rucc-target 0.11.9

//! What each AArch64 machine instruction is in assembly, and what it hands the encoder.
//!
//! Design: `spec/11-asm-objects-debug.md` section 11.1.
//!
//! The same arrangement as `rucc_target::x86_64::written`: an opcode is a list of instructions and
//! each argument of each instruction names the operand it wants by index. What is different is the
//! last step. The x86 description hands its arguments to a writer and to an encoder that each know
//! how to spell them, and this one turns every argument into the [`Value`] the encoder takes with
//! [`fill`], so the object writer hands the values to [`encode`](crate::aarch64::encode) and `-S`
//! hands the same values to [`write`](crate::aarch64::write). What is printed is then what was
//! encoded, and not only what the table says ought to be.
//!
//! # Operand order
//!
//! The order the machine reads them, destination first, which is also the order the operand vector
//! holds them in. So an argument names its operand by index for the other reasons the x86 table
//! gives: an instruction may name one operand twice, a choice names its two sources the other way
//! round, and an instruction in the middle of an opcode may name none of them.
//!
//! # Register 31
//!
//! The register file numbers the stack pointer 31, and an instruction that writes 31 into a
//! register field means the stack pointer in some places and the zero register in others. [`fill`]
//! reads an allocated 31 as the stack pointer, since the zero register is not one the allocator
//! hands out, and the encoder refuses the stack pointer where the machine has no way to say it.

use crate::aarch64::encode::{
    Addr, Arrangement, Cond, Extend, Mode, Offset, Operator, Scalar, Shift, Value, Width,
};
use crate::aarch64::read::system_field;

use Arg::{
    At, Barrier, Base, Disp, Fixed, Fp, GotPage, GotSlot, Imm, Label, Lit, Low, Mem, Near, Page,
    Pop, Push, Reg, Symbol, Thread, Through, TlsPage, TlsSlot, TprelHi, TprelLo, Vector,
};
use Scalar::{D, Q, S};
use Width::{W, X};

/// One argument of one instruction.
#[derive(Debug, Clone, Copy, PartialEq, Eq)]
pub enum Arg {
    /// The general register the operand at that index was given, at that width.
    Reg(u8, Width),
    /// The register in the other file the operand at that index was given, at that size.
    Fp(u8, Scalar),
    /// All sixteen bytes of the register in the other file the operand at that index was given.
    Vector(u8),
    /// The instruction's constant.
    Imm,
    /// A constant that is part of the opcode rather than of the instruction.
    Lit(i64),
    /// A shift that is part of the opcode.
    Shift(Shift, u8),
    /// How the register before this one is widened, as part of the opcode.
    Extend(Extend),
    /// A condition that is part of the opcode.
    Cond(Cond),
    /// A barrier option that is part of the opcode.
    Barrier(u8),
    /// A general register that is part of the opcode rather than one the allocator chose, where
    /// thirty one is the zero register.
    Fixed(u8, Width),
    /// The instruction's addressing mode.
    Mem,
    /// The base register of the instruction's addressing mode, as a register on its own.
    Base,
    /// The constant of the instruction's addressing mode, as a constant on its own.
    Disp,
    /// The page the instruction's symbol is on.
    Page,
    /// The low twelve bits of the instruction's symbol.
    Low,
    /// The page the global offset table slot of the instruction's symbol is on.
    GotPage,
    /// The global offset table slot of the instruction's symbol, reached from the page in the
    /// register the operand at that index was given.
    GotSlot(u8),
    /// The page of the global offset table slot that holds the offset of the instruction's symbol
    /// from the thread pointer.
    TlsPage,
    /// That slot, reached from the page in the register the operand at that index was given.
    TlsSlot(u8),
    /// The high twelve bits of the offset of the instruction's symbol from the thread pointer.
    TprelHi,
    /// The low twelve bits of the same offset.
    TprelLo,
    /// The register that holds the thread pointer.
    Thread,
    /// Sixteen bytes below the stack pointer, moving the stack pointer there first.
    Push,
    /// The stack pointer, moving it sixteen bytes up afterwards.
    Pop,
    /// The instruction's symbol, as a branch or a call reaches it.
    Symbol,
    /// The block a branch goes to.
    Label,
    /// The first register the instruction reads, which is what a call through a register reads.
    Through,
    /// The address in the register the operand at that index was given, with nothing added, which
    /// is the only address an acquiring or exclusive access takes.
    At(u8),
    /// The instruction that many places away in the same opcode, counted in instructions, which is
    /// where a loop inside one opcode branches to. The listing writes it as a distance from `.`,
    /// since an opcode has no label of its own to name.
    Near(i8),
}

/// One machine instruction an opcode is written as.
#[derive(Debug, Clone, Copy, PartialEq, Eq)]
pub struct Written {
    /// The mnemonic, as GNU as reads it.
    pub mnemonic: &'static str,
    /// The arguments, in the order GNU as reads them.
    pub args: &'static [Arg],
}

const fn spell(mnemonic: &'static str, args: &'static [Arg]) -> Written {
    Written { mnemonic, args }
}

