miri/intrinsics/x86/

mod.rs

use rustc_abi::{FieldIdx, Size};
use rustc_apfloat::Float;
use rustc_apfloat::ieee::Single;
use rustc_middle::ty::Ty;
use rustc_middle::{mir, ty};
use rustc_span::Symbol;
use rustc_target::spec::Arch;

use self::helpers::bool_to_simd_element;
use crate::*;

mod aesni;
mod avx;
mod avx2;
mod avx512;
mod bmi;
mod gfni;
mod sha;
mod sse;
mod sse2;
mod sse3;
mod sse41;
mod sse42;
mod ssse3;

impl<'tcx> EvalContextExt<'tcx> for crate::MiriInterpCx<'tcx> {}
pub(super) trait EvalContextExt<'tcx>: crate::MiriInterpCxExt<'tcx> {
    fn emulate_x86_intrinsic(
        &mut self,
        link_name: Symbol,
        args: &[OpTy<'tcx>],
        dest: &MPlaceTy<'tcx>,
    ) -> InterpResult<'tcx, EmulateItemResult> {
        let this = self.eval_context_mut();
        // Prefix should have already been checked.
        let unprefixed_name = link_name.as_str().strip_prefix("llvm.x86.").unwrap();
        match unprefixed_name {
            // Used to implement the `_addcarry_u{32, 64}` and the `_subborrow_u{32, 64}` functions.
            // Computes a + b or a - b with input and output carry/borrow. The input carry/borrow is an 8-bit
            // value, which is interpreted as 1 if it is non-zero. The output carry/borrow is an 8-bit value that will be 0 or 1.
            // https://www.intel.com/content/www/us/en/docs/cpp-compiler/developer-guide-reference/2021-8/addcarry-u32-addcarry-u64.html
            // https://www.intel.com/content/www/us/en/docs/cpp-compiler/developer-guide-reference/2021-8/subborrow-u32-subborrow-u64.html
            "addcarry.32" | "addcarry.64" | "subborrow.32" | "subborrow.64" => {
                if unprefixed_name.ends_with("64") && this.tcx.sess.target.arch != Arch::X86_64 {
                    return interp_ok(EmulateItemResult::NotSupported);
                }

                let [cb_in, a, b] = this.check_shim_sig_llvm_intrinsic(link_name, args)?;
                let op = if unprefixed_name.starts_with("add") {
                    mir::BinOp::AddWithOverflow
                } else {
                    mir::BinOp::SubWithOverflow
                };

                let (sum, cb_out) = carrying_add(this, cb_in, a, b, op)?;
                this.write_scalar(cb_out, &this.project_field(dest, FieldIdx::ZERO)?)?;
                this.write_immediate(*sum, &this.project_field(dest, FieldIdx::ONE)?)?;
            }

            // Used to implement the `_mm_pause` function.
            // The intrinsic is used to hint the processor that the code is in a spin-loop.
            // It is compiled down to a `pause` instruction. When SSE2 is not available,
            // the instruction behaves like a no-op, so it is always safe to call the
            // intrinsic.
            "sse2.pause" => {
                let [] = this.check_shim_sig_llvm_intrinsic(link_name, args)?;
                // Only exhibit the spin-loop hint behavior when SSE2 is enabled.
                if this.tcx.sess.internal_target_features.contains(&Symbol::intern("sse2")) {
                    this.yield_active_thread();
                }
            }

            "pclmulqdq" | "pclmulqdq.256" | "pclmulqdq.512" => {
                let mut len = 2; // in units of 64bits
                this.expect_target_feature_for_intrinsic(link_name, "pclmulqdq")?;
                if unprefixed_name.ends_with(".256") {
                    this.expect_target_feature_for_intrinsic(link_name, "vpclmulqdq")?;
                    len = 4;
                } else if unprefixed_name.ends_with(".512") {
                    this.expect_target_feature_for_intrinsic(link_name, "vpclmulqdq")?;
                    this.expect_target_feature_for_intrinsic(link_name, "avx512f")?;
                    len = 8;
                }

                let [left, right, imm] = this.check_shim_sig_llvm_intrinsic(link_name, args)?;

                pclmulqdq(this, left, right, imm, dest, len)?;
            }

            name if name.starts_with("bmi.") => {
                return bmi::EvalContextExt::emulate_x86_bmi_intrinsic(this, link_name, args, dest);
            }
            // The GFNI extension does not get its own namespace.
            // Check for instruction names instead.
            name if name.starts_with("vgf2p8affine") || name.starts_with("vgf2p8mulb") => {
                return gfni::EvalContextExt::emulate_x86_gfni_intrinsic(
                    this, link_name, args, dest,
                );
            }
            name if name.starts_with("sha") => {
                return sha::EvalContextExt::emulate_x86_sha_intrinsic(this, link_name, args, dest);
            }
            name if name.starts_with("sse.") => {
                return sse::EvalContextExt::emulate_x86_sse_intrinsic(this, link_name, args, dest);
            }
            name if name.starts_with("sse2.") => {
                return sse2::EvalContextExt::emulate_x86_sse2_intrinsic(
                    this, link_name, args, dest,
                );
            }
            name if name.starts_with("sse3.") => {
                return sse3::EvalContextExt::emulate_x86_sse3_intrinsic(
                    this, link_name, args, dest,
                );
            }
            name if name.starts_with("ssse3.") => {
                return ssse3::EvalContextExt::emulate_x86_ssse3_intrinsic(
                    this, link_name, args, dest,
                );
            }
            name if name.starts_with("sse41.") => {
                return sse41::EvalContextExt::emulate_x86_sse41_intrinsic(
                    this, link_name, args, dest,
                );
            }
            name if name.starts_with("sse42.") => {
                return sse42::EvalContextExt::emulate_x86_sse42_intrinsic(
                    this, link_name, args, dest,
                );
            }
            name if name.starts_with("aesni.") => {
                return aesni::EvalContextExt::emulate_x86_aesni_intrinsic(
                    this, link_name, args, dest,
                );
            }
            name if name.starts_with("avx.") => {
                return avx::EvalContextExt::emulate_x86_avx_intrinsic(this, link_name, args, dest);
            }
            name if name.starts_with("avx2.") => {
                return avx2::EvalContextExt::emulate_x86_avx2_intrinsic(
                    this, link_name, args, dest,
                );
            }
            name if name.starts_with("avx512.") => {
                return avx512::EvalContextExt::emulate_x86_avx512_intrinsic(
                    this, link_name, args, dest,
                );
            }

            _ => return interp_ok(EmulateItemResult::NotSupported),
        }
        interp_ok(EmulateItemResult::NeedsReturn)
    }
}

#[derive(Copy, Clone)]
enum FloatBinOp {
    /// Comparison
    ///
    /// The semantics of this operator is a case distinction: we compare the two operands,
    /// and then we return one of the four booleans `gt`, `lt`, `eq`, `unord` depending on
    /// which class they fall into.
    ///
    /// AVX supports all 16 combinations, SSE only a subset
    ///
    /// <https://www.felixcloutier.com/x86/cmpss>
    /// <https://www.felixcloutier.com/x86/cmpps>
    /// <https://www.felixcloutier.com/x86/cmpsd>
    /// <https://www.felixcloutier.com/x86/cmppd>
    Cmp {
        /// Result when lhs < rhs
        gt: bool,
        /// Result when lhs > rhs
        lt: bool,
        /// Result when lhs == rhs
        eq: bool,
        /// Result when lhs is NaN or rhs is NaN
        unord: bool,
    },
    /// Minimum value (with SSE semantics)
    ///
    /// <https://www.felixcloutier.com/x86/minss>
    /// <https://www.felixcloutier.com/x86/minps>
    /// <https://www.felixcloutier.com/x86/minsd>
    /// <https://www.felixcloutier.com/x86/minpd>
    Min,
    /// Maximum value (with SSE semantics)
    ///
    /// <https://www.felixcloutier.com/x86/maxss>
    /// <https://www.felixcloutier.com/x86/maxps>
    /// <https://www.felixcloutier.com/x86/maxsd>
    /// <https://www.felixcloutier.com/x86/maxpd>
    Max,
}

