rucc_types/

lib.rs

//! The C type system, interned, and layout computation.
//!
//! Design: `spec/07-types-and-semantics.md`. Layer rank 3, see `spec/18-package-layout.md`.
//!
//! There is one [`Types`] per translation unit and it owns every type in it. A [`TypeId`] is
//! four bytes and two of them are equal exactly when they are the same type, which turns the
//! question the compiler asks more often than any other into an integer comparison.
//!
//! Two ideas shape the rest of it.
//!
//! **Sugar is kept and never decided on.** `typedef int32_t;` gives a node that remembers the
//! name and points at canonical `int`. Every semantic rule reads [`Types::canonical`] and sees
//! `int`; every diagnostic reads the type as it was written and says `int32_t`. Compilers that
//! throw the name away produce messages nobody can act on, and compilers that decide on the
//! name produce wrong answers, and both are common. Sugar is not only at the outermost node,
//! so `int32_t *` and `int32_t[4]` are sugar too and canonicalising rebuilds them.
//!
//! **`_Atomic` is a type, not a qualifier.** `const` and `volatile` and `restrict` are a
//! bitmask in the interning key, because nothing about them changes what an object is. C lets
//! `_Atomic` be written in the same position, but `_Atomic(T)` can have a different alignment
//! from `T`, so it is a type constructor here and the parser is what maps the spelling onto
//! it. Document 01 recorded a compiler that treated it as a qualifier and lost track of it,
//! which is exactly the shortcut that makes atomics silently wrong.
//!
//! Layout comes out of [`TargetInfo`](rucc_target::TargetInfo) and never out of the host.
//! `long` is four bytes on Windows and eight on Linux, and `long double` is eight bytes on
//! Apple and sixteen on SysV x86-64, so a cross compiler that asks its own platform is wrong
//! twice before it has read a line of C.
//!
//! ```
//! use rucc_target::{TargetInfo, Triple};
//! use rucc_types::{IntKind, Types, layout};
//!
//! let mut types = Types::new();
//! let linux = TargetInfo::new("x86_64-unknown-linux-gnu".parse::<Triple>().unwrap());
//! let windows = TargetInfo::new("x86_64-pc-windows-msvc".parse::<Triple>().unwrap());
//!
//! let long = types.int(IntKind::Long);
//! assert_eq!(layout(&types, long, &linux).unwrap().size, 8);
//! assert_eq!(layout(&types, long, &windows).unwrap().size, 4);
//! ```
//!
//! Records are laid out by [`layout_record`], which takes the members and gives back their
//! offsets, and the result is handed to [`Types::complete_record`] so that the record then has
//! a size like any other type. Bit-fields, `packed`, `#pragma pack`, `aligned`, zero width
//! bit-fields and flexible array members are all in there, and every one of their rules was
//! measured against gcc and clang rather than read off a document.
//!
//! [`promote`] and [`usual_arithmetic`] are 6.3.1.1 and 6.3.1.8, the rules that decide what
//! type an arithmetic expression has. Their answers were read out of gcc and clang with
//! `_Generic` naming the type of every interesting pair, which is also how the C23 changes were
//! pinned down: `_BitInt` does not promote, and an enumeration promotes through whatever it is
//! represented in.
//!
//! `__int128` is one of the integer kinds rather than a `_BitInt(128)` in disguise. The two are
//! different types: `__int128` is sixteen bytes aligned to sixteen everywhere, `_BitInt(128)` is
//! aligned to its granule, and `__int128` outranks `long long` where a `_BitInt` is ranked by
//! width alone. It is available on every target here, because all three architectures are
//! 64-bit and GCC has it on every 64-bit target it supports.
//!
//! [`compatible`] and [`composite`] are 6.2.7, the relation that decides whether two
//! declarations of one name are talking about the same thing and the type that is left when they
//! are. Identity is not that relation: `int f(int a[3])` and `int f(int *a)` are different types
//! and the same function. The composite is what a caller merging two declarations should keep,
//! because it is the only one of the three types in play that knows both the array size and the
//! parameter list.
//!
//! # Status
//!
//! The type universe, the interner, the canonical and sugar split, the qualifier rules, layout
//! with records included, the arithmetic conversions, compatibility with the composite type, and
//! [`spell`], which writes a type back as the C declaration it is, are implemented.
//!
//! Not here yet, and named so that the gap is not mistaken for a decision: the decimal floating
//! types.
//!
//! Every crate in the workspace is published, and publishing implies a promise. This one is
//! tier 3: its Rust API is explicitly unstable and will change without a major version bump.
//! Depend on the `rucc` binary's behaviour, not on this.

#![doc(html_root_url = "https://docs.rs/rucc-types/0.10.73")]

mod classify;
mod compat;
mod convert;
mod granule;
mod kind;
mod layout;
mod print;
mod record;
mod types;

pub use crate::classify::{
    element, is_aggregate, is_arithmetic, is_array, is_atomic, is_complete, is_complex,
    is_floating, is_function, is_integer, is_modifiable, is_object, is_pointer, is_real,
    is_real_floating, is_record, is_scalar, is_vector, is_void, lanes, pointee, pointee_as_written,
    real_part,
};
pub use crate::compat::{adjust_parameter, compatible, composite};
pub use crate::convert::{
    mask_of, promote, promote_bit_field, usual_arithmetic, vectors_convertible,
};
pub use crate::granule::{GRANULE, Keying, Tally, measure, measure_all, report as granule_report};
pub use crate::kind::{
    ArrayLen, EnumId, FloatKind, FunctionId, FunctionType, IntKind, Qualifiers, RecordId,
    RecordKind, Type, TypeKind, VlaId,
};
pub use crate::layout::{
    IntegerInfo, Layout, LayoutError, align, float_format, float_width, int_width, integer_info,
    layout,
};
pub use crate::print::{declare, spell};
pub use crate::record::{
    Extent, Field, FieldDecl, RecordError, RecordLayout, RecordOptions, VariableLayout,
    layout_record,
};
pub use crate::types::{EnumInfo, RecordInfo, TypeId, Types};

/// The milestone in `spec/17-milestones.md` that fills this crate in.
pub const MILESTONE: &str = "M2";

#[cfg(test)]
mod tests {
    use std::num::NonZeroU32;

    use rucc_base::{Interner, Symbol};
    use rucc_target::{TargetInfo, Triple};

    use super::*;

    fn target(triple: &str) -> TargetInfo {
        TargetInfo::new(triple.parse::<Triple>().expect("a triple the compiler supports"))
    }

    fn linux() -> TargetInfo {
        target("x86_64-unknown-linux-gnu")
    }

    /// The one target in the table that runs Microsoft's bit-field rule. So does
    /// `x86_64-pc-windows-gnu`, and the tests below say so where it matters, because a rule keyed
    /// on the environment rather than on the operating system would pass every one of them.
    fn windows() -> TargetInfo {
        target("x86_64-pc-windows-msvc")
    }

    /// AAPCS64 proper, where an unnamed bit-field raises the record's alignment. Apple's AArch64
    /// dropped that, so `aarch64-apple-darwin` answers the way x86-64 does and is the pair to
    /// this one wherever the difference is being measured.
    fn aapcs() -> TargetInfo {
        target("aarch64-unknown-linux-gnu")
    }

    /// Lays out a record with no attributes on it, on x86-64 Linux.
    fn lay_out(types: &Types, kind: RecordKind, fields: &[FieldDecl]) -> RecordLayout {
        lay_out_on(&linux(), types, kind, fields)
    }

    /// Lays out a record with no attributes on it, on a named target.
    fn lay_out_on(
        target: &TargetInfo,
        types: &Types,
        kind: RecordKind,
        fields: &[FieldDecl],
    ) -> RecordLayout {
        layout_record(types, kind, fields, &RecordOptions::default(), target)
            .expect("a record every member of which has a layout")
    }

