rustc_const_eval/interpret/

intrinsics.rs

//! Intrinsics and other functions that the interpreter executes without
//! looking at their MIR. Intrinsics/functions supported here are shared by CTFE
//! and miri.

mod atomic;
mod simd;

use std::assert_matches;

use rustc_abi::{FieldIdx, HasDataLayout, Size, VariantIdx};
use rustc_apfloat::ieee::{Double, Half, Quad, Single};
use rustc_ast::{IntTy, UintTy};
use rustc_middle::mir::interpret::{CTFE_ALLOC_SALT, read_target_uint, write_target_uint};
use rustc_middle::mir::{self, BinOp, ConstValue, NonDivergingIntrinsic};
use rustc_middle::ty::layout::TyAndLayout;
use rustc_middle::ty::{self, FloatTy, ScalarInt, Ty, TyCtxt, TypeVisitableExt};
use rustc_span::{Symbol, bug, kw, span_bug, sym};
use tracing::trace;

use super::memory::MemoryKind;
use super::util::ensure_monomorphic_enough;
use super::{
    AllocId, AtomicRmwOp, CheckInAllocMsg, ImmTy, Immediate, InterpCx, InterpResult, Machine, OpTy,
    PlaceTy, Pointer, PointerArithmetic, Projectable, Provenance, Scalar, err_ub_format,
    err_unsup_format, interp_ok, throw_inval, throw_ub, throw_ub_format,
};
use crate::interpret::{MPlaceTy, Writeable};

#[derive(Copy, Clone, Debug, PartialEq, Eq)]
enum MulAddType {
    /// Used with `fma` and `simd_fma`, always uses fused-multiply-add
    Fused,
    /// Used with `fmuladd` and `simd_relaxed_fma`, nondeterministically determines whether to use
    /// fma or simple multiply-add
    Nondeterministic,
}

#[derive(Copy, Clone)]
pub(crate) enum MinMax {
    /// The IEEE-2019 `minimum` operation - see `f32::minimum` etc.
    /// In particular, `-0.0` is considered smaller than `+0.0` and
    /// if either input is NaN, the result is NaN.
    Minimum,
    /// The IEEE-2019 `minimumNumber` operation but with non-deterministic signed zero handling
    /// (like in IEEE-2008 `minNum`) - see `f32::min` etc.
    /// In particular, if the inputs are `-0.0` and `+0.0`, the result is non-deterministic,
    /// and if one argument is NaN (quiet or signaling), the other one is returned.
    MinimumNumberNsz,
    /// The IEEE-2019 `maximum` operation - see `f32::maximum` etc.
    /// In particular, `-0.0` is considered smaller than `+0.0` and
    /// if either input is NaN, the result is NaN.
    Maximum,
    /// The IEEE-2019 `maximumNumber` operation but with non-deterministic signed zero handling
    /// (like in IEEE-2008 `maxNum`) - see `f32::max` etc.
    /// In particular, if the inputs are `-0.0` and `+0.0`, the result is non-deterministic,
    /// and if one argument is NaN (quiet or signaling), the other one is returned.
    MaximumNumberNsz,
}

/// Whether two types `T` and `U` are compatible when a value of type `T` is passed as a c-variadic
/// argument and read as a value of type `U`.
pub enum VarArgCompatible {
    /// `T` and `U` are compatible, e.g.
    ///
    /// - They're the same type.
    /// - One is `usize`/`isize`, the other an integer type of the same width
    /// and sign on the current target.
    /// - They are compatible pointer types (see the exact rules below).
    Compatible,
    /// `T` and `U` are definitely not compatible.
    Incompatible,
    /// `T` and `U` are corresponding signed and unsigned integer types.
    /// This is compatible only if the value can be represented in both types.
    CastIntTo { source_is_signed: bool },
}

/// Directly returns an `Allocation` containing an absolute path representation of the given type.
pub(crate) fn alloc_type_name<'tcx>(tcx: TyCtxt<'tcx>, ty: Ty<'tcx>) -> (AllocId, u64) {
    let path = crate::util::type_name(tcx, ty);
    let bytes = path.into_bytes();
    let len = bytes.len().try_into().unwrap();
    (tcx.allocate_bytes_dedup(bytes, CTFE_ALLOC_SALT), len)
}
impl<'tcx, M: Machine<'tcx>> InterpCx<'tcx, M> {
    /// Generates a value of `TypeId` for `ty` in-place.
    pub(crate) fn write_type_id(
        &mut self,
        ty: Ty<'tcx>,
        dest: &impl Writeable<'tcx, M::Provenance>,
    ) -> InterpResult<'tcx, ()> {
        debug_assert!(
            !ty.has_erasable_regions(),
            "type {ty:?} has regions that need erasing before writing a TypeId",
        );

        let tcx = self.tcx;
        let type_id_hash = tcx.type_id_hash(ty).as_u128();
        let op = self.const_val_to_op(
            ConstValue::Scalar(Scalar::from_u128(type_id_hash)),
            tcx.types.u128,
            None,
        )?;
        self.copy_op_allow_transmute(&op, dest)?;

        // Give the each pointer-sized chunk provenance that knows about the type id.
        // Here we rely on `TypeId` being a newtype around an array of pointers, so we
        // first project to its only field and then the array elements.
        let alloc_id = tcx.reserve_and_set_type_id_alloc(ty);
        let arr = self.project_field(dest, FieldIdx::ZERO)?;
        let mut elem_iter = self.project_array_fields(&arr)?;
        while let Some((_, elem)) = elem_iter.next(self)? {
            // Decorate this part of the hash with provenance; leave the integer part unchanged.
            let hash_fragment = self.read_scalar(&elem)?.to_target_usize(&tcx)?;
            let ptr = Pointer::new(alloc_id.into(), Size::from_bytes(hash_fragment));
            let ptr = self.global_root_pointer(ptr)?;
            let val = Scalar::from_pointer(ptr, &tcx);
            self.write_scalar(val, &elem)?;
        }
        interp_ok(())
    }

    /// Read a value of type `TypeId`, returning the type it represents.
    pub(crate) fn read_type_id(
        &self,
        op: &OpTy<'tcx, M::Provenance>,
    ) -> InterpResult<'tcx, Ty<'tcx>> {
        // `TypeId` is a newtype around an array of pointers. All pointers must have the same
        // provenance, and that provenance represents the type.
        let ptr_size = self.pointer_size().bytes_usize();
        let arr = self.project_field(op, FieldIdx::ZERO)?;

        let mut ty_and_hash = None;
        let mut elem_iter = self.project_array_fields(&arr)?;
        while let Some((idx, elem)) = elem_iter.next(self)? {
            let elem = self.read_pointer(&elem)?;
            let (elem_ty, elem_hash) = self.get_ptr_type_id(elem)?;
            // If this is the first element, remember the type and its hash.
            // If this is not the first element, ensure it is consistent with the previous ones.
            let full_hash = match ty_and_hash {
                None => {
                    let hash = self.tcx.type_id_hash(elem_ty).as_u128();
                    let mut hash_bytes = [0u8; 16];
                    write_target_uint(self.data_layout().endian, &mut hash_bytes, hash).unwrap();
                    ty_and_hash = Some((elem_ty, hash_bytes));
                    hash_bytes
                }
                Some((ty, hash_bytes)) => {
                    if ty != elem_ty {
                        throw_ub_format!(
                            "invalid `TypeId` value: not all bytes carry the same type id metadata"
                        );
                    }
                    hash_bytes
                }
            };
            // Ensure the elem_hash matches the corresponding part of the full hash.
            let hash_frag = &full_hash[(idx as usize) * ptr_size..][..ptr_size];
            if read_target_uint(self.data_layout().endian, hash_frag).unwrap() != elem_hash.into() {
                throw_ub_format!(
                    "invalid `TypeId` value: the hash does not match the type id metadata"
                );
            }
        }