impl FloatBinOp {
    /// Convert from the `imm` argument used to specify the comparison
    /// operation in intrinsics such as `llvm.x86.sse.cmp.ss`.
    fn cmp_from_imm<'tcx>(
        ecx: &crate::MiriInterpCx<'tcx>,
        imm: i8,
        intrinsic: Symbol,
    ) -> InterpResult<'tcx, Self> {
        // Only bits 0..=4 are used, remaining should be zero.
        if imm & !0b1_1111 != 0 {
            panic!("invalid `imm` parameter of {intrinsic}: 0x{imm:x}");
        }
        // Bit 4 specifies whether the operation is quiet or signaling, which
        // we do not care in Miri.
        // Bits 0..=2 specifies the operation.
        // `gt` indicates the result to be returned when the LHS is strictly
        // greater than the RHS, and so on.
        let (gt, lt, eq, mut unord) = match imm & 0b111 {
            // Equal
            0x0 => (false, false, true, false),
            // Less-than
            0x1 => (false, true, false, false),
            // Less-or-equal
            0x2 => (false, true, true, false),
            // Unordered (either is NaN)
            0x3 => (false, false, false, true),
            // Not equal
            0x4 => (true, true, false, true),
            // Not less-than
            0x5 => (true, false, true, true),
            // Not less-or-equal
            0x6 => (true, false, false, true),
            // Ordered (neither is NaN)
            0x7 => (true, true, true, false),
            _ => unreachable!(),
        };
        // When bit 3 is 1 (only possible in AVX), unord is toggled.
        if imm & 0b1000 != 0 {
            ecx.expect_target_feature_for_intrinsic(intrinsic, "avx")?;
            unord = !unord;
        }
        interp_ok(Self::Cmp { gt, lt, eq, unord })
    }
}

/// Performs `which` scalar operation on `left` and `right` and returns
/// the result.
fn bin_op_float<'tcx, F: rustc_apfloat::Float>(
    which: FloatBinOp,
    left: &ImmTy<'tcx>,
    right: &ImmTy<'tcx>,
) -> InterpResult<'tcx, Scalar> {
    match which {
        FloatBinOp::Cmp { gt, lt, eq, unord } => {
            let left = left.to_scalar().to_float::<F>()?;
            let right = right.to_scalar().to_float::<F>()?;

            let res = match left.partial_cmp(&right) {
                None => unord,
                Some(std::cmp::Ordering::Less) => lt,
                Some(std::cmp::Ordering::Equal) => eq,
                Some(std::cmp::Ordering::Greater) => gt,
            };
            interp_ok(bool_to_simd_element(res, Size::from_bits(F::BITS)))
        }
        FloatBinOp::Min => {
            let left_scalar = left.to_scalar();
            let left = left_scalar.to_float::<F>()?;
            let right_scalar = right.to_scalar();
            let right = right_scalar.to_float::<F>()?;
            // SSE semantics to handle zero and NaN. Note that `x == F::ZERO`
            // is true when `x` is either +0 or -0.
            if (left == F::ZERO && right == F::ZERO)
                || left.is_nan()
                || right.is_nan()
                || left >= right
            {
                interp_ok(right_scalar)
            } else {
                interp_ok(left_scalar)
            }
        }
        FloatBinOp::Max => {
            let left_scalar = left.to_scalar();
            let left = left_scalar.to_float::<F>()?;
            let right_scalar = right.to_scalar();
            let right = right_scalar.to_float::<F>()?;
            // SSE semantics to handle zero and NaN. Note that `x == F::ZERO`
            // is true when `x` is either +0 or -0.
            if (left == F::ZERO && right == F::ZERO)
                || left.is_nan()
                || right.is_nan()
                || left <= right
            {
                interp_ok(right_scalar)
            } else {
                interp_ok(left_scalar)
            }
        }
    }
}

/// Performs `which` operation on the first component of `left` and `right`
/// and copies the other components from `left`. The result is stored in `dest`.
fn bin_op_simd_float_first<'tcx, F: rustc_apfloat::Float>(
    ecx: &mut crate::MiriInterpCx<'tcx>,
    which: FloatBinOp,
    left: &OpTy<'tcx>,
    right: &OpTy<'tcx>,
    dest: &MPlaceTy<'tcx>,
) -> InterpResult<'tcx, ()> {
    let (left, left_len) = ecx.project_to_simd(left)?;
    let (right, right_len) = ecx.project_to_simd(right)?;
    let (dest, dest_len) = ecx.project_to_simd(dest)?;

    assert_eq!(dest_len, left_len);
    assert_eq!(dest_len, right_len);

    let res0 = bin_op_float::<F>(
        which,
        &ecx.read_immediate(&ecx.project_index(&left, 0)?)?,
        &ecx.read_immediate(&ecx.project_index(&right, 0)?)?,
    )?;
    ecx.write_scalar(res0, &ecx.project_index(&dest, 0)?)?;

    for i in 1..dest_len {
        ecx.copy_op(&ecx.project_index(&left, i)?, &ecx.project_index(&dest, i)?)?;
    }

    interp_ok(())
}

/// Performs `which` operation on each component of `left` and
/// `right`, storing the result is stored in `dest`.
fn bin_op_simd_float_all<'tcx, F: rustc_apfloat::Float>(
    ecx: &mut crate::MiriInterpCx<'tcx>,
    which: FloatBinOp,
    left: &OpTy<'tcx>,
    right: &OpTy<'tcx>,
    dest: &MPlaceTy<'tcx>,
) -> InterpResult<'tcx, ()> {
    let (left, left_len) = ecx.project_to_simd(left)?;
    let (right, right_len) = ecx.project_to_simd(right)?;
    let (dest, dest_len) = ecx.project_to_simd(dest)?;

    assert_eq!(dest_len, left_len);
    assert_eq!(dest_len, right_len);

    for i in 0..dest_len {
        let left = ecx.read_immediate(&ecx.project_index(&left, i)?)?;
        let right = ecx.read_immediate(&ecx.project_index(&right, i)?)?;
        let dest = ecx.project_index(&dest, i)?;

        let res = bin_op_float::<F>(which, &left, &right)?;
        ecx.write_scalar(res, &dest)?;
    }

    interp_ok(())
}

#[derive(Copy, Clone)]
enum FloatUnaryOp {
    /// Approximation of 1/x
    ///
    /// <https://www.felixcloutier.com/x86/rcpss>
    /// <https://www.felixcloutier.com/x86/rcpps>
    Rcp,
    /// Approximation of 1/sqrt(x)
    ///
    /// <https://www.felixcloutier.com/x86/rsqrtss>
    /// <https://www.felixcloutier.com/x86/rsqrtps>
    Rsqrt,
}

/// Performs `which` scalar operation on `op` and returns the result.
fn unary_op_f32<'tcx>(
    ecx: &mut crate::MiriInterpCx<'tcx>,
    which: FloatUnaryOp,
    op: &ImmTy<'tcx>,
) -> InterpResult<'tcx, Scalar> {
    match which {
        FloatUnaryOp::Rcp => {
            let op = op.to_scalar().to_f32()?;
            let div = (Single::from_u128(1).value / op).value;
            // Apply a relative error with a magnitude on the order of 2^-12 to simulate the
            // inaccuracy of RCP.
            let res = math::apply_random_float_error(ecx, div, -12);
            interp_ok(Scalar::from_f32(res))
        }
        FloatUnaryOp::Rsqrt => {
            let op = op.to_scalar().to_f32()?;
            let rsqrt = (Single::from_u128(1).value / math::sqrt(op)).value;
            // Apply a relative error with a magnitude on the order of 2^-12 to simulate the
            // inaccuracy of RSQRT.
            let res = math::apply_random_float_error(ecx, rsqrt, -12);
            interp_ok(Scalar::from_f32(res))
        }
    }
}