    /// The offsets of the members, in bits, which is what a measurement of a real compiler
    /// gives back once its byte offsets and its bit dumps are put together.
    fn offsets(laid_out: &RecordLayout) -> Vec<u128> {
        laid_out.fields.iter().map(Field::bit_offset).collect()
    }

    /// A complete record type built out of the given members.
    fn record(types: &mut Types, kind: RecordKind, fields: &[FieldDecl]) -> TypeId {
        let id = types.declare_record(kind, None);
        let laid_out = lay_out(types, kind, fields);
        types.complete_record(id, laid_out);
        types.record(id)
    }

    /// An ordinary member of the given type, unnamed, which is all most of these tests need.
    fn member(ty: TypeId) -> FieldDecl {
        FieldDecl::new(None, ty)
    }

    /// A named bit-field, which is what a measurement of a real compiler has to use to be able
    /// to read the field back.
    fn bits(interner: &mut Interner, name: &str, ty: TypeId, width: u32) -> FieldDecl {
        FieldDecl::bit_field(Some(interner.intern(name)), ty, width)
    }

    /// An unnamed bit-field, which occupies bits and raises nothing.
    fn unnamed_bits(ty: TypeId, width: u32) -> FieldDecl {
        FieldDecl::bit_field(None, ty, width)
    }

    #[test]
    fn milestone_is_recorded() {
        assert!(MILESTONE.starts_with('M'));
    }

    #[test]
    fn an_integer_type_answers_with_the_width_of_its_value_and_not_of_its_object() {
        let mut interner = Interner::new();
        let mut types = Types::new();
        let target = linux();

        // A `bool` is one byte and holds one bit, and a `_BitInt(37)` is eight bytes and holds
        // thirty seven. Folding a constant in the size rather than the width gets both wrong.
        let boolean = types.boolean();
        let bits = types.bit_int(true, 37);
        // Through the sugar, the qualifiers and `_Atomic`, none of which is part of a value.
        let short = types.int(IntKind::Short);
        let alias = types.typedef(interner.intern("word"), short);
        let unsigned_char = types.int(IntKind::UChar);
        let atomic = types.atomic(unsigned_char);

        let shape = |ty| integer_info(&types, ty, &target).expect("an integer type");
        assert_eq!(shape(boolean), IntegerInfo::new(false, 1));
        assert_eq!(shape(bits), IntegerInfo::new(true, 37));
        assert_eq!(shape(types.int(IntKind::Int)), IntegerInfo::new(true, 32));
        assert_eq!(shape(types.int(IntKind::ULong)), IntegerInfo::new(false, 64));
        assert_eq!(shape(alias), IntegerInfo::new(true, 16));
        assert_eq!(shape(atomic), IntegerInfo::new(false, 8));

        assert_eq!(integer_info(&types, types.float(FloatKind::Double), &target), None);
    }

    #[test]
    fn an_enumeration_answers_with_the_type_the_enumerators_are_kept_in() {
        let mut interner = Interner::new();
        let mut types = Types::new();
        let target = linux();

        // An enumeration that has not been completed has no underlying type yet, and the answer
        // is that there is no answer rather than a guess at `int` that a later `: long` unsays.
        let colour = types.declare_enum(Some(interner.intern("colour")));
        let ty = types.enumeration(colour);
        assert_eq!(integer_info(&types, ty, &target), None);

        let underlying = types.int(IntKind::ULong);
        types.complete_enum(colour, underlying, true);
        assert_eq!(integer_info(&types, ty, &target), Some(IntegerInfo::new(false, 64)));
    }

    #[test]
    fn a_value_stored_in_an_integer_type_keeps_the_bits_the_type_has_room_for() {
        let char_type = IntegerInfo::new(true, 8);
        assert_eq!(char_type.wrap(300), 44);
        assert!(!char_type.holds(300));
        assert!(char_type.holds(-128));

        assert_eq!(IntegerInfo::new(false, 32).wrap(-1), 4_294_967_295);
        assert_eq!(IntegerInfo::new(false, 8).wrap(-1), 255);

        // Every pattern is a value of a hundred and twenty eight bit type, of either signedness,
        // which is what stops the folding from inventing an overflow at the widest type there is.
        assert!(IntegerInfo::new(false, 128).holds(i128::MIN));
        assert!(IntegerInfo::new(true, 128).holds(i128::MIN));
        assert_eq!(IntegerInfo::new(true, 128).wrap(i128::MAX), i128::MAX);
    }

    #[test]
    fn a_long_double_has_a_format_the_size_does_not_give_away() {
        let target = linux();
        // Sixteen bytes on SysV x86-64 and eighty bits of x87 inside them. A compiler that
        // picked the format by the size would fold every one of those constants too finely.
        assert_eq!(float_width(FloatKind::LongDouble, &target), 128);
        assert_eq!(
            float_format(FloatKind::LongDouble, &target),
            rucc_base::float::Format::X87Extended
        );
        assert_eq!(float_format(FloatKind::Float, &target), rucc_base::float::Format::Single);
    }

    #[test]
    fn an_interchange_type_names_a_format_and_an_extended_one_names_the_target() {
        use rucc_base::float::Format;

        // The four `_FloatN` types are the same format everywhere, which is the point of them,
        // so a program that wants binary128 can say so and get it or get told it cannot.
        for target in [&linux(), &target("aarch64-apple-darwin")] {
            assert_eq!(float_format(FloatKind::Float16, target), Format::Half);
            assert_eq!(float_format(FloatKind::Float32, target), Format::Single);
            assert_eq!(float_format(FloatKind::Float64, target), Format::Double);
            assert_eq!(float_format(FloatKind::Float128, target), Format::Quad);
            assert_eq!(float_width(FloatKind::Float16, target), 16);
            assert_eq!(float_width(FloatKind::Float32, target), 32);
            assert_eq!(float_width(FloatKind::Float64, target), 64);
            assert_eq!(float_width(FloatKind::Float128, target), 128);
            // `_Float32x` is `double` on every target this compiles for.
            assert_eq!(float_format(FloatKind::Float32x, target), Format::Double);
        }

        // `_Float64x` is the one that moves, and it moves with the processor rather than with
        // the operating system, so it stays eighty bits of x87 on x86-64 where `long double`
        // is the same thing and is quad on Apple where `long double` is only a `double`.
        let x86 = linux();
        assert_eq!(float_format(FloatKind::Float64x, &x86), Format::X87Extended);
        assert_eq!(float_format(FloatKind::LongDouble, &x86), Format::X87Extended);
        let mac = target("aarch64-apple-darwin");
        assert_eq!(float_format(FloatKind::Float64x, &mac), Format::Quad);
        assert_eq!(float_format(FloatKind::LongDouble, &mac), Format::Double);
        // Sixteen bytes either way, because the x87 eighty bits are stored padded, which is
        // the same reason `long double` is sixteen bytes on x86-64 and not ten.
        assert_eq!(float_width(FloatKind::Float64x, &x86), 128);
        assert_eq!(float_width(FloatKind::Float64x, &mac), 128);
    }

    #[test]
    fn every_floating_type_is_as_wide_as_the_format_it_is_stored_in() {
        let types = Types::new();
        let sizes = |target: &TargetInfo| -> Vec<(u64, u64)> {
            FloatKind::ALL
                .iter()
                .map(|&kind| {
                    let found = layout(&types, types.float(kind), target).expect("a complete type");
                    (found.size, found.align)
                })
                .collect()
        };
        // Read off gcc 16 with `sizeof` and `_Alignof`, in the order of `FloatKind::ALL`. The
        // two targets differ in one place, which is `long double`, and the eighty bit x87 value
        // that `long double` and `_Float64x` hold on x86-64 takes sixteen bytes to store.
        assert_eq!(
            sizes(&linux()),
            [(2, 2), (4, 4), (4, 4), (8, 8), (8, 8), (8, 8), (16, 16), (16, 16), (16, 16)]
        );
        assert_eq!(
            sizes(&target("aarch64-apple-darwin")),
            [(2, 2), (4, 4), (4, 4), (8, 8), (8, 8), (8, 8), (8, 8), (16, 16), (16, 16)]
        );
    }