static TEXT: &[(&str, &[Written])] = &[
    // Constants. A constant that is not one `mov` writes is a `mov` of one sixteen bit piece and a
    // `movk` for each piece after it.
    ("mov_ri_32", &[spell("mov", &[Reg(0, W), Imm])]),
    ("mov_ri_64", &[spell("mov", &[Reg(0, X), Imm])]),
    ("movk_ri_16_32", &[spell("movk", &[Reg(0, W), Imm, Arg::Shift(Shift::Lsl, 16)])]),
    ("movk_ri_16_64", &[spell("movk", &[Reg(0, X), Imm, Arg::Shift(Shift::Lsl, 16)])]),
    ("movk_ri_32_64", &[spell("movk", &[Reg(0, X), Imm, Arg::Shift(Shift::Lsl, 32)])]),
    ("movk_ri_48_64", &[spell("movk", &[Reg(0, X), Imm, Arg::Shift(Shift::Lsl, 48)])]),
    // Copies.
    ("mov_rr_32", &[spell("mov", &[Reg(0, W), Reg(1, W)])]),
    ("mov_rr_64", &[spell("mov", &[Reg(0, X), Reg(1, X)])]),
    // Arithmetic on two registers.
    ("add_rr_32", &[spell("add", &[Reg(0, W), Reg(1, W), Reg(2, W)])]),
    ("add_rr_64", &[spell("add", &[Reg(0, X), Reg(1, X), Reg(2, X)])]),
    ("sub_rr_32", &[spell("sub", &[Reg(0, W), Reg(1, W), Reg(2, W)])]),
    ("sub_rr_64", &[spell("sub", &[Reg(0, X), Reg(1, X), Reg(2, X)])]),
    ("and_rr_32", &[spell("and", &[Reg(0, W), Reg(1, W), Reg(2, W)])]),
    ("and_rr_64", &[spell("and", &[Reg(0, X), Reg(1, X), Reg(2, X)])]),
    ("orr_rr_32", &[spell("orr", &[Reg(0, W), Reg(1, W), Reg(2, W)])]),
    ("orr_rr_64", &[spell("orr", &[Reg(0, X), Reg(1, X), Reg(2, X)])]),
    ("eor_rr_32", &[spell("eor", &[Reg(0, W), Reg(1, W), Reg(2, W)])]),
    ("eor_rr_64", &[spell("eor", &[Reg(0, X), Reg(1, X), Reg(2, X)])]),
    ("bic_rr_32", &[spell("bic", &[Reg(0, W), Reg(1, W), Reg(2, W)])]),
    ("bic_rr_64", &[spell("bic", &[Reg(0, X), Reg(1, X), Reg(2, X)])]),
    ("orn_rr_32", &[spell("orn", &[Reg(0, W), Reg(1, W), Reg(2, W)])]),
    ("orn_rr_64", &[spell("orn", &[Reg(0, X), Reg(1, X), Reg(2, X)])]),
    ("mul_rr_32", &[spell("mul", &[Reg(0, W), Reg(1, W), Reg(2, W)])]),
    ("mul_rr_64", &[spell("mul", &[Reg(0, X), Reg(1, X), Reg(2, X)])]),
    ("sdiv_rr_32", &[spell("sdiv", &[Reg(0, W), Reg(1, W), Reg(2, W)])]),
    ("sdiv_rr_64", &[spell("sdiv", &[Reg(0, X), Reg(1, X), Reg(2, X)])]),
    ("udiv_rr_32", &[spell("udiv", &[Reg(0, W), Reg(1, W), Reg(2, W)])]),
    ("udiv_rr_64", &[spell("udiv", &[Reg(0, X), Reg(1, X), Reg(2, X)])]),
    ("lsl_rr_32", &[spell("lslv", &[Reg(0, W), Reg(1, W), Reg(2, W)])]),
    ("lsl_rr_64", &[spell("lslv", &[Reg(0, X), Reg(1, X), Reg(2, X)])]),
    ("lsr_rr_32", &[spell("lsrv", &[Reg(0, W), Reg(1, W), Reg(2, W)])]),
    ("lsr_rr_64", &[spell("lsrv", &[Reg(0, X), Reg(1, X), Reg(2, X)])]),
    ("asr_rr_32", &[spell("asrv", &[Reg(0, W), Reg(1, W), Reg(2, W)])]),
    ("asr_rr_64", &[spell("asrv", &[Reg(0, X), Reg(1, X), Reg(2, X)])]),
    ("ror_rr_32", &[spell("rorv", &[Reg(0, W), Reg(1, W), Reg(2, W)])]),
    ("ror_rr_64", &[spell("rorv", &[Reg(0, X), Reg(1, X), Reg(2, X)])]),
    ("smulh_rr_64", &[spell("smulh", &[Reg(0, X), Reg(1, X), Reg(2, X)])]),
    ("umulh_rr_64", &[spell("umulh", &[Reg(0, X), Reg(1, X), Reg(2, X)])]),
    // Arithmetic on a register and a constant. The constant is one the instruction can hold, which
    // for `add` and `sub` is twelve bits and for the three logical operations is a bitmask.
    ("add_ri_32", &[spell("add", &[Reg(0, W), Reg(1, W), Imm])]),
    ("add_ri_64", &[spell("add", &[Reg(0, X), Reg(1, X), Imm])]),
    ("sub_ri_32", &[spell("sub", &[Reg(0, W), Reg(1, W), Imm])]),
    ("sub_ri_64", &[spell("sub", &[Reg(0, X), Reg(1, X), Imm])]),
    ("and_ri_32", &[spell("and", &[Reg(0, W), Reg(1, W), Imm])]),
    ("and_ri_64", &[spell("and", &[Reg(0, X), Reg(1, X), Imm])]),
    ("orr_ri_32", &[spell("orr", &[Reg(0, W), Reg(1, W), Imm])]),
    ("orr_ri_64", &[spell("orr", &[Reg(0, X), Reg(1, X), Imm])]),
    ("eor_ri_32", &[spell("eor", &[Reg(0, W), Reg(1, W), Imm])]),
    ("eor_ri_64", &[spell("eor", &[Reg(0, X), Reg(1, X), Imm])]),
    ("lsl_ri_32", &[spell("lsl", &[Reg(0, W), Reg(1, W), Imm])]),
    ("lsl_ri_64", &[spell("lsl", &[Reg(0, X), Reg(1, X), Imm])]),
    ("lsr_ri_32", &[spell("lsr", &[Reg(0, W), Reg(1, W), Imm])]),
    ("lsr_ri_64", &[spell("lsr", &[Reg(0, X), Reg(1, X), Imm])]),
    ("asr_ri_32", &[spell("asr", &[Reg(0, W), Reg(1, W), Imm])]),
    ("asr_ri_64", &[spell("asr", &[Reg(0, X), Reg(1, X), Imm])]),
    ("ror_ri_32", &[spell("ror", &[Reg(0, W), Reg(1, W), Imm])]),
    ("ror_ri_64", &[spell("ror", &[Reg(0, X), Reg(1, X), Imm])]),
    // Multiply and add, and multiply and subtract, which is how a remainder is taken: the quotient
    // times the divisor, subtracted from the dividend.
    ("madd_rrr_32", &[spell("madd", &[Reg(0, W), Reg(1, W), Reg(2, W), Reg(3, W)])]),
    ("madd_rrr_64", &[spell("madd", &[Reg(0, X), Reg(1, X), Reg(2, X), Reg(3, X)])]),
    ("msub_rrr_32", &[spell("msub", &[Reg(0, W), Reg(1, W), Reg(2, W), Reg(3, W)])]),
    ("msub_rrr_64", &[spell("msub", &[Reg(0, X), Reg(1, X), Reg(2, X), Reg(3, X)])]),
    // One register to one register.
    ("neg_r_32", &[spell("neg", &[Reg(0, W), Reg(1, W)])]),
    ("neg_r_64", &[spell("neg", &[Reg(0, X), Reg(1, X)])]),
    ("mvn_r_32", &[spell("mvn", &[Reg(0, W), Reg(1, W)])]),
    ("mvn_r_64", &[spell("mvn", &[Reg(0, X), Reg(1, X)])]),
    ("clz_r_32", &[spell("clz", &[Reg(0, W), Reg(1, W)])]),
    ("clz_r_64", &[spell("clz", &[Reg(0, X), Reg(1, X)])]),
    ("rbit_r_32", &[spell("rbit", &[Reg(0, W), Reg(1, W)])]),
    ("rbit_r_64", &[spell("rbit", &[Reg(0, X), Reg(1, X)])]),
    ("rev_r_32", &[spell("rev", &[Reg(0, W), Reg(1, W)])]),
    ("rev_r_64", &[spell("rev", &[Reg(0, X), Reg(1, X)])]),
    ("rev16_r_32", &[spell("rev16", &[Reg(0, W), Reg(1, W)])]),
    // Widening. Each reads the low bits of its source and agrees with it about every one of them,
    // which is what `copies_low` asks.
    // A value of eight or sixteen bits lives in a W register, so widening to sixteen bits is the
    // thirty two bit instruction under a head that says how much of it the result is.
    ("sxtb_16", &[spell("sxtb", &[Reg(0, W), Reg(1, W)])]),
    ("sxtb_32", &[spell("sxtb", &[Reg(0, W), Reg(1, W)])]),
    ("sxtb_64", &[spell("sxtb", &[Reg(0, X), Reg(1, W)])]),
    ("sxth_32", &[spell("sxth", &[Reg(0, W), Reg(1, W)])]),
    ("sxth_64", &[spell("sxth", &[Reg(0, X), Reg(1, W)])]),
    ("uxtb_16", &[spell("uxtb", &[Reg(0, W), Reg(1, W)])]),
    ("uxtb_32", &[spell("uxtb", &[Reg(0, W), Reg(1, W)])]),
    // Writing a W register clears the top half of the X register, so the widening to sixty four
    // bits is the thirty two bit instruction under a head that says so.
    ("uxtb_64", &[spell("uxtb", &[Reg(0, W), Reg(1, W)])]),
    ("uxth_32", &[spell("uxth", &[Reg(0, W), Reg(1, W)])]),
    ("uxth_64", &[spell("uxth", &[Reg(0, W), Reg(1, W)])]),
    ("sxtw_64", &[spell("sxtw", &[Reg(0, X), Reg(1, W)])]),
    ("uxtw_64", &[spell("mov", &[Reg(0, W), Reg(1, W)])]),
    ("bit_to_8", &[spell("and", &[Reg(0, W), Reg(1, W), Lit(1)])]),
    ("bit_to_16", &[spell("and", &[Reg(0, W), Reg(1, W), Lit(1)])]),
    ("bit_to_32", &[spell("and", &[Reg(0, W), Reg(1, W), Lit(1)])]),
    ("bit_to_64", &[spell("and", &[Reg(0, W), Reg(1, W), Lit(1)])]),
    ("low_8", &[spell("mov", &[Reg(0, W), Reg(1, W)])]),
    ("low_16", &[spell("mov", &[Reg(0, W), Reg(1, W)])]),
    ("low_32", &[spell("mov", &[Reg(0, W), Reg(1, W)])]),
    ("bit_of_32", &[spell("and", &[Reg(0, W), Reg(1, W), Imm])]),
    ("bit_of_64", &[spell("and", &[Reg(0, W), Reg(1, W), Imm])]),
    // Comparisons that keep nothing but the condition state, which is what a branch on one is
    // folded into.
    ("cmp_rr_32", &[spell("cmp", &[Reg(0, W), Reg(1, W)])]),
    ("cmp_ri_32", &[spell("cmp", &[Reg(0, W), Imm])]),
    ("test_32", &[spell("cmp", &[Reg(0, W), Lit(0)])]),
    ("cmp_rr_64", &[spell("cmp", &[Reg(0, X), Reg(1, X)])]),
    ("cmp_ri_64", &[spell("cmp", &[Reg(0, X), Imm])]),
    ("test_64", &[spell("cmp", &[Reg(0, X), Lit(0)])]),
    // Comparisons that keep the answer, which is `cset` after the comparison.
    (
        "cmp_set_eq_32",
        &[spell("cmp", &[Reg(1, W), Reg(2, W)]), spell("cset", &[Reg(0, W), Arg::Cond(Cond::Eq)])],
    ),
    (
        "cmp_set_eq_64",
        &[spell("cmp", &[Reg(1, X), Reg(2, X)]), spell("cset", &[Reg(0, W), Arg::Cond(Cond::Eq)])],
    ),
    (
        "cmp_set_ne_32",
        &[spell("cmp", &[Reg(1, W), Reg(2, W)]), spell("cset", &[Reg(0, W), Arg::Cond(Cond::Ne)])],
    ),
    (
        "cmp_set_ne_64",
        &[spell("cmp", &[Reg(1, X), Reg(2, X)]), spell("cset", &[Reg(0, W), Arg::Cond(Cond::Ne)])],
    ),
    (
        "cmp_set_lt_32",
        &[spell("cmp", &[Reg(1, W), Reg(2, W)]), spell("cset", &[Reg(0, W), Arg::Cond(Cond::Lt)])],
    ),
    (
        "cmp_set_lt_64",
        &[spell("cmp", &[Reg(1, X), Reg(2, X)]), spell("cset", &[Reg(0, W), Arg::Cond(Cond::Lt)])],
    ),
    (
        "cmp_set_le_32",
        &[spell("cmp", &[Reg(1, W), Reg(2, W)]), spell("cset", &[Reg(0, W), Arg::Cond(Cond::Le)])],
    ),
    (
        "cmp_set_le_64",
        &[spell("cmp", &[Reg(1, X), Reg(2, X)]), spell("cset", &[Reg(0, W), Arg::Cond(Cond::Le)])],
    ),
    (
        "cmp_set_gt_32",
        &[spell("cmp", &[Reg(1, W), Reg(2, W)]), spell("cset", &[Reg(0, W), Arg::Cond(Cond::Gt)])],