        interp_ok(ty_and_hash.unwrap().0)
    }

    /// 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 eval_intrinsic(
        &mut self,
        instance: ty::Instance<'tcx>,
        args: &[OpTy<'tcx, M::Provenance>],
        dest: &PlaceTy<'tcx, M::Provenance>,
        ret: Option<mir::BasicBlock>,
    ) -> InterpResult<'tcx, bool> {
        let instance_args = instance.args;
        let intrinsic_name = self.tcx.item_name(instance.def_id());

        if intrinsic_name.as_str().starts_with("atomic_") {
            return self.eval_atomic_intrinsic(intrinsic_name, instance_args, args, dest, ret);
        }
        if intrinsic_name.as_str().starts_with("simd_") {
            return self.eval_simd_intrinsic(intrinsic_name, instance_args, args, dest, ret);
        }

        let tcx = self.tcx.tcx;

        match intrinsic_name {
            sym::type_name => {
                let tp_ty = instance.args.type_at(0);
                ensure_monomorphic_enough(tp_ty)?;
                let (alloc_id, meta) = alloc_type_name(tcx, tp_ty);
                let val = ConstValue::Slice { alloc_id, meta };
                let val = self.const_val_to_op(val, dest.layout.ty, Some(dest.layout))?;
                self.copy_op(&val, dest)?;
            }
            sym::needs_drop => {
                let tp_ty = instance.args.type_at(0);
                ensure_monomorphic_enough(tp_ty)?;
                let val = ConstValue::from_bool(tp_ty.needs_drop(tcx, self.typing_env));
                let val = self.const_val_to_op(val, tcx.types.bool, Some(dest.layout))?;
                self.copy_op(&val, dest)?;
            }
            sym::type_id => {
                let tp_ty = instance.args.type_at(0);
                ensure_monomorphic_enough(tp_ty)?;
                self.write_type_id(tp_ty, dest)?;
            }
            sym::type_id_eq => {
                let a_ty = self.read_type_id(&args[0])?;
                let b_ty = self.read_type_id(&args[1])?;
                self.write_scalar(Scalar::from_bool(a_ty == b_ty), dest)?;
            }
            sym::size_of => {
                let tp_ty = instance.args.type_at(0);
                let layout = self.layout_of(tp_ty)?;
                if !layout.is_sized() {
                    span_bug!(self.cur_span(), "unsized type for `size_of`");
                }
                let val = layout.size.bytes();
                self.write_scalar(Scalar::from_target_usize(val, self), dest)?;
            }
            sym::align_of => {
                let tp_ty = instance.args.type_at(0);
                let layout = self.layout_of(tp_ty)?;
                if !layout.is_sized() {
                    span_bug!(self.cur_span(), "unsized type for `align_of`");
                }
                let val = layout.align.bytes();
                self.write_scalar(Scalar::from_target_usize(val, self), dest)?;
            }
            kw::OffsetOf => {
                let tp_ty = instance.args.type_at(0);

                let variant = self.read_scalar(&args[0])?.to_u32()?;
                let field = self.read_scalar(&args[1])?.to_u32()? as usize;

                let layout = self.layout_of(tp_ty)?;
                let cx = ty::layout::LayoutCx::new(*self.tcx, self.typing_env);

                let layout = layout.for_variant(&cx, VariantIdx::from_u32(variant));
                let offset = layout.fields.offset(field).bytes();

                self.write_scalar(Scalar::from_target_usize(offset, self), dest)?;
            }
            sym::variant_count => {
                let tp_ty = instance.args.type_at(0);
                let ty = match tp_ty.kind() {
                    // Pattern types have the same number of variants as their base type.
                    // Even if we restrict e.g. which variants are valid, the variants are essentially just uninhabited.
                    // And `Result<(), !>` still has two variants according to `variant_count`.
                    ty::Pat(base, _) => *base,
                    _ => tp_ty,
                };
                let val = match ty.kind() {
                    // Correctly handles non-monomorphic calls, so there is no need for ensure_monomorphic_enough.
                    ty::Adt(adt, _) => {
                        ConstValue::from_target_usize(adt.variants().len() as u64, &tcx)
                    }
                    ty::Alias(..) | ty::Param(_) | ty::Placeholder(_) | ty::Infer(_) => {
                        throw_inval!(TooGeneric)
                    }
                    ty::Pat(..) => unreachable!(),
                    ty::Bound(_, _) => bug!("bound ty during ctfe"),
                    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::CoroutineClosure(_, _)
                    | ty::Coroutine(_, _)
                    | ty::CoroutineWitness(..)
                    | ty::UnsafeBinder(_)
                    | ty::Never
                    | ty::Tuple(_)
                    | ty::Error(_) => ConstValue::from_target_usize(0u64, &tcx),
                };
                let val = self.const_val_to_op(val, dest.layout.ty, Some(dest.layout))?;
                self.copy_op(&val, dest)?;
            }

            sym::caller_location => {
                let span = self.caller_location();
                let val = self.tcx.span_as_caller_location(span);
                let val =
                    self.const_val_to_op(val, self.tcx.caller_location_ty(), Some(dest.layout))?;
                self.copy_op(&val, dest)?;
            }

            sym::align_of_val | sym::size_of_val => {
                // Avoid `deref_pointer` -- this is not a deref, the ptr does not have to be
                // dereferenceable!
                let place = self.imm_ptr_to_mplace(&self.read_immediate(&args[0])?)?;
                let (size, align) = self
                    .size_and_align_of_val(&place)?
                    .ok_or_else(|| err_unsup_format!("`extern type` does not have known layout"))?;

                let result = match intrinsic_name {
                    sym::align_of_val => align.bytes(),
                    sym::size_of_val => size.bytes(),
                    _ => bug!(),
                };

                self.write_scalar(Scalar::from_target_usize(result, self), dest)?;
            }

            sym::fadd_algebraic
            | sym::fsub_algebraic
            | sym::fmul_algebraic
            | sym::fdiv_algebraic
            | sym::frem_algebraic => {
                let a = self.read_immediate(&args[0])?;
                let b = self.read_immediate(&args[1])?;

                let op = match intrinsic_name {
                    sym::fadd_algebraic => BinOp::Add,
                    sym::fsub_algebraic => BinOp::Sub,
                    sym::fmul_algebraic => BinOp::Mul,
                    sym::fdiv_algebraic => BinOp::Div,
                    sym::frem_algebraic => BinOp::Rem,

