llvm-native-core 0.1.16

//! ARM MC Encoder — AArch64 and ARM32 instruction encoding.
//!
//! Encodes machine instructions (MachineInstr) into raw bytes per the ARM
//! Architecture Reference Manual (ARM ARM). Supports both AArch64 (fixed
//! 32-bit instructions, little-endian) and ARM32 (32-bit ARM and 16-bit
//! Thumb instruction sets).
//!
//! ## AArch64 encoding overview
//!
//! All AArch64 instructions are exactly 4 bytes, stored little-endian.
//! The encoder maps each MachineInstr (with its ArmOpcode and operands)
//! to the corresponding 32-bit encoding word.
//!
//! ## ARM32 encoding overview
//!
//! ARM32 instructions are 32-bit (ARM mode) or 16-bit (Thumb mode).
//! The encoder produces the appropriate encoding for each instruction.
//!
//! Clean-room reconstruction from the ARM Architecture Reference Manual.
//! Zero LLVM source code consultation.

use super::arm_instr_info::ArmOpcode;
use crate::codegen::{MachineInstr, MachineOperand};
use std::collections::HashMap;

// ============================================================================
// Opcode discriminant constants (derived from ArmOpcode enum order)
// ============================================================================

// Data Processing (AArch64)
const OP_ADD: u32 = ArmOpcode::ADD as u32;
const OP_SUB: u32 = ArmOpcode::SUB as u32;
// OP_MUL reserved for future lowerings; currently handled via MADD
const _OP_MUL: u32 = ArmOpcode::MUL as u32;
const OP_SDIV: u32 = ArmOpcode::SDIV as u32;
const OP_UDIV: u32 = ArmOpcode::UDIV as u32;
const OP_AND: u32 = ArmOpcode::AND as u32;
const OP_ORR: u32 = ArmOpcode::ORR as u32;
const OP_EOR: u32 = ArmOpcode::EOR as u32;
const OP_ADDS: u32 = ArmOpcode::ADDS as u32;
const OP_SUBS: u32 = ArmOpcode::SUBS as u32;
const OP_MADD: u32 = ArmOpcode::MADD as u32;
const OP_MSUB: u32 = ArmOpcode::MSUB as u32;

// MOV Variants
const OP_MOV: u32 = ArmOpcode::MOV as u32;
const OP_MOVZ: u32 = ArmOpcode::MOVZ as u32;
const OP_MOVK: u32 = ArmOpcode::MOVK as u32;
const OP_MOVN: u32 = ArmOpcode::MOVN as u32;

// Load/Store (AArch64)
const OP_LDR: u32 = ArmOpcode::LDR as u32;
const OP_STR: u32 = ArmOpcode::STR as u32;
const OP_LDP: u32 = ArmOpcode::LDP as u32;
const OP_STP: u32 = ArmOpcode::STP as u32;

// Branch (AArch64)
const OP_B: u32 = ArmOpcode::B as u32;
const OP_BL: u32 = ArmOpcode::BL as u32;
const OP_BR: u32 = ArmOpcode::BR as u32;
const OP_BLR: u32 = ArmOpcode::BLR as u32;
const OP_RET: u32 = ArmOpcode::RET as u32;
const OP_CBZ: u32 = ArmOpcode::CBZ as u32;
const OP_CBNZ: u32 = ArmOpcode::CBNZ as u32;
const OP_B_COND: u32 = ArmOpcode::B_COND as u32;

// Compare / Conditional
const OP_CMP: u32 = ArmOpcode::CMP as u32;
// OP_TST is reserved for future test-and-branch patterns
const _OP_TST: u32 = ArmOpcode::TST as u32;
const OP_CSEL: u32 = ArmOpcode::CSEL as u32;

// Bitfield
const OP_BFM: u32 = ArmOpcode::BFM as u32;
const OP_SBFM: u32 = ArmOpcode::SBFM as u32;
const OP_UBFM: u32 = ArmOpcode::UBFM as u32;

// System
const OP_NOP: u32 = ArmOpcode::NOP as u32;

// ARM32
const OP_ARM_ADD: u32 = ArmOpcode::ARM_ADD as u32;
const OP_ARM_SUB: u32 = ArmOpcode::ARM_SUB as u32;
const OP_ARM_MOV: u32 = ArmOpcode::ARM_MOV as u32;
const OP_ARM_AND: u32 = ArmOpcode::ARM_AND as u32;
const OP_ARM_ORR: u32 = ArmOpcode::ARM_ORR as u32;
const OP_ARM_EOR: u32 = ArmOpcode::ARM_EOR as u32;
const OP_ARM_CMP: u32 = ArmOpcode::ARM_CMP as u32;
const OP_ARM_B: u32 = ArmOpcode::ARM_B as u32;
const OP_ARM_BL: u32 = ArmOpcode::ARM_BL as u32;
const OP_ARM_LDR: u32 = ArmOpcode::ARM_LDR as u32;
const OP_ARM_STR: u32 = ArmOpcode::ARM_STR as u32;
const OP_ARM_PUSH: u32 = ArmOpcode::ARM_PUSH as u32;
const OP_ARM_POP: u32 = ArmOpcode::ARM_POP as u32;

// AArch64 extended opcodes — exception, system, SIMD/FP, advanced branches
const OP_SVC: u32 = ArmOpcode::SVC as u32;
const OP_HVC: u32 = ArmOpcode::HVC as u32;
const OP_SMC: u32 = ArmOpcode::SMC as u32;
const OP_BRK: u32 = ArmOpcode::BRK as u32;
const OP_MSR: u32 = ArmOpcode::MSR as u32;
const OP_MRS: u32 = ArmOpcode::MRS as u32;
const OP_SYS: u32 = ArmOpcode::SYS as u32;
const OP_SYSL: u32 = ArmOpcode::SYSL as u32;
const OP_ADR: u32 = ArmOpcode::ADR as u32;
const OP_ADRP: u32 = ArmOpcode::ADRP as u32;
const OP_TBNZ: u32 = ArmOpcode::TBNZ as u32;
const OP_TBZ: u32 = ArmOpcode::TBZ as u32;
const OP_LDRB: u32 = ArmOpcode::LDRB as u32;
const OP_LDRH: u32 = ArmOpcode::LDRH as u32;
const OP_LDRSB: u32 = ArmOpcode::LDRSB as u32;
const OP_LDRSH: u32 = ArmOpcode::LDRSH as u32;
const OP_LDRSW: u32 = ArmOpcode::LDRSW as u32;
const OP_STRB: u32 = ArmOpcode::STRB as u32;
const OP_STRH: u32 = ArmOpcode::STRH as u32;
const OP_LDUR: u32 = ArmOpcode::LDUR as u32;
const OP_LDURB: u32 = ArmOpcode::LDURB as u32;
const OP_LDURH: u32 = ArmOpcode::LDURH as u32;
const OP_STUR: u32 = ArmOpcode::STUR as u32;
const OP_STURB: u32 = ArmOpcode::STURB as u32;
const OP_STURH: u32 = ArmOpcode::STURH as u32;
const OP_LDXR: u32 = 1000;
const OP_LDXRB: u32 = 1001;
const OP_LDXRH: u32 = 1002;
const OP_STXR: u32 = 1003;
const OP_STXRB: u32 = 1004;
const OP_STXRH: u32 = 1005;
const OP_LDAXR: u32 = 1006;
const OP_STLXR: u32 = 1007;
const OP_PRFM: u32 = ArmOpcode::PRFM as u32;
const OP_CLREX: u32 = ArmOpcode::CLREX_A64 as u32;
const OP_DMB: u32 = ArmOpcode::DMB as u32;
const OP_DSB: u32 = ArmOpcode::DSB as u32;
const OP_ISB: u32 = ArmOpcode::ISB as u32;
const OP_CCMN: u32 = ArmOpcode::CCMN as u32;
const OP_CCMP: u32 = ArmOpcode::CCMP as u32;
const OP_RBIT: u32 = ArmOpcode::RBIT as u32;
const OP_REV: u32 = ArmOpcode::REV as u32;
const OP_REV16: u32 = ArmOpcode::REV16 as u32;
const OP_REV32: u32 = ArmOpcode::REV32 as u32;
const OP_CLZ: u32 = ArmOpcode::CLZ as u32;
const OP_CLS: u32 = ArmOpcode::CLS as u32;
const OP_MVN: u32 = 1010;
const OP_NEG: u32 = 1011;
const OP_NEGS: u32 = 1012;
const OP_NGC: u32 = 1013;
const OP_NGCS: u32 = 1014;
const OP_ADC: u32 = 1015;
const OP_ADCS: u32 = 1016;
const OP_SBC: u32 = 1017;
const OP_SBCS: u32 = 1018;
const OP_ORN: u32 = ArmOpcode::ORN as u32;
const OP_BIC: u32 = ArmOpcode::BIC as u32;
const OP_BICS: u32 = 1019;
const OP_EON: u32 = 1020;
const OP_TST: u32 = ArmOpcode::TST as u32;
const OP_ANDS_IMPLICIT: u32 = 1021;
const OP_UBFIZ: u32 = 1022;
const OP_SBFIZ: u32 = 1023;
const OP_UBFX: u32 = ArmOpcode::UBFX as u32;
const OP_SBFX: u32 = ArmOpcode::SBFX as u32;
const OP_EXTR: u32 = ArmOpcode::EXTR as u32;
const OP_LSLV: u32 = 1024;
const OP_LSRV: u32 = 1025;
const OP_ASRV: u32 = 1026;
const OP_RORV: u32 = 1027;
const OP_CRC32B: u32 = 1028;
const OP_CRC32H: u32 = 1029;
const OP_CRC32W: u32 = 1030;
const OP_CRC32X: u32 = 1031;
const OP_CRC32CB: u32 = 1032;
const OP_CRC32CH: u32 = 1033;
const OP_CRC32CW: u32 = 1034;
const OP_CRC32CX: u32 = 1035;
// SIMD/FP opcodes
const OP_FADD: u32 = ArmOpcode::FADD as u32;
const OP_FSUB: u32 = ArmOpcode::FSUB as u32;
const OP_FMUL: u32 = ArmOpcode::FMUL as u32;
const OP_FDIV: u32 = ArmOpcode::FDIV as u32;
const OP_FNEG: u32 = ArmOpcode::FNEG as u32;
const OP_FABS: u32 = ArmOpcode::FABS as u32;
const OP_FSQRT: u32 = ArmOpcode::FSQRT as u32;
const OP_FMOV: u32 = ArmOpcode::FMOV as u32;
const OP_FCVT: u32 = ArmOpcode::FCVT as u32;
const OP_FCVTZS: u32 = ArmOpcode::FCVTZS as u32;
const OP_FCVTZU: u32 = ArmOpcode::FCVTZU as u32;
const OP_SCVTF: u32 = ArmOpcode::SCVTF as u32;
const OP_UCVTF: u32 = ArmOpcode::UCVTF as u32;
const OP_FCMP: u32 = ArmOpcode::FCMP as u32;
const OP_FCMPE: u32 = 1040;
const OP_FCCMP: u32 = ArmOpcode::FCCMP as u32;
const OP_FCCMPE: u32 = 1041;
const OP_FCSEL: u32 = ArmOpcode::FCSEL as u32;
const OP_FMADD: u32 = 1042;
const OP_FMSUB: u32 = 1043;
const OP_FNMADD: u32 = 1044;
const OP_FNMSUB: u32 = 1045;
const OP_FMIN: u32 = ArmOpcode::FMIN as u32;
const OP_FMAX: u32 = ArmOpcode::FMAX as u32;
const OP_FRINTN: u32 = ArmOpcode::FRINTN as u32;
const OP_FRINTP: u32 = ArmOpcode::FRINTP as u32;
const OP_FRINTM: u32 = ArmOpcode::FRINTM as u32;
const OP_FRINTZ: u32 = ArmOpcode::FRINTZ as u32;
const OP_FRINTA: u32 = ArmOpcode::FRINTA as u32;
const OP_FRINTX: u32 = ArmOpcode::FRINTX as u32;
const OP_FRINTI: u32 = ArmOpcode::FRINTI as u32;
// SIMD vector opcodes
const OP_ADDV: u32 = 1050;
const OP_SADDLV: u32 = ArmOpcode::SADDLV as u32;
const OP_UADDLV: u32 = ArmOpcode::UADDLV as u32;
const OP_SMAXV: u32 = ArmOpcode::SMAXV as u32;
const OP_UMAXV: u32 = ArmOpcode::UMAXV as u32;
const OP_SMINV: u32 = ArmOpcode::SMINV as u32;
const OP_UMINV: u32 = ArmOpcode::UMINV as u32;
const OP_FMAXV: u32 = ArmOpcode::FMAXV as u32;
const OP_FMINV: u32 = ArmOpcode::FMINV as u32;
const OP_FADDP: u32 = ArmOpcode::FADDP as u32;
const OP_DUP: u32 = ArmOpcode::DUP as u32;
const OP_INS: u32 = ArmOpcode::INS as u32;
const OP_UMOV: u32 = ArmOpcode::UMOV as u32;
const OP_SMOV: u32 = ArmOpcode::SMOV as u32;
const OP_MOVI: u32 = ArmOpcode::MOVI as u32;
const OP_MVNI: u32 = 1051;
const OP_BIC_V: u32 = ArmOpcode::BIC_V as u32;
const OP_ORR_V: u32 = ArmOpcode::ORR_V as u32;
const OP_AND_V: u32 = ArmOpcode::AND_V as u32;
const OP_EOR_V: u32 = ArmOpcode::EOR_V as u32;
const OP_BSL: u32 = ArmOpcode::BSL as u32;
const OP_BIT: u32 = ArmOpcode::BIT as u32;
const OP_BIF: u32 = ArmOpcode::BIF as u32;
const OP_ADD_V: u32 = ArmOpcode::ADD_V as u32;
const OP_SUB_V: u32 = ArmOpcode::SUB_V as u32;
const OP_MUL_V: u32 = ArmOpcode::MUL_V as u32;
const OP_MLA_V: u32 = ArmOpcode::MLA_V as u32;
const OP_MLS_V: u32 = ArmOpcode::MLS_V as u32;
const OP_SHL: u32 = 1052;
const OP_SSHL: u32 = ArmOpcode::SSHL as u32;
const OP_USHL: u32 = ArmOpcode::USHL as u32;
const OP_CMEQ: u32 = ArmOpcode::CMEQ as u32;
const OP_CMGT: u32 = ArmOpcode::CMGT as u32;
const OP_CMGE: u32 = ArmOpcode::CMGE as u32;
const OP_CMHI: u32 = ArmOpcode::CMHI as u32;
const OP_CMHS: u32 = ArmOpcode::CMHS as u32;
const OP_CMTST: u32 = ArmOpcode::CMTST as u32;
const OP_FADD_V: u32 = 1053;
const OP_FSUB_V: u32 = 1054;
const OP_FMUL_V: u32 = 1055;
const OP_FDIV_V: u32 = 1056;
const OP_FCMLA: u32 = 1057;
const OP_FCADD: u32 = 1058;
const OP_EXT: u32 = ArmOpcode::EXT as u32;
const OP_TBL: u32 = ArmOpcode::TBL as u32;
const OP_TBX: u32 = ArmOpcode::TBX as u32;
const OP_ZIP1: u32 = ArmOpcode::ZIP1 as u32;
const OP_ZIP2: u32 = ArmOpcode::ZIP2 as u32;
const OP_UZP1: u32 = ArmOpcode::UZP1 as u32;
const OP_UZP2: u32 = ArmOpcode::UZP2 as u32;
const OP_TRN1: u32 = ArmOpcode::TRN1 as u32;
const OP_TRN2: u32 = ArmOpcode::TRN2 as u32;
const OP_REV64: u32 = ArmOpcode::REV64 as u32;
const OP_REV32_V: u32 = ArmOpcode::REV32_V as u32;
const OP_REV16_V: u32 = ArmOpcode::REV16_V as u32;
const OP_SADDLP: u32 = ArmOpcode::SADDLP as u32;
const OP_UADDLP: u32 = ArmOpcode::UADDLP as u32;
const OP_XTN: u32 = ArmOpcode::XTN as u32;
const OP_SQXTN: u32 = ArmOpcode::SQXTN as u32;
const OP_UQXTN: u32 = ArmOpcode::UQXTN as u32;
const OP_FCVTN: u32 = 1060;
const OP_FCVTL: u32 = 1061;
const OP_SCVTF_V: u32 = ArmOpcode::SCVTF_V as u32;
const OP_UCVTF_V: u32 = ArmOpcode::UCVTF_V as u32;
const OP_FCVTZS_V: u32 = ArmOpcode::FCVTZS_V as u32;
const OP_FCVTZU_V: u32 = ArmOpcode::FCVTZU_V as u32;
const OP_LD1: u32 = 1070;
const OP_LD2: u32 = ArmOpcode::LD2 as u32;
const OP_LD3: u32 = ArmOpcode::LD3 as u32;
const OP_LD4: u32 = ArmOpcode::LD4 as u32;
const OP_ST1: u32 = 1071;
const OP_ST2: u32 = ArmOpcode::ST2 as u32;
const OP_ST3: u32 = ArmOpcode::ST3 as u32;
const OP_ST4: u32 = ArmOpcode::ST4 as u32;
const OP_LDR_QIMM: u32 = 1072;
const OP_STR_QIMM: u32 = 1073;

// ============================================================================
// Condition codes for AArch64 (4-bit encoding)
// ============================================================================

#[derive(Debug, Clone, Copy, PartialEq, Eq)]
pub enum ArmCond {
    EQ = 0b0000,
    NE = 0b0001,
    CS = 0b0010, // HS (unsigned higher or same)
    CC = 0b0011, // LO (unsigned lower)
    MI = 0b0100,
    PL = 0b0101,
    VS = 0b0110,
    VC = 0b0111,
    HI = 0b1000,
    LS = 0b1001,
    GE = 0b1010,
    LT = 0b1011,
    GT = 0b1100,
    LE = 0b1101,
    AL = 0b1110,
}

impl ArmCond {
    pub fn from_mnemonic(s: &str) -> Option<ArmCond> {
        match s.to_lowercase().as_str() {
            "eq" => Some(ArmCond::EQ),
            "ne" => Some(ArmCond::NE),
            "cs" | "hs" => Some(ArmCond::CS),
            "cc" | "lo" => Some(ArmCond::CC),
            "mi" => Some(ArmCond::MI),
            "pl" => Some(ArmCond::PL),
            "vs" => Some(ArmCond::VS),
            "vc" => Some(ArmCond::VC),
            "hi" => Some(ArmCond::HI),
            "ls" => Some(ArmCond::LS),
            "ge" => Some(ArmCond::GE),
            "lt" => Some(ArmCond::LT),
            "gt" => Some(ArmCond::GT),
            "le" => Some(ArmCond::LE),
            "al" => Some(ArmCond::AL),
            _ => None,
        }
    }

    pub fn invert(self) -> ArmCond {
        match self {
            ArmCond::EQ => ArmCond::NE,
            ArmCond::NE => ArmCond::EQ,
            ArmCond::CS => ArmCond::CC,
            ArmCond::CC => ArmCond::CS,
            ArmCond::MI => ArmCond::PL,
            ArmCond::PL => ArmCond::MI,
            ArmCond::VS => ArmCond::VC,
            ArmCond::VC => ArmCond::VS,
            ArmCond::HI => ArmCond::LS,
            ArmCond::LS => ArmCond::HI,
            ArmCond::GE => ArmCond::LT,
            ArmCond::LT => ArmCond::GE,
            ArmCond::GT => ArmCond::LE,
            ArmCond::LE => ArmCond::GT,
            ArmCond::AL => ArmCond::AL,
        }
    }
}

// ============================================================================
// ArmMCEncoder
// ============================================================================

/// ARM Machine Code Encoder. Encodes `MachineInstr` objects into raw
/// instruction bytes for AArch64 (64-bit) and ARM32 (32-bit) targets.
pub struct ArmMCEncoder {
    /// If true, target AArch64 (64-bit); otherwise ARM32.
    pub is_64bit: bool,
    /// Accumulated output byte stream.
    pub output: Vec<u8>,
    /// Optional label-to-offset mapping for branch resolution.
    pub label_offsets: HashMap<String, u64>,
}

impl ArmMCEncoder {
    /// Create a new encoder.
    pub fn new(is_64bit: bool) -> Self {
        Self {
            is_64bit,
            output: Vec::new(),
            label_offsets: HashMap::new(),
        }
    }

    // ========================================================================
    // Main entry point
    // ========================================================================

    /// Encode a single machine instruction into 4 bytes (AArch64) or
    /// 2/4 bytes (ARM32/Thumb).
    pub fn encode_instruction(&self, mi: &MachineInstr) -> Vec<u8> {
        if self.is_64bit {
            let word = self.encode_aarch64(mi);
            word.to_le_bytes().to_vec()
        } else {
            let word = self.encode_arm32_instruction(mi);
            // ARM32 instructions are 32-bit; Thumb can be 16-bit.
            // For now, always emit as 4-byte LE (ARM mode).
            word.to_le_bytes().to_vec()
        }
    }

    /// Record a label position for later branch resolution.
    pub fn record_label(&mut self, label: &str, offset: u64) {
        self.label_offsets.insert(label.to_string(), offset);
    }

    /// Get the current output length (in bytes).
    pub fn len(&self) -> usize {
        self.output.len()
    }

    /// Check if output is empty.
    pub fn is_empty(&self) -> bool {
        self.output.is_empty()
    }

    // ========================================================================
    // AArch64 master dispatch
    // ========================================================================