    #[test]
    fn every_floating_type_has_a_slot_of_its_own_and_a_name_of_its_own() {
        // Nine types and nine ids, which is what makes `_Float64` and `double` two types that
        // `_Generic` can tell apart rather than one type with two spellings.
        let types = Types::new();
        let mut seen = Vec::new();
        for kind in FloatKind::ALL {
            seen.push(types.float(kind));
        }
        let mut sorted = seen.clone();
        sorted.sort_unstable();
        sorted.dedup();
        assert_eq!(sorted.len(), seen.len(), "two floating types share an id");

        let names: Vec<&str> = FloatKind::ALL.iter().map(|kind| kind.as_str()).collect();
        assert_eq!(
            names,
            [
                "_Float16",
                "float",
                "_Float32",
                "double",
                "_Float32x",
                "_Float64",
                "long double",
                "_Float64x",
                "_Float128",
            ]
        );
    }

    #[test]
    fn the_same_type_asked_for_twice_is_the_same_id() {
        let mut types = Types::new();
        let a = types.pointer(types.int(IntKind::Int));
        let b = types.pointer(types.int(IntKind::Int));
        assert_eq!(a, b, "interning is what makes type identity an integer comparison");
        let c = types.pointer(types.int(IntKind::Long));
        assert_ne!(a, c);
    }

    #[test]
    fn a_qualifier_makes_a_different_type_with_the_same_shape() {
        let mut types = Types::new();
        let int = types.int(IntKind::Int);
        let konst = types.qualified(int, Qualifiers::CONST);
        assert_ne!(int, konst);
        assert_eq!(types.kind(konst), types.kind(int));
        assert!(types.quals(konst).has(Qualifiers::CONST));
        assert_eq!(types.unqualified(konst), int);
    }

    #[test]
    fn qualifiers_accumulate_and_do_not_depend_on_the_order_they_were_written() {
        let mut types = Types::new();
        let int = types.int(IntKind::Int);
        let a = types.qualified(int, Qualifiers::CONST);
        let a = types.qualified(a, Qualifiers::VOLATILE);
        let b = types.qualified(int, Qualifiers::VOLATILE);
        let b = types.qualified(b, Qualifiers::CONST);
        assert_eq!(a, b, "`const volatile int` and `volatile const int` are one type");
    }

    #[test]
    fn qualifying_an_array_qualifies_its_element() {
        // 6.7.3p10, and not a shortcut. An array type has no qualifiers of its own, so if this
        // put the `const` on the array then `const` on an array parameter would mean nothing.
        let mut types = Types::new();
        let int = types.int(IntKind::Int);
        let array = types.array(int, ArrayLen::Fixed(4));
        let konst = types.qualified(array, Qualifiers::CONST);
        assert!(types.quals(konst).is_none(), "the array itself is unqualified");
        let TypeKind::Array { elem, len } = types.kind(konst) else {
            panic!("still an array");
        };
        assert_eq!(len, ArrayLen::Fixed(4));
        assert!(types.quals(elem).has(Qualifiers::CONST));
    }

    #[test]
    fn a_typedef_is_a_different_type_that_means_the_same_thing() {
        let mut interner = Interner::new();
        let mut types = Types::new();
        let int = types.int(IntKind::Int);
        let name = types.typedef(interner.intern("int32_t"), int);
        assert_ne!(name, int, "the sugar survives, so a diagnostic can print it");
        assert_eq!(types.canonical(name), int, "and no rule ever sees it");
        assert!(types.is_sugar(name));
        assert!(!types.is_sugar(int));
    }

    #[test]
    fn sugar_below_the_outermost_node_is_resolved_too() {
        // The bug this is here for: canonicalising only the top node leaves `int32_t *` and
        // `int *` as different types, and then every rule stated on pointers stops firing.
        let mut interner = Interner::new();
        let mut types = Types::new();
        let int = types.int(IntKind::Int);
        let name = types.typedef(interner.intern("int32_t"), int);
        let sugar_pointer = types.pointer(name);
        let plain_pointer = types.pointer(int);
        assert_ne!(sugar_pointer, plain_pointer);
        assert_eq!(types.canonical(sugar_pointer), plain_pointer);

        let sugar_array = types.array(name, ArrayLen::Fixed(3));
        let plain_array = types.array(int, ArrayLen::Fixed(3));
        assert_eq!(types.canonical(sugar_array), plain_array);
    }

    #[test]
    fn a_typedef_of_a_typedef_canonicalises_all_the_way_down() {
        let mut interner = Interner::new();
        let mut types = Types::new();
        let int = types.int(IntKind::Int);
        let mut current = int;
        for i in 0..8 {
            current = types.typedef(interner.intern(&format!("t{i}")), current);
        }
        assert_eq!(types.canonical(current), int);
    }

    #[test]
    fn a_typedef_that_asked_for_an_alignment_says_what_it_is_and_not_what_it_is_at_least() {
        // `__attribute__((aligned(n)))` in this one position replaces the alignment rather than
        // raising it, which is what lets `typedef int L __attribute__((aligned(2)))` really be an
        // `int` at a multiple of two. The size is left where it was, which is gcc's answer and the
        // reason an array of an over aligned typedef is refused rather than padded.
        let mut interner = Interner::new();
        let mut types = Types::new();
        let target = linux();
        let int = types.int(IntKind::Int);
        let low = types.aligned_typedef(interner.intern("L"), int, NonZeroU32::new(2).unwrap());
        let high = types.aligned_typedef(interner.intern("H"), int, NonZeroU32::new(16).unwrap());

        assert_eq!(layout(&types, low, &target), Ok(Layout::new(4, 2)));
        assert_eq!(layout(&types, high, &target), Ok(Layout::new(4, 16)));
        // And the type behind them is what it always was, since the alignment belongs to the name
        // and not to the `int`.
        assert_eq!(layout(&types, int, &target), Ok(Layout::new(4, 4)));

        // Two names for one type that asked for different alignments are two types, which is why
        // the alignment is part of what the table interns them by.
        assert_ne!(low, high);

        // The nearest one wins, because the outer typedef is the one a declaration was written
        // with, and one that asked for nothing keeps whatever the one below it asked for.
        let outer = types.aligned_typedef(interner.intern("M"), low, NonZeroU32::new(8).unwrap());
        assert_eq!(types.align_override(outer), NonZeroU32::new(8));
        let plain = types.typedef(interner.intern("N"), low);
        assert_eq!(types.align_override(plain), NonZeroU32::new(2));
        // Below the sugar there is nothing to find, since only a typedef can carry one of these.
        assert_eq!(types.align_override(int), None);
    }

    #[test]
    fn a_qualified_typedef_keeps_the_name_and_canonicalises_to_the_qualified_type() {
        let mut interner = Interner::new();
        let mut types = Types::new();
        let int = types.int(IntKind::Int);
        let name = types.typedef(interner.intern("int32_t"), int);
        let konst = types.qualified(name, Qualifiers::CONST);
        assert!(matches!(types.kind(konst), TypeKind::Typedef { .. }), "still prints as int32_t");
        let want = types.qualified(int, Qualifiers::CONST);
        assert_eq!(types.canonical(konst), want);
    }

    #[test]
    fn a_typedef_of_an_array_pushes_a_qualifier_to_the_element_when_it_canonicalises() {
        // `typedef int A[4]; const A x;` declares an array of `const int`, which is where the
        // array rule and the sugar rule have to agree with each other.
        let mut interner = Interner::new();
        let mut types = Types::new();
        let int = types.int(IntKind::Int);
        let array = types.array(int, ArrayLen::Fixed(4));
        let name = types.typedef(interner.intern("A"), array);
        let konst = types.qualified(name, Qualifiers::CONST);
        let konst_int = types.qualified(int, Qualifiers::CONST);
        let want = types.array(konst_int, ArrayLen::Fixed(4));
        assert_eq!(types.canonical(konst), want);
    }

