_name == sym::rotate_left {
(val_bits << shift_bits) | (val_bits >> inv_shift_bits)
} else {
(val_bits >> shift_bits) | (val_bits << inv_shift_bits)
};
let truncated_bits = self.truncate(result_bits, layout);
let result = Scalar::from_uint(truncated_bits, layout.size);
self.write_scalar(result, dest)?;
}
sym::copy => {
self.copy_intrinsic(&args[0], &args[1], &args[2], /*nonoverlapping*/ false)?;
}
sym::offset => {
let ptr = self.read_pointer(&args[0])?;
let offset_count = self.read_scalar(&args[1])?.to_machine_isize(self)?;
let pointee_ty = substs.type_at(0);
let offset_ptr = self.ptr_offset_inbounds(ptr, pointee_ty, offset_count)?;
self.write_pointer(offset_ptr, dest)?;
}
sym::arith_offset => {
let ptr = self.read_pointer(&args[0])?;
let offset_count = self.read_scalar(&args[1])?.to_machine_isize(self)?;
let pointee_ty = substs.type_at(0);
let pointee_size = i64::try_from(self.layout_of(pointee_ty)?.size.bytes()).unwrap();
let offset_bytes = offset_count.wrapping_mul(pointee_size);
let offset_ptr = ptr.wrapping_signed_offset(offset_bytes, self);
self.write_pointer(offset_ptr, dest)?;
}
sym::ptr_offset_from => {
let a = self.read_immediate(&args[0])?.to_scalar()?;
let b = self.read_immediate(&args[1])?.to_scalar()?;
// Special case: if both scalars are *equal integers*
// and not null, we pretend there is an allocation of size 0 right there,
// and their offset is 0. (There's never a valid object at null, making it an
// exception from the exception.)
// This is the dual to the special exception for offset-by-0
// in the inbounds pointer offset operation (see the Miri code, `src/operator.rs`).
//
// Control flow is weird because we cannot early-return (to reach the
// `go_to_block` at the end).
let done = if let (Ok(a), Ok(b)) = (a.try_to_int(), b.try_to_int()) {
let a = a.try_to_machine_usize(*self.tcx).unwrap();
let b = b.try_to_machine_usize(*self.tcx).unwrap();
if a == b && a != 0 {
self.write_scalar(Scalar::from_machine_isize(0, self), dest)?;
true
} else {
false
}
} else {
false
};
if !done {
// General case: we need two pointers.
let a = self.scalar_to_ptr(a);
let b = self.scalar_to_ptr(b);
let (a_alloc_id, a_offset, _) = self.memory.ptr_get_alloc(a)?;
let (b_alloc_id, b_offset, _) = self.memory.ptr_get_alloc(b)?;
if a_alloc_id != b_alloc_id {
throw_ub_format!(
"ptr_offset_from cannot compute offset of pointers into different \
allocations.",
);
}
let usize_layout = self.layout_of(self.tcx.types.usize)?;
let isize_layout = self.layout_of(self.tcx.types.isize)?;
let a_offset = ImmTy::from_uint(a_offset.bytes(), usize_layout);
let b_offset = ImmTy::from_uint(b_offset.bytes(), usize_layout);
let (val, _overflowed, _ty) =
self.overflowing_binary_op(BinOp::Sub, &a_offset, &b_offset)?;
let pointee_layout = self.layout_of(substs.type_at(0))?;
let val = ImmTy::from_scalar(val, isize_layout);
let size = ImmTy::from_int(pointee_layout.size.bytes(), isize_layout);
self.exact_div(&val, &size, dest)?;
}
}
sym::transmute => {
self.copy_op_transmute(&args[0], dest)?;
}
sym::assert_inhabited => {
let ty = instance.substs.type_at(0);
let layout = self.layout_of(ty)?;
if layout.abi.is_uninhabited() {
// The run-time intrinsic panics just to get a good backtrace; here we abort
// since there is no problem showing a backtrace even for aborts.
M::abort(
self,
format!(
"aborted execution: attempted to instantiate uninhabited type `{}`",
ty
),
)?;
}
}
sym::simd_insert => {
let index = u64::from(self.read_scalar(&args[1])?.to_u32()?);
let elem = &args[2];
let input = &args[0];
let (len, e_ty) = input.layout.ty.simd_size_and_type(*self.tcx);
assert!(
index < len,
"Index `{}` must be in bounds of vector type `{}`: `[0, {})`",
index,
e_ty,
len
);
assert_eq!(
input.layout, dest.layout,
"Return type `{}` must match vector type `{}`",
dest.layout.ty, input.layout.ty
);
assert_eq!(
elem.layout.ty, e_ty,
"Scalar element type `{}` must match vector element type `{}`",
elem.layout.ty, e_ty
);
for i in 0..len {
let place = self.place_index(dest, i)?;
let value = if i == index { *elem } else { self.operand_index(input, i)? };
self.copy_op(&value, &place)?;
}
}
sym::simd_extract => {
let index = u64::from(self.read_scalar(&args[1])?.to_u32()?);
let (len, e_ty) = args[0].layout.ty.simd_size_and_type(*self.tcx);
assert!(
index < len,
"index `{}` is out-of-bounds of vector type `{}` with length `{}`",
index,
e_ty,
len
);
assert_eq!(
e_ty, dest.layout.ty,
"Return type `{}` must match vector element type `{}`",
dest.layout.ty, e_ty
);
self.copy_op(&self.operand_index(&args[0], index)?, dest)?;
}
sym::likely | sym::unlikely | sym::black_box => {
// These just return their argument
self.copy_op(&args[0], dest)?;
}
sym::assume => {
let cond = self.read_scalar(&args[0])?.check_init()?.to_bool()?;
if !cond {
throw_ub_format!("`assume` intrinsic called with `false`");
}
}
sym::raw_eq => {
let result = self.raw_eq_intrinsic(&args[0], &args[1])?;
self.write_scalar(result, dest)?;
}
_ => return Ok(false),
}
trace!("{:?}", self.dump_place(**dest));
self.go_to_block(ret);
Ok(true)
}
pub fn exact_div(
&mut self,
a: &ImmTy<'tcx, M::PointerTag>,
b: &ImmTy<'tcx, M::PointerTag>,
dest: &PlaceTy<'tcx, M::PointerTag>,
) -> InterpResult<'tcx> {
// Performs an exact division, resulting in undefined behavior where
// `x % y != 0` or `y == 0` or `x == T::MIN && y == -1`.
// First, check x % y != 0 (or if that computation overflows).
let (res, overflow, _ty) = self.overflowing_binary_op(BinOp::Rem, &a, &b)?;
if overflow || res.assert_bits(a.layout.size) != 0 {
// Then, check if `b` is -1, which is the "MIN / -1" case.
let minus1 = Scalar::from_int(-1, dest.layout.size);
let b_scalar = b.to_scalar().unwrap();
if b_scalar == minus1 {
throw_ub_format!("exact_div: result of dividing MIN by -1 cannot be represented")
} else {
throw_ub_format!("exact_div: {} cannot be divided by {} without remainder", a, b,)
}
}
// `Rem` says this is all right, so we can let `Div` do its job.
self.binop_ignore_overflow(BinOp::Div, &a, &b, dest)
}
/// Offsets a pointer by some multiple of its type, returning an error if the pointer leaves its
/// allocation. For integer pointers, we consider each of them their own tiny allocation of size
/// 0, so offset-by-0 (and only 0) is okay -- except that null cannot be offset by _any_ value.
pub fn ptr_offset_inbounds(
&self,
ptr: Pointer<Option<M::PointerTag>>,
pointee_ty: Ty<'tcx>,
offset_count: i64,
) -> InterpResult<'tcx, Pointer<Option<M::PointerTag>>> {
// We cannot overflow i64 as a type's size must be <= isize::MAX.
let pointee_size = i64::try_from(self.layout_of(pointee_ty)?.size.bytes()).unwrap();
// The computed offset, in bytes, cannot overflow an isize.
let offset_bytes =
offset_count.checked_mul(pointee_size).ok_or(err_ub!(PointerArithOverflow))?;
// The offset being in bounds cannot rely on "wrapping around" the address space.
// So, first rule out overflows in the pointer arithmetic.
let offset_ptr = ptr.signed_offset(offset_bytes, self)?;
// ptr and offset_ptr must be in bounds of the same allocated object. This means all of the
// memory between these pointers must be accessible. Note that we do not require the
// pointers to be properly aligned (unlike a read/write operation).
let min_ptr = if offset_bytes >= 0 { ptr } else { offset_ptr };
let size = offset_bytes.unsigned_abs();
// This call handles checking for integer/null pointers.
self.memory.check_ptr_access_align(
min_ptr,
Size::from_bytes(size),
Align::ONE,
CheckInAllocMsg::PointerArithmeticTest,
)?;
Ok(offset_ptr)
}
/// Copy `count*size_of::<T>()` many bytes from `*src` to `*dst`.
pub(crate) fn copy_intrinsic(
&mut self,
src: &OpTy<'tcx, <M as Machine<'mir, 'tcx>>::PointerTag>,
dst: &OpTy<'tcx, <M as Machine<'mir, 'tcx>>::PointerTag>,
count: &OpTy<'tcx, <M as Machine<'mir, 'tcx>>::PointerTag>,
nonoverlapping: bool,
) -> InterpResult<'tcx> {
let count = self.read_scalar(&count)?.to_machine_usize(self)?;
let layout = self.layout_of(src.layout.ty.builtin_deref(true).unwrap().ty)?;
let (size, align) = (layout.size, layout.align.abi);
let size = size.checked_mul(count, self).ok_or_else(|| {
err_ub_format!(
"overflow computing total size of `{}`",
if nonoverlapping { "copy_nonoverlapping" } else { "copy" }
)
})?;
let src = self.read_pointer(&src)?;
let dst = self.read_pointer(&dst)?;
self.memory.copy(src, align, dst, align, size, nonoverlapping)
}
pub(crate) fn raw_eq_intrinsic(
&mut self,
lhs: &OpTy<'tcx, <M as Machine<'mir, 'tcx>>::PointerTag>,
rhs: &OpTy<'tcx, <M as Machine<'mir, 'tcx>>::PointerTag>,
) -> InterpResult<'tcx, Scalar<M::PointerTag>> {
let layout = self.layout_of(lhs.layout.ty.builtin_deref(true).unwrap().ty)?;
assert!(!layout.is_unsized());
let lhs = self.read_pointer(lhs)?;
let rhs = self.read_pointer(rhs)?;
let lhs_bytes = self.memory.read_bytes(lhs, layout.size)?;
let rhs_bytes = self.memory.read_bytes(rhs, layout.size)?;
Ok(Scalar::from_bool(lhs_bytes == rhs_bytes))
}
}

123
compiler/rustc_const_eval/src/interpret/intrinsics/caller_location.rs Normal file
View File

@@ -0,0 +1,123 @@
use std::convert::TryFrom;
use rustc_ast::Mutability;
use rustc_hir::lang_items::LangItem;
use rustc_middle::mir::TerminatorKind;
use rustc_middle::ty::layout::LayoutOf;
use rustc_middle::ty::subst::Subst;
use rustc_span::{Span, Symbol};
use crate::interpret::{
intrinsics::{InterpCx, Machine},
MPlaceTy, MemoryKind, Scalar,
};
impl<'mir, 'tcx: 'mir, M: Machine<'mir, 'tcx>> InterpCx<'mir, 'tcx, M> {
/// Walks up the callstack from the intrinsic's callsite, searching for the first callsite in a
/// frame which is not `#[track_caller]`.
crate fn find_closest_untracked_caller_location(&self) -> Span {
for frame in self.stack().iter().rev() {
debug!("find_closest_untracked_caller_location: checking frame {:?}", frame.instance);
// Assert that the frame we look at is actually executing code currently
// (`loc` is `Err` when we are unwinding and the frame does not require cleanup).
let loc = frame.loc.unwrap();
// This could be a non-`Call` terminator (such as `Drop`), or not a terminator at all
// (such as `box`). Use the normal span by default.
let mut source_info = *frame.body.source_info(loc);
// If this is a `Call` terminator, use the `fn_span` instead.
let block = &frame.body.basic_blocks()[loc.block];
if loc.statement_index == block.statements.len() {
debug!(
"find_closest_untracked_caller_location: got terminator {:?} ({:?})",
block.terminator(),
block.terminator().kind
);
if let TerminatorKind::Call { fn_span, .. } = block.terminator().kind {
source_info.span = fn_span;
}
}
// Walk up the `SourceScope`s, in case some of them are from MIR inlining.
// If so, the starting `source_info.span` is in the innermost inlined
// function, and will be replaced with outer callsite spans as long
// as the inlined functions were `#[track_caller]`.
loop {
let scope_data = &frame.body.source_scopes[source_info.scope];
if let Some((callee, callsite_span)) = scope_data.inlined {
// Stop inside the most nested non-`#[track_caller]` function,
// before ever reaching its caller (which is irrelevant).
if !callee.def.requires_caller_location(*self.tcx) {
return source_info.span;
}
source_info.span = callsite_span;
}
// Skip past all of the parents with `inlined: None`.
match scope_data.inlined_parent_scope {
Some(parent) => source_info.scope = parent,
None => break,
}
}
// Stop inside the most nested non-`#[track_caller]` function,
// before ever reaching its caller (which is irrelevant).
if !frame.instance.def.requires_caller_location(*self.tcx) {
return source_info.span;
}
}
bug!("no non-`#[track_caller]` frame found")
}
/// Allocate a `const core::panic::Location` with the provided filename and line/column numbers.
crate fn alloc_caller_location(
&mut self,
filename: Symbol,
line: u32,
col: u32,
) -> MPlaceTy<'tcx, M::PointerTag> {
let file =
self.allocate_str(&filename.as_str(), MemoryKind::CallerLocation, Mutability::Not);
let line = Scalar::from_u32(line);
let col = Scalar::from_u32(col);
// Allocate memory for `CallerLocation` struct.
let loc_ty = self
.tcx
.type_of(self.tcx.require_lang_item(LangItem::PanicLocation, None))
.subst(*self.tcx, self.tcx.mk_substs([self.tcx.lifetimes.re_erased.into()].iter()));
let loc_layout = self.layout_of(loc_ty).unwrap();
// This can fail if rustc runs out of memory right here. Trying to emit an error would be
// pointless, since that would require allocating more memory than a Location.
let location = self.allocate(loc_layout, MemoryKind::CallerLocation).unwrap();
// Initialize fields.
self.write_immediate(file.to_ref(self), &self.mplace_field(&location, 0).unwrap().into())
.expect("writing to memory we just allocated cannot fail");
self.write_scalar(line, &self.mplace_field(&location, 1).unwrap().into())
.expect("writing to memory we just allocated cannot fail");
self.write_scalar(col, &self.mplace_field(&location, 2).unwrap().into())
.expect("writing to memory we just allocated cannot fail");
location
}
crate fn location_triple_for_span(&self, span: Span) -> (Symbol, u32, u32) {
let topmost = span.ctxt().outer_expn().expansion_cause().unwrap_or(span);
let caller = self.tcx.sess.source_map().lookup_char_pos(topmost.lo());
(
Symbol::intern(&caller.file.name.prefer_remapped().to_string_lossy()),
u32::try_from(caller.line).unwrap(),
u32::try_from(caller.col_display).unwrap().checked_add(1).unwrap(),
)
}
pub fn alloc_caller_location_for_span(&mut self, span: Span) -> MPlaceTy<'tcx, M::PointerTag> {
let (file, line, column) = self.location_triple_for_span(span);
self.alloc_caller_location(file, line, column)
}
}

197
compiler/rustc_const_eval/src/interpret/intrinsics/type_name.rs Normal file
View File

@@ -0,0 +1,197 @@
use rustc_hir::def_id::CrateNum;
use rustc_hir::definitions::DisambiguatedDefPathData;
use rustc_middle::mir::interpret::Allocation;
use rustc_middle::ty::{
self,
print::{PrettyPrinter, Print, Printer},
subst::{GenericArg, GenericArgKind},
Ty, TyCtxt,
};
use std::fmt::Write;
struct AbsolutePathPrinter<'tcx> {
tcx: TyCtxt<'tcx>,
path: String,
}
impl<'tcx> Printer<'tcx> for AbsolutePathPrinter<'tcx> {
type Error = std::fmt::Error;
type Path = Self;
type Region = Self;
type Type = Self;
type DynExistential = Self;
type Const = Self;
fn tcx(&self) -> TyCtxt<'tcx> {
self.tcx
}
fn print_region(self, _region: ty::Region<'_>) -> Result<Self::Region, Self::Error> {
Ok(self)
}
fn print_type(mut self, ty: Ty<'tcx>) -> Result<Self::Type, Self::Error> {
match *ty.kind() {
// Types without identity.
ty::Bool
| ty::Char
| ty::Int(_)
| ty::Uint(_)
| ty::Float(_)
| ty::Str
| ty::Array(_, _)
| ty::Slice(_)
| ty::RawPtr(_)
| ty::Ref(_, _, _)
| ty::FnPtr(_)
| ty::Never
| ty::Tuple(_)
| ty::Dynamic(_, _) => self.pretty_print_type(ty),
// Placeholders (all printed as `_` to uniformize them).
ty::Param(_) | ty::Bound(..) | ty::Placeholder(_) | ty::Infer(_) | ty::Error(_) => {
write!(self, "_")?;
Ok(self)
}
// Types with identity (print the module path).
ty::Adt(&ty::AdtDef { did: def_id, .. }, substs)
| ty::FnDef(def_id, substs)
| ty::Opaque(def_id, substs)
| ty::Projection(ty::ProjectionTy { item_def_id: def_id, substs })
| ty::Closure(def_id, substs)
| ty::Generator(def_id, substs, _) => self.print_def_path(def_id, substs),
ty::Foreign(def_id) => self.print_def_path(def_id, &[]),
ty::GeneratorWitness(_) => bug!("type_name: unexpected `GeneratorWitness`"),
}
}
fn print_const(self, ct: &'tcx ty::Const<'tcx>) -> Result<Self::Const, Self::Error> {
self.pretty_print_const(ct, false)
}
fn print_dyn_existential(
mut self,
predicates: &'tcx ty::List<ty::Binder<'tcx, ty::ExistentialPredicate<'tcx>>>,
) -> Result<Self::DynExistential, Self::Error> {
let mut first = true;
for p in predicates {
if !first {
write!(self, "+")?;
}
first = false;
self = p.print(self)?;
}
Ok(self)
}
fn path_crate(mut self, cnum: CrateNum) -> Result<Self::Path, Self::Error> {
self.path.push_str(&self.tcx.crate_name(cnum).as_str());
Ok(self)
}
fn path_qualified(
self,
self_ty: Ty<'tcx>,
trait_ref: Option<ty::TraitRef<'tcx>>,
) -> Result<Self::Path, Self::Error> {
self.pretty_path_qualified(self_ty, trait_ref)
}
fn path_append_impl(
self,
print_prefix: impl FnOnce(Self) -> Result<Self::Path, Self::Error>,
_disambiguated_data: &DisambiguatedDefPathData,
self_ty: Ty<'tcx>,
trait_ref: Option<ty::TraitRef<'tcx>>,
) -> Result<Self::Path, Self::Error> {
self.pretty_path_append_impl(
|mut cx| {
cx = print_prefix(cx)?;
cx.path.push_str("::");
Ok(cx)
},
self_ty,
trait_ref,
)
}
fn path_append(
mut self,
print_prefix: impl FnOnce(Self) -> Result<Self::Path, Self::Error>,
disambiguated_data: &DisambiguatedDefPathData,
) -> Result<Self::Path, Self::Error> {
self = print_prefix(self)?;
write!(self.path, "::{}", disambiguated_data.data).unwrap();
Ok(self)
}
fn path_generic_args(
mut self,
print_prefix: impl FnOnce(Self) -> Result<Self::Path, Self::Error>,
args: &[GenericArg<'tcx>],
) -> Result<Self::Path, Self::Error> {
self = print_prefix(self)?;
let args = args.iter().cloned().filter(|arg| match arg.unpack() {
GenericArgKind::Lifetime(_) => false,
_ => true,
});
if args.clone().next().is_some() {
self.generic_delimiters(|cx| cx.comma_sep(args))
} else {
Ok(self)
}
}
}
impl PrettyPrinter<'tcx> for AbsolutePathPrinter<'tcx> {
fn region_should_not_be_omitted(&self, _region: ty::Region<'_>) -> bool {
false
}
fn comma_sep<T>(mut self, mut elems: impl Iterator<Item = T>) -> Result<Self, Self::Error>
where
T: Print<'tcx, Self, Output = Self, Error = Self::Error>,
{
if let Some(first) = elems.next() {
self = first.print(self)?;
for elem in elems {
self.path.push_str(", ");
self = elem.print(self)?;
}
}
Ok(self)
}
fn generic_delimiters(
mut self,
f: impl FnOnce(Self) -> Result<Self, Self::Error>,
) -> Result<Self, Self::Error> {
write!(self, "<")?;
self = f(self)?;
write!(self, ">")?;
Ok(self)
}
}
impl Write for AbsolutePathPrinter<'_> {
fn write_str(&mut self, s: &str) -> std::fmt::Result {
self.path.push_str(s);
Ok(())
}
}
/// Directly returns an `Allocation` containing an absolute path representation of the given type.
crate fn alloc_type_name<'tcx>(tcx: TyCtxt<'tcx>, ty: Ty<'tcx>) -> &'tcx Allocation {
let path = AbsolutePathPrinter { tcx, path: String::new() }.print_type(ty).unwrap().path;
let alloc = Allocation::from_bytes_byte_aligned_immutable(path.into_bytes());
tcx.intern_const_alloc(alloc)
}