/// Performs `which` operation on the first component of `op` and copies
/// the other components. The result is stored in `dest`.
fn unary_op_ss<'tcx>(
    ecx: &mut crate::MiriInterpCx<'tcx>,
    which: FloatUnaryOp,
    op: &OpTy<'tcx>,
    dest: &MPlaceTy<'tcx>,
) -> InterpResult<'tcx, ()> {
    let (op, op_len) = ecx.project_to_simd(op)?;
    let (dest, dest_len) = ecx.project_to_simd(dest)?;

    assert_eq!(dest_len, op_len);

    let res0 = unary_op_f32(ecx, which, &ecx.read_immediate(&ecx.project_index(&op, 0)?)?)?;
    ecx.write_scalar(res0, &ecx.project_index(&dest, 0)?)?;

    for i in 1..dest_len {
        ecx.copy_op(&ecx.project_index(&op, i)?, &ecx.project_index(&dest, i)?)?;
    }

    interp_ok(())
}

/// Performs `which` operation on each component of `op`, storing the
/// result is stored in `dest`.
fn unary_op_ps<'tcx>(
    ecx: &mut crate::MiriInterpCx<'tcx>,
    which: FloatUnaryOp,
    op: &OpTy<'tcx>,
    dest: &MPlaceTy<'tcx>,
) -> InterpResult<'tcx, ()> {
    let (op, op_len) = ecx.project_to_simd(op)?;
    let (dest, dest_len) = ecx.project_to_simd(dest)?;

    assert_eq!(dest_len, op_len);

    for i in 0..dest_len {
        let op = ecx.read_immediate(&ecx.project_index(&op, i)?)?;
        let dest = ecx.project_index(&dest, i)?;

        let res = unary_op_f32(ecx, which, &op)?;
        ecx.write_scalar(res, &dest)?;
    }

    interp_ok(())
}

enum ShiftOp {
    /// Shift left, logically (shift in zeros) -- same as shift left, arithmetically
    Left,
    /// Shift right, logically (shift in zeros)
    RightLogic,
    /// Shift right, arithmetically (shift in sign)
    RightArith,
}

/// Shifts each element of `left` by a scalar amount. The shift amount
/// is determined by the lowest 64 bits of `right` (which is a 128-bit vector).
///
/// For logic shifts, when right is larger than BITS - 1, zero is produced.
/// For arithmetic right-shifts, when right is larger than BITS - 1, the sign
/// bit is copied to all bits.
fn shift_simd_by_scalar<'tcx>(
    ecx: &mut crate::MiriInterpCx<'tcx>,
    left: &OpTy<'tcx>,
    right: &OpTy<'tcx>,
    which: ShiftOp,
    dest: &MPlaceTy<'tcx>,
) -> InterpResult<'tcx, ()> {
    let (left, left_len) = ecx.project_to_simd(left)?;
    let (dest, dest_len) = ecx.project_to_simd(dest)?;

    assert_eq!(dest_len, left_len);
    // `right` may have a different length, and we only care about its
    // lowest 64bit anyway.

    // Get the 64-bit shift operand and convert it to the type expected
    // by checked_{shl,shr} (u32).
    // It is ok to saturate the value to u32::MAX because any value
    // above BITS - 1 will produce the same result.
    let shift = u32::try_from(extract_first_u64(ecx, right)?).unwrap_or(u32::MAX);

    for i in 0..dest_len {
        let left = ecx.read_scalar(&ecx.project_index(&left, i)?)?;
        let dest = ecx.project_index(&dest, i)?;

        let res = match which {
            ShiftOp::Left => {
                let left = left.to_uint(dest.layout.size)?;
                let res = left.checked_shl(shift).unwrap_or(0);
                // `truncate` is needed as left-shift can make the absolute value larger.
                Scalar::from_uint(dest.layout.size.truncate(res), dest.layout.size)
            }
            ShiftOp::RightLogic => {
                let left = left.to_uint(dest.layout.size)?;
                let res = left.checked_shr(shift).unwrap_or(0);
                // No `truncate` needed as right-shift can only make the absolute value smaller.
                Scalar::from_uint(res, dest.layout.size)
            }
            ShiftOp::RightArith => {
                let left = left.to_int(dest.layout.size)?;
                // On overflow, copy the sign bit to the remaining bits
                let res = left.checked_shr(shift).unwrap_or(left >> 127);
                // No `truncate` needed as right-shift can only make the absolute value smaller.
                Scalar::from_int(res, dest.layout.size)
            }
        };
        ecx.write_scalar(res, &dest)?;
    }

    interp_ok(())
}

fn shift_simd_by_simd<'tcx>(
    ecx: &mut crate::MiriInterpCx<'tcx>,
    left: &OpTy<'tcx>,
    right: &OpTy<'tcx>,
    which: ShiftOp,
    dest: &MPlaceTy<'tcx>,
) -> InterpResult<'tcx, ()> {
    let (left, left_len) = ecx.project_to_simd(left)?;
    let (right, right_len) = ecx.project_to_simd(right)?;
    let (dest, dest_len) = ecx.project_to_simd(dest)?;

    assert_eq!(dest_len, left_len);
    assert_eq!(dest_len, right_len);

    for i in 0..dest_len {
        let left = ecx.read_scalar(&ecx.project_index(&left, i)?)?;
        let right = ecx.read_scalar(&ecx.project_index(&right, i)?)?;
        let dest = ecx.project_index(&dest, i)?;

        // It is ok to saturate the value to u32::MAX because any value
        // above BITS - 1 will produce the same result.
        let shift = u32::try_from(right.to_uint(dest.layout.size)?).unwrap_or(u32::MAX);

        let res = match which {
            ShiftOp::Left => {
                let left = left.to_uint(dest.layout.size)?;
                let res = left.checked_shl(shift).unwrap_or(0);
                // `truncate` is needed as left-shift can make the absolute value larger.
                Scalar::from_uint(dest.layout.size.truncate(res), dest.layout.size)
            }
            ShiftOp::RightLogic => {
                let left = left.to_uint(dest.layout.size)?;
                let res = left.checked_shr(shift).unwrap_or(0);
                // No `truncate` needed as right-shift can only make the absolute value smaller.
                Scalar::from_uint(res, dest.layout.size)
            }
            ShiftOp::RightArith => {
                let left = left.to_int(dest.layout.size)?;
                // On overflow, copy the sign bit to the remaining bits
                let res = left.checked_shr(shift).unwrap_or(left >> 127);
                // No `truncate` needed as right-shift can only make the absolute value smaller.
                Scalar::from_int(res, dest.layout.size)
            }
        };
        ecx.write_scalar(res, &dest)?;
    }

    interp_ok(())
}

/// Takes a 128-bit vector, transmutes it to `[u64; 2]` and extracts
/// the first value.
fn extract_first_u64<'tcx>(
    ecx: &crate::MiriInterpCx<'tcx>,
    op: &OpTy<'tcx>,
) -> InterpResult<'tcx, u64> {
    // Transmute vector to `[u64; 2]`
    let array_layout = ecx.layout_of(Ty::new_array(ecx.tcx.tcx, ecx.tcx.types.u64, 2))?;
    let op = op.transmute(array_layout, ecx)?;

    // Get the first u64 from the array
    ecx.read_scalar(&ecx.project_index(&op, 0)?)?.to_u64()
}