                    _ => bug!(),
                };

                let res = self.binary_op(op, &a, &b)?;
                // `binary_op` already called `generate_nan` if needed.
                let res = M::apply_float_nondet(self, res)?;
                self.write_immediate(*res, dest)?;
            }

            sym::ctpop
            | sym::cttz
            | sym::cttz_nonzero
            | sym::ctlz
            | sym::ctlz_nonzero
            | sym::bswap
            | sym::bitreverse => {
                let ty = instance_args.type_at(0);
                let layout = self.layout_of(ty)?;
                let val = self.read_scalar(&args[0])?;

                let out_val = self.numeric_intrinsic(intrinsic_name, val, layout, dest.layout)?;
                self.write_scalar(out_val, dest)?;
            }
            sym::saturating_add | sym::saturating_sub => {
                let l = self.read_immediate(&args[0])?;
                let r = self.read_immediate(&args[1])?;
                let val = self.saturating_arith(
                    if intrinsic_name == sym::saturating_add { BinOp::Add } else { BinOp::Sub },
                    &l,
                    &r,
                )?;
                self.write_scalar(val, dest)?;
            }
            sym::discriminant_value => {
                let place = self.deref_pointer(&args[0])?;
                let variant = self.read_discriminant(&place)?;
                let discr = self.discriminant_for_variant(place.layout.ty, variant)?;
                self.write_immediate(*discr, dest)?;
            }
            sym::exact_div => {
                let l = self.read_immediate(&args[0])?;
                let r = self.read_immediate(&args[1])?;
                self.exact_div(&l, &r, dest)?;
            }
            sym::unchecked_funnel_shl | sym::unchecked_funnel_shr => {
                let shift_left = intrinsic_name == sym::unchecked_funnel_shl;

                let l = self.read_immediate(&args[0])?;
                let r = self.read_immediate(&args[1])?;
                let shift = self.read_immediate(&args[2])?;

                let result = self.funnel_shift(&l, &r, &shift, shift_left)?;
                self.write_scalar(result, dest)?;
            }

            sym::copy => {
                self.copy_intrinsic(&args[0], &args[1], &args[2], /*nonoverlapping*/ false)?;
            }
            sym::write_bytes => {
                self.write_bytes_intrinsic(&args[0], &args[1], &args[2], "write_bytes")?;
            }
            sym::compare_bytes => {
                let result = self.compare_bytes_intrinsic(&args[0], &args[1], &args[2])?;
                self.write_scalar(result, dest)?;
            }
            sym::arith_offset => {
                let ptr = self.read_pointer(&args[0])?;
                let offset_count = self.read_target_isize(&args[1])?;
                let pointee_ty = instance_args.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 | sym::ptr_offset_from_unsigned => {
                let a = self.read_pointer(&args[0])?;
                let b = self.read_pointer(&args[1])?;

                let usize_layout = self.layout_of(self.tcx.types.usize)?;
                let isize_layout = self.layout_of(self.tcx.types.isize)?;

                // Get offsets for both that are at least relative to the same base.
                // With `OFFSET_IS_ADDR` this is trivial; without it we need either
                // two integers or two pointers into the same allocation.
                let (a_offset, b_offset, is_addr) = if M::Provenance::OFFSET_IS_ADDR {
                    (a.addr().bytes(), b.addr().bytes(), /*is_addr*/ true)
                } else {
                    match (self.ptr_try_get_alloc_id(a, 0), self.ptr_try_get_alloc_id(b, 0)) {
                        (Err(a), Err(b)) => {
                            // Neither pointer points to an allocation, so they are both absolute.
                            (a, b, /*is_addr*/ true)
                        }
                        (Ok((a_alloc_id, a_offset, _)), Ok((b_alloc_id, b_offset, _)))
                            if a_alloc_id == b_alloc_id =>
                        {
                            // Found allocation for both, and it's the same.
                            // Use these offsets for distance calculation.
                            (a_offset.bytes(), b_offset.bytes(), /*is_addr*/ false)
                        }
                        _ => {
                            // Not into the same allocation -- this is UB.
                            throw_ub_format!(
                                "`{name}` called on two different pointers that are not both derived from the same allocation",
                                name = intrinsic_name,
                            );
                        }
                    }
                };

                // Compute distance: a - b.
                let dist = {
                    // Addresses are unsigned, so this is a `usize` computation. We have to do the
                    // overflow check separately anyway.
                    let (val, overflowed) = {
                        let a_offset = ImmTy::from_uint(a_offset, usize_layout);
                        let b_offset = ImmTy::from_uint(b_offset, usize_layout);
                        self.binary_op(BinOp::SubWithOverflow, &a_offset, &b_offset)?
                            .to_scalar_pair()
                    };
                    if overflowed.to_bool()? {
                        // a < b
                        if intrinsic_name == sym::ptr_offset_from_unsigned {
                            throw_ub_format!(
                                "`ptr_offset_from_unsigned` called when first pointer has smaller {is_addr} than second: {a_offset} < {b_offset}",
                                a_offset = a_offset,
                                b_offset = b_offset,
                                is_addr = if is_addr { "address" } else { "offset" },
                            );
                        }
                        // The signed form of the intrinsic allows this. If we interpret the
                        // difference as isize, we'll get the proper signed difference. If that
                        // seems *positive* or equal to isize::MIN, they were more than isize::MAX apart.
                        let dist = val.to_target_isize(self)?;
                        if dist >= 0 || i128::from(dist) == self.pointer_size().signed_int_min() {
                            throw_ub_format!(
                                "`{intrinsic_name}` called when first pointer is too far before second"
                            );
                        }
                        dist
                    } else {
                        // b >= a
                        let dist = val.to_target_isize(self)?;
                        // If converting to isize produced a *negative* result, we had an overflow
                        // because they were more than isize::MAX apart.
                        if dist < 0 {
                            throw_ub_format!(
                                "`{intrinsic_name}` called when first pointer is too far ahead of second"
                            );
                        }
                        dist
                    }
                };

                // Check that the memory between them is dereferenceable at all, starting from the
                // origin pointer: `dist` is `a - b`, so it is based on `b`.
                self.check_ptr_access_signed(b, dist, CheckInAllocMsg::Dereferenceable("pointer"))
                    .map_err_kind(|_| {
                        // This could mean they point to different allocations, or they point to the same allocation
                        // but not the entire range between the pointers is in-bounds.
                        if let Ok((a_alloc_id, ..)) = self.ptr_try_get_alloc_id(a, 0)
                            && let Ok((b_alloc_id, ..)) = self.ptr_try_get_alloc_id(b, 0)
                            && a_alloc_id == b_alloc_id
                        {
                            err_ub_format!(
                                "`{intrinsic_name}` called on two different pointers where the memory range between them is not in-bounds of an allocation"
                            )
                        } else {
                            err_ub_format!(
                                "`{intrinsic_name}` called on two different pointers that are not both derived from the same allocation"
                            )
                        }
                    })?;
                // Then check that this is also dereferenceable from `a`. This ensures that they are
                // derived from the same allocation.
                self.check_ptr_access_signed(
                    a,
                    dist.checked_neg().unwrap(), // i64::MIN is impossible as no allocation can be that large
                    CheckInAllocMsg::Dereferenceable("pointer"),
                )
                .map_err_kind(|_| {
                    // Make the error more specific.
                    err_ub_format!(
                        "`{intrinsic_name}` called on two different pointers that are not both derived from the same allocation"
                    )
                })?;