479
compiler/rustc_const_eval/src/interpret/machine.rs Normal file
View File

@@ -0,0 +1,479 @@
//! This module contains everything needed to instantiate an interpreter.
//! This separation exists to ensure that no fancy miri features like
//! interpreting common C functions leak into CTFE.
use std::borrow::{Borrow, Cow};
use std::fmt::Debug;
use std::hash::Hash;
use rustc_middle::mir;
use rustc_middle::ty::{self, Ty};
use rustc_span::def_id::DefId;
use rustc_target::abi::Size;
use rustc_target::spec::abi::Abi;
use super::{
AllocId, AllocRange, Allocation, Frame, ImmTy, InterpCx, InterpResult, LocalValue, MemPlace,
Memory, MemoryKind, OpTy, Operand, PlaceTy, Pointer, Provenance, Scalar, StackPopUnwind,
};
/// Data returned by Machine::stack_pop,
/// to provide further control over the popping of the stack frame
#[derive(Eq, PartialEq, Debug, Copy, Clone)]
pub enum StackPopJump {
/// Indicates that no special handling should be
/// done - we'll either return normally or unwind
/// based on the terminator for the function
/// we're leaving.
Normal,
/// Indicates that we should *not* jump to the return/unwind address, as the callback already
/// took care of everything.
NoJump,
}
/// Whether this kind of memory is allowed to leak
pub trait MayLeak: Copy {
fn may_leak(self) -> bool;
}
/// The functionality needed by memory to manage its allocations
pub trait AllocMap<K: Hash + Eq, V> {
/// Tests if the map contains the given key.
/// Deliberately takes `&mut` because that is sufficient, and some implementations
/// can be more efficient then (using `RefCell::get_mut`).
fn contains_key<Q: ?Sized + Hash + Eq>(&mut self, k: &Q) -> bool
where
K: Borrow<Q>;
/// Inserts a new entry into the map.
fn insert(&mut self, k: K, v: V) -> Option<V>;
/// Removes an entry from the map.
fn remove<Q: ?Sized + Hash + Eq>(&mut self, k: &Q) -> Option<V>
where
K: Borrow<Q>;
/// Returns data based on the keys and values in the map.
fn filter_map_collect<T>(&self, f: impl FnMut(&K, &V) -> Option<T>) -> Vec<T>;
/// Returns a reference to entry `k`. If no such entry exists, call
/// `vacant` and either forward its error, or add its result to the map
/// and return a reference to *that*.
fn get_or<E>(&self, k: K, vacant: impl FnOnce() -> Result<V, E>) -> Result<&V, E>;
/// Returns a mutable reference to entry `k`. If no such entry exists, call
/// `vacant` and either forward its error, or add its result to the map
/// and return a reference to *that*.
fn get_mut_or<E>(&mut self, k: K, vacant: impl FnOnce() -> Result<V, E>) -> Result<&mut V, E>;
/// Read-only lookup.
fn get(&self, k: K) -> Option<&V> {
self.get_or(k, || Err(())).ok()
}
/// Mutable lookup.
fn get_mut(&mut self, k: K) -> Option<&mut V> {
self.get_mut_or(k, || Err(())).ok()
}
}
/// Methods of this trait signifies a point where CTFE evaluation would fail
/// and some use case dependent behaviour can instead be applied.
pub trait Machine<'mir, 'tcx>: Sized {
/// Additional memory kinds a machine wishes to distinguish from the builtin ones
type MemoryKind: Debug + std::fmt::Display + MayLeak + Eq + 'static;
/// Pointers are "tagged" with provenance information; typically the `AllocId` they belong to.
type PointerTag: Provenance + Eq + Hash + 'static;
/// Machines can define extra (non-instance) things that represent values of function pointers.
/// For example, Miri uses this to return a function pointer from `dlsym`
/// that can later be called to execute the right thing.
type ExtraFnVal: Debug + Copy;
/// Extra data stored in every call frame.
type FrameExtra;
/// Extra data stored in memory. A reference to this is available when `AllocExtra`
/// gets initialized, so you can e.g., have an `Rc` here if there is global state you
/// need access to in the `AllocExtra` hooks.
type MemoryExtra;
/// Extra data stored in every allocation.
type AllocExtra: Debug + Clone + 'static;
/// Memory's allocation map
type MemoryMap: AllocMap<
AllocId,
(MemoryKind<Self::MemoryKind>, Allocation<Self::PointerTag, Self::AllocExtra>),
> + Default
+ Clone;
/// The memory kind to use for copied global memory (held in `tcx`) --
/// or None if such memory should not be mutated and thus any such attempt will cause
/// a `ModifiedStatic` error to be raised.
/// Statics are copied under two circumstances: When they are mutated, and when
/// `tag_allocation` (see below) returns an owned allocation
/// that is added to the memory so that the work is not done twice.
const GLOBAL_KIND: Option<Self::MemoryKind>;
/// Should the machine panic on allocation failures?
const PANIC_ON_ALLOC_FAIL: bool;
/// Whether memory accesses should be alignment-checked.
fn enforce_alignment(memory_extra: &Self::MemoryExtra) -> bool;
/// Whether, when checking alignment, we should `force_int` and thus support
/// custom alignment logic based on whatever the integer address happens to be.
fn force_int_for_alignment_check(memory_extra: &Self::MemoryExtra) -> bool;
/// Whether to enforce the validity invariant
fn enforce_validity(ecx: &InterpCx<'mir, 'tcx, Self>) -> bool;
/// Whether function calls should be [ABI](Abi)-checked.
fn enforce_abi(_ecx: &InterpCx<'mir, 'tcx, Self>) -> bool {
true
}
/// Entry point for obtaining the MIR of anything that should get evaluated.
/// So not just functions and shims, but also const/static initializers, anonymous
/// constants, ...
fn load_mir(
ecx: &InterpCx<'mir, 'tcx, Self>,
instance: ty::InstanceDef<'tcx>,
) -> InterpResult<'tcx, &'tcx mir::Body<'tcx>> {
Ok(ecx.tcx.instance_mir(instance))
}
/// Entry point to all function calls.
///
/// Returns either the mir to use for the call, or `None` if execution should
/// just proceed (which usually means this hook did all the work that the
/// called function should usually have done). In the latter case, it is
/// this hook's responsibility to advance the instruction pointer!
/// (This is to support functions like `__rust_maybe_catch_panic` that neither find a MIR
/// nor just jump to `ret`, but instead push their own stack frame.)
/// Passing `dest`and `ret` in the same `Option` proved very annoying when only one of them
/// was used.
fn find_mir_or_eval_fn(
ecx: &mut InterpCx<'mir, 'tcx, Self>,
instance: ty::Instance<'tcx>,
abi: Abi,
args: &[OpTy<'tcx, Self::PointerTag>],
ret: Option<(&PlaceTy<'tcx, Self::PointerTag>, mir::BasicBlock)>,
unwind: StackPopUnwind,
) -> InterpResult<'tcx, Option<&'mir mir::Body<'tcx>>>;
/// Execute `fn_val`. It is the hook's responsibility to advance the instruction
/// pointer as appropriate.
fn call_extra_fn(
ecx: &mut InterpCx<'mir, 'tcx, Self>,
fn_val: Self::ExtraFnVal,
abi: Abi,
args: &[OpTy<'tcx, Self::PointerTag>],
ret: Option<(&PlaceTy<'tcx, Self::PointerTag>, mir::BasicBlock)>,
unwind: StackPopUnwind,
) -> InterpResult<'tcx>;
/// Directly process an intrinsic without pushing a stack frame. It is the hook's
/// responsibility to advance the instruction pointer as appropriate.
fn call_intrinsic(
ecx: &mut InterpCx<'mir, 'tcx, Self>,
instance: ty::Instance<'tcx>,
args: &[OpTy<'tcx, Self::PointerTag>],
ret: Option<(&PlaceTy<'tcx, Self::PointerTag>, mir::BasicBlock)>,
unwind: StackPopUnwind,
) -> InterpResult<'tcx>;
/// Called to evaluate `Assert` MIR terminators that trigger a panic.
fn assert_panic(
ecx: &mut InterpCx<'mir, 'tcx, Self>,
msg: &mir::AssertMessage<'tcx>,
unwind: Option<mir::BasicBlock>,
) -> InterpResult<'tcx>;
/// Called to evaluate `Abort` MIR terminator.
fn abort(_ecx: &mut InterpCx<'mir, 'tcx, Self>, _msg: String) -> InterpResult<'tcx, !> {
throw_unsup_format!("aborting execution is not supported")
}
/// Called for all binary operations where the LHS has pointer type.
///
/// Returns a (value, overflowed) pair if the operation succeeded
fn binary_ptr_op(
ecx: &InterpCx<'mir, 'tcx, Self>,
bin_op: mir::BinOp,
left: &ImmTy<'tcx, Self::PointerTag>,
right: &ImmTy<'tcx, Self::PointerTag>,
) -> InterpResult<'tcx, (Scalar<Self::PointerTag>, bool, Ty<'tcx>)>;
/// Heap allocations via the `box` keyword.
fn box_alloc(
ecx: &mut InterpCx<'mir, 'tcx, Self>,
dest: &PlaceTy<'tcx, Self::PointerTag>,
) -> InterpResult<'tcx>;
/// Called to read the specified `local` from the `frame`.
/// Since reading a ZST is not actually accessing memory or locals, this is never invoked
/// for ZST reads.
#[inline]
fn access_local(
_ecx: &InterpCx<'mir, 'tcx, Self>,
frame: &Frame<'mir, 'tcx, Self::PointerTag, Self::FrameExtra>,
local: mir::Local,
) -> InterpResult<'tcx, Operand<Self::PointerTag>> {
frame.locals[local].access()
}
/// Called to write the specified `local` from the `frame`.
/// Since writing a ZST is not actually accessing memory or locals, this is never invoked
/// for ZST reads.
#[inline]
fn access_local_mut<'a>(
ecx: &'a mut InterpCx<'mir, 'tcx, Self>,
frame: usize,
local: mir::Local,
) -> InterpResult<'tcx, Result<&'a mut LocalValue<Self::PointerTag>, MemPlace<Self::PointerTag>>>
where
'tcx: 'mir,
{
ecx.stack_mut()[frame].locals[local].access_mut()
}
/// Called before a basic block terminator is executed.
/// You can use this to detect endlessly running programs.
#[inline]
fn before_terminator(_ecx: &mut InterpCx<'mir, 'tcx, Self>) -> InterpResult<'tcx> {
Ok(())
}
/// Called before a global allocation is accessed.
/// `def_id` is `Some` if this is the "lazy" allocation of a static.
#[inline]
fn before_access_global(
_memory_extra: &Self::MemoryExtra,
_alloc_id: AllocId,
_allocation: &Allocation,
_static_def_id: Option<DefId>,
_is_write: bool,
) -> InterpResult<'tcx> {
Ok(())
}
/// Return the `AllocId` for the given thread-local static in the current thread.
fn thread_local_static_base_pointer(
_ecx: &mut InterpCx<'mir, 'tcx, Self>,
def_id: DefId,
) -> InterpResult<'tcx, Pointer<Self::PointerTag>> {
throw_unsup!(ThreadLocalStatic(def_id))
}
/// Return the root pointer for the given `extern static`.
fn extern_static_base_pointer(
mem: &Memory<'mir, 'tcx, Self>,
def_id: DefId,
) -> InterpResult<'tcx, Pointer<Self::PointerTag>>;
/// Return a "base" pointer for the given allocation: the one that is used for direct
/// accesses to this static/const/fn allocation, or the one returned from the heap allocator.
///
/// Not called on `extern` or thread-local statics (those use the methods above).
fn tag_alloc_base_pointer(
mem: &Memory<'mir, 'tcx, Self>,
ptr: Pointer,
) -> Pointer<Self::PointerTag>;
/// "Int-to-pointer cast"
fn ptr_from_addr(
mem: &Memory<'mir, 'tcx, Self>,
addr: u64,
) -> Pointer<Option<Self::PointerTag>>;
/// Convert a pointer with provenance into an allocation-offset pair.
fn ptr_get_alloc(
mem: &Memory<'mir, 'tcx, Self>,
ptr: Pointer<Self::PointerTag>,
) -> (AllocId, Size);
/// Called to initialize the "extra" state of an allocation and make the pointers
/// it contains (in relocations) tagged. The way we construct allocations is
/// to always first construct it without extra and then add the extra.
/// This keeps uniform code paths for handling both allocations created by CTFE
/// for globals, and allocations created by Miri during evaluation.
///
/// `kind` is the kind of the allocation being tagged; it can be `None` when
/// it's a global and `GLOBAL_KIND` is `None`.
///
/// This should avoid copying if no work has to be done! If this returns an owned
/// allocation (because a copy had to be done to add tags or metadata), machine memory will
/// cache the result. (This relies on `AllocMap::get_or` being able to add the
/// owned allocation to the map even when the map is shared.)
fn init_allocation_extra<'b>(
mem: &Memory<'mir, 'tcx, Self>,
id: AllocId,
alloc: Cow<'b, Allocation>,
kind: Option<MemoryKind<Self::MemoryKind>>,
) -> Cow<'b, Allocation<Self::PointerTag, Self::AllocExtra>>;
/// Hook for performing extra checks on a memory read access.
///
/// Takes read-only access to the allocation so we can keep all the memory read
/// operations take `&self`. Use a `RefCell` in `AllocExtra` if you
/// need to mutate.
#[inline(always)]
fn memory_read(
_memory_extra: &Self::MemoryExtra,
_alloc_extra: &Self::AllocExtra,
_tag: Self::PointerTag,
_range: AllocRange,
) -> InterpResult<'tcx> {
Ok(())
}
/// Hook for performing extra checks on a memory write access.
#[inline(always)]
fn memory_written(
_memory_extra: &mut Self::MemoryExtra,
_alloc_extra: &mut Self::AllocExtra,
_tag: Self::PointerTag,
_range: AllocRange,
) -> InterpResult<'tcx> {
Ok(())
}
/// Hook for performing extra operations on a memory deallocation.
#[inline(always)]
fn memory_deallocated(
_memory_extra: &mut Self::MemoryExtra,
_alloc_extra: &mut Self::AllocExtra,
_tag: Self::PointerTag,
_range: AllocRange,
) -> InterpResult<'tcx> {
Ok(())
}
/// Executes a retagging operation.
#[inline]
fn retag(
_ecx: &mut InterpCx<'mir, 'tcx, Self>,
_kind: mir::RetagKind,
_place: &PlaceTy<'tcx, Self::PointerTag>,
) -> InterpResult<'tcx> {
Ok(())
}
/// Called immediately before a new stack frame gets pushed.
fn init_frame_extra(
ecx: &mut InterpCx<'mir, 'tcx, Self>,
frame: Frame<'mir, 'tcx, Self::PointerTag>,
) -> InterpResult<'tcx, Frame<'mir, 'tcx, Self::PointerTag, Self::FrameExtra>>;
/// Borrow the current thread's stack.
fn stack(
ecx: &'a InterpCx<'mir, 'tcx, Self>,
) -> &'a [Frame<'mir, 'tcx, Self::PointerTag, Self::FrameExtra>];
/// Mutably borrow the current thread's stack.
fn stack_mut(
ecx: &'a mut InterpCx<'mir, 'tcx, Self>,
) -> &'a mut Vec<Frame<'mir, 'tcx, Self::PointerTag, Self::FrameExtra>>;
/// Called immediately after a stack frame got pushed and its locals got initialized.
fn after_stack_push(_ecx: &mut InterpCx<'mir, 'tcx, Self>) -> InterpResult<'tcx> {
Ok(())
}
/// Called immediately after a stack frame got popped, but before jumping back to the caller.
fn after_stack_pop(
_ecx: &mut InterpCx<'mir, 'tcx, Self>,
_frame: Frame<'mir, 'tcx, Self::PointerTag, Self::FrameExtra>,
_unwinding: bool,
) -> InterpResult<'tcx, StackPopJump> {
// By default, we do not support unwinding from panics
Ok(StackPopJump::Normal)
}
}
// A lot of the flexibility above is just needed for `Miri`, but all "compile-time" machines
// (CTFE and ConstProp) use the same instance. Here, we share that code.
pub macro compile_time_machine(<$mir: lifetime, $tcx: lifetime>) {
type PointerTag = AllocId;
type ExtraFnVal = !;
type MemoryMap =
rustc_data_structures::fx::FxHashMap<AllocId, (MemoryKind<Self::MemoryKind>, Allocation)>;
const GLOBAL_KIND: Option<Self::MemoryKind> = None; // no copying of globals from `tcx` to machine memory
type AllocExtra = ();
type FrameExtra = ();
#[inline(always)]
fn enforce_alignment(_memory_extra: &Self::MemoryExtra) -> bool {
// We do not check for alignment to avoid having to carry an `Align`
// in `ConstValue::ByRef`.
false
}
#[inline(always)]
fn force_int_for_alignment_check(_memory_extra: &Self::MemoryExtra) -> bool {
// We do not support `force_int`.
false
}
#[inline(always)]
fn enforce_validity(_ecx: &InterpCx<$mir, $tcx, Self>) -> bool {
false // for now, we don't enforce validity
}
#[inline(always)]
fn call_extra_fn(
_ecx: &mut InterpCx<$mir, $tcx, Self>,
fn_val: !,
_abi: Abi,
_args: &[OpTy<$tcx>],
_ret: Option<(&PlaceTy<$tcx>, mir::BasicBlock)>,
_unwind: StackPopUnwind,
) -> InterpResult<$tcx> {
match fn_val {}
}
#[inline(always)]
fn init_allocation_extra<'b>(
_mem: &Memory<$mir, $tcx, Self>,
_id: AllocId,
alloc: Cow<'b, Allocation>,
_kind: Option<MemoryKind<Self::MemoryKind>>,
) -> Cow<'b, Allocation<Self::PointerTag>> {
// We do not use a tag so we can just cheaply forward the allocation
alloc
}
fn extern_static_base_pointer(
mem: &Memory<$mir, $tcx, Self>,
def_id: DefId,
) -> InterpResult<$tcx, Pointer> {
// Use the `AllocId` associated with the `DefId`. Any actual *access* will fail.
Ok(Pointer::new(mem.tcx.create_static_alloc(def_id), Size::ZERO))
}
#[inline(always)]
fn tag_alloc_base_pointer(
_mem: &Memory<$mir, $tcx, Self>,
ptr: Pointer<AllocId>,
) -> Pointer<AllocId> {
ptr
}
#[inline(always)]
fn ptr_from_addr(_mem: &Memory<$mir, $tcx, Self>, addr: u64) -> Pointer<Option<AllocId>> {
Pointer::new(None, Size::from_bytes(addr))
}
#[inline(always)]
fn ptr_get_alloc(_mem: &Memory<$mir, $tcx, Self>, ptr: Pointer<AllocId>) -> (AllocId, Size) {
// We know `offset` is relative to the allocation, so we can use `into_parts`.
let (alloc_id, offset) = ptr.into_parts();
(alloc_id, offset)
}
}

1186
compiler/rustc_const_eval/src/interpret/memory.rs Normal file
View File

File diff suppressed because it is too large Load Diff

33
compiler/rustc_const_eval/src/interpret/mod.rs Normal file
View File

@@ -0,0 +1,33 @@
//! An interpreter for MIR used in CTFE and by miri
mod cast;
mod eval_context;
mod intern;
mod intrinsics;
mod machine;
mod memory;
mod operand;
mod operator;
mod place;
mod step;
mod terminator;
mod traits;
mod util;
mod validity;
mod visitor;
pub use rustc_middle::mir::interpret::*; // have all the `interpret` symbols in one place: here
pub use self::eval_context::{
Frame, FrameInfo, InterpCx, LocalState, LocalValue, StackPopCleanup, StackPopUnwind,
};
pub use self::intern::{intern_const_alloc_recursive, InternKind};
pub use self::machine::{compile_time_machine, AllocMap, Machine, MayLeak, StackPopJump};
pub use self::memory::{AllocCheck, AllocRef, AllocRefMut, FnVal, Memory, MemoryKind};
pub use self::operand::{ImmTy, Immediate, OpTy, Operand};
pub use self::place::{MPlaceTy, MemPlace, MemPlaceMeta, Place, PlaceTy};
pub use self::validity::{CtfeValidationMode, RefTracking};
pub use self::visitor::{MutValueVisitor, ValueVisitor};
crate use self::intrinsics::eval_nullary_intrinsic;
use eval_context::{from_known_layout, mir_assign_valid_types};