// Rounds the first element of `right` according to `rounding`
// and copies the remaining elements from `left`.
fn round_first<'tcx, F: rustc_apfloat::Float>(
    ecx: &mut crate::MiriInterpCx<'tcx>,
    left: &OpTy<'tcx>,
    right: &OpTy<'tcx>,
    rounding: &OpTy<'tcx>,
    dest: &MPlaceTy<'tcx>,
) -> InterpResult<'tcx, ()> {
    let (left, left_len) = ecx.project_to_simd(left)?;
    let (right, right_len) = ecx.project_to_simd(right)?;
    let (dest, dest_len) = ecx.project_to_simd(dest)?;

    assert_eq!(dest_len, left_len);
    assert_eq!(dest_len, right_len);

    let rounding = rounding_from_imm(ecx.read_scalar(rounding)?.to_i32()?)?;

    let op0: F = ecx.read_scalar(&ecx.project_index(&right, 0)?)?.to_float()?;
    let res = op0.round_to_integral(rounding).value;
    ecx.write_scalar(
        Scalar::from_uint(res.to_bits(), Size::from_bits(F::BITS)),
        &ecx.project_index(&dest, 0)?,
    )?;

    for i in 1..dest_len {
        ecx.copy_op(&ecx.project_index(&left, i)?, &ecx.project_index(&dest, i)?)?;
    }

    interp_ok(())
}

// Rounds all elements of `op` according to `rounding`.
fn round_all<'tcx, F: rustc_apfloat::Float>(
    ecx: &mut crate::MiriInterpCx<'tcx>,
    op: &OpTy<'tcx>,
    rounding: &OpTy<'tcx>,
    dest: &MPlaceTy<'tcx>,
) -> InterpResult<'tcx, ()> {
    let (op, op_len) = ecx.project_to_simd(op)?;
    let (dest, dest_len) = ecx.project_to_simd(dest)?;

    assert_eq!(dest_len, op_len);

    let rounding = rounding_from_imm(ecx.read_scalar(rounding)?.to_i32()?)?;

    for i in 0..dest_len {
        let op: F = ecx.read_scalar(&ecx.project_index(&op, i)?)?.to_float()?;
        let res = op.round_to_integral(rounding).value;
        ecx.write_scalar(
            Scalar::from_uint(res.to_bits(), Size::from_bits(F::BITS)),
            &ecx.project_index(&dest, i)?,
        )?;
    }

    interp_ok(())
}

/// Gets equivalent `rustc_apfloat::Round` from rounding mode immediate of
/// `round.{ss,sd,ps,pd}` intrinsics.
fn rounding_from_imm<'tcx>(rounding: i32) -> InterpResult<'tcx, rustc_apfloat::Round> {
    // The fourth bit of `rounding` only affects the SSE status
    // register, which cannot be accessed from Miri (or from Rust,
    // for that matter), so we can ignore it.
    match rounding & !0b1000 {
        // When the third bit is 0, the rounding mode is determined by the
        // first two bits.
        0b000 => interp_ok(rustc_apfloat::Round::NearestTiesToEven),
        0b001 => interp_ok(rustc_apfloat::Round::TowardNegative),
        0b010 => interp_ok(rustc_apfloat::Round::TowardPositive),
        0b011 => interp_ok(rustc_apfloat::Round::TowardZero),
        // When the third bit is 1, the rounding mode is determined by the
        // SSE status register. Since we do not support modifying it from
        // Miri (or Rust), we assume it to be at its default mode (round-to-nearest).
        0b100..=0b111 => interp_ok(rustc_apfloat::Round::NearestTiesToEven),
        rounding => panic!("invalid rounding mode 0x{rounding:02x}"),
    }
}

/// Converts each element of `op` from floating point to signed integer.
///
/// When the input value is NaN or out of range, fall back to minimum value.
///
/// If `op` has more elements than `dest`, extra elements are ignored. If `op`
/// has less elements than `dest`, the rest is filled with zeros.
fn convert_float_to_int<'tcx>(
    ecx: &mut crate::MiriInterpCx<'tcx>,
    op: &OpTy<'tcx>,
    rnd: rustc_apfloat::Round,
    dest: &MPlaceTy<'tcx>,
) -> InterpResult<'tcx, ()> {
    let (op, op_len) = ecx.project_to_simd(op)?;
    let (dest, dest_len) = ecx.project_to_simd(dest)?;

    // Output must be *signed* integers.
    assert!(matches!(dest.layout.field(ecx, 0).ty.kind(), ty::Int(_)));

    for i in 0..op_len.min(dest_len) {
        let op = ecx.read_immediate(&ecx.project_index(&op, i)?)?;
        let dest = ecx.project_index(&dest, i)?;

        let res = ecx.float_to_int_checked(&op, dest.layout, rnd)?.unwrap_or_else(|| {
            // Fallback to minimum according to SSE/AVX semantics.
            ImmTy::from_int(dest.layout.size.signed_int_min(), dest.layout)
        });
        ecx.write_immediate(*res, &dest)?;
    }
    // Fill remainder with zeros
    for i in op_len..dest_len {
        let dest = ecx.project_index(&dest, i)?;
        ecx.write_scalar(Scalar::from_int(0, dest.layout.size), &dest)?;
    }

    interp_ok(())
}

/// Splits `op` (which must be a SIMD vector) into 128-bit chunks.
///
/// Returns a tuple where:
/// * The first element is the number of 128-bit chunks (let's call it `N`).
/// * The second element is the number of elements per chunk (let's call it `M`).
/// * The third element is the `op` vector split into chunks, i.e, it's
///   type is `[[T; M]; N]` where `T` is the element type of `op`.
fn split_simd_to_128bit_chunks<'tcx, P: Projectable<'tcx, Provenance>>(
    ecx: &mut crate::MiriInterpCx<'tcx>,
    op: &P,
) -> InterpResult<'tcx, (u64, u64, P)> {
    let simd_layout = op.layout();
    let (simd_len, element_ty) = simd_layout.ty.simd_size_and_type(ecx.tcx.tcx);

    assert_eq!(simd_layout.size.bits() % 128, 0);
    let num_chunks = simd_layout.size.bits() / 128;
    let items_per_chunk = simd_len.strict_div(num_chunks);

    // Transmute to `[[T; items_per_chunk]; num_chunks]`
    let chunked_layout = ecx
        .layout_of(Ty::new_array(
            ecx.tcx.tcx,
            Ty::new_array(ecx.tcx.tcx, element_ty, items_per_chunk),
            num_chunks,
        ))
        .unwrap();
    let chunked_op = op.transmute(chunked_layout, ecx)?;

    interp_ok((num_chunks, items_per_chunk, chunked_op))
}

/// Conditionally multiplies the packed floating-point elements in
/// `left` and `right` using the high 4 bits in `imm`, sums the calculated
/// products (up to 4), and conditionally stores the sum in `dest` using
/// the low 4 bits of `imm`.
///
/// Each 128-bit chunk is treated independently (i.e., the value for
/// the is i-th 128-bit chunk of `dest` is calculated with the i-th
/// 128-bit blocks of `left` and `right`).
fn conditional_dot_product<'tcx>(
    ecx: &mut crate::MiriInterpCx<'tcx>,
    left: &OpTy<'tcx>,
    right: &OpTy<'tcx>,
    imm: &OpTy<'tcx>,
    dest: &MPlaceTy<'tcx>,
) -> InterpResult<'tcx, ()> {
    assert_eq!(left.layout, dest.layout);
    assert_eq!(right.layout, dest.layout);

    let (num_chunks, items_per_chunk, left) = split_simd_to_128bit_chunks(ecx, left)?;
    let (_, _, right) = split_simd_to_128bit_chunks(ecx, right)?;
    let (_, _, dest) = split_simd_to_128bit_chunks(ecx, dest)?;

    let element_layout = left.layout.field(ecx, 0).field(ecx, 0);
    assert!(items_per_chunk <= 4);

    // `imm` is a `u8` for SSE4.1 or an `i32` for AVX :/
    let imm = ecx.read_scalar(imm)?.to_uint(imm.layout.size)?;

    for i in 0..num_chunks {
        let left = ecx.project_index(&left, i)?;
        let right = ecx.project_index(&right, i)?;
        let dest = ecx.project_index(&dest, i)?;

        // Calculate dot product
        // Elements are floating point numbers, but we can use `from_int`
        // for the initial value because the representation of 0.0 is all zero bits.
        let mut sum = ImmTy::from_int(0u8, element_layout);
        for j in 0..items_per_chunk {
            if imm & (1 << j.strict_add(4)) != 0 {
                let left = ecx.read_immediate(&ecx.project_index(&left, j)?)?;
                let right = ecx.read_immediate(&ecx.project_index(&right, j)?)?;

                let mul = ecx.binary_op(mir::BinOp::Mul, &left, &right)?;
                sum = ecx.binary_op(mir::BinOp::Add, &sum, &mul)?;
            }
        }