    ),
    (
        "cmp_set_gt_64",
        &[spell("cmp", &[Reg(1, X), Reg(2, X)]), spell("cset", &[Reg(0, W), Arg::Cond(Cond::Gt)])],
    ),
    (
        "cmp_set_ge_32",
        &[spell("cmp", &[Reg(1, W), Reg(2, W)]), spell("cset", &[Reg(0, W), Arg::Cond(Cond::Ge)])],
    ),
    (
        "cmp_set_ge_64",
        &[spell("cmp", &[Reg(1, X), Reg(2, X)]), spell("cset", &[Reg(0, W), Arg::Cond(Cond::Ge)])],
    ),
    (
        "cmp_set_lo_32",
        &[spell("cmp", &[Reg(1, W), Reg(2, W)]), spell("cset", &[Reg(0, W), Arg::Cond(Cond::Lo)])],
    ),
    (
        "cmp_set_lo_64",
        &[spell("cmp", &[Reg(1, X), Reg(2, X)]), spell("cset", &[Reg(0, W), Arg::Cond(Cond::Lo)])],
    ),
    (
        "cmp_set_ls_32",
        &[spell("cmp", &[Reg(1, W), Reg(2, W)]), spell("cset", &[Reg(0, W), Arg::Cond(Cond::Ls)])],
    ),
    (
        "cmp_set_ls_64",
        &[spell("cmp", &[Reg(1, X), Reg(2, X)]), spell("cset", &[Reg(0, W), Arg::Cond(Cond::Ls)])],
    ),
    (
        "cmp_set_hi_32",
        &[spell("cmp", &[Reg(1, W), Reg(2, W)]), spell("cset", &[Reg(0, W), Arg::Cond(Cond::Hi)])],
    ),
    (
        "cmp_set_hi_64",
        &[spell("cmp", &[Reg(1, X), Reg(2, X)]), spell("cset", &[Reg(0, W), Arg::Cond(Cond::Hi)])],
    ),
    (
        "cmp_set_hs_32",
        &[spell("cmp", &[Reg(1, W), Reg(2, W)]), spell("cset", &[Reg(0, W), Arg::Cond(Cond::Hs)])],
    ),
    (
        "cmp_set_hs_64",
        &[spell("cmp", &[Reg(1, X), Reg(2, X)]), spell("cset", &[Reg(0, W), Arg::Cond(Cond::Hs)])],
    ),
    (
        "cmp_set_eq_ri_32",
        &[spell("cmp", &[Reg(1, W), Imm]), spell("cset", &[Reg(0, W), Arg::Cond(Cond::Eq)])],
    ),
    (
        "cmp_set_eq_ri_64",
        &[spell("cmp", &[Reg(1, X), Imm]), spell("cset", &[Reg(0, W), Arg::Cond(Cond::Eq)])],
    ),
    (
        "cmp_set_ne_ri_32",
        &[spell("cmp", &[Reg(1, W), Imm]), spell("cset", &[Reg(0, W), Arg::Cond(Cond::Ne)])],
    ),
    (
        "cmp_set_ne_ri_64",
        &[spell("cmp", &[Reg(1, X), Imm]), spell("cset", &[Reg(0, W), Arg::Cond(Cond::Ne)])],
    ),
    (
        "cmp_set_lt_ri_32",
        &[spell("cmp", &[Reg(1, W), Imm]), spell("cset", &[Reg(0, W), Arg::Cond(Cond::Lt)])],
    ),
    (
        "cmp_set_lt_ri_64",
        &[spell("cmp", &[Reg(1, X), Imm]), spell("cset", &[Reg(0, W), Arg::Cond(Cond::Lt)])],
    ),
    (
        "cmp_set_le_ri_32",
        &[spell("cmp", &[Reg(1, W), Imm]), spell("cset", &[Reg(0, W), Arg::Cond(Cond::Le)])],
    ),
    (
        "cmp_set_le_ri_64",
        &[spell("cmp", &[Reg(1, X), Imm]), spell("cset", &[Reg(0, W), Arg::Cond(Cond::Le)])],
    ),
    (
        "cmp_set_gt_ri_32",
        &[spell("cmp", &[Reg(1, W), Imm]), spell("cset", &[Reg(0, W), Arg::Cond(Cond::Gt)])],
    ),
    (
        "cmp_set_gt_ri_64",
        &[spell("cmp", &[Reg(1, X), Imm]), spell("cset", &[Reg(0, W), Arg::Cond(Cond::Gt)])],
    ),
    (
        "cmp_set_ge_ri_32",
        &[spell("cmp", &[Reg(1, W), Imm]), spell("cset", &[Reg(0, W), Arg::Cond(Cond::Ge)])],
    ),
    (
        "cmp_set_ge_ri_64",
        &[spell("cmp", &[Reg(1, X), Imm]), spell("cset", &[Reg(0, W), Arg::Cond(Cond::Ge)])],
    ),
    (
        "cmp_set_lo_ri_32",
        &[spell("cmp", &[Reg(1, W), Imm]), spell("cset", &[Reg(0, W), Arg::Cond(Cond::Lo)])],
    ),
    (
        "cmp_set_lo_ri_64",
        &[spell("cmp", &[Reg(1, X), Imm]), spell("cset", &[Reg(0, W), Arg::Cond(Cond::Lo)])],
    ),
    (
        "cmp_set_ls_ri_32",
        &[spell("cmp", &[Reg(1, W), Imm]), spell("cset", &[Reg(0, W), Arg::Cond(Cond::Ls)])],
    ),
    (
        "cmp_set_ls_ri_64",
        &[spell("cmp", &[Reg(1, X), Imm]), spell("cset", &[Reg(0, W), Arg::Cond(Cond::Ls)])],
    ),
    (
        "cmp_set_hi_ri_32",
        &[spell("cmp", &[Reg(1, W), Imm]), spell("cset", &[Reg(0, W), Arg::Cond(Cond::Hi)])],
    ),
    (
        "cmp_set_hi_ri_64",
        &[spell("cmp", &[Reg(1, X), Imm]), spell("cset", &[Reg(0, W), Arg::Cond(Cond::Hi)])],
    ),
    (
        "cmp_set_hs_ri_32",
        &[spell("cmp", &[Reg(1, W), Imm]), spell("cset", &[Reg(0, W), Arg::Cond(Cond::Hs)])],
    ),
    (
        "cmp_set_hs_ri_64",
        &[spell("cmp", &[Reg(1, X), Imm]), spell("cset", &[Reg(0, W), Arg::Cond(Cond::Hs)])],
    ),
    // A choice on a byte, which is the second source when the byte is not zero and the first when
    // it is, and the same choice straight off a condition some comparison left.
    (
        "sel_32",
        &[
            spell("cmp", &[Reg(3, W), Lit(0)]),
            spell("csel", &[Reg(0, W), Reg(2, W), Reg(1, W), Arg::Cond(Cond::Ne)]),
        ],
    ),
    (
        "sel_64",
        &[
            spell("cmp", &[Reg(3, W), Lit(0)]),
            spell("csel", &[Reg(0, X), Reg(2, X), Reg(1, X), Arg::Cond(Cond::Ne)]),
        ],
    ),
    ("csel_eq_32", &[spell("csel", &[Reg(0, W), Reg(2, W), Reg(1, W), Arg::Cond(Cond::Eq)])]),
    ("csel_eq_64", &[spell("csel", &[Reg(0, X), Reg(2, X), Reg(1, X), Arg::Cond(Cond::Eq)])]),
    ("csel_ne_32", &[spell("csel", &[Reg(0, W), Reg(2, W), Reg(1, W), Arg::Cond(Cond::Ne)])]),
    ("csel_ne_64", &[spell("csel", &[Reg(0, X), Reg(2, X), Reg(1, X), Arg::Cond(Cond::Ne)])]),
    ("csel_lt_32", &[spell("csel", &[Reg(0, W), Reg(2, W), Reg(1, W), Arg::Cond(Cond::Lt)])]),
    ("csel_lt_64", &[spell("csel", &[Reg(0, X), Reg(2, X), Reg(1, X), Arg::Cond(Cond::Lt)])]),
    ("csel_le_32", &[spell("csel", &[Reg(0, W), Reg(2, W), Reg(1, W), Arg::Cond(Cond::Le)])]),
    ("csel_le_64", &[spell("csel", &[Reg(0, X), Reg(2, X), Reg(1, X), Arg::Cond(Cond::Le)])]),
    ("csel_gt_32", &[spell("csel", &[Reg(0, W), Reg(2, W), Reg(1, W), Arg::Cond(Cond::Gt)])]),
    ("csel_gt_64", &[spell("csel", &[Reg(0, X), Reg(2, X), Reg(1, X), Arg::Cond(Cond::Gt)])]),
    ("csel_ge_32", &[spell("csel", &[Reg(0, W), Reg(2, W), Reg(1, W), Arg::Cond(Cond::Ge)])]),
    ("csel_ge_64", &[spell("csel", &[Reg(0, X), Reg(2, X), Reg(1, X), Arg::Cond(Cond::Ge)])]),
    ("csel_lo_32", &[spell("csel", &[Reg(0, W), Reg(2, W), Reg(1, W), Arg::Cond(Cond::Lo)])]),
    ("csel_lo_64", &[spell("csel", &[Reg(0, X), Reg(2, X), Reg(1, X), Arg::Cond(Cond::Lo)])]),
    ("csel_ls_32", &[spell("csel", &[Reg(0, W), Reg(2, W), Reg(1, W), Arg::Cond(Cond::Ls)])]),
    ("csel_ls_64", &[spell("csel", &[Reg(0, X), Reg(2, X), Reg(1, X), Arg::Cond(Cond::Ls)])]),
    ("csel_hi_32", &[spell("csel", &[Reg(0, W), Reg(2, W), Reg(1, W), Arg::Cond(Cond::Hi)])]),
    ("csel_hi_64", &[spell("csel", &[Reg(0, X), Reg(2, X), Reg(1, X), Arg::Cond(Cond::Hi)])]),
    ("csel_hs_32", &[spell("csel", &[Reg(0, W), Reg(2, W), Reg(1, W), Arg::Cond(Cond::Hs)])]),
    ("csel_hs_64", &[spell("csel", &[Reg(0, X), Reg(2, X), Reg(1, X), Arg::Cond(Cond::Hs)])]),
    ("cset_eq", &[spell("cset", &[Reg(0, W), Arg::Cond(Cond::Eq)])]),
    ("cset_ne", &[spell("cset", &[Reg(0, W), Arg::Cond(Cond::Ne)])]),
    ("cset_lt", &[spell("cset", &[Reg(0, W), Arg::Cond(Cond::Lt)])]),
    ("cset_le", &[spell("cset", &[Reg(0, W), Arg::Cond(Cond::Le)])]),
    ("cset_gt", &[spell("cset", &[Reg(0, W), Arg::Cond(Cond::Gt)])]),
    ("cset_ge", &[spell("cset", &[Reg(0, W), Arg::Cond(Cond::Ge)])]),
    ("cset_lo", &[spell("cset", &[Reg(0, W), Arg::Cond(Cond::Lo)])]),
    ("cset_ls", &[spell("cset", &[Reg(0, W), Arg::Cond(Cond::Ls)])]),
    ("cset_hi", &[spell("cset", &[Reg(0, W), Arg::Cond(Cond::Hi)])]),
    ("cset_hs", &[spell("cset", &[Reg(0, W), Arg::Cond(Cond::Hs)])]),
    ("cset_mi", &[spell("cset", &[Reg(0, W), Arg::Cond(Cond::Mi)])]),
    // Addresses. A base and a constant is one `add`, and a symbol is the page it is on and then the
    // offset into that page.
    ("lea_64", &[spell("add", &[Reg(0, X), Base, Disp])]),
    ("addr_64", &[spell("adrp", &[Reg(0, X), Page]), spell("add", &[Reg(0, X), Reg(0, X), Low])]),
    ("adr_64", &[spell("adr", &[Reg(0, X), Symbol])]),
    ("got_64", &[spell("adrp", &[Reg(0, X), GotPage]), spell("ldr", &[Reg(0, X), GotSlot(0)])]),
    // A thread-local variable's offset from the thread pointer sits in a slot of the global offset
    // table the link fills in, and the thread pointer is read on its own. Adding the two is the
    // initial exec model, which works in a shared library as well as in the program.
    (
        "gottprel_64",
        &[spell("adrp", &[Reg(0, X), TlsPage]), spell("ldr", &[Reg(0, X), TlsSlot(0)])],
    ),
    ("thread_64", &[spell("mrs", &[Reg(0, X), Thread])]),
    (
        "tls_64",
        &[
            spell("mrs", &[Reg(0, X), Thread]),
            spell("add", &[Reg(0, X), Reg(0, X), TprelHi, Arg::Shift(Shift::Lsl, 12)]),
            spell("add", &[Reg(0, X), Reg(0, X), TprelLo]),
        ],
    ),
    // Loads, which write a whole register whatever they read.
    ("ldr_8", &[spell("ldrb", &[Reg(0, W), Mem])]),
    ("ldr_16", &[spell("ldrh", &[Reg(0, W), Mem])]),
    ("ldr_32", &[spell("ldr", &[Reg(0, W), Mem])]),
    ("ldr_64", &[spell("ldr", &[Reg(0, X), Mem])]),
    ("ldrs_8_32", &[spell("ldrsb", &[Reg(0, W), Mem])]),
    ("ldrs_8_64", &[spell("ldrsb", &[Reg(0, X), Mem])]),
    ("ldrs_16_32", &[spell("ldrsh", &[Reg(0, W), Mem])]),
    ("ldrs_16_64", &[spell("ldrsh", &[Reg(0, X), Mem])]),
    ("ldrs_32_64", &[spell("ldrsw", &[Reg(0, X), Mem])]),