                // Perform division by size to compute return value.
                let ret_layout = if intrinsic_name == sym::ptr_offset_from_unsigned {
                    assert!(0 <= dist && dist <= self.target_isize_max());
                    usize_layout
                } else {
                    assert!(self.target_isize_min() <= dist && dist <= self.target_isize_max());
                    isize_layout
                };
                let pointee_layout = self.layout_of(instance_args.type_at(0))?;
                // If ret_layout is unsigned, we checked that so is the distance, so we are good.
                let val = ImmTy::from_int(dist, ret_layout);
                let size = ImmTy::from_int(pointee_layout.size.bytes(), ret_layout);
                self.exact_div(&val, &size, dest)?;
            }

            sym::black_box => {
                // These just return their argument
                self.copy_op(&args[0], dest)?;
            }
            sym::raw_eq => {
                let result = self.raw_eq_intrinsic(&args[0], &args[1])?;
                self.write_scalar(result, dest)?;
            }
            sym::typed_swap_nonoverlapping => {
                self.typed_swap_nonoverlapping_intrinsic(&args[0], &args[1])?;
            }

            sym::volatile_load => {
                let [ptr] = args else {
                    span_bug!(self.cur_span(), "invalid `volatile_load` call")
                };
                let place = self.deref_pointer(ptr)?;
                self.copy_op(&place, dest)?;
            }
            sym::volatile_store => {
                let [ptr, val] = args else {
                    span_bug!(self.cur_span(), "invalid `volatile_store` call")
                };
                let place = self.deref_pointer(ptr)?;
                self.copy_op(val, &place)?;
            }
            sym::volatile_set_memory => {
                let [ptr, val_byte, count] = args else {
                    span_bug!(self.cur_span(), "invalid `volatile_set_memory` call")
                };
                self.write_bytes_intrinsic(ptr, val_byte, count, "volatile_set_memory")?;
            }

            sym::vtable_size => {
                let ptr = self.read_pointer(&args[0])?;
                // `None` because we don't know which trait to expect here; any vtable is okay.
                let (size, _align) = self.get_vtable_size_and_align(ptr, None)?;
                self.write_scalar(Scalar::from_target_usize(size.bytes(), self), dest)?;
            }
            sym::vtable_align => {
                let ptr = self.read_pointer(&args[0])?;
                // `None` because we don't know which trait to expect here; any vtable is okay.
                let (_size, align) = self.get_vtable_size_and_align(ptr, None)?;
                self.write_scalar(Scalar::from_target_usize(align.bytes(), self), dest)?;
            }

            sym::minimum_number_nsz_f16 => {
                self.float_minmax_intrinsic::<Half>(args, MinMax::MinimumNumberNsz, dest)?
            }
            sym::minimum_number_nsz_f32 => {
                self.float_minmax_intrinsic::<Single>(args, MinMax::MinimumNumberNsz, dest)?
            }
            sym::minimum_number_nsz_f64 => {
                self.float_minmax_intrinsic::<Double>(args, MinMax::MinimumNumberNsz, dest)?
            }
            sym::minimum_number_nsz_f128 => {
                self.float_minmax_intrinsic::<Quad>(args, MinMax::MinimumNumberNsz, dest)?
            }

            sym::minimumf16 => self.float_minmax_intrinsic::<Half>(args, MinMax::Minimum, dest)?,
            sym::minimumf32 => {
                self.float_minmax_intrinsic::<Single>(args, MinMax::Minimum, dest)?
            }
            sym::minimumf64 => {
                self.float_minmax_intrinsic::<Double>(args, MinMax::Minimum, dest)?
            }
            sym::minimumf128 => self.float_minmax_intrinsic::<Quad>(args, MinMax::Minimum, dest)?,

            sym::maximum_number_nsz_f16 => {
                self.float_minmax_intrinsic::<Half>(args, MinMax::MaximumNumberNsz, dest)?
            }
            sym::maximum_number_nsz_f32 => {
                self.float_minmax_intrinsic::<Single>(args, MinMax::MaximumNumberNsz, dest)?
            }
            sym::maximum_number_nsz_f64 => {
                self.float_minmax_intrinsic::<Double>(args, MinMax::MaximumNumberNsz, dest)?
            }
            sym::maximum_number_nsz_f128 => {
                self.float_minmax_intrinsic::<Quad>(args, MinMax::MaximumNumberNsz, dest)?
            }

            sym::maximumf16 => self.float_minmax_intrinsic::<Half>(args, MinMax::Maximum, dest)?,
            sym::maximumf32 => {
                self.float_minmax_intrinsic::<Single>(args, MinMax::Maximum, dest)?
            }
            sym::maximumf64 => {
                self.float_minmax_intrinsic::<Double>(args, MinMax::Maximum, dest)?
            }
            sym::maximumf128 => self.float_minmax_intrinsic::<Quad>(args, MinMax::Maximum, dest)?,

            sym::copysignf16 => self.float_copysign_intrinsic::<Half>(args, dest)?,
            sym::copysignf32 => self.float_copysign_intrinsic::<Single>(args, dest)?,
            sym::copysignf64 => self.float_copysign_intrinsic::<Double>(args, dest)?,
            sym::copysignf128 => self.float_copysign_intrinsic::<Quad>(args, dest)?,

            sym::fabs | sym::floor | sym::ceil | sym::trunc | sym::round | sym::round_ties_even => {
                let arg = self.read_immediate(&args[0])?;
                let ty::Float(float_ty) = arg.layout.ty.kind() else {
                    span_bug!(
                        self.cur_span(),
                        "non-float type for float intrinsic: {}",
                        arg.layout.ty,
                    );
                };
                let out_val = match float_ty {
                    FloatTy::F16 => self.unop_float_intrinsic::<Half>(intrinsic_name, arg)?,
                    FloatTy::F32 => self.unop_float_intrinsic::<Single>(intrinsic_name, arg)?,
                    FloatTy::F64 => self.unop_float_intrinsic::<Double>(intrinsic_name, arg)?,
                    FloatTy::F128 => self.unop_float_intrinsic::<Quad>(intrinsic_name, arg)?,
                };
                self.write_scalar(out_val, dest)?;
            }

            sym::fma | sym::fmuladd => {
                let fma_type = if intrinsic_name == sym::fma {
                    MulAddType::Fused
                } else {
                    MulAddType::Nondeterministic
                };
                let ty = args[0].layout.ty;
                let ty::Float(float_ty) = ty.kind() else {
                    span_bug!(self.cur_span(), "non-float type for float intrinsic: {ty}");
                };
                match float_ty {
                    FloatTy::F16 => self.float_muladd_intrinsic::<Half>(args, dest, fma_type)?,
                    FloatTy::F32 => self.float_muladd_intrinsic::<Single>(args, dest, fma_type)?,
                    FloatTy::F64 => self.float_muladd_intrinsic::<Double>(args, dest, fma_type)?,
                    FloatTy::F128 => self.float_muladd_intrinsic::<Quad>(args, dest, fma_type)?,
                }
            }

            sym::va_copy => {
                let va_list = self.deref_pointer(&args[0])?;

                let key_mplace = self.va_list_key_field(&va_list)?;
                let key = self.read_pointer(&key_mplace)?;

                let varargs = self.get_ptr_va_list(key)?;
                let copy_key = self.va_list_ptr(varargs.clone());

                // Zero the destination VaList, so it is fully initialized.
                let dest = self.force_allocation(dest)?;
                let zeros = std::iter::repeat_n(0u8, dest.layout.size.bytes_usize());
                self.write_bytes_ptr(dest.ptr(), zeros)?;

                let copy_key_mplace = self.va_list_key_field(&dest)?;
                self.write_pointer(copy_key, &copy_key_mplace)?;
            }

            sym::va_end => {
                let va_list = self.deref_pointer(&args[0])?;
                let key_mplace = self.va_list_key_field(&va_list)?;
                let key = self.read_pointer(&key_mplace)?;

                self.deallocate_va_list(key)?;
            }

            sym::va_arg => {
                let va_list = self.deref_pointer(&args[0])?;
                let key_mplace = self.va_list_key_field(&va_list)?;
                let key = self.read_pointer(&key_mplace)?;