        // Write to destination (conditioned to imm)
        for j in 0..items_per_chunk {
            let dest = ecx.project_index(&dest, j)?;

            if imm & (1 << j) != 0 {
                ecx.write_immediate(*sum, &dest)?;
            } else {
                ecx.write_scalar(Scalar::from_int(0u8, element_layout.size), &dest)?;
            }
        }
    }

    interp_ok(())
}

/// Calculates two booleans.
///
/// The first is true when all the bits of `op & mask` are zero.
/// The second is true when `(op & mask) == mask`
fn test_bits_masked<'tcx>(
    ecx: &crate::MiriInterpCx<'tcx>,
    op: &OpTy<'tcx>,
    mask: &OpTy<'tcx>,
) -> InterpResult<'tcx, (bool, bool)> {
    assert_eq!(op.layout, mask.layout);

    let (op, op_len) = ecx.project_to_simd(op)?;
    let (mask, mask_len) = ecx.project_to_simd(mask)?;

    assert_eq!(op_len, mask_len);

    let mut all_zero = true;
    let mut masked_set = true;
    for i in 0..op_len {
        let op = ecx.project_index(&op, i)?;
        let mask = ecx.project_index(&mask, i)?;

        let op = ecx.read_scalar(&op)?.to_uint(op.layout.size)?;
        let mask = ecx.read_scalar(&mask)?.to_uint(mask.layout.size)?;
        all_zero &= (op & mask) == 0;
        masked_set &= (op & mask) == mask;
    }

    interp_ok((all_zero, masked_set))
}

/// Calculates two booleans.
///
/// The first is true when the highest bit of each element of `op & mask` is zero.
/// The second is true when the highest bit of each element of `!op & mask` is zero.
fn test_high_bits_masked<'tcx>(
    ecx: &crate::MiriInterpCx<'tcx>,
    op: &OpTy<'tcx>,
    mask: &OpTy<'tcx>,
) -> InterpResult<'tcx, (bool, bool)> {
    assert_eq!(op.layout, mask.layout);

    let (op, op_len) = ecx.project_to_simd(op)?;
    let (mask, mask_len) = ecx.project_to_simd(mask)?;

    assert_eq!(op_len, mask_len);

    let high_bit_offset = op.layout.field(ecx, 0).size.bits().strict_sub(1);

    let mut direct = true;
    let mut negated = true;
    for i in 0..op_len {
        let op = ecx.project_index(&op, i)?;
        let mask = ecx.project_index(&mask, i)?;

        let op = ecx.read_scalar(&op)?.to_uint(op.layout.size)?;
        let mask = ecx.read_scalar(&mask)?.to_uint(mask.layout.size)?;
        direct &= (op & mask) >> high_bit_offset == 0;
        negated &= (!op & mask) >> high_bit_offset == 0;
    }

    interp_ok((direct, negated))
}

/// Compute the sum of absolute differences of quadruplets of unsigned
/// 8-bit integers in `left` and `right`, and store the 16-bit results
/// in `right`. Quadruplets are selected from `left` and `right` with
/// offsets specified in `imm`.
///
/// <https://www.intel.com/content/www/us/en/docs/intrinsics-guide/index.html#text=_mm_maddubs_epi16>
/// <https://www.intel.com/content/www/us/en/docs/intrinsics-guide/index.html#text=_mm256_mpsadbw_epu8>
///
/// Each 128-bit chunk is treated independently (i.e., the value for
/// the is i-th 128-bit chunk of `dest` is calculated with the i-th
/// 128-bit chunks of `left` and `right`).
fn mpsadbw<'tcx>(
    ecx: &mut crate::MiriInterpCx<'tcx>,
    left: &OpTy<'tcx>,
    right: &OpTy<'tcx>,
    imm: &OpTy<'tcx>,
    dest: &MPlaceTy<'tcx>,
) -> InterpResult<'tcx, ()> {
    assert_eq!(left.layout, right.layout);
    assert_eq!(left.layout.size, dest.layout.size);

    let (num_chunks, op_items_per_chunk, left) = split_simd_to_128bit_chunks(ecx, left)?;
    assert!(num_chunks <= 2);

    let (_, _, right) = split_simd_to_128bit_chunks(ecx, right)?;
    let (_, dest_items_per_chunk, dest) = split_simd_to_128bit_chunks(ecx, dest)?;

    assert_eq!(op_items_per_chunk, dest_items_per_chunk.strict_mul(2));

    let imm = ecx.read_scalar(imm)?.to_uint(imm.layout.size)?;

    for i in 0..num_chunks {
        let left = ecx.project_index(&left, i)?;
        let right = ecx.project_index(&right, i)?;
        let dest = ecx.project_index(&dest, i)?;

        // The first 128-bit chunk uses the low 3 bits of IMM, the second chunk uses bits 3..6.
        let lane_imm = imm.strict_shr(i.strict_mul(3).try_into().unwrap());

        // Bit 2 of `lane_imm` specifies the offset for indices of `left`.
        // The offset is 0 when the bit is 0 or 4 when the bit is 1.
        let left_base = u64::try_from((lane_imm >> 2) & 1).unwrap().strict_mul(4);
        // Bits 0..=1 of `lane_imm` specify the offset for indices of
        // `right` in blocks of 4 elements.
        let right_base = u64::try_from(lane_imm & 0b11).unwrap().strict_mul(4);

        for j in 0..dest_items_per_chunk {
            let left_offset = left_base.strict_add(j);
            let mut res: u16 = 0;
            for k in 0..4 {
                let left = ecx
                    .read_scalar(&ecx.project_index(&left, left_offset.strict_add(k))?)?
                    .to_u8()?;
                let right = ecx
                    .read_scalar(&ecx.project_index(&right, right_base.strict_add(k))?)?
                    .to_u8()?;
                res = res.strict_add(left.abs_diff(right).into());
            }
            ecx.write_scalar(Scalar::from_u16(res), &ecx.project_index(&dest, j)?)?;
        }
    }

    interp_ok(())
}

/// Compute the absolute differences of packed unsigned 8-bit integers
/// in `left` and `right`, then horizontally sum each consecutive 8
/// differences to produce unsigned 16-bit integers, and pack
/// these unsigned 16-bit integers in the low 16 bits of 64-bit elements
/// in `dest`.
///
/// <https://www.intel.com/content/www/us/en/docs/intrinsics-guide/index.html#text=_mm_sad_epu8>
/// <https://www.intel.com/content/www/us/en/docs/intrinsics-guide/index.html#text=_mm256_sad_epu8>
/// <https://www.intel.com/content/www/us/en/docs/intrinsics-guide/index.html#text=_mm512_sad_epu8>
fn psadbw<'tcx>(
    ecx: &mut crate::MiriInterpCx<'tcx>,
    left: &OpTy<'tcx>,
    right: &OpTy<'tcx>,
    dest: &MPlaceTy<'tcx>,
) -> InterpResult<'tcx, ()> {
    let (left, left_len) = ecx.project_to_simd(left)?;
    let (right, right_len) = ecx.project_to_simd(right)?;
    let (dest, dest_len) = ecx.project_to_simd(dest)?;

    // fn psadbw(a: u8x16, b: u8x16) -> u64x2;
    // fn psadbw(a: u8x32, b: u8x32) -> u64x4;
    // fn vpsadbw(a: u8x64, b: u8x64) -> u64x8;
    assert_eq!(left_len, right_len);
    assert_eq!(left_len, left.layout.layout.size().bytes());
    assert_eq!(dest_len, left_len.strict_div(8));

    for i in 0..dest_len {
        let dest = ecx.project_index(&dest, i)?;

        let mut acc: u16 = 0;
        for j in 0..8 {
            let src_index = i.strict_mul(8).strict_add(j);

            let left = ecx.project_index(&left, src_index)?;
            let left = ecx.read_scalar(&left)?.to_u8()?;

            let right = ecx.project_index(&right, src_index)?;
            let right = ecx.read_scalar(&right)?.to_u8()?;

            acc = acc.strict_add(left.abs_diff(right).into());
        }

        ecx.write_scalar(Scalar::from_u64(acc.into()), &dest)?;
    }

    interp_ok(())
}

/// Multiply packed signed 16-bit integers in `left` and `right`, producing intermediate signed 32-bit integers.
/// Horizontally add adjacent pairs of intermediate 32-bit integers, and pack the results in `dest`.
///
/// <https://www.intel.com/content/www/us/en/docs/intrinsics-guide/index.html#text=_mm_madd_epi16>
/// <https://www.intel.com/content/www/us/en/docs/intrinsics-guide/index.html#text=_mm256_madd_epi16>
/// <https://www.intel.com/content/www/us/en/docs/intrinsics-guide/index.html#text=_mm512_madd_epi16>
fn pmaddwd<'tcx>(
    ecx: &mut crate::MiriInterpCx<'tcx>,
    left: &OpTy<'tcx>,
    right: &OpTy<'tcx>,
    dest: &MPlaceTy<'tcx>,
) -> InterpResult<'tcx, ()> {
    let (left, left_len) = ecx.project_to_simd(left)?;
    let (right, right_len) = ecx.project_to_simd(right)?;
    let (dest, dest_len) = ecx.project_to_simd(dest)?;