    // Stores.
    ("str_8", &[spell("strb", &[Reg(0, W), Mem])]),
    ("str_16", &[spell("strh", &[Reg(0, W), Mem])]),
    ("str_32", &[spell("str", &[Reg(0, W), Mem])]),
    ("str_64", &[spell("str", &[Reg(0, X), Mem])]),
    // Loads and stores of the other register file.
    ("ldr_f32", &[spell("ldr", &[Fp(0, S), Mem])]),
    ("ldr_f64", &[spell("ldr", &[Fp(0, D), Mem])]),
    ("ldr_f128", &[spell("ldr", &[Fp(0, Q), Mem])]),
    ("str_f32", &[spell("str", &[Fp(0, S), Mem])]),
    ("str_f64", &[spell("str", &[Fp(0, D), Mem])]),
    ("str_f128", &[spell("str", &[Fp(0, Q), Mem])]),
    // Floating point arithmetic.
    ("fadd_f32", &[spell("fadd", &[Fp(0, S), Fp(1, S), Fp(2, S)])]),
    ("fadd_f64", &[spell("fadd", &[Fp(0, D), Fp(1, D), Fp(2, D)])]),
    ("fsub_f32", &[spell("fsub", &[Fp(0, S), Fp(1, S), Fp(2, S)])]),
    ("fsub_f64", &[spell("fsub", &[Fp(0, D), Fp(1, D), Fp(2, D)])]),
    ("fmul_f32", &[spell("fmul", &[Fp(0, S), Fp(1, S), Fp(2, S)])]),
    ("fmul_f64", &[spell("fmul", &[Fp(0, D), Fp(1, D), Fp(2, D)])]),
    ("fdiv_f32", &[spell("fdiv", &[Fp(0, S), Fp(1, S), Fp(2, S)])]),
    ("fdiv_f64", &[spell("fdiv", &[Fp(0, D), Fp(1, D), Fp(2, D)])]),
    ("fmin_f32", &[spell("fmin", &[Fp(0, S), Fp(1, S), Fp(2, S)])]),
    ("fmin_f64", &[spell("fmin", &[Fp(0, D), Fp(1, D), Fp(2, D)])]),
    ("fmax_f32", &[spell("fmax", &[Fp(0, S), Fp(1, S), Fp(2, S)])]),
    ("fmax_f64", &[spell("fmax", &[Fp(0, D), Fp(1, D), Fp(2, D)])]),
    ("fminnm_f32", &[spell("fminnm", &[Fp(0, S), Fp(1, S), Fp(2, S)])]),
    ("fminnm_f64", &[spell("fminnm", &[Fp(0, D), Fp(1, D), Fp(2, D)])]),
    ("fmaxnm_f32", &[spell("fmaxnm", &[Fp(0, S), Fp(1, S), Fp(2, S)])]),
    ("fmaxnm_f64", &[spell("fmaxnm", &[Fp(0, D), Fp(1, D), Fp(2, D)])]),
    ("fneg_f32", &[spell("fneg", &[Fp(0, S), Fp(1, S)])]),
    ("fneg_f64", &[spell("fneg", &[Fp(0, D), Fp(1, D)])]),
    ("fabs_f32", &[spell("fabs", &[Fp(0, S), Fp(1, S)])]),
    ("fabs_f64", &[spell("fabs", &[Fp(0, D), Fp(1, D)])]),
    ("fsqrt_f32", &[spell("fsqrt", &[Fp(0, S), Fp(1, S)])]),
    ("fsqrt_f64", &[spell("fsqrt", &[Fp(0, D), Fp(1, D)])]),
    ("frintz_f32", &[spell("frintz", &[Fp(0, S), Fp(1, S)])]),
    ("frintz_f64", &[spell("frintz", &[Fp(0, D), Fp(1, D)])]),
    ("frintm_f32", &[spell("frintm", &[Fp(0, S), Fp(1, S)])]),
    ("frintm_f64", &[spell("frintm", &[Fp(0, D), Fp(1, D)])]),
    ("frintp_f32", &[spell("frintp", &[Fp(0, S), Fp(1, S)])]),
    ("frintp_f64", &[spell("frintp", &[Fp(0, D), Fp(1, D)])]),
    ("frinta_f32", &[spell("frinta", &[Fp(0, S), Fp(1, S)])]),
    ("frinta_f64", &[spell("frinta", &[Fp(0, D), Fp(1, D)])]),
    ("frintn_f32", &[spell("frintn", &[Fp(0, S), Fp(1, S)])]),
    ("frintn_f64", &[spell("frintn", &[Fp(0, D), Fp(1, D)])]),
    ("frintx_f32", &[spell("frintx", &[Fp(0, S), Fp(1, S)])]),
    ("frintx_f64", &[spell("frintx", &[Fp(0, D), Fp(1, D)])]),
    ("frinti_f32", &[spell("frinti", &[Fp(0, S), Fp(1, S)])]),
    ("frinti_f64", &[spell("frinti", &[Fp(0, D), Fp(1, D)])]),
    // Copies within the other register file. A 128 bit value is copied as sixteen bytes, since
    // there is no scalar move that wide.
    ("fmov_rr_f32", &[spell("fmov", &[Fp(0, S), Fp(1, S)])]),
    ("fmov_rr_f64", &[spell("fmov", &[Fp(0, D), Fp(1, D)])]),
    ("mov_rr_f128", &[spell("mov", &[Vector(0), Vector(1)])]),
    // Between the two floating point widths.
    ("fcvt_f64_f32", &[spell("fcvt", &[Fp(0, S), Fp(1, D)])]),
    ("fcvt_f32_f64", &[spell("fcvt", &[Fp(0, D), Fp(1, S)])]),
    // From an integer to a float, and back, rounding toward zero the way C does. The name is the
    // source and then the destination.
    ("scvtf_32_f32", &[spell("scvtf", &[Fp(0, S), Reg(1, W)])]),
    ("scvtf_32_f64", &[spell("scvtf", &[Fp(0, D), Reg(1, W)])]),
    ("scvtf_64_f32", &[spell("scvtf", &[Fp(0, S), Reg(1, X)])]),
    ("scvtf_64_f64", &[spell("scvtf", &[Fp(0, D), Reg(1, X)])]),
    ("ucvtf_32_f32", &[spell("ucvtf", &[Fp(0, S), Reg(1, W)])]),
    ("ucvtf_32_f64", &[spell("ucvtf", &[Fp(0, D), Reg(1, W)])]),
    ("ucvtf_64_f32", &[spell("ucvtf", &[Fp(0, S), Reg(1, X)])]),
    ("ucvtf_64_f64", &[spell("ucvtf", &[Fp(0, D), Reg(1, X)])]),
    ("fcvtzs_f32_32", &[spell("fcvtzs", &[Reg(0, W), Fp(1, S)])]),
    ("fcvtzs_f32_64", &[spell("fcvtzs", &[Reg(0, X), Fp(1, S)])]),
    ("fcvtzs_f64_32", &[spell("fcvtzs", &[Reg(0, W), Fp(1, D)])]),
    ("fcvtzs_f64_64", &[spell("fcvtzs", &[Reg(0, X), Fp(1, D)])]),
    ("fcvtzu_f32_32", &[spell("fcvtzu", &[Reg(0, W), Fp(1, S)])]),
    ("fcvtzu_f32_64", &[spell("fcvtzu", &[Reg(0, X), Fp(1, S)])]),
    ("fcvtzu_f64_32", &[spell("fcvtzu", &[Reg(0, W), Fp(1, D)])]),
    ("fcvtzu_f64_64", &[spell("fcvtzu", &[Reg(0, X), Fp(1, D)])]),
    // The bits of a register moved across to the other file unchanged, which is how a float
    // constant is built and how a union is read.
    ("fmov_to_f32", &[spell("fmov", &[Fp(0, S), Reg(1, W)])]),
    ("fmov_to_f64", &[spell("fmov", &[Fp(0, D), Reg(1, X)])]),
    ("fmov_from_f32", &[spell("fmov", &[Reg(0, W), Fp(1, S)])]),
    ("fmov_from_f64", &[spell("fmov", &[Reg(0, X), Fp(1, D)])]),
    // Floating point comparisons. The conditions are chosen so that every ordered one is false when
    // either side is not a number and `ne` is true, which is what C asks of each of them.
    ("fcmp_f32", &[spell("fcmp", &[Fp(0, S), Fp(1, S)])]),
    ("fcmp_f64", &[spell("fcmp", &[Fp(0, D), Fp(1, D)])]),
    (
        "fcmp_set_eq_f32",
        &[spell("fcmp", &[Fp(1, S), Fp(2, S)]), spell("cset", &[Reg(0, W), Arg::Cond(Cond::Eq)])],
    ),
    (
        "fcmp_set_eq_f64",
        &[spell("fcmp", &[Fp(1, D), Fp(2, D)]), spell("cset", &[Reg(0, W), Arg::Cond(Cond::Eq)])],
    ),
    (
        "fcmp_set_ne_f32",
        &[spell("fcmp", &[Fp(1, S), Fp(2, S)]), spell("cset", &[Reg(0, W), Arg::Cond(Cond::Ne)])],
    ),
    (
        "fcmp_set_ne_f64",
        &[spell("fcmp", &[Fp(1, D), Fp(2, D)]), spell("cset", &[Reg(0, W), Arg::Cond(Cond::Ne)])],
    ),
    (
        "fcmp_set_lt_f32",
        &[spell("fcmp", &[Fp(1, S), Fp(2, S)]), spell("cset", &[Reg(0, W), Arg::Cond(Cond::Mi)])],
    ),
    (
        "fcmp_set_lt_f64",
        &[spell("fcmp", &[Fp(1, D), Fp(2, D)]), spell("cset", &[Reg(0, W), Arg::Cond(Cond::Mi)])],
    ),
    (
        "fcmp_set_le_f32",
        &[spell("fcmp", &[Fp(1, S), Fp(2, S)]), spell("cset", &[Reg(0, W), Arg::Cond(Cond::Ls)])],
    ),
    (
        "fcmp_set_le_f64",
        &[spell("fcmp", &[Fp(1, D), Fp(2, D)]), spell("cset", &[Reg(0, W), Arg::Cond(Cond::Ls)])],
    ),
    (
        "fcmp_set_gt_f32",
        &[spell("fcmp", &[Fp(1, S), Fp(2, S)]), spell("cset", &[Reg(0, W), Arg::Cond(Cond::Gt)])],
    ),
    (
        "fcmp_set_gt_f64",
        &[spell("fcmp", &[Fp(1, D), Fp(2, D)]), spell("cset", &[Reg(0, W), Arg::Cond(Cond::Gt)])],
    ),
    (
        "fcmp_set_ge_f32",
        &[spell("fcmp", &[Fp(1, S), Fp(2, S)]), spell("cset", &[Reg(0, W), Arg::Cond(Cond::Ge)])],
    ),
    (
        "fcmp_set_ge_f64",
        &[spell("fcmp", &[Fp(1, D), Fp(2, D)]), spell("cset", &[Reg(0, W), Arg::Cond(Cond::Ge)])],
    ),
    // After `fcmp` a NaN on either side sets C and V and clears N and Z, so `lt`, `le`, `hi` and
    // `pl` are the unordered forms of the four orderings, `vs` is unordered and `vc` is ordered.
    // Ordered and not equal is `mi` or `gt`, and unordered or equal is `eq` or `vs`, and each of
    // those is a `cset` of one and a `csinc` that makes it one when the other held, as gcc does.
    (
        "fcmp_set_ult_f32",
        &[spell("fcmp", &[Fp(1, S), Fp(2, S)]), spell("cset", &[Reg(0, W), Arg::Cond(Cond::Lt)])],
    ),
    (
        "fcmp_set_ult_f64",
        &[spell("fcmp", &[Fp(1, D), Fp(2, D)]), spell("cset", &[Reg(0, W), Arg::Cond(Cond::Lt)])],
    ),
    (
        "fcmp_set_ule_f32",
        &[spell("fcmp", &[Fp(1, S), Fp(2, S)]), spell("cset", &[Reg(0, W), Arg::Cond(Cond::Le)])],
    ),
    (
        "fcmp_set_ule_f64",
        &[spell("fcmp", &[Fp(1, D), Fp(2, D)]), spell("cset", &[Reg(0, W), Arg::Cond(Cond::Le)])],
    ),
    (
        "fcmp_set_ugt_f32",
        &[spell("fcmp", &[Fp(1, S), Fp(2, S)]), spell("cset", &[Reg(0, W), Arg::Cond(Cond::Hi)])],
    ),
    (
        "fcmp_set_ugt_f64",
        &[spell("fcmp", &[Fp(1, D), Fp(2, D)]), spell("cset", &[Reg(0, W), Arg::Cond(Cond::Hi)])],
    ),
    (
        "fcmp_set_uge_f32",
        &[spell("fcmp", &[Fp(1, S), Fp(2, S)]), spell("cset", &[Reg(0, W), Arg::Cond(Cond::Pl)])],
    ),
    (
        "fcmp_set_uge_f64",
        &[spell("fcmp", &[Fp(1, D), Fp(2, D)]), spell("cset", &[Reg(0, W), Arg::Cond(Cond::Pl)])],
    ),
    (
        "fcmp_set_uno_f32",
        &[spell("fcmp", &[Fp(1, S), Fp(2, S)]), spell("cset", &[Reg(0, W), Arg::Cond(Cond::Vs)])],
    ),
    (
        "fcmp_set_uno_f64",
        &[spell("fcmp", &[Fp(1, D), Fp(2, D)]), spell("cset", &[Reg(0, W), Arg::Cond(Cond::Vs)])],
    ),
    (
        "fcmp_set_ord_f32",
        &[spell("fcmp", &[Fp(1, S), Fp(2, S)]), spell("cset", &[Reg(0, W), Arg::Cond(Cond::Vc)])],
    ),
    (
        "fcmp_set_ord_f64",
        &[spell("fcmp", &[Fp(1, D), Fp(2, D)]), spell("cset", &[Reg(0, W), Arg::Cond(Cond::Vc)])],
    ),
    (