762
compiler/rustc_const_eval/src/interpret/operand.rs Normal file
View File

@@ -0,0 +1,762 @@
//! Functions concerning immediate values and operands, and reading from operands.
//! All high-level functions to read from memory work on operands as sources.
use std::convert::TryFrom;
use std::fmt::Write;
use rustc_errors::ErrorReported;
use rustc_hir::def::Namespace;
use rustc_macros::HashStable;
use rustc_middle::ty::layout::{LayoutOf, PrimitiveExt, TyAndLayout};
use rustc_middle::ty::print::{FmtPrinter, PrettyPrinter, Printer};
use rustc_middle::ty::{ConstInt, Ty};
use rustc_middle::{mir, ty};
use rustc_target::abi::{Abi, HasDataLayout, Size, TagEncoding};
use rustc_target::abi::{VariantIdx, Variants};
use super::{
alloc_range, from_known_layout, mir_assign_valid_types, AllocId, ConstValue, GlobalId,
InterpCx, InterpResult, MPlaceTy, Machine, MemPlace, Place, PlaceTy, Pointer, Provenance,
Scalar, ScalarMaybeUninit,
};
/// An `Immediate` represents a single immediate self-contained Rust value.
///
/// For optimization of a few very common cases, there is also a representation for a pair of
/// primitive values (`ScalarPair`). It allows Miri to avoid making allocations for checked binary
/// operations and wide pointers. This idea was taken from rustc's codegen.
/// In particular, thanks to `ScalarPair`, arithmetic operations and casts can be entirely
/// defined on `Immediate`, and do not have to work with a `Place`.
#[derive(Copy, Clone, PartialEq, Eq, HashStable, Hash, Debug)]
pub enum Immediate<Tag: Provenance = AllocId> {
Scalar(ScalarMaybeUninit<Tag>),
ScalarPair(ScalarMaybeUninit<Tag>, ScalarMaybeUninit<Tag>),
}
#[cfg(all(target_arch = "x86_64", target_pointer_width = "64"))]
rustc_data_structures::static_assert_size!(Immediate, 56);
impl<Tag: Provenance> From<ScalarMaybeUninit<Tag>> for Immediate<Tag> {
#[inline(always)]
fn from(val: ScalarMaybeUninit<Tag>) -> Self {
Immediate::Scalar(val)
}
}
impl<Tag: Provenance> From<Scalar<Tag>> for Immediate<Tag> {
#[inline(always)]
fn from(val: Scalar<Tag>) -> Self {
Immediate::Scalar(val.into())
}
}
impl<'tcx, Tag: Provenance> Immediate<Tag> {
pub fn from_pointer(p: Pointer<Tag>, cx: &impl HasDataLayout) -> Self {
Immediate::Scalar(ScalarMaybeUninit::from_pointer(p, cx))
}
pub fn from_maybe_pointer(p: Pointer<Option<Tag>>, cx: &impl HasDataLayout) -> Self {
Immediate::Scalar(ScalarMaybeUninit::from_maybe_pointer(p, cx))
}
pub fn new_slice(val: Scalar<Tag>, len: u64, cx: &impl HasDataLayout) -> Self {
Immediate::ScalarPair(val.into(), Scalar::from_machine_usize(len, cx).into())
}
pub fn new_dyn_trait(
val: Scalar<Tag>,
vtable: Pointer<Option<Tag>>,
cx: &impl HasDataLayout,
) -> Self {
Immediate::ScalarPair(val.into(), ScalarMaybeUninit::from_maybe_pointer(vtable, cx))
}
#[inline]
pub fn to_scalar_or_uninit(self) -> ScalarMaybeUninit<Tag> {
match self {
Immediate::Scalar(val) => val,
Immediate::ScalarPair(..) => bug!("Got a scalar pair where a scalar was expected"),
}
}
#[inline]
pub fn to_scalar(self) -> InterpResult<'tcx, Scalar<Tag>> {
self.to_scalar_or_uninit().check_init()
}
#[inline]
pub fn to_scalar_pair(self) -> InterpResult<'tcx, (Scalar<Tag>, Scalar<Tag>)> {
match self {
Immediate::ScalarPair(val1, val2) => Ok((val1.check_init()?, val2.check_init()?)),
Immediate::Scalar(..) => {
bug!("Got a scalar where a scalar pair was expected")
}
}
}
}
// ScalarPair needs a type to interpret, so we often have an immediate and a type together
// as input for binary and cast operations.
#[derive(Copy, Clone, Debug)]
pub struct ImmTy<'tcx, Tag: Provenance = AllocId> {
imm: Immediate<Tag>,
pub layout: TyAndLayout<'tcx>,
}
#[cfg(all(target_arch = "x86_64", target_pointer_width = "64"))]
rustc_data_structures::static_assert_size!(ImmTy<'_>, 72);
impl<Tag: Provenance> std::fmt::Display for ImmTy<'tcx, Tag> {
fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
/// Helper function for printing a scalar to a FmtPrinter
fn p<'a, 'tcx, F: std::fmt::Write, Tag: Provenance>(
cx: FmtPrinter<'a, 'tcx, F>,
s: ScalarMaybeUninit<Tag>,
ty: Ty<'tcx>,
) -> Result<FmtPrinter<'a, 'tcx, F>, std::fmt::Error> {
match s {
ScalarMaybeUninit::Scalar(Scalar::Int(int)) => {
cx.pretty_print_const_scalar_int(int, ty, true)
}
ScalarMaybeUninit::Scalar(Scalar::Ptr(ptr, _sz)) => {
// Just print the ptr value. `pretty_print_const_scalar_ptr` would also try to
// print what is points to, which would fail since it has no access to the local
// memory.
cx.pretty_print_const_pointer(ptr, ty, true)
}
ScalarMaybeUninit::Uninit => cx.typed_value(
|mut this| {
this.write_str("uninit ")?;
Ok(this)
},
|this| this.print_type(ty),
" ",
),
}
}
ty::tls::with(|tcx| {
match self.imm {
Immediate::Scalar(s) => {
if let Some(ty) = tcx.lift(self.layout.ty) {
let cx = FmtPrinter::new(tcx, f, Namespace::ValueNS);
p(cx, s, ty)?;
return Ok(());
}
write!(f, "{}: {}", s, self.layout.ty)
}
Immediate::ScalarPair(a, b) => {
// FIXME(oli-obk): at least print tuples and slices nicely
write!(f, "({}, {}): {}", a, b, self.layout.ty,)
}
}
})
}
}
impl<'tcx, Tag: Provenance> std::ops::Deref for ImmTy<'tcx, Tag> {
type Target = Immediate<Tag>;
#[inline(always)]
fn deref(&self) -> &Immediate<Tag> {
&self.imm
}
}
/// An `Operand` is the result of computing a `mir::Operand`. It can be immediate,
/// or still in memory. The latter is an optimization, to delay reading that chunk of
/// memory and to avoid having to store arbitrary-sized data here.
#[derive(Copy, Clone, PartialEq, Eq, HashStable, Hash, Debug)]
pub enum Operand<Tag: Provenance = AllocId> {
Immediate(Immediate<Tag>),
Indirect(MemPlace<Tag>),
}
#[derive(Copy, Clone, PartialEq, Eq, Hash, Debug)]
pub struct OpTy<'tcx, Tag: Provenance = AllocId> {
op: Operand<Tag>, // Keep this private; it helps enforce invariants.
pub layout: TyAndLayout<'tcx>,
}
#[cfg(all(target_arch = "x86_64", target_pointer_width = "64"))]
rustc_data_structures::static_assert_size!(OpTy<'_>, 80);
impl<'tcx, Tag: Provenance> std::ops::Deref for OpTy<'tcx, Tag> {
type Target = Operand<Tag>;
#[inline(always)]
fn deref(&self) -> &Operand<Tag> {
&self.op
}
}
impl<'tcx, Tag: Provenance> From<MPlaceTy<'tcx, Tag>> for OpTy<'tcx, Tag> {
#[inline(always)]
fn from(mplace: MPlaceTy<'tcx, Tag>) -> Self {
OpTy { op: Operand::Indirect(*mplace), layout: mplace.layout }
}
}
impl<'tcx, Tag: Provenance> From<&'_ MPlaceTy<'tcx, Tag>> for OpTy<'tcx, Tag> {
#[inline(always)]
fn from(mplace: &MPlaceTy<'tcx, Tag>) -> Self {
OpTy { op: Operand::Indirect(**mplace), layout: mplace.layout }
}
}
impl<'tcx, Tag: Provenance> From<ImmTy<'tcx, Tag>> for OpTy<'tcx, Tag> {
#[inline(always)]
fn from(val: ImmTy<'tcx, Tag>) -> Self {
OpTy { op: Operand::Immediate(val.imm), layout: val.layout }
}
}
impl<'tcx, Tag: Provenance> ImmTy<'tcx, Tag> {
#[inline]
pub fn from_scalar(val: Scalar<Tag>, layout: TyAndLayout<'tcx>) -> Self {
ImmTy { imm: val.into(), layout }
}
#[inline]
pub fn from_immediate(imm: Immediate<Tag>, layout: TyAndLayout<'tcx>) -> Self {
ImmTy { imm, layout }
}
#[inline]
pub fn try_from_uint(i: impl Into<u128>, layout: TyAndLayout<'tcx>) -> Option<Self> {
Some(Self::from_scalar(Scalar::try_from_uint(i, layout.size)?, layout))
}
#[inline]
pub fn from_uint(i: impl Into<u128>, layout: TyAndLayout<'tcx>) -> Self {
Self::from_scalar(Scalar::from_uint(i, layout.size), layout)
}
#[inline]
pub fn try_from_int(i: impl Into<i128>, layout: TyAndLayout<'tcx>) -> Option<Self> {
Some(Self::from_scalar(Scalar::try_from_int(i, layout.size)?, layout))
}
#[inline]
pub fn from_int(i: impl Into<i128>, layout: TyAndLayout<'tcx>) -> Self {
Self::from_scalar(Scalar::from_int(i, layout.size), layout)
}
#[inline]
pub fn to_const_int(self) -> ConstInt {
assert!(self.layout.ty.is_integral());
let int = self.to_scalar().expect("to_const_int doesn't work on scalar pairs").assert_int();
ConstInt::new(int, self.layout.ty.is_signed(), self.layout.ty.is_ptr_sized_integral())
}
}
impl<'mir, 'tcx: 'mir, M: Machine<'mir, 'tcx>> InterpCx<'mir, 'tcx, M> {
/// Try reading an immediate in memory; this is interesting particularly for `ScalarPair`.
/// Returns `None` if the layout does not permit loading this as a value.
fn try_read_immediate_from_mplace(
&self,
mplace: &MPlaceTy<'tcx, M::PointerTag>,
) -> InterpResult<'tcx, Option<ImmTy<'tcx, M::PointerTag>>> {
if mplace.layout.is_unsized() {
// Don't touch unsized
return Ok(None);
}
let alloc = match self.get_alloc(mplace)? {
Some(ptr) => ptr,
None => {
return Ok(Some(ImmTy {
// zero-sized type
imm: Scalar::ZST.into(),
layout: mplace.layout,
}));
}
};
match mplace.layout.abi {
Abi::Scalar(..) => {
let scalar = alloc.read_scalar(alloc_range(Size::ZERO, mplace.layout.size))?;
Ok(Some(ImmTy { imm: scalar.into(), layout: mplace.layout }))
}
Abi::ScalarPair(ref a, ref b) => {
// We checked `ptr_align` above, so all fields will have the alignment they need.
// We would anyway check against `ptr_align.restrict_for_offset(b_offset)`,
// which `ptr.offset(b_offset)` cannot possibly fail to satisfy.
let (a, b) = (&a.value, &b.value);
let (a_size, b_size) = (a.size(self), b.size(self));
let b_offset = a_size.align_to(b.align(self).abi);
assert!(b_offset.bytes() > 0); // we later use the offset to tell apart the fields
let a_val = alloc.read_scalar(alloc_range(Size::ZERO, a_size))?;
let b_val = alloc.read_scalar(alloc_range(b_offset, b_size))?;
Ok(Some(ImmTy { imm: Immediate::ScalarPair(a_val, b_val), layout: mplace.layout }))
}
_ => Ok(None),
}
}
/// Try returning an immediate for the operand.
/// If the layout does not permit loading this as an immediate, return where in memory
/// we can find the data.
/// Note that for a given layout, this operation will either always fail or always
/// succeed! Whether it succeeds depends on whether the layout can be represented
/// in an `Immediate`, not on which data is stored there currently.
pub fn try_read_immediate(
&self,
src: &OpTy<'tcx, M::PointerTag>,
) -> InterpResult<'tcx, Result<ImmTy<'tcx, M::PointerTag>, MPlaceTy<'tcx, M::PointerTag>>> {
Ok(match src.try_as_mplace() {
Ok(ref mplace) => {
if let Some(val) = self.try_read_immediate_from_mplace(mplace)? {
Ok(val)
} else {
Err(*mplace)
}
}
Err(val) => Ok(val),
})
}
/// Read an immediate from a place, asserting that that is possible with the given layout.
#[inline(always)]
pub fn read_immediate(
&self,
op: &OpTy<'tcx, M::PointerTag>,
) -> InterpResult<'tcx, ImmTy<'tcx, M::PointerTag>> {
if let Ok(imm) = self.try_read_immediate(op)? {
Ok(imm)
} else {
span_bug!(self.cur_span(), "primitive read failed for type: {:?}", op.layout.ty);
}
}
/// Read a scalar from a place
pub fn read_scalar(
&self,
op: &OpTy<'tcx, M::PointerTag>,
) -> InterpResult<'tcx, ScalarMaybeUninit<M::PointerTag>> {
Ok(self.read_immediate(op)?.to_scalar_or_uninit())
}
/// Read a pointer from a place.
pub fn read_pointer(
&self,
op: &OpTy<'tcx, M::PointerTag>,
) -> InterpResult<'tcx, Pointer<Option<M::PointerTag>>> {
Ok(self.scalar_to_ptr(self.read_scalar(op)?.check_init()?))
}
// Turn the wide MPlace into a string (must already be dereferenced!)
pub fn read_str(&self, mplace: &MPlaceTy<'tcx, M::PointerTag>) -> InterpResult<'tcx, &str> {
let len = mplace.len(self)?;
let bytes = self.memory.read_bytes(mplace.ptr, Size::from_bytes(len))?;
let str = std::str::from_utf8(bytes).map_err(|err| err_ub!(InvalidStr(err)))?;
Ok(str)
}
/// Projection functions
pub fn operand_field(
&self,
op: &OpTy<'tcx, M::PointerTag>,
field: usize,
) -> InterpResult<'tcx, OpTy<'tcx, M::PointerTag>> {
let base = match op.try_as_mplace() {
Ok(ref mplace) => {
// We can reuse the mplace field computation logic for indirect operands.
let field = self.mplace_field(mplace, field)?;
return Ok(field.into());
}
Err(value) => value,
};
let field_layout = op.layout.field(self, field);
if field_layout.is_zst() {
let immediate = Scalar::ZST.into();
return Ok(OpTy { op: Operand::Immediate(immediate), layout: field_layout });
}
let offset = op.layout.fields.offset(field);
let immediate = match *base {
// the field covers the entire type
_ if offset.bytes() == 0 && field_layout.size == op.layout.size => *base,
// extract fields from types with `ScalarPair` ABI
Immediate::ScalarPair(a, b) => {
let val = if offset.bytes() == 0 { a } else { b };
Immediate::from(val)
}
Immediate::Scalar(val) => span_bug!(
self.cur_span(),
"field access on non aggregate {:#?}, {:#?}",
val,
op.layout
),
};
Ok(OpTy { op: Operand::Immediate(immediate), layout: field_layout })
}
pub fn operand_index(
&self,
op: &OpTy<'tcx, M::PointerTag>,
index: u64,
) -> InterpResult<'tcx, OpTy<'tcx, M::PointerTag>> {
if let Ok(index) = usize::try_from(index) {
// We can just treat this as a field.
self.operand_field(op, index)
} else {
// Indexing into a big array. This must be an mplace.
let mplace = op.assert_mem_place();
Ok(self.mplace_index(&mplace, index)?.into())
}
}
pub fn operand_downcast(
&self,
op: &OpTy<'tcx, M::PointerTag>,
variant: VariantIdx,
) -> InterpResult<'tcx, OpTy<'tcx, M::PointerTag>> {
// Downcasts only change the layout
Ok(match op.try_as_mplace() {
Ok(ref mplace) => self.mplace_downcast(mplace, variant)?.into(),
Err(..) => {
let layout = op.layout.for_variant(self, variant);
OpTy { layout, ..*op }
}
})
}
pub fn operand_projection(
&self,
base: &OpTy<'tcx, M::PointerTag>,
proj_elem: mir::PlaceElem<'tcx>,
) -> InterpResult<'tcx, OpTy<'tcx, M::PointerTag>> {
use rustc_middle::mir::ProjectionElem::*;
Ok(match proj_elem {
Field(field, _) => self.operand_field(base, field.index())?,
Downcast(_, variant) => self.operand_downcast(base, variant)?,
Deref => self.deref_operand(base)?.into(),
Subslice { .. } | ConstantIndex { .. } | Index(_) => {
// The rest should only occur as mplace, we do not use Immediates for types
// allowing such operations. This matches place_projection forcing an allocation.
let mplace = base.assert_mem_place();
self.mplace_projection(&mplace, proj_elem)?.into()
}
})
}
/// Read from a local. Will not actually access the local if reading from a ZST.
/// Will not access memory, instead an indirect `Operand` is returned.
///
/// This is public because it is used by [priroda](https://github.com/oli-obk/priroda) to get an
/// OpTy from a local
pub fn access_local(
&self,
frame: &super::Frame<'mir, 'tcx, M::PointerTag, M::FrameExtra>,
local: mir::Local,
layout: Option<TyAndLayout<'tcx>>,
) -> InterpResult<'tcx, OpTy<'tcx, M::PointerTag>> {
let layout = self.layout_of_local(frame, local, layout)?;
let op = if layout.is_zst() {
// Do not read from ZST, they might not be initialized
Operand::Immediate(Scalar::ZST.into())
} else {
M::access_local(&self, frame, local)?
};
Ok(OpTy { op, layout })
}
/// Every place can be read from, so we can turn them into an operand.
/// This will definitely return `Indirect` if the place is a `Ptr`, i.e., this
/// will never actually read from memory.
#[inline(always)]
pub fn place_to_op(
&self,
place: &PlaceTy<'tcx, M::PointerTag>,
) -> InterpResult<'tcx, OpTy<'tcx, M::PointerTag>> {
let op = match **place {
Place::Ptr(mplace) => Operand::Indirect(mplace),
Place::Local { frame, local } => {
*self.access_local(&self.stack()[frame], local, None)?
}
};
Ok(OpTy { op, layout: place.layout })
}
// Evaluate a place with the goal of reading from it. This lets us sometimes
// avoid allocations.
pub fn eval_place_to_op(
&self,
place: mir::Place<'tcx>,
layout: Option<TyAndLayout<'tcx>>,
) -> InterpResult<'tcx, OpTy<'tcx, M::PointerTag>> {
// Do not use the layout passed in as argument if the base we are looking at
// here is not the entire place.
let layout = if place.projection.is_empty() { layout } else { None };
let base_op = self.access_local(self.frame(), place.local, layout)?;
let op = place
.projection
.iter()
.try_fold(base_op, |op, elem| self.operand_projection(&op, elem))?;
trace!("eval_place_to_op: got {:?}", *op);
// Sanity-check the type we ended up with.
debug_assert!(mir_assign_valid_types(
*self.tcx,
self.param_env,
self.layout_of(self.subst_from_current_frame_and_normalize_erasing_regions(
place.ty(&self.frame().body.local_decls, *self.tcx).ty
))?,
op.layout,
));
Ok(op)
}
/// Evaluate the operand, returning a place where you can then find the data.
/// If you already know the layout, you can save two table lookups
/// by passing it in here.
#[inline]
pub fn eval_operand(
&self,
mir_op: &mir::Operand<'tcx>,
layout: Option<TyAndLayout<'tcx>>,
) -> InterpResult<'tcx, OpTy<'tcx, M::PointerTag>> {
use rustc_middle::mir::Operand::*;
let op = match *mir_op {
// FIXME: do some more logic on `move` to invalidate the old location
Copy(place) | Move(place) => self.eval_place_to_op(place, layout)?,
Constant(ref constant) => {
let val =
self.subst_from_current_frame_and_normalize_erasing_regions(constant.literal);
// This can still fail:
// * During ConstProp, with `TooGeneric` or since the `requried_consts` were not all
// checked yet.
// * During CTFE, since promoteds in `const`/`static` initializer bodies can fail.
self.mir_const_to_op(&val, layout)?
}
};
trace!("{:?}: {:?}", mir_op, *op);
Ok(op)
}
/// Evaluate a bunch of operands at once
pub(super) fn eval_operands(
&self,
ops: &[mir::Operand<'tcx>],
) -> InterpResult<'tcx, Vec<OpTy<'tcx, M::PointerTag>>> {
ops.iter().map(|op| self.eval_operand(op, None)).collect()
}
// Used when the miri-engine runs into a constant and for extracting information from constants
// in patterns via the `const_eval` module
/// The `val` and `layout` are assumed to already be in our interpreter
/// "universe" (param_env).
pub fn const_to_op(
&self,
val: &ty::Const<'tcx>,
layout: Option<TyAndLayout<'tcx>>,
) -> InterpResult<'tcx, OpTy<'tcx, M::PointerTag>> {
match val.val {
ty::ConstKind::Param(_) | ty::ConstKind::Bound(..) => throw_inval!(TooGeneric),
ty::ConstKind::Error(_) => throw_inval!(AlreadyReported(ErrorReported)),
ty::ConstKind::Unevaluated(uv) => {
let instance = self.resolve(uv.def, uv.substs(*self.tcx))?;
Ok(self.eval_to_allocation(GlobalId { instance, promoted: uv.promoted })?.into())
}
ty::ConstKind::Infer(..) | ty::ConstKind::Placeholder(..) => {
span_bug!(self.cur_span(), "const_to_op: Unexpected ConstKind {:?}", val)
}
ty::ConstKind::Value(val_val) => self.const_val_to_op(val_val, val.ty, layout),
}
}
pub fn mir_const_to_op(
&self,
val: &mir::ConstantKind<'tcx>,
layout: Option<TyAndLayout<'tcx>>,
) -> InterpResult<'tcx, OpTy<'tcx, M::PointerTag>> {
match val {
mir::ConstantKind::Ty(ct) => self.const_to_op(ct, layout),
mir::ConstantKind::Val(val, ty) => self.const_val_to_op(*val, ty, layout),
}
}
crate fn const_val_to_op(
&self,
val_val: ConstValue<'tcx>,
ty: Ty<'tcx>,
layout: Option<TyAndLayout<'tcx>>,
) -> InterpResult<'tcx, OpTy<'tcx, M::PointerTag>> {
// Other cases need layout.
let tag_scalar = |scalar| -> InterpResult<'tcx, _> {
Ok(match scalar {
Scalar::Ptr(ptr, size) => Scalar::Ptr(self.global_base_pointer(ptr)?, size),
Scalar::Int(int) => Scalar::Int(int),
})
};
let layout = from_known_layout(self.tcx, self.param_env, layout, || self.layout_of(ty))?;
let op = match val_val {
ConstValue::ByRef { alloc, offset } => {
let id = self.tcx.create_memory_alloc(alloc);
// We rely on mutability being set correctly in that allocation to prevent writes
// where none should happen.
let ptr = self.global_base_pointer(Pointer::new(id, offset))?;
Operand::Indirect(MemPlace::from_ptr(ptr.into(), layout.align.abi))
}
ConstValue::Scalar(x) => Operand::Immediate(tag_scalar(x)?.into()),
ConstValue::Slice { data, start, end } => {
// We rely on mutability being set correctly in `data` to prevent writes
// where none should happen.
let ptr = Pointer::new(
self.tcx.create_memory_alloc(data),
Size::from_bytes(start), // offset: `start`
);
Operand::Immediate(Immediate::new_slice(
Scalar::from_pointer(self.global_base_pointer(ptr)?, &*self.tcx),
u64::try_from(end.checked_sub(start).unwrap()).unwrap(), // len: `end - start`
self,
))
}
};
Ok(OpTy { op, layout })
}
/// Read discriminant, return the runtime value as well as the variant index.
pub fn read_discriminant(
&self,
op: &OpTy<'tcx, M::PointerTag>,
) -> InterpResult<'tcx, (Scalar<M::PointerTag>, VariantIdx)> {
trace!("read_discriminant_value {:#?}", op.layout);
// Get type and layout of the discriminant.
let discr_layout = self.layout_of(op.layout.ty.discriminant_ty(*self.tcx))?;
trace!("discriminant type: {:?}", discr_layout.ty);
// We use "discriminant" to refer to the value associated with a particular enum variant.
// This is not to be confused with its "variant index", which is just determining its position in the
// declared list of variants -- they can differ with explicitly assigned discriminants.
// We use "tag" to refer to how the discriminant is encoded in memory, which can be either
// straight-forward (`TagEncoding::Direct`) or with a niche (`TagEncoding::Niche`).
let (tag_scalar_layout, tag_encoding, tag_field) = match op.layout.variants {
Variants::Single { index } => {
let discr = match op.layout.ty.discriminant_for_variant(*self.tcx, index) {
Some(discr) => {
// This type actually has discriminants.
assert_eq!(discr.ty, discr_layout.ty);
Scalar::from_uint(discr.val, discr_layout.size)
}
None => {
// On a type without actual discriminants, variant is 0.
assert_eq!(index.as_u32(), 0);
Scalar::from_uint(index.as_u32(), discr_layout.size)
}
};
return Ok((discr, index));
}
Variants::Multiple { ref tag, ref tag_encoding, tag_field, .. } => {
(tag, tag_encoding, tag_field)
}
};
// There are *three* layouts that come into play here:
// - The discriminant has a type for typechecking. This is `discr_layout`, and is used for
// the `Scalar` we return.
// - The tag (encoded discriminant) has layout `tag_layout`. This is always an integer type,
// and used to interpret the value we read from the tag field.
// For the return value, a cast to `discr_layout` is performed.
// - The field storing the tag has a layout, which is very similar to `tag_layout` but
// may be a pointer. This is `tag_val.layout`; we just use it for sanity checks.
// Get layout for tag.
let tag_layout = self.layout_of(tag_scalar_layout.value.to_int_ty(*self.tcx))?;
// Read tag and sanity-check `tag_layout`.
let tag_val = self.read_immediate(&self.operand_field(op, tag_field)?)?;
assert_eq!(tag_layout.size, tag_val.layout.size);
assert_eq!(tag_layout.abi.is_signed(), tag_val.layout.abi.is_signed());
let tag_val = tag_val.to_scalar()?;
trace!("tag value: {:?}", tag_val);
// Figure out which discriminant and variant this corresponds to.
Ok(match *tag_encoding {
TagEncoding::Direct => {
let tag_bits = tag_val
.try_to_int()
.map_err(|dbg_val| err_ub!(InvalidTag(dbg_val)))?
.assert_bits(tag_layout.size);
// Cast bits from tag layout to discriminant layout.
let discr_val = self.cast_from_scalar(tag_bits, tag_layout, discr_layout.ty);
let discr_bits = discr_val.assert_bits(discr_layout.size);
// Convert discriminant to variant index, and catch invalid discriminants.
let index = match *op.layout.ty.kind() {
ty::Adt(adt, _) => {
adt.discriminants(*self.tcx).find(|(_, var)| var.val == discr_bits)
}
ty::Generator(def_id, substs, _) => {
let substs = substs.as_generator();
substs
.discriminants(def_id, *self.tcx)
.find(|(_, var)| var.val == discr_bits)
}
_ => span_bug!(self.cur_span(), "tagged layout for non-adt non-generator"),
}
.ok_or_else(|| err_ub!(InvalidTag(Scalar::from_uint(tag_bits, tag_layout.size))))?;
// Return the cast value, and the index.
(discr_val, index.0)
}
TagEncoding::Niche { dataful_variant, ref niche_variants, niche_start } => {
// Compute the variant this niche value/"tag" corresponds to. With niche layout,
// discriminant (encoded in niche/tag) and variant index are the same.
let variants_start = niche_variants.start().as_u32();
let variants_end = niche_variants.end().as_u32();
let variant = match tag_val.try_to_int() {
Err(dbg_val) => {
// So this is a pointer then, and casting to an int failed.
// Can only happen during CTFE.
let ptr = self.scalar_to_ptr(tag_val);
// The niche must be just 0, and the ptr not null, then we know this is
// okay. Everything else, we conservatively reject.
let ptr_valid = niche_start == 0
&& variants_start == variants_end
&& !self.memory.ptr_may_be_null(ptr);
if !ptr_valid {
throw_ub!(InvalidTag(dbg_val))
}
dataful_variant
}
Ok(tag_bits) => {
let tag_bits = tag_bits.assert_bits(tag_layout.size);
// We need to use machine arithmetic to get the relative variant idx:
// variant_index_relative = tag_val - niche_start_val
let tag_val = ImmTy::from_uint(tag_bits, tag_layout);
let niche_start_val = ImmTy::from_uint(niche_start, tag_layout);
let variant_index_relative_val =
self.binary_op(mir::BinOp::Sub, &tag_val, &niche_start_val)?;
let variant_index_relative = variant_index_relative_val
.to_scalar()?
.assert_bits(tag_val.layout.size);
// Check if this is in the range that indicates an actual discriminant.
if variant_index_relative <= u128::from(variants_end - variants_start) {
let variant_index_relative = u32::try_from(variant_index_relative)
.expect("we checked that this fits into a u32");
// Then computing the absolute variant idx should not overflow any more.
let variant_index = variants_start
.checked_add(variant_index_relative)
.expect("overflow computing absolute variant idx");
let variants_len = op
.layout
.ty
.ty_adt_def()
.expect("tagged layout for non adt")
.variants
.len();
assert!(usize::try_from(variant_index).unwrap() < variants_len);
VariantIdx::from_u32(variant_index)
} else {
dataful_variant
}
}
};
// Compute the size of the scalar we need to return.
// No need to cast, because the variant index directly serves as discriminant and is
// encoded in the tag.
(Scalar::from_uint(variant.as_u32(), discr_layout.size), variant)
}
})
}
}