    #[test]
    fn a_function_type_is_deduplicated_by_its_signature() {
        let mut types = Types::new();
        let int = types.int(IntKind::Int);
        let long = types.int(IntKind::Long);
        let make = |types: &mut Types, params: Vec<TypeId>, variadic| {
            types.function(FunctionType { ret: int, params, variadic, prototyped: true })
        };
        let a = make(&mut types, vec![int, long], false);
        let b = make(&mut types, vec![int, long], false);
        assert_eq!(a, b);
        assert_ne!(a, make(&mut types, vec![int, long], true), "`...` is part of the type");
        assert_ne!(a, make(&mut types, vec![long, int], false));
    }

    #[test]
    fn a_function_type_written_with_a_typedef_canonicalises_through_its_signature() {
        let mut interner = Interner::new();
        let mut types = Types::new();
        let int = types.int(IntKind::Int);
        let name = types.typedef(interner.intern("int32_t"), int);
        let sugar = types.function(FunctionType {
            ret: name,
            params: vec![name],
            variadic: false,
            prototyped: true,
        });
        let plain = types.function(FunctionType {
            ret: int,
            params: vec![int],
            variadic: false,
            prototyped: true,
        });
        assert_ne!(sugar, plain);
        assert_eq!(types.canonical(sugar), plain);
    }

    #[test]
    fn a_record_is_its_declaration_and_not_its_members() {
        // Two structs written the same way in one translation unit are different types. The
        // looser relation that does hold between them is compatibility, which is a separate
        // question from identity and is answered elsewhere.
        let mut interner = Interner::new();
        let mut types = Types::new();
        let tag = interner.intern("point");
        let first = types.declare_record(RecordKind::Struct, Some(tag));
        let second = types.declare_record(RecordKind::Struct, Some(tag));
        assert_ne!(types.record(first), types.record(second));
        assert_eq!(types.record(first), types.record(first));
    }

    #[test]
    fn a_record_has_no_layout_until_it_has_been_completed() {
        let mut types = Types::new();
        let id = types.declare_record(RecordKind::Struct, None);
        let ty = types.record(id);
        assert_eq!(layout(&types, ty, &linux()), Err(LayoutError::Incomplete));
        let long_long = types.int(IntKind::LongLong);
        let laid_out = lay_out(&types, RecordKind::Struct, &[member(long_long); 2]);
        types.complete_record(id, laid_out);
        assert_eq!(layout(&types, ty, &linux()).unwrap(), Layout::new(16, 8));
    }

    #[test]
    fn an_enum_takes_the_layout_of_its_underlying_type() {
        let mut types = Types::new();
        let id = types.declare_enum(None);
        let ty = types.enumeration(id);
        assert_eq!(layout(&types, ty, &linux()), Err(LayoutError::Incomplete));
        let int = types.int(IntKind::Int);
        types.complete_enum(id, int, false);
        assert_eq!(layout(&types, ty, &linux()).unwrap(), Layout::new(4, 4));
    }

    #[test]
    fn the_scalar_widths_come_from_the_target() {
        let mut types = Types::new();
        let linux = linux();
        let windows = target("x86_64-pc-windows-msvc");
        let darwin = target("aarch64-apple-darwin");

        let long = types.int(IntKind::Long);
        assert_eq!(layout(&types, long, &linux).unwrap(), Layout::new(8, 8));
        assert_eq!(layout(&types, long, &windows).unwrap(), Layout::new(4, 4), "LLP64");

        let ldouble = types.float(FloatKind::LongDouble);
        assert_eq!(layout(&types, ldouble, &linux).unwrap(), Layout::new(16, 16));
        assert_eq!(layout(&types, ldouble, &darwin).unwrap(), Layout::new(8, 8));

        let pointer = types.pointer(types.void());
        assert_eq!(layout(&types, pointer, &linux).unwrap(), Layout::new(8, 8));

        let boolean = types.boolean();
        assert_eq!(layout(&types, boolean, &linux).unwrap(), Layout::new(1, 1));
    }

    #[test]
    fn a_complex_type_is_two_of_its_component_with_the_components_alignment() {
        // `_Complex long double` on SysV x86-64 is thirty two bytes aligned to sixteen, which
        // is the case that catches an implementation that aligns the pair to its own size.
        let mut types = Types::new();
        let linux = linux();
        let cfloat = types.complex_float(FloatKind::Float);
        assert_eq!(layout(&types, cfloat, &linux).unwrap(), Layout::new(8, 4));
        let cdouble = types.complex_float(FloatKind::Double);
        assert_eq!(layout(&types, cdouble, &linux).unwrap(), Layout::new(16, 8));
        let cldouble = types.complex_float(FloatKind::LongDouble);
        assert_eq!(layout(&types, cldouble, &linux).unwrap(), Layout::new(32, 16));
        let darwin = target("aarch64-apple-darwin");
        assert_eq!(layout(&types, cldouble, &darwin).unwrap(), Layout::new(16, 8));
    }

    #[test]
    fn an_atomic_type_can_be_more_aligned_than_the_type_it_wraps() {
        // The whole reason `_Atomic` is a type here rather than a qualifier. A sixteen byte
        // record is aligned to eight and the atomic version of it is aligned to sixteen.
        let mut types = Types::new();
        let linux = linux();
        let long_long = types.int(IntKind::LongLong);
        let plain = record(&mut types, RecordKind::Struct, &[member(long_long); 2]);
        let atomic = types.atomic(plain);
        assert_eq!(layout(&types, plain, &linux).unwrap(), Layout::new(16, 8));
        assert_eq!(layout(&types, atomic, &linux).unwrap(), Layout::new(16, 16));

        // An odd size cannot be accessed atomically in one go, so nothing is raised.
        let odd = record(&mut types, RecordKind::Struct, &[member(long_long); 3]);
        let atomic_odd = types.atomic(odd);
        assert_eq!(layout(&types, atomic_odd, &linux).unwrap(), Layout::new(24, 8));

        let int = types.int(IntKind::Int);
        let atomic_int = types.atomic(int);
        assert_eq!(layout(&types, atomic_int, &linux).unwrap(), Layout::new(4, 4));
    }

    #[test]
    fn a_bit_int_is_laid_out_like_a_standard_integer_until_it_outgrows_one() {
        // Measured with clang 18 on x86-64 Linux and clang on AArch64 Darwin. The two disagree
        // above sixty four bits, which is why the granule is a target fact.
        let mut types = Types::new();
        let linux = linux();
        let darwin = target("aarch64-apple-darwin");
        let cases = [(7, 1, 1), (8, 1, 1), (9, 2, 2), (17, 4, 4), (33, 8, 8), (64, 8, 8)];
        for (width, size, align) in cases {
            let ty = types.bit_int(true, width);
            assert_eq!(layout(&types, ty, &linux).unwrap(), Layout::new(size, align), "{width}");
            assert_eq!(layout(&types, ty, &darwin).unwrap(), Layout::new(size, align), "{width}");
        }
        for width in [65, 96, 128] {
            let ty = types.bit_int(false, width);
            assert_eq!(layout(&types, ty, &linux).unwrap(), Layout::new(16, 8), "{width}");
            assert_eq!(layout(&types, ty, &darwin).unwrap(), Layout::new(16, 16), "{width}");
        }
        let wide = types.bit_int(true, 129);
        assert_eq!(layout(&types, wide, &linux).unwrap(), Layout::new(24, 8));
        assert_eq!(layout(&types, wide, &darwin).unwrap(), Layout::new(32, 16));
    }