                // Invalidate the old list and get its content. We'll recreate the
                // new list (one element shorter) below.
                let mut varargs = self.deallocate_va_list(key)?;

                let Some(arg_mplace) = varargs.pop_front() else {
                    throw_ub!(VaArgOutOfBounds);
                };

                // Error when the caller's argument is not c-variadic compatible with the type
                // requested by the callee.
                self.validate_c_variadic_argument(&arg_mplace, dest.layout)?;

                // Copy the argument, allowing a transmute and relying on the compatibility check
                // rejecting conversions between types of different size.
                self.copy_op_allow_transmute(&arg_mplace, dest)?;

                // Update the VaList pointer.
                let new_key = self.va_list_ptr(varargs);
                self.write_pointer(new_key, &key_mplace)?;
            }

            // Unsupported intrinsic: skip the return_to_block below.
            _ => return interp_ok(false),
        }

        trace!("{:?}", self.dump_place(&dest));
        self.return_to_block(ret)?;
        interp_ok(true)
    }

    /// Validate whether the value and type passed by the caller are compatible with the type
    /// requested by the callee. Based on section 7.16.1.1 of the C23 specification.
    ///
    /// The callee requesting a value of a type is valid when that type is compatible with the type
    /// provided by the caller (see `validate_c_variadic_compatible_ty`) and, if both types are
    /// integers of the same size but different signedness, the passed value must be representable
    /// in both types.
    fn validate_c_variadic_argument(
        &self,
        arg_mplace: &MPlaceTy<'tcx, M::Provenance>,
        callee_type: TyAndLayout<'tcx>,
    ) -> InterpResult<'tcx> {
        let callee_ty = callee_type.ty;
        let caller_ty = arg_mplace.layout.ty;

        match self.validate_c_variadic_compatible_ty(arg_mplace.layout.ty, callee_type.ty)? {
            VarArgCompatible::Compatible => interp_ok(()),
            VarArgCompatible::Incompatible => throw_ub_format!(
                "va_arg type mismatch: requested `{}` is incompatible with next argument of type `{}`",
                callee_ty,
                caller_ty,
            ),
            VarArgCompatible::CastIntTo { source_is_signed } => {
                // Check that the value can be represented in the target type.
                let size = arg_mplace.layout.size;
                let scalar = self.read_scalar(arg_mplace)?;
                if scalar.to_int(size)? < 0 {
                    throw_ub_format!(
                        "va_arg value mismatch: value `{value}_{caller_ty}` cannot be represented by type `{callee_ty}`",
                        value = if source_is_signed {
                            scalar.to_int(size)?.to_string()
                        } else {
                            scalar.to_uint(size)?.to_string()
                        }
                    )
                }

                interp_ok(())
            }
        }
    }

    /// Check whether the caller and callee type are compatible for c-variadic calls. Further
    /// validation of the argument value may be needed to detect all UB.
    ///
    /// Types `T` and `U` are compatible when:
    ///
    /// - `T` and `U` are the same type.
    /// - `T` and `U` are integer types of the same size.
    /// - `T` and `U` are both pointers, and their target types are compatible.
    /// - `T` is a pointer to [`std::ffi::c_void`] and `U` is a pointer to [`i8`] or [`u8`],
    /// or vice versa.
    ///
    /// This is designed to match the C rules for variadics, it is incomparable to what we allow in
    /// terms of ABI mismatches for regular (fixed) function arguments.
    pub fn validate_c_variadic_compatible_ty(
        &self,
        caller_type: Ty<'tcx>,
        callee_type: Ty<'tcx>,
    ) -> InterpResult<'tcx, VarArgCompatible> {
        // FIXME: Accepting two copies of the same repr(C) type as compatible is currently not
        // guaranteed by our `VaList::next_arg` docs, but it is needed for Miri itself when it
        // checks whether shims were called with the right arguments. We should eventually put this
        // into the docs as well.
        if self.identical_c_types(caller_type, callee_type) {
            return interp_ok(VarArgCompatible::Compatible);
        }

        if self.layout_of(caller_type)?.size != self.layout_of(callee_type)?.size {
            return interp_ok(VarArgCompatible::Incompatible);
        }

        // Any character type (`char`, `unsigned char` and `signed char`) is compatible with
        // `void*`, so the signedness of `c_char` is irrelevant here.
        let is_c_char = |ty: Ty<'_>| matches!(ty.kind(), ty::Uint(UintTy::U8) | ty::Int(IntTy::I8));

        match (caller_type.kind(), callee_type.kind()) {
            // Some types look different but are actually the same for ABI purposes.
            (ty::Int(_), ty::Int(_)) | (ty::Uint(_), ty::Uint(_)) => {
                // E.g. cast between `usize` and `u64` on a 64-bit platform.
                interp_ok(VarArgCompatible::Compatible)
            }
            // C allows different types if...
            // - "both types are pointers to qualified or unqualified versions of compatible types"
            // - "one type is pointer to qualified or unqualified void and the other is a pointer to a qualified or
            //   unqualified character type"
            //
            // As usual for the ABI, we treat references and raw pointers alike.
            (
                ty::RawPtr(caller_target_ty, _) | ty::Ref(_, caller_target_ty, _),
                ty::RawPtr(callee_target_ty, _) | ty::Ref(_, callee_target_ty, _),
            ) => {
                // In C, types can be qualified by a combination of `const`, `volatile` and
                // `restrict`. These properties are irrelevant for the ABI, and don't have an
                // equivalent in rust.

                // Accept the cast if one type is pointer to void, and the other is a pointer to
                // a character type (`char`, `unsigned char` and `signed char`).
                if caller_target_ty.is_c_void(self.tcx.tcx) && is_c_char(*callee_target_ty) {
                    return interp_ok(VarArgCompatible::Compatible);
                }
                if callee_target_ty.is_c_void(self.tcx.tcx) && is_c_char(*caller_target_ty) {
                    return interp_ok(VarArgCompatible::Compatible);
                }

                // Accept the cast if both types are pointers to compatible types.
                match self
                    .validate_c_variadic_compatible_ty(*caller_target_ty, *callee_target_ty)?
                {
                    VarArgCompatible::Incompatible => interp_ok(VarArgCompatible::Incompatible),
                    VarArgCompatible::Compatible => interp_ok(VarArgCompatible::Compatible),
                    VarArgCompatible::CastIntTo { source_is_signed: _ } => {
                        // The integer cast check is not needed when the value is behind a pointer.
                        interp_ok(VarArgCompatible::Compatible)
                    }
                }
            }
            // - "one type is a signed integer type, the other type is the corresponding unsigned integer type,
            //   and the value is representable in both types"
            (ty::Int(_), ty::Uint(_)) => {
                interp_ok(VarArgCompatible::CastIntTo { source_is_signed: true })
            }
            (ty::Uint(_), ty::Int(_)) => {
                interp_ok(VarArgCompatible::CastIntTo { source_is_signed: false })
            }
            // - "or, the type of the next argument is nullptr_t and type is a pointer type that has the same
            //   representation and alignment requirements as a pointer to a character type"
            //   This one does not have an equivalent form in Rust.
            _ => interp_ok(VarArgCompatible::Incompatible),
        }
    }