    // fn  pmaddwd(a: i16x8,  b: i16x8)  -> i32x4;
    // fn  pmaddwd(a: i16x16, b: i16x16) -> i32x8;
    // fn vpmaddwd(a: i16x32, b: i16x32) -> i32x16;
    assert_eq!(left_len, right_len);
    assert_eq!(dest_len.strict_mul(2), left_len);

    for i in 0..dest_len {
        let j1 = i.strict_mul(2);
        let left1 = ecx.read_scalar(&ecx.project_index(&left, j1)?)?.to_i16()?;
        let right1 = ecx.read_scalar(&ecx.project_index(&right, j1)?)?.to_i16()?;

        let j2 = j1.strict_add(1);
        let left2 = ecx.read_scalar(&ecx.project_index(&left, j2)?)?.to_i16()?;
        let right2 = ecx.read_scalar(&ecx.project_index(&right, j2)?)?.to_i16()?;

        let dest = ecx.project_index(&dest, i)?;

        // Multiplications are i16*i16->i32, which will not overflow.
        let mul1 = i32::from(left1).strict_mul(right1.into());
        let mul2 = i32::from(left2).strict_mul(right2.into());
        // However, this addition can overflow in the most extreme case
        // (-0x8000)*(-0x8000)+(-0x8000)*(-0x8000) = 0x80000000
        let res = mul1.wrapping_add(mul2);

        ecx.write_scalar(Scalar::from_i32(res), &dest)?;
    }

    interp_ok(())
}

/// Multiplies packed 8-bit unsigned integers from `left` and packed
/// signed 8-bit integers from `right` into 16-bit signed integers. Then,
/// the saturating sum of the products with indices `2*i` and `2*i+1`
/// produces the output at index `i`.
///
/// <https://www.intel.com/content/www/us/en/docs/intrinsics-guide/index.html#text=_mm_maddubs_epi16>
/// <https://www.intel.com/content/www/us/en/docs/intrinsics-guide/index.html#text=_mm256_maddubs_epi16>
/// <https://www.intel.com/content/www/us/en/docs/intrinsics-guide/index.html#text=_mm512_maddubs_epi16>
fn pmaddbw<'tcx>(
    ecx: &mut crate::MiriInterpCx<'tcx>,
    left: &OpTy<'tcx>,
    right: &OpTy<'tcx>,
    dest: &MPlaceTy<'tcx>,
) -> InterpResult<'tcx, ()> {
    let (left, left_len) = ecx.project_to_simd(left)?;
    let (right, right_len) = ecx.project_to_simd(right)?;
    let (dest, dest_len) = ecx.project_to_simd(dest)?;

    // fn pmaddubsw128(a: u8x16, b: i8x16) -> i16x8;
    // fn pmaddubsw(   a: u8x32, b: i8x32) -> i16x16;
    // fn vpmaddubsw(  a: u8x64, b: i8x64) -> i16x32;
    assert_eq!(left_len, right_len);
    assert_eq!(dest_len.strict_mul(2), left_len);

    for i in 0..dest_len {
        let j1 = i.strict_mul(2);
        let left1 = ecx.read_scalar(&ecx.project_index(&left, j1)?)?.to_u8()?;
        let right1 = ecx.read_scalar(&ecx.project_index(&right, j1)?)?.to_i8()?;

        let j2 = j1.strict_add(1);
        let left2 = ecx.read_scalar(&ecx.project_index(&left, j2)?)?.to_u8()?;
        let right2 = ecx.read_scalar(&ecx.project_index(&right, j2)?)?.to_i8()?;

        let dest = ecx.project_index(&dest, i)?;

        // Multiplication of a u8 and an i8 into an i16 cannot overflow.
        let mul1 = i16::from(left1).strict_mul(right1.into());
        let mul2 = i16::from(left2).strict_mul(right2.into());
        let res = mul1.saturating_add(mul2);

        ecx.write_scalar(Scalar::from_i16(res), &dest)?;
    }

    interp_ok(())
}

/// Shuffle elements in `values` across lanes using the corresponding index in
/// `indices`, and store the results in `dest`.
///
/// This helper is shared by both the 32-bit-lane and 64-bit-lane AVX
/// permute-by-index intrinsics. The element type is taken from `values` and
/// `dest`, while the index lanes are interpreted at their full width (`i32` or
/// `i64`, depending on the intrinsic).
///
/// For a vector with `N` lanes, only the low `log2(N)` bits of each index are
/// used. Equivalently, lane `i` of the result is copied from
/// `values[indices[i] & (N - 1)]`.
///
/// <https://www.intel.com/content/www/us/en/docs/intrinsics-guide/index.html#text=_mm256_permutevar8x32_epi32>
/// <https://www.intel.com/content/www/us/en/docs/intrinsics-guide/index.html#text=_mm256_permutevar8x32_ps>
/// <https://www.intel.com/content/www/us/en/docs/intrinsics-guide/index.html#text=_mm512_permutexvar_epi32>
/// <https://www.intel.com/content/www/us/en/docs/intrinsics-guide/index.html#text=_mm512_permutexvar_epi64>
fn permute<'tcx>(
    ecx: &mut crate::MiriInterpCx<'tcx>,
    values: &OpTy<'tcx>,
    indices: &OpTy<'tcx>,
    dest: &MPlaceTy<'tcx>,
) -> InterpResult<'tcx, ()> {
    let (values, values_len) = ecx.project_to_simd(values)?;
    let (indices, indices_len) = ecx.project_to_simd(indices)?;
    let (dest, dest_len) = ecx.project_to_simd(dest)?;

    // fn permd(a: u32x8, b: u32x8) -> u32x8;
    // fn permps(a: __m256, b: i32x8) -> __m256;
    // fn vpermd(a: i32x16, idx: i32x16) -> i32x16;
    // fn vpermq(a: i64x8, b: i64x8) -> i64x8;
    assert_eq!(dest_len, values_len);
    assert_eq!(dest_len, indices_len);

    // Only use the lower 3 bits to index into a vector with 8 lanes,
    // or the lower 4 bits when indexing into a 16-lane vector.
    assert!(dest_len.is_power_of_two());
    let mask = u128::from(dest_len).strict_sub(1);

    for i in 0..dest_len {
        let dest = ecx.project_index(&dest, i)?;
        let index_place = ecx.project_index(&indices, i)?;
        let index = ecx.read_scalar(&index_place)?.to_uint(index_place.layout.size)?;
        // `mask` is at most `dest_len - 1` which fits in a `u64`, so this cannot fail.
        let element = ecx.project_index(&values, u64::try_from(index & mask).unwrap())?;

        ecx.copy_op(&element, &dest)?;
    }

    interp_ok(())
}

/// Shuffle elements from *two* source registers (`left` and `right`) using
/// the corresponding index in `indices`, and store the results in `dest`.
///
/// For indexing, we basically concatenate `left` and `right`, and index into the concatenation.
/// More precisely: For a vector with `N` lanes, the low `log2(N)` bits of each index select a
/// lane within a source vector. Bit `log2(N)` selects the source vector (`0` =>
/// `left`, `1` => `right`), and all higher bits are ignored.
/// Equivalently, lane `i` of the result is copied from
/// `src[indices[i] & (N - 1)]` where
/// `src = if indices[i] & N == 0 { left } else { right }`.
///
/// <https://www.intel.com/content/www/us/en/docs/intrinsics-guide/index.html#text=_mm512_permutex2var_epi64>
/// <https://www.intel.com/content/www/us/en/docs/intrinsics-guide/index.html#text=_mm512_permutex2var_epi8>
fn permute2<'tcx>(
    ecx: &mut crate::MiriInterpCx<'tcx>,
    left: &OpTy<'tcx>,
    indices: &OpTy<'tcx>,
    right: &OpTy<'tcx>,
    dest: &MPlaceTy<'tcx>,
) -> InterpResult<'tcx, ()> {
    let (left, left_len) = ecx.project_to_simd(left)?;
    let (indices, indices_len) = ecx.project_to_simd(indices)?;
    let (right, right_len) = ecx.project_to_simd(right)?;
    let (dest, dest_len) = ecx.project_to_simd(dest)?;

    assert_eq!(dest_len, left_len);
    assert_eq!(dest_len, indices_len);
    assert_eq!(dest_len, right_len);