    #[test]
    fn an_array_is_its_element_repeated_and_keeps_its_elements_alignment() {
        let mut types = Types::new();
        let linux = linux();
        let int = types.int(IntKind::Int);
        let ty = types.array(int, ArrayLen::Fixed(10));
        assert_eq!(layout(&types, ty, &linux).unwrap(), Layout::new(40, 4));
        let nested = types.array(ty, ArrayLen::Fixed(3));
        assert_eq!(layout(&types, nested, &linux).unwrap(), Layout::new(120, 4));
    }

    #[test]
    fn an_array_without_a_size_is_incomplete_and_an_impossible_one_says_so() {
        let mut types = Types::new();
        let linux = linux();
        let int = types.int(IntKind::Int);
        for len in [ArrayLen::Unknown, ArrayLen::Star] {
            let ty = types.array(int, len);
            assert_eq!(layout(&types, ty, &linux), Err(LayoutError::Incomplete));
        }
        // An array whose length the program computes is a different answer from an incomplete
        // one, because it is not a mistake: there is a size and this is not the place that
        // knows it. A caller that only wants a number treats the two the same and a caller
        // building the arithmetic asks for the members instead.
        let measured = types.array(int, ArrayLen::Variable(VlaId(0)));
        assert_eq!(layout(&types, measured, &linux), Err(LayoutError::Variable));
        assert_eq!(align(&types, measured, &linux), Ok(4));
        let huge = types.array(int, ArrayLen::Fixed(u64::MAX));
        assert_eq!(layout(&types, huge, &linux), Err(LayoutError::TooLarge));
    }

    #[test]
    fn the_largest_array_is_the_largest_object_and_not_the_largest_number() {
        // The limit is `PTRDIFF_MAX` rather than wherever the multiplication happens to
        // overflow, so an array of a byte may be every byte an object may have and one more
        // than that is refused. gcc 16 gives the same two answers.
        let mut types = Types::new();
        let linux = linux();
        let max = linux.max_object_size();
        let ch = types.int(IntKind::Char);
        let fits = types.array(ch, ArrayLen::Fixed(max));
        assert_eq!(layout(&types, fits, &linux), Ok(Layout::new(max, 1)));
        let over = types.array(ch, ArrayLen::Fixed(max + 1));
        assert_eq!(layout(&types, over, &linux), Err(LayoutError::TooLarge));
    }

    #[test]
    fn a_record_may_be_as_large_as_an_object_may_be_and_no_larger() {
        // The shape `991014-1.c` in the gcc.c-torture execution suite asks about: a type
        // nothing is ever an object of is still a type `sizeof` has to answer about. Counting
        // the record in bits made the largest one an eighth of this, with the multiply by eight
        // overflowing rather than any rule saying so.
        let mut types = Types::new();
        let linux = linux();
        let max = linux.max_object_size();
        let ch = types.int(IntKind::Char);
        let int = types.int(IntKind::Int);
        let short = types.int(IntKind::Short);

        let huge = types.array(short, ArrayLen::Fixed((1 << 62) - 256));
        let members = [member(huge), member(int), member(int), member(int), member(int)];
        let laid_out = lay_out(&types, RecordKind::Struct, &members);
        assert_eq!(laid_out.layout, Layout::new((1 << 63) - 496, 4));

        let brim = types.array(ch, ArrayLen::Fixed(max));
        let laid_out = lay_out(&types, RecordKind::Struct, &[member(brim)]);
        assert_eq!(laid_out.layout, Layout::new(max, 1));

        let over = [member(brim), member(ch)];
        let options = RecordOptions::default();
        let error = layout_record(&types, RecordKind::Struct, &over, &options, &linux);
        assert_eq!(error, Err(RecordError::TooLarge));
    }

    #[test]
    fn a_bit_field_past_where_a_bit_count_fits_is_still_placed() {
        // Eight times the largest object is more than a `u64` holds, so a bit-field at the end
        // of a record that large has a bit offset no bit count can name. It is a byte offset
        // and a bit within it here, which is what lets this be laid out at all, and gcc 16
        // gives the same size for it.
        let mut types = Types::new();
        let linux = linux();
        let ch = types.int(IntKind::Char);
        let int = types.int(IntKind::Int);
        let mut interner = Interner::new();
        let buf = types.array(ch, ArrayLen::Fixed(linux.max_object_size() - 7));
        let members = [member(buf), bits(&mut interner, "x", int, 1)];
        let laid_out = lay_out(&types, RecordKind::Struct, &members);
        assert_eq!(laid_out.layout, Layout::new(9_223_372_036_854_775_804, 4));
        let last = laid_out.fields[1];
        assert_eq!((last.offset, last.bit), (9_223_372_036_854_775_800, 0));
        assert_eq!(last.bit_offset(), 73_786_976_294_838_206_400);
    }

    #[test]
    fn two_variable_length_arrays_of_the_same_element_are_still_different_types() {
        let mut types = Types::new();
        let int = types.int(IntKind::Int);
        let a = types.array(int, ArrayLen::Variable(VlaId(0)));
        let b = types.array(int, ArrayLen::Variable(VlaId(1)));
        assert_ne!(a, b);
    }

    #[test]
    fn a_vector_is_rounded_up_to_a_power_of_two_and_aligned_to_the_whole_thing() {
        // What GCC does with a `vector_size` that is not already one, checked against clang on
        // AArch64 Darwin, which accepts the three element case that GCC rejects outright.
        let mut types = Types::new();
        let linux = linux();
        let int = types.int(IntKind::Int);
        let four = types.vector(int, 4);
        assert_eq!(layout(&types, four, &linux).unwrap(), Layout::new(16, 16));
        let three = types.vector(int, 3);
        assert_eq!(layout(&types, three, &linux).unwrap(), Layout::new(16, 16));
        let three_chars = types.vector(types.int(IntKind::Char), 3);
        assert_eq!(layout(&types, three_chars, &linux).unwrap(), Layout::new(4, 4));
    }

    #[test]
    fn the_types_without_a_size_say_which_kind_of_without_they_are() {
        // Kept apart because GNU C gives both of them a size of one and a different warning,
        // and because a caller that cannot tell them apart cannot write either message.
        let mut types = Types::new();
        let linux = linux();
        let void = types.void();
        assert_eq!(layout(&types, void, &linux), Err(LayoutError::Incomplete));
        let int = types.int(IntKind::Int);
        let function = types.function(FunctionType {
            ret: int,
            params: Vec::new(),
            variadic: false,
            prototyped: true,
        });
        assert_eq!(layout(&types, function, &linux), Err(LayoutError::Function));
        let pointer_to_function = types.pointer(function);
        assert_eq!(layout(&types, pointer_to_function, &linux).unwrap(), Layout::new(8, 8));
    }

    #[test]
    fn a_struct_puts_each_member_at_the_next_offset_it_is_allowed_to_start_at() {
        let types = Types::new();
        let char_ = types.int(IntKind::Char);
        let int = types.int(IntKind::Int);
        let laid_out = lay_out(&types, RecordKind::Struct, &[member(char_), member(int)]);
        assert_eq!(laid_out.layout, Layout::new(8, 4));
        assert_eq!(offsets(&laid_out), [0, 32]);
        assert_eq!(laid_out.fields[1].offset, 4);

        // And the tail is padded, which is what makes an array of the thing work.
        let long_long = types.int(IntKind::LongLong);
        let laid_out = lay_out(&types, RecordKind::Struct, &[member(long_long), member(char_)]);
        assert_eq!(laid_out.layout, Layout::new(16, 8));
    }

    #[test]
    fn a_union_starts_every_member_at_zero_and_is_as_large_as_the_largest() {
        let mut types = Types::new();
        let char_ = types.int(IntKind::Char);
        let int = types.int(IntKind::Int);
        let laid_out = lay_out(&types, RecordKind::Union, &[member(char_), member(int)]);
        assert_eq!(laid_out.layout, Layout::new(4, 4));
        assert_eq!(offsets(&laid_out), [0, 0]);