    pub(super) fn eval_nondiverging_intrinsic(
        &mut self,
        intrinsic: &NonDivergingIntrinsic<'tcx>,
    ) -> InterpResult<'tcx> {
        match intrinsic {
            NonDivergingIntrinsic::Assume(op) => {
                let op = self.eval_operand(op, None)?;
                let cond = self.read_scalar(&op)?.to_bool()?;
                if !cond {
                    throw_ub_format!("`assume` called with `false`");
                }
                interp_ok(())
            }
            NonDivergingIntrinsic::CopyNonOverlapping(mir::CopyNonOverlapping {
                count,
                src,
                dst,
            }) => {
                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)
            }
        }
    }

    pub fn numeric_intrinsic(
        &self,
        name: Symbol,
        val: Scalar<M::Provenance>,
        layout: TyAndLayout<'tcx>,
        ret_layout: TyAndLayout<'tcx>,
    ) -> InterpResult<'tcx, Scalar<M::Provenance>> {
        assert!(layout.ty.is_integral(), "invalid type for numeric intrinsic: {}", layout.ty);
        let bits = val.to_bits(layout.size)?; // these operations all ignore the sign
        let extra = 128 - u128::from(layout.size.bits());
        let bits_out = match name {
            sym::ctpop => u128::from(bits.count_ones()),
            sym::ctlz_nonzero | sym::cttz_nonzero if bits == 0 => {
                throw_ub_format!("`{name}` called on 0");
            }
            sym::ctlz | sym::ctlz_nonzero => u128::from(bits.leading_zeros()) - extra,
            sym::cttz | sym::cttz_nonzero => u128::from((bits << extra).trailing_zeros()) - extra,
            sym::bswap => {
                assert_eq!(layout, ret_layout);
                (bits << extra).swap_bytes()
            }
            sym::bitreverse => {
                assert_eq!(layout, ret_layout);
                (bits << extra).reverse_bits()
            }
            _ => bug!("not a numeric intrinsic: {}", name),
        };
        interp_ok(Scalar::from_uint(bits_out, ret_layout.size))
    }

    pub fn exact_div(
        &mut self,
        a: &ImmTy<'tcx, M::Provenance>,
        b: &ImmTy<'tcx, M::Provenance>,
        dest: &PlaceTy<'tcx, M::Provenance>,
    ) -> InterpResult<'tcx> {
        assert_eq!(a.layout.ty, b.layout.ty);
        assert_matches!(a.layout.ty.kind(), ty::Int(..) | ty::Uint(..));

        // 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 rem = self.binary_op(BinOp::Rem, a, b)?;
        // sign does not matter for 0 test, so `to_bits` is fine
        if rem.to_scalar().to_bits(a.layout.size)? != 0 {
            throw_ub_format!("exact_div: {a} cannot be divided by {b} without remainder")
        }
        // `Rem` says this is all right, so we can let `Div` do its job.
        let res = self.binary_op(BinOp::Div, a, b)?;
        self.write_immediate(*res, dest)
    }

    pub fn saturating_arith(
        &self,
        mir_op: BinOp,
        l: &ImmTy<'tcx, M::Provenance>,
        r: &ImmTy<'tcx, M::Provenance>,
    ) -> InterpResult<'tcx, Scalar<M::Provenance>> {
        assert_eq!(l.layout.ty, r.layout.ty);
        assert_matches!(l.layout.ty.kind(), ty::Int(..) | ty::Uint(..));
        assert_matches!(mir_op, BinOp::Add | BinOp::Sub);

        let (val, overflowed) =
            self.binary_op(mir_op.wrapping_to_overflowing().unwrap(), l, r)?.to_scalar_pair();
        interp_ok(if overflowed.to_bool()? {
            let size = l.layout.size;
            if l.layout.backend_repr.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: i128 = l.to_scalar().to_int(l.layout.size)?;
                if first_term >= 0 {
                    // 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_int(size.signed_int_max(), size)
                } else {
                    // Positive overflow not possible for similar reason.
                    Scalar::from_int(size.signed_int_min(), size)
                }
            } else {
                // unsigned
                if mir_op == BinOp::Add {
                    // max unsigned
                    Scalar::from_uint(size.unsigned_int_max(), size)
                } else {
                    // underflow to 0
                    Scalar::from_uint(0u128, size)
                }
            }
        } else {
            val
        })
    }

    pub fn funnel_shift(
        &self,
        left: &ImmTy<'tcx, M::Provenance>,
        right: &ImmTy<'tcx, M::Provenance>,
        shift_bits: &ImmTy<'tcx, M::Provenance>,
        shift_left: bool,
    ) -> InterpResult<'tcx, Scalar<M::Provenance>> {
        let size = left.layout.size;
        let left = left.to_scalar().to_bits(size)?;
        let right = right.to_scalar().to_bits(size)?;
        let shift_bits = shift_bits.to_scalar().to_bits(shift_bits.layout.size)?; // can have different size

        if shift_bits >= size.bits().into() {
            throw_ub_format!(
                "overflowing shift by {shift_bits} in {} funnel shift",
                if shift_left { "left" } else { "right" }
            );
        }
        let inv_shift_bits = u32::try_from(u128::from(size.bits()) - shift_bits).unwrap();

        // A funnel shift left by S can be implemented as `(x << S) | y.unbounded_shr(SIZE - S)`.
        // The `unbounded_shr` is needed because otherwise if `S = 0`, it would be `x | y`
        // when it should be `x`.
        //
        // This selects the least-significant `SIZE - S` bits of `x`, followed by the `S` most
        // significant bits of `y`. As `left` and `right` both occupy the lower `SIZE` bits,
        // we can treat the lower `SIZE` bits as an integer of the right width and use
        // the same implementation, but on a zero-extended `x` and `y`. This works because
        // `x << S` just pushes the `SIZE-S` MSBs out, and `y >> (SIZE - S)` shifts in
        // zeros, as it is zero-extended. To the lower `SIZE` bits, this looks just like a
        // funnel shift left.
        //
        // Note that the `unbounded_sh{l,r}`s are needed only in case we are using this on
        // `u128xN` and `inv_shift_bits == 128`.
        let result_bits = if shift_left {
            (left << shift_bits) | right.unbounded_shr(inv_shift_bits)
        } else {
            left.unbounded_shl(inv_shift_bits) | (right >> shift_bits)
        };

        let (result, _overflow) = ScalarInt::truncate_from_uint(result_bits, size);
        interp_ok(result.into())
    }

    /// Offsets a pointer by some multiple of its type, returning an error if the pointer leaves its
    /// allocation.
    pub fn ptr_offset_inbounds(
        &self,
        ptr: Pointer<Option<M::Provenance>>,
        offset_bytes: i64,
    ) -> InterpResult<'tcx, Pointer<Option<M::Provenance>>> {
        // The offset must be in bounds starting from `ptr`.
        self.check_ptr_access_signed(
            ptr,
            offset_bytes,
            CheckInAllocMsg::InboundsPointerArithmetic,
        )?;
        // This also implies that there is no overflow, so we are done.
        interp_ok(ptr.wrapping_signed_offset(offset_bytes, self))
    }