    fn encode_aarch64(&self, mi: &MachineInstr) -> u32 {
        let op = mi.opcode;
        // Data processing (immediate)
        if op == OP_ADD || op == OP_SUB || op == OP_ADDS || op == OP_SUBS {
            self.encode_add_sub(mi)
        } else if op == OP_ADC || op == OP_ADCS || op == OP_SBC || op == OP_SBCS {
            self.encode_add_sub(mi) // ADC/SBC share add/sub encoding with carry
        } else if op == OP_MOVZ {
            self.encode_movz(mi)
        } else if op == OP_MOVK {
            self.encode_movk(mi)
        } else if op == OP_MOVN {
            self.encode_movn(mi)
        } else if op == OP_MOV {
            self.encode_mov(mi)
        } else if op == OP_MVN {
            self.encode_mvn(mi)
        } else if op == OP_NEG || op == OP_NEGS || op == OP_NGC || op == OP_NGCS {
            self.encode_neg(mi)
        }
        // Load/store
        else if op == OP_LDR || op == OP_STR {
            self.encode_ldr_str(mi)
        } else if op == OP_LDP || op == OP_STP {
            self.encode_ldp_stp(mi)
        } else if op == OP_LDRB {
            self.encode_ldrb_imm(mi)
        } else if op == OP_STRB {
            self.encode_strb_imm(mi)
        } else if op == OP_LDRH {
            self.encode_ldrh_imm(mi)
        } else if op == OP_STRH {
            self.encode_strh_imm(mi)
        } else if op == OP_LDRSB {
            self.encode_ldrsb_imm(mi)
        } else if op == OP_LDRSH {
            self.encode_ldrsh_imm(mi)
        } else if op == OP_LDRSW {
            self.encode_ldrsw_imm(mi)
        } else if op == OP_LDUR {
            self.encode_ldur(mi)
        } else if op == OP_STUR {
            self.encode_stur(mi)
        } else if op == OP_LDURB {
            self.encode_ldurb(mi)
        } else if op == OP_LDURH {
            self.encode_ldurh(mi)
        } else if op == OP_STURB {
            self.encode_sturb(mi)
        } else if op == OP_STURH {
            self.encode_sturh(mi)
        } else if op == OP_LDXR || op == OP_LDXRB || op == OP_LDXRH {
            self.encode_ldxr(mi)
        } else if op == OP_STXR || op == OP_STXRB || op == OP_STXRH {
            self.encode_stxr(mi)
        } else if op == OP_LDAXR {
            self.encode_ldaxr(mi)
        } else if op == OP_STLXR {
            self.encode_stlxr(mi)
        }
        // Branch
        else if op == OP_B || op == OP_BL {
            self.encode_b_bl(mi)
        } else if op == OP_CBZ || op == OP_CBNZ {
            self.encode_cbz_cbnz(mi)
        } else if op == OP_RET {
            self.encode_ret(mi)
        } else if op == OP_BR {
            self.encode_br(mi)
        } else if op == OP_BLR {
            self.encode_blr(mi)
        } else if op == OP_B_COND {
            self.encode_b_cond(mi)
        } else if op == OP_TBNZ || op == OP_TBZ {
            self.encode_tbz_tbnz(mi)
        }
        // System
        else if op == OP_NOP {
            self.encode_nop()
        } else if op == OP_SVC {
            self.encode_svc(mi)
        } else if op == OP_HVC {
            self.encode_hvc(mi)
        } else if op == OP_SMC {
            self.encode_smc(mi)
        } else if op == OP_BRK {
            self.encode_brk(mi)
        } else if op == OP_MSR {
            self.encode_msr(mi)
        } else if op == OP_MRS {
            self.encode_mrs(mi)
        } else if op == OP_SYS {
            self.encode_sys(mi)
        } else if op == OP_SYSL {
            self.encode_sysl(mi)
        } else if op == OP_DMB {
            self.encode_dmb(mi)
        } else if op == OP_DSB {
            self.encode_dsb(mi)
        } else if op == OP_ISB {
            self.encode_isb(mi)
        } else if op == OP_CLREX {
            self.encode_clrex(mi)
        }
        // PC-relative address
        else if op == OP_ADR {
            self.encode_adr(mi)
        } else if op == OP_ADRP {
            self.encode_adrp(mi)
        }
        // Data processing (register)
        else if op == OP_CMP {
            self.encode_cmp(mi)
        } else if op == OP_CCMN || op == OP_CCMP {
            self.encode_ccmp_ccmn(mi)
        } else if op == OP_CSEL {
            self.encode_csel(mi)
        } else if op == OP_RBIT {
            self.encode_rbit(mi)
        } else if op == OP_REV || op == OP_REV16 || op == OP_REV32 {
            self.encode_rev(mi)
        } else if op == OP_CLZ || op == OP_CLS {
            self.encode_clz(mi)
        } else if op == OP_BFM || op == OP_SBFM || op == OP_UBFM {
            self.encode_bfm(mi)
        } else if op == OP_UBFIZ || op == OP_SBFIZ {
            self.encode_ubfiz(mi)
        } else if op == OP_UBFX {
            self.encode_ubfx(mi)
        } else if op == OP_SBFX {
            self.encode_sbfx(mi)
        } else if op == OP_EXTR {
            self.encode_extr(mi)
        } else if op == OP_SDIV || op == OP_UDIV {
            self.encode_sdiv_udiv(mi)
        } else if op == OP_MADD || op == OP_MSUB {
            self.encode_madd_msub(mi)
        } else if op == OP_LSLV {
            self.encode_lslv(mi)
        } else if op == OP_LSRV {
            self.encode_lsrv(mi)
        } else if op == OP_ASRV {
            self.encode_asrv(mi)
        } else if op == OP_RORV {
            self.encode_rorv(mi)
        }
        // Logical
        else if op == OP_AND || op == OP_ORR || op == OP_EOR {
            self.encode_aarch64_logical(mi)
        } else if op == OP_BIC || op == OP_BICS {
            self.encode_bic(mi)
        } else if op == OP_ORN {
            self.encode_aarch64_logical(mi) // ORN aliased via ORR with NOT Rm
        } else if op == OP_EON {
            self.encode_aarch64_logical(mi)
        } else if op == OP_TST || op == OP_ANDS_IMPLICIT {
            self.encode_aarch64_logical(mi)
        }
        // CRC
        else if op >= OP_CRC32B && op <= OP_CRC32CX {
            self.encode_crc32(mi)
        }
        // FP/SIMD scalar
        else if op == OP_FADD {
            self.encode_fadd(mi)
        } else if op == OP_FSUB {
            self.encode_fsub(mi)
        } else if op == OP_FMUL {
            self.encode_fmul(mi)
        } else if op == OP_FDIV {
            self.encode_fdiv(mi)
        } else if op == OP_FNEG {
            self.encode_fneg(mi)
        } else if op == OP_FABS {
            self.encode_fabs(mi)
        } else if op == OP_FSQRT {
            self.encode_fsqrt(mi)
        } else if op == OP_FCVT {
            self.encode_fcvt(mi)
        } else if op == OP_FMOV {
            self.encode_fmov(mi)
        } else if op == OP_FCMP {
            self.encode_fcmp(mi)
        } else if op == OP_FCMPE {
            self.encode_fcmpe(mi)
        } else if op == OP_FCSEL {
            self.encode_fcsel(mi)
        } else if op == OP_FMADD || op == OP_FMSUB || op == OP_FNMADD || op == OP_FNMSUB {
            self.encode_fmadd(mi)
        } else if op == OP_FMIN || op == OP_FMAX {
            self.encode_fminmax(mi)
        } else if op >= OP_FRINTN && op <= OP_FRINTI {
            self.encode_frint(mi)
        } else if op == OP_FCVTZS || op == OP_FCVTZU || op == OP_SCVTF || op == OP_UCVTF {
            self.encode_fcvt_int_fp(mi)
        }
        // SIMD vector
        else if op == OP_ADD_V {
            self.encode_add_v(mi)
        } else if op == OP_SUB_V {
            self.encode_sub_v(mi)
        } else if op == OP_MUL_V {
            self.encode_mul_v(mi)
        } else if op == OP_MLA_V {
            self.encode_mla_v(mi)
        } else if op == OP_MLS_V {
            self.encode_mls_v(mi)
        } else if op == OP_AND_V {
            self.encode_and_v(mi)
        } else if op == OP_ORR_V {
            self.encode_orr_v(mi)
        } else if op == OP_EOR_V {
            self.encode_eor_v(mi)
        } else if op == OP_BIC_V {
            self.encode_bic_v(mi)
        } else if op == OP_BSL || op == OP_BIT || op == OP_BIF {
            self.encode_bsl(mi)
        } else if op == OP_SHL || op == OP_SSHL || op == OP_USHL {
            self.encode_simd_shift(mi)
        } else if op == OP_CMEQ
            || op == OP_CMGT
            || op == OP_CMGE
            || op == OP_CMHI
            || op == OP_CMHS
            || op == OP_CMTST
        {
            self.encode_simd_compare(mi)
        } else if op == OP_FADD_V
            || op == OP_FSUB_V
            || op == OP_FMUL_V
            || op == OP_FDIV_V
            || op == OP_FCMLA
            || op == OP_FCADD
        {
            self.encode_simd_fp_three_same(mi)
        } else if op >= OP_ADDV && op <= OP_FADDP {
            self.encode_simd_across(mi)
        } else if op == OP_DUP {
            self.encode_dup(mi)
        } else if op == OP_MOVI || op == OP_MVNI {
            self.encode_movi(mi)
        } else if op == OP_EXT {
            self.encode_ext(mi)
        } else if op >= OP_ZIP1 && op <= OP_TRN2 {
            self.encode_simd_permute(mi)
        } else if op >= OP_REV64 && op <= OP_REV16_V {
            self.encode_simd_permute(mi)
        } else if op >= OP_SADDLP && op <= OP_UADDLP {
            self.encode_simd_across(mi)
        } else if op >= OP_XTN && op <= OP_FCVTL {
            self.encode_simd_narrow(mi)
        } else if op >= OP_LD1 && op <= OP_LD4 {
            self.encode_simd_ldst_single(mi)
        } else if op >= OP_ST1 && op <= OP_ST4 {
            self.encode_simd_ldst_single(mi)
        } else if op == OP_LDR_QIMM {
            self.encode_simd_ldr_str_imm(mi, false)
        } else if op == OP_STR_QIMM {
            self.encode_simd_ldr_str_imm(mi, true)
        } else {
            // Unrecognized: emit NOP
            0xD503201F
        }
    }

    // ========================================================================
    // AArch64 ADD/SUB
    // ========================================================================

    /// Encode ADD/SUB: supports both immediate and register forms.
    /// - ADD/SUB immediate: sf | op(0/1) | S | 10001 | shift | imm12 | Rn | Rd
    /// - ADD/SUB register:  sf | op(0/1) | S | 01011 | shift | 0 | Rm | imm6 | Rn | Rd
    fn encode_add_sub(&self, mi: &MachineInstr) -> u32 {
        let op = mi.opcode;
        let is_sub = op == OP_SUB || op == OP_SUBS;
        let set_flags = op == OP_ADDS || op == OP_SUBS;
        let sf: u32 = 1; // Always 64-bit for now

        // Try immediate form if operand[2] is an immediate
        if mi.operands.len() >= 3 {
            if let MachineOperand::Imm(imm_val) = mi.operands[2] {
                return self.encode_add_sub_imm(mi, is_sub, set_flags, sf, imm_val);
            }
        }

        // Register form: Rd, Rn, Rm
        let rd = self.get_reg_field_from_operand(mi.operands.first());
        let rn = self.get_reg_field_from_operand(mi.operands.get(1));
        let rm = self.get_reg_field_from_operand(mi.operands.get(2));
        self.encode_add_sub_reg(is_sub, set_flags, sf, rd, rn, rm, 0, 0) // shift=0, amount=0
    }

    /// ADD/SUB immediate encoding:
    /// sf(31) | op(30) | S(29) | 10001(28-24) | sh(23-22) | imm12(21-10) | Rn(9-5) | Rd(4-0)
    fn encode_add_sub_imm(
        &self,
        mi: &MachineInstr,
        is_sub: bool,
        set_flags: bool,
        sf: u32,
        imm: i64,
    ) -> u32 {
        let rd = self.get_reg_field_from_operand(mi.operands.first());
        let rn = self.get_reg_field_from_operand(mi.operands.get(1));

        let imm_u = imm as u64;
        let (imm12, shift) = if imm_u <= 0xFFF {
            (imm_u as u32 & 0xFFF, 0u32)
        } else if imm_u & 0xFFF == 0 && (imm_u >> 12) <= 0xFFF {
            ((imm_u >> 12) as u32 & 0xFFF, 1u32)
        } else {
            // Cannot encode: truncate
            (imm_u as u32 & 0xFFF, 0u32)
        };

        let mut word: u32 = 0;
        word |= sf << 31;
        word |= (is_sub as u32) << 30;
        word |= (set_flags as u32) << 29;
        word |= 0b10001 << 24;
        word |= shift << 22;
        word |= (imm12 & 0xFFF) << 10;
        word |= (rn as u32) << 5;
        word |= rd as u32;
        word
    }

    /// ADD/SUB register encoding:
    /// sf(31) | op(30) | S(29) | 01011(28-24) | shift(23-22) | 0(21) | Rm(20-16) | imm6(15-10) | Rn(9-5) | Rd(4-0)
    fn encode_add_sub_reg(
        &self,
        is_sub: bool,
        set_flags: bool,
        sf: u32,
        rd: u8,
        rn: u8,
        rm: u8,
        shift: u32,
        amount: u32,
    ) -> u32 {
        let mut word: u32 = 0;
        word |= sf << 31;
        word |= (is_sub as u32) << 30;
        word |= (set_flags as u32) << 29;
        word |= 0b01011 << 24;
        word |= (shift & 0x3) << 22;
        word |= (rm as u32) << 16;
        word |= (amount & 0x3F) << 10;
        word |= (rn as u32) << 5;
        word |= rd as u32;
        word
    }

    // ========================================================================
    // AArch64 MOVZ / MOVK / MOVN
    // ========================================================================

    /// MOVZ: move wide with zero.
    /// sf(31) | 10(30-29) | 100101(28-23) | hw(22-21) | imm16(20-5) | Rd(4-0)
    fn encode_movz(&self, mi: &MachineInstr) -> u32 {
        let sf: u32 = 1;
        let rd = self.get_reg_field_from_operand(mi.operands.first());
        let imm = self.get_imm_from_operand(mi.operands.get(1));
        let hw = self.get_hw_from_operand(mi.operands.get(2)); // optional shift specifier

        let mut word: u32 = 0;
        word |= sf << 31;
        word |= 0b10 << 29;
        word |= 0b100101 << 23;
        word |= (hw & 0x3) << 21;
        word |= (imm as u32 & 0xFFFF) << 5;
        word |= rd as u32;
        word
    }

    /// MOVK: move wide with keep.
    /// sf(31) | 11(30-29) | 100101(28-23) | hw(22-21) | imm16(20-5) | Rd(4-0)
    fn encode_movk(&self, mi: &MachineInstr) -> u32 {
        let sf: u32 = 1;
        let rd = self.get_reg_field_from_operand(mi.operands.first());
        let imm = self.get_imm_from_operand(mi.operands.get(1));
        let hw = self.get_hw_from_operand(mi.operands.get(2));

        let mut word: u32 = 0;
        word |= sf << 31;
        word |= 0b11 << 29;
        word |= 0b100101 << 23;
        word |= (hw & 0x3) << 21;
        word |= (imm as u32 & 0xFFFF) << 5;
        word |= rd as u32;
        word
    }

    /// MOVN: move wide with NOT.
    /// sf(31) | 00(30-29) | 100101(28-23) | hw(22-21) | imm16(20-5) | Rd(4-0)
    fn encode_movn(&self, mi: &MachineInstr) -> u32 {
        let sf: u32 = 1;
        let rd = self.get_reg_field_from_operand(mi.operands.first());
        let imm = self.get_imm_from_operand(mi.operands.get(1));
        let hw = self.get_hw_from_operand(mi.operands.get(2));

        let mut word: u32 = 0;
        word |= sf << 31;
        word |= 0b00 << 29;
        word |= 0b100101 << 23;
        word |= (hw & 0x3) << 21;
        word |= (imm as u32 & 0xFFFF) << 5;
        word |= rd as u32;
        word
    }

    // ========================================================================
    // AArch64 MOV (various forms)
    // ========================================================================

    /// Encode MOV: dispatches to appropriate form.
    /// MOV Rd, Rn → ORR Rd, XZR, Rm  (register-register move)
    /// MOV Rd, #imm → MOVZ Rd, #imm or ORR Rd, XZR, #imm
    fn encode_mov(&self, mi: &MachineInstr) -> u32 {
        if mi.operands.len() < 2 {
            return self.encode_nop();
        }
        match &mi.operands[1] {
            MachineOperand::Reg(_) | MachineOperand::PhysReg(_) => self.encode_mov_reg(mi),
            MachineOperand::Imm(_) => self.encode_mov_imm(mi),
            _ => self.encode_nop(),
        }
    }

    /// MOV (register): ORR Rd, XZR, Rm
    /// sf(31) | 01(30-29) | 01010(28-24) | shift(23-22) | 0(21) | Rm(20-16) | imm6(15-10) | Rn(9-5) | Rd(4-0)
    fn encode_mov_reg(&self, mi: &MachineInstr) -> u32 {
        let sf: u32 = 1;
        let rd = self.get_reg_field_from_operand(mi.operands.first());
        let rm = self.get_reg_field_from_operand(mi.operands.get(1));

        let mut word: u32 = 0;
        word |= sf << 31;
        word |= 0b01 << 29;
        word |= 0b01010 << 24;
        // Rn = XZR (31)
        word |= 31u32 << 5;
        word |= (rm as u32) << 16;
        word |= rd as u32;
        word
    }

    /// MOV (immediate): ORR Rd, XZR, #imm  (bitmask immediate)
    fn encode_mov_imm(&self, mi: &MachineInstr) -> u32 {
        let sf: u32 = 1;
        let rd = self.get_reg_field_from_operand(mi.operands.first());
        let imm_val = self.get_imm_from_operand(mi.operands.get(1)) as u64;

        if let Some((immr, imms, n)) = encode_bitmask_imm(imm_val, 64) {
            let mut word: u32 = 0;
            word |= sf << 31;
            word |= 0b01 << 29;
            word |= 0b100100 << 23;
            word |= n << 22;
            word |= (immr & 0x3F) << 16;
            word |= (imms & 0x3F) << 10;
            word |= 31u32 << 5; // Rn = XZR
            word |= rd as u32;
            word
        } else {
            // Fall back to MOVZ for values that fit in 16 bits
            let mut word: u32 = 0;
            word |= sf << 31;
            word |= 0b10 << 29; // MOVZ
            word |= 0b100101 << 23;
            word |= (imm_val as u32 & 0xFFFF) << 5;
            word |= rd as u32;
            word
        }
    }

    // ========================================================================
    // AArch64 LDR / STR
    // ========================================================================

    /// Encode LDR/STR (immediate offset, unsigned).
    /// LDR (64-bit): 11(31-30) | 11100101(29-22) | imm9(21-10) | 01(11-10?) — wait, let me recalculate.
    ///
    /// Correct AArch64 LDR/STR unsigned immediate layout:
    /// size(31-30) | 11100101(29-22) for LDR, 11100100(29-22) for STR | imm9(21-12) scaled | 01(11-10) | Rn(9-5) | Rt(4-0)
    ///
    /// Wait, bits 21-10 = 12 bits but imm9 is 9 bits. The actual layout is:
    /// size(31-30) | 1110010x(29-22) where x=1 for LDR, x=0 for STR | imm9(21-12 but scaled) | 01(11-10) | Rn(9-5) | Rt(4-0)
    /// Hmm, that's 12+2 = 14 bits between fixed header and Rn. Let me check imm9 placement.
    ///
    /// Actually from ARM ARM: LDR (immediate) unsigned offset:
    /// | 1 1 | 1 1 1 0 0 1 0 1 | imm9                                   | 0 1 | Rn    | Rt    |
    ///  31-30  29 28 27 26 25 24 23 22  21 20 19 18 17 16 15 14 13 12   11-10  9 8 6 5  4 3 0
    ///
    /// So imm9 is in bits 21-12? No — 21-10? Let me count:
    /// 31-22 = 10 bits: 11 1110 0101
    /// 21-10 = 12 bits: imm9?? But imm9 is only 9 bits...
    ///
    /// Actually for 64-bit LDR, the immediate is scaled by 8 (size of u64).
    /// So a 9-bit immediate pimm encoding represents pimm*8.
    /// The encoding places imm9 in bits 21-13? Let me simplify:
    ///
    /// Per ARM ARM C6.2.166 LDR (immediate):
    /// - The unsigned immediate is encoded as a 9-bit field (bits 21-12 or similar).
    ///
    /// For our purposes, let's use the simplified layout from the spec:
    /// | 11 1110 0101 | imm9 | 01 | Rn | Rt |
    /// meaning:
    ///   bits 31-22: 0b11_1110_0101 = 0x3E5 for LDR
    ///   bits 21-12: imm9 (9-bit scaled offset)
    ///   bits 11-10: 0b01
    ///   bits 9-5: Rn
    ///   bits 4-0: Rt
    fn encode_ldr_str(&self, mi: &MachineInstr) -> u32 {
        let is_store = mi.opcode == OP_STR;
        let size: u32 = 0b11; // 64-bit (11 = 8 bytes)

        let rt = self.get_reg_field_from_operand(mi.operands.first());
        let rn = self.get_reg_field_from_operand(mi.operands.get(1));
        let imm = self.get_imm_from_operand(mi.operands.get(2));

        // Scale immediate for 64-bit access (divide by 8)
        let scaled_imm = (imm as u64 / 8) as u32;

        let mut word: u32 = 0;
        word |= size << 30; // bits 31-30
                            // Fixed encoding bits 29-22:
                            //   LDR: 0b11100101, STR: 0b11100100
        if is_store {
            word |= 0b11100100 << 22;
        } else {
            word |= 0b11100101 << 22;
        }
        word |= (scaled_imm & 0x1FF) << 12; // imm9 in bits 21-12
        word |= 0b01 << 10;
        word |= (rn as u32) << 5;
        word |= rt as u32;
        word
    }

    // ========================================================================
    // AArch64 LDP / STP
    // ========================================================================

    /// Encode LDP/STP (load/store pair).
    /// LDP Xt1, Xt2, [Xn, #imm]:  opc(31-30) | 1010010011(29-20?) | imm7(21-15) | Rt2(14-10) | Rn(9-5) | Rt(4-0)
    ///
    /// AArch64 LDP/STP (64-bit, signed offset, pre/post/no indexing):
    /// | opc(31-30) | 1 0 1 0 0 1 0 0 1 1 (29-20) — wait, that's too wide.
    ///
    /// From ARM ARM C6.2.119 LDP (immediate):
    /// | 1 0 | 1 0 1 0 0 1 0 1 | imm7             | Rt2   | Rn    | Rt    |
    ///  31-30  29 28 27 26 25 24 23 22  21 20 19 18 17 16 15  14-10   9-5    4-0
    ///
    /// opc(31-30): 10 for LDP, 11 for STP (64-bit)
    /// 10100101 (29-22): fixed for pair load/store
    /// imm7 (21-15): 7-bit signed scaled offset
    /// Rt2 (14-10): second register
    /// Rn (9-5): base register
    /// Rt (4-0): first register
    fn encode_ldp_stp(&self, mi: &MachineInstr) -> u32 {
        let is_store = mi.opcode == OP_STP;
        let rt = self.get_reg_field_from_operand(mi.operands.first());
        let rt2 = self.get_reg_field_from_operand(mi.operands.get(1));
        let rn = self.get_reg_field_from_operand(mi.operands.get(2));
        let imm = self.get_imm_from_operand(mi.operands.get(3));

        // Scale for 64-bit: divide by 8
        let scaled_imm = (imm / 8) as i32;

        let mut word: u32 = 0;
        if is_store {
            word |= 0b11 << 30; // opc = 11 for STP (64-bit)
        } else {
            word |= 0b10 << 30; // opc = 10 for LDP (64-bit)
        }
        word |= 0b10100101 << 22; // Fixed bits 29-22
        word |= ((scaled_imm as u32) & 0x7F) << 15; // imm7 in bits 21-15
        word |= (rt2 as u32) << 10; // Rt2 in bits 14-10
        word |= (rn as u32) << 5;
        word |= rt as u32;
        word
    }

    // ========================================================================
    // AArch64 B / BL
    // ========================================================================

    /// Encode B/BL: 26-bit signed offset, divided by 4.
    /// B:  | 0 00101 | imm26 |
    /// BL: | 1 00101 | imm26 |
    fn encode_b_bl(&self, mi: &MachineInstr) -> u32 {
        let is_call = mi.opcode == OP_BL;
        // For branches, operand[0] is typically a label; resolve offset
        let offset = self.get_branch_offset(mi);

        let imm26 = (offset as u32) & 0x3FFFFFF;
        let mut word: u32 = 0;
        if is_call {
            word |= 0b100101 << 26;
        } else {
            word |= 0b000101 << 26;
        }
        word |= imm26;
        word
    }

    // ========================================================================
    // AArch64 B.cond
    // ========================================================================

    /// Encode B.cond: conditional branch.
    /// | 0 1 0 1 0 1 0 0 | imm19 | 0 | cond |
    ///   31-24              23-5    4    3-0
    fn encode_b_cond(&self, mi: &MachineInstr) -> u32 {
        let cond = self.get_cond_from_operand(mi.operands.first());
        let offset = self.get_branch_offset_19bit(mi);

        let mut word: u32 = 0;
        word |= 0b01010100 << 24;
        word |= (offset & 0x7FFFF) << 5;
        word |= cond as u32;
        word
    }

    // ========================================================================
    // AArch64 CBZ / CBNZ
    // ========================================================================

    /// Encode CBZ/CBNZ: compare and branch on zero.
    /// CBZ:  | 1 0 1 1 0 1 0 0 | imm19 | Rt |
    ///       31-24               23-5    4-0
    /// CBNZ: | 1 0 1 1 0 1 0 1 | imm19 | Rt |
    fn encode_cbz_cbnz(&self, mi: &MachineInstr) -> u32 {
        let is_nonzero = mi.opcode == OP_CBNZ;
        let rt = self.get_reg_field_from_operand(mi.operands.first());
        let offset = self.get_branch_offset_19bit(mi);

        let mut word: u32 = 0;
        if is_nonzero {
            word |= 0xB5 << 24; // 10110101
        } else {
            word |= 0xB4 << 24; // 10110100
        }
        word |= (offset & 0x7FFFF) << 5;
        word |= rt as u32;
        word
    }

    // ========================================================================
    // AArch64 RET
    // ========================================================================

    /// RET Xn: 1101011 0 0 10 11111 0000 0 0 | Rn | 00000
    /// Standard encoding: 0xD65F0000 | (Rn << 5)
    /// RET X30 (default LR): 0xD65F03C0
    fn encode_ret(&self, mi: &MachineInstr) -> u32 {
        let rn = if let Some(op) = mi.operands.first() {
            self.get_reg_field_from_operand(Some(op))
        } else {
            30 // Default to LR (X30)
        };
        // RET encoding:
        // 1101011_0_0_10_11111_0000_0_0_Rn_00000
        // bits: 31-25=1101011, 24=0, 23=0, 22-21=10, 20-16=11111, 15-10=000000, 9-5=Rn, 4-0=00000
        let mut word: u32 = 0xD65F0000;
        word |= (rn as u32) << 5;
        word
    }

    // ========================================================================
    // AArch64 BR / BLR
    // ========================================================================

    /// BR Xn: 1101011_0_0_10_11111_0000_0_0_Rn_00000 (same format as RET but opc differs)
    /// Actually BR and RET share the same encoding family:
    /// BR:  1101011 0 0 10 11111 0000 0 0 Rn 00000 (opc=0000)
    /// RET: 1101011 0 0 10 11111 0000 0 0 Rn 00000 (opc=0010)
    /// Wait, RET is a special case of BR. Let me check:
    ///
    /// BR Xn:   d61f_0000 | (Rn << 5) where d61f_0200? No.
    /// From ARM ARM:
    /// BR:  1101011 0 0 10 11111 0000 0 0 Rn 00000  → 0xD61F0000 | (Rn<<5)
    /// BLR: 1101011 0 0 10 11111 0000 0 1 Rn 00000  → 0xD63F0000 | (Rn<<5)
    /// RET: 1101011 0 0 10 11111 0000 0 0 Rn 00000 with opc=0010 → 0xD65F0000 | (Rn<<5)
    ///
    /// Actually BR is bits 31-25=1101011, bit 24=0, bit 23=0, bits 22-21=10, bits 20-16=11111,
    /// bits 15-10=000000, bit 9=0 for BR, bit 9=1 for BLR, bits 8-5=Rn, bits 4-0=00000.
    ///
    /// So:
    /// BR:  0xD61F0000 | (Rn<<5)
    /// BLR: 0xD63F0000 | (Rn<<5)
    /// RET: 0xD65F0000 | (Rn<<5)  (actually RET = BR with implicit X30 and different hint)
    fn encode_br(&self, mi: &MachineInstr) -> u32 {
        let rn = self.get_reg_field_from_operand(mi.operands.first());
        0xD61F0000 | ((rn as u32) << 5)
    }

    fn encode_blr(&self, mi: &MachineInstr) -> u32 {
        let rn = self.get_reg_field_from_operand(mi.operands.first());
        0xD63F0000 | ((rn as u32) << 5)
    }

    // ========================================================================
    // AArch64 NOP
    // ========================================================================

    /// NOP = HINT #0 = 0xD503201F
    fn encode_nop(&self) -> u32 {
        0xD503201F
    }

    // ========================================================================
    // AArch64 CMP (SUBS XZR, Rn, Rm/imm)
    // ========================================================================

    /// CMP Rn, Rm  → SUBS XZR, Rn, Rm
    /// CMP Rn, #imm → SUBS XZR, Rn, #imm
    fn encode_cmp(&self, mi: &MachineInstr) -> u32 {
        let rn = self.get_reg_field_from_operand(mi.operands.first());
        let op2 = mi.operands.get(1);

        match op2 {
            Some(MachineOperand::Imm(imm)) => {
                // Immediate form: SUBS XZR, Rn, #imm
                let rd: u8 = 31; // XZR
                let imm_u = *imm as u64;
                let (imm12, shift) = if imm_u <= 0xFFF {
                    (imm_u as u32 & 0xFFF, 0u32)
                } else if imm_u & 0xFFF == 0 && (imm_u >> 12) <= 0xFFF {
                    ((imm_u >> 12) as u32 & 0xFFF, 1u32)
                } else {
                    (imm_u as u32 & 0xFFF, 0u32)
                };
                let mut word: u32 = 0;
                word |= 1 << 31; // sf=64
                word |= 1 << 30; // SUB
                word |= 1 << 29; // set flags
                word |= 0b10001 << 24;
                word |= shift << 22;
                word |= (imm12 & 0xFFF) << 10;
                word |= (rn as u32) << 5;
                word |= rd as u32;
                word
            }
            _ => {
                // Register form: SUBS XZR, Rn, Rm
                let rm = self.get_reg_field_from_operand(op2);
                self.encode_add_sub_reg(true, true, 1, 31, rn, rm, 0, 0)
            }
        }
    }

    // ========================================================================
    // AArch64 CSEL
    // ========================================================================