417
compiler/rustc_const_eval/src/interpret/operator.rs Normal file
View File

@@ -0,0 +1,417 @@
use std::convert::TryFrom;
use rustc_apfloat::Float;
use rustc_middle::mir;
use rustc_middle::mir::interpret::{InterpResult, Scalar};
use rustc_middle::ty::layout::{LayoutOf, TyAndLayout};
use rustc_middle::ty::{self, FloatTy, Ty};
use super::{ImmTy, Immediate, InterpCx, Machine, PlaceTy};
impl<'mir, 'tcx: 'mir, M: Machine<'mir, 'tcx>> InterpCx<'mir, 'tcx, M> {
/// Applies the binary operation `op` to the two operands and writes a tuple of the result
/// and a boolean signifying the potential overflow to the destination.
pub fn binop_with_overflow(
&mut self,
op: mir::BinOp,
left: &ImmTy<'tcx, M::PointerTag>,
right: &ImmTy<'tcx, M::PointerTag>,
dest: &PlaceTy<'tcx, M::PointerTag>,
) -> InterpResult<'tcx> {
let (val, overflowed, ty) = self.overflowing_binary_op(op, &left, &right)?;
debug_assert_eq!(
self.tcx.intern_tup(&[ty, self.tcx.types.bool]),
dest.layout.ty,
"type mismatch for result of {:?}",
op,
);
let val = Immediate::ScalarPair(val.into(), Scalar::from_bool(overflowed).into());
self.write_immediate(val, dest)
}
/// Applies the binary operation `op` to the arguments and writes the result to the
/// destination.
pub fn binop_ignore_overflow(
&mut self,
op: mir::BinOp,
left: &ImmTy<'tcx, M::PointerTag>,
right: &ImmTy<'tcx, M::PointerTag>,
dest: &PlaceTy<'tcx, M::PointerTag>,
) -> InterpResult<'tcx> {
let (val, _overflowed, ty) = self.overflowing_binary_op(op, left, right)?;
assert_eq!(ty, dest.layout.ty, "type mismatch for result of {:?}", op);
self.write_scalar(val, dest)
}
}
impl<'mir, 'tcx: 'mir, M: Machine<'mir, 'tcx>> InterpCx<'mir, 'tcx, M> {
fn binary_char_op(
&self,
bin_op: mir::BinOp,
l: char,
r: char,
) -> (Scalar<M::PointerTag>, bool, Ty<'tcx>) {
use rustc_middle::mir::BinOp::*;
let res = match bin_op {
Eq => l == r,
Ne => l != r,
Lt => l < r,
Le => l <= r,
Gt => l > r,
Ge => l >= r,
_ => span_bug!(self.cur_span(), "Invalid operation on char: {:?}", bin_op),
};
(Scalar::from_bool(res), false, self.tcx.types.bool)
}
fn binary_bool_op(
&self,
bin_op: mir::BinOp,
l: bool,
r: bool,
) -> (Scalar<M::PointerTag>, bool, Ty<'tcx>) {
use rustc_middle::mir::BinOp::*;
let res = match bin_op {
Eq => l == r,
Ne => l != r,
Lt => l < r,
Le => l <= r,
Gt => l > r,
Ge => l >= r,
BitAnd => l & r,
BitOr => l | r,
BitXor => l ^ r,
_ => span_bug!(self.cur_span(), "Invalid operation on bool: {:?}", bin_op),
};
(Scalar::from_bool(res), false, self.tcx.types.bool)
}
fn binary_float_op<F: Float + Into<Scalar<M::PointerTag>>>(
&self,
bin_op: mir::BinOp,
ty: Ty<'tcx>,
l: F,
r: F,
) -> (Scalar<M::PointerTag>, bool, Ty<'tcx>) {
use rustc_middle::mir::BinOp::*;
let (val, ty) = match bin_op {
Eq => (Scalar::from_bool(l == r), self.tcx.types.bool),
Ne => (Scalar::from_bool(l != r), self.tcx.types.bool),
Lt => (Scalar::from_bool(l < r), self.tcx.types.bool),
Le => (Scalar::from_bool(l <= r), self.tcx.types.bool),
Gt => (Scalar::from_bool(l > r), self.tcx.types.bool),
Ge => (Scalar::from_bool(l >= r), self.tcx.types.bool),
Add => ((l + r).value.into(), ty),
Sub => ((l - r).value.into(), ty),
Mul => ((l * r).value.into(), ty),
Div => ((l / r).value.into(), ty),
Rem => ((l % r).value.into(), ty),
_ => span_bug!(self.cur_span(), "invalid float op: `{:?}`", bin_op),
};
(val, false, ty)
}
fn binary_int_op(
&self,
bin_op: mir::BinOp,
// passing in raw bits
l: u128,
left_layout: TyAndLayout<'tcx>,
r: u128,
right_layout: TyAndLayout<'tcx>,
) -> InterpResult<'tcx, (Scalar<M::PointerTag>, bool, Ty<'tcx>)> {
use rustc_middle::mir::BinOp::*;
// Shift ops can have an RHS with a different numeric type.
if bin_op == Shl || bin_op == Shr {
let signed = left_layout.abi.is_signed();
let size = u128::from(left_layout.size.bits());
let overflow = r >= size;
let r = r % size; // mask to type size
let r = u32::try_from(r).unwrap(); // we masked so this will always fit
let result = if signed {
let l = self.sign_extend(l, left_layout) as i128;
let result = match bin_op {
Shl => l.checked_shl(r).unwrap(),
Shr => l.checked_shr(r).unwrap(),
_ => bug!("it has already been checked that this is a shift op"),
};
result as u128
} else {
match bin_op {
Shl => l.checked_shl(r).unwrap(),
Shr => l.checked_shr(r).unwrap(),
_ => bug!("it has already been checked that this is a shift op"),
}
};
let truncated = self.truncate(result, left_layout);
return Ok((Scalar::from_uint(truncated, left_layout.size), overflow, left_layout.ty));
}
// For the remaining ops, the types must be the same on both sides
if left_layout.ty != right_layout.ty {
span_bug!(
self.cur_span(),
"invalid asymmetric binary op {:?}: {:?} ({:?}), {:?} ({:?})",
bin_op,
l,
left_layout.ty,
r,
right_layout.ty,
)
}
let size = left_layout.size;
// Operations that need special treatment for signed integers
if left_layout.abi.is_signed() {
let op: Option<fn(&i128, &i128) -> bool> = match bin_op {
Lt => Some(i128::lt),
Le => Some(i128::le),
Gt => Some(i128::gt),
Ge => Some(i128::ge),
_ => None,
};
if let Some(op) = op {
let l = self.sign_extend(l, left_layout) as i128;
let r = self.sign_extend(r, right_layout) as i128;
return Ok((Scalar::from_bool(op(&l, &r)), false, self.tcx.types.bool));
}
let op: Option<fn(i128, i128) -> (i128, bool)> = match bin_op {
Div if r == 0 => throw_ub!(DivisionByZero),
Rem if r == 0 => throw_ub!(RemainderByZero),
Div => Some(i128::overflowing_div),
Rem => Some(i128::overflowing_rem),
Add => Some(i128::overflowing_add),
Sub => Some(i128::overflowing_sub),
Mul => Some(i128::overflowing_mul),
_ => None,
};
if let Some(op) = op {
let r = self.sign_extend(r, right_layout) as i128;
// We need a special check for overflowing remainder:
// "int_min % -1" overflows and returns 0, but after casting things to a larger int
// type it does *not* overflow nor give an unrepresentable result!
if bin_op == Rem {
if r == -1 && l == (1 << (size.bits() - 1)) {
return Ok((Scalar::from_int(0, size), true, left_layout.ty));
}
}
let l = self.sign_extend(l, left_layout) as i128;
let (result, oflo) = op(l, r);
// This may be out-of-bounds for the result type, so we have to truncate ourselves.
// If that truncation loses any information, we have an overflow.
let result = result as u128;
let truncated = self.truncate(result, left_layout);
return Ok((
Scalar::from_uint(truncated, size),
oflo || self.sign_extend(truncated, left_layout) != result,
left_layout.ty,
));
}
}
let (val, ty) = match bin_op {
Eq => (Scalar::from_bool(l == r), self.tcx.types.bool),
Ne => (Scalar::from_bool(l != r), self.tcx.types.bool),
Lt => (Scalar::from_bool(l < r), self.tcx.types.bool),
Le => (Scalar::from_bool(l <= r), self.tcx.types.bool),
Gt => (Scalar::from_bool(l > r), self.tcx.types.bool),
Ge => (Scalar::from_bool(l >= r), self.tcx.types.bool),
BitOr => (Scalar::from_uint(l | r, size), left_layout.ty),
BitAnd => (Scalar::from_uint(l & r, size), left_layout.ty),
BitXor => (Scalar::from_uint(l ^ r, size), left_layout.ty),
Add | Sub | Mul | Rem | Div => {
assert!(!left_layout.abi.is_signed());
let op: fn(u128, u128) -> (u128, bool) = match bin_op {
Add => u128::overflowing_add,
Sub => u128::overflowing_sub,
Mul => u128::overflowing_mul,
Div if r == 0 => throw_ub!(DivisionByZero),
Rem if r == 0 => throw_ub!(RemainderByZero),
Div => u128::overflowing_div,
Rem => u128::overflowing_rem,
_ => bug!(),
};
let (result, oflo) = op(l, r);
// Truncate to target type.
// If that truncation loses any information, we have an overflow.
let truncated = self.truncate(result, left_layout);
return Ok((
Scalar::from_uint(truncated, size),
oflo || truncated != result,
left_layout.ty,
));
}
_ => span_bug!(
self.cur_span(),
"invalid binary op {:?}: {:?}, {:?} (both {:?})",
bin_op,
l,
r,
right_layout.ty,
),
};
Ok((val, false, ty))
}
/// Returns the result of the specified operation, whether it overflowed, and
/// the result type.
pub fn overflowing_binary_op(
&self,
bin_op: mir::BinOp,
left: &ImmTy<'tcx, M::PointerTag>,
right: &ImmTy<'tcx, M::PointerTag>,
) -> InterpResult<'tcx, (Scalar<M::PointerTag>, bool, Ty<'tcx>)> {
trace!(
"Running binary op {:?}: {:?} ({:?}), {:?} ({:?})",
bin_op,
*left,
left.layout.ty,
*right,
right.layout.ty
);
match left.layout.ty.kind() {
ty::Char => {
assert_eq!(left.layout.ty, right.layout.ty);
let left = left.to_scalar()?;
let right = right.to_scalar()?;
Ok(self.binary_char_op(bin_op, left.to_char()?, right.to_char()?))
}
ty::Bool => {
assert_eq!(left.layout.ty, right.layout.ty);
let left = left.to_scalar()?;
let right = right.to_scalar()?;
Ok(self.binary_bool_op(bin_op, left.to_bool()?, right.to_bool()?))
}
ty::Float(fty) => {
assert_eq!(left.layout.ty, right.layout.ty);
let ty = left.layout.ty;
let left = left.to_scalar()?;
let right = right.to_scalar()?;
Ok(match fty {
FloatTy::F32 => {
self.binary_float_op(bin_op, ty, left.to_f32()?, right.to_f32()?)
}
FloatTy::F64 => {
self.binary_float_op(bin_op, ty, left.to_f64()?, right.to_f64()?)
}
})
}
_ if left.layout.ty.is_integral() => {
// the RHS type can be different, e.g. for shifts -- but it has to be integral, too
assert!(
right.layout.ty.is_integral(),
"Unexpected types for BinOp: {:?} {:?} {:?}",
left.layout.ty,
bin_op,
right.layout.ty
);
let l = left.to_scalar()?.to_bits(left.layout.size)?;
let r = right.to_scalar()?.to_bits(right.layout.size)?;
self.binary_int_op(bin_op, l, left.layout, r, right.layout)
}
_ if left.layout.ty.is_any_ptr() => {
// The RHS type must be the same *or an integer type* (for `Offset`).
assert!(
right.layout.ty == left.layout.ty || right.layout.ty.is_integral(),
"Unexpected types for BinOp: {:?} {:?} {:?}",
left.layout.ty,
bin_op,
right.layout.ty
);
M::binary_ptr_op(self, bin_op, left, right)
}
_ => span_bug!(
self.cur_span(),
"Invalid MIR: bad LHS type for binop: {:?}",
left.layout.ty
),
}
}
/// Typed version of `overflowing_binary_op`, returning an `ImmTy`. Also ignores overflows.
#[inline]
pub fn binary_op(
&self,
bin_op: mir::BinOp,
left: &ImmTy<'tcx, M::PointerTag>,
right: &ImmTy<'tcx, M::PointerTag>,
) -> InterpResult<'tcx, ImmTy<'tcx, M::PointerTag>> {
let (val, _overflow, ty) = self.overflowing_binary_op(bin_op, left, right)?;
Ok(ImmTy::from_scalar(val, self.layout_of(ty)?))
}
/// Returns the result of the specified operation, whether it overflowed, and
/// the result type.
pub fn overflowing_unary_op(
&self,
un_op: mir::UnOp,
val: &ImmTy<'tcx, M::PointerTag>,
) -> InterpResult<'tcx, (Scalar<M::PointerTag>, bool, Ty<'tcx>)> {
use rustc_middle::mir::UnOp::*;
let layout = val.layout;
let val = val.to_scalar()?;
trace!("Running unary op {:?}: {:?} ({:?})", un_op, val, layout.ty);
match layout.ty.kind() {
ty::Bool => {
let val = val.to_bool()?;
let res = match un_op {
Not => !val,
_ => span_bug!(self.cur_span(), "Invalid bool op {:?}", un_op),
};
Ok((Scalar::from_bool(res), false, self.tcx.types.bool))
}
ty::Float(fty) => {
let res = match (un_op, fty) {
(Neg, FloatTy::F32) => Scalar::from_f32(-val.to_f32()?),
(Neg, FloatTy::F64) => Scalar::from_f64(-val.to_f64()?),
_ => span_bug!(self.cur_span(), "Invalid float op {:?}", un_op),
};
Ok((res, false, layout.ty))
}
_ => {
assert!(layout.ty.is_integral());
let val = val.to_bits(layout.size)?;
let (res, overflow) = match un_op {
Not => (self.truncate(!val, layout), false), // bitwise negation, then truncate
Neg => {
// arithmetic negation
assert!(layout.abi.is_signed());
let val = self.sign_extend(val, layout) as i128;
let (res, overflow) = val.overflowing_neg();
let res = res as u128;
// Truncate to target type.
// If that truncation loses any information, we have an overflow.
let truncated = self.truncate(res, layout);
(truncated, overflow || self.sign_extend(truncated, layout) != res)
}
};
Ok((Scalar::from_uint(res, layout.size), overflow, layout.ty))
}
}
}
pub fn unary_op(
&self,
un_op: mir::UnOp,
val: &ImmTy<'tcx, M::PointerTag>,
) -> InterpResult<'tcx, ImmTy<'tcx, M::PointerTag>> {
let (val, _overflow, ty) = self.overflowing_unary_op(un_op, val)?;
Ok(ImmTy::from_scalar(val, self.layout_of(ty)?))
}
}