        "fcmp_set_one_f32",
        &[
            spell("fcmp", &[Fp(1, S), Fp(2, S)]),
            spell("cset", &[Reg(0, W), Arg::Cond(Cond::Mi)]),
            spell("csinc", &[Reg(0, W), Reg(0, W), Fixed(31, W), Arg::Cond(Cond::Le)]),
        ],
    ),
    (
        "fcmp_set_one_f64",
        &[
            spell("fcmp", &[Fp(1, D), Fp(2, D)]),
            spell("cset", &[Reg(0, W), Arg::Cond(Cond::Mi)]),
            spell("csinc", &[Reg(0, W), Reg(0, W), Fixed(31, W), Arg::Cond(Cond::Le)]),
        ],
    ),
    (
        "fcmp_set_ueq_f32",
        &[
            spell("fcmp", &[Fp(1, S), Fp(2, S)]),
            spell("cset", &[Reg(0, W), Arg::Cond(Cond::Eq)]),
            spell("csinc", &[Reg(0, W), Reg(0, W), Fixed(31, W), Arg::Cond(Cond::Vc)]),
        ],
    ),
    (
        "fcmp_set_ueq_f64",
        &[
            spell("fcmp", &[Fp(1, D), Fp(2, D)]),
            spell("cset", &[Reg(0, W), Arg::Cond(Cond::Eq)]),
            spell("csinc", &[Reg(0, W), Reg(0, W), Fixed(31, W), Arg::Cond(Cond::Vc)]),
        ],
    ),
    // Values arriving and leaving in the registers the convention names. None is an instruction,
    // and each is here so the allocator knows which register the value has to be in.
    ("ret_val_32", &[]),
    ("ret_val_64", &[]),
    ("ret_val_f32", &[]),
    ("ret_val_f64", &[]),
    ("ret_val_f128", &[]),
    ("ret_val2_32", &[]),
    ("ret_val2_64", &[]),
    ("ret_val2_f32", &[]),
    ("ret_val2_f64", &[]),
    ("ret_val2_f128", &[]),
    ("ret_val3_f32", &[]),
    ("ret_val3_f64", &[]),
    ("ret_val3_f128", &[]),
    ("ret_val4_f32", &[]),
    ("ret_val4_f64", &[]),
    ("ret_val4_f128", &[]),
    ("arg_val_32", &[]),
    ("arg_val_64", &[]),
    ("arg_val_f32", &[]),
    ("arg_val_f64", &[]),
    ("arg_val_f128", &[]),
    ("br_cond_32", &[]),
    // Branches, calls and the return.
    ("b", &[spell("b", &[Label])]),
    ("b_eq", &[spell("b.eq", &[Label])]),
    ("b_ne", &[spell("b.ne", &[Label])]),
    ("b_lt", &[spell("b.lt", &[Label])]),
    ("b_le", &[spell("b.le", &[Label])]),
    ("b_gt", &[spell("b.gt", &[Label])]),
    ("b_ge", &[spell("b.ge", &[Label])]),
    ("b_lo", &[spell("b.lo", &[Label])]),
    ("b_ls", &[spell("b.ls", &[Label])]),
    ("b_hi", &[spell("b.hi", &[Label])]),
    ("b_hs", &[spell("b.hs", &[Label])]),
    ("b_mi", &[spell("b.mi", &[Label])]),
    ("b_pl", &[spell("b.pl", &[Label])]),
    ("b_vs", &[spell("b.vs", &[Label])]),
    ("b_vc", &[spell("b.vc", &[Label])]),
    ("b_away", &[spell("b", &[Symbol])]),
    ("br", &[spell("br", &[Reg(0, X)])]),
    ("bl", &[spell("bl", &[Symbol])]),
    ("blr", &[spell("blr", &[Through])]),
    ("ret", &[spell("ret", &[])]),
    // A register saved and restored sixteen bytes at a time, since the stack pointer has to stay a
    // multiple of sixteen whenever it is used to reach memory.
    ("push_64", &[spell("str", &[Reg(0, X), Push])]),
    ("pop_64", &[spell("ldr", &[Reg(0, X), Pop])]),
    ("push_pair_64", &[spell("stp", &[Reg(0, X), Reg(1, X), Push])]),
    ("pop_pair_64", &[spell("ldp", &[Reg(0, X), Reg(1, X), Pop])]),
    (
        "align_sp_64",
        &[spell("mov", &[Fixed(16, X), Reg(1, X)]), spell("and", &[Reg(0, X), Fixed(16, X), Imm])],
    ),
    ("probe_64", &[spell("str", &[Fixed(31, X), Mem])]),
    // Everything else.
    ("nop", &[spell("nop", &[])]),
    ("trap", &[spell("brk", &[Lit(1)])]),
    ("fence", &[spell("dmb", &[Barrier(0b1011)])]),
    ("fence_acquire", &[spell("dmb", &[Barrier(0b1001)])]),
    // The ordered accesses, and the read modify writes as loops of an exclusive load and an
    // exclusive store that go round again when something else wrote the address in between. Both
    // halves of every loop acquire and release, which is right at every ordering and is what gcc
    // 16.2.0 writes for the strongest one. The narrow widths compare the expected value as the
    // byte or the half the load zero extended, since what is above it in the register means
    // nothing.
    ("ldar_8", &[spell("ldarb", &[Reg(0, W), At(1)])]),
    ("ldar_16", &[spell("ldarh", &[Reg(0, W), At(1)])]),
    ("ldar_32", &[spell("ldar", &[Reg(0, W), At(1)])]),
    ("ldar_64", &[spell("ldar", &[Reg(0, X), At(1)])]),
    ("stlr_8", &[spell("stlrb", &[Reg(0, W), At(1)])]),
    ("stlr_16", &[spell("stlrh", &[Reg(0, W), At(1)])]),
    ("stlr_32", &[spell("stlr", &[Reg(0, W), At(1)])]),
    ("stlr_64", &[spell("stlr", &[Reg(0, X), At(1)])]),
    (
        "xchg_8",
        &[
            spell("ldaxrb", &[Reg(0, W), At(2)]),
            spell("stlxrb", &[Reg(1, W), Reg(3, W), At(2)]),
            spell("cbnz", &[Reg(1, W), Near(-2)]),
        ],
    ),
    (
        "xchg_16",
        &[
            spell("ldaxrh", &[Reg(0, W), At(2)]),
            spell("stlxrh", &[Reg(1, W), Reg(3, W), At(2)]),
            spell("cbnz", &[Reg(1, W), Near(-2)]),
        ],
    ),
    (
        "xchg_32",
        &[
            spell("ldaxr", &[Reg(0, W), At(2)]),
            spell("stlxr", &[Reg(1, W), Reg(3, W), At(2)]),
            spell("cbnz", &[Reg(1, W), Near(-2)]),
        ],
    ),
    (
        "xchg_64",
        &[
            spell("ldaxr", &[Reg(0, X), At(2)]),
            spell("stlxr", &[Reg(1, W), Reg(3, X), At(2)]),
            spell("cbnz", &[Reg(1, W), Near(-2)]),
        ],
    ),
    (
        "xadd_8",
        &[
            spell("ldaxrb", &[Reg(0, W), At(3)]),
            spell("add", &[Reg(1, W), Reg(0, W), Reg(4, W)]),
            spell("stlxrb", &[Reg(2, W), Reg(1, W), At(3)]),
            spell("cbnz", &[Reg(2, W), Near(-3)]),
        ],
    ),
    (
        "xadd_16",
        &[
            spell("ldaxrh", &[Reg(0, W), At(3)]),
            spell("add", &[Reg(1, W), Reg(0, W), Reg(4, W)]),
            spell("stlxrh", &[Reg(2, W), Reg(1, W), At(3)]),
            spell("cbnz", &[Reg(2, W), Near(-3)]),
        ],
    ),
    (
        "xadd_32",
        &[
            spell("ldaxr", &[Reg(0, W), At(3)]),
            spell("add", &[Reg(1, W), Reg(0, W), Reg(4, W)]),
            spell("stlxr", &[Reg(2, W), Reg(1, W), At(3)]),
            spell("cbnz", &[Reg(2, W), Near(-3)]),
        ],
    ),
    (
        "xadd_64",
        &[
            spell("ldaxr", &[Reg(0, X), At(3)]),
            spell("add", &[Reg(1, X), Reg(0, X), Reg(4, X)]),
            spell("stlxr", &[Reg(2, W), Reg(1, X), At(3)]),
            spell("cbnz", &[Reg(2, W), Near(-3)]),
        ],
    ),
    (
        "xsub_8",
        &[
            spell("ldaxrb", &[Reg(0, W), At(3)]),
            spell("sub", &[Reg(1, W), Reg(0, W), Reg(4, W)]),
            spell("stlxrb", &[Reg(2, W), Reg(1, W), At(3)]),
            spell("cbnz", &[Reg(2, W), Near(-3)]),
        ],
    ),
    (
        "xsub_16",
        &[
            spell("ldaxrh", &[Reg(0, W), At(3)]),
            spell("sub", &[Reg(1, W), Reg(0, W), Reg(4, W)]),
            spell("stlxrh", &[Reg(2, W), Reg(1, W), At(3)]),
            spell("cbnz", &[Reg(2, W), Near(-3)]),
        ],
    ),
    (
        "xsub_32",
        &[
            spell("ldaxr", &[Reg(0, W), At(3)]),
            spell("sub", &[Reg(1, W), Reg(0, W), Reg(4, W)]),
            spell("stlxr", &[Reg(2, W), Reg(1, W), At(3)]),
            spell("cbnz", &[Reg(2, W), Near(-3)]),
        ],
    ),
    (
        "xsub_64",
        &[
            spell("ldaxr", &[Reg(0, X), At(3)]),
            spell("sub", &[Reg(1, X), Reg(0, X), Reg(4, X)]),
            spell("stlxr", &[Reg(2, W), Reg(1, X), At(3)]),
            spell("cbnz", &[Reg(2, W), Near(-3)]),
        ],
    ),
    (
        "cmpxchg_8",
        &[
            spell("ldaxrb", &[Reg(0, W), At(2)]),
            spell("cmp", &[Reg(0, W), Reg(3, W), Arg::Extend(Extend::Uxtb)]),
            spell("b.ne", &[Near(3)]),
            spell("stlxrb", &[Reg(1, W), Reg(4, W), At(2)]),
            spell("cbnz", &[Reg(1, W), Near(-4)]),
            spell("cset", &[Reg(1, W), Arg::Cond(Cond::Eq)]),
        ],
    ),
    (
        "cmpxchg_16",
        &[
            spell("ldaxrh", &[Reg(0, W), At(2)]),
            spell("cmp", &[Reg(0, W), Reg(3, W), Arg::Extend(Extend::Uxth)]),
            spell("b.ne", &[Near(3)]),
            spell("stlxrh", &[Reg(1, W), Reg(4, W), At(2)]),
            spell("cbnz", &[Reg(1, W), Near(-4)]),
            spell("cset", &[Reg(1, W), Arg::Cond(Cond::Eq)]),
        ],
    ),
    (
        "cmpxchg_32",
        &[
            spell("ldaxr", &[Reg(0, W), At(2)]),
            spell("cmp", &[Reg(0, W), Reg(3, W)]),
            spell("b.ne", &[Near(3)]),
            spell("stlxr", &[Reg(1, W), Reg(4, W), At(2)]),
            spell("cbnz", &[Reg(1, W), Near(-4)]),
            spell("cset", &[Reg(1, W), Arg::Cond(Cond::Eq)]),
        ],
    ),
    (
        "cmpxchg_64",
        &[
            spell("ldaxr", &[Reg(0, X), At(2)]),
            spell("cmp", &[Reg(0, X), Reg(3, X)]),
            spell("b.ne", &[Near(3)]),
            spell("stlxr", &[Reg(1, W), Reg(4, X), At(2)]),
            spell("cbnz", &[Reg(1, W), Near(-4)]),
            spell("cset", &[Reg(1, W), Arg::Cond(Cond::Eq)]),
        ],
    ),
    // Nothing from the table. The text it carries is the program's and the writer puts it down.
    ("template", &[]),
];