    // Use the low bits to select a lane within either input vector, and the next bit to
    // choose between the two vectors.
    assert!(dest_len.is_power_of_two());
    let lane_mask = u128::from(dest_len).strict_sub(1);
    let vector_select_bit = u128::from(dest_len);

    for i in 0..dest_len {
        let dest = ecx.project_index(&dest, i)?;
        let index_place = ecx.project_index(&indices, i)?;
        let index = ecx.read_scalar(&index_place)?.to_uint(index_place.layout.size)?;
        // `lane_mask` is at most `dest_len - 1` which fits in a `u64`, so this cannot fail.
        let lane = u64::try_from(index & lane_mask).unwrap();
        let src = if index & vector_select_bit == 0 { &left } else { &right };
        let element = ecx.project_index(src, lane)?;

        ecx.copy_op(&element, &dest)?;
    }

    interp_ok(())
}

/// Multiplies packed 16-bit signed integer values, truncates the 32-bit
/// product to the 18 most significant bits by right-shifting, and then
/// divides the 18-bit value by 2 (rounding to nearest) by first adding
/// 1 and then taking the bits `1..=16`.
///
/// <https://www.intel.com/content/www/us/en/docs/intrinsics-guide/index.html#text=_mm_mulhrs_epi16>
/// <https://www.intel.com/content/www/us/en/docs/intrinsics-guide/index.html#text=_mm256_mulhrs_epi16>
fn pmulhrsw<'tcx>(
    ecx: &mut crate::MiriInterpCx<'tcx>,
    left: &OpTy<'tcx>,
    right: &OpTy<'tcx>,
    dest: &MPlaceTy<'tcx>,
) -> InterpResult<'tcx, ()> {
    let (left, left_len) = ecx.project_to_simd(left)?;
    let (right, right_len) = ecx.project_to_simd(right)?;
    let (dest, dest_len) = ecx.project_to_simd(dest)?;

    assert_eq!(dest_len, left_len);
    assert_eq!(dest_len, right_len);

    for i in 0..dest_len {
        let left = ecx.read_scalar(&ecx.project_index(&left, i)?)?.to_i16()?;
        let right = ecx.read_scalar(&ecx.project_index(&right, i)?)?.to_i16()?;
        let dest = ecx.project_index(&dest, i)?;

        let res = (i32::from(left).strict_mul(right.into()) >> 14).strict_add(1) >> 1;

        // The result of this operation can overflow a signed 16-bit integer.
        // When `left` and `right` are -0x8000, the result is 0x8000.
        #[expect(clippy::as_conversions)]
        let res = res as i16;

        ecx.write_scalar(Scalar::from_i16(res), &dest)?;
    }

    interp_ok(())
}

/// Perform a carry-less multiplication of two 64-bit integers, selected from `left` and `right` according to `imm8`,
/// and store the results in `dst`.
///
/// `left` and `right` are both vectors of type `len` x i64. Only bits 0 and 4 of `imm8` matter;
/// they select the element of `left` and `right`, respectively.
///
/// `len` is the SIMD vector length (in counts of `i64` values). It is expected to be one of
/// `2`, `4`, or `8`.
///
/// <https://www.intel.com/content/www/us/en/docs/intrinsics-guide/index.html#text=_mm_clmulepi64_si128>
fn pclmulqdq<'tcx>(
    ecx: &mut MiriInterpCx<'tcx>,
    left: &OpTy<'tcx>,
    right: &OpTy<'tcx>,
    imm8: &OpTy<'tcx>,
    dest: &MPlaceTy<'tcx>,
    len: u64,
) -> InterpResult<'tcx, ()> {
    assert_eq!(left.layout, right.layout);
    assert_eq!(left.layout.size, dest.layout.size);
    assert!([2u64, 4, 8].contains(&len));

    // Transmute the input into arrays of `[u64; len]`.
    // Transmute the output into an array of `[u128, len / 2]`.

    let src_layout = ecx.layout_of(Ty::new_array(ecx.tcx.tcx, ecx.tcx.types.u64, len))?;
    let dest_layout = ecx.layout_of(Ty::new_array(ecx.tcx.tcx, ecx.tcx.types.u128, len / 2))?;

    let left = left.transmute(src_layout, ecx)?;
    let right = right.transmute(src_layout, ecx)?;
    let dest = dest.transmute(dest_layout, ecx)?;

    let imm8 = ecx.read_scalar(imm8)?.to_u8()?;

    for i in 0..(len / 2) {
        let lo = i.strict_mul(2);
        let hi = i.strict_mul(2).strict_add(1);

        // select the 64-bit integer from left that the user specified (low or high)
        let index = if (imm8 & 0x01) == 0 { lo } else { hi };
        let left = ecx.read_scalar(&ecx.project_index(&left, index)?)?.to_u64()?;

        // select the 64-bit integer from right that the user specified (low or high)
        let index = if (imm8 & 0x10) == 0 { lo } else { hi };
        let right = ecx.read_scalar(&ecx.project_index(&right, index)?)?.to_u64()?;

        let result = left.widening_carryless_mul(right);

        let dest = ecx.project_index(&dest, i)?;
        ecx.write_scalar(Scalar::from_u128(result), &dest)?;
    }

    interp_ok(())
}

/// Shuffles bytes from `left` using `right` as pattern. Each 16-byte block is shuffled independently.
///
/// `left` and `right` are both vectors of type `len` x i8.
///
/// If the highest bit of a byte in `right` is not set, the corresponding byte in `dest` is taken
/// from the current 16-byte block of `left` at the position indicated by the lowest 4 bits of this
/// byte in `right`. If the highest bit of a byte in `right` is set, the corresponding byte in
/// `dest` is set to `0`.
///
/// <https://www.intel.com/content/www/us/en/docs/intrinsics-guide/index.html#text=_mm_shuffle_epi8>
/// <https://www.intel.com/content/www/us/en/docs/intrinsics-guide/index.html#text=_mm256_shuffle_epi8>
/// <https://www.intel.com/content/www/us/en/docs/intrinsics-guide/index.html#text=_mm512_shuffle_epi8>
fn pshufb<'tcx>(
    ecx: &mut crate::MiriInterpCx<'tcx>,
    left: &OpTy<'tcx>,
    right: &OpTy<'tcx>,
    dest: &MPlaceTy<'tcx>,
) -> InterpResult<'tcx, ()> {
    let (left, left_len) = ecx.project_to_simd(left)?;
    let (right, right_len) = ecx.project_to_simd(right)?;
    let (dest, dest_len) = ecx.project_to_simd(dest)?;

    assert_eq!(dest_len, left_len);
    assert_eq!(dest_len, right_len);

    for i in 0..dest_len {
        let right = ecx.read_scalar(&ecx.project_index(&right, i)?)?.to_u8()?;
        let dest = ecx.project_index(&dest, i)?;

        let res = if right & 0x80 == 0 {
            // Shuffle each 128-bit (16-byte) block independently.
            let block_offset = i & !15; // round down to previous multiple of 16
            let j = block_offset.strict_add((right % 16).into());
            ecx.read_scalar(&ecx.project_index(&left, j)?)?
        } else {
            // If the highest bit in `right` is 1, write zero.
            Scalar::from_u8(0)
        };