1091
compiler/rustc_const_eval/src/interpret/place.rs Normal file
View File

File diff suppressed because it is too large Load Diff

316
compiler/rustc_const_eval/src/interpret/step.rs Normal file
View File

@@ -0,0 +1,316 @@
//! This module contains the `InterpCx` methods for executing a single step of the interpreter.
//!
//! The main entry point is the `step` method.
use rustc_middle::mir;
use rustc_middle::mir::interpret::{InterpResult, Scalar};
use rustc_middle::ty::layout::LayoutOf;
use super::{InterpCx, Machine};
/// Classify whether an operator is "left-homogeneous", i.e., the LHS has the
/// same type as the result.
#[inline]
fn binop_left_homogeneous(op: mir::BinOp) -> bool {
use rustc_middle::mir::BinOp::*;
match op {
Add | Sub | Mul | Div | Rem | BitXor | BitAnd | BitOr | Offset | Shl | Shr => true,
Eq | Ne | Lt | Le | Gt | Ge => false,
}
}
/// Classify whether an operator is "right-homogeneous", i.e., the RHS has the
/// same type as the LHS.
#[inline]
fn binop_right_homogeneous(op: mir::BinOp) -> bool {
use rustc_middle::mir::BinOp::*;
match op {
Add | Sub | Mul | Div | Rem | BitXor | BitAnd | BitOr | Eq | Ne | Lt | Le | Gt | Ge => true,
Offset | Shl | Shr => false,
}
}
impl<'mir, 'tcx: 'mir, M: Machine<'mir, 'tcx>> InterpCx<'mir, 'tcx, M> {
pub fn run(&mut self) -> InterpResult<'tcx> {
while self.step()? {}
Ok(())
}
/// Returns `true` as long as there are more things to do.
///
/// This is used by [priroda](https://github.com/oli-obk/priroda)
///
/// This is marked `#inline(always)` to work around adverserial codegen when `opt-level = 3`
#[inline(always)]
pub fn step(&mut self) -> InterpResult<'tcx, bool> {
if self.stack().is_empty() {
return Ok(false);
}
let loc = match self.frame().loc {
Ok(loc) => loc,
Err(_) => {
// We are unwinding and this fn has no cleanup code.
// Just go on unwinding.
trace!("unwinding: skipping frame");
self.pop_stack_frame(/* unwinding */ true)?;
return Ok(true);
}
};
let basic_block = &self.body().basic_blocks()[loc.block];
let old_frames = self.frame_idx();
if let Some(stmt) = basic_block.statements.get(loc.statement_index) {
assert_eq!(old_frames, self.frame_idx());
self.statement(stmt)?;
return Ok(true);
}
M::before_terminator(self)?;
let terminator = basic_block.terminator();
assert_eq!(old_frames, self.frame_idx());
self.terminator(terminator)?;
Ok(true)
}
/// Runs the interpretation logic for the given `mir::Statement` at the current frame and
/// statement counter. This also moves the statement counter forward.
pub fn statement(&mut self, stmt: &mir::Statement<'tcx>) -> InterpResult<'tcx> {
info!("{:?}", stmt);
use rustc_middle::mir::StatementKind::*;
// Some statements (e.g., box) push new stack frames.
// We have to record the stack frame number *before* executing the statement.
let frame_idx = self.frame_idx();
match &stmt.kind {
Assign(box (place, rvalue)) => self.eval_rvalue_into_place(rvalue, *place)?,
SetDiscriminant { place, variant_index } => {
let dest = self.eval_place(**place)?;
self.write_discriminant(*variant_index, &dest)?;
}
// Mark locals as alive
StorageLive(local) => {
self.storage_live(*local)?;
}
// Mark locals as dead
StorageDead(local) => {
self.storage_dead(*local)?;
}
// No dynamic semantics attached to `FakeRead`; MIR
// interpreter is solely intended for borrowck'ed code.
FakeRead(..) => {}
// Stacked Borrows.
Retag(kind, place) => {
let dest = self.eval_place(**place)?;
M::retag(self, *kind, &dest)?;
}
// Call CopyNonOverlapping
CopyNonOverlapping(box rustc_middle::mir::CopyNonOverlapping { src, dst, count }) => {
let src = self.eval_operand(src, None)?;
let dst = self.eval_operand(dst, None)?;
let count = self.eval_operand(count, None)?;
self.copy_intrinsic(&src, &dst, &count, /* nonoverlapping */ true)?;
}
// Statements we do not track.
AscribeUserType(..) => {}
// Currently, Miri discards Coverage statements. Coverage statements are only injected
// via an optional compile time MIR pass and have no side effects. Since Coverage
// statements don't exist at the source level, it is safe for Miri to ignore them, even
// for undefined behavior (UB) checks.
//
// A coverage counter inside a const expression (for example, a counter injected in a
// const function) is discarded when the const is evaluated at compile time. Whether
// this should change, and/or how to implement a const eval counter, is a subject of the
// following issue:
//
// FIXME(#73156): Handle source code coverage in const eval
Coverage(..) => {}
// Defined to do nothing. These are added by optimization passes, to avoid changing the
// size of MIR constantly.
Nop => {}
LlvmInlineAsm { .. } => throw_unsup_format!("inline assembly is not supported"),
}
self.stack_mut()[frame_idx].loc.as_mut().unwrap().statement_index += 1;
Ok(())
}
/// Evaluate an assignment statement.
///
/// There is no separate `eval_rvalue` function. Instead, the code for handling each rvalue
/// type writes its results directly into the memory specified by the place.
pub fn eval_rvalue_into_place(
&mut self,
rvalue: &mir::Rvalue<'tcx>,
place: mir::Place<'tcx>,
) -> InterpResult<'tcx> {
let dest = self.eval_place(place)?;
use rustc_middle::mir::Rvalue::*;
match *rvalue {
ThreadLocalRef(did) => {
let ptr = M::thread_local_static_base_pointer(self, did)?;
self.write_pointer(ptr, &dest)?;
}
Use(ref operand) => {
// Avoid recomputing the layout
let op = self.eval_operand(operand, Some(dest.layout))?;
self.copy_op(&op, &dest)?;
}
BinaryOp(bin_op, box (ref left, ref right)) => {
let layout = binop_left_homogeneous(bin_op).then_some(dest.layout);
let left = self.read_immediate(&self.eval_operand(left, layout)?)?;
let layout = binop_right_homogeneous(bin_op).then_some(left.layout);
let right = self.read_immediate(&self.eval_operand(right, layout)?)?;
self.binop_ignore_overflow(bin_op, &left, &right, &dest)?;
}
CheckedBinaryOp(bin_op, box (ref left, ref right)) => {
// Due to the extra boolean in the result, we can never reuse the `dest.layout`.
let left = self.read_immediate(&self.eval_operand(left, None)?)?;
let layout = binop_right_homogeneous(bin_op).then_some(left.layout);
let right = self.read_immediate(&self.eval_operand(right, layout)?)?;
self.binop_with_overflow(bin_op, &left, &right, &dest)?;
}
UnaryOp(un_op, ref operand) => {
// The operand always has the same type as the result.
let val = self.read_immediate(&self.eval_operand(operand, Some(dest.layout))?)?;
let val = self.unary_op(un_op, &val)?;
assert_eq!(val.layout, dest.layout, "layout mismatch for result of {:?}", un_op);
self.write_immediate(*val, &dest)?;
}
Aggregate(ref kind, ref operands) => {
let (dest, active_field_index) = match **kind {
mir::AggregateKind::Adt(adt_def, variant_index, _, _, active_field_index) => {
self.write_discriminant(variant_index, &dest)?;
if adt_def.is_enum() {
(self.place_downcast(&dest, variant_index)?, active_field_index)
} else {
(dest, active_field_index)
}
}
_ => (dest, None),
};
for (i, operand) in operands.iter().enumerate() {
let op = self.eval_operand(operand, None)?;
// Ignore zero-sized fields.
if !op.layout.is_zst() {
let field_index = active_field_index.unwrap_or(i);
let field_dest = self.place_field(&dest, field_index)?;
self.copy_op(&op, &field_dest)?;
}
}
}
Repeat(ref operand, _) => {
let src = self.eval_operand(operand, None)?;
assert!(!src.layout.is_unsized());
let dest = self.force_allocation(&dest)?;
let length = dest.len(self)?;
if length == 0 {
// Nothing to copy... but let's still make sure that `dest` as a place is valid.
self.get_alloc_mut(&dest)?;
} else {
// Write the src to the first element.
let first = self.mplace_field(&dest, 0)?;
self.copy_op(&src, &first.into())?;
// This is performance-sensitive code for big static/const arrays! So we
// avoid writing each operand individually and instead just make many copies
// of the first element.
let elem_size = first.layout.size;
let first_ptr = first.ptr;
let rest_ptr = first_ptr.offset(elem_size, self)?;
self.memory.copy_repeatedly(
first_ptr,
first.align,
rest_ptr,
first.align,
elem_size,
length - 1,
/*nonoverlapping:*/ true,
)?;
}
}
Len(place) => {
// FIXME(CTFE): don't allow computing the length of arrays in const eval
let src = self.eval_place(place)?;
let mplace = self.force_allocation(&src)?;
let len = mplace.len(self)?;
self.write_scalar(Scalar::from_machine_usize(len, self), &dest)?;
}
AddressOf(_, place) | Ref(_, _, place) => {
let src = self.eval_place(place)?;
let place = self.force_allocation(&src)?;
self.write_immediate(place.to_ref(self), &dest)?;
}
NullaryOp(mir::NullOp::Box, _) => {
M::box_alloc(self, &dest)?;
}
NullaryOp(mir::NullOp::SizeOf, ty) => {
let ty = self.subst_from_current_frame_and_normalize_erasing_regions(ty);
let layout = self.layout_of(ty)?;
if layout.is_unsized() {
// FIXME: This should be a span_bug (#80742)
self.tcx.sess.delay_span_bug(
self.frame().current_span(),
&format!("SizeOf nullary MIR operator called for unsized type {}", ty),
);
throw_inval!(SizeOfUnsizedType(ty));
}
self.write_scalar(Scalar::from_machine_usize(layout.size.bytes(), self), &dest)?;
}
Cast(cast_kind, ref operand, cast_ty) => {
let src = self.eval_operand(operand, None)?;
let cast_ty = self.subst_from_current_frame_and_normalize_erasing_regions(cast_ty);
self.cast(&src, cast_kind, cast_ty, &dest)?;
}
Discriminant(place) => {
let op = self.eval_place_to_op(place, None)?;
let discr_val = self.read_discriminant(&op)?.0;
self.write_scalar(discr_val, &dest)?;
}
}
trace!("{:?}", self.dump_place(*dest));
Ok(())
}
fn terminator(&mut self, terminator: &mir::Terminator<'tcx>) -> InterpResult<'tcx> {
info!("{:?}", terminator.kind);
self.eval_terminator(terminator)?;
if !self.stack().is_empty() {
if let Ok(loc) = self.frame().loc {
info!("// executing {:?}", loc.block);
}
}
Ok(())
}
}