        // Nine bytes and a short is ten, not nine and not sixteen: the size is rounded up to
        // the alignment rather than to the largest member.
        let nine = types.array(char_, ArrayLen::Fixed(9));
        let short = types.int(IntKind::Short);
        let laid_out = lay_out(&types, RecordKind::Union, &[member(nine), member(short)]);
        assert_eq!(laid_out.layout, Layout::new(10, 2));
    }

    #[test]
    fn bit_fields_share_a_unit_until_one_of_them_would_span_two() {
        // Measured with gcc 13.3 on x86-64 Linux and clang on AArch64 Darwin, including where
        // the bits landed, by setting each field to all ones and dumping the bytes.
        let mut interner = Interner::new();
        let types = Types::new();
        let char_ = types.int(IntKind::Char);
        let int = types.int(IntKind::Int);
        let long_long = types.int(IntKind::LongLong);

        let fields = [bits(&mut interner, "a", int, 3), bits(&mut interner, "b", int, 5)];
        let laid_out = lay_out(&types, RecordKind::Struct, &fields);
        assert_eq!(laid_out.layout, Layout::new(4, 4));
        assert_eq!(offsets(&laid_out), [0, 3]);

        // Thirty bits do not fit in what is left of the first int, so they start a new one.
        let fields = [member(char_), bits(&mut interner, "b", int, 30)];
        let laid_out = lay_out(&types, RecordKind::Struct, &fields);
        assert_eq!(laid_out.layout, Layout::new(8, 4));
        assert_eq!(offsets(&laid_out), [0, 32]);

        // Thirty three bits of a `long long` do fit in what is left of the first one, because
        // the unit is eight bytes rather than four, so they stay where they are.
        let fields = [member(char_), bits(&mut interner, "b", long_long, 33)];
        let laid_out = lay_out(&types, RecordKind::Struct, &fields);
        assert_eq!(laid_out.layout, Layout::new(8, 8));
        assert_eq!(offsets(&laid_out), [0, 8]);

        // An ordinary member after a bit-field starts at the next byte it is allowed to.
        let fields = [bits(&mut interner, "a", int, 3), member(char_)];
        let laid_out = lay_out(&types, RecordKind::Struct, &fields);
        assert_eq!(offsets(&laid_out), [0, 8]);
    }

    #[test]
    fn a_zero_width_bit_field_moves_the_next_member_on_and_nothing_else() {
        let types = Types::new();
        let char_ = types.int(IntKind::Char);
        let int = types.int(IntKind::Int);
        let fields = [member(char_), unnamed_bits(int, 0), member(char_)];
        let laid_out = lay_out(&types, RecordKind::Struct, &fields);
        // Five bytes aligned to one: the zero width field pushed the second `char` to offset
        // four without giving the record the alignment of an `int`. Both compilers report that.
        assert_eq!(laid_out.layout, Layout::new(5, 1));
        assert_eq!(offsets(&laid_out), [0, 32, 32]);
        assert_eq!(laid_out.fields.len(), 3, "one field per declaration, so indices line up");

        // With nothing after it the padding is still the record's, which is the half of the rule
        // the case above hides: the second `char` ends further along than the zero width field
        // does, so whether the field moved the size or only the next member never showed.
        let trailing = [member(char_), unnamed_bits(int, 0)];
        assert_eq!(lay_out(&types, RecordKind::Struct, &trailing).layout, Layout::new(4, 1));

        // And a record that is nothing but the zero width field has nothing to pad, so it is the
        // empty structure with the alignment of whatever the field's type was.
        let only = [unnamed_bits(int, 0)];
        assert_eq!(lay_out(&types, RecordKind::Struct, &only).layout, Layout::new(0, 1));
    }

    #[test]
    fn an_unnamed_bit_field_does_not_raise_the_records_alignment_but_a_named_one_does() {
        let mut interner = Interner::new();
        let types = Types::new();
        let char_ = types.int(IntKind::Char);
        let int = types.int(IntKind::Int);

        let unnamed = [member(char_), unnamed_bits(int, 20)];
        let unnamed = lay_out(&types, RecordKind::Struct, &unnamed);
        assert_eq!(unnamed.layout, Layout::new(4, 1));

        let named = [member(char_), bits(&mut interner, "b", int, 20)];
        let named = lay_out(&types, RecordKind::Struct, &named);
        assert_eq!(named.layout, Layout::new(4, 4));
        assert_eq!(offsets(&named), [0, 8], "the same place either way");

        // The unit an unnamed field has to fit inside is still its own type's, so this one
        // moves to bit thirty two and the record is eight bytes aligned to one.
        let wider = [member(char_), unnamed_bits(int, 30)];
        let wider = lay_out(&types, RecordKind::Struct, &wider);
        assert_eq!(wider.layout, Layout::new(8, 1));
        assert_eq!(offsets(&wider), [0, 32]);
    }

    #[test]
    fn aapcs64_lets_an_unnamed_bit_field_raise_the_records_alignment() {
        let types = Types::new();
        let char_ = types.int(IntKind::Char);
        let uint = types.int(IntKind::UInt);

        // The same structure as the test above, on the one ABI in the table that disagrees.
        let fields = [member(char_), unnamed_bits(uint, 20)];
        let arm = lay_out_on(&aapcs(), &types, RecordKind::Struct, &fields);
        assert_eq!(arm.layout, Layout::new(4, 4));

        // The zero width member is the case a program actually writes, and it is where the rule
        // is visible with nothing else in the record at all.
        let only = [unnamed_bits(uint, 0)];
        assert_eq!(
            lay_out_on(&aapcs(), &types, RecordKind::Struct, &only).layout,
            Layout::new(0, 4)
        );
        assert_eq!(lay_out(&types, RecordKind::Struct, &only).layout, Layout::new(0, 1));

        // And it changes a size rather than only an alignment, because the record is rounded up
        // to the alignment it ends with. Five bytes on x86-64 and eight here.
        let pushed = [member(char_), unnamed_bits(uint, 0), member(char_)];
        let pushed = lay_out_on(&aapcs(), &types, RecordKind::Struct, &pushed);
        assert_eq!(pushed.layout, Layout::new(8, 4));
        assert_eq!(offsets(&pushed), [0, 32, 32]);
    }

    #[test]
    fn windows_allocates_a_bit_field_into_a_unit_of_its_declared_type() {
        let mut interner = Interner::new();
        let types = Types::new();
        let char_ = types.int(IntKind::Char);
        let uint = types.int(IntKind::UInt);
        let ushort = types.int(IntKind::UShort);
        let longlong = types.int(IntKind::LongLong);

        // An ordinary member closes the unit, and the unit costs its whole four bytes, so the
        // `char` is at offset four rather than at offset one.
        let then_member = [bits(&mut interner, "m0", uint, 3), member(char_)];
        let ms = lay_out_on(&windows(), &types, RecordKind::Struct, &then_member);
        assert_eq!(ms.layout, Layout::new(8, 4));
        assert_eq!(offsets(&ms), [0, 32]);
        let itanium = lay_out(&types, RecordKind::Struct, &then_member);
        assert_eq!(itanium.layout, Layout::new(4, 4));
        assert_eq!(offsets(&itanium), [0, 8]);

        // A declared type of a different size closes it too, although five bits were free.
        let narrower = [bits(&mut interner, "m0", uint, 3), bits(&mut interner, "m1", ushort, 5)];
        let ms = lay_out_on(&windows(), &types, RecordKind::Struct, &narrower);
        assert_eq!(ms.layout, Layout::new(8, 4));
        assert_eq!(offsets(&ms), [0, 32]);
        assert_eq!(lay_out(&types, RecordKind::Struct, &narrower).layout, Layout::new(4, 4));