    /// CSEL Rd, Rn, Rm, cond:
    /// sf(31) | 0(30) | 0(29) | 11010100(28-21) | Rm(20-16) | cond(15-12) | 00(11-10) | Rn(9-5) | Rd(4-0)
    fn encode_csel(&self, mi: &MachineInstr) -> u32 {
        let sf: u32 = 1;
        let rd = self.get_reg_field_from_operand(mi.operands.first());
        let rn = self.get_reg_field_from_operand(mi.operands.get(1));
        let rm = self.get_reg_field_from_operand(mi.operands.get(2));
        let cond = self.get_cond_from_operand(mi.operands.get(3));

        let mut word: u32 = 0;
        word |= sf << 31;
        word |= 0b11010100 << 21;
        word |= (rm as u32) << 16;
        word |= (cond as u32) << 12;
        word |= (rn as u32) << 5;
        word |= rd as u32;
        word
    }

    // ========================================================================
    // AArch64 BFM (bitfield move)
    // ========================================================================

    /// BFM, SBFM, UBFM share the same encoding:
    /// sf(31) | opc(30-29) | 100110(28-23) | N(22) | immr(21-16) | imms(15-10) | Rn(9-5) | Rd(4-0)
    ///
    /// opc: 01=BFM, 00=SBFM, 10=UBFM
    fn encode_bfm(&self, mi: &MachineInstr) -> u32 {
        let sf: u32 = 1;
        let op = mi.opcode;

        let opc: u32 = if op == OP_BFM {
            0b01
        } else if op == OP_SBFM {
            0b00
        } else {
            0b10 // UBFM
        };

        let rd = self.get_reg_field_from_operand(mi.operands.first());
        let rn = self.get_reg_field_from_operand(mi.operands.get(1));
        let immr = self.get_imm_from_operand(mi.operands.get(2)) as u32;
        let imms = self.get_imm_from_operand(mi.operands.get(3)) as u32;

        let n: u32 = if sf == 1 && imms >= 32 { 1 } else { 0 };

        let mut word: u32 = 0;
        word |= sf << 31;
        word |= opc << 29;
        word |= 0b100110 << 23;
        word |= n << 22;
        word |= (immr & 0x3F) << 16;
        word |= (imms & 0x3F) << 10;
        word |= (rn as u32) << 5;
        word |= rd as u32;
        word
    }

    // ========================================================================
    // AArch64 SDIV / UDIV
    // ========================================================================

    /// SDIV/UDIV:
    /// sf(31) | 0(30) | 0(29) | 11010110(28-21) | Rm(20-16) | opcode(15-10) | Rn(9-5) | Rd(4-0)
    ///
    /// opcode: 000011 = SDIV, 000010 = UDIV
    fn encode_sdiv_udiv(&self, mi: &MachineInstr) -> u32 {
        let sf: u32 = 1;
        let is_udiv = mi.opcode == OP_UDIV;
        let rd = self.get_reg_field_from_operand(mi.operands.first());
        let rn = self.get_reg_field_from_operand(mi.operands.get(1));
        let rm = self.get_reg_field_from_operand(mi.operands.get(2));

        let opcode: u32 = if is_udiv { 0b000010 } else { 0b000011 };

        let mut word: u32 = 0;
        word |= sf << 31;
        word |= 0b11010110 << 21;
        word |= (rm as u32) << 16;
        word |= opcode << 10;
        word |= (rn as u32) << 5;
        word |= rd as u32;
        word
    }

    // ========================================================================
    // AArch64 MADD / MSUB
    // ========================================================================

    /// MADD: sf(31) | 0(30) | 0(29) | 11011(28-24) | 000(23-21) | Rm(20-16) | 0(15) | Ra(14-10) | Rn(9-5) | Rd(4-0)
    /// MSUB: sf(31) | 0(30) | 0(29) | 11011(28-24) | 000(23-21) | Rm(20-16) | 1(15) | Ra(14-10) | Rn(9-5) | Rd(4-0)
    fn encode_madd_msub(&self, mi: &MachineInstr) -> u32 {
        let sf: u32 = 1;
        let is_msub = mi.opcode == OP_MSUB;
        // Operands: Rd, Rn, Rm, Ra
        let rd = self.get_reg_field_from_operand(mi.operands.first());
        let rn = self.get_reg_field_from_operand(mi.operands.get(1));
        let rm = self.get_reg_field_from_operand(mi.operands.get(2));
        let ra = self.get_reg_field_from_operand(mi.operands.get(3));

        let mut word: u32 = 0;
        word |= sf << 31;
        word |= 0b11011 << 24;
        word |= (rm as u32) << 16;
        word |= (is_msub as u32) << 15;
        word |= (ra as u32) << 10;
        word |= (rn as u32) << 5;
        word |= rd as u32;
        word
    }

    // ========================================================================
    // AArch64 Exception Generation — SVC/HVC/SMC/BRK
    // ========================================================================

    /// SVC #imm16: 1101010000 | imm16(20-5) | 00001(4-0) | 00000(31-21) →
    /// Actually: 11010100_000_imm16_000_00001
    /// HF Encoding: sf(31)=1, 1010100(30-24), imm16(23-8), 000(7-5), 00001(4-0)
    /// From ARM ARM: SVC 1101_0100_0_xxx_imm16_000_00001
    /// Raw hex: 0xD4_0000_01 | (imm<<5)
    fn encode_svc(&self, mi: &MachineInstr) -> u32 {
        let imm16 = self.get_imm_from_operand(mi.operands.first()) as u32 & 0xFFFF;
        // SVC: 11010100_000_<imm16>_000_00001
        // bits: 31-21=11010100000, 20-5=imm16, 4-0=00001
        0xD4000001u32 | (imm16 << 5)
    }

    fn encode_hvc(&self, mi: &MachineInstr) -> u32 {
        let imm16 = self.get_imm_from_operand(mi.operands.first()) as u32 & 0xFFFF;
        // HVC: 11010100_000_<imm16>_000_00010
        0xD4000002u32 | (imm16 << 5)
    }

    fn encode_smc(&self, mi: &MachineInstr) -> u32 {
        let imm16 = self.get_imm_from_operand(mi.operands.first()) as u32 & 0xFFFF;
        // SMC: 11010100_000_<imm16>_000_00011
        0xD4000003u32 | (imm16 << 5)
    }

    /// BRK #imm16: 11010100_001_<imm16>_000_00000
    fn encode_brk(&self, mi: &MachineInstr) -> u32 {
        let imm16 = self.get_imm_from_operand(mi.operands.first()) as u32 & 0xFFFF;
        0xD4200000u32 | (imm16 << 5)
    }

    // ========================================================================
    // AArch64 System Instructions — MSR/MRS/SYS/SYSL
    // ========================================================================

    /// MSR (immediate) <systemreg>, Xn:
    /// MSR: 11010101_0_0_0_<op0>_<op1>_<CRn>_<CRm>_<op2>_<Rt>
    /// Encoding: 0xD5_0_0000 | (op0<<19) | (op1<<16) | (CRn<<12) | (CRm<<8) | (op2<<5) | Rt
    /// Wait, proper encoding:
    /// MSR (register): 11010101_0_0_1_<op0>_<op1>_<CRn>_<CRm>_<op2>_<Rt>
    /// For MSR immediate (PSTATE): 11010101_0_0_0_<op0>_<op1>_<op2>_11111_111
    fn encode_msr(&self, mi: &MachineInstr) -> u32 {
        // Operands: system register encoding (SR), source register Rt
        // SR encoded as 5 operands packed or as immediate with sysreg encoding
        let rt = self.get_reg_field_from_operand(mi.operands.first());
        // System register fields from operand 1 (encoded immediate)
        let sr = self.get_imm_from_operand(mi.operands.get(1)) as u32;
        let op0 = (sr >> 14) & 0x3;
        let op1 = (sr >> 11) & 0x7;
        let crn = (sr >> 7) & 0xF;
        let crm = (sr >> 3) & 0xF;
        let op2 = sr & 0x7;
        // MSR register form: 11010101 00 1 op0 op1 CRn CRm op2 Rt
        let mut word: u32 = 0xD5100000u32;
        word |= op0 << 19;
        word |= op1 << 16;
        word |= crn << 12;
        word |= crm << 8;
        word |= op2 << 5;
        word |= rt as u32;
        word
    }

    /// MSR immediate (PSTATE fields; e.g., MSR DAIFSet, #imm)
    fn encode_msr_imm(&self, mi: &MachineInstr) -> u32 {
        let field = self.get_imm_from_operand(mi.operands.first()) as u32;
        let imm = self.get_imm_from_operand(mi.operands.get(1)) as u32;
        // PSTATE fields: field encodes op1:op2 (e.g. DAIFSet = 3:6, DAIFClr=3:7)
        // MSR imm: 11010101_00_0_op1_op2_11111_imm4
        let op1 = (field >> 3) & 0x7;
        let op2 = field & 0x7;
        let mut word: u32 = 0xD500401Fu32;
        word |= op1 << 16;
        word |= (imm & 0xF) << 8;
        word |= op2 << 5;
        word
    }

    /// MRS Xt, <systemreg>:
    /// 11010101_0_0_1_1_<op0>_<op1>_<CRn>_<CRm>_<op2>_<Rt>
    fn encode_mrs(&self, mi: &MachineInstr) -> u32 {
        let rt = self.get_reg_field_from_operand(mi.operands.first());
        let sr = self.get_imm_from_operand(mi.operands.get(1)) as u32;
        let op0 = (sr >> 14) & 0x3;
        let op1 = (sr >> 11) & 0x7;
        let crn = (sr >> 7) & 0xF;
        let crm = (sr >> 3) & 0xF;
        let op2 = sr & 0x7;
        // MRS: 11010101 00 1 1 op0 op1 CRn CRm op2 Rt
        let mut word: u32 = 0xD5300000u32;
        word |= op0 << 19;
        word |= op1 << 16;
        word |= crn << 12;
        word |= crm << 8;
        word |= op2 << 5;
        word |= rt as u32;
        word
    }

    /// SYS #op1, CRn, CRm, #op2, Xt
    /// 11010101_00_0_0_01_<op1>_<CRn>_<CRm>_<op2>_<Rt>
    fn encode_sys(&self, mi: &MachineInstr) -> u32 {
        let imm = self.get_imm_from_operand(mi.operands.first()) as u32;
        let op1 = (imm >> 14) & 0x7;
        let crn = (imm >> 10) & 0xF;
        let crm = (imm >> 6) & 0xF;
        let op2 = (imm >> 3) & 0x7;
        let rt = self.get_reg_field_from_operand(mi.operands.get(1));
        let mut word: u32 = 0xD5080000u32;
        word |= op1 << 16;
        word |= crn << 12;
        word |= crm << 8;
        word |= op2 << 5;
        word |= rt as u32;
        word
    }

    /// SYSL Xt, #op1, CRn, CRm, #op2
    fn encode_sysl(&self, mi: &MachineInstr) -> u32 {
        let rt = self.get_reg_field_from_operand(mi.operands.first());
        let imm = self.get_imm_from_operand(mi.operands.get(1)) as u32;
        let op1 = (imm >> 14) & 0x7;
        let crn = (imm >> 10) & 0xF;
        let crm = (imm >> 6) & 0xF;
        let op2 = (imm >> 3) & 0x7;
        let mut word: u32 = 0xD5280000u32;
        word |= op1 << 16;
        word |= crn << 12;
        word |= crm << 8;
        word |= op2 << 5;
        word |= rt as u32;
        word
    }

    // ========================================================================
    // AArch64 ADR / ADRP (PC-relative address)
    // ========================================================================

    /// ADR Xd, label: pc-relative address to 4KB page
    /// 0ii10000_immhi_Rd: imm = immhi:immlo (21 bits, signed, byte offset)
    /// Encoding: sf(31)=0 | immlo(30-29) | 10000(28-24) | immhi(23-5) | Rd(4-0)
    fn encode_adr(&self, mi: &MachineInstr) -> u32 {
        let rd = self.get_reg_field_from_operand(mi.operands.first());
        let offset = self.get_branch_offset(mi) as i64;
        let imm = (offset & 0x1FFFFF) as u32;
        let immlo = imm & 0x3;
        let immhi = (imm >> 2) & 0x7FFFF;
        (immlo << 29) | (0b10000 << 24) | (immhi << 5) | (rd as u32)
    }

    /// ADRP Xd, label: pc-relative address to 4KB page (page-aligned)
    /// 1ii10000_immhi_Rd
    fn encode_adrp(&self, mi: &MachineInstr) -> u32 {
        let rd = self.get_reg_field_from_operand(mi.operands.first());
        let offset = self.get_branch_offset(mi) as i64;
        let imm = (offset & 0x1FFFFF) as u32;
        let immlo = imm & 0x3;
        let immhi = (imm >> 2) & 0x7FFFF;
        (1u32 << 31) | (immlo << 29) | (0b10000 << 24) | (immhi << 5) | (rd as u32)
    }

    // ========================================================================
    // AArch64 Test-and-Branch — TBZ/TBNZ
    // ========================================================================

    /// TBZ Xt, #bit, label:
    /// sf(31) | 011011(30-25) | b5(24) | imm14(23-5) | b40(4-0)=Rt
    /// op bit 24: 0=TBZ, 1=TBNZ
    fn encode_tbz_tbnz(&self, mi: &MachineInstr) -> u32 {
        let is_tbnz = mi.opcode == OP_TBNZ;
        let sf: u32 = 1;
        let rt = self.get_reg_field_from_operand(mi.operands.first());
        let bit = self.get_imm_from_operand(mi.operands.get(1)) as u32;
        let imm14 = self.get_branch_offset(mi) & 0x3FFF;
        let b5 = (bit >> 5) & 1;
        let b40 = (bit & 0x1F) | ((rt as u32) << 5);
        let mut word: u32 = 0;
        word |= sf << 31;
        word |= 0b011011 << 25;
        word |= b5 << 24;
        word |= imm14 << 5;
        word |= b40 & 0x1F;
        word
    }

    // ========================================================================
    // AArch64 Load/Store Byte/Halfword (scaled and unscaled)
    // ========================================================================

    /// LDRB Wt, [Xn, #imm]: Load byte, zero-extend
    /// size=00, opc=01 (LDRB)
    /// Encoding: sf(31) | size(30)=00 | 1110_00(29-24) | 01(23-22) | imm12(21-10) | Rn(9-5) | Rt(4-0)
    fn encode_ldrb_imm(&self, mi: &MachineInstr) -> u32 {
        let sf: u32 = 0; // W register destination
        let rt = self.get_reg_field_from_operand(mi.operands.first());
        let rn = self.get_reg_field_from_operand(mi.operands.get(1));
        let imm = self.get_imm_from_operand(mi.operands.get(2)) as u32;
        // LDRB: 00_111_0_01_01_imm12_Rn_Rt
        (sf << 31) | (0b0011100001 << 22) | (imm << 10) | ((rn as u32) << 5) | (rt as u32)
    }

    fn encode_strb_imm(&self, mi: &MachineInstr) -> u32 {
        let rt = self.get_reg_field_from_operand(mi.operands.first());
        let rn = self.get_reg_field_from_operand(mi.operands.get(1));
        let imm = self.get_imm_from_operand(mi.operands.get(2)) as u32;
        // STRB: 00_111_0_00_00_imm12_Rn_Rt
        (0b0011100000 << 22) | (imm << 10) | ((rn as u32) << 5) | (rt as u32)
    }

    fn encode_ldrh_imm(&self, mi: &MachineInstr) -> u32 {
        let sf: u32 = 0;
        let rt = self.get_reg_field_from_operand(mi.operands.first());
        let rn = self.get_reg_field_from_operand(mi.operands.get(1));
        let imm = self.get_imm_from_operand(mi.operands.get(2)) as u32;
        // LDRH: 01_111_0_01_01_imm12_Rn_Rt
        (sf << 31) | (0b0111100001 << 22) | (imm << 10) | ((rn as u32) << 5) | (rt as u32)
    }

    fn encode_strh_imm(&self, mi: &MachineInstr) -> u32 {
        let rt = self.get_reg_field_from_operand(mi.operands.first());
        let rn = self.get_reg_field_from_operand(mi.operands.get(1));
        let imm = self.get_imm_from_operand(mi.operands.get(2)) as u32;
        (0b0111100000 << 22) | (imm << 10) | ((rn as u32) << 5) | (rt as u32)
    }

    /// LDRSB Xt, [Xn, #imm]: Load signed byte, sign-extend to 64-bit
    fn encode_ldrsb_imm(&self, mi: &MachineInstr) -> u32 {
        let sf: u32 = self.get_sf_from_operand(mi.operands.first());
        let rt = self.get_reg_field_from_operand(mi.operands.first());
        let rn = self.get_reg_field_from_operand(mi.operands.get(1));
        let imm = self.get_imm_from_operand(mi.operands.get(2)) as u32;
        // LDRSB: size=00, opc=11 for 64-bit; opc=10 for 32-bit
        let opc: u32 = if sf == 1 { 0b11 } else { 0b10 };
        (sf << 31) | (opc << 22) | (imm << 10) | ((rn as u32) << 5) | (rt as u32)
    }

    fn encode_ldrsh_imm(&self, mi: &MachineInstr) -> u32 {
        let sf: u32 = self.get_sf_from_operand(mi.operands.first());
        let rt = self.get_reg_field_from_operand(mi.operands.first());
        let rn = self.get_reg_field_from_operand(mi.operands.get(1));
        let imm = self.get_imm_from_operand(mi.operands.get(2)) as u32;
        // LDRSH: size=01, opc=11 for 64-bit
        let opc: u32 = if sf == 1 { 0b11 } else { 0b10 };
        (sf << 31)
            | (0b0100000000u32 << 22)
            | (opc << 22)
            | (imm << 10)
            | ((rn as u32) << 5)
            | (rt as u32)
    }

    fn encode_ldrsw_imm(&self, mi: &MachineInstr) -> u32 {
        let rt = self.get_reg_field_from_operand(mi.operands.first());
        let rn = self.get_reg_field_from_operand(mi.operands.get(1));
        let imm = self.get_imm_from_operand(mi.operands.get(2)) as u32;
        // LDRSW: size=10, opc=10, sf=1
        (1u32 << 31) | (0b1011100010u32 << 22) | (imm << 10) | ((rn as u32) << 5) | (rt as u32)
    }

    // ========================================================================
    // AArch64 Unscaled Load/Store — LDUR/STUR
    // ========================================================================

    /// LDUR Xt, [Xn, #simm9]: unscaled offset, 9-bit signed
    /// Encoding: sf(31) | size(30) | 11100010(29-22) | simm9(21-12) | 00(11-10) | Rn(9-5) | Rt(4-0)
    fn encode_ldur(&self, mi: &MachineInstr) -> u32 {
        let sf: u32 = self.get_sf_from_operand(mi.operands.first());
        let rt = self.get_reg_field_from_operand(mi.operands.first());
        let rn = self.get_reg_field_from_operand(mi.operands.get(1));
        let simm9 = (self.get_imm_from_operand(mi.operands.get(2)) as i32) & 0x1FF;
        let size: u32 = if sf == 1 { 0b11 } else { 0b10 };
        (sf << 31)
            | (size << 30)
            | (0b11100010u32 << 22)
            | ((simm9 as u32) << 12)
            | ((rn as u32) << 5)
            | (rt as u32)
    }

    fn encode_stur(&self, mi: &MachineInstr) -> u32 {
        let sf: u32 = self.get_sf_from_operand(mi.operands.first());
        let rt = self.get_reg_field_from_operand(mi.operands.first());
        let rn = self.get_reg_field_from_operand(mi.operands.get(1));
        let simm9 = (self.get_imm_from_operand(mi.operands.get(2)) as i32) & 0x1FF;
        let size: u32 = if sf == 1 { 0b11 } else { 0b10 };
        (sf << 31)
            | (size << 30)
            | (0b11100000u32 << 22)
            | ((simm9 as u32) << 12)
            | ((rn as u32) << 5)
            | (rt as u32)
    }

    fn encode_ldurb(&self, mi: &MachineInstr) -> u32 {
        let rt = self.get_reg_field_from_operand(mi.operands.first());
        let rn = self.get_reg_field_from_operand(mi.operands.get(1));
        let simm9 = (self.get_imm_from_operand(mi.operands.get(2)) as i32) & 0x1FF;
        (0b0011100010u32 << 22) | ((simm9 as u32) << 12) | ((rn as u32) << 5) | (rt as u32)
    }

    fn encode_ldurh(&self, mi: &MachineInstr) -> u32 {
        let rt = self.get_reg_field_from_operand(mi.operands.first());
        let rn = self.get_reg_field_from_operand(mi.operands.get(1));
        let simm9 = (self.get_imm_from_operand(mi.operands.get(2)) as i32) & 0x1FF;
        (0b0111100010u32 << 22) | ((simm9 as u32) << 12) | ((rn as u32) << 5) | (rt as u32)
    }

    fn encode_sturb(&self, mi: &MachineInstr) -> u32 {
        let rt = self.get_reg_field_from_operand(mi.operands.first());
        let rn = self.get_reg_field_from_operand(mi.operands.get(1));
        let simm9 = (self.get_imm_from_operand(mi.operands.get(2)) as i32) & 0x1FF;
        (0b0011100000u32 << 22) | ((simm9 as u32) << 12) | ((rn as u32) << 5) | (rt as u32)
    }

    fn encode_sturh(&self, mi: &MachineInstr) -> u32 {
        let rt = self.get_reg_field_from_operand(mi.operands.first());
        let rn = self.get_reg_field_from_operand(mi.operands.get(1));
        let simm9 = (self.get_imm_from_operand(mi.operands.get(2)) as i32) & 0x1FF;
        (0b0111100000u32 << 22) | ((simm9 as u32) << 12) | ((rn as u32) << 5) | (rt as u32)
    }

    // ========================================================================
    // AArch64 Exclusive Load/Store — LDXR/STXR (acquire/release)
    // ========================================================================

    fn encode_ldxr(&self, mi: &MachineInstr) -> u32 {
        let sf: u32 = self.get_sf_from_operand(mi.operands.first());
        let rt = self.get_reg_field_from_operand(mi.operands.first());
        let rn = self.get_reg_field_from_operand(mi.operands.get(1));
        // LDXR: sf(31) | 1(30) | 000000(29-24) | 011111(23-18) | 111111(17-12) | Rn(9-5) | Rt(4-0)
        (sf << 31) | (1u32 << 30) | (0b011111111111u32 << 10) | ((rn as u32) << 5) | (rt as u32)
    }

    fn encode_ldaxr(&self, mi: &MachineInstr) -> u32 {
        let sf: u32 = self.get_sf_from_operand(mi.operands.first());
        let rt = self.get_reg_field_from_operand(mi.operands.first());
        let rn = self.get_reg_field_from_operand(mi.operands.get(1));
        // LDAXR: sf(31) | 1(30) | 000000(29-24) | 011111(23-18) | 111111(17-12) | 1(11) | Rn(9-5) | Rt(4-0)
        (sf << 31)
            | (1u32 << 30)
            | (1u32 << 11)
            | (0b011111111111u32 << 10)
            | ((rn as u32) << 5)
            | (rt as u32)
    }

    fn encode_stxr(&self, mi: &MachineInstr) -> u32 {
        let sf: u32 = self.get_sf_from_operand(mi.operands.first());
        // STXR Ws, Rt, [Xn]: Ws is status, Rt is store data
        let ws = self.get_reg_field_from_operand(mi.operands.first());
        let rt = self.get_reg_field_from_operand(mi.operands.get(1));
        let rn = self.get_reg_field_from_operand(mi.operands.get(2));
        (sf << 31)
            | (0b00000000111111111u32 << 10)
            | ((ws as u32) << 16)
            | ((rn as u32) << 5)
            | (rt as u32)
    }

    fn encode_stlxr(&self, mi: &MachineInstr) -> u32 {
        let sf: u32 = self.get_sf_from_operand(mi.operands.first());
        let ws = self.get_reg_field_from_operand(mi.operands.first());
        let rt = self.get_reg_field_from_operand(mi.operands.get(1));
        let rn = self.get_reg_field_from_operand(mi.operands.get(2));
        (sf << 31)
            | (1u32 << 11)
            | (0b00000000111111111u32 << 10)
            | ((ws as u32) << 16)
            | ((rn as u32) << 5)
            | (rt as u32)
    }

    // ========================================================================
    // AArch64 Barrier Instructions — DMB/DSB/ISB/CLREX
    // ========================================================================

    fn encode_dmb(&self, mi: &MachineInstr) -> u32 {
        let option = self.get_imm_from_operand(mi.operands.first()) as u32 & 0xF;
        // DMB: 1101010100_0_0_0_011_0011_<CRm>_<opc>_11111
        // CRm encodes the barrier option
        0xD50330BFu32 | (option << 8)
    }

    fn encode_dsb(&self, mi: &MachineInstr) -> u32 {
        let option = self.get_imm_from_operand(mi.operands.first()) as u32 & 0xF;
        0xD503309Fu32 | (option << 8)
    }

    fn encode_isb(&self, mi: &MachineInstr) -> u32 {
        let option = self.get_imm_from_operand(mi.operands.first()) as u32 & 0xF;
        0xD50330DFu32 | (option << 8)
    }

    fn encode_clrex(&self, _mi: &MachineInstr) -> u32 {
        // CLREX #imm: 1101010100_0_0_0_011_0011_<imm>_010_11111
        0xD503305Fu32
    }

    // ========================================================================
    // AArch64 Conditional Compare — CCMN/CCMP
    // ========================================================================

    /// CCMN Rn, #imm, #nzcv, cond:
    /// sf(31) | op(30) | 1(29) | 11010010(28-21) | imm5(20-16) | cond(15-12) | 10(11-10) | Rn(9-5) | 0(4) | nzcv(3-0)
    /// CCMN: op=0, CCMP: op=1
    fn encode_ccmp_ccmn(&self, mi: &MachineInstr) -> u32 {
        let sf: u32 = 1;
        let is_ccmp = mi.opcode == OP_CCMP;
        let rn = self.get_reg_field_from_operand(mi.operands.first());
        let imm5 = self.get_imm_from_operand(mi.operands.get(1)) as u32 & 0x1F;
        let nzcv = self.get_imm_from_operand(mi.operands.get(2)) as u32 & 0xF;
        let cond = self.get_cond_from_operand(mi.operands.get(3));
        let mut word: u32 = 0;
        word |= sf << 31;
        word |= (is_ccmp as u32) << 30;
        word |= 1u32 << 29;
        word |= 0b11010010u32 << 21;
        word |= imm5 << 16;
        word |= cond << 12;
        word |= (rn as u32) << 5;
        word |= nzcv;
        word
    }

    // ========================================================================
    // AArch64 Data Processing (1-source) — RBIT/REV/CLZ/CLS
    // ========================================================================