517
compiler/rustc_const_eval/src/interpret/terminator.rs Normal file
View File

@@ -0,0 +1,517 @@
use std::borrow::Cow;
use std::convert::TryFrom;
use rustc_middle::middle::codegen_fn_attrs::CodegenFnAttrFlags;
use rustc_middle::ty::layout::{self, LayoutOf as _, TyAndLayout};
use rustc_middle::ty::Instance;
use rustc_middle::{
mir,
ty::{self, Ty},
};
use rustc_target::abi;
use rustc_target::spec::abi::Abi;
use super::{
FnVal, ImmTy, InterpCx, InterpResult, MPlaceTy, Machine, OpTy, PlaceTy, Scalar,
StackPopCleanup, StackPopUnwind,
};
impl<'mir, 'tcx: 'mir, M: Machine<'mir, 'tcx>> InterpCx<'mir, 'tcx, M> {
fn fn_can_unwind(&self, attrs: CodegenFnAttrFlags, abi: Abi) -> bool {
layout::fn_can_unwind(*self.tcx, attrs, abi)
}
pub(super) fn eval_terminator(
&mut self,
terminator: &mir::Terminator<'tcx>,
) -> InterpResult<'tcx> {
use rustc_middle::mir::TerminatorKind::*;
match terminator.kind {
Return => {
self.pop_stack_frame(/* unwinding */ false)?
}
Goto { target } => self.go_to_block(target),
SwitchInt { ref discr, ref targets, switch_ty } => {
let discr = self.read_immediate(&self.eval_operand(discr, None)?)?;
trace!("SwitchInt({:?})", *discr);
assert_eq!(discr.layout.ty, switch_ty);
// Branch to the `otherwise` case by default, if no match is found.
assert!(!targets.iter().is_empty());
let mut target_block = targets.otherwise();
for (const_int, target) in targets.iter() {
// Compare using binary_op, to also support pointer values
let res = self
.overflowing_binary_op(
mir::BinOp::Eq,
&discr,
&ImmTy::from_uint(const_int, discr.layout),
)?
.0;
if res.to_bool()? {
target_block = target;
break;
}
}
self.go_to_block(target_block);
}
Call { ref func, ref args, destination, ref cleanup, from_hir_call: _, fn_span: _ } => {
let old_stack = self.frame_idx();
let old_loc = self.frame().loc;
let func = self.eval_operand(func, None)?;
let (fn_val, abi, caller_can_unwind) = match *func.layout.ty.kind() {
ty::FnPtr(sig) => {
let caller_abi = sig.abi();
let fn_ptr = self.read_pointer(&func)?;
let fn_val = self.memory.get_fn(fn_ptr)?;
(
fn_val,
caller_abi,
self.fn_can_unwind(CodegenFnAttrFlags::empty(), caller_abi),
)
}
ty::FnDef(def_id, substs) => {
let sig = func.layout.ty.fn_sig(*self.tcx);
(
FnVal::Instance(
self.resolve(ty::WithOptConstParam::unknown(def_id), substs)?,
),
sig.abi(),
self.fn_can_unwind(self.tcx.codegen_fn_attrs(def_id).flags, sig.abi()),
)
}
_ => span_bug!(
terminator.source_info.span,
"invalid callee of type {:?}",
func.layout.ty
),
};
let args = self.eval_operands(args)?;
let dest_place;
let ret = match destination {
Some((dest, ret)) => {
dest_place = self.eval_place(dest)?;
Some((&dest_place, ret))
}
None => None,
};
self.eval_fn_call(
fn_val,
abi,
&args[..],
ret,
match (cleanup, caller_can_unwind) {
(Some(cleanup), true) => StackPopUnwind::Cleanup(*cleanup),
(None, true) => StackPopUnwind::Skip,
(_, false) => StackPopUnwind::NotAllowed,
},
)?;
// Sanity-check that `eval_fn_call` either pushed a new frame or
// did a jump to another block.
if self.frame_idx() == old_stack && self.frame().loc == old_loc {
span_bug!(terminator.source_info.span, "evaluating this call made no progress");
}
}
Drop { place, target, unwind } => {
let place = self.eval_place(place)?;
let ty = place.layout.ty;
trace!("TerminatorKind::drop: {:?}, type {}", place, ty);
let instance = Instance::resolve_drop_in_place(*self.tcx, ty);
self.drop_in_place(&place, instance, target, unwind)?;
}
Assert { ref cond, expected, ref msg, target, cleanup } => {
let cond_val =
self.read_immediate(&self.eval_operand(cond, None)?)?.to_scalar()?.to_bool()?;
if expected == cond_val {
self.go_to_block(target);
} else {
M::assert_panic(self, msg, cleanup)?;
}
}
Abort => {
M::abort(self, "the program aborted execution".to_owned())?;
}
// When we encounter Resume, we've finished unwinding
// cleanup for the current stack frame. We pop it in order
// to continue unwinding the next frame
Resume => {
trace!("unwinding: resuming from cleanup");
// By definition, a Resume terminator means
// that we're unwinding
self.pop_stack_frame(/* unwinding */ true)?;
return Ok(());
}
// It is UB to ever encounter this.
Unreachable => throw_ub!(Unreachable),
// These should never occur for MIR we actually run.
DropAndReplace { .. }
| FalseEdge { .. }
| FalseUnwind { .. }
| Yield { .. }
| GeneratorDrop => span_bug!(
terminator.source_info.span,
"{:#?} should have been eliminated by MIR pass",
terminator.kind
),
// Inline assembly can't be interpreted.
InlineAsm { .. } => throw_unsup_format!("inline assembly is not supported"),
}
Ok(())
}
fn check_argument_compat(
rust_abi: bool,
caller: TyAndLayout<'tcx>,
callee: TyAndLayout<'tcx>,
) -> bool {
if caller.ty == callee.ty {
// No question
return true;
}
if !rust_abi {
// Don't risk anything
return false;
}
// Compare layout
match (&caller.abi, &callee.abi) {
// Different valid ranges are okay (once we enforce validity,
// that will take care to make it UB to leave the range, just
// like for transmute).
(abi::Abi::Scalar(ref caller), abi::Abi::Scalar(ref callee)) => {
caller.value == callee.value
}
(
abi::Abi::ScalarPair(ref caller1, ref caller2),
abi::Abi::ScalarPair(ref callee1, ref callee2),
) => caller1.value == callee1.value && caller2.value == callee2.value,
// Be conservative
_ => false,
}
}
/// Pass a single argument, checking the types for compatibility.
fn pass_argument(
&mut self,
rust_abi: bool,
caller_arg: &mut impl Iterator<Item = OpTy<'tcx, M::PointerTag>>,
callee_arg: &PlaceTy<'tcx, M::PointerTag>,
) -> InterpResult<'tcx> {
if rust_abi && callee_arg.layout.is_zst() {
// Nothing to do.
trace!("Skipping callee ZST");
return Ok(());
}
let caller_arg = caller_arg.next().ok_or_else(|| {
err_ub_format!("calling a function with fewer arguments than it requires")
})?;
if rust_abi {
assert!(!caller_arg.layout.is_zst(), "ZSTs must have been already filtered out");
}
// Now, check
if !Self::check_argument_compat(rust_abi, caller_arg.layout, callee_arg.layout) {
throw_ub_format!(
"calling a function with argument of type {:?} passing data of type {:?}",
callee_arg.layout.ty,
caller_arg.layout.ty
)
}
// We allow some transmutes here
self.copy_op_transmute(&caller_arg, callee_arg)
}
/// Call this function -- pushing the stack frame and initializing the arguments.
fn eval_fn_call(
&mut self,
fn_val: FnVal<'tcx, M::ExtraFnVal>,
caller_abi: Abi,
args: &[OpTy<'tcx, M::PointerTag>],
ret: Option<(&PlaceTy<'tcx, M::PointerTag>, mir::BasicBlock)>,
mut unwind: StackPopUnwind,
) -> InterpResult<'tcx> {
trace!("eval_fn_call: {:#?}", fn_val);
let instance = match fn_val {
FnVal::Instance(instance) => instance,
FnVal::Other(extra) => {
return M::call_extra_fn(self, extra, caller_abi, args, ret, unwind);
}
};
let get_abi = |this: &Self, instance_ty: Ty<'tcx>| match instance_ty.kind() {
ty::FnDef(..) => instance_ty.fn_sig(*this.tcx).abi(),
ty::Closure(..) => Abi::RustCall,
ty::Generator(..) => Abi::Rust,
_ => span_bug!(this.cur_span(), "unexpected callee ty: {:?}", instance_ty),
};
// ABI check
let check_abi = |callee_abi: Abi| -> InterpResult<'tcx> {
let normalize_abi = |abi| match abi {
Abi::Rust | Abi::RustCall | Abi::RustIntrinsic | Abi::PlatformIntrinsic =>
// These are all the same ABI, really.
{
Abi::Rust
}
abi => abi,
};
if normalize_abi(caller_abi) != normalize_abi(callee_abi) {
throw_ub_format!(
"calling a function with ABI {} using caller ABI {}",
callee_abi.name(),
caller_abi.name()
)
}
Ok(())
};
match instance.def {
ty::InstanceDef::Intrinsic(..) => {
if M::enforce_abi(self) {
check_abi(get_abi(self, instance.ty(*self.tcx, self.param_env)))?;
}
assert!(caller_abi == Abi::RustIntrinsic || caller_abi == Abi::PlatformIntrinsic);
M::call_intrinsic(self, instance, args, ret, unwind)
}
ty::InstanceDef::VtableShim(..)
| ty::InstanceDef::ReifyShim(..)
| ty::InstanceDef::ClosureOnceShim { .. }
| ty::InstanceDef::FnPtrShim(..)
| ty::InstanceDef::DropGlue(..)
| ty::InstanceDef::CloneShim(..)
| ty::InstanceDef::Item(_) => {
// We need MIR for this fn
let body =
match M::find_mir_or_eval_fn(self, instance, caller_abi, args, ret, unwind)? {
Some(body) => body,
None => return Ok(()),
};
// Check against the ABI of the MIR body we are calling (not the ABI of `instance`;
// these can differ when `find_mir_or_eval_fn` does something clever like resolve
// exported symbol names).
let callee_def_id = body.source.def_id();
let callee_abi = get_abi(self, self.tcx.type_of(callee_def_id));
if M::enforce_abi(self) {
check_abi(callee_abi)?;
}
if !matches!(unwind, StackPopUnwind::NotAllowed)
&& !self
.fn_can_unwind(self.tcx.codegen_fn_attrs(callee_def_id).flags, callee_abi)
{
// The callee cannot unwind.
unwind = StackPopUnwind::NotAllowed;
}
self.push_stack_frame(
instance,
body,
ret.map(|p| p.0),
StackPopCleanup::Goto { ret: ret.map(|p| p.1), unwind },
)?;
// If an error is raised here, pop the frame again to get an accurate backtrace.
// To this end, we wrap it all in a `try` block.
let res: InterpResult<'tcx> = try {
trace!(
"caller ABI: {:?}, args: {:#?}",
caller_abi,
args.iter()
.map(|arg| (arg.layout.ty, format!("{:?}", **arg)))
.collect::<Vec<_>>()
);
trace!(
"spread_arg: {:?}, locals: {:#?}",
body.spread_arg,
body.args_iter()
.map(|local| (
local,
self.layout_of_local(self.frame(), local, None).unwrap().ty
))
.collect::<Vec<_>>()
);
// Figure out how to pass which arguments.
// The Rust ABI is special: ZST get skipped.
let rust_abi = match caller_abi {
Abi::Rust | Abi::RustCall => true,
_ => false,
};
// We have two iterators: Where the arguments come from,
// and where they go to.
// For where they come from: If the ABI is RustCall, we untuple the
// last incoming argument. These two iterators do not have the same type,
// so to keep the code paths uniform we accept an allocation
// (for RustCall ABI only).
let caller_args: Cow<'_, [OpTy<'tcx, M::PointerTag>]> =
if caller_abi == Abi::RustCall && !args.is_empty() {
// Untuple
let (untuple_arg, args) = args.split_last().unwrap();
trace!("eval_fn_call: Will pass last argument by untupling");
Cow::from(
args.iter()
.map(|&a| Ok(a))
.chain(
(0..untuple_arg.layout.fields.count())
.map(|i| self.operand_field(untuple_arg, i)),
)
.collect::<InterpResult<'_, Vec<OpTy<'tcx, M::PointerTag>>>>(
)?,
)
} else {
// Plain arg passing
Cow::from(args)
};
// Skip ZSTs
let mut caller_iter =
caller_args.iter().filter(|op| !rust_abi || !op.layout.is_zst()).copied();
// Now we have to spread them out across the callee's locals,
// taking into account the `spread_arg`. If we could write
// this is a single iterator (that handles `spread_arg`), then
// `pass_argument` would be the loop body. It takes care to
// not advance `caller_iter` for ZSTs.
for local in body.args_iter() {
let dest = self.eval_place(mir::Place::from(local))?;
if Some(local) == body.spread_arg {
// Must be a tuple
for i in 0..dest.layout.fields.count() {
let dest = self.place_field(&dest, i)?;
self.pass_argument(rust_abi, &mut caller_iter, &dest)?;
}
} else {
// Normal argument
self.pass_argument(rust_abi, &mut caller_iter, &dest)?;
}
}
// Now we should have no more caller args
if caller_iter.next().is_some() {
throw_ub_format!("calling a function with more arguments than it expected")
}
// Don't forget to check the return type!
if let Some((caller_ret, _)) = ret {
let callee_ret = self.eval_place(mir::Place::return_place())?;
if !Self::check_argument_compat(
rust_abi,
caller_ret.layout,
callee_ret.layout,
) {
throw_ub_format!(
"calling a function with return type {:?} passing \
return place of type {:?}",
callee_ret.layout.ty,
caller_ret.layout.ty
)
}
} else {
let local = mir::RETURN_PLACE;
let callee_layout = self.layout_of_local(self.frame(), local, None)?;
if !callee_layout.abi.is_uninhabited() {
throw_ub_format!("calling a returning function without a return place")
}
}
};
match res {
Err(err) => {
self.stack_mut().pop();
Err(err)
}
Ok(()) => Ok(()),
}
}
// cannot use the shim here, because that will only result in infinite recursion
ty::InstanceDef::Virtual(_, idx) => {
let mut args = args.to_vec();
// We have to implement all "object safe receivers". Currently we
// support built-in pointers `(&, &mut, Box)` as well as unsized-self. We do
// not yet support custom self types.
// Also see `compiler/rustc_codegen_llvm/src/abi.rs` and `compiler/rustc_codegen_ssa/src/mir/block.rs`.
let receiver_place = match args[0].layout.ty.builtin_deref(true) {
Some(_) => {
// Built-in pointer.
self.deref_operand(&args[0])?
}
None => {
// Unsized self.
args[0].assert_mem_place()
}
};
// Find and consult vtable
let vtable = self.scalar_to_ptr(receiver_place.vtable());
let fn_val = self.get_vtable_slot(vtable, u64::try_from(idx).unwrap())?;
// `*mut receiver_place.layout.ty` is almost the layout that we
// want for args[0]: We have to project to field 0 because we want
// a thin pointer.
assert!(receiver_place.layout.is_unsized());
let receiver_ptr_ty = self.tcx.mk_mut_ptr(receiver_place.layout.ty);
let this_receiver_ptr = self.layout_of(receiver_ptr_ty)?.field(self, 0);
// Adjust receiver argument.
args[0] = OpTy::from(ImmTy::from_immediate(
Scalar::from_maybe_pointer(receiver_place.ptr, self).into(),
this_receiver_ptr,
));
trace!("Patched self operand to {:#?}", args[0]);
// recurse with concrete function
self.eval_fn_call(fn_val, caller_abi, &args, ret, unwind)
}
}
}
fn drop_in_place(
&mut self,
place: &PlaceTy<'tcx, M::PointerTag>,
instance: ty::Instance<'tcx>,
target: mir::BasicBlock,
unwind: Option<mir::BasicBlock>,
) -> InterpResult<'tcx> {
trace!("drop_in_place: {:?},\n {:?}, {:?}", *place, place.layout.ty, instance);
// We take the address of the object. This may well be unaligned, which is fine
// for us here. However, unaligned accesses will probably make the actual drop
// implementation fail -- a problem shared by rustc.
let place = self.force_allocation(place)?;
let (instance, place) = match place.layout.ty.kind() {
ty::Dynamic(..) => {
// Dropping a trait object.
self.unpack_dyn_trait(&place)?
}
_ => (instance, place),
};
let arg = ImmTy::from_immediate(
place.to_ref(self),
self.layout_of(self.tcx.mk_mut_ptr(place.layout.ty))?,
);
let ty = self.tcx.mk_unit(); // return type is ()
let dest = MPlaceTy::dangling(self.layout_of(ty)?);
self.eval_fn_call(
FnVal::Instance(instance),
Abi::Rust,
&[arg.into()],
Some((&dest.into(), target)),
match unwind {
Some(cleanup) => StackPopUnwind::Cleanup(cleanup),
None => StackPopUnwind::Skip,
},
)
}
}

142
compiler/rustc_const_eval/src/interpret/traits.rs Normal file
View File

@@ -0,0 +1,142 @@
use std::convert::TryFrom;
use rustc_middle::mir::interpret::{InterpResult, Pointer, PointerArithmetic};
use rustc_middle::ty::{
self, Ty, COMMON_VTABLE_ENTRIES, COMMON_VTABLE_ENTRIES_ALIGN,
COMMON_VTABLE_ENTRIES_DROPINPLACE, COMMON_VTABLE_ENTRIES_SIZE,
};
use rustc_target::abi::{Align, Size};
use super::util::ensure_monomorphic_enough;
use super::{FnVal, InterpCx, Machine};
impl<'mir, 'tcx: 'mir, M: Machine<'mir, 'tcx>> InterpCx<'mir, 'tcx, M> {
/// Creates a dynamic vtable for the given type and vtable origin. This is used only for
/// objects.
///
/// The `trait_ref` encodes the erased self type. Hence, if we are
/// making an object `Foo<Trait>` from a value of type `Foo<T>`, then
/// `trait_ref` would map `T: Trait`.
pub fn get_vtable(
&mut self,
ty: Ty<'tcx>,
poly_trait_ref: Option<ty::PolyExistentialTraitRef<'tcx>>,
) -> InterpResult<'tcx, Pointer<Option<M::PointerTag>>> {
trace!("get_vtable(trait_ref={:?})", poly_trait_ref);
let (ty, poly_trait_ref) = self.tcx.erase_regions((ty, poly_trait_ref));
// All vtables must be monomorphic, bail out otherwise.
ensure_monomorphic_enough(*self.tcx, ty)?;
ensure_monomorphic_enough(*self.tcx, poly_trait_ref)?;
let vtable_allocation = self.tcx.vtable_allocation(ty, poly_trait_ref);
let vtable_ptr = self.memory.global_base_pointer(Pointer::from(vtable_allocation))?;
Ok(vtable_ptr.into())
}
/// Resolves the function at the specified slot in the provided
/// vtable. Currently an index of '3' (`COMMON_VTABLE_ENTRIES.len()`)
/// corresponds to the first method declared in the trait of the provided vtable.
pub fn get_vtable_slot(
&self,
vtable: Pointer<Option<M::PointerTag>>,
idx: u64,
) -> InterpResult<'tcx, FnVal<'tcx, M::ExtraFnVal>> {
let ptr_size = self.pointer_size();
let vtable_slot = vtable.offset(ptr_size * idx, self)?;
let vtable_slot = self
.memory
.get(vtable_slot, ptr_size, self.tcx.data_layout.pointer_align.abi)?
.expect("cannot be a ZST");
let fn_ptr = self.scalar_to_ptr(vtable_slot.read_ptr_sized(Size::ZERO)?.check_init()?);
self.memory.get_fn(fn_ptr)
}
/// Returns the drop fn instance as well as the actual dynamic type.
pub fn read_drop_type_from_vtable(
&self,
vtable: Pointer<Option<M::PointerTag>>,
) -> InterpResult<'tcx, (ty::Instance<'tcx>, Ty<'tcx>)> {
let pointer_size = self.pointer_size();
// We don't care about the pointee type; we just want a pointer.
let vtable = self
.memory
.get(
vtable,
pointer_size * u64::try_from(COMMON_VTABLE_ENTRIES.len()).unwrap(),
self.tcx.data_layout.pointer_align.abi,
)?
.expect("cannot be a ZST");
let drop_fn = vtable
.read_ptr_sized(
pointer_size * u64::try_from(COMMON_VTABLE_ENTRIES_DROPINPLACE).unwrap(),
)?
.check_init()?;
// We *need* an instance here, no other kind of function value, to be able
// to determine the type.
let drop_instance = self.memory.get_fn(self.scalar_to_ptr(drop_fn))?.as_instance()?;
trace!("Found drop fn: {:?}", drop_instance);
let fn_sig = drop_instance.ty(*self.tcx, self.param_env).fn_sig(*self.tcx);
let fn_sig = self.tcx.normalize_erasing_late_bound_regions(self.param_env, fn_sig);
// The drop function takes `*mut T` where `T` is the type being dropped, so get that.
let args = fn_sig.inputs();
if args.len() != 1 {
throw_ub!(InvalidVtableDropFn(fn_sig));
}
let ty =
args[0].builtin_deref(true).ok_or_else(|| err_ub!(InvalidVtableDropFn(fn_sig)))?.ty;
Ok((drop_instance, ty))
}
pub fn read_size_and_align_from_vtable(
&self,
vtable: Pointer<Option<M::PointerTag>>,
) -> InterpResult<'tcx, (Size, Align)> {
let pointer_size = self.pointer_size();
// We check for `size = 3 * ptr_size`, which covers the drop fn (unused here),
// the size, and the align (which we read below).
let vtable = self
.memory
.get(
vtable,
pointer_size * u64::try_from(COMMON_VTABLE_ENTRIES.len()).unwrap(),
self.tcx.data_layout.pointer_align.abi,
)?
.expect("cannot be a ZST");
let size = vtable
.read_ptr_sized(pointer_size * u64::try_from(COMMON_VTABLE_ENTRIES_SIZE).unwrap())?
.check_init()?;
let size = size.to_machine_usize(self)?;
let align = vtable
.read_ptr_sized(pointer_size * u64::try_from(COMMON_VTABLE_ENTRIES_ALIGN).unwrap())?
.check_init()?;
let align = align.to_machine_usize(self)?;
let align = Align::from_bytes(align).map_err(|e| err_ub!(InvalidVtableAlignment(e)))?;
if size >= self.tcx.data_layout.obj_size_bound() {
throw_ub!(InvalidVtableSize);
}
Ok((Size::from_bytes(size), align))
}
pub fn read_new_vtable_after_trait_upcasting_from_vtable(
&self,
vtable: Pointer<Option<M::PointerTag>>,
idx: u64,
) -> InterpResult<'tcx, Pointer<Option<M::PointerTag>>> {
let pointer_size = self.pointer_size();
let vtable_slot = vtable.offset(pointer_size * idx, self)?;
let new_vtable = self
.memory
.get(vtable_slot, pointer_size, self.tcx.data_layout.pointer_align.abi)?
.expect("cannot be a ZST");
let new_vtable = self.scalar_to_ptr(new_vtable.read_ptr_sized(Size::ZERO)?.check_init()?);
Ok(new_vtable)
}
}

84
compiler/rustc_const_eval/src/interpret/util.rs Normal file
View File

@@ -0,0 +1,84 @@
use rustc_middle::mir::interpret::InterpResult;
use rustc_middle::ty::{self, Ty, TyCtxt, TypeFoldable, TypeVisitor};
use std::convert::TryInto;
use std::ops::ControlFlow;
/// Returns `true` if a used generic parameter requires substitution.
crate fn ensure_monomorphic_enough<'tcx, T>(tcx: TyCtxt<'tcx>, ty: T) -> InterpResult<'tcx>
where
T: TypeFoldable<'tcx>,
{
debug!("ensure_monomorphic_enough: ty={:?}", ty);
if !ty.potentially_needs_subst() {
return Ok(());
}
struct FoundParam;
struct UsedParamsNeedSubstVisitor<'tcx> {
tcx: TyCtxt<'tcx>,
}
impl<'tcx> TypeVisitor<'tcx> for UsedParamsNeedSubstVisitor<'tcx> {
type BreakTy = FoundParam;
fn tcx_for_anon_const_substs(&self) -> Option<TyCtxt<'tcx>> {
Some(self.tcx)
}
fn visit_ty(&mut self, ty: Ty<'tcx>) -> ControlFlow<Self::BreakTy> {
if !ty.potentially_needs_subst() {
return ControlFlow::CONTINUE;
}
match *ty.kind() {
ty::Param(_) => ControlFlow::Break(FoundParam),
ty::Closure(def_id, substs)
| ty::Generator(def_id, substs, ..)
| ty::FnDef(def_id, substs) => {
let unused_params = self.tcx.unused_generic_params(def_id);
for (index, subst) in substs.into_iter().enumerate() {
let index = index
.try_into()
.expect("more generic parameters than can fit into a `u32`");
let is_used = unused_params.contains(index).map_or(true, |unused| !unused);
// Only recurse when generic parameters in fns, closures and generators
// are used and require substitution.
match (is_used, subst.definitely_needs_subst(self.tcx)) {
// Just in case there are closures or generators within this subst,
// recurse.
(true, true) => return subst.super_visit_with(self),
// Confirm that polymorphization replaced the parameter with
// `ty::Param`/`ty::ConstKind::Param`.
(false, true) if cfg!(debug_assertions) => match subst.unpack() {
ty::subst::GenericArgKind::Type(ty) => {
assert!(matches!(ty.kind(), ty::Param(_)))
}
ty::subst::GenericArgKind::Const(ct) => {
assert!(matches!(ct.val, ty::ConstKind::Param(_)))
}
ty::subst::GenericArgKind::Lifetime(..) => (),
},
_ => {}
}
}
ControlFlow::CONTINUE
}
_ => ty.super_visit_with(self),
}
}
fn visit_const(&mut self, c: &'tcx ty::Const<'tcx>) -> ControlFlow<Self::BreakTy> {
match c.val {
ty::ConstKind::Param(..) => ControlFlow::Break(FoundParam),
_ => c.super_visit_with(self),
}
}
}
let mut vis = UsedParamsNeedSubstVisitor { tcx };
if matches!(ty.visit_with(&mut vis), ControlFlow::Break(FoundParam)) {
throw_inval!(TooGeneric);
} else {
Ok(())
}
}