/// The instructions the opcode of that name is written as.
///
/// `None` for a name this target does not have, and an empty slice for one that is an opcode and
/// not an instruction, which are two different answers.
#[must_use]
pub fn written(name: &str) -> Option<&'static [Written]> {
    TEXT.iter().find(|(known, _)| *known == name).map(|&(_, insts)| insts)
}

/// How many bits of the operand at that index the opcode of that name names, or `None` for all
/// of it.
///
/// `None` for an operand no instruction of the opcode names as a general register, which is one
/// in the other file, one inside an addressing mode, one of an opcode that is written as nothing,
/// and every operand of a name this target does not have. The widest where one operand is named
/// more than once, since that is how much of it something reads.
///
/// This is [`crate::BitInsts::width`] for this target, taken from the table the encoder and the
/// listing both read, as the x86 one is.
#[must_use]
pub fn operand_width(name: &str, operand: u8) -> Option<u32> {
    let mut found: Option<u32> = None;
    for inst in written(name)? {
        for arg in inst.args {
            let bits = match *arg {
                Reg(at, width) if at == operand => width.bits(),
                GotSlot(at) | TlsSlot(at) | At(at) if at == operand => 64,
                Fp(at, _) | Vector(at) if at == operand => return None,
                _ => continue,
            };
            found = Some(found.map_or(bits, |had: u32| had.max(bits)));
        }
    }
    found
}