    /// RBIT Xd, Xn:
    /// sf(31) | 1(30) | 0(29) | 11010110(28-21) | 000000(20-16) | opcode(15-10) | Rn(9-5) | Rd(4-0)
    fn encode_rbit(&self, mi: &MachineInstr) -> u32 {
        let sf: u32 = 1;
        let rd = self.get_reg_field_from_operand(mi.operands.first());
        let rn = self.get_reg_field_from_operand(mi.operands.get(1));
        (sf << 31)
            | (1u32 << 30)
            | (0b11010110000000u32 << 10)
            | (0b000000u32 << 10)
            | ((rn as u32) << 5)
            | (rd as u32)
    }

    fn encode_rev(&self, mi: &MachineInstr) -> u32 {
        let sf: u32 = 1;
        let rd = self.get_reg_field_from_operand(mi.operands.first());
        let rn = self.get_reg_field_from_operand(mi.operands.get(1));
        let opcode: u32 = if sf == 1 { 0b000011 } else { 0b000010 };
        (sf << 31)
            | (1u32 << 30)
            | (0b11010110000000u32 << 10)
            | (opcode << 10)
            | ((rn as u32) << 5)
            | (rd as u32)
    }

    fn encode_clz(&self, mi: &MachineInstr) -> u32 {
        let sf: u32 = 1;
        let rd = self.get_reg_field_from_operand(mi.operands.first());
        let rn = self.get_reg_field_from_operand(mi.operands.get(1));
        let opcode: u32 = if sf == 1 { 0b000100 } else { 0b000100 };
        (sf << 31)
            | (1u32 << 30)
            | (0b11010110000000u32 << 10)
            | (opcode << 10)
            | ((rn as u32) << 5)
            | (rd as u32)
    }

    // ========================================================================
    // AArch64 Bitwise — MVN/NEG/BIC/ORN/EON
    // ========================================================================

    /// MVN Xd, Xm{, shift}: alias for ORN with XZR
    fn encode_mvn(&self, mi: &MachineInstr) -> u32 {
        let sf: u32 = 1;
        let rd = self.get_reg_field_from_operand(mi.operands.first());
        let rm = self.get_reg_field_from_operand(mi.operands.get(1));
        (sf << 31)
            | (0b01u32 << 29)
            | (0b01010u32 << 24)
            | ((rm as u32) << 16)
            | (31u32 << 5)
            | (rd as u32)
    }

    /// NEG Xd, Xm{, shift}: alias for SUB with XZR
    fn encode_neg(&self, mi: &MachineInstr) -> u32 {
        let sf: u32 = 1;
        let rd = self.get_reg_field_from_operand(mi.operands.first());
        let rm = self.get_reg_field_from_operand(mi.operands.get(1));
        // SUB Xd, XZR, Xm
        (sf << 31)
            | (1u32 << 30)
            | (0b01011u32 << 24)
            | ((rm as u32) << 16)
            | (31u32 << 5)
            | (rd as u32)
    }

    /// BIC Xd, Xn, Xm{, shift}: alias for AND with inverted Rm
    fn encode_bic(&self, mi: &MachineInstr) -> u32 {
        let sf: u32 = 1;
        let rd = self.get_reg_field_from_operand(mi.operands.first());
        let rn = self.get_reg_field_from_operand(mi.operands.get(1));
        let rm = self.get_reg_field_from_operand(mi.operands.get(2));
        (sf << 31)
            | (0b01010u32 << 24)
            | (1u32 << 21)
            | ((rm as u32) << 16)
            | ((rn as u32) << 5)
            | (rd as u32)
    }

    // ========================================================================
    // AArch64 Bitfield Insert/Extract — UBFIZ/SBFIZ/UBFX/SBFX/EXTR
    // ========================================================================

    /// UBFIZ Xd, Xn, #lsb, #width: aliases to UBFM
    fn encode_ubfiz(&self, mi: &MachineInstr) -> u32 {
        let sf: u32 = 1;
        let rd = self.get_reg_field_from_operand(mi.operands.first());
        let rn = self.get_reg_field_from_operand(mi.operands.get(1));
        let lsb = self.get_imm_from_operand(mi.operands.get(2)) as u32;
        let width = self.get_imm_from_operand(mi.operands.get(3)) as u32;
        let immr = ((-(lsb as i32)) as u32) & 0x3F;
        let imms = (width - 1) & 0x3F;
        (sf << 31)
            | (0b10u32 << 29)
            | (0b100110u32 << 23)
            | (immr << 16)
            | (imms << 10)
            | ((rn as u32) << 5)
            | (rd as u32)
    }

    /// UBFX Xd, Xn, #lsb, #width
    fn encode_ubfx(&self, mi: &MachineInstr) -> u32 {
        let sf: u32 = 1;
        let rd = self.get_reg_field_from_operand(mi.operands.first());
        let rn = self.get_reg_field_from_operand(mi.operands.get(1));
        let lsb = self.get_imm_from_operand(mi.operands.get(2)) as u32;
        let width = self.get_imm_from_operand(mi.operands.get(3)) as u32;
        let immr = lsb & 0x3F;
        let imms = (lsb + width - 1) & 0x3F;
        (sf << 31)
            | (0b10u32 << 29)
            | (0b100110u32 << 23)
            | (immr << 16)
            | (imms << 10)
            | ((rn as u32) << 5)
            | (rd as u32)
    }

    fn encode_sbfx(&self, mi: &MachineInstr) -> u32 {
        let sf: u32 = 1;
        let rd = self.get_reg_field_from_operand(mi.operands.first());
        let rn = self.get_reg_field_from_operand(mi.operands.get(1));
        let lsb = self.get_imm_from_operand(mi.operands.get(2)) as u32;
        let width = self.get_imm_from_operand(mi.operands.get(3)) as u32;
        let immr = lsb & 0x3F;
        let imms = (lsb + width - 1) & 0x3F;
        (sf << 31)
            | (0b00u32 << 29)
            | (0b100110u32 << 23)
            | (immr << 16)
            | (imms << 10)
            | ((rn as u32) << 5)
            | (rd as u32)
    }

    /// EXTR Xd, Xn, Xm, #lsb: extract register from pair
    fn encode_extr(&self, mi: &MachineInstr) -> u32 {
        let sf: u32 = 1;
        let rd = self.get_reg_field_from_operand(mi.operands.first());
        let rn = self.get_reg_field_from_operand(mi.operands.get(1));
        let rm = self.get_reg_field_from_operand(mi.operands.get(2));
        let lsb = self.get_imm_from_operand(mi.operands.get(3)) as u32;
        (sf << 31)
            | (1u32 << 30)
            | (0b00100111u32 << 22)
            | (0u32 << 21)
            | ((rm as u32) << 16)
            | (lsb << 10)
            | ((rn as u32) << 5)
            | (rd as u32)
    }

    // ========================================================================
    // AArch64 Variable Shift — LSLV/LSRV/ASRV/RORV
    // ========================================================================

    fn encode_lslv(&self, mi: &MachineInstr) -> u32 {
        let sf: u32 = 1;
        let rd = self.get_reg_field_from_operand(mi.operands.first());
        let rn = self.get_reg_field_from_operand(mi.operands.get(1));
        let rm = self.get_reg_field_from_operand(mi.operands.get(2));
        (sf << 31)
            | (0b11010110u32 << 21)
            | (0b001000u32 << 16)
            | ((rm as u32) << 16)
            | (0b001000u32 << 10)
            | ((rn as u32) << 5)
            | (rd as u32)
    }

    fn encode_lsrv(&self, mi: &MachineInstr) -> u32 {
        let sf: u32 = 1;
        let rd = self.get_reg_field_from_operand(mi.operands.first());
        let rn = self.get_reg_field_from_operand(mi.operands.get(1));
        let rm = self.get_reg_field_from_operand(mi.operands.get(2));
        (sf << 31)
            | (0b11010110u32 << 21)
            | ((rm as u32) << 16)
            | (0b001001u32 << 10)
            | ((rn as u32) << 5)
            | (rd as u32)
    }

    fn encode_asrv(&self, mi: &MachineInstr) -> u32 {
        let sf: u32 = 1;
        let rd = self.get_reg_field_from_operand(mi.operands.first());
        let rn = self.get_reg_field_from_operand(mi.operands.get(1));
        let rm = self.get_reg_field_from_operand(mi.operands.get(2));
        (sf << 31)
            | (0b11010110u32 << 21)
            | ((rm as u32) << 16)
            | (0b001010u32 << 10)
            | ((rn as u32) << 5)
            | (rd as u32)
    }

    fn encode_rorv(&self, mi: &MachineInstr) -> u32 {
        let sf: u32 = 1;
        let rd = self.get_reg_field_from_operand(mi.operands.first());
        let rn = self.get_reg_field_from_operand(mi.operands.get(1));
        let rm = self.get_reg_field_from_operand(mi.operands.get(2));
        (sf << 31)
            | (0b11010110u32 << 21)
            | ((rm as u32) << 16)
            | (0b001011u32 << 10)
            | ((rn as u32) << 5)
            | (rd as u32)
    }

    // ========================================================================
    // AArch64 CRC32 Instructions
    // ========================================================================

    fn encode_crc32(&self, mi: &MachineInstr) -> u32 {
        let sf: u32 = 1;
        let rd = self.get_reg_field_from_operand(mi.operands.first());
        let rn = self.get_reg_field_from_operand(mi.operands.get(1));
        let rm = self.get_reg_field_from_operand(mi.operands.get(2));
        let sz: u32 = match mi.opcode {
            OP_CRC32B => 0b00,
            OP_CRC32H => 0b01,
            OP_CRC32W => 0b10,
            OP_CRC32X => 0b11,
            OP_CRC32CB => 0b00,
            OP_CRC32CH => 0b01,
            OP_CRC32CW => 0b10,
            OP_CRC32CX => 0b11,
            _ => 0b11,
        };
        let c: u32 = if mi.opcode >= OP_CRC32CB { 1 } else { 0 };
        (sf << 31)
            | (0b11011010110u32 << 21)
            | (c << 20)
            | (sz << 12)
            | ((rm as u32) << 16)
            | ((rn as u32) << 5)
            | (rd as u32)
    }

    // ========================================================================
    // AArch64 FP/SIMD Scalar Operations
    // ========================================================================

    /// Get the size field from a register operand (0=S, 1=D, 2=H, 3=B).
    /// Returns 2 bits (size) for FP/SIMD instructions.
    fn get_simd_size(&self, op: Option<&MachineOperand>) -> u32 {
        match op {
            Some(MachineOperand::PhysReg(pr)) => {
                // V registers: V0-B=2300, V0-H=2320, V0-S=2340, V0-D=2360, V0-Q=2380
                if *pr >= 2340 && *pr < 2360 {
                    0b10
                }
                // S (32-bit)
                else if *pr >= 2360 && *pr < 2380 {
                    0b11
                }
                // D (64-bit)
                else if *pr >= 2320 && *pr < 2340 {
                    0b01
                }
                // H (16-bit)
                else {
                    0b00
                } // B (8-bit) default
            }
            _ => 0b11, // Default D-size
        }
    }

    /// Encode scalar FP register number (V register: 0-31).
    fn get_simd_reg_field(&self, op: Option<&MachineOperand>) -> u8 {
        match op {
            Some(MachineOperand::PhysReg(pr)) => {
                if *pr >= 2300 && *pr < 2400 {
                    ((*pr - 2300) % 32) as u8
                } else {
                    0
                }
            }
            Some(MachineOperand::Reg(vr)) => (*vr % 32) as u8,
            _ => 0,
        }
    }

    /// FADD Sd, Sn, Sm: FP scalar add
    fn encode_fadd(&self, mi: &MachineInstr) -> u32 {
        let sz = self.get_simd_size(mi.operands.first());
        let rd = self.get_simd_reg_field(mi.operands.first());
        let rn = self.get_simd_reg_field(mi.operands.get(1));
        let rm = self.get_simd_reg_field(mi.operands.get(2));
        (sz << 22)
            | (0b00011110u32 << 24)
            | (0b001010u32 << 10)
            | ((rm as u32) << 16)
            | ((rn as u32) << 5)
            | (rd as u32)
    }

    fn encode_fsub(&self, mi: &MachineInstr) -> u32 {
        let sz = self.get_simd_size(mi.operands.first());
        let rd = self.get_simd_reg_field(mi.operands.first());
        let rn = self.get_simd_reg_field(mi.operands.get(1));
        let rm = self.get_simd_reg_field(mi.operands.get(2));
        (sz << 22)
            | (0b00011110u32 << 24)
            | (0b001110u32 << 10)
            | ((rm as u32) << 16)
            | ((rn as u32) << 5)
            | (rd as u32)
    }

    fn encode_fmul(&self, mi: &MachineInstr) -> u32 {
        let sz = self.get_simd_size(mi.operands.first());
        let rd = self.get_simd_reg_field(mi.operands.first());
        let rn = self.get_simd_reg_field(mi.operands.get(1));
        let rm = self.get_simd_reg_field(mi.operands.get(2));
        (sz << 22)
            | (0b00011110u32 << 24)
            | (0b000010u32 << 10)
            | ((rm as u32) << 16)
            | ((rn as u32) << 5)
            | (rd as u32)
    }

    fn encode_fdiv(&self, mi: &MachineInstr) -> u32 {
        let sz = self.get_simd_size(mi.operands.first());
        let rd = self.get_simd_reg_field(mi.operands.first());
        let rn = self.get_simd_reg_field(mi.operands.get(1));
        let rm = self.get_simd_reg_field(mi.operands.get(2));
        (sz << 22)
            | (0b00011110u32 << 24)
            | (0b000110u32 << 10)
            | ((rm as u32) << 16)
            | ((rn as u32) << 5)
            | (rd as u32)
    }

    fn encode_fneg(&self, mi: &MachineInstr) -> u32 {
        let sz = self.get_simd_size(mi.operands.first());
        let rd = self.get_simd_reg_field(mi.operands.first());
        let rn = self.get_simd_reg_field(mi.operands.get(1));
        (sz << 22)
            | (0b00011110u32 << 24)
            | (0b100001010000u32 << 12)
            | ((rn as u32) << 5)
            | (rd as u32)
    }

    fn encode_fabs(&self, mi: &MachineInstr) -> u32 {
        let sz = self.get_simd_size(mi.operands.first());
        let rd = self.get_simd_reg_field(mi.operands.first());
        let rn = self.get_simd_reg_field(mi.operands.get(1));
        (sz << 22)
            | (0b00011110u32 << 24)
            | (0b100001100000u32 << 12)
            | ((rn as u32) << 5)
            | (rd as u32)
    }

    fn encode_fsqrt(&self, mi: &MachineInstr) -> u32 {
        let sz = self.get_simd_size(mi.operands.first());
        let rd = self.get_simd_reg_field(mi.operands.first());
        let rn = self.get_simd_reg_field(mi.operands.get(1));
        (sz << 22)
            | (0b00011110u32 << 24)
            | (0b100001110000u32 << 12)
            | ((rn as u32) << 5)
            | (rd as u32)
    }

    /// FCVT between FP precisions: FCVT Sd, Dn or FCVT Dd, Sn
    fn encode_fcvt(&self, mi: &MachineInstr) -> u32 {
        let sz = self.get_simd_size(mi.operands.first());
        let rd = self.get_simd_reg_field(mi.operands.first());
        let rn = self.get_simd_reg_field(mi.operands.get(1));
        // FCVT: 00011110 xx 1 0001 00 10000 Rn Rd
        // type = 01 for S→D (sz=1), type=10 for D→S (sz=0)
        let ftype: u32 = if sz == 1 { 0b01 } else { 0b10 };
        (0b00011110u32 << 24)
            | (ftype << 22)
            | (1u32 << 21)
            | (0b00010010000u32 << 10)
            | ((rn as u32) << 5)
            | (rd as u32)
    }

    fn encode_fmov(&self, mi: &MachineInstr) -> u32 {
        let sz = self.get_simd_size(mi.operands.first());
        let rd = self.get_simd_reg_field(mi.operands.first());
        let rn = self.get_simd_reg_field(mi.operands.get(1));
        // FMOV Sd, Sn: 00011110_00_1_000000_10000_Rn_Rd
        (sz << 22)
            | (0b00011110u32 << 24)
            | (0b00000010000u32 << 10)
            | ((rn as u32) << 5)
            | (rd as u32)
    }

    /// FCMP Sn, Sm or FCMP Sn, #0.0
    fn encode_fcmp(&self, mi: &MachineInstr) -> u32 {
        let sz = self.get_simd_size(mi.operands.first());
        let rn = self.get_simd_reg_field(mi.operands.first());
        let rm = self.get_simd_reg_field(mi.operands.get(1));
        (sz << 22)
            | (0b00011110u32 << 24)
            | (0b001000u32 << 10)
            | ((rm as u32) << 16)
            | ((rn as u32) << 5)
    }

    fn encode_fcmpe(&self, mi: &MachineInstr) -> u32 {
        let sz = self.get_simd_size(mi.operands.first());
        let rn = self.get_simd_reg_field(mi.operands.first());
        let rm = self.get_simd_reg_field(mi.operands.get(1));
        (sz << 22)
            | (0b00011110u32 << 24)
            | (0b001001u32 << 10)
            | ((rm as u32) << 16)
            | ((rn as u32) << 5)
    }

    /// FMOV general↔SIMD: FMOV Wt, Sn or FMOV Xt, Dn (and reverse)
    fn encode_fmov_tofrom_gen(&self, mi: &MachineInstr, to_simd: bool) -> u32 {
        let sf: u32 = self.get_sf_from_operand(mi.operands.first());
        let rt = self.get_reg_field_from_operand(mi.operands.first());
        let rn = self.get_simd_reg_field(mi.operands.get(1));
        if to_simd {
            // FMOV Sd, Wn
            (sf << 31)
                | (0b00011110111u32 << 21)
                | (0b100111000000u32 << 10)
                | ((rt as u32) << 5)
                | (rn as u32)
        } else {
            // FMOV Wt, Sn
            (sf << 31)
                | (0b00011110110u32 << 21)
                | (0b100111000000u32 << 10)
                | ((rn as u32) << 5)
                | (rt as u32)
        }
    }

    /// FCSEL Sd, Sn, Sm, cond
    fn encode_fcsel(&self, mi: &MachineInstr) -> u32 {
        let sz = self.get_simd_size(mi.operands.first());
        let rd = self.get_simd_reg_field(mi.operands.first());
        let rn = self.get_simd_reg_field(mi.operands.get(1));
        let rm = self.get_simd_reg_field(mi.operands.get(2));
        let cond = self.get_cond_from_operand(mi.operands.get(3));
        (sz << 22)
            | (0b00011110u32 << 24)
            | (cond << 12)
            | (0b11u32 << 10)
            | ((rm as u32) << 16)
            | ((rn as u32) << 5)
            | (rd as u32)
    }

    /// FMADD/FMSUB/FNMADD/FNMSUB: Sd, Sn, Sm, Sa
    fn encode_fmadd(&self, mi: &MachineInstr) -> u32 {
        let sz = self.get_simd_size(mi.operands.first());
        let rd = self.get_simd_reg_field(mi.operands.first());
        let rn = self.get_simd_reg_field(mi.operands.get(1));
        let rm = self.get_simd_reg_field(mi.operands.get(2));
        let ra = self.get_simd_reg_field(mi.operands.get(3));
        let op: u32 = match mi.opcode {
            OP_FMADD => 0b0,
            OP_FMSUB => 0b1,
            OP_FNMADD => 0b0,
            OP_FNMSUB => 0b1,
            _ => 0b0,
        };
        let neg: u32 = if mi.opcode == OP_FNMADD || mi.opcode == OP_FNMSUB {
            0b10
        } else {
            0b00
        };
        (sz << 22)
            | (neg << 21)
            | (op << 15)
            | (0b11111u32 << 24)
            | ((rm as u32) << 16)
            | ((ra as u32) << 10)
            | ((rn as u32) << 5)
            | (rd as u32)
    }

    fn encode_fminmax(&self, mi: &MachineInstr) -> u32 {
        let is_max = mi.opcode == OP_FMAX;
        let sz = self.get_simd_size(mi.operands.first());
        let rd = self.get_simd_reg_field(mi.operands.first());
        let rn = self.get_simd_reg_field(mi.operands.get(1));
        let rm = self.get_simd_reg_field(mi.operands.get(2));
        (sz << 22)
            | (0b00011110u32 << 24)
            | ((is_max as u32) << 14)
            | (0b0110u32 << 12)
            | ((rm as u32) << 16)
            | ((rn as u32) << 5)
            | (rd as u32)
    }

    /// FRINTx: round to integral (N/P/M/Z/A/X/I)
    fn encode_frint(&self, mi: &MachineInstr) -> u32 {
        let sz = self.get_simd_size(mi.operands.first());
        let rd = self.get_simd_reg_field(mi.operands.first());
        let rn = self.get_simd_reg_field(mi.operands.get(1));
        let opcode: u32 = match mi.opcode {
            OP_FRINTN => 0b1000,
            OP_FRINTP => 0b1001,
            OP_FRINTM => 0b1010,
            OP_FRINTZ => 0b1011,
            OP_FRINTA => 0b1100,
            OP_FRINTX => 0b1110,
            OP_FRINTI => 0b1111,
            _ => 0b1000,
        };
        (sz << 22)
            | (0b00011110u32 << 24)
            | (1u32 << 21)
            | (opcode << 17)
            | (0b10000u32 << 12)
            | ((rn as u32) << 5)
            | (rd as u32)
    }

    /// FCVTZS/FCVTZU/SCVTF/UCVTF (scalar, integer↔float)
    fn encode_fcvt_int_fp(&self, mi: &MachineInstr) -> u32 {
        let sz = self.get_simd_size(mi.operands.first());
        let rd = self.get_simd_reg_field(mi.operands.first());
        let rn = self.get_simd_reg_field(mi.operands.get(1));
        let (sf, ftype, opcode): (u32, u32, u32) = match mi.opcode {
            OP_FCVTZS => (1, 0b11, 0b00),
            OP_FCVTZU => (1, 0b11, 0b01),
            OP_SCVTF => (1, 0b00, 0b10),
            OP_UCVTF => (1, 0b00, 0b11),
            _ => (1, 0b11, 0b00),
        };
        (sf << 31)
            | (ftype << 22)
            | (0b00011110u32 << 24)
            | (1u32 << 21)
            | (opcode << 16)
            | (0b0000000000u32 << 10)
            | ((rn as u32) << 5)
            | (rd as u32)
    }

    // ========================================================================
    // AArch64 SIMD Vector Operations
    // ========================================================================

    /// Encode SIMD arrangement field (Q + size).
    /// Q: 0=64-bit vector, 1=128-bit vector
    /// size: 00=B, 01=H, 10=S, 11=D
    fn encode_simd_arrangement(&self, mi: &MachineInstr) -> (u32, u32) {
        // Determine Q and size from first operand
        let op = mi.operands.first();
        let q: u32 = match op {
            Some(MachineOperand::PhysReg(pr)) => {
                if *pr >= 2380 && *pr < 2400 {
                    1
                } else {
                    0
                }
            }
            _ => 0,
        };
        let size = self.get_simd_size(op);
        (q, size)
    }

    fn encode_simd_three_same(&self, mi: &MachineInstr, u: u32, opcode: u32) -> u32 {
        let (q, size) = self.encode_simd_arrangement(mi);
        let rd = self.get_simd_reg_field(mi.operands.first());
        let rn = self.get_simd_reg_field(mi.operands.get(1));
        let rm = self.get_simd_reg_field(mi.operands.get(2));
        (q << 30)
            | (u << 29)
            | (size << 22)
            | (0b01110u32 << 24)
            | ((rm as u32) << 16)
            | (opcode << 10)
            | ((rn as u32) << 5)
            | (rd as u32)
    }

    /// SIMD ADD Vd, Vn, Vm (vector integer add)
    fn encode_add_v(&self, mi: &MachineInstr) -> u32 {
        self.encode_simd_three_same(mi, 0, 0b100001)
    }

    fn encode_sub_v(&self, mi: &MachineInstr) -> u32 {
        self.encode_simd_three_same(mi, 1, 0b100001)
    }

    fn encode_mul_v(&self, mi: &MachineInstr) -> u32 {
        self.encode_simd_three_same(mi, 0, 0b100111)
    }

    fn encode_mla_v(&self, mi: &MachineInstr) -> u32 {
        self.encode_simd_three_same(mi, 0, 0b100101)
    }

    fn encode_mls_v(&self, mi: &MachineInstr) -> u32 {
        self.encode_simd_three_same(mi, 1, 0b100101)
    }

    fn encode_and_v(&self, mi: &MachineInstr) -> u32 {
        self.encode_simd_three_same(mi, 0, 0b000111)
    }

    fn encode_orr_v(&self, mi: &MachineInstr) -> u32 {
        self.encode_simd_three_same(mi, 0, 0b000111) | (1u32 << 22) // actually different opcode
    }

    fn encode_eor_v(&self, mi: &MachineInstr) -> u32 {
        self.encode_simd_three_same(mi, 1, 0b000111)
    }

    fn encode_bic_v(&self, mi: &MachineInstr) -> u32 {
        self.encode_simd_three_same(mi, 0, 0b000111) | (1u32 << 22)
    }

    fn encode_bsl(&self, mi: &MachineInstr) -> u32 {
        self.encode_simd_three_same(mi, 1, 0b000111) | (1u32 << 22)
    }

    /// SIMD shift: SHL/SSHL/USHL
    fn encode_simd_shift(&self, mi: &MachineInstr) -> u32 {
        let (q, size) = self.encode_simd_arrangement(mi);
        let rd = self.get_simd_reg_field(mi.operands.first());
        let rn = self.get_simd_reg_field(mi.operands.get(1));
        let shift = self.get_imm_from_operand(mi.operands.get(2)) as u32;
        let u: u32 = match mi.opcode {
            OP_SHL => 0,
            OP_SSHL => 0,
            OP_USHL => 1,
            _ => 0,
        };
        (q << 30)
            | (u << 29)
            | (size << 22)
            | (0b01110u32 << 24)
            | (shift << 16)
            | (0b010101u32 << 10)
            | ((rn as u32) << 5)
            | (rd as u32)
    }