965
compiler/rustc_const_eval/src/interpret/validity.rs Normal file
View File

@@ -0,0 +1,965 @@
//! Check the validity invariant of a given value, and tell the user
//! where in the value it got violated.
//! In const context, this goes even further and tries to approximate const safety.
//! That's useful because it means other passes (e.g. promotion) can rely on `const`s
//! to be const-safe.
use std::convert::TryFrom;
use std::fmt::Write;
use std::num::NonZeroUsize;
use rustc_data_structures::fx::FxHashSet;
use rustc_hir as hir;
use rustc_middle::mir::interpret::InterpError;
use rustc_middle::ty;
use rustc_middle::ty::layout::{LayoutOf, TyAndLayout};
use rustc_span::symbol::{sym, Symbol};
use rustc_target::abi::{Abi, Scalar as ScalarAbi, Size, VariantIdx, Variants, WrappingRange};
use std::hash::Hash;
use super::{
alloc_range, CheckInAllocMsg, GlobalAlloc, InterpCx, InterpResult, MPlaceTy, Machine,
MemPlaceMeta, OpTy, ScalarMaybeUninit, ValueVisitor,
};
macro_rules! throw_validation_failure {
($where:expr, { $( $what_fmt:expr ),+ } $( expected { $( $expected_fmt:expr ),+ } )?) => {{
let mut msg = String::new();
msg.push_str("encountered ");
write!(&mut msg, $($what_fmt),+).unwrap();
$(
msg.push_str(", but expected ");
write!(&mut msg, $($expected_fmt),+).unwrap();
)?
let path = rustc_middle::ty::print::with_no_trimmed_paths(|| {
let where_ = &$where;
if !where_.is_empty() {
let mut path = String::new();
write_path(&mut path, where_);
Some(path)
} else {
None
}
});
throw_ub!(ValidationFailure { path, msg })
}};
}
/// If $e throws an error matching the pattern, throw a validation failure.
/// Other errors are passed back to the caller, unchanged -- and if they reach the root of
/// the visitor, we make sure only validation errors and `InvalidProgram` errors are left.
/// This lets you use the patterns as a kind of validation list, asserting which errors
/// can possibly happen:
///
/// ```
/// let v = try_validation!(some_fn(), some_path, {
/// Foo | Bar | Baz => { "some failure" },
/// });
/// ```
///
/// An additional expected parameter can also be added to the failure message:
///
/// ```
/// let v = try_validation!(some_fn(), some_path, {
/// Foo | Bar | Baz => { "some failure" } expected { "something that wasn't a failure" },
/// });
/// ```
///
/// An additional nicety is that both parameters actually take format args, so you can just write
/// the format string in directly:
///
/// ```
/// let v = try_validation!(some_fn(), some_path, {
/// Foo | Bar | Baz => { "{:?}", some_failure } expected { "{}", expected_value },
/// });
/// ```
///
macro_rules! try_validation {
($e:expr, $where:expr,
$( $( $p:pat )|+ => { $( $what_fmt:expr ),+ } $( expected { $( $expected_fmt:expr ),+ } )? ),+ $(,)?
) => {{
match $e {
Ok(x) => x,
// We catch the error and turn it into a validation failure. We are okay with
// allocation here as this can only slow down builds that fail anyway.
Err(e) => match e.kind() {
$(
$($p)|+ =>
throw_validation_failure!(
$where,
{ $( $what_fmt ),+ } $( expected { $( $expected_fmt ),+ } )?
)
),+,
#[allow(unreachable_patterns)]
_ => Err::<!, _>(e)?,
}
}
}};
}
/// We want to show a nice path to the invalid field for diagnostics,
/// but avoid string operations in the happy case where no error happens.
/// So we track a `Vec<PathElem>` where `PathElem` contains all the data we
/// need to later print something for the user.
#[derive(Copy, Clone, Debug)]
pub enum PathElem {
Field(Symbol),
Variant(Symbol),
GeneratorState(VariantIdx),
CapturedVar(Symbol),
ArrayElem(usize),
TupleElem(usize),
Deref,
EnumTag,
GeneratorTag,
DynDowncast,
}
/// Extra things to check for during validation of CTFE results.
pub enum CtfeValidationMode {
/// Regular validation, nothing special happening.
Regular,
/// Validation of a `const`.
/// `inner` says if this is an inner, indirect allocation (as opposed to the top-level const
/// allocation). Being an inner allocation makes a difference because the top-level allocation
/// of a `const` is copied for each use, but the inner allocations are implicitly shared.
/// `allow_static_ptrs` says if pointers to statics are permitted (which is the case for promoteds in statics).
Const { inner: bool, allow_static_ptrs: bool },
}
/// State for tracking recursive validation of references
pub struct RefTracking<T, PATH = ()> {
pub seen: FxHashSet<T>,
pub todo: Vec<(T, PATH)>,
}
impl<T: Copy + Eq + Hash + std::fmt::Debug, PATH: Default> RefTracking<T, PATH> {
pub fn empty() -> Self {
RefTracking { seen: FxHashSet::default(), todo: vec![] }
}
pub fn new(op: T) -> Self {
let mut ref_tracking_for_consts =
RefTracking { seen: FxHashSet::default(), todo: vec![(op, PATH::default())] };
ref_tracking_for_consts.seen.insert(op);
ref_tracking_for_consts
}
pub fn track(&mut self, op: T, path: impl FnOnce() -> PATH) {
if self.seen.insert(op) {
trace!("Recursing below ptr {:#?}", op);
let path = path();
// Remember to come back to this later.
self.todo.push((op, path));
}
}
}
/// Format a path
fn write_path(out: &mut String, path: &[PathElem]) {
use self::PathElem::*;
for elem in path.iter() {
match elem {
Field(name) => write!(out, ".{}", name),
EnumTag => write!(out, ".<enum-tag>"),
Variant(name) => write!(out, ".<enum-variant({})>", name),
GeneratorTag => write!(out, ".<generator-tag>"),
GeneratorState(idx) => write!(out, ".<generator-state({})>", idx.index()),
CapturedVar(name) => write!(out, ".<captured-var({})>", name),
TupleElem(idx) => write!(out, ".{}", idx),
ArrayElem(idx) => write!(out, "[{}]", idx),
// `.<deref>` does not match Rust syntax, but it is more readable for long paths -- and
// some of the other items here also are not Rust syntax. Actually we can't
// even use the usual syntax because we are just showing the projections,
// not the root.
Deref => write!(out, ".<deref>"),
DynDowncast => write!(out, ".<dyn-downcast>"),
}
.unwrap()
}
}
// Formats such that a sentence like "expected something {}" to mean
// "expected something <in the given range>" makes sense.
fn wrapping_range_format(r: WrappingRange, max_hi: u128) -> String {
let WrappingRange { start: lo, end: hi } = r;
assert!(hi <= max_hi);
if lo > hi {
format!("less or equal to {}, or greater or equal to {}", hi, lo)
} else if lo == hi {
format!("equal to {}", lo)
} else if lo == 0 {
assert!(hi < max_hi, "should not be printing if the range covers everything");
format!("less or equal to {}", hi)
} else if hi == max_hi {
assert!(lo > 0, "should not be printing if the range covers everything");
format!("greater or equal to {}", lo)
} else {
format!("in the range {:?}", r)
}
}
struct ValidityVisitor<'rt, 'mir, 'tcx, M: Machine<'mir, 'tcx>> {
/// The `path` may be pushed to, but the part that is present when a function
/// starts must not be changed! `visit_fields` and `visit_array` rely on
/// this stack discipline.
path: Vec<PathElem>,
ref_tracking: Option<&'rt mut RefTracking<MPlaceTy<'tcx, M::PointerTag>, Vec<PathElem>>>,
/// `None` indicates this is not validating for CTFE (but for runtime).
ctfe_mode: Option<CtfeValidationMode>,
ecx: &'rt InterpCx<'mir, 'tcx, M>,
}
impl<'rt, 'mir, 'tcx: 'mir, M: Machine<'mir, 'tcx>> ValidityVisitor<'rt, 'mir, 'tcx, M> {
fn aggregate_field_path_elem(&mut self, layout: TyAndLayout<'tcx>, field: usize) -> PathElem {
// First, check if we are projecting to a variant.
match layout.variants {
Variants::Multiple { tag_field, .. } => {
if tag_field == field {
return match layout.ty.kind() {
ty::Adt(def, ..) if def.is_enum() => PathElem::EnumTag,
ty::Generator(..) => PathElem::GeneratorTag,
_ => bug!("non-variant type {:?}", layout.ty),
};
}
}
Variants::Single { .. } => {}
}
// Now we know we are projecting to a field, so figure out which one.
match layout.ty.kind() {
// generators and closures.
ty::Closure(def_id, _) | ty::Generator(def_id, _, _) => {
let mut name = None;
// FIXME this should be more descriptive i.e. CapturePlace instead of CapturedVar
// https://github.com/rust-lang/project-rfc-2229/issues/46
if let Some(local_def_id) = def_id.as_local() {
let tables = self.ecx.tcx.typeck(local_def_id);
if let Some(captured_place) =
tables.closure_min_captures_flattened(*def_id).nth(field)
{
// Sometimes the index is beyond the number of upvars (seen
// for a generator).
let var_hir_id = captured_place.get_root_variable();
let node = self.ecx.tcx.hir().get(var_hir_id);
if let hir::Node::Binding(pat) = node {
if let hir::PatKind::Binding(_, _, ident, _) = pat.kind {
name = Some(ident.name);
}
}
}
}
PathElem::CapturedVar(name.unwrap_or_else(|| {
// Fall back to showing the field index.
sym::integer(field)
}))
}
// tuples
ty::Tuple(_) => PathElem::TupleElem(field),
// enums
ty::Adt(def, ..) if def.is_enum() => {
// we might be projecting *to* a variant, or to a field *in* a variant.
match layout.variants {
Variants::Single { index } => {
// Inside a variant
PathElem::Field(def.variants[index].fields[field].ident.name)
}
Variants::Multiple { .. } => bug!("we handled variants above"),
}
}
// other ADTs
ty::Adt(def, _) => PathElem::Field(def.non_enum_variant().fields[field].ident.name),
// arrays/slices
ty::Array(..) | ty::Slice(..) => PathElem::ArrayElem(field),
// dyn traits
ty::Dynamic(..) => PathElem::DynDowncast,
// nothing else has an aggregate layout
_ => bug!("aggregate_field_path_elem: got non-aggregate type {:?}", layout.ty),
}
}
fn with_elem<R>(
&mut self,
elem: PathElem,
f: impl FnOnce(&mut Self) -> InterpResult<'tcx, R>,
) -> InterpResult<'tcx, R> {
// Remember the old state
let path_len = self.path.len();
// Record new element
self.path.push(elem);
// Perform operation
let r = f(self)?;
// Undo changes
self.path.truncate(path_len);
// Done
Ok(r)
}
fn check_wide_ptr_meta(
&mut self,
meta: MemPlaceMeta<M::PointerTag>,
pointee: TyAndLayout<'tcx>,
) -> InterpResult<'tcx> {
let tail = self.ecx.tcx.struct_tail_erasing_lifetimes(pointee.ty, self.ecx.param_env);
match tail.kind() {
ty::Dynamic(..) => {
let vtable = self.ecx.scalar_to_ptr(meta.unwrap_meta());
// Direct call to `check_ptr_access_align` checks alignment even on CTFE machines.
try_validation!(
self.ecx.memory.check_ptr_access_align(
vtable,
3 * self.ecx.tcx.data_layout.pointer_size, // drop, size, align
self.ecx.tcx.data_layout.pointer_align.abi,
CheckInAllocMsg::InboundsTest, // will anyway be replaced by validity message
),
self.path,
err_ub!(DanglingIntPointer(..)) |
err_ub!(PointerUseAfterFree(..)) =>
{ "dangling vtable pointer in wide pointer" },
err_ub!(AlignmentCheckFailed { .. }) =>
{ "unaligned vtable pointer in wide pointer" },
err_ub!(PointerOutOfBounds { .. }) =>
{ "too small vtable" },
);
try_validation!(
self.ecx.read_drop_type_from_vtable(vtable),
self.path,
err_ub!(DanglingIntPointer(..)) |
err_ub!(InvalidFunctionPointer(..)) =>
{ "invalid drop function pointer in vtable (not pointing to a function)" },
err_ub!(InvalidVtableDropFn(..)) =>
{ "invalid drop function pointer in vtable (function has incompatible signature)" },
);
try_validation!(
self.ecx.read_size_and_align_from_vtable(vtable),
self.path,
err_ub!(InvalidVtableSize) =>
{ "invalid vtable: size is bigger than largest supported object" },
err_ub!(InvalidVtableAlignment(msg)) =>
{ "invalid vtable: alignment {}", msg },
err_unsup!(ReadPointerAsBytes) => { "invalid size or align in vtable" },
);
// FIXME: More checks for the vtable.
}
ty::Slice(..) | ty::Str => {
let _len = try_validation!(
meta.unwrap_meta().to_machine_usize(self.ecx),
self.path,
err_unsup!(ReadPointerAsBytes) => { "non-integer slice length in wide pointer" },
);
// We do not check that `len * elem_size <= isize::MAX`:
// that is only required for references, and there it falls out of the
// "dereferenceable" check performed by Stacked Borrows.
}
ty::Foreign(..) => {
// Unsized, but not wide.
}
_ => bug!("Unexpected unsized type tail: {:?}", tail),
}
Ok(())
}
/// Check a reference or `Box`.
fn check_safe_pointer(
&mut self,
value: &OpTy<'tcx, M::PointerTag>,
kind: &str,
) -> InterpResult<'tcx> {
let value = try_validation!(
self.ecx.read_immediate(value),
self.path,
err_unsup!(ReadPointerAsBytes) => { "part of a pointer" } expected { "a proper pointer or integer value" },
);
// Handle wide pointers.
// Check metadata early, for better diagnostics
let place = try_validation!(
self.ecx.ref_to_mplace(&value),
self.path,
err_ub!(InvalidUninitBytes(None)) => { "uninitialized {}", kind },
);
if place.layout.is_unsized() {
self.check_wide_ptr_meta(place.meta, place.layout)?;
}
// Make sure this is dereferenceable and all.
let size_and_align = try_validation!(
self.ecx.size_and_align_of_mplace(&place),
self.path,
err_ub!(InvalidMeta(msg)) => { "invalid {} metadata: {}", kind, msg },
);
let (size, align) = size_and_align
// for the purpose of validity, consider foreign types to have
// alignment and size determined by the layout (size will be 0,
// alignment should take attributes into account).
.unwrap_or_else(|| (place.layout.size, place.layout.align.abi));
// Direct call to `check_ptr_access_align` checks alignment even on CTFE machines.
try_validation!(
self.ecx.memory.check_ptr_access_align(
place.ptr,
size,
align,
CheckInAllocMsg::InboundsTest, // will anyway be replaced by validity message
),
self.path,
err_ub!(AlignmentCheckFailed { required, has }) =>
{
"an unaligned {} (required {} byte alignment but found {})",
kind,
required.bytes(),
has.bytes()
},
err_ub!(DanglingIntPointer(0, _)) =>
{ "a null {}", kind },
err_ub!(DanglingIntPointer(i, _)) =>
{ "a dangling {} (address 0x{:x} is unallocated)", kind, i },
err_ub!(PointerOutOfBounds { .. }) =>
{ "a dangling {} (going beyond the bounds of its allocation)", kind },
// This cannot happen during const-eval (because interning already detects
// dangling pointers), but it can happen in Miri.
err_ub!(PointerUseAfterFree(..)) =>
{ "a dangling {} (use-after-free)", kind },
);
// Recursive checking
if let Some(ref mut ref_tracking) = self.ref_tracking {
// Proceed recursively even for ZST, no reason to skip them!
// `!` is a ZST and we want to validate it.
// Skip validation entirely for some external statics
if let Ok((alloc_id, _offset, _ptr)) = self.ecx.memory.ptr_try_get_alloc(place.ptr) {
// not a ZST
let alloc_kind = self.ecx.tcx.get_global_alloc(alloc_id);
if let Some(GlobalAlloc::Static(did)) = alloc_kind {
assert!(!self.ecx.tcx.is_thread_local_static(did));
assert!(self.ecx.tcx.is_static(did));
if matches!(
self.ctfe_mode,
Some(CtfeValidationMode::Const { allow_static_ptrs: false, .. })
) {
// See const_eval::machine::MemoryExtra::can_access_statics for why
// this check is so important.
// This check is reachable when the const just referenced the static,
// but never read it (so we never entered `before_access_global`).
throw_validation_failure!(self.path,
{ "a {} pointing to a static variable", kind }
);
}
// We skip checking other statics. These statics must be sound by
// themselves, and the only way to get broken statics here is by using
// unsafe code.
// The reasons we don't check other statics is twofold. For one, in all
// sound cases, the static was already validated on its own, and second, we
// trigger cycle errors if we try to compute the value of the other static
// and that static refers back to us.
// We might miss const-invalid data,
// but things are still sound otherwise (in particular re: consts
// referring to statics).
return Ok(());
}
}
let path = &self.path;
ref_tracking.track(place, || {
// We need to clone the path anyway, make sure it gets created
// with enough space for the additional `Deref`.
let mut new_path = Vec::with_capacity(path.len() + 1);
new_path.clone_from(path);
new_path.push(PathElem::Deref);
new_path
});
}
Ok(())
}
fn read_scalar(
&self,
op: &OpTy<'tcx, M::PointerTag>,
) -> InterpResult<'tcx, ScalarMaybeUninit<M::PointerTag>> {
Ok(try_validation!(
self.ecx.read_scalar(op),
self.path,
err_unsup!(ReadPointerAsBytes) => { "(potentially part of) a pointer" } expected { "plain (non-pointer) bytes" },
))
}
/// Check if this is a value of primitive type, and if yes check the validity of the value
/// at that type. Return `true` if the type is indeed primitive.
fn try_visit_primitive(
&mut self,
value: &OpTy<'tcx, M::PointerTag>,
) -> InterpResult<'tcx, bool> {
// Go over all the primitive types
let ty = value.layout.ty;
match ty.kind() {
ty::Bool => {
let value = self.read_scalar(value)?;
try_validation!(
value.to_bool(),
self.path,
err_ub!(InvalidBool(..)) | err_ub!(InvalidUninitBytes(None)) =>
{ "{}", value } expected { "a boolean" },
);
Ok(true)
}
ty::Char => {
let value = self.read_scalar(value)?;
try_validation!(
value.to_char(),
self.path,
err_ub!(InvalidChar(..)) | err_ub!(InvalidUninitBytes(None)) =>
{ "{}", value } expected { "a valid unicode scalar value (in `0..=0x10FFFF` but not in `0xD800..=0xDFFF`)" },
);
Ok(true)
}
ty::Float(_) | ty::Int(_) | ty::Uint(_) => {
let value = self.read_scalar(value)?;
// NOTE: Keep this in sync with the array optimization for int/float
// types below!
if self.ctfe_mode.is_some() {
// Integers/floats in CTFE: Must be scalar bits, pointers are dangerous
let is_bits = value.check_init().map_or(false, |v| v.try_to_int().is_ok());
if !is_bits {
throw_validation_failure!(self.path,
{ "{}", value } expected { "initialized plain (non-pointer) bytes" }
)
}
} else {
// At run-time, for now, we accept *anything* for these types, including
// uninit. We should fix that, but let's start low.
}
Ok(true)
}
ty::RawPtr(..) => {
// We are conservative with uninit for integers, but try to
// actually enforce the strict rules for raw pointers (mostly because
// that lets us re-use `ref_to_mplace`).
let place = try_validation!(
self.ecx.read_immediate(value).and_then(|ref i| self.ecx.ref_to_mplace(i)),
self.path,
err_ub!(InvalidUninitBytes(None)) => { "uninitialized raw pointer" },
err_unsup!(ReadPointerAsBytes) => { "part of a pointer" } expected { "a proper pointer or integer value" },
);
if place.layout.is_unsized() {
self.check_wide_ptr_meta(place.meta, place.layout)?;
}
Ok(true)
}
ty::Ref(_, ty, mutbl) => {
if matches!(self.ctfe_mode, Some(CtfeValidationMode::Const { .. }))
&& *mutbl == hir::Mutability::Mut
{
// A mutable reference inside a const? That does not seem right (except if it is
// a ZST).
let layout = self.ecx.layout_of(ty)?;
if !layout.is_zst() {
throw_validation_failure!(self.path, { "mutable reference in a `const`" });
}
}
self.check_safe_pointer(value, "reference")?;
Ok(true)
}
ty::Adt(def, ..) if def.is_box() => {
self.check_safe_pointer(value, "box")?;
Ok(true)
}
ty::FnPtr(_sig) => {
let value = try_validation!(
self.ecx.read_immediate(value),
self.path,
err_unsup!(ReadPointerAsBytes) => { "part of a pointer" } expected { "a proper pointer or integer value" },
);
// Make sure we print a `ScalarMaybeUninit` (and not an `ImmTy`) in the error
// message below.
let value = value.to_scalar_or_uninit();
let _fn = try_validation!(
value.check_init().and_then(|ptr| self.ecx.memory.get_fn(self.ecx.scalar_to_ptr(ptr))),
self.path,
err_ub!(DanglingIntPointer(..)) |
err_ub!(InvalidFunctionPointer(..)) |
err_ub!(InvalidUninitBytes(None)) =>
{ "{}", value } expected { "a function pointer" },
);
// FIXME: Check if the signature matches
Ok(true)
}
ty::Never => throw_validation_failure!(self.path, { "a value of the never type `!`" }),
ty::Foreign(..) | ty::FnDef(..) => {
// Nothing to check.
Ok(true)
}
// The above should be all the primitive types. The rest is compound, we
// check them by visiting their fields/variants.
ty::Adt(..)
| ty::Tuple(..)
| ty::Array(..)
| ty::Slice(..)
| ty::Str
| ty::Dynamic(..)
| ty::Closure(..)
| ty::Generator(..) => Ok(false),
// Some types only occur during typechecking, they have no layout.
// We should not see them here and we could not check them anyway.
ty::Error(_)
| ty::Infer(..)
| ty::Placeholder(..)
| ty::Bound(..)
| ty::Param(..)
| ty::Opaque(..)
| ty::Projection(..)
| ty::GeneratorWitness(..) => bug!("Encountered invalid type {:?}", ty),
}
}
fn visit_scalar(
&mut self,
op: &OpTy<'tcx, M::PointerTag>,
scalar_layout: &ScalarAbi,
) -> InterpResult<'tcx> {
let value = self.read_scalar(op)?;
let valid_range = scalar_layout.valid_range.clone();
let WrappingRange { start: lo, end: hi } = valid_range;
// Determine the allowed range
// `max_hi` is as big as the size fits
let max_hi = u128::MAX >> (128 - op.layout.size.bits());
assert!(hi <= max_hi);
// We could also write `(hi + 1) % (max_hi + 1) == lo` but `max_hi + 1` overflows for `u128`
if (lo == 0 && hi == max_hi) || (hi + 1 == lo) {
// Nothing to check
return Ok(());
}
// At least one value is excluded. Get the bits.
let value = try_validation!(
value.check_init(),
self.path,
err_ub!(InvalidUninitBytes(None)) => { "{}", value }
expected { "something {}", wrapping_range_format(valid_range, max_hi) },
);
let bits = match value.try_to_int() {
Err(_) => {
// So this is a pointer then, and casting to an int failed.
// Can only happen during CTFE.
let ptr = self.ecx.scalar_to_ptr(value);
if lo == 1 && hi == max_hi {
// Only null is the niche. So make sure the ptr is NOT null.
if self.ecx.memory.ptr_may_be_null(ptr) {
throw_validation_failure!(self.path,
{ "a potentially null pointer" }
expected {
"something that cannot possibly fail to be {}",
wrapping_range_format(valid_range, max_hi)
}
)
}
return Ok(());
} else {
// Conservatively, we reject, because the pointer *could* have a bad
// value.
throw_validation_failure!(self.path,
{ "a pointer" }
expected {
"something that cannot possibly fail to be {}",
wrapping_range_format(valid_range, max_hi)
}
)
}
}
Ok(int) => int.assert_bits(op.layout.size),
};
// Now compare. This is slightly subtle because this is a special "wrap-around" range.
if valid_range.contains(bits) {
Ok(())
} else {
throw_validation_failure!(self.path,
{ "{}", bits }
expected { "something {}", wrapping_range_format(valid_range, max_hi) }
)
}
}
}
impl<'rt, 'mir, 'tcx: 'mir, M: Machine<'mir, 'tcx>> ValueVisitor<'mir, 'tcx, M>
for ValidityVisitor<'rt, 'mir, 'tcx, M>
{
type V = OpTy<'tcx, M::PointerTag>;
#[inline(always)]
fn ecx(&self) -> &InterpCx<'mir, 'tcx, M> {
&self.ecx
}
fn read_discriminant(
&mut self,
op: &OpTy<'tcx, M::PointerTag>,
) -> InterpResult<'tcx, VariantIdx> {
self.with_elem(PathElem::EnumTag, move |this| {
Ok(try_validation!(
this.ecx.read_discriminant(op),
this.path,
err_ub!(InvalidTag(val)) =>
{ "{}", val } expected { "a valid enum tag" },
err_ub!(InvalidUninitBytes(None)) =>
{ "uninitialized bytes" } expected { "a valid enum tag" },
err_unsup!(ReadPointerAsBytes) =>
{ "a pointer" } expected { "a valid enum tag" },
)
.1)
})
}
#[inline]
fn visit_field(
&mut self,
old_op: &OpTy<'tcx, M::PointerTag>,
field: usize,
new_op: &OpTy<'tcx, M::PointerTag>,
) -> InterpResult<'tcx> {
let elem = self.aggregate_field_path_elem(old_op.layout, field);
self.with_elem(elem, move |this| this.visit_value(new_op))
}
#[inline]
fn visit_variant(
&mut self,
old_op: &OpTy<'tcx, M::PointerTag>,
variant_id: VariantIdx,
new_op: &OpTy<'tcx, M::PointerTag>,
) -> InterpResult<'tcx> {
let name = match old_op.layout.ty.kind() {
ty::Adt(adt, _) => PathElem::Variant(adt.variants[variant_id].ident.name),
// Generators also have variants
ty::Generator(..) => PathElem::GeneratorState(variant_id),
_ => bug!("Unexpected type with variant: {:?}", old_op.layout.ty),
};
self.with_elem(name, move |this| this.visit_value(new_op))
}
#[inline(always)]
fn visit_union(
&mut self,
_op: &OpTy<'tcx, M::PointerTag>,
_fields: NonZeroUsize,
) -> InterpResult<'tcx> {
Ok(())
}
#[inline]
fn visit_value(&mut self, op: &OpTy<'tcx, M::PointerTag>) -> InterpResult<'tcx> {
trace!("visit_value: {:?}, {:?}", *op, op.layout);
// Check primitive types -- the leafs of our recursive descend.
if self.try_visit_primitive(op)? {
return Ok(());
}
// Sanity check: `builtin_deref` does not know any pointers that are not primitive.
assert!(op.layout.ty.builtin_deref(true).is_none());
// Special check preventing `UnsafeCell` in the inner part of constants
if let Some(def) = op.layout.ty.ty_adt_def() {
if matches!(self.ctfe_mode, Some(CtfeValidationMode::Const { inner: true, .. }))
&& Some(def.did) == self.ecx.tcx.lang_items().unsafe_cell_type()
{
throw_validation_failure!(self.path, { "`UnsafeCell` in a `const`" });
}
}
// Recursively walk the value at its type.
self.walk_value(op)?;
// *After* all of this, check the ABI. We need to check the ABI to handle
// types like `NonNull` where the `Scalar` info is more restrictive than what
// the fields say (`rustc_layout_scalar_valid_range_start`).
// But in most cases, this will just propagate what the fields say,
// and then we want the error to point at the field -- so, first recurse,
// then check ABI.
//
// FIXME: We could avoid some redundant checks here. For newtypes wrapping
// scalars, we do the same check on every "level" (e.g., first we check
// MyNewtype and then the scalar in there).
match op.layout.abi {
Abi::Uninhabited => {
throw_validation_failure!(self.path,
{ "a value of uninhabited type {:?}", op.layout.ty }
);
}
Abi::Scalar(ref scalar_layout) => {
self.visit_scalar(op, scalar_layout)?;
}
Abi::ScalarPair { .. } | Abi::Vector { .. } => {
// These have fields that we already visited above, so we already checked
// all their scalar-level restrictions.
// There is also no equivalent to `rustc_layout_scalar_valid_range_start`
// that would make skipping them here an issue.
}
Abi::Aggregate { .. } => {
// Nothing to do.
}
}
Ok(())
}
fn visit_aggregate(
&mut self,
op: &OpTy<'tcx, M::PointerTag>,
fields: impl Iterator<Item = InterpResult<'tcx, Self::V>>,
) -> InterpResult<'tcx> {
match op.layout.ty.kind() {
ty::Str => {
let mplace = op.assert_mem_place(); // strings are never immediate
let len = mplace.len(self.ecx)?;
try_validation!(
self.ecx.memory.read_bytes(mplace.ptr, Size::from_bytes(len)),
self.path,
err_ub!(InvalidUninitBytes(..)) => { "uninitialized data in `str`" },
err_unsup!(ReadPointerAsBytes) => { "a pointer in `str`" },
);
}
ty::Array(tys, ..) | ty::Slice(tys)
// This optimization applies for types that can hold arbitrary bytes (such as
// integer and floating point types) or for structs or tuples with no fields.
// FIXME(wesleywiser) This logic could be extended further to arbitrary structs
// or tuples made up of integer/floating point types or inhabited ZSTs with no
// padding.
if matches!(tys.kind(), ty::Int(..) | ty::Uint(..) | ty::Float(..))
=>
{
// Optimized handling for arrays of integer/float type.
// Arrays cannot be immediate, slices are never immediate.
let mplace = op.assert_mem_place();
// This is the length of the array/slice.
let len = mplace.len(self.ecx)?;
// This is the element type size.
let layout = self.ecx.layout_of(tys)?;
// This is the size in bytes of the whole array. (This checks for overflow.)
let size = layout.size * len;
// Optimization: we just check the entire range at once.
// NOTE: Keep this in sync with the handling of integer and float
// types above, in `visit_primitive`.
// In run-time mode, we accept pointers in here. This is actually more
// permissive than a per-element check would be, e.g., we accept
// a &[u8] that contains a pointer even though bytewise checking would
// reject it. However, that's good: We don't inherently want
// to reject those pointers, we just do not have the machinery to
// talk about parts of a pointer.
// We also accept uninit, for consistency with the slow path.
let alloc = match self.ecx.memory.get(mplace.ptr, size, mplace.align)? {
Some(a) => a,
None => {
// Size 0, nothing more to check.
return Ok(());
}
};
match alloc.check_bytes(
alloc_range(Size::ZERO, size),
/*allow_uninit_and_ptr*/ self.ctfe_mode.is_none(),
) {
// In the happy case, we needn't check anything else.
Ok(()) => {}
// Some error happened, try to provide a more detailed description.
Err(err) => {
// For some errors we might be able to provide extra information.
// (This custom logic does not fit the `try_validation!` macro.)
match err.kind() {
err_ub!(InvalidUninitBytes(Some((_alloc_id, access)))) => {
// Some byte was uninitialized, determine which
// element that byte belongs to so we can
// provide an index.
let i = usize::try_from(
access.uninit_offset.bytes() / layout.size.bytes(),
)
.unwrap();
self.path.push(PathElem::ArrayElem(i));
throw_validation_failure!(self.path, { "uninitialized bytes" })
}
err_unsup!(ReadPointerAsBytes) => {
throw_validation_failure!(self.path, { "a pointer" } expected { "plain (non-pointer) bytes" })
}
// Propagate upwards (that will also check for unexpected errors).
_ => return Err(err),
}
}
}
}
// Fast path for arrays and slices of ZSTs. We only need to check a single ZST element
// of an array and not all of them, because there's only a single value of a specific
// ZST type, so either validation fails for all elements or none.
ty::Array(tys, ..) | ty::Slice(tys) if self.ecx.layout_of(tys)?.is_zst() => {
// Validate just the first element (if any).
self.walk_aggregate(op, fields.take(1))?
}
_ => {
self.walk_aggregate(op, fields)? // default handler
}
}
Ok(())
}
}
impl<'mir, 'tcx: 'mir, M: Machine<'mir, 'tcx>> InterpCx<'mir, 'tcx, M> {
fn validate_operand_internal(
&self,
op: &OpTy<'tcx, M::PointerTag>,
path: Vec<PathElem>,
ref_tracking: Option<&mut RefTracking<MPlaceTy<'tcx, M::PointerTag>, Vec<PathElem>>>,
ctfe_mode: Option<CtfeValidationMode>,
) -> InterpResult<'tcx> {
trace!("validate_operand_internal: {:?}, {:?}", *op, op.layout.ty);
// Construct a visitor
let mut visitor = ValidityVisitor { path, ref_tracking, ctfe_mode, ecx: self };
// Run it.
match visitor.visit_value(&op) {
Ok(()) => Ok(()),
// Pass through validation failures.
Err(err) if matches!(err.kind(), err_ub!(ValidationFailure { .. })) => Err(err),
// Also pass through InvalidProgram, those just indicate that we could not
// validate and each caller will know best what to do with them.
Err(err) if matches!(err.kind(), InterpError::InvalidProgram(_)) => Err(err),
// Avoid other errors as those do not show *where* in the value the issue lies.
Err(err) => {
err.print_backtrace();
bug!("Unexpected error during validation: {}", err);
}
}
}
/// This function checks the data at `op` to be const-valid.
/// `op` is assumed to cover valid memory if it is an indirect operand.
/// It will error if the bits at the destination do not match the ones described by the layout.
///
/// `ref_tracking` is used to record references that we encounter so that they
/// can be checked recursively by an outside driving loop.
///
/// `constant` controls whether this must satisfy the rules for constants:
/// - no pointers to statics.
/// - no `UnsafeCell` or non-ZST `&mut`.
#[inline(always)]
pub fn const_validate_operand(
&self,
op: &OpTy<'tcx, M::PointerTag>,
path: Vec<PathElem>,
ref_tracking: &mut RefTracking<MPlaceTy<'tcx, M::PointerTag>, Vec<PathElem>>,
ctfe_mode: CtfeValidationMode,
) -> InterpResult<'tcx> {
self.validate_operand_internal(op, path, Some(ref_tracking), Some(ctfe_mode))
}
/// This function checks the data at `op` to be runtime-valid.
/// `op` is assumed to cover valid memory if it is an indirect operand.
/// It will error if the bits at the destination do not match the ones described by the layout.
#[inline(always)]
pub fn validate_operand(&self, op: &OpTy<'tcx, M::PointerTag>) -> InterpResult<'tcx> {
self.validate_operand_internal(op, vec![], None, None)
}
}