    /// Copy `count*size_of::<T>()` many bytes from `*src` to `*dst`.
    pub(crate) fn copy_intrinsic(
        &mut self,
        src: &OpTy<'tcx, <M as Machine<'tcx>>::Provenance>, // pointer to source memory
        dst: &OpTy<'tcx, <M as Machine<'tcx>>::Provenance>, // pointer to destination memory
        count: &OpTy<'tcx, <M as Machine<'tcx>>::Provenance>,
        nonoverlapping: bool,
    ) -> InterpResult<'tcx> {
        let count = self.read_target_usize(count)?;
        let layout = self.layout_of(src.layout.ty.builtin_deref(true).unwrap())?;
        let (size, align) = (layout.size, layout.align);

        let size = self.compute_size_in_bytes(size, count).ok_or_else(|| {
            err_ub_format!(
                "overflow computing total size of `{name}`",
                name = if nonoverlapping { "copy_nonoverlapping" } else { "copy" }
            )
        })?;

        let src = self.read_pointer(src)?;
        let dst = self.read_pointer(dst)?;

        self.check_ptr_align(src, align)?;
        self.check_ptr_align(dst, align)?;

        self.mem_copy(src, dst, size, nonoverlapping)
    }

    /// Does a *typed* swap of `*left` and `*right`.
    fn typed_swap_nonoverlapping_intrinsic(
        &mut self,
        left: &OpTy<'tcx, <M as Machine<'tcx>>::Provenance>,
        right: &OpTy<'tcx, <M as Machine<'tcx>>::Provenance>,
    ) -> InterpResult<'tcx> {
        let left = self.deref_pointer(left)?;
        let right = self.deref_pointer(right)?;
        assert_eq!(left.layout, right.layout);
        assert!(left.layout.is_sized());
        let kind = MemoryKind::Stack;
        let temp = self.allocate(left.layout, kind)?;
        self.copy_op(&left, &temp)?; // checks alignment of `left`

        // We want to always enforce non-overlapping, even if this is a scalar type.
        // Therefore we directly use the underlying `mem_copy` here.
        self.mem_copy(right.ptr(), left.ptr(), left.layout.size, /*nonoverlapping*/ true)?;
        // This means we also need to do the validation of the value that used to be in `right`
        // ourselves. This value is now in `left.` The one that started out in `left` already got
        // validated by the copy above.
        if M::enforce_validity(self, left.layout) {
            self.validate_place(
                &left.clone().into(),
                M::enforce_validity_recursively(self, left.layout),
                /*reset_provenance_and_padding*/ true,
            )?;
        }

        self.copy_op(&temp, &right)?; // checks alignment of `right`

        self.deallocate_ptr(temp.ptr(), None, kind)?;
        interp_ok(())
    }

    pub fn write_bytes_intrinsic(
        &mut self,
        dst: &OpTy<'tcx, <M as Machine<'tcx>>::Provenance>, // pointer to destination memory
        byte: &OpTy<'tcx, <M as Machine<'tcx>>::Provenance>,
        count: &OpTy<'tcx, <M as Machine<'tcx>>::Provenance>,
        name: &'static str,
    ) -> InterpResult<'tcx> {
        let layout = self.layout_of(dst.layout.ty.builtin_deref(true).unwrap())?;

        let dst = self.read_pointer(dst)?;
        let byte = self.read_scalar(byte)?.to_u8()?;
        let count = self.read_target_usize(count)?;

        // `checked_mul` enforces a too small bound (the correct one would probably be target_isize_max),
        // but no actual allocation can be big enough for the difference to be noticeable.
        let len = self
            .compute_size_in_bytes(layout.size, count)
            .ok_or_else(|| err_ub_format!("overflow computing total size of `{name}`"))?;

        // Check alignment as well. We set the attribute in LLVM.
        self.check_ptr_align(dst, layout.align)?;

        let bytes = std::iter::repeat_n(byte, len.bytes_usize());
        self.write_bytes_ptr(dst, bytes)
    }

    pub(crate) fn compare_bytes_intrinsic(
        &mut self,
        left: &OpTy<'tcx, <M as Machine<'tcx>>::Provenance>,
        right: &OpTy<'tcx, <M as Machine<'tcx>>::Provenance>,
        byte_count: &OpTy<'tcx, <M as Machine<'tcx>>::Provenance>,
    ) -> InterpResult<'tcx, Scalar<M::Provenance>> {
        let left = self.read_pointer(left)?;
        let right = self.read_pointer(right)?;
        let n = Size::from_bytes(self.read_target_usize(byte_count)?);

        let left_bytes = self.read_bytes_ptr_strip_provenance(left, n)?;
        let right_bytes = self.read_bytes_ptr_strip_provenance(right, n)?;

        // `Ordering`'s discriminants are -1/0/+1, so casting does the right thing.
        let result = Ord::cmp(left_bytes, right_bytes) as i32;
        interp_ok(Scalar::from_i32(result))
    }

    pub(crate) fn raw_eq_intrinsic(
        &mut self,
        lhs: &OpTy<'tcx, <M as Machine<'tcx>>::Provenance>,
        rhs: &OpTy<'tcx, <M as Machine<'tcx>>::Provenance>,
    ) -> InterpResult<'tcx, Scalar<M::Provenance>> {
        let layout = self.layout_of(lhs.layout.ty.builtin_deref(true).unwrap())?;
        assert!(layout.is_sized());

        let get_bytes = |this: &InterpCx<'tcx, M>,
                         op: &OpTy<'tcx, <M as Machine<'tcx>>::Provenance>|
         -> InterpResult<'tcx, &[u8]> {
            let ptr = this.read_pointer(op)?;
            this.check_ptr_align(ptr, layout.align)?;
            let Some(alloc_ref) = self.get_ptr_alloc(ptr, layout.size)? else {
                // zero-sized access
                return interp_ok(&[]);
            };
            alloc_ref.get_bytes_strip_provenance()
        };

        let lhs_bytes = get_bytes(self, lhs)?;
        let rhs_bytes = get_bytes(self, rhs)?;
        interp_ok(Scalar::from_bool(lhs_bytes == rhs_bytes))
    }

    fn unop_float_intrinsic<F>(
        &mut self,
        name: Symbol,
        arg: ImmTy<'tcx, M::Provenance>,
    ) -> InterpResult<'tcx, Scalar<M::Provenance>>
    where
        F: rustc_apfloat::Float + rustc_apfloat::FloatConvert<F> + Into<Scalar<M::Provenance>>,
    {
        let x: F = arg.to_scalar().to_float()?;
        match name {
            // bitwise, no NaN adjustments
            sym::fabs => interp_ok(x.abs().into()),

            sym::floor => self.float_round(x, rustc_apfloat::Round::TowardNegative),
            sym::ceil => self.float_round(x, rustc_apfloat::Round::TowardPositive),
            sym::trunc => self.float_round(x, rustc_apfloat::Round::TowardZero),
            sym::round => self.float_round(x, rustc_apfloat::Round::NearestTiesToAway),
            sym::round_ties_even => self.float_round(x, rustc_apfloat::Round::NearestTiesToEven),