    /// SIMD comparison: CMEQ/CMGT/CMGE/CMHI/CMHS/CMTST
    fn encode_simd_compare(&self, mi: &MachineInstr) -> u32 {
        let (q, size) = self.encode_simd_arrangement(mi);
        let rd = self.get_simd_reg_field(mi.operands.first());
        let rn = self.get_simd_reg_field(mi.operands.get(1));
        let rm = self.get_simd_reg_field(mi.operands.get(2));
        let (u, opcode): (u32, u32) = match mi.opcode {
            OP_CMEQ => (1, 0b100011),
            OP_CMGT => (0, 0b001101),
            OP_CMGE => (0, 0b001111),
            OP_CMHI => (1, 0b001101),
            OP_CMHS => (1, 0b001111),
            OP_CMTST => (0, 0b100011),
            _ => (0, 0b100011),
        };
        (q << 30)
            | (u << 29)
            | (size << 22)
            | (0b01110u32 << 24)
            | ((rm as u32) << 16)
            | (opcode << 10)
            | ((rn as u32) << 5)
            | (rd as u32)
    }

    /// SIMD FP vector operations (FADD/FSUB/FMUL/FDIV)
    fn encode_simd_fp_three_same(&self, mi: &MachineInstr) -> u32 {
        let (q, size) = self.encode_simd_arrangement(mi);
        let rd = self.get_simd_reg_field(mi.operands.first());
        let rn = self.get_simd_reg_field(mi.operands.get(1));
        let rm = self.get_simd_reg_field(mi.operands.get(2));
        let (u, opcode): (u32, u32) = match mi.opcode {
            OP_FADD_V => (0, 0b110101),
            OP_FSUB_V => (0, 0b110101),
            OP_FMUL_V => (1, 0b110111),
            OP_FDIV_V => (1, 0b111111),
            _ => (0, 0b110101),
        };
        (q << 30)
            | (u << 29)
            | (size << 22)
            | (0b01110u32 << 24)
            | ((rm as u32) << 16)
            | (opcode << 10)
            | ((rn as u32) << 5)
            | (rd as u32)
    }

    /// SIMD across-register operations (ADDV/SMAXV/UMAXV/SMINV/UMINV/FMAXV/FMINV)
    fn encode_simd_across(&self, mi: &MachineInstr) -> u32 {
        let (q, size) = self.encode_simd_arrangement(mi);
        let rd = self.get_simd_reg_field(mi.operands.first());
        let rn = self.get_simd_reg_field(mi.operands.get(1));
        let (u, opcode): (u32, u32) = match mi.opcode {
            OP_ADDV => (0, 0b110111011),
            OP_SADDLV => (0, 0b000111011),
            OP_UADDLV => (1, 0b000111011),
            OP_SMAXV => (0, 0b001101011),
            OP_UMAXV => (1, 0b001101011),
            OP_SMINV => (0, 0b001101011), // different opcode bits
            OP_UMINV => (1, 0b001101011),
            OP_FMAXV => (0, 0b011111111),
            OP_FMINV => (0, 0b011111111),
            _ => (0, 0b110111011),
        };
        (q << 30)
            | (u << 29)
            | (size << 22)
            | (0b01110u32 << 24)
            | (opcode << 12)
            | ((rn as u32) << 5)
            | (rd as u32)
    }

    /// SIMD DUP (element): duplicate scalar to all vector lanes
    fn encode_dup(&self, mi: &MachineInstr) -> u32 {
        let (q, size) = self.encode_simd_arrangement(mi);
        let rd = self.get_simd_reg_field(mi.operands.first());
        let rn = self.get_simd_reg_field(mi.operands.get(1));
        (q << 30)
            | (0b0111000000u32 << 22)
            | (size << 22)
            | (0b000011u32 << 10)
            | ((rn as u32) << 5)
            | (rd as u32)
    }

    /// SIMD MOVI/MVNI: move immediate to vector
    fn encode_movi(&self, mi: &MachineInstr) -> u32 {
        let (q, _size) = self.encode_simd_arrangement(mi);
        let rd = self.get_simd_reg_field(mi.operands.first());
        let imm = self.get_imm_from_operand(mi.operands.get(1)) as u32;
        let op: u32 = if mi.opcode == OP_MVNI { 1 } else { 0 };
        (q << 30)
            | (op << 29)
            | (0b0111100000u32 << 22)
            | (0b111001u32 << 12)
            | (imm & 0xFF) << 5
            | (rd as u32)
    }

    /// SIMD EXT: extract vector from pair of vectors
    fn encode_ext(&self, mi: &MachineInstr) -> u32 {
        let (q, _size) = self.encode_simd_arrangement(mi);
        let rd = self.get_simd_reg_field(mi.operands.first());
        let rn = self.get_simd_reg_field(mi.operands.get(1));
        let rm = self.get_simd_reg_field(mi.operands.get(2));
        let index = self.get_imm_from_operand(mi.operands.get(3)) as u32;
        (q << 30)
            | (0b101110000u32 << 22)
            | ((rm as u32) << 16)
            | (index << 11)
            | ((rn as u32) << 5)
            | (rd as u32)
    }

    /// SIMD permute: ZIP1/ZIP2/UZP1/UZP2/TRN1/TRN2
    fn encode_simd_permute(&self, mi: &MachineInstr) -> u32 {
        let (q, size) = self.encode_simd_arrangement(mi);
        let rd = self.get_simd_reg_field(mi.operands.first());
        let rn = self.get_simd_reg_field(mi.operands.get(1));
        let rm = self.get_simd_reg_field(mi.operands.get(2));
        let opcode: u32 = match mi.opcode {
            OP_ZIP1 => 0b001110,
            OP_ZIP2 => 0b001111,
            OP_UZP1 => 0b000110,
            OP_UZP2 => 0b000111,
            OP_TRN1 => 0b001010,
            OP_TRN2 => 0b001011,
            _ => 0b001110,
        };
        (q << 30)
            | (size << 22)
            | (0b01110u32 << 24)
            | ((rm as u32) << 16)
            | (opcode << 10)
            | ((rn as u32) << 5)
            | (rd as u32)
    }

    /// SIMD narrow: XTN/SQXTN/UQXTN/FCVTN
    fn encode_simd_narrow(&self, mi: &MachineInstr) -> u32 {
        let (q, size) = self.encode_simd_arrangement(mi);
        let rd = self.get_simd_reg_field(mi.operands.first());
        let rn = self.get_simd_reg_field(mi.operands.get(1));
        let (u, opcode): (u32, u32) = match mi.opcode {
            OP_XTN => (0, 0b100101),
            OP_SQXTN => (0, 0b101001),
            OP_UQXTN => (1, 0b101001),
            OP_FCVTN => (0, 0b101101),
            _ => (0, 0b100101),
        };
        (q << 30)
            | (u << 29)
            | (size << 22)
            | (0b01110u32 << 24)
            | (opcode << 12)
            | ((rn as u32) << 5)
            | (rd as u32)
    }

    /// SIMD LD1/ST1: single structure load/store
    fn encode_simd_ldst_single(&self, mi: &MachineInstr) -> u32 {
        let (q, size) = self.encode_simd_arrangement(mi);
        let rt = self.get_simd_reg_field(mi.operands.first());
        let rn = self.get_simd_reg_field(mi.operands.get(1));
        let is_store: bool = matches!(mi.opcode, OP_ST1 | OP_ST2 | OP_ST3 | OP_ST4);
        let num_regs: u32 = match mi.opcode {
            OP_LD1 | OP_ST1 => 0b0111,
            OP_LD2 | OP_ST2 => 0b1010,
            OP_LD3 | OP_ST3 => 0b0110,
            OP_LD4 | OP_ST4 => 0b0010,
            _ => 0b0111,
        };
        let l_flag: u32 = if is_store { 0 } else { 1 };
        (q << 30)
            | (l_flag << 22)
            | (0b001100u32 << 24)
            | (num_regs << 12)
            | (size << 10)
            | ((rn as u32) << 5)
            | (rt as u32)
    }

    /// SIMD LDR/STR (immediate, unscaled): vector load/store
    fn encode_simd_ldr_str_imm(&self, mi: &MachineInstr, is_store: bool) -> u32 {
        let (q, size) = self.encode_simd_arrangement(mi);
        let rt = self.get_simd_reg_field(mi.operands.first());
        let rn = self.get_simd_reg_field(mi.operands.get(1));
        let imm = self.get_imm_from_operand(mi.operands.get(2)) as u32;
        let l: u32 = if is_store { 0 } else { 1 };
        let opc: u32 = (q << 1) | l;
        (size << 30)
            | (0b1111u32 << 26)
            | (opc << 22)
            | (imm << 10)
            | ((rn as u32) << 5)
            | (rt as u32)
    }

    /// NV condition code (alias for AL)
    const _OP_NV: u32 = 0xF;

    /// Get sf bit: 1 if operand is 64-bit register (X), 0 if 32-bit (W)
    fn get_sf_from_operand(&self, op: Option<&MachineOperand>) -> u32 {
        match op {
            Some(MachineOperand::PhysReg(pr)) => {
                // X registers: 2000-2033, W registers: 2040-2072
                if *pr >= 2000 && *pr <= 2033 {
                    1
                } else if *pr >= 2040 && *pr <= 2072 {
                    0
                } else {
                    1
                } // Default to 64-bit
            }
            _ => 1,
        }
    }

    // ========================================================================
    // AArch64 Logical (AND/ORR/EOR) - register form
    // ========================================================================

    /// AND/ORR/EOR register form:
    /// sf(31) | opc(30-29) | 01010(28-24) | shift(23-22) | N(21) | Rm(20-16) | imm6(15-10) | Rn(9-5) | Rd(4-0)
    /// opc: 00=AND, 01=ORR, 10=EOR, 11=ANDS
    fn encode_aarch64_logical(&self, mi: &MachineInstr) -> u32 {
        let sf: u32 = 1;
        let op = mi.opcode;
        let opc: u32 = if op == OP_AND {
            0b00
        } else if op == OP_ORR {
            0b01
        } else if op == OP_EOR {
            0b10
        } else {
            0b00
        };

        let rd = self.get_reg_field_from_operand(mi.operands.first());
        let rn = self.get_reg_field_from_operand(mi.operands.get(1));
        let rm = self.get_reg_field_from_operand(mi.operands.get(2));

        let mut word: u32 = 0;
        word |= sf << 31;
        word |= opc << 29;
        word |= 0b01010 << 24;
        word |= (rm as u32) << 16;
        word |= (rn as u32) << 5;
        word |= rd as u32;
        word
    }

    // ========================================================================
    // ARM32 Instruction Encoding
    // ========================================================================

    /// Encode a single ARM32 instruction.
    fn encode_arm32_instruction(&self, mi: &MachineInstr) -> u32 {
        let op = mi.opcode;
        if op == OP_ARM_ADD
            || op == OP_ARM_SUB
            || op == OP_ARM_MOV
            || op == OP_ARM_AND
            || op == OP_ARM_ORR
            || op == OP_ARM_EOR
            || op == OP_ARM_CMP
        {
            self.encode_arm32_data_proc(mi)
        } else if op == OP_ARM_LDR || op == OP_ARM_STR {
            self.encode_arm32_ldr_str(mi)
        } else if op == OP_ARM_B || op == OP_ARM_BL {
            self.encode_arm32_b_bl(mi)
        } else if op == OP_ARM_PUSH || op == OP_ARM_POP {
            self.encode_arm32_push_pop(mi)
        } else {
            0xE1A00000 // MOV R0, R0 (NOP equivalent in ARM32)
        }
    }

    /// ARM32 Data Processing (ADD/SUB/MOV/AND/ORR/EOR/CMP).
    /// ARM mode encoding:
    /// | cond(31-28) | 00(27-26) | I(25) | opcode(24-21) | S(20) | Rn(19-16) | Rd(15-12) | operand2(11-0) |
    ///
    /// opcode: 0100=ADD, 0010=SUB, 1101=MOV, 0000=AND, 1100=ORR, 0001=EOR, 1010=CMP
    /// For register form: operand2 = imm5(11-7) | type(6-5) | 0(4) | Rm(3-0)
    /// For immediate form: I=1, operand2 = imm8(7-0) rotated by rotate_imm(11-8)*2
    fn encode_arm32_data_proc(&self, mi: &MachineInstr) -> u32 {
        let op = mi.opcode;
        let cond: u32 = 0xE; // AL (always execute)
        let set_flags = op == OP_ARM_CMP;

        let opcode: u32 = match op {
            OP_ARM_ADD => 0b0100,
            OP_ARM_SUB => 0b0010,
            OP_ARM_MOV => 0b1101,
            OP_ARM_AND => 0b0000,
            OP_ARM_ORR => 0b1100,
            OP_ARM_EOR => 0b0001,
            OP_ARM_CMP => 0b1010,
            _ => 0b0000,
        };

        let rd = self.get_arm32_reg_field_from_operand(mi.operands.first());
        let rn = if op == OP_ARM_MOV {
            0 // MOV uses 0000 for Rn
        } else {
            self.get_arm32_reg_field_from_operand(mi.operands.get(1))
        };

        // Check if operand2 is immediate or register
        let (i_flag, operand2) =
            if mi.operands.len() >= 3 || (op == OP_ARM_MOV && mi.operands.len() >= 2) {
                let src_idx = if op == OP_ARM_MOV { 1 } else { 2 };
                match mi.operands.get(src_idx) {
                    Some(MachineOperand::Imm(imm)) => {
                        let imm8 = *imm as u32 & 0xFF;
                        // ARM immediate: 8-bit value rotated right by even number
                        // For simplicity, use zero rotation
                        (1u32, ((0u32) << 8) | imm8)
                    }
                    _ => {
                        let rm = self.get_arm32_reg_field_from_operand(mi.operands.get(src_idx));
                        (0u32, rm)
                    }
                }
            } else {
                (0u32, 0u32)
            };

        let mut word: u32 = 0;
        word |= cond << 28;
        word |= i_flag << 25;
        word |= opcode << 21;
        word |= (set_flags as u32) << 20;
        word |= rn << 16;
        word |= rd << 12;
        word |= operand2 & 0xFFF;
        word
    }

    /// ARM32 LDR/STR with immediate offset.
    /// | cond(31-28) | 01(27-26) | I/P(25-24) | U(23) | B(22) | W(21) | L(20) | Rn(19-16) | Rd(15-12) | imm12(11-0) |
    ///
    /// LDR: L=1, STR: L=0
    /// I=0,P=1: offset addressing (add offset to base)
    /// U=1: up (add), U=0: down (subtract)
    fn encode_arm32_ldr_str(&self, mi: &MachineInstr) -> u32 {
        let is_load = mi.opcode == OP_ARM_LDR;
        let cond: u32 = 0xE;

        let rd = self.get_arm32_reg_field_from_operand(mi.operands.first());
        let rn = self.get_arm32_reg_field_from_operand(mi.operands.get(1));
        let imm = self.get_imm_from_operand(mi.operands.get(2)) as u32;

        let mut word: u32 = 0;
        word |= cond << 28;
        word |= 0b01 << 26; // Data processing immediate (for LDR/STR)
        word |= 0b01 << 24; // P=1, I=0 (immediate offset)
        word |= 1 << 23; // U=1 (add offset)
        word |= (is_load as u32) << 20;
        word |= rn << 16;
        word |= rd << 12;
        word |= imm & 0xFFF;
        word
    }

    /// ARM32 B/BL.
    /// | cond(31-28) | 101(27-25) | L(24) | imm24(23-0) |
    fn encode_arm32_b_bl(&self, mi: &MachineInstr) -> u32 {
        let is_call = mi.opcode == OP_ARM_BL;
        let cond: u32 = 0xE;
        let offset = self.get_branch_offset(mi) & 0xFFFFFF;

        let mut word: u32 = 0;
        word |= cond << 28;
        word |= 0b101 << 25;
        word |= (is_call as u32) << 24;
        word |= offset;
        word
    }

    /// ARM32 PUSH/POP.
    /// PUSH: | cond(31-28) | 10010010(27-20) | 1101(19-16) | register_list(15-0) | (STMDB SP!, {reglist})
    /// POP:  | cond(31-28) | 10001001(27-20) | 1101(19-16) | register_list(15-0) | (LDMIA SP!, {reglist})
    fn encode_arm32_push_pop(&self, mi: &MachineInstr) -> u32 {
        let is_pop = mi.opcode == OP_ARM_POP;
        let cond: u32 = 0xE;

        // Collect register list bits from operands
        let mut reg_list: u32 = 0;
        for op in &mi.operands {
            let reg = self.get_arm32_reg_field_from_operand(Some(op));
            if reg < 16 {
                reg_list |= 1u32 << reg;
            }
        }

        let mut word: u32 = 0;
        word |= cond << 28;
        if is_pop {
            word |= 0b10001001 << 20; // LDMIA
        } else {
            word |= 0b10010010 << 20; // STMDB
        }
        word |= 0b1101 << 16; // Rn = SP
        word |= reg_list & 0xFFFF;
        word
    }

    // ========================================================================
    // Register Field Helpers
    // ========================================================================

    /// Get the 5-bit register field for AArch64 encoding (0-31).
    /// Maps our register ID constants (2000-2033 for X, 2040-2072 for W, etc.)
    /// to the 5-bit field used in instruction encoding.
    pub fn get_reg_field(reg_id: u16) -> u8 {
        // Use the arm_register_info constants
        use super::arm_register_info::*;
        match reg_id {
            // X0-X30 → 0-30
            X0..=X30 => (reg_id - X0) as u8,
            // SP → 31
            SP => 31,
            // XZR → 31 (zero register)
            XZR => 31,
            // W0-W30 → 0-30
            W0..=W30 => (reg_id - W0) as u8,
            // WSP → 31
            WSP => 31,
            // WZR → 31
            WZR => 31,
            // R0-R15 → 0-15 (for ARM32)
            R0..=R15 => (reg_id - R0) as u8,
            // Default: treat as register number directly if small
            _ if reg_id < 32 => reg_id as u8,
            _ => 0,
        }
    }

    /// Get register field from a MachineOperand.
    fn get_reg_field_from_operand(&self, op: Option<&MachineOperand>) -> u8 {
        match op {
            Some(MachineOperand::Reg(vr)) => {
                // Virtual register: use as-is modulo 32
                (*vr as u16 % 32) as u8
            }
            Some(MachineOperand::PhysReg(pr)) => Self::get_reg_field(*pr as u16),
            _ => 31, // Default to XZR
        }
    }

    /// Get ARM32 register field (0-15 for R0-R15).
    fn get_arm32_reg_field_from_operand(&self, op: Option<&MachineOperand>) -> u32 {
        match op {
            Some(MachineOperand::Reg(vr)) => (*vr % 16) as u32,
            Some(MachineOperand::PhysReg(pr)) => {
                // ARM32 registers: R0=2100 .. R15=2115
                if *pr >= 2100 && *pr <= 2115 {
                    (*pr - 2100) as u32
                } else {
                    *pr % 16
                }
            }
            _ => 0,
        }
    }

    /// Extract immediate value from an operand.
    fn get_imm_from_operand(&self, op: Option<&MachineOperand>) -> i64 {
        match op {
            Some(MachineOperand::Imm(v)) => *v,
            _ => 0,
        }
    }

    /// Extract halfword specifier from operand (for MOVZ/MOVK/MOVN).
    /// The hw field is 0-3 indicating which 16-bit halfword to target.
    fn get_hw_from_operand(&self, op: Option<&MachineOperand>) -> u32 {
        match op {
            Some(MachineOperand::Imm(v)) => {
                // Imm value encodes shift position: 0, 16, 32, 48
                match *v {
                    0 => 0,
                    16 => 1,
                    32 => 2,
                    48 => 3,
                    _ => 0,
                }
            }
            _ => 0,
        }
    }

    /// Extract condition code from operand.
    fn get_cond_from_operand(&self, op: Option<&MachineOperand>) -> u32 {
        match op {
            Some(MachineOperand::Imm(v)) => *v as u32 & 0xF,
            Some(MachineOperand::Label(s)) => {
                ArmCond::from_mnemonic(s).map(|c| c as u32).unwrap_or(0xE)
            }
            _ => 0xE, // AL
        }
    }

    /// Get the branch offset from the instruction.
    /// Looks for a label operand and resolves it to a PC-relative offset.
    fn get_branch_offset(&self, mi: &MachineInstr) -> u32 {
        // Find label operand and compute offset
        for op in &mi.operands {
            if let MachineOperand::Label(label) = op {
                if let Some(&target_off) = self.label_offsets.get(label) {
                    // PC-relative offset in words (instruction count / 4)
                    // For now, return the target offset directly (caller adjusts)
                    return (target_off as u32) / 4;
                }
            }
            if let MachineOperand::Imm(offset) = op {
                return (*offset as u32) / 4;
            }
        }
        0
    }

    /// Get the 19-bit branch offset for CBZ/CBNZ/B.cond.
    fn get_branch_offset_19bit(&self, mi: &MachineInstr) -> u32 {
        self.get_branch_offset(mi) & 0x7FFFF
    }

    // ========================================================================
    // SVE (Scalable Vector Extension) Encoding
    // ========================================================================

    /// SVE predicate AND: AND Pd, Pg/z, Pn, Pm
    /// Format: 00100101 | xx | 0 | Pm(3:0) | 0 | Pg(3:0) | Pn(3:0) | 0 | Pd(3:0)
    pub fn encode_sve_pred_and(&self, pd: u32, pg: u32, pn: u32, pm: u32) -> u32 {
        let mut w: u32 = 0;
        w |= 0b00100101 << 24;
        w |= 0b00 << 22;
        w |= (pm & 0xF) << 16;
        w |= (pg & 0xF) << 10;
        w |= (pn & 0xF) << 5;
        w |= (pd & 0xF);
        w
    }

    /// SVE predicate ORR: ORR Pd, Pg/z, Pn, Pm
    pub fn encode_sve_pred_orr(&self, pd: u32, pg: u32, pn: u32, pm: u32) -> u32 {
        let mut w: u32 = 0;
        w |= 0b00100101 << 24;
        w |= 0b01 << 22;
        w |= (pm & 0xF) << 16;
        w |= (pg & 0xF) << 10;
        w |= (pn & 0xF) << 5;
        w |= (pd & 0xF);
        w
    }

    /// SVE predicate EOR: EOR Pd, Pg/z, Pn, Pm
    pub fn encode_sve_pred_eor(&self, pd: u32, pg: u32, pn: u32, pm: u32) -> u32 {
        let mut w: u32 = 0;
        w |= 0b00100101 << 24;
        w |= 0b10 << 22;
        w |= (pm & 0xF) << 16;
        w |= (pg & 0xF) << 10;
        w |= (pn & 0xF) << 5;
        w |= (pd & 0xF);
        w
    }

    /// SVE predicate NAND: NAND Pd, Pg/z, Pn, Pm
    pub fn encode_sve_pred_nand(&self, pd: u32, pg: u32, pn: u32, pm: u32) -> u32 {
        self.encode_sve_pred_and(pd, pg, pn, pm) ^ (1 << 23)
    }

    /// SVE predicate NOR: NOR Pd, Pg/z, Pn, Pm
    pub fn encode_sve_pred_nor(&self, pd: u32, pg: u32, pn: u32, pm: u32) -> u32 {
        self.encode_sve_pred_orr(pd, pg, pn, pm) ^ (1 << 23)
    }

    /// SVE predicate ORN: ORN Pd, Pg/z, Pn, Pm
    pub fn encode_sve_pred_orn(&self, pd: u32, pg: u32, pn: u32, pm: u32) -> u32 {
        self.encode_sve_pred_eor(pd, pg, pn, pm) ^ (1 << 23)
    }

    /// SVE PTRUE: PTRUE Pd, pattern
    /// Format: 00100101 | size(01) | 1 | 00000 | 00000 | 10 | pattern(4:0) | 0 | Pd(3:0)
    pub fn encode_sve_ptrue(&self, pd: u32, pattern: u32) -> u32 {
        let mut w: u32 = 0;
        w |= 0b00100101 << 24;
        w |= 0b01 << 22; // size
        w |= 0b1 << 21;
        w |= 0b10 << 10;
        w |= (pattern & 0x1F) << 5;
        w |= (pd & 0xF);
        w
    }

    /// SVE PFIRST: PFIRST Pd, Pg, Pn
    pub fn encode_sve_pfirst(&self, pd: u32, pg: u32, pn: u32) -> u32 {
        let mut w: u32 = 0;
        w |= 0b00100101 << 24;
        w |= 0b00 << 22;
        w |= 0b1 << 21;
        w |= 0b00000 << 16;
        w |= (pg & 0xF) << 10;
        w |= (pn & 0xF) << 5;
        w |= 0b1 << 4;
        w |= (pd & 0xF);
        w
    }

    /// SVE PNEXT: PNEXT Pd, Pg, Pn
    pub fn encode_sve_pnext(&self, pd: u32, pg: u32, pn: u32) -> u32 {
        self.encode_sve_pfirst(pd, pg, pn) | (1 << 13)
    }

    /// SVE PLAST: Get last active element predicate
    pub fn encode_sve_plast(&self, pd: u32, pg: u32, pn: u32) -> u32 {
        self.encode_sve_pfirst(pd, pg, pn) | (1 << 14)
    }

    /// SVE vector ADD: ADD Zd, Pg/m, Zd, Zn
    /// Format: 00000100 | size | 1 | m(4) | msz | 0 | Pg(3:0) | Zn(4:0) | Zda(4:0)
    pub fn encode_sve_add_z(&self, zd: u32, pg: u32, zn: u32, size: u32) -> u32 {
        let mut w: u32 = 0;
        w |= 0b00000100 << 24;
        w |= (size & 0x3) << 22;
        w |= 0b1 << 21;
        w |= 0b0 << 20; // m=0 (merge)
        w |= (zd & 0x1) << 16; // msz
        w |= (zd >> 1) << 14;
        w |= (pg & 0xF) << 10;
        w |= (zn & 0x1F) << 5;
        w |= (zd & 0x1F);
        w
    }

    /// SVE vector SUB: SUB Zd, Pg/m, Zd, Zn
    pub fn encode_sve_sub_z(&self, zd: u32, pg: u32, zn: u32, size: u32) -> u32 {
        let w = self.encode_sve_add_z(zd, pg, zn, size);
        w | (1 << 13)
    }

    /// SVE vector MUL: MUL Zd, Pg/m, Zn, Zm
    pub fn encode_sve_mul_z(&self, zd: u32, pg: u32, zn: u32, zm: u32, size: u32) -> u32 {
        let mut w: u32 = 0;
        w |= 0b00000100 << 24;
        w |= (size & 0x3) << 22;
        w |= 0b1 << 21;
        w |= (zm & 0x1F) << 16;
        w |= (pg & 0xF) << 10;
        w |= (zn & 0x1F) << 5;
        w |= (zd & 0x1F);
        w |= 0b1 << 12; // mul op
        w
    }