278
compiler/rustc_const_eval/src/interpret/visitor.rs Normal file
View File

@@ -0,0 +1,278 @@
//! Visitor for a run-time value with a given layout: Traverse enums, structs and other compound
//! types until we arrive at the leaves, with custom handling for primitive types.
use rustc_middle::mir::interpret::InterpResult;
use rustc_middle::ty;
use rustc_middle::ty::layout::TyAndLayout;
use rustc_target::abi::{FieldsShape, VariantIdx, Variants};
use std::num::NonZeroUsize;
use super::{InterpCx, MPlaceTy, Machine, OpTy};
// A thing that we can project into, and that has a layout.
// This wouldn't have to depend on `Machine` but with the current type inference,
// that's just more convenient to work with (avoids repeating all the `Machine` bounds).
pub trait Value<'mir, 'tcx, M: Machine<'mir, 'tcx>>: Copy {
/// Gets this value's layout.
fn layout(&self) -> TyAndLayout<'tcx>;
/// Makes this into an `OpTy`.
fn to_op(&self, ecx: &InterpCx<'mir, 'tcx, M>)
-> InterpResult<'tcx, OpTy<'tcx, M::PointerTag>>;
/// Creates this from an `MPlaceTy`.
fn from_mem_place(mplace: MPlaceTy<'tcx, M::PointerTag>) -> Self;
/// Projects to the given enum variant.
fn project_downcast(
&self,
ecx: &InterpCx<'mir, 'tcx, M>,
variant: VariantIdx,
) -> InterpResult<'tcx, Self>;
/// Projects to the n-th field.
fn project_field(
&self,
ecx: &InterpCx<'mir, 'tcx, M>,
field: usize,
) -> InterpResult<'tcx, Self>;
}
// Operands and memory-places are both values.
// Places in general are not due to `place_field` having to do `force_allocation`.
impl<'mir, 'tcx: 'mir, M: Machine<'mir, 'tcx>> Value<'mir, 'tcx, M> for OpTy<'tcx, M::PointerTag> {
#[inline(always)]
fn layout(&self) -> TyAndLayout<'tcx> {
self.layout
}
#[inline(always)]
fn to_op(
&self,
_ecx: &InterpCx<'mir, 'tcx, M>,
) -> InterpResult<'tcx, OpTy<'tcx, M::PointerTag>> {
Ok(*self)
}
#[inline(always)]
fn from_mem_place(mplace: MPlaceTy<'tcx, M::PointerTag>) -> Self {
mplace.into()
}
#[inline(always)]
fn project_downcast(
&self,
ecx: &InterpCx<'mir, 'tcx, M>,
variant: VariantIdx,
) -> InterpResult<'tcx, Self> {
ecx.operand_downcast(self, variant)
}
#[inline(always)]
fn project_field(
&self,
ecx: &InterpCx<'mir, 'tcx, M>,
field: usize,
) -> InterpResult<'tcx, Self> {
ecx.operand_field(self, field)
}
}
impl<'mir, 'tcx: 'mir, M: Machine<'mir, 'tcx>> Value<'mir, 'tcx, M>
for MPlaceTy<'tcx, M::PointerTag>
{
#[inline(always)]
fn layout(&self) -> TyAndLayout<'tcx> {
self.layout
}
#[inline(always)]
fn to_op(
&self,
_ecx: &InterpCx<'mir, 'tcx, M>,
) -> InterpResult<'tcx, OpTy<'tcx, M::PointerTag>> {
Ok((*self).into())
}
#[inline(always)]
fn from_mem_place(mplace: MPlaceTy<'tcx, M::PointerTag>) -> Self {
mplace
}
#[inline(always)]
fn project_downcast(
&self,
ecx: &InterpCx<'mir, 'tcx, M>,
variant: VariantIdx,
) -> InterpResult<'tcx, Self> {
ecx.mplace_downcast(self, variant)
}
#[inline(always)]
fn project_field(
&self,
ecx: &InterpCx<'mir, 'tcx, M>,
field: usize,
) -> InterpResult<'tcx, Self> {
ecx.mplace_field(self, field)
}
}
macro_rules! make_value_visitor {
($visitor_trait_name:ident, $($mutability:ident)?) => {
// How to traverse a value and what to do when we are at the leaves.
pub trait $visitor_trait_name<'mir, 'tcx: 'mir, M: Machine<'mir, 'tcx>>: Sized {
type V: Value<'mir, 'tcx, M>;
/// The visitor must have an `InterpCx` in it.
fn ecx(&$($mutability)? self)
-> &$($mutability)? InterpCx<'mir, 'tcx, M>;
/// `read_discriminant` can be hooked for better error messages.
#[inline(always)]
fn read_discriminant(
&mut self,
op: &OpTy<'tcx, M::PointerTag>,
) -> InterpResult<'tcx, VariantIdx> {
Ok(self.ecx().read_discriminant(op)?.1)
}
// Recursive actions, ready to be overloaded.
/// Visits the given value, dispatching as appropriate to more specialized visitors.
#[inline(always)]
fn visit_value(&mut self, v: &Self::V) -> InterpResult<'tcx>
{
self.walk_value(v)
}
/// Visits the given value as a union. No automatic recursion can happen here.
#[inline(always)]
fn visit_union(&mut self, _v: &Self::V, _fields: NonZeroUsize) -> InterpResult<'tcx>
{
Ok(())
}
/// Visits this value as an aggregate, you are getting an iterator yielding
/// all the fields (still in an `InterpResult`, you have to do error handling yourself).
/// Recurses into the fields.
#[inline(always)]
fn visit_aggregate(
&mut self,
v: &Self::V,
fields: impl Iterator<Item=InterpResult<'tcx, Self::V>>,
) -> InterpResult<'tcx> {
self.walk_aggregate(v, fields)
}
/// Called each time we recurse down to a field of a "product-like" aggregate
/// (structs, tuples, arrays and the like, but not enums), passing in old (outer)
/// and new (inner) value.
/// This gives the visitor the chance to track the stack of nested fields that
/// we are descending through.
#[inline(always)]
fn visit_field(
&mut self,
_old_val: &Self::V,
_field: usize,
new_val: &Self::V,
) -> InterpResult<'tcx> {
self.visit_value(new_val)
}
/// Called when recursing into an enum variant.
/// This gives the visitor the chance to track the stack of nested fields that
/// we are descending through.
#[inline(always)]
fn visit_variant(
&mut self,
_old_val: &Self::V,
_variant: VariantIdx,
new_val: &Self::V,
) -> InterpResult<'tcx> {
self.visit_value(new_val)
}
// Default recursors. Not meant to be overloaded.
fn walk_aggregate(
&mut self,
v: &Self::V,
fields: impl Iterator<Item=InterpResult<'tcx, Self::V>>,
) -> InterpResult<'tcx> {
// Now iterate over it.
for (idx, field_val) in fields.enumerate() {
self.visit_field(v, idx, &field_val?)?;
}
Ok(())
}
fn walk_value(&mut self, v: &Self::V) -> InterpResult<'tcx>
{
trace!("walk_value: type: {}", v.layout().ty);
// Special treatment for special types, where the (static) layout is not sufficient.
match *v.layout().ty.kind() {
// If it is a trait object, switch to the real type that was used to create it.
ty::Dynamic(..) => {
// immediate trait objects are not a thing
let op = v.to_op(self.ecx())?;
let dest = op.assert_mem_place();
let inner = self.ecx().unpack_dyn_trait(&dest)?.1;
trace!("walk_value: dyn object layout: {:#?}", inner.layout);
// recurse with the inner type
return self.visit_field(&v, 0, &Value::from_mem_place(inner));
},
// Slices do not need special handling here: they have `Array` field
// placement with length 0, so we enter the `Array` case below which
// indirectly uses the metadata to determine the actual length.
_ => {},
};
// Visit the fields of this value.
match v.layout().fields {
FieldsShape::Primitive => {},
FieldsShape::Union(fields) => {
self.visit_union(v, fields)?;
},
FieldsShape::Arbitrary { ref offsets, .. } => {
// FIXME: We collect in a vec because otherwise there are lifetime
// errors: Projecting to a field needs access to `ecx`.
let fields: Vec<InterpResult<'tcx, Self::V>> =
(0..offsets.len()).map(|i| {
v.project_field(self.ecx(), i)
})
.collect();
self.visit_aggregate(v, fields.into_iter())?;
},
FieldsShape::Array { .. } => {
// Let's get an mplace first.
let op = v.to_op(self.ecx())?;
let mplace = op.assert_mem_place();
// Now we can go over all the fields.
// This uses the *run-time length*, i.e., if we are a slice,
// the dynamic info from the metadata is used.
let iter = self.ecx().mplace_array_fields(&mplace)?
.map(|f| f.and_then(|f| {
Ok(Value::from_mem_place(f))
}));
self.visit_aggregate(v, iter)?;
}
}
match v.layout().variants {
// If this is a multi-variant layout, find the right variant and proceed
// with *its* fields.
Variants::Multiple { .. } => {
let op = v.to_op(self.ecx())?;
let idx = self.read_discriminant(&op)?;
let inner = v.project_downcast(self.ecx(), idx)?;
trace!("walk_value: variant layout: {:#?}", inner.layout());
// recurse with the inner type
self.visit_variant(v, idx, &inner)
}
// For single-variant layouts, we already did anything there is to do.
Variants::Single { .. } => Ok(())
}
}
}
}
}
make_value_visitor!(ValueVisitor,);
make_value_visitor!(MutValueVisitor, mut);