        ecx.write_scalar(res, &dest)?;
    }

    interp_ok(())
}

/// Packs two N-bit integer vectors to a single N/2-bit integers.
///
/// The conversion from N-bit to N/2-bit should be provided by `f`.
///
/// Each 128-bit chunk is treated independently (i.e., the value for
/// the is i-th 128-bit chunk of `dest` is calculated with the i-th
/// 128-bit chunks of `left` and `right`).
fn pack_generic<'tcx>(
    ecx: &mut crate::MiriInterpCx<'tcx>,
    left: &OpTy<'tcx>,
    right: &OpTy<'tcx>,
    dest: &MPlaceTy<'tcx>,
    f: impl Fn(Scalar) -> InterpResult<'tcx, Scalar>,
) -> InterpResult<'tcx, ()> {
    assert_eq!(left.layout, right.layout);
    assert_eq!(left.layout.size, dest.layout.size);

    let (num_chunks, op_items_per_chunk, left) = split_simd_to_128bit_chunks(ecx, left)?;
    let (_, _, right) = split_simd_to_128bit_chunks(ecx, right)?;
    let (_, dest_items_per_chunk, dest) = split_simd_to_128bit_chunks(ecx, dest)?;

    assert_eq!(dest_items_per_chunk, op_items_per_chunk.strict_mul(2));

    for i in 0..num_chunks {
        let left = ecx.project_index(&left, i)?;
        let right = ecx.project_index(&right, i)?;
        let dest = ecx.project_index(&dest, i)?;

        for j in 0..op_items_per_chunk {
            let left = ecx.read_scalar(&ecx.project_index(&left, j)?)?;
            let right = ecx.read_scalar(&ecx.project_index(&right, j)?)?;
            let left_dest = ecx.project_index(&dest, j)?;
            let right_dest = ecx.project_index(&dest, j.strict_add(op_items_per_chunk))?;

            let left_res = f(left)?;
            let right_res = f(right)?;

            ecx.write_scalar(left_res, &left_dest)?;
            ecx.write_scalar(right_res, &right_dest)?;
        }
    }

    interp_ok(())
}

/// Converts two 16-bit integer vectors to a single 8-bit integer
/// vector with signed saturation.
///
/// Each 128-bit chunk is treated independently (i.e., the value for
/// the is i-th 128-bit chunk of `dest` is calculated with the i-th
/// 128-bit chunks of `left` and `right`).
fn packsswb<'tcx>(
    ecx: &mut crate::MiriInterpCx<'tcx>,
    left: &OpTy<'tcx>,
    right: &OpTy<'tcx>,
    dest: &MPlaceTy<'tcx>,
) -> InterpResult<'tcx, ()> {
    pack_generic(ecx, left, right, dest, |op| {
        let op = op.to_i16()?;
        let res = i8::try_from(op).unwrap_or(if op < 0 { i8::MIN } else { i8::MAX });
        interp_ok(Scalar::from_i8(res))
    })
}

/// Converts two 16-bit signed integer vectors to a single 8-bit
/// unsigned integer vector with saturation.
///
/// Each 128-bit chunk is treated independently (i.e., the value for
/// the is i-th 128-bit chunk of `dest` is calculated with the i-th
/// 128-bit chunks of `left` and `right`).
fn packuswb<'tcx>(
    ecx: &mut crate::MiriInterpCx<'tcx>,
    left: &OpTy<'tcx>,
    right: &OpTy<'tcx>,
    dest: &MPlaceTy<'tcx>,
) -> InterpResult<'tcx, ()> {
    pack_generic(ecx, left, right, dest, |op| {
        let op = op.to_i16()?;
        let res = u8::try_from(op).unwrap_or(if op < 0 { 0 } else { u8::MAX });
        interp_ok(Scalar::from_u8(res))
    })
}

/// Converts two 32-bit integer vectors to a single 16-bit integer
/// vector with signed saturation.
///
/// Each 128-bit chunk is treated independently (i.e., the value for
/// the is i-th 128-bit chunk of `dest` is calculated with the i-th
/// 128-bit chunks of `left` and `right`).
fn packssdw<'tcx>(
    ecx: &mut crate::MiriInterpCx<'tcx>,
    left: &OpTy<'tcx>,
    right: &OpTy<'tcx>,
    dest: &MPlaceTy<'tcx>,
) -> InterpResult<'tcx, ()> {
    pack_generic(ecx, left, right, dest, |op| {
        let op = op.to_i32()?;
        let res = i16::try_from(op).unwrap_or(if op < 0 { i16::MIN } else { i16::MAX });
        interp_ok(Scalar::from_i16(res))
    })
}

/// Converts two 32-bit integer vectors to a single 16-bit integer
/// vector with unsigned saturation.
///
/// Each 128-bit chunk is treated independently (i.e., the value for
/// the is i-th 128-bit chunk of `dest` is calculated with the i-th
/// 128-bit chunks of `left` and `right`).
fn packusdw<'tcx>(
    ecx: &mut crate::MiriInterpCx<'tcx>,
    left: &OpTy<'tcx>,
    right: &OpTy<'tcx>,
    dest: &MPlaceTy<'tcx>,
) -> InterpResult<'tcx, ()> {
    pack_generic(ecx, left, right, dest, |op| {
        let op = op.to_i32()?;
        let res = u16::try_from(op).unwrap_or(if op < 0 { 0 } else { u16::MAX });
        interp_ok(Scalar::from_u16(res))
    })
}

/// Negates elements from `left` when the corresponding element in
/// `right` is negative. If an element from `right` is zero, zero
/// is written to the corresponding output element.
/// In other words, multiplies `left` with `right.signum()`.
fn psign<'tcx>(
    ecx: &mut crate::MiriInterpCx<'tcx>,
    left: &OpTy<'tcx>,
    right: &OpTy<'tcx>,
    dest: &MPlaceTy<'tcx>,
) -> InterpResult<'tcx, ()> {
    let (left, left_len) = ecx.project_to_simd(left)?;
    let (right, right_len) = ecx.project_to_simd(right)?;
    let (dest, dest_len) = ecx.project_to_simd(dest)?;

    assert_eq!(dest_len, left_len);
    assert_eq!(dest_len, right_len);

    for i in 0..dest_len {
        let dest = ecx.project_index(&dest, i)?;
        let left = ecx.read_immediate(&ecx.project_index(&left, i)?)?;
        let right = ecx.read_scalar(&ecx.project_index(&right, i)?)?.to_int(dest.layout.size)?;

        let res =
            ecx.binary_op(mir::BinOp::Mul, &left, &ImmTy::from_int(right.signum(), dest.layout))?;

        ecx.write_immediate(*res, &dest)?;
    }

    interp_ok(())
}

/// Calcultates either `a + b + cb_in` or `a - b - cb_in` depending on the value
/// of `op` and returns both the sum and the overflow bit. `op` is expected to be
/// either one of `mir::BinOp::AddWithOverflow` and `mir::BinOp::SubWithOverflow`.
fn carrying_add<'tcx>(
    ecx: &mut crate::MiriInterpCx<'tcx>,
    cb_in: &OpTy<'tcx>,
    a: &OpTy<'tcx>,
    b: &OpTy<'tcx>,
    op: mir::BinOp,
) -> InterpResult<'tcx, (ImmTy<'tcx>, Scalar)> {
    assert!(op == mir::BinOp::AddWithOverflow || op == mir::BinOp::SubWithOverflow);

    let cb_in = ecx.read_scalar(cb_in)?.to_u8()? != 0;
    let a = ecx.read_immediate(a)?;
    let b = ecx.read_immediate(b)?;

    let (sum, overflow1) = ecx.binary_op(op, &a, &b)?.to_pair(ecx);
    let (sum, overflow2) =
        ecx.binary_op(op, &sum, &ImmTy::from_uint(cb_in, a.layout))?.to_pair(ecx);
    let cb_out = overflow1.to_scalar().to_bool()? | overflow2.to_scalar().to_bool()?;

    interp_ok((sum, Scalar::from_u8(cb_out.into())))
}