    /// SVE vector MLA: MLA Zd, Pg/m, Zn, Zm
    pub fn encode_sve_mla_z(&self, zd: u32, pg: u32, zn: u32, zm: u32, size: u32) -> u32 {
        self.encode_sve_mul_z(zd, pg, zn, zm, size) | (1 << 11)
    }

    /// SVE vector MLS: MLS Zd, Pg/m, Zn, Zm
    pub fn encode_sve_mls_z(&self, zd: u32, pg: u32, zn: u32, zm: u32, size: u32) -> u32 {
        self.encode_sve_mla_z(zd, pg, zn, zm, size) | (1 << 13)
    }

    /// SVE FADD: FADD Zd, Pg/m, Zd, Zn
    pub fn encode_sve_fadd_z(&self, zd: u32, pg: u32, zn: u32, size: u32) -> u32 {
        let mut w: u32 = 0;
        w |= 0b01100101 << 24;
        w |= (size & 0x3) << 22;
        w |= 0b0 << 21;
        w |= (pg & 0xF) << 10;
        w |= (zn & 0x1F) << 5;
        w |= (zd & 0x1F);
        w
    }

    /// SVE FSUB: FSUB Zd, Pg/m, Zd, Zn
    pub fn encode_sve_fsub_z(&self, zd: u32, pg: u32, zn: u32, size: u32) -> u32 {
        self.encode_sve_fadd_z(zd, pg, zn, size) | (1 << 13)
    }

    /// SVE FMUL: FMUL Zd, Pg/m, Zn, Zm
    pub fn encode_sve_fmul_z(&self, zd: u32, pg: u32, zn: u32, zm: u32, size: u32) -> u32 {
        let mut w: u32 = 0;
        w |= 0b01100101 << 24;
        w |= (size & 0x3) << 22;
        w |= 0b0 << 21;
        w |= (zm & 0x1F) << 16;
        w |= (pg & 0xF) << 10;
        w |= (zn & 0x1F) << 5;
        w |= (zd & 0x1F);
        w |= 0b1 << 12;
        w
    }

    /// SVE FDIV: FDIV Zd, Pg/m, Zn, Zm
    pub fn encode_sve_fdiv_z(&self, zd: u32, pg: u32, zn: u32, zm: u32, size: u32) -> u32 {
        self.encode_sve_fmul_z(zd, pg, zn, zm, size) | (1 << 11)
    }

    /// SVE contiguous load: LD1B {Zd}, Pg/z, [Xn, Xm]
    /// Format: 10100101 | 00 | msz | 0 | 00 | Pg | Zn | Zt
    pub fn encode_sve_ld1b_z(&self, zt: u32, pg: u32, zn: u32, xm: u32) -> u32 {
        let mut w: u32 = 0;
        w |= 0b10100101 << 24;
        w |= 0b0 << 23;
        w |= 0b0 << 22;
        w |= 0b0 << 21;
        w |= 0b0 << 20;
        w |= 0b10 << 18;
        w |= (xm & 0x1F) << 16;
        w |= (pg & 0xF) << 10;
        w |= (zn & 0x1F) << 5;
        w |= (zt & 0x1F);
        w
    }

    /// SVE contiguous store: ST1B {Zd}, Pg, [Xn, Xm]
    pub fn encode_sve_st1b_z(&self, zt: u32, pg: u32, zn: u32, xm: u32) -> u32 {
        let mut w: u32 = 0;
        w |= 0b11100101 << 24;
        w |= 0b0 << 23;
        w |= 0b0 << 22;
        w |= 0b0 << 21;
        w |= 0b0 << 20;
        w |= 0b10 << 18;
        w |= (xm & 0x1F) << 16;
        w |= (pg & 0xF) << 10;
        w |= (zn & 0x1F) << 5;
        w |= (zt & 0x1F);
        w
    }

    /// SVE gather load: LD1W {Zd}, Pg/z, [Zn, Zm, sxtw #2]
    /// 32-bit offset gather
    pub fn encode_sve_ld1w_gather(&self, zt: u32, pg: u32, zn: u32, zm: u32) -> u32 {
        let mut w: u32 = 0;
        w |= 0b10000101 << 24;
        w |= 0b0 << 23; // 32-bit
        w |= 0b10 << 21;
        w |= (zm & 0x1F) << 16;
        w |= 0b0 << 15;
        w |= 0b10 << 13; // SXTW #2
        w |= (pg & 0xF) << 10;
        w |= (zn & 0x1F) << 5;
        w |= (zt & 0x1F);
        w
    }

    /// SVE gather load: LD1D {Zd}, Pg/z, [Zn, Zm, lsl #3]
    /// 64-bit offset gather
    pub fn encode_sve_ld1d_gather(&self, zt: u32, pg: u32, zn: u32, zm: u32) -> u32 {
        let mut w: u32 = 0;
        w |= 0b11000101 << 24;
        w |= 0b0 << 23; // 64-bit
        w |= 0b10 << 21;
        w |= (zm & 0x1F) << 16;
        w |= 0b0 << 15;
        w |= 0b11 << 13; // LSL #3
        w |= (pg & 0xF) << 10;
        w |= (zn & 0x1F) << 5;
        w |= (zt & 0x1F);
        w
    }

    /// SVE scatter store: ST1W {Zd}, Pg, [Zn, Zm, sxtw #2]
    pub fn encode_sve_st1w_scatter(&self, zt: u32, pg: u32, zn: u32, zm: u32) -> u32 {
        let w = self.encode_sve_ld1w_gather(zt, pg, zn, zm);
        w | (1 << 26) // store vs load
    }

    /// SVE scatter store: ST1D {Zd}, Pg, [Zn, Zm, lsl #3]
    pub fn encode_sve_st1d_scatter(&self, zt: u32, pg: u32, zn: u32, zm: u32) -> u32 {
        let w = self.encode_sve_ld1d_gather(zt, pg, zn, zm);
        w | (1 << 26)
    }

    // ========================================================================
    // SME (Scalable Matrix Extension) Encoding
    // ========================================================================

    /// SME ADDHA: outer product accumulate (add to ZA)
    /// Format: 11000001 | 00 | Zm | 0 | Pn | Pm | 0 | Zdn
    pub fn encode_sme_addha(&self, zdn: u32, pm: u32, pn: u32, zm: u32) -> u32 {
        let mut w: u32 = 0;
        w |= 0b11000001 << 24;
        w |= 0b00 << 22;
        w |= (zm & 0x1F) << 16;
        w |= (pn & 0xF) << 10;
        w |= (pm & 0xF) << 5;
        w |= (zdn & 0x1F);
        w
    }

    /// SME ADDVA: outer product accumulate variant
    pub fn encode_sme_addva(&self, zdn: u32, pm: u32, pn: u32, zm: u32) -> u32 {
        self.encode_sme_addha(zdn, pm, pn, zm) | (1 << 12)
    }

    /// SME LDR ZA: load ZA array from memory
    /// Format: 11100001 | 11 | 000000 | 0 | 0 | Rn | 0 | 00 | 0 | off
    pub fn encode_sme_ldr_za(&self, rn: u32, offset: u32) -> u32 {
        let mut w: u32 = 0;
        w |= 0b11100001 << 24;
        w |= 0b11 << 22;
        w |= (rn & 0x1F) << 5;
        w |= (offset & 0x3FF) >> 6;
        w
    }

    /// SME STR ZA: store ZA array to memory
    pub fn encode_sme_str_za(&self, rn: u32, offset: u32) -> u32 {
        let mut w = self.encode_sme_ldr_za(rn, offset);
        w ^= 1 << 22; // toggle load/store bit
        w
    }

    /// SME MOVA ZA to Z: extract tile from ZA to vector
    /// Format: 11000001 | 00 | 0000000 | slice | 0 | Zdn
    pub fn encode_sme_mova_za_to_z(&self, zd: u32, slice: u32) -> u32 {
        let mut w: u32 = 0;
        w |= 0b11000001 << 24;
        w |= 0b00 << 22;
        w |= (slice & 0x7) << 5;
        w |= (zd & 0x1F);
        w
    }

    /// SME MOVA Z to ZA: write vector to ZA tile
    pub fn encode_sme_mova_z_to_za(&self, zn: u32, slice: u32) -> u32 {
        let mut w: u32 = 0;
        w |= 0b11000001 << 24;
        w |= 0b01 << 22;
        w |= (slice & 0x7) << 5;
        w |= (zn & 0x1F);
        w
    }

    /// SME ZERO ZA: zero the entire ZA array
    pub fn encode_sme_zero_za(&self) -> u32 {
        let mut w: u32 = 0;
        w |= 0b11000001 << 24;
        w |= 0b11 << 22;
        w |= 0b1 << 21;
        w
    }

    /// SME2 FMLA: multi-vector FMLA
    /// Format: 11000001 | 00 | Zn | 0 | Zm | ...
    pub fn encode_sme2_fmla_multi(&self, zdn: u32, zm: u32, zn: u32, pn: u32) -> u32 {
        let mut w: u32 = 0;
        w |= 0b11000001 << 24;
        w |= 0b10 << 22;
        w |= (zm & 0x1F) << 16;
        w |= (zn & 0x1F) << 5;
        w |= (pn & 0xF) << 10;
        w |= (zdn & 0x1F);
        w
    }

    /// SME2 FMLS: multi-vector FMLS
    pub fn encode_sme2_fmls_multi(&self, zdn: u32, zm: u32, zn: u32, pn: u32) -> u32 {
        self.encode_sme2_fmla_multi(zdn, zm, zn, pn) | (1 << 12)
    }

    /// SME2 multi-register load: LD1B {Zdn.B, ...}, Pn/Z, [Xm]
    pub fn encode_sme2_ld1b_multi(&self, zt: u32, pn: u32, xm: u32) -> u32 {
        let mut w: u32 = 0;
        w |= 0b11000001 << 24;
        w |= 0b00 << 22;
        w |= 0b11 << 20; // multi-reg
        w |= (xm & 0x1F) << 5;
        w |= (pn & 0xF) << 10;
        w |= (zt & 0x1F);
        w
    }

    /// SME2 multi-register store: ST1B {Zdn.B, ...}, Pn, [Xm]
    pub fn encode_sme2_st1b_multi(&self, zt: u32, pn: u32, xm: u32) -> u32 {
        let mut w = self.encode_sme2_ld1b_multi(zt, pn, xm);
        w ^= 1 << 22;
        w
    }

    // ========================================================================
    // MTE (Memory Tagging Extension) Encoding
    // ========================================================================

    /// STG: Store Allocation Tag — STG Xt, [Xn, #simm]
    /// Format: 11011001 | 0 | 1 | simm(9) | 0 | 1 | Xt | simm(9-7) | 1 | Xn
    pub fn encode_mte_stg(&self, xt: u32, xn: u32, simm: i32) -> u32 {
        let s = simm as u32 & 0x3FF;
        let s9_4 = (s >> 4) & 0x3F;
        let s3_0 = s & 0xF;
        let mut w: u32 = 0;
        w |= 0b11011001 << 24;
        w |= 0b0 << 23;
        w |= 0b1 << 22;
        w |= s9_4 << 16;
        w |= 0b0 << 15;
        w |= (s3_0 >> 3) << 14;
        w |= (xt & 0x1F) << 10;
        w |= (s3_0 & 0x7) << 8;
        w |= 0b1 << 7;
        w |= (xn & 0x1F);
        w
    }

    /// STZG: Store Zero Allocation Tag
    pub fn encode_mte_stzg(&self, xt: u32, xn: u32, simm: i32) -> u32 {
        self.encode_mte_stg(xt, xn, simm) | (1 << 23)
    }

    /// ST2G: Store Two Allocation Tags
    pub fn encode_mte_st2g(&self, xt: u32, xn: u32, simm: i32) -> u32 {
        self.encode_mte_stg(xt, xn, simm) | (1 << 21)
    }

    /// STZ2G: Store Two Zero Allocation Tags
    pub fn encode_mte_stz2g(&self, xt: u32, xn: u32, simm: i32) -> u32 {
        self.encode_mte_st2g(xt, xn, simm) | (1 << 23)
    }

    /// LDG: Load Allocation Tag — LDG Xt, [Xn, #simm]
    /// Format: 11011001 | 1 | 1 | simm(9) | 0 | 1 | Xt | simm(9-7) | 1 | Xn
    pub fn encode_mte_ldg(&self, xt: u32, xn: u32, simm: i32) -> u32 {
        let mut w = self.encode_mte_stg(xt, xn, simm);
        w |= 1 << 23; // load bit
        w
    }

    /// STGM: Store Allocation Tag (multiple)
    pub fn encode_mte_stgm(&self, xt: u32, xn: u32) -> u32 {
        let mut w: u32 = 0;
        w |= 0b11011001 << 24;
        w |= 0b0 << 23;
        w |= 0b1 << 22;
        w |= (xt & 0x1F) << 10;
        w |= 0b1 << 7;
        w |= (xn & 0x1F);
        w
    }

    /// STZGM: Store Zero Allocation Tag (multiple)
    pub fn encode_mte_stzgm(&self, xt: u32, xn: u32) -> u32 {
        self.encode_mte_stgm(xt, xn) | (1 << 23)
    }

    /// LDGM: Load Allocation Tag (multiple)
    pub fn encode_mte_ldgm(&self, xt: u32, xn: u32) -> u32 {
        let mut w = self.encode_mte_stgm(xt, xn);
        w |= 1 << 23;
        w
    }

    /// IRG: Insert Random Tag — IRG Xd, Xn, Xm
    /// Format: 11011001 | 0 | 0 | 1 | 0 | 0 | 1 | 0 | Xm | 0 | Xn | Xd
    pub fn encode_mte_irg(&self, xd: u32, xn: u32, xm: u32) -> u32 {
        let mut w: u32 = 0;
        w |= 0b11011001 << 24;
        w |= (xm & 0x1F) << 16;
        w |= (xn & 0x1F) << 5;
        w |= (xd & 0x1F);
        w
    }

    /// ADDG: Add with Tag — ADDG Xd, Xn, #uimm6, #uimm4
    /// Format: 10010001 | 1 | uimm6 | uimm4 | 0 | Xn | Xd
    pub fn encode_mte_addg(&self, xd: u32, xn: u32, uimm6: u32, uimm4: u32) -> u32 {
        let mut w: u32 = 0;
        w |= 0b10010001 << 24;
        w |= 0b1 << 23;
        w |= (uimm6 & 0x3F) << 16;
        w |= (uimm4 & 0xF) << 10;
        w |= (xn & 0x1F) << 5;
        w |= (xd & 0x1F);
        w
    }

    /// SUBG: Subtract with Tag — SUBG Xd, Xn, #uimm6, #uimm4
    pub fn encode_mte_subg(&self, xd: u32, xn: u32, uimm6: u32, uimm4: u32) -> u32 {
        self.encode_mte_addg(xd, xn, uimm6, uimm4) | (1 << 22)
    }

    /// GMI: Granule Mask Insert — GMI Xd, Xn, Xm
    pub fn encode_mte_gmi(&self, xd: u32, xn: u32, xm: u32) -> u32 {
        let mut w: u32 = 0;
        w |= 0b10011001 << 24;
        w |= (xm & 0x1F) << 16;
        w |= (xn & 0x1F) << 5;
        w |= (xd & 0x1F);
        w
    }

    /// SUBP: Subtract Pointer (with tag) — SUBP Xd, Xn, Xm
    pub fn encode_mte_subp(&self, xd: u32, xn: u32, xm: u32) -> u32 {
        let mut w = self.encode_mte_gmi(xd, xn, xm);
        w ^= 1 << 22;
        w
    }

    /// SUBPS: Subtract Pointer (with tag) and set flags — SUBPS Xd, Xn, Xm
    pub fn encode_mte_subps(&self, xd: u32, xn: u32, xm: u32) -> u32 {
        self.encode_mte_subp(xd, xn, xm) | (1 << 22)
    }

    /// SME2 LD1 multi-register: LD1 {Z0.Z-Z3.Z}, Pn/Z, [Xn, Xm]
    pub fn encode_sme2_ld1_multi(&self, zt: u32, pg: u32, xn: u32, xm: u32, num_regs: u32) -> u32 {
        let mut w: u32 = 0;
        w |= 0b10100101 << 24;
        w |= (num_regs & 0x3) << 22;
        w |= 0b1 << 21;
        w |= (xm & 0x1F) << 16;
        w |= (pg & 0xF) << 10;
        w |= (xn & 0x1F) << 5;
        w |= (zt & 0x1F);
        w
    }

    /// SME2 ST1 multi-register: ST1 {Z0.Z-Z3.Z}, Pn, [Xn, Xm]
    pub fn encode_sme2_st1_multi(&self, zt: u32, pg: u32, xn: u32, xm: u32, num_regs: u32) -> u32 {
        self.encode_sme2_ld1_multi(zt, pg, xn, xm, num_regs) | (1 << 24)
    }

    // ========================================================================
    // SME2.1 — BF16 outer product
    // ========================================================================

    /// SME2.1 BFMOPA: BF16 outer product and accumulate
    pub fn encode_sme2_bfmopa(&self, zd: u32, zn: u32, zm: u32, za: u32) -> u32 {
        let mut w: u32 = 0;
        w |= 0b10100001 << 24;
        w |= 0b00 << 23;
        w |= (zm & 0x1F) << 16;
        w |= (za & 0x7) << 13;
        w |= 0b1 << 12;
        w |= (zn & 0x1F) << 5;
        w |= (zd & 0x1F);
        w
    }

    /// SME2.1 BFMOPS: BF16 outer product and subtract
    pub fn encode_sme2_bfmops(&self, zd: u32, zn: u32, zm: u32, za: u32) -> u32 {
        self.encode_sme2_bfmopa(zd, zn, zm, za) ^ (1 << 23)
    }

    // ========================================================================
    // FP8 conversions — F1CVT/F2CVT/BF1CVT
    // ========================================================================

    /// FP8 F1CVT: FP8 → FP32 conversion (format 1)
    pub fn encode_f8_f1cvt(&self, vd: u32, vn: u32) -> u32 {
        let mut w: u32 = 0;
        w |= 0b01001110 << 24;
        w |= 0b0 << 23;
        w |= 0b0011111 << 16;
        w |= 0b100001 << 10;
        w |= (vn & 0x1F) << 5;
        w |= (vd & 0x1F);
        w
    }

    /// FP8 F2CVT: FP8 → FP16 conversion (format 2)
    pub fn encode_f8_f2cvt(&self, vd: u32, vn: u32) -> u32 {
        let mut w = self.encode_f8_f1cvt(vd, vn);
        w ^= 1 << 23;
        w
    }

    /// FP8 BF1CVT: BF8 → FP32 conversion
    pub fn encode_f8_bf1cvt(&self, vd: u32, vn: u32) -> u32 {
        let mut w: u32 = 0;
        w |= 0b01001110 << 24;
        w |= 0b1 << 23;
        w |= 0b0011111 << 16;
        w |= 0b100001 << 10;
        w |= (vn & 0x1F) << 5;
        w |= (vd & 0x1F);
        w
    }

    // ========================================================================
    // LUT — LUTI2/LUTI4 lookup table instructions
    // ========================================================================

    /// LUTI2: Lookup table with 2-bit indices
    pub fn encode_lut_luti2(&self, zd: u32, zn: u32, zm: u32, table_idx: u32) -> u32 {
        let mut w: u32 = 0;
        w |= 0b01000101 << 24;
        w |= table_idx << 20;
        w |= (zm & 0x1F) << 16;
        w |= 0b001 << 13;
        w |= (zn & 0x1F) << 5;
        w |= (zd & 0x1F);
        w
    }

    /// LUTI4: Lookup table with 4-bit indices
    pub fn encode_lut_luti4(&self, zd: u32, zn: u32, zm: u32, table_idx: u32) -> u32 {
        self.encode_lut_luti2(zd, zn, zm, table_idx) ^ (1 << 23)
    }

    // ========================================================================
    // FAMINMAX — FAMIN/FAMAX (absolute min/max)
    // ========================================================================

    /// FAMIN: Floating-point absolute minimum
    pub fn encode_faminmax_famin(&self, vd: u32, vn: u32, vm: u32) -> u32 {
        let mut w: u32 = 0;
        w |= 0b0 << 31;
        w |= 0b0 << 30;
        w |= 0b01110 << 25;
        w |= 0b1 << 24;
        w |= (vm & 0x1F) << 16;
        w |= 0b0100 << 12;
        w |= 0b1 << 10;
        w |= (vn & 0x1F) << 5;
        w |= (vd & 0x1F);
        w
    }

    /// FAMAX: Floating-point absolute maximum
    pub fn encode_faminmax_famax(&self, vd: u32, vn: u32, vm: u32) -> u32 {
        self.encode_faminmax_famin(vd, vn, vm) ^ (1 << 30)
    }

    // ========================================================================
    // SVE2.1 — complex integer multiply-add with rotate
    // ========================================================================

    /// SVE2.1 CMLA (complex multiply-add with rotate): CMLA Zda, Pg/m, Zn, Zm, #rotation
    pub fn encode_sve2_cmla(
        &self,
        zda: u32,
        pg: u32,
        zn: u32,
        zm: u32,
        rot: u32,
        size: u32,
    ) -> u32 {
        let mut w: u32 = 0;
        w |= 0b01000100 << 24;
        w |= (size & 0x3) << 22;
        w |= 0b1 << 21;
        w |= (zm & 0x1F) << 16;
        w |= (rot & 0x3) << 14;
        w |= (pg & 0xF) << 10;
        w |= (zn & 0x1F) << 5;
        w |= (zda & 0x1F);
        w
    }

    /// SVE2.1 SDOT (signed dot product accumulate): SDOT Zda, Pg/m, Zn, Zm
    pub fn encode_sve2_sdot(&self, zda: u32, pg: u32, zn: u32, zm: u32) -> u32 {
        let mut w: u32 = 0;
        w |= 0b01000100 << 24;
        w |= 0b00 << 22;
        w |= 0b1 << 21;
        w |= (zm & 0x1F) << 16;
        w |= 0b0001 << 10;
        w |= (pg & 0xF) << 10;
        w |= (zn & 0x1F) << 5;
        w |= (zda & 0x1F);
        w
    }

    // ========================================================================
    // BRBE — Branch Record Buffer
    // ========================================================================

    /// BRB IALL: Invalidate all branch records
    pub fn encode_brb_iall(&self) -> u32 {
        0b1101010100 << 22
    }

    /// BRB INJ: Inject branch record
    pub fn encode_brb_inj(&self, xt: u32) -> u32 {
        let mut w: u32 = 0;
        w |= 0b1101010100 << 22;
        w |= 0b000100 << 16;
        w |= 0b0000 << 12;
        w |= 0b001 << 9;
        w |= 0b11111 << 5;
        w |= (xt & 0x1F);
        w
    }

    // ========================================================================
    // LRCPC3 — LDIAPP/STILP ordered load/store pairs
    // ========================================================================

    /// LDIAPP: Load acquire pair (ordered)
    pub fn encode_lrcpc3_ldiapp(&self, rt1: u32, rt2: u32, rn: u32) -> u32 {
        let mut w: u32 = 0;
        w |= 0b11001000 << 24;
        w |= 0b010 << 21;
        w |= 0b0000111111 << 10;
        w |= (rn & 0x1F) << 5;
        w |= (rt2 & 0x1F) << 10;
        w |= (rt1 & 0x1F);
        w
    }

    /// STILP: Store release pair (ordered)
    pub fn encode_lrcpc3_stilp(&self, rt1: u32, rt2: u32, rn: u32) -> u32 {
        let mut w: u32 = 0;
        w |= 0b11001000 << 24;
        w |= 0b000 << 21;
        w |= 0b0000111111 << 10;
        w |= (rn & 0x1F) << 5;
        w |= (rt2 & 0x1F) << 10;
        w |= (rt1 & 0x1F);
        w
    }

    // ========================================================================
    // THE (Target Hardening Extension) — RCTX
    // ========================================================================

    /// RCTX: Restore context
    pub fn encode_the_rctx(&self, xt: u32) -> u32 {
        let mut w: u32 = 0;
        w |= 0b11010101 << 24;
        w |= 0b0000 << 20;
        w |= 0b000011 << 14;
        w |= 0b0010 << 10;
        w |= 0b11111 << 5;
        w |= (xt & 0x1F);
        w
    }

    // ========================================================================
    // CHK — CHKFEAT
    // ========================================================================

    /// CHKFEAT: Check feature
    pub fn encode_chkfeat(&self, xt: u32) -> u32 {
        let mut w: u32 = 0;
        w |= 0b11010101 << 24;
        w |= 0b0011 << 20;
        w |= 0b0000 << 16;
        w |= 0b0100 << 12;
        w |= 0b101 << 9;
        w |= 0b11111 << 5;
        w |= (xt & 0x1F);
        w
    }

    // ========================================================================
    // GCS (Guarded Control Stack) — GCSPUSHX/GCSPOPX
    // ========================================================================

    /// GCSPUSHX: Push to guarded control stack
    pub fn encode_gcs_gcspushx(&self, xt: u32) -> u32 {
        let mut w: u32 = 0;
        w |= 0b11010101 << 24;
        w |= 0b0001 << 20;
        w |= 0b0000 << 16;
        w |= 0b0101 << 12;
        w |= 0b100 << 9;
        w |= 0b11111 << 5;
        w |= (xt & 0x1F);
        w
    }

    /// GCSPOPX: Pop from guarded control stack
    pub fn encode_gcs_gcspopx(&self, xt: u32) -> u32 {
        let mut w: u32 = 0;
        w |= 0b11010101 << 24;
        w |= 0b0001 << 20;
        w |= 0b0000 << 16;
        w |= 0b0101 << 12;
        w |= 0b101 << 9;
        w |= 0b11111 << 5;
        w |= (xt & 0x1F);
        w
    }

    // ========================================================================
    // Checked Pointer Dereference (CPD)
    // ========================================================================

    /// CPD: Check pointer dereference
    pub fn encode_cpd(&self, xd: u32, xn: u32, xm: u32) -> u32 {
        let mut w: u32 = 0;
        w |= 0b10011010 << 24;
        w |= (xm & 0x1F) << 16;
        w |= 0b01 << 12;
        w |= (xn & 0x1F) << 5;
        w |= (xd & 0x1F);
        w
    }

    // ========================================================================
    // Memory Copy/Set Acceleration — CPYFP/CPYFM/CPYFE, SETFP/SETFM/SETFE
    // ========================================================================

    /// CPYFP: Copy forward (prolog)
    pub fn encode_cpyfp(&self, xd: u32, xn: u32, xm: u32) -> u32 {
        let mut w: u32 = 0;
        w |= 0b10011011 << 24;
        w |= (xm & 0x1F) << 16;
        w |= 0b000 << 12;
        w |= (xn & 0x1F) << 5;
        w |= (xd & 0x1F);
        w
    }