            _ => bug!("not a unary float intrinsic: {}", name),
        }
    }

    fn float_minmax<F>(
        &self,
        a: Scalar<M::Provenance>,
        b: Scalar<M::Provenance>,
        op: MinMax,
    ) -> InterpResult<'tcx, Scalar<M::Provenance>>
    where
        F: rustc_apfloat::Float + rustc_apfloat::FloatConvert<F> + Into<Scalar<M::Provenance>>,
    {
        let a: F = a.to_float()?;
        let b: F = b.to_float()?;
        let res = if matches!(op, MinMax::MinimumNumberNsz | MinMax::MaximumNumberNsz) && a == b {
            // They are definitely not NaN (those are never equal), but they could be `+0` and `-0`.
            // Let the machine decide which one to return.
            M::equal_float_min_max(self, a, b)
        } else {
            let result = match op {
                MinMax::Minimum => a.minimum(b),
                MinMax::MinimumNumberNsz => a.min(b),
                MinMax::Maximum => a.maximum(b),
                MinMax::MaximumNumberNsz => a.max(b),
            };
            self.adjust_nan(result, &[a, b])
        };

        interp_ok(res.into())
    }

    fn float_minmax_intrinsic<F>(
        &mut self,
        args: &[OpTy<'tcx, M::Provenance>],
        op: MinMax,
        dest: &PlaceTy<'tcx, M::Provenance>,
    ) -> InterpResult<'tcx, ()>
    where
        F: rustc_apfloat::Float + rustc_apfloat::FloatConvert<F> + Into<Scalar<M::Provenance>>,
    {
        let res =
            self.float_minmax::<F>(self.read_scalar(&args[0])?, self.read_scalar(&args[1])?, op)?;
        self.write_scalar(res, dest)?;
        interp_ok(())
    }

    fn float_copysign_intrinsic<F>(
        &mut self,
        args: &[OpTy<'tcx, M::Provenance>],
        dest: &PlaceTy<'tcx, M::Provenance>,
    ) -> InterpResult<'tcx, ()>
    where
        F: rustc_apfloat::Float + rustc_apfloat::FloatConvert<F> + Into<Scalar<M::Provenance>>,
    {
        let a: F = self.read_scalar(&args[0])?.to_float()?;
        let b: F = self.read_scalar(&args[1])?.to_float()?;
        // bitwise, no NaN adjustments
        self.write_scalar(a.copy_sign(b), dest)?;
        interp_ok(())
    }

    fn float_round<F>(
        &mut self,
        x: F,
        mode: rustc_apfloat::Round,
    ) -> InterpResult<'tcx, Scalar<M::Provenance>>
    where
        F: rustc_apfloat::Float + rustc_apfloat::FloatConvert<F> + Into<Scalar<M::Provenance>>,
    {
        let res = x.round_to_integral(mode).value;
        let res = self.adjust_nan(res, &[x]);
        interp_ok(res.into())
    }

    fn float_muladd<F>(
        &self,
        a: Scalar<M::Provenance>,
        b: Scalar<M::Provenance>,
        c: Scalar<M::Provenance>,
        typ: MulAddType,
    ) -> InterpResult<'tcx, Scalar<M::Provenance>>
    where
        F: rustc_apfloat::Float + rustc_apfloat::FloatConvert<F> + Into<Scalar<M::Provenance>>,
    {
        let a: F = a.to_float()?;
        let b: F = b.to_float()?;
        let c: F = c.to_float()?;

        let fuse = typ == MulAddType::Fused || M::float_fuse_mul_add(self);

        let res = if fuse { a.mul_add(b, c).value } else { ((a * b).value + c).value };
        let res = self.adjust_nan(res, &[a, b, c]);
        interp_ok(res.into())
    }

    fn float_muladd_intrinsic<F>(
        &mut self,
        args: &[OpTy<'tcx, M::Provenance>],
        dest: &PlaceTy<'tcx, M::Provenance>,
        typ: MulAddType,
    ) -> InterpResult<'tcx, ()>
    where
        F: rustc_apfloat::Float + rustc_apfloat::FloatConvert<F> + Into<Scalar<M::Provenance>>,
    {
        let a = self.read_scalar(&args[0])?;
        let b = self.read_scalar(&args[1])?;
        let c = self.read_scalar(&args[2])?;

        let res = self.float_muladd::<F>(a, b, c, typ)?;
        self.write_scalar(res, dest)?;
        interp_ok(())
    }

    /// Converts `src` from floating point to integer type `dest_ty`
    /// after rounding with mode `round`.
    /// Returns `None` if `f` is NaN or out of range.
    pub fn float_to_int_checked(
        &self,
        src: &ImmTy<'tcx, M::Provenance>,
        cast_to: TyAndLayout<'tcx>,
        round: rustc_apfloat::Round,
    ) -> InterpResult<'tcx, Option<ImmTy<'tcx, M::Provenance>>> {
        fn float_to_int_inner<'tcx, F: rustc_apfloat::Float, M: Machine<'tcx>>(
            ecx: &InterpCx<'tcx, M>,
            src: F,
            cast_to: TyAndLayout<'tcx>,
            round: rustc_apfloat::Round,
        ) -> (Scalar<M::Provenance>, rustc_apfloat::Status) {
            let int_size = cast_to.layout.size;
            match cast_to.ty.kind() {
                // Unsigned
                ty::Uint(_) => {
                    let res = src.to_u128_r(int_size.bits_usize(), round, &mut false);
                    (Scalar::from_uint(res.value, int_size), res.status)
                }
                // Signed
                ty::Int(_) => {
                    let res = src.to_i128_r(int_size.bits_usize(), round, &mut false);
                    (Scalar::from_int(res.value, int_size), res.status)
                }
                // Nothing else
                _ => span_bug!(
                    ecx.cur_span(),
                    "attempted float-to-int conversion with non-int output type {}",
                    cast_to.ty,
                ),
            }
        }

        let ty::Float(fty) = src.layout.ty.kind() else {
            bug!("float_to_int_checked: non-float input type {}", src.layout.ty)
        };

        let (val, status) = match fty {
            FloatTy::F16 => float_to_int_inner(self, src.to_scalar().to_f16()?, cast_to, round),
            FloatTy::F32 => float_to_int_inner(self, src.to_scalar().to_f32()?, cast_to, round),
            FloatTy::F64 => float_to_int_inner(self, src.to_scalar().to_f64()?, cast_to, round),
            FloatTy::F128 => float_to_int_inner(self, src.to_scalar().to_f128()?, cast_to, round),
        };

        if status.intersects(
            rustc_apfloat::Status::INVALID_OP
                | rustc_apfloat::Status::OVERFLOW
                | rustc_apfloat::Status::UNDERFLOW,
        ) {
            // Floating point value is NaN (flagged with INVALID_OP) or outside the range
            // of values of the integer type (flagged with OVERFLOW or UNDERFLOW).
            interp_ok(None)
        } else {
            // Floating point value can be represented by the integer type after rounding.
            // The INEXACT flag is ignored on purpose to allow rounding.
            interp_ok(Some(ImmTy::from_scalar(val, cast_to)))
        }
    }

    /// Get the MPlace of the key from the place storing the VaList.
    pub(super) fn va_list_key_field<P: Projectable<'tcx, M::Provenance>>(
        &self,
        va_list: &P,
    ) -> InterpResult<'tcx, P> {
        // The struct wrapped by VaList.
        let va_list_inner = self.project_field(va_list, FieldIdx::ZERO)?;

        // Find the first pointer field in this struct. The exact index is target-specific.
        let ty::Adt(adt, substs) = va_list_inner.layout().ty.kind() else {
            bug!("invalid VaListImpl layout");
        };

        for (i, field) in adt.non_enum_variant().fields.iter().enumerate() {
            if field.ty(*self.tcx, *substs).skip_norm_wip().is_raw_ptr() {
                return self.project_field(&va_list_inner, FieldIdx::from_usize(i));
            }
        }

        bug!("no VaListImpl field is a pointer");
    }
}