/// What an allocated instruction has that its arguments are filled in from.
#[derive(Debug, Clone, Copy, PartialEq)]
pub struct Operands<'a> {
    /// The number of the register each operand was given, in the operand's own class.
    pub regs: &'a [u8],
    /// Where the operands the instruction reads start, which is after every one it writes.
    pub reads: usize,
    /// The instruction's constant, or zero for one that has none.
    pub imm: i64,
    /// The instruction's addressing mode, with its registers already allocated.
    pub mem: Option<Addr>,
}

/// Why an argument could not be filled in.
#[derive(Debug, Clone, Copy, PartialEq, Eq)]
pub enum Missing {
    /// The argument names an operand the instruction does not have.
    Operand(u8),
    /// The argument is part of an addressing mode and the instruction has none.
    Mem,
}

/// The fifteen bits `mrs` carries for `tpidr_el0`, which is where the thread pointer is.
const TPIDR_EL0: u16 = system_field([3, 3, 13, 0, 2]);

/// One argument as the value [`encode`](crate::aarch64::encode) and
/// [`write`](crate::aarch64::write) both take.
///
/// A symbol or a block comes back as [`Value::Symbol`] and which one it is, and where it is, is
/// the caller's to keep: the encoder answers with the relocation that fills it in.
///
/// # Errors
///
/// When the argument names an operand or an addressing mode the instruction does not have, which
/// is a table that disagrees with the form it is written against.
pub fn fill(arg: Arg, with: &Operands<'_>) -> Result<Value, Missing> {
    let reg = |at: u8| with.regs.get(usize::from(at)).copied().ok_or(Missing::Operand(at));
    let gpr = |width: Width, number: u8| {
        if number == 31 { Value::Sp(width) } else { Value::Gpr(width, number) }
    };
    Ok(match arg {
        Reg(at, width) => gpr(width, reg(at)?),
        Fp(at, scalar) => Value::Fp(scalar, reg(at)?),
        Vector(at) => Value::Vector(Arrangement::B16, reg(at)?),
        Imm => Value::Imm(with.imm),
        Lit(imm) => Value::Imm(imm),
        Arg::Shift(shift, amount) => Value::Shift(shift, amount),
        Arg::Extend(extend) => Value::Extend(extend, None),
        Arg::Cond(cond) => Value::Cond(cond),
        Barrier(option) => Value::Barrier(option),
        Fixed(number, width) => Value::Gpr(width, number),
        Mem => Value::Mem(with.mem.ok_or(Missing::Mem)?),
        Base => gpr(X, with.mem.ok_or(Missing::Mem)?.base),
        Disp => match with.mem.ok_or(Missing::Mem)?.offset {
            Offset::Imm(imm) => Value::Imm(imm),
            _ => return Err(Missing::Mem),
        },
        Page | Symbol | Label | Near(_) => Value::Symbol(Operator::Plain),
        At(at) => Value::Mem(Addr { base: reg(at)?, offset: Offset::Imm(0), mode: Mode::Offset }),
        Low => Value::Symbol(Operator::Lo12),
        GotPage => Value::Symbol(Operator::Got),
        GotSlot(at) => Value::Mem(Addr {
            base: reg(at)?,
            offset: Offset::Symbol(Operator::GotLo12),
            mode: Mode::Offset,
        }),
        TlsPage => Value::Symbol(Operator::GotTprel),
        TlsSlot(at) => Value::Mem(Addr {
            base: reg(at)?,
            offset: Offset::Symbol(Operator::GotTprelLo12),
            mode: Mode::Offset,
        }),
        TprelHi => Value::Symbol(Operator::TprelHi12),
        TprelLo => Value::Symbol(Operator::TprelLo12Nc),
        Thread => Value::System(TPIDR_EL0),
        Push => Value::Mem(Addr { base: 31, offset: Offset::Imm(-16), mode: Mode::Pre }),
        Pop => Value::Mem(Addr { base: 31, offset: Offset::Imm(16), mode: Mode::Post }),
        Through => {
            let at = u8::try_from(with.reads).map_err(|_| Missing::Operand(u8::MAX))?;
            gpr(X, reg(at)?)
        }
    })
}