        // And the unit is opened at its own alignment, so a `long long` bit-field after a `char`
        // starts at offset eight where the Itanium rule leaves it at bit eight.
        let wide = [member(char_), bits(&mut interner, "b", longlong, 33)];
        let ms = lay_out_on(&windows(), &types, RecordKind::Struct, &wide);
        assert_eq!(ms.layout, Layout::new(16, 8));
        assert_eq!(offsets(&ms), [0, 64]);
        let itanium = lay_out(&types, RecordKind::Struct, &wide);
        assert_eq!(itanium.layout, Layout::new(8, 8));
        assert_eq!(offsets(&itanium), [0, 8]);
    }

    #[test]
    fn microsofts_zero_width_bit_field_closes_a_unit_and_does_nothing_without_one() {
        let mut interner = Interner::new();
        let types = Types::new();
        let char_ = types.int(IntKind::Char);
        let uint = types.int(IntKind::UInt);

        // Nothing before it is a bit-field, so there is no run to end and the member is free.
        let alone = [member(char_), unnamed_bits(uint, 0)];
        assert_eq!(
            lay_out_on(&windows(), &types, RecordKind::Struct, &alone).layout,
            Layout::new(1, 1)
        );
        assert_eq!(lay_out(&types, RecordKind::Struct, &alone).layout, Layout::new(4, 1));

        // With a unit open it ends it, and the member after starts a unit of its own.
        let between = [
            bits(&mut interner, "m0", uint, 3),
            unnamed_bits(uint, 0),
            bits(&mut interner, "m1", uint, 5),
            member(char_),
        ];
        let ms = lay_out_on(&windows(), &types, RecordKind::Struct, &between);
        assert_eq!(ms.layout, Layout::new(12, 4));
        assert_eq!(offsets(&ms), [0, 32, 32, 64]);
        let itanium = lay_out(&types, RecordKind::Struct, &between);
        assert_eq!(itanium.layout, Layout::new(8, 4));
        assert_eq!(offsets(&itanium), [0, 32, 32, 40]);
    }

    #[test]
    fn microsoft_gives_a_unions_bit_field_storage_and_no_say_in_the_alignment() {
        let mut interner = Interner::new();
        let types = Types::new();
        let char_ = types.int(IntKind::Char);
        let uint = types.int(IntKind::UInt);

        // Four bytes because the unit is an `unsigned`, aligned to one because the only member
        // that gets a say is the `char`. An alignment smaller than either member would have.
        let fields = [bits(&mut interner, "m0", uint, 3), member(char_)];
        assert_eq!(
            lay_out_on(&windows(), &types, RecordKind::Union, &fields).layout,
            Layout::new(4, 1)
        );
        assert_eq!(lay_out(&types, RecordKind::Union, &fields).layout, Layout::new(4, 4));
    }

    #[test]
    fn a_record_with_no_storage_in_it_is_four_bytes_under_msvc_and_nothing_anywhere_else() {
        let mut types = Types::new();
        let uint = types.int(IntKind::UInt);
        let mingw = target("x86_64-pc-windows-gnu");

        // Three shapes that hold nothing, and the reference gives all three the same answer.
        let none: [FieldDecl; 0] = [];
        let zero_width = [unnamed_bits(uint, 0)];
        let flexible = [member(types.array(uint, ArrayLen::Unknown))];
        for fields in [&none[..], &zero_width[..], &flexible[..]] {
            let msvc = lay_out_on(&windows(), &types, RecordKind::Struct, fields);
            assert_eq!(msvc.layout.size, 4, "four bytes under MSVC");
            assert_eq!(lay_out_on(&mingw, &types, RecordKind::Struct, fields).layout.size, 0);
            assert_eq!(lay_out(&types, RecordKind::Struct, fields).layout.size, 0);
        }

        // It is the environment that decides and not the operating system, so the two Windows
        // targets disagree with each other and mingw agrees with Linux. A rule keyed on the
        // operating system would have put both of them at four.
        assert_eq!(windows().empty_record_size, 4);
        assert_eq!(mingw.empty_record_size, 0);
    }

    #[test]
    fn packed_drops_every_member_to_a_byte_and_bit_fields_to_the_next_free_bit() {
        let mut interner = Interner::new();
        let types = Types::new();
        let char_ = types.int(IntKind::Char);
        let int = types.int(IntKind::Int);
        let packed = RecordOptions { packed: true, ..RecordOptions::default() };

        let fields = [member(char_), member(int)];
        let laid_out = layout_record(&types, RecordKind::Struct, &fields, &packed, &linux())
            .expect("a packed struct of two complete members");
        assert_eq!(laid_out.layout, Layout::new(5, 1));
        assert_eq!(offsets(&laid_out), [0, 8]);

        let fields = [member(char_), bits(&mut interner, "b", int, 30)];
        let laid_out = layout_record(&types, RecordKind::Struct, &fields, &packed, &linux())
            .expect("a packed struct with a bit-field");
        assert_eq!(laid_out.layout, Layout::new(5, 1));
        assert_eq!(offsets(&laid_out), [0, 8], "no boundary left to move to");

        // A zero width bit-field still rounds to its own type, packed or not, which is the
        // whole reason a program writes one inside a packed structure.
        let fields = [member(char_), unnamed_bits(int, 0), member(char_)];
        let laid_out = layout_record(&types, RecordKind::Struct, &fields, &packed, &linux())
            .expect("a packed struct with a zero width bit-field");
        assert_eq!(laid_out.layout, Layout::new(5, 1));
        assert_eq!(offsets(&laid_out), [0, 32, 32]);
    }

    #[test]
    fn pragma_pack_caps_alignment_and_leaves_a_bit_field_where_it_already_is() {
        let mut interner = Interner::new();
        let types = Types::new();
        let char_ = types.int(IntKind::Char);
        let int = types.int(IntKind::Int);
        let pack = RecordOptions { pack: Some(2), ..RecordOptions::default() };

        let fields = [member(char_), member(int)];
        let laid_out = layout_record(&types, RecordKind::Struct, &fields, &pack, &linux())
            .expect("a packed struct of two complete members");
        assert_eq!(laid_out.layout, Layout::new(6, 2));
        assert_eq!(offsets(&laid_out), [0, 16]);

        // Six bytes with the field at bit eight, not at bit sixteen. Once the alignment has
        // been capped below the type's own there is no boundary to move to, so the field stays
        // put. Measured, because moving it is at least as plausible a reading.
        let fields = [member(char_), bits(&mut interner, "b", int, 30)];
        let laid_out = layout_record(&types, RecordKind::Struct, &fields, &pack, &linux())
            .expect("a packed struct with a bit-field");
        assert_eq!(laid_out.layout, Layout::new(6, 2));
        assert_eq!(offsets(&laid_out), [0, 8]);

        // The same structure with the field unnamed is five bytes aligned to one, because the
        // capped alignment reached it through the record and an unnamed field gives none back.
        let fields = [member(char_), unnamed_bits(int, 30)];
        let laid_out = layout_record(&types, RecordKind::Struct, &fields, &pack, &linux())
            .expect("a packed struct with an unnamed bit-field");
        assert_eq!(laid_out.layout, Layout::new(5, 1));
    }

    #[test]
    fn an_alignment_the_program_asked_for_raises_the_member_and_the_record() {
        let types = Types::new();
        let char_ = types.int(IntKind::Char);
        let int = types.int(IntKind::Int);

        let aligned = FieldDecl { align: Some(16), ..member(int) };
        let laid_out = lay_out(&types, RecordKind::Struct, &[member(char_), aligned]);
        assert_eq!(laid_out.layout, Layout::new(32, 16));
        assert_eq!(offsets(&laid_out), [0, 128]);