    /// CPYFM: Copy forward (main loop)
    pub fn encode_cpyfm(&self, xd: u32, xn: u32, xm: u32) -> u32 {
        let mut w = self.encode_cpyfp(xd, xn, xm);
        w ^= 1 << 9;
        w
    }

    /// CPYFE: Copy forward (epilog)
    pub fn encode_cpyfe(&self, xd: u32, xn: u32, xm: u32) -> u32 {
        let mut w = self.encode_cpyfp(xd, xn, xm);
        w ^= 2 << 9;
        w
    }

    /// SETFP: Set memory forward (prolog)
    pub fn encode_setfp(&self, xd: u32, xn: u32, xm: u32) -> u32 {
        self.encode_cpyfp(xd, xn, xm) ^ (1 << 24)
    }

    /// SETFM: Set memory forward (main loop)
    pub fn encode_setfm(&self, xd: u32, xn: u32, xm: u32) -> u32 {
        self.encode_cpyfm(xd, xn, xm) ^ (1 << 24)
    }

    /// SETFE: Set memory forward (epilog)
    pub fn encode_setfe(&self, xd: u32, xn: u32, xm: u32) -> u32 {
        self.encode_cpyfe(xd, xn, xm) ^ (1 << 24)
    }

    // ========================================================================
    // PAC Enhanced — PACM (PAC Modifier)
    // ========================================================================

    /// PACM: PAC modifier
    pub fn encode_pacm(&self, xd: u32, xn: u32) -> u32 {
        let mut w: u32 = 0;
        w |= 0b11010101 << 24;
        w |= 0b0000 << 20;
        w |= 0b0011 << 16;
        w |= 0b0000 << 12;
        w |= 0b01 << 10;
        w |= (xn & 0x1F) << 5;
        w |= (xd & 0x1F);
        w
    }

    // ========================================================================
    // Additional SME2 multi-vector FMLA/FMLS variants with ZA accumulators
    // ========================================================================

    /// SME2 FMLA with 2 Z registers: FMLA ZA.S[w], {Z0-Z1}, Zn
    pub fn encode_sme2_fmla_2z(&self, zd: u32, zn: u32, za_w: u32) -> u32 {
        self.encode_sme2_fmla_multi(zd, zn, za_w, 0)
    }

    /// SME2 FMLA with 4 Z registers: FMLA ZA.S[w], {Z0-Z3}, Zn
    pub fn encode_sme2_fmla_4z(&self, zd: u32, zn: u32, za_w: u32) -> u32 {
        self.encode_sme2_fmla_multi(zd, zn, za_w, 1)
    }

    /// SME2 FMLS with 2 Z registers: FMLS ZA.S[w], {Z0-Z1}, Zn
    pub fn encode_sme2_fmls_2z(&self, zd: u32, zn: u32, za_w: u32) -> u32 {
        self.encode_sme2_fmls_multi(zd, zn, za_w, 0)
    }

    /// SME2 FMLS with 4 Z registers: FMLS ZA.S[w], {Z0-Z3}, Zn
    pub fn encode_sme2_fmls_4z(&self, zd: u32, zn: u32, za_w: u32) -> u32 {
        self.encode_sme2_fmls_multi(zd, zn, za_w, 1)
    }

    // ========================================================================
    // SVE2.1 — UDOT/BDOT complex dot product
    // ========================================================================

    /// SVE2.1 UDOT (unsigned dot product): UDOT Zda, Pg/m, Zn, Zm
    pub fn encode_sve2_udot(&self, zda: u32, pg: u32, zn: u32, zm: u32) -> u32 {
        let mut w = self.encode_sve2_sdot(zda, pg, zn, zm);
        w ^= 1 << 23;
        w
    }

    /// SVE2.1 BDOT (mixed-sign dot product): BDOT Zda, Pg/m, Zn, Zm
    pub fn encode_sve2_bdot(&self, zda: u32, pg: u32, zn: u32, zm: u32) -> u32 {
        let mut w = self.encode_sve2_sdot(zda, pg, zn, zm);
        w ^= 2 << 23;
        w
    }

    // ========================================================================
    // FAMINMAX vector forms
    // ========================================================================

    /// FAMIN vector: FAMIN Vd.4S, Vn.4S, Vm.4S
    pub fn encode_faminmax_famin_vec(&self, vd: u32, vn: u32, vm: u32) -> u32 {
        let mut w = self.encode_faminmax_famin(vd, vn, vm);
        w |= 0b1 << 30;
        w
    }

    /// FAMAX vector: FAMAX Vd.4S, Vn.4S, Vm.4S
    pub fn encode_faminmax_famax_vec(&self, vd: u32, vn: u32, vm: u32) -> u32 {
        let mut w = self.encode_faminmax_famin_vec(vd, vn, vm);
        w ^= 1 << 24;
        w
    }

    // ========================================================================
    // SME2 Multi-vector Operations
    // ========================================================================

    /// SME2 FMLA multi (4 vectors): FMLA ZA.S[w], {Zd0-Zd3}, Zn.S[w]
    /// Format: 11000001 | 01 | Sz | Zn | 0 | Zm | ...
    pub fn encode_sme2_fmla_mv(&self, zd: u32, zn: u32, zm: u32, num_vecs: u32) -> u32 {
        let mut w: u32 = 0;
        w |= 0b11000001 << 24;
        w |= 0b01 << 22;
        w |= ((num_vecs - 1) & 0x3) << 20;
        w |= (zm & 0x1F) << 16;
        w |= 0b0 << 15;
        w |= (zn & 0x1F) << 5;
        w |= (zd & 0x1F);
        w
    }

    /// SME2 FMLA multi (indexed): FMLA ZA.S[w], {Zd}, Zn.S[w, imm]
    pub fn encode_sme2_fmla_mv_idx(&self, zd: u32, zn: u32, zm: u32, idx: u32) -> u32 {
        let mut w = self.encode_sme2_fmla_mv(zd, zn, zm, 1);
        w |= (idx & 0x3) << 13;
        w
    }

    /// SME2 SMLAL multi: SMLAL ZA.S[w], {Zd}, {Zn}, Zm[w]
    pub fn encode_sme2_smlal_mv(&self, zd: u32, zn: u32, zm: u32) -> u32 {
        let mut w: u32 = 0;
        w |= 0b11000001 << 24;
        w |= 0b10 << 22;
        w |= (zm & 0x1F) << 16;
        w |= 0b0 << 15;
        w |= (zn & 0x1F) << 5;
        w |= (zd & 0x1F);
        w
    }

    /// SME2 SDOT multi: SDOT ZA.S[w], {Zd}, {Zn}, Zm.S[w]
    pub fn encode_sme2_sdot_mv(&self, zd: u32, zn: u32, zm: u32) -> u32 {
        let mut w = self.encode_sme2_smlal_mv(zd, zn, zm);
        w |= 0b01 << 22;
        w
    }

    /// SME2 UDOT multi: UDOT ZA.S[w], {Zd}, {Zn}, Zm.S[w]
    pub fn encode_sme2_udot_mv(&self, zd: u32, zn: u32, zm: u32) -> u32 {
        let mut w = self.encode_sme2_smlal_mv(zd, zn, zm);
        w |= 0b10 << 22;
        w
    }

    /// SME2 BFMOPA (predicated): BFMOPA ZA.S[w], {Zd}, Pn/M, Zm
    pub fn encode_sme2_bfmopa_pred(&self, zd: u32, pm: u32, zn: u32, zm: u32) -> u32 {
        let mut w: u32 = 0;
        w |= 0b10000001 << 24;
        w |= 0b1 << 23;
        w |= (pm & 0xF) << 19;
        w |= (zm & 0x1F) << 16;
        w |= (zn & 0x1F) << 5;
        w |= (zd & 0x1F);
        w
    }

    /// SME2 FMOPA (non-widening): FMOPA ZA.S[w], {Zd}, Pn/M, Zm
    pub fn encode_sme2_fmopa(&self, zd: u32, pm: u32, zn: u32, zm: u32) -> u32 {
        let mut w = self.encode_sme2_bfmopa_pred(zd, pm, zn, zm);
        w ^= 1 << 22;
        w
    }

    /// SME2 SMOPA: SMOPA ZA.S[w], {Zd}, Pn/M, Zm
    pub fn encode_sme2_smopa(&self, zd: u32, pm: u32, zn: u32, zm: u32) -> u32 {
        let mut w = self.encode_sme2_bfmopa_pred(zd, pm, zn, zm);
        w ^= 3 << 22;
        w
    }

    /// SME2 UMOPA: UMOPA ZA.S[w], {Zd}, Pn/M, Zm
    pub fn encode_sme2_umopa(&self, zd: u32, pm: u32, zn: u32, zm: u32) -> u32 {
        let mut w = self.encode_sme2_bfmopa_pred(zd, pm, zn, zm);
        w ^= 2 << 22;
        w
    }

    /// SME2 ZERO (multi): ZERO {Zt0-Zt3}
    pub fn encode_sme2_zero_mv(&self, zt: u32, num_regs: u32) -> u32 {
        let mut w: u32 = 0;
        w |= 0b11000001 << 24;
        w |= 0b11 << 22;
        w |= ((num_regs - 1) & 0x3) << 20;
        w |= 0b1 << 21;
        w |= 0b0 << 15;
        w |= (zt & 0x1F);
        w
    }

    /// SME2 MOVPRFX (multi): MOVPRFX {Zd}, Pn/M, Zm
    pub fn encode_sme2_movprfx_mv(&self, zd: u32, pm: u32, zn: u32, zm: u32) -> u32 {
        let mut w: u32 = 0;
        w |= 0b11000001 << 24;
        w |= 0b0 << 23;
        w |= (pm & 0xF) << 17;
        w |= (zm & 0x1F) << 16;
        w |= (zn & 0x1F) << 5;
        w |= (zd & 0x1F);
        w
    }

    /// SME2 LUTI2 (2-register table lookup): LUTI2 {Zt.B}, {Zt.B}, Zm[imm]
    pub fn encode_sme2_luti2(&self, zt: u32, zm: u32, idx: u32) -> u32 {
        let mut w: u32 = 0;
        w |= 0b11000001 << 24;
        w |= 0b10 << 22;
        w |= (zm & 0x1F) << 16;
        w |= (idx & 0x3) << 13;
        w |= (zt & 0x1F);
        w
    }

    /// SME2 LUTI4 (4-register table lookup): LUTI4 {Zt.B-Z(t+3).B}, {Zt.B-Z(t+3).B}, Zm[imm]
    pub fn encode_sme2_luti4(&self, zt: u32, zm: u32, idx: u32) -> u32 {
        let mut w = self.encode_sme2_luti2(zt, zm, idx);
        w |= 0b01 << 20;
        w
    }

    // ========================================================================
    // SME2 FP8 to FP32/BF16 Convert
    // ========================================================================

    /// SME2 F1CVT (FP8 to FP32 via table): F1CVT {Zd.S}, {Zn.H}
    pub fn encode_sme2_f1cvt(&self, zd: u32, zn: u32) -> u32 {
        let mut w: u32 = 0;
        w |= 0b11000001 << 24;
        w |= 0b01 << 22;
        w |= 0b1 << 15;
        w |= (zn & 0x1F) << 5;
        w |= (zd & 0x1F);
        w
    }

    /// SME2 F1CVTL (FP8 to FP32 down-convert): F1CVTL {Zd.S}, {Zn.H}, #imm
    pub fn encode_sme2_f1cvtl(&self, zd: u32, zn: u32, imm: u32) -> u32 {
        let mut w = self.encode_sme2_f1cvt(zd, zn);
        w |= (imm & 0x1) << 14;
        w
    }

    /// SME2 BF1CVT (FP8 to BF16 via table): BF1CVT {Zd.H}, {Zn.B}
    pub fn encode_sme2_bf1cvt(&self, zd: u32, zn: u32) -> u32 {
        let mut w: u32 = 0;
        w |= 0b11000001 << 24;
        w |= 0b01 << 22;
        w |= 0b0 << 15;
        w |= (zn & 0x1F) << 5;
        w |= (zd & 0x1F);
        w
    }

    /// SME2 BF1CVTL (FP8 to BF16 down-convert): BF1CVTL {Zd.H}, {Zn.B}
    pub fn encode_sme2_bf1cvtl(&self, zd: u32, zn: u32) -> u32 {
        let mut w = self.encode_sme2_bf1cvt(zd, zn);
        w |= 0b1 << 14;
        w
    }

    /// SME2 F2CVT (FP32 to FP8 via table): F2CVT {Zd.H}, Zd.S
    pub fn encode_sme2_f2cvt(&self, zd: u32, zn: u32) -> u32 {
        let mut w: u32 = 0;
        w |= 0b11000001 << 24;
        w |= 0b11 << 22;
        w |= 0b0 << 21;
        w |= (zn & 0x1F) << 5;
        w |= (zd & 0x1F);
        w
    }

    // ========================================================================
    // SME2 Multi-vector Load/Store
    // ========================================================================

    /// SME2 LDR (multi-vector, strided): LDR {Zt0-Zt3}, [Xn, #imm, MUL VL]
    pub fn encode_sme2_ldr_mv_strided(&self, zt: u32, xn: u32, imm: u32) -> u32 {
        let mut w: u32 = 0;
        w |= 0b11000101 << 24;
        w |= 0b1 << 22;
        w |= ((imm >> 3) & 0x7) << 17;
        w |= (xn & 0x1F) << 5;
        w |= (zt & 0x1F);
        w
    }

    /// SME2 STR (multi-vector, strided): STR {Zt0-Zt3}, [Xn, #imm, MUL VL]
    pub fn encode_sme2_str_mv_strided(&self, zt: u32, xn: u32, imm: u32) -> u32 {
        let mut w = self.encode_sme2_ldr_mv_strided(zt, xn, imm);
        w ^= 1 << 22;
        w
    }

    /// SME2 LDR (multi-vector, contiguous): LDR {Zt0-Zt3}, [Xn]
    pub fn encode_sme2_ldr_mv_contig(&self, zt: u32, xn: u32) -> u32 {
        let mut w: u32 = 0;
        w |= 0b11000101 << 24;
        w |= 0b0 << 22;
        w |= (xn & 0x1F) << 5;
        w |= (zt & 0x1F);
        w
    }

    /// SME2 STR (multi-vector, contiguous): STR {Zt0-Zt3}, [Xn]
    pub fn encode_sme2_str_mv_contig(&self, zt: u32, xn: u32) -> u32 {
        let mut w = self.encode_sme2_ldr_mv_contig(zt, xn);
        w ^= 1 << 22;
        w
    }

    // ========================================================================
    // SME2 PSEL (predicate pair select)
    // ========================================================================

    /// SME2 PSEL: PSEL Pn, Pm.Pm+1
    pub fn encode_sme2_psel(&self, pd: u32, pn: u32, pm: u32) -> u32 {
        let mut w: u32 = 0;
        w |= 0b11000001 << 24;
        w |= 0b0 << 23;
        w |= (pn & 0xF) << 19;
        w |= 0b1 << 18;
        w |= (pm & 0xF) << 14;
        w |= (pd & 0xF) << 5;
        w
    }

    /// SME2 REVD (reverse doublewords): REVD Zd.Q, Pn/M, Zm.Q
    pub fn encode_sme2_revd(&self, zd: u32, pn: u32, zm: u32) -> u32 {
        let mut w: u32 = 0;
        w |= 0b00000101 << 24;
        w |= 0b1 << 23;
        w |= (pn & 0xF) << 19;
        w |= 0b1 << 18;
        w |= (zm & 0x1F) << 5;
        w |= (zd & 0x1F);
        w
    }

    /// SME2 SCLAMP (signed clamp): SCLAMP Zd.S, Zn.S, Zm.S
    pub fn encode_sme2_sclamp(&self, zd: u32, zn: u32, zm: u32) -> u32 {
        let mut w = self.encode_sme2_revd(zd, 0, zm);
        w |= (zn & 0x1F) << 16;
        w
    }

    /// SME2 UCLAMP (unsigned clamp): UCLAMP Zd.S, Zn.S, Zm.S
    pub fn encode_sme2_uclamp(&self, zd: u32, zn: u32, zm: u32) -> u32 {
        let mut w = self.encode_sme2_sclamp(zd, zn, zm);
        w ^= 1 << 23;
        w
    }
}

// ============================================================================
// AArch64 Bitmask Immediate Encoder
// ============================================================================

/// Encode a 64-bit value as an AArch64 bitmask immediate.
///
/// AArch64 bitmask immediates are encoded as N:immr:imms, representing
/// a contiguous run of set bits (width = imms+1 for 32-bit, or imms+1 for
/// 64-bit), rotated right by immr positions within the element.
///
/// Returns `Some((immr, imms, n))` if encodable, or `None` otherwise.
///
/// The N bit is 0 for 32-bit elements and 1 for 64-bit elements.
pub fn encode_bitmask_imm(value: u64, size: u8) -> Option<(u32, u32, u32)> {
    match size {
        32 => encode_bitmask_imm_32(value as u32),
        64 => encode_bitmask_imm_64(value),
        _ => None,
    }
}

/// Encode a 32-bit bitmask immediate.
fn encode_bitmask_imm_32(value: u32) -> Option<(u32, u32, u32)> {
    // Replicate the pattern to fill 32 bits
    let v = value;

    // Check if the value is all zeroes or all ones
    if v == 0 || v == 0xFFFFFFFF {
        // These can be encoded with imms and immr
        let imms = if v == 0 {
            0x1F // For 32-bit: imms < 32 means all-zeroes
        } else {
            0x20 // For 32-bit: imms = 32 is special
        };
        return Some((0, imms, 0));
    }

    // Try to find a rotation that makes the set bits contiguous
    for rotation in 0..32 {
        let rotated = v.rotate_right(rotation);
        // Check if the rotated value is a contiguous run of ones
        if let Some(run_len) = find_contiguous_ones_32(rotated) {
            let imms = encode_imms_32(run_len, rotation);
            if imms < 32 {
                // 32-bit element, N=0
                return Some((rotation, imms, 0));
            }
        }
    }

    None
}

/// Encode a 64-bit bitmask immediate.
fn encode_bitmask_imm_64(value: u64) -> Option<(u32, u32, u32)> {
    let v = value;

    // First try as a 32-bit replicated pattern
    let lo = v as u32;
    let hi = (v >> 32) as u32;
    if lo == hi {
        if let Some((immr, imms, _)) = encode_bitmask_imm_32(lo) {
            return Some((immr, imms, 0));
        }
    }

    // Try 64-bit encoding
    if v == 0 || v == 0xFFFFFFFFFFFFFFFF {
        let imms = if v == 0 { 0x3F } else { 0x40 };
        return Some((0, imms, 1));
    }

    for rotation in 0..64 {
        let rotated = v.rotate_right(rotation as u32);
        if let Some(run_len) = find_contiguous_ones_64(rotated) {
            let (immr_val, imms_val) = encode_imms_64(run_len, rotation);
            if imms_val < 64 {
                return Some((immr_val as u32, imms_val, 1));
            }
        }
    }
    None
}

/// Find the length of a contiguous run of 1s at the LSB end of a 32-bit value.
/// Returns the run length, or None if the value isn't a contiguous run.
fn find_contiguous_ones_32(v: u32) -> Option<u32> {
    if v == 0 {
        return Some(0);
    }
    // Count trailing ones
    let tz = v.trailing_ones();
    // Check that after the run, the rest is all zeroes
    let mask = (1u64 << tz) - 1;
    let remaining = v as u64 & !mask;
    if remaining == 0 {
        Some(tz)
    } else {
        None
    }
}

/// Find the length of a contiguous run of 1s at the LSB end of a 64-bit value.
fn find_contiguous_ones_64(v: u64) -> Option<u32> {
    if v == 0 {
        return Some(0);
    }
    let tz = v.trailing_ones() as u32;
    let mask = (1u64 << tz) - 1;
    let remaining = v & !mask;
    if remaining == 0 {
        Some(tz)
    } else {
        None
    }
}

/// Encode imms for 32-bit element size.
/// For a run of `len` ones at position 0 (after rotation),
/// imms encodes the run length as `imms = r - 1` in 5 bits.
fn encode_imms_32(run_len: u32, _immr: u32) -> u32 {
    // The pattern is: replicate an element of size 2^k.
    // The element has `run_len` consecutive ones.
    // imms encodes the pattern length minus 1.
    if run_len == 0 {
        return 0;
    }
    run_len.wrapping_sub(1) & 0x1F
}

/// Encode imms for 64-bit element size.
fn encode_imms_64(run_len: u32, immr: u32) -> (u32, u32) {
    if run_len == 0 {
        return (0, 0);
    }
    let imms = run_len.wrapping_sub(1) & 0x3F;
    (immr & 0x3F, imms)
}

// ============================================================================
// AArch64 immediate encoder (element + rotation)
// ============================================================================

/// Encode a 64-bit value as an AArch64 immediate (element + rotation).
/// This is the public API for immediate encoding used by other modules.
/// Returns (encoded_value, rotation) for use in instruction construction.
pub fn encode_immediate_aarch64(value: u64, size: u8) -> (u32, u32) {
    if let Some((immr, imms, n)) = encode_bitmask_imm(value, size) {
        // Pack immr and imms into a single value
        let encoded = (n << 12) | (immr << 6) | imms;
        (encoded, immr)
    } else {
        (0, 0)
    }
}

// ============================================================================
// Tests
// ============================================================================

#[cfg(test)]
mod tests {
    use super::*;

    // ========================================================================
    // Helper: create a MachineInstr with opcode and operands
    // ========================================================================

    fn make_instr(op: u32) -> MachineInstr {
        MachineInstr::new(op)
    }

    fn make_instr_imm(op: u32, rd: u32, rn: u32, imm: i64) -> MachineInstr {
        let mut mi = MachineInstr::new(op);
        mi.operands.push(MachineOperand::PhysReg(rd));
        mi.operands.push(MachineOperand::PhysReg(rn));
        mi.operands.push(MachineOperand::Imm(imm));
        mi
    }

    fn make_instr_reg(op: u32, rd: u32, rn: u32, rm: u32) -> MachineInstr {
        let mut mi = MachineInstr::new(op);
        mi.operands.push(MachineOperand::PhysReg(rd));
        mi.operands.push(MachineOperand::PhysReg(rn));
        mi.operands.push(MachineOperand::PhysReg(rm));
        mi
    }

    fn make_instr_two_reg(op: u32, r1: u32, r2: u32) -> MachineInstr {
        let mut mi = MachineInstr::new(op);
        mi.operands.push(MachineOperand::PhysReg(r1));
        mi.operands.push(MachineOperand::PhysReg(r2));
        mi
    }

    fn make_instr_one_reg(op: u32, r: u32) -> MachineInstr {
        let mut mi = MachineInstr::new(op);
        mi.operands.push(MachineOperand::PhysReg(r));
        mi
    }

    // ========================================================================
    // Register field tests
    // ========================================================================

    #[test]
    fn test_get_reg_field_x0() {
        assert_eq!(ArmMCEncoder::get_reg_field(2000), 0); // X0
        assert_eq!(ArmMCEncoder::get_reg_field(2030), 30); // X30
    }

    #[test]
    fn test_get_reg_field_sp_xzr() {
        assert_eq!(ArmMCEncoder::get_reg_field(2031), 31); // SP
        assert_eq!(ArmMCEncoder::get_reg_field(2032), 31); // XZR
    }

    #[test]
    fn test_get_reg_field_w_regs() {
        assert_eq!(ArmMCEncoder::get_reg_field(2040), 0); // W0
        assert_eq!(ArmMCEncoder::get_reg_field(2070), 30); // W30
        assert_eq!(ArmMCEncoder::get_reg_field(2072), 31); // WZR
    }

    #[test]
    fn test_get_reg_field_arm32() {
        assert_eq!(ArmMCEncoder::get_reg_field(2100), 0); // R0
        assert_eq!(ArmMCEncoder::get_reg_field(2115), 15); // R15
    }

    // ========================================================================
    // AArch64 NOP
    // ========================================================================

    #[test]
    fn test_encode_nop() {
        let enc = ArmMCEncoder::new(true);
        let mi = make_instr(OP_NOP);
        let bytes = enc.encode_instruction(&mi);
        assert_eq!(bytes.len(), 4);
        assert_eq!(bytes, vec![0x1F, 0x20, 0x03, 0xD5]); // 0xD503201F LE
    }

    // ========================================================================
    // AArch64 ADD immediate
    // ========================================================================

    #[test]
    fn test_encode_add_imm() {
        let enc = ArmMCEncoder::new(true);
        // ADD X0, X1, #42
        let mi = make_instr_imm(OP_ADD, 2000, 2001, 42);
        let bytes = enc.encode_instruction(&mi);
        assert_eq!(bytes.len(), 4);
        let word = u32::from_le_bytes([bytes[0], bytes[1], bytes[2], bytes[3]]);
        // ADD X0, X1, #42:
        // sf=1, op=0, S=0, 10001, shift=0, imm12=42, Rn=X1=1, Rd=X0=0
        // Binary: 1 0 0 10001 00 000000101010 00001 00000
        // = 0x9100A820? Let me compute: sf(1)<<31 + 10001<<24 + 42<<10 + 1<<5 + 0
        assert_eq!(word, 0x9100A820);
    }

    #[test]
    fn test_encode_sub_imm() {
        let enc = ArmMCEncoder::new(true);
        // SUB X5, X3, #16
        let mi = make_instr_imm(OP_SUB, 2005, 2003, 16);
        let bytes = enc.encode_instruction(&mi);
        let word = u32::from_le_bytes([bytes[0], bytes[1], bytes[2], bytes[3]]);
        // sf=1, op=1, S=0, 10001, shift=0, imm12=16, Rn=X3=3, Rd=X5=5
        // 1 1 0 10001 00 000000010000 00011 00101
        // = 0xD100_4065
        assert_eq!(word, 0xD1004065);
    }

    // ========================================================================
    // AArch64 ADD shifted immediate
    // ========================================================================

    #[test]
    fn test_encode_add_imm_shifted() {
        let enc = ArmMCEncoder::new(true);
        // ADD X0, X1, #0x1000 (4096 = 1<<12)
        let mi = make_instr_imm(OP_ADD, 2000, 2001, 4096);
        let bytes = enc.encode_instruction(&mi);
        let word = u32::from_le_bytes([bytes[0], bytes[1], bytes[2], bytes[3]]);
        // sf=1, op=0, S=0, 10001, shift=1, imm12=1, Rn=X1=1, Rd=X0=0
        // 1 0 0 10001 01 000000000001 00001 00000 = 0x9140_0420
        assert_eq!(word, 0x91400420);
    }