#[cfg(test)]
mod tests {
    use super::*;
    use crate::aarch64::insts::{Form, INSTS, form};
    use crate::aarch64::{FPR, GPR, Spelling, encode, write};
    use crate::operand::Role;

    #[test]
    fn every_opcode_is_described_and_written_and_nothing_else_is() {
        for &(name, _) in INSTS {
            assert!(written(name).is_some(), "{name} has a form and is not written");
        }
        for &(name, _) in TEXT {
            assert!(form(name).is_some(), "{name} is written and has no form");
        }
    }

    #[test]
    fn every_argument_names_an_operand_of_the_file_it_is_written_as() {
        for &(name, form) in INSTS {
            let operands = form.operands();
            for inst in written(name).unwrap() {
                for arg in inst.args {
                    let (at, class) = match *arg {
                        Reg(at, _) | GotSlot(at) | TlsSlot(at) | At(at) => (at, GPR),
                        Fp(at, _) | Vector(at) => (at, FPR),
                        Mem | Base | Disp => {
                            assert!(form.takes_mem(), "{name} names an address it has not got");
                            continue;
                        }
                        Imm => {
                            assert!(form.takes_imm(), "{name} names a constant it has not got");
                            continue;
                        }
                        _ => continue,
                    };
                    let operand = operands.get(usize::from(at));
                    assert_eq!(operand.map(|op| op.class), Some(class), "{name} operand {at}");
                }
            }
        }
    }

    /// The operands an instruction of that form is filled in from in the test below: every
    /// register a different one and none of them the stack pointer, and a constant and an address
    /// every instruction here can hold.
    ///
    /// Five registers whatever the form says, because a call's operands are the arguments it
    /// passes and those are the lowering's to add rather than the form's, and five is what the
    /// longest of the atomic loops names.
    fn sample(form: Form, regs: &mut Vec<u8>) -> Operands<'_> {
        regs.clear();
        regs.extend(1..=5);
        let reads = form.operands().iter().take_while(|op| op.role != Role::Use).count();
        let mem = Addr { base: 9, offset: Offset::Imm(16), mode: Mode::Offset };
        Operands { regs, reads, imm: 1, mem: Some(mem) }
    }

    #[test]
    fn every_instruction_every_opcode_is_written_as_is_one_the_encoder_writes() {
        let mut wrong = Vec::new();
        let mut regs = Vec::new();
        for &(name, form) in INSTS {
            let with = sample(form, &mut regs);
            for inst in written(name).unwrap() {
                let values: Vec<Value> =
                    inst.args.iter().map(|&arg| fill(arg, &with).unwrap()).collect();
                if let Err(e) = encode(inst.mnemonic, &values) {
                    wrong.push(format!(
                        "{name}: {}: {e}",
                        write(inst.mnemonic, &values, Some("s"), Spelling::Gnu)
                    ));
                }
            }
        }
        assert!(wrong.is_empty(), "{} do not encode:\n{}", wrong.len(), wrong.join("\n"));
    }

    #[test]
    fn every_instruction_is_the_word_gnu_as_writes_for_the_line_it_is_listed_as() {
        let mut want = include_str!("listing.txt").lines().filter(|line| !line.starts_with('#'));
        let mut wrong = Vec::new();
        let mut regs = Vec::new();
        for &(name, form) in INSTS {
            let with = sample(form, &mut regs);
            for inst in written(name).unwrap() {
                let values: Vec<Value> =
                    inst.args.iter().map(|&arg| fill(arg, &with).unwrap()).collect();
                let word = encode(inst.mnemonic, &values).unwrap().word;
                let got = format!(
                    "{word:08x}\t{}\t{name}",
                    write(inst.mnemonic, &values, Some("s"), Spelling::Gnu)
                );
                match want.next() {
                    Some(line) if line == got => {}
                    line => wrong.push(format!("{got}, GNU as: {line:?}")),
                }
            }
        }
        assert_eq!(want.next(), None, "listing.txt has lines no opcode is written as");
        assert!(wrong.is_empty(), "{} lines differ:\n{}", wrong.len(), wrong.join("\n"));
    }

    fn listing(name: &str, with: &Operands<'_>) -> Vec<String> {
        written(name)
            .unwrap()
            .iter()
            .map(|inst| {
                let values: Vec<Value> =
                    inst.args.iter().map(|&arg| fill(arg, with).unwrap()).collect();
                write(inst.mnemonic, &values, Some("s"), Spelling::Gnu)
            })
            .collect()
    }

    #[test]
    fn an_opcode_is_listed_as_the_instructions_a_reader_expects() {
        let with = Operands { regs: &[0, 1, 2, 3], reads: 1, imm: 42, mem: None };
        assert_eq!(listing("cmp_set_lt_64", &with), ["cmp x1, x2", "cset w0, lt"]);
        assert_eq!(listing("fcmp_set_lt_f64", &with), ["fcmp d1, d2", "cset w0, mi"]);
        assert_eq!(listing("sel_32", &with), ["cmp w3, #0", "csel w0, w2, w1, ne"]);
        assert_eq!(listing("movk_ri_32_64", &with), ["movk x0, #42, lsl #32"]);
        assert_eq!(listing("addr_64", &with), ["adrp x0, s", "add x0, x0, :lo12:s"]);
        assert_eq!(listing("got_64", &with), ["adrp x0, :got:s", "ldr x0, [x0, :got_lo12:s]"]);
        assert_eq!(
            listing("gottprel_64", &with),
            ["adrp x0, :gottprel:s", "ldr x0, [x0, :gottprel_lo12:s]"]
        );
        assert_eq!(listing("thread_64", &with), ["mrs x0, tpidr_el0"]);
        assert_eq!(
            listing("tls_64", &with),
            [
                "mrs x0, tpidr_el0",
                "add x0, x0, :tprel_hi12:s, lsl #12",
                "add x0, x0, :tprel_lo12_nc:s"
            ]
        );
        assert_eq!(listing("blr", &with), ["blr x1"]);
        assert_eq!(listing("push_64", &with), ["str x0, [sp, #-16]!"]);
        assert_eq!(listing("pop_64", &with), ["ldr x0, [sp], #16"]);
        assert_eq!(listing("cset_hs", &with), ["cset w0, hs"]);
        let sp = Addr { base: 31, offset: Offset::Imm(0), mode: Mode::Offset };
        let touched = Operands { regs: &[], reads: 0, imm: 0, mem: Some(sp) };
        assert_eq!(listing("probe_64", &touched), ["str xzr, [sp]"]);
        let aligned = Operands { regs: &[31, 31], reads: 1, imm: -64, mem: None };
        assert_eq!(listing("align_sp_64", &aligned), ["mov x16, sp", "and sp, x16, #-64"]);
        assert!(listing("arg_val_64", &with).is_empty());
    }

    #[test]
    fn the_stack_pointer_is_written_where_the_machine_has_a_way_to_say_it() {
        let sp = Operands {
            regs: &[31, 31, 9],
            reads: 1,
            imm: 32,
            mem: Some(Addr { base: 31, offset: Offset::Imm(24), mode: Mode::Offset }),
        };
        assert_eq!(listing("add_ri_64", &sp), ["add sp, sp, #32"]);
        assert_eq!(listing("sub_ri_64", &sp), ["sub sp, sp, #32"]);
        assert_eq!(listing("mov_rr_64", &sp), ["mov sp, sp"]);
        assert_eq!(listing("lea_64", &sp), ["add sp, sp, #24"]);
        for name in ["add_ri_64", "sub_ri_64", "sub_rr_64", "mov_rr_64", "lea_64"] {
            for inst in written(name).unwrap() {
                let values: Vec<Value> =
                    inst.args.iter().map(|&arg| fill(arg, &sp).unwrap()).collect();
                assert!(encode(inst.mnemonic, &values).is_ok(), "{name} with the stack pointer");
            }
        }
        let load = Operands { regs: &[9], ..sp };
        assert_eq!(listing("ldr_64", &load), ["ldr x9, [sp, #24]"]);
        // The frame record and the way back from the frame pointer to it, which is the one address
        // with a negative distance a prologue or epilogue hands this machine.
        let record = Operands { regs: &[29, 30], ..sp };
        assert_eq!(listing("push_pair_64", &record), ["stp x29, x30, [sp, #-16]!"]);
        assert_eq!(listing("pop_pair_64", &record), ["ldp x29, x30, [sp], #16"]);
        let back = Operands {
            regs: &[31, 29],
            mem: Some(Addr { base: 29, offset: Offset::Imm(-16), mode: Mode::Offset }),
            ..sp
        };
        for name in ["push_pair_64", "pop_pair_64", "lea_64"] {
            let with = if name == "lea_64" { &back } else { &record };
            for inst in written(name).unwrap() {
                let values: Vec<Value> =
                    inst.args.iter().map(|&arg| fill(arg, with).unwrap()).collect();
                assert!(encode(inst.mnemonic, &values).is_ok(), "{name}");
            }
        }
        assert_eq!(listing("str_f64", &load), ["str d9, [sp, #24]"]);
    }

    #[test]
    fn the_thread_pointer_is_tpidr_el0() {
        let line = crate::aarch64::read("mrs x0, tpidr_el0").unwrap();
        assert_eq!(line.values[1], Value::System(TPIDR_EL0));
    }
}