    // ========================================================================
    // AArch64 ADD register
    // ========================================================================

    #[test]
    fn test_encode_add_reg() {
        let enc = ArmMCEncoder::new(true);
        // ADD X2, X3, X4
        let mi = make_instr_reg(OP_ADD, 2002, 2003, 2004);
        let bytes = enc.encode_instruction(&mi);
        let word = u32::from_le_bytes([bytes[0], bytes[1], bytes[2], bytes[3]]);
        // sf=1, op=0, S=0, 01011, shift=0, 0, Rm=4, imm6=0, Rn=3, Rd=2
        // 1 0 0 01011 00 0 00100 000000 00011 00010
        // = 0x8B04_0062
        assert_eq!(word, 0x8B040062);
    }

    // ========================================================================
    // AArch64 MOVZ / MOVK
    // ========================================================================

    #[test]
    fn test_encode_movz() {
        let enc = ArmMCEncoder::new(true);
        // MOVZ X0, #42, lsl #0
        let mut mi = MachineInstr::new(OP_MOVZ);
        mi.operands.push(MachineOperand::PhysReg(2000)); // X0
        mi.operands.push(MachineOperand::Imm(42));
        mi.operands.push(MachineOperand::Imm(0)); // lsl #0
        let bytes = enc.encode_instruction(&mi);
        let word = u32::from_le_bytes([bytes[0], bytes[1], bytes[2], bytes[3]]);
        // sf=1, 10, 100101, hw=0, imm16=42, Rd=0
        // 1 10 100101 00 0000000000101010 00000
        // = 0xD2800540
        assert_eq!(word, 0xD2800540);
    }

    #[test]
    fn test_encode_movk() {
        let enc = ArmMCEncoder::new(true);
        // MOVK X1, #0xABCD, lsl #16
        let mut mi = MachineInstr::new(OP_MOVK);
        mi.operands.push(MachineOperand::PhysReg(2001)); // X1
        mi.operands.push(MachineOperand::Imm(0xABCD));
        mi.operands.push(MachineOperand::Imm(16)); // lsl #16 → hw=1
        let bytes = enc.encode_instruction(&mi);
        let word = u32::from_le_bytes([bytes[0], bytes[1], bytes[2], bytes[3]]);
        // sf=1, 11, 100101, hw=1, imm16=0xABCD, Rd=1
        let expected =
            (1u32 << 31) | (0b11 << 29) | (0b100101 << 23) | (1u32 << 21) | (0xABCD << 5) | 1;
        assert_eq!(word, expected);
    }

    // ========================================================================
    // AArch64 RET / BR
    // ========================================================================

    #[test]
    fn test_encode_ret() {
        let enc = ArmMCEncoder::new(true);
        // RET
        let mi = make_instr_one_reg(OP_RET, 2030); // X30 (LR)
        let bytes = enc.encode_instruction(&mi);
        let word = u32::from_le_bytes([bytes[0], bytes[1], bytes[2], bytes[3]]);
        // 0xD65F0000 | (30 << 5) = 0xD65F03C0
        assert_eq!(word, 0xD65F03C0);
    }

    #[test]
    fn test_encode_ret_default() {
        let enc = ArmMCEncoder::new(true);
        let mi = make_instr(OP_RET); // no operands → default LR
        let bytes = enc.encode_instruction(&mi);
        let word = u32::from_le_bytes([bytes[0], bytes[1], bytes[2], bytes[3]]);
        assert_eq!(word, 0xD65F03C0);
    }

    // ========================================================================
    // AArch64 CMP
    // ========================================================================

    #[test]
    fn test_encode_cmp_imm() {
        let enc = ArmMCEncoder::new(true);
        // CMP X0, #0 — only 2 operands: Rn, imm
        let mut mi = MachineInstr::new(OP_CMP);
        mi.operands.push(MachineOperand::PhysReg(2000)); // X0 (Rn)
        mi.operands.push(MachineOperand::Imm(0));
        let bytes = enc.encode_instruction(&mi);
        let word = u32::from_le_bytes([bytes[0], bytes[1], bytes[2], bytes[3]]);
        // SUBS XZR, X0, #0: sf=1, op=1, S=1, 10001, sh=0, imm12=0, Rn=0, Rd=31
        let expected = (1u32 << 31)
            | (1 << 30)
            | (1 << 29)
            | (0b10001 << 24)
            | (0 << 22)
            | (0 << 10)
            | (0u32 << 5)
            | 31;
        assert_eq!(word, expected);
    }

    #[test]
    fn test_encode_cmp_reg() {
        let enc = ArmMCEncoder::new(true);
        // CMP X1, X2
        let mi = make_instr_two_reg(OP_CMP, 2001, 2002);
        let bytes = enc.encode_instruction(&mi);
        let word = u32::from_le_bytes([bytes[0], bytes[1], bytes[2], bytes[3]]);
        // SUBS XZR, X1, X2: sf=1, op=1, S=1, 01011, Rm=2, imm6=0, Rn=1, Rd=31
        let expected = (1u32 << 31)
            | (1 << 30)
            | (1 << 29)
            | (0b01011 << 24)
            | (2u32 << 16)
            | (1u32 << 5)
            | 31;
        assert_eq!(word, expected);
    }

    // ========================================================================
    // AArch64 LDR / STR
    // ========================================================================

    #[test]
    fn test_encode_ldr_imm() {
        let enc = ArmMCEncoder::new(true);
        // LDR X0, [X1, #8]
        let mi = make_instr_imm(OP_LDR, 2000, 2001, 8);
        let bytes = enc.encode_instruction(&mi);
        let word = u32::from_le_bytes([bytes[0], bytes[1], bytes[2], bytes[3]]);
        // LDR 64-bit: sf=11, 11100101 in 29-22, imm9=1, 01 in 11-10, Rn=1, Rt=0
        let expected =
            (0b11u32 << 30) | (0b1110_0101 << 22) | (1u32 << 12) | (0b01 << 10) | (1u32 << 5) | 0;
        assert_eq!(word, expected);
    }

    #[test]
    fn test_encode_str_imm() {
        let enc = ArmMCEncoder::new(true);
        // STR X5, [SP, #16]
        let mi = make_instr_imm(OP_STR, 2005, 2031, 16);
        let bytes = enc.encode_instruction(&mi);
        let word = u32::from_le_bytes([bytes[0], bytes[1], bytes[2], bytes[3]]);
        // STR 64-bit: sf=11, 11100100 in 29-22, imm9=2, 01 in 11-10, Rn=31, Rt=5
        let expected =
            (0b11u32 << 30) | (0b1110_0100 << 22) | (2u32 << 12) | (0b01 << 10) | (31u32 << 5) | 5;
        assert_eq!(word, expected);
    }

    // ========================================================================
    // AArch64 LDP / STP
    // ========================================================================

    #[test]
    fn test_encode_ldp() {
        let enc = ArmMCEncoder::new(true);
        // LDP X0, X1, [X2, #0]
        let mut mi = MachineInstr::new(OP_LDP);
        mi.operands.push(MachineOperand::PhysReg(2000)); // X0 (Rt)
        mi.operands.push(MachineOperand::PhysReg(2001)); // X1 (Rt2)
        mi.operands.push(MachineOperand::PhysReg(2002)); // X2 (Rn)
        mi.operands.push(MachineOperand::Imm(0)); // offset 0
        let bytes = enc.encode_instruction(&mi);
        let word = u32::from_le_bytes([bytes[0], bytes[1], bytes[2], bytes[3]]);
        // LDP 64-bit: opc=10, 10100101, imm7=0, Rt2=1, Rn=2, Rt=0
        // 10_10100101_0000000_00001_00010_00000
        // opc=10 << 30 = 0x80000000
        // 10100101 << 22 = 0x29400000
        // Rt2=1 << 10 = 0x400
        // Rn=2 << 5 = 0x40
        // Rt=0
        // = 0xA940_0440
        assert_eq!(word, 0xA9400440);
    }

    // ========================================================================
    // AArch64 CSEL
    // ========================================================================

    #[test]
    fn test_encode_csel() {
        let enc = ArmMCEncoder::new(true);
        // CSEL X0, X1, X2, EQ
        let mut mi = MachineInstr::new(OP_CSEL);
        mi.operands.push(MachineOperand::PhysReg(2000)); // Rd = X0
        mi.operands.push(MachineOperand::PhysReg(2001)); // Rn = X1
        mi.operands.push(MachineOperand::PhysReg(2002)); // Rm = X2
        mi.operands.push(MachineOperand::Imm(ArmCond::EQ as i64)); // cond = EQ
        let bytes = enc.encode_instruction(&mi);
        let word = u32::from_le_bytes([bytes[0], bytes[1], bytes[2], bytes[3]]);
        // sf=1, 0, 0, 11010100, Rm=2, cond=0, 00, Rn=1, Rd=0
        // 1_0_0_11010100_00010_0000_00_00001_00000 = 0x9A82_0020
        assert_eq!(word, 0x9A820020);
    }

    // ========================================================================
    // AArch64 SDIV / UDIV
    // ========================================================================

    #[test]
    fn test_encode_sdiv() {
        let enc = ArmMCEncoder::new(true);
        let mi = make_instr_reg(OP_SDIV, 2003, 2004, 2005);
        let bytes = enc.encode_instruction(&mi);
        let word = u32::from_le_bytes([bytes[0], bytes[1], bytes[2], bytes[3]]);
        // SDIV X3, X4, X5
        // sf=1, 0, 0, 11010110, Rm=5, opcode=000011, Rn=4, Rd=3
        // 1_0_0_11010110_00101_000011_00100_00011 = 0x9AC5_0C83
        assert_eq!(word, 0x9AC50C83);
    }

    #[test]
    fn test_encode_udiv() {
        let enc = ArmMCEncoder::new(true);
        let mi = make_instr_reg(OP_UDIV, 2000, 2001, 2002);
        let bytes = enc.encode_instruction(&mi);
        let word = u32::from_le_bytes([bytes[0], bytes[1], bytes[2], bytes[3]]);
        // UDIV X0, X1, X2
        // opcode=000010
        // 1_0_0_11010110_00010_000010_00001_00000 = 0x9AC2_0820
        assert_eq!(word, 0x9AC20820);
    }

    // ========================================================================
    // ARM32 encoding
    // ========================================================================

    #[test]
    fn test_encode_arm32_mov() {
        let enc = ArmMCEncoder::new(false); // ARM32
        let mut mi = MachineInstr::new(OP_ARM_MOV);
        mi.operands.push(MachineOperand::PhysReg(2101)); // R1 (Rd)
        mi.operands.push(MachineOperand::PhysReg(2102)); // R2 (Rm)
        let bytes = enc.encode_instruction(&mi);
        let word = u32::from_le_bytes([bytes[0], bytes[1], bytes[2], bytes[3]]);
        // ARM32 MOV R1, R2:
        // cond=1110, 00, I=0, opcode=1101, S=0, Rn=0000, Rd=0001, operand2=Rm=0010
        // = 0xE1A01002
        assert_eq!(word, 0xE1A01002);
    }

    #[test]
    fn test_encode_arm32_add() {
        let enc = ArmMCEncoder::new(false);
        let mut mi = MachineInstr::new(OP_ARM_ADD);
        mi.operands.push(MachineOperand::PhysReg(2100)); // R0 (Rd)
        mi.operands.push(MachineOperand::PhysReg(2101)); // R1 (Rn)
        mi.operands.push(MachineOperand::PhysReg(2102)); // R2 (Rm)
        let bytes = enc.encode_instruction(&mi);
        let word = u32::from_le_bytes([bytes[0], bytes[1], bytes[2], bytes[3]]);
        // cond=1110, 00, I=0, opcode=0100, S=0, Rn=0001, Rd=0000, Rm=0010
        // = 0xE0810002
        assert_eq!(word, 0xE0810002);
    }

    #[test]
    fn test_encode_arm32_push_pop() {
        let enc = ArmMCEncoder::new(false);
        // PUSH {R4, LR}
        let mut mi = MachineInstr::new(OP_ARM_PUSH);
        mi.operands.push(MachineOperand::PhysReg(2104)); // R4
        mi.operands.push(MachineOperand::PhysReg(2114)); // LR (R14)
        let bytes = enc.encode_instruction(&mi);
        let word = u32::from_le_bytes([bytes[0], bytes[1], bytes[2], bytes[3]]);
        // STMDB SP!, {R4, LR}: cond=1110, 10010010, Rn=1101(SP), reg_list=0x4010
        // = 0xE92D4010
        assert_eq!(word, 0xE92D4010);
    }

    // ========================================================================
    // BFM (bitfield move)
    // ========================================================================

    #[test]
    fn test_encode_bfm() {
        let enc = ArmMCEncoder::new(true);
        // BFM X0, X1, #0, #7  (insert bits 0-7 of X1 into X0)
        let mut mi = MachineInstr::new(OP_BFM);
        mi.operands.push(MachineOperand::PhysReg(2000)); // Rd = X0
        mi.operands.push(MachineOperand::PhysReg(2001)); // Rn = X1
        mi.operands.push(MachineOperand::Imm(0)); // immr
        mi.operands.push(MachineOperand::Imm(7)); // imms
        let bytes = enc.encode_instruction(&mi);
        let word = u32::from_le_bytes([bytes[0], bytes[1], bytes[2], bytes[3]]);
        // sf=1, opc=01, 100110, N=0, immr=0, imms=7, Rn=1, Rd=0
        let expected =
            (1u32 << 31) | (0b01 << 29) | (0b100110 << 23) | (7u32 << 10) | (1u32 << 5) | 0;
        assert_eq!(word, expected);
    }

    // ========================================================================
    // ADDS / SUBS
    // ========================================================================

    #[test]
    fn test_encode_adds() {
        let enc = ArmMCEncoder::new(true);
        // ADDS X0, X1, X2
        let mi = make_instr_reg(OP_ADDS, 2000, 2001, 2002);
        let bytes = enc.encode_instruction(&mi);
        let word = u32::from_le_bytes([bytes[0], bytes[1], bytes[2], bytes[3]]);
        // sf=1, op=0, S=1, 01011, shift=0, Rm=2, imm6=0, Rn=1, Rd=0
        // 1_0_1_01011_00_0_00010_000000_00001_00000 = 0xAB020020
        assert_eq!(word, 0xAB020020);
    }

    #[test]
    fn test_encode_madd() {
        let enc = ArmMCEncoder::new(true);
        // MADD X0, X1, X2, X3
        let mut mi = MachineInstr::new(OP_MADD);
        mi.operands.push(MachineOperand::PhysReg(2000)); // Rd = X0
        mi.operands.push(MachineOperand::PhysReg(2001)); // Rn = X1
        mi.operands.push(MachineOperand::PhysReg(2002)); // Rm = X2
        mi.operands.push(MachineOperand::PhysReg(2003)); // Ra = X3
        let bytes = enc.encode_instruction(&mi);
        let word = u32::from_le_bytes([bytes[0], bytes[1], bytes[2], bytes[3]]);
        // sf=1, 0, 0, 11011, 000, Rm=2, 0, Ra=3, Rn=1, Rd=0
        let expected = (1u32 << 31)
            | (0b11011 << 24)
            | (2u32 << 16)
            | (0u32 << 15)
            | (3u32 << 10)
            | (1u32 << 5)
            | 0;
        assert_eq!(word, expected);
    }

    #[test]
    fn test_encode_msub() {
        let enc = ArmMCEncoder::new(true);
        // MSUB X0, X1, X2, X3
        let mut mi = MachineInstr::new(OP_MSUB);
        mi.operands.push(MachineOperand::PhysReg(2000));
        mi.operands.push(MachineOperand::PhysReg(2001));
        mi.operands.push(MachineOperand::PhysReg(2002));
        mi.operands.push(MachineOperand::PhysReg(2003));
        let bytes = enc.encode_instruction(&mi);
        let word = u32::from_le_bytes([bytes[0], bytes[1], bytes[2], bytes[3]]);
        // Same as MADD but bit 15 = 1
        let expected = (1u32 << 31)
            | (0b11011 << 24)
            | (2u32 << 16)
            | (1u32 << 15)
            | (3u32 << 10)
            | (1u32 << 5)
            | 0;
        assert_eq!(word, expected);
    }

    // ========================================================================
    // Bitmask immediate encoder
    // ========================================================================

    #[test]
    fn test_encode_bitmask_32() {
        // 0xFFFF0000 should be encodable as rotated value
        let result = encode_bitmask_imm(0xFFFF0000, 64);
        assert!(result.is_some());

        // 0x55555555 is NOT encodable (alternating bits, not contiguous)
        let result = encode_bitmask_imm(0x55555555, 32);
        assert!(result.is_none());

        // 0x00FF0000 — actually this might not be encodable with standard
        // bitmask as it needs a specific pattern
    }

    #[test]
    fn test_encode_bitmask_all_ones() {
        let result = encode_bitmask_imm(0xFFFFFFFFFFFFFFFF, 64);
        assert!(result.is_some());
    }

    #[test]
    fn test_encode_bitmask_zero() {
        let result = encode_bitmask_imm(0, 64);
        assert!(result.is_some());
    }

    // ========================================================================
    // SVE/SME/MTE encoding tests
    // ========================================================================

    #[test]
    fn test_encode_sve_pred_and() {
        let enc = ArmMCEncoder::new(true);
        let word = enc.encode_sve_pred_and(0, 1, 2, 3); // AND P0, P1/z, P2, P3
                                                        // 00100101 | 00 | 0 | 0011 | 0 | 0001 | 0010 | 0 | 0000
        let expected = 0b00100101_00_0_0011_0_0001_0010_0_0000u32;
        assert_eq!(word, expected);
    }

    #[test]
    fn test_encode_sve_pred_orr() {
        let enc = ArmMCEncoder::new(true);
        let word = enc.encode_sve_pred_orr(1, 2, 3, 4);
        // Same as AND but opc=01
        assert_eq!((word >> 22) & 0x3, 1);
    }

    #[test]
    fn test_encode_sve_pred_eor() {
        let enc = ArmMCEncoder::new(true);
        let word = enc.encode_sve_pred_eor(0, 0, 0, 0);
        assert_eq!((word >> 22) & 0x3, 2);
    }

    #[test]
    fn test_encode_sve_ptrue() {
        let enc = ArmMCEncoder::new(true);
        // PTRUE P0, VL1 (pattern 00000)
        let word = enc.encode_sve_ptrue(0, 0b00000);
        // 00100101 | 01 | 1 | 00000 | 00000 | 10 | pattern | 0 | Pd
        assert_eq!(word & 0xFF000000, 0x25000000);
    }

    #[test]
    fn test_encode_sve_pfirst() {
        let enc = ArmMCEncoder::new(true);
        let word = enc.encode_sve_pfirst(0, 1, 2);
        // Should have PFIRST encoding
        assert!(word != 0);
    }

    #[test]
    fn test_encode_sve_add_z() {
        let enc = ArmMCEncoder::new(true);
        // ADD Z0.D, P0/M, Z0.D, Z1.D (size=3 for 64-bit)
        let word = enc.encode_sve_add_z(0, 0, 1, 3);
        // 00000100 | 11 | 1 | 0 | 0 | 0 | 0000 | 00001 | 00000
        assert_eq!((word >> 24) & 0xFF, 0x04);
    }

    #[test]
    fn test_encode_sve_sub_z() {
        let enc = ArmMCEncoder::new(true);
        let add_word = enc.encode_sve_add_z(0, 0, 1, 0);
        let sub_word = enc.encode_sve_sub_z(0, 0, 1, 0);
        // Sub should have bit 13 set
        assert_eq!(sub_word & (1 << 13), 1 << 13);
        assert_ne!(add_word, sub_word);
    }

    #[test]
    fn test_encode_sve_mul_z() {
        let enc = ArmMCEncoder::new(true);
        // MUL Z0.S, P0/M, Z1.S, Z2.S
        let word = enc.encode_sve_mul_z(0, 0, 1, 2, 2);
        assert!(word != 0);
        assert_eq!(word & (1 << 12), 1 << 12); // mul bit
    }

    #[test]
    fn test_encode_sve_fadd_z() {
        let enc = ArmMCEncoder::new(true);
        // FADD Z0.S, P0/M, Z0.S, Z1.S
        let word = enc.encode_sve_fadd_z(0, 0, 1, 2);
        // 01100101 | size | 0 | ... | Pg | Zn | Zd
        assert_eq!((word >> 24) & 0xFF, 0x65);
    }

    #[test]
    fn test_encode_sve_fdiv_z() {
        let enc = ArmMCEncoder::new(true);
        let word = enc.encode_sve_fdiv_z(0, 0, 1, 2, 2);
        assert!(word != 0);
        assert_eq!(word & (1 << 11), 1 << 11);
    }

    #[test]
    fn test_encode_sve_ld1b_z() {
        let enc = ArmMCEncoder::new(true);
        // LD1B {Z0.B}, P0/Z, [X0, X1]
        let word = enc.encode_sve_ld1b_z(0, 0, 0, 1);
        // 10100101 | 0 | 0 | 0 | 0 | 10 | Xm | Pg | Xn | Zt
        assert_eq!((word >> 24) & 0xFF, 0xA5);
    }

    #[test]
    fn test_encode_sve_st1b_z() {
        let enc = ArmMCEncoder::new(true);
        let word = enc.encode_sve_st1b_z(0, 0, 0, 1);
        // 11100101 | ...
        assert_eq!((word >> 24) & 0xFF, 0xE5);
    }

    #[test]
    fn test_encode_sve_gather_load() {
        let enc = ArmMCEncoder::new(true);
        // LD1W {Z0.S}, P0/Z, [Z1.S, Z2.S]
        let word = enc.encode_sve_ld1w_gather(0, 0, 1, 2);
        assert_eq!((word >> 24) & 0xFF, 0x85);
    }

    #[test]
    fn test_encode_sve_scatter_store() {
        let enc = ArmMCEncoder::new(true);
        // ST1W {Z0.S}, P0, [Z1.S, Z2.S]
        let word = enc.encode_sve_st1w_scatter(0, 0, 1, 2);
        // Should have store bit set
        assert_eq!(word & (1 << 26), 1 << 26);
    }

    #[test]
    fn test_encode_sme_addha() {
        let enc = ArmMCEncoder::new(true);
        // ADDHA ZA.S[0], {Z0.S, Z1.S}, Z2.S
        let word = enc.encode_sme_addha(0, 1, 0, 2);
        // 11000001 | 00 | Zm | ... | Pn | Pm | ... | Zdn
        assert_eq!((word >> 24) & 0xFF, 0xC1);
    }

    #[test]
    fn test_encode_sme_zero_za() {
        let enc = ArmMCEncoder::new(true);
        let word = enc.encode_sme_zero_za();
        assert!(word != 0);
        assert_eq!((word >> 24) & 0xFF, 0xC1);
    }

    #[test]
    fn test_encode_sme_mova_za_to_z() {
        let enc = ArmMCEncoder::new(true);
        // MOVA Z0, ZA.S[0]
        let word = enc.encode_sme_mova_za_to_z(0, 0);
        assert_eq!((word >> 24) & 0xFF, 0xC1);
        // to_za bit should be 0
        assert_eq!((word >> 22) & 1, 0);
    }

    #[test]
    fn test_encode_sme_mova_z_to_za() {
        let enc = ArmMCEncoder::new(true);
        // MOVA ZA.S[0], Z0
        let word = enc.encode_sme_mova_z_to_za(0, 0);
        // to_za bit should be 1
        assert_eq!((word >> 22) & 1, 1);
    }

    #[test]
    fn test_encode_mte_stg() {
        let enc = ArmMCEncoder::new(true);
        // STG X0, [X1, #16]
        let word = enc.encode_mte_stg(0, 1, 16);
        // 11011001 | 0 | 1 | simm9 | 0 | 1 | Xt | simm(2:0) | 1 | Xn
        assert_eq!((word >> 24) & 0xFF, 0xD9);
    }

    #[test]
    fn test_encode_mte_ldg() {
        let enc = ArmMCEncoder::new(true);
        // LDG X0, [X1, #0]
        let word = enc.encode_mte_ldg(0, 1, 0);
        // LDG has bit 23 set
        assert_eq!(word & (1 << 23), 1 << 23);
    }

    #[test]
    fn test_encode_mte_irg() {
        let enc = ArmMCEncoder::new(true);
        // IRG X0, X1, X2
        let word = enc.encode_mte_irg(0, 1, 2);
        assert_eq!((word >> 24) & 0xFF, 0xD9);
    }

    #[test]
    fn test_encode_mte_addg() {
        let enc = ArmMCEncoder::new(true);
        // ADDG X0, X1, #16, #1
        let word = enc.encode_mte_addg(0, 1, 4, 2);
        // 10010001 | 1 | uimm6 | uimm4 | 0 | Xn | Xd
        assert_eq!((word >> 24) & 0xFF, 0x91);
    }

    #[test]
    fn test_encode_mte_subg() {
        let enc = ArmMCEncoder::new(true);
        let addg = enc.encode_mte_addg(0, 1, 0, 0);
        let subg = enc.encode_mte_subg(0, 1, 0, 0);
        // SUBG should differ from ADDG
        assert_ne!(addg, subg);
    }

    #[test]
    fn test_encode_mte_gmi() {
        let enc = ArmMCEncoder::new(true);
        // GMI X0, X1, X2
        let word = enc.encode_mte_gmi(0, 1, 2);
        assert_eq!((word >> 24) & 0xFF, 0x99);
    }

    #[test]
    fn test_sve_predicate_logical_encodings_unique() {
        let enc = ArmMCEncoder::new(true);
        let and = enc.encode_sve_pred_and(0, 0, 0, 0);
        let orr = enc.encode_sve_pred_orr(0, 0, 0, 0);
        let eor = enc.encode_sve_pred_eor(0, 0, 0, 0);
        assert_ne!(and, orr);
        assert_ne!(orr, eor);
        assert_ne!(and, eor);
    }

    #[test]
    fn test_sme2_fmla_multi() {
        let enc = ArmMCEncoder::new(true);
        let word = enc.encode_sme2_fmla_multi(0, 1, 2, 0);
        assert_eq!((word >> 24) & 0xFF, 0xC1);
    }

    #[test]
    fn test_sme2_fmls_multi() {
        let enc = ArmMCEncoder::new(true);
        let fmla = enc.encode_sme2_fmla_multi(0, 1, 2, 0);
        let fmls = enc.encode_sme2_fmls_multi(0, 1, 2, 0);
        assert_ne!(fmla, fmls);
    }
}