        // `packed, aligned(4)` together: the members pack and the record does not, which is
        // the combination the attribute pair exists for.
        let options = RecordOptions { packed: true, align: Some(4), pack: None };
        let fields = [member(char_), member(int)];
        let laid_out = layout_record(&types, RecordKind::Struct, &fields, &options, &linux())
            .expect("a packed struct with an alignment asked for");
        assert_eq!(laid_out.layout, Layout::new(8, 4));
        assert_eq!(offsets(&laid_out), [0, 8]);
    }

    #[test]
    fn a_flexible_array_member_costs_nothing_but_its_alignment() {
        // What makes `malloc(sizeof(struct S) + n)` the idiom it is.
        let mut types = Types::new();
        let char_ = types.int(IntKind::Char);
        let int = types.int(IntKind::Int);
        let long_long = types.int(IntKind::LongLong);

        let chars = types.array(char_, ArrayLen::Unknown);
        let laid_out = lay_out(&types, RecordKind::Struct, &[member(int), member(chars)]);
        assert_eq!(laid_out.layout, Layout::new(4, 4));
        assert_eq!(offsets(&laid_out), [0, 32]);

        // The alignment still applies, so this is eight bytes of which one is the `char`.
        let longs = types.array(long_long, ArrayLen::Unknown);
        let laid_out = lay_out(&types, RecordKind::Struct, &[member(char_), member(longs)]);
        assert_eq!(laid_out.layout, Layout::new(8, 8));
        assert_eq!(offsets(&laid_out), [0, 64]);

        // Anywhere but last it is an incomplete member, and which member is part of the answer.
        let fields = [member(chars), member(int)];
        let error =
            layout_record(&types, RecordKind::Struct, &fields, &RecordOptions::default(), &linux());
        assert_eq!(error, Err(RecordError::Member { index: 0, error: LayoutError::Incomplete }));
    }

    #[test]
    fn a_record_with_no_members_is_zero_bytes_aligned_to_one() {
        // The GNU empty structure, which C itself does not have and which real headers do.
        let types = Types::new();
        let laid_out = lay_out(&types, RecordKind::Struct, &[]);
        assert_eq!(laid_out.layout, Layout::new(0, 1));
    }

    #[test]
    fn a_bit_field_wider_than_the_type_it_is_declared_with_is_refused() {
        let types = Types::new();
        let int = types.int(IntKind::Int);
        let fields = [unnamed_bits(int, 33)];
        let error =
            layout_record(&types, RecordKind::Struct, &fields, &RecordOptions::default(), &linux());
        let want = RecordError::BitFieldTooWide { index: 0, width: 33, capacity: 32 };
        assert_eq!(error, Err(want));
    }

    #[test]
    fn a_record_reports_its_members_once_it_has_been_completed() {
        let mut interner = Interner::new();
        let mut types = Types::new();
        let char_ = types.int(IntKind::Char);
        let int = types.int(IntKind::Int);
        let name = interner.intern("count");
        let fields = [member(char_), FieldDecl::new(Some(name), int)];
        let id = types.declare_record(RecordKind::Struct, None);
        let laid_out = lay_out(&types, RecordKind::Struct, &fields);
        types.complete_record(id, laid_out);
        let ty = types.record(id);
        assert_eq!(layout(&types, ty, &linux()).unwrap(), Layout::new(8, 4));
        let field = types.field(id, name).expect("the member that was declared");
        assert_eq!(field.offset, 4);
        assert!(!field.is_bit_field());
        assert_eq!(types.field(id, interner.intern("missing")), None);
    }

    #[test]
    fn a_nested_record_brings_its_own_alignment_with_it() {
        let mut types = Types::new();
        let char_ = types.int(IntKind::Char);
        let int = types.int(IntKind::Int);
        let inner = record(&mut types, RecordKind::Struct, &[member(char_)]);
        let laid_out = lay_out(&types, RecordKind::Struct, &[member(inner), member(int)]);
        assert_eq!(laid_out.layout, Layout::new(8, 4));
        assert_eq!(offsets(&laid_out), [0, 32]);

        // An anonymous member is an ordinary member with no name, so the same code lays it out
        // and the four bytes of padding after the `char` are there either way.
        let anonymous = record(&mut types, RecordKind::Struct, &[member(int), member(char_)]);
        let laid_out = lay_out(&types, RecordKind::Struct, &[member(char_), member(anonymous)]);
        assert_eq!(laid_out.layout, Layout::new(12, 4));
        assert_eq!(offsets(&laid_out), [0, 32]);
    }

    #[test]
    fn everything_narrower_than_an_int_promotes_to_one() {
        // Measured by naming the type of `+x` with `_Generic` in gcc 13.3 and clang 18. Every
        // one of these answers `int`, including the unsigned ones, because an `int` holds every
        // value a sixteen bit unsigned type has.
        let mut types = Types::new();
        let linux = linux();
        let int = types.int(IntKind::Int);
        let narrow =
            [IntKind::Char, IntKind::SChar, IntKind::UChar, IntKind::Short, IntKind::UShort];
        for kind in narrow {
            let ty = types.int(kind);
            assert_eq!(promote(&mut types, ty, &linux), int, "{}", kind.as_str());
        }
        let boolean = types.boolean();
        assert_eq!(promote(&mut types, boolean, &linux), int, "C23 made bool a real type");

        // From `int` up, a type is its own promotion.
        for kind in [IntKind::Int, IntKind::UInt, IntKind::Long, IntKind::ULongLong] {
            let ty = types.int(kind);
            assert_eq!(promote(&mut types, ty, &linux), ty, "{}", kind.as_str());
        }
    }

    #[test]
    fn a_bit_int_is_not_promoted_at_all() {
        // C23 6.3.1.1p2, and the point of the type. `_BitInt(8) + _BitInt(8)` stays eight bits
        // wide where `char + char` is an `int`, which is what makes the width mean something.
        let mut types = Types::new();
        let linux = linux();
        let small = types.bit_int(true, 8);
        assert_eq!(promote(&mut types, small, &linux), small);
        assert_eq!(usual_arithmetic(&mut types, small, small, &linux), Some(small));
    }

    #[test]
    fn a_bit_field_is_promoted_by_its_width_and_not_by_its_type() {
        let mut types = Types::new();
        let linux = linux();
        let int = types.int(IntKind::Int);
        let uint = types.int(IntKind::UInt);
        let ullong = types.int(IntKind::ULongLong);

        // Three bits of an unsigned field all fit in an `int`, so it is signed afterwards.
        assert_eq!(promote_bit_field(&mut types, uint, 3, &linux), int);
        // Thirty two of them do not.
        assert_eq!(promote_bit_field(&mut types, uint, 32, &linux), uint);
        // Twenty bits of a signed field, which is an `int` either way.
        assert_eq!(promote_bit_field(&mut types, int, 20, &linux), int);
        // Forty bits are forty bits of value and nothing more. The C17 wording says `unsigned
        // int` here, which would silently drop eight of them, and C23 says the declared type,
        // which would silently add twenty four. Both compilers give the width instead, so
        // `x.b << 32` on such a field is zero rather than a value with a bit above the fortieth.
        let forty = types.bit_int(false, 40);
        assert_eq!(promote_bit_field(&mut types, ullong, 40, &linux), forty);
        // A field as wide as its type is that type, since there is no precision to lose.
        assert_eq!(promote_bit_field(&mut types, ullong, 64, &linux), ullong);
    }

    #[test]
    fn an_enumeration_promotes_through_what_it_is_represented_in() {
        let mut types = Types::new();
        let linux = linux();
        let int = types.int(IntKind::Int);
        let short = types.int(IntKind::Short);
        let uint = types.int(IntKind::UInt);

        // `enum E : short` promotes the same way a `short` does, which is to `int`.
        let fixed = types.declare_enum(None);
        types.complete_enum(fixed, short, true);
        let fixed = types.enumeration(fixed);
        assert_eq!(promote(&mut types, fixed, &linux), int);

        // An enumeration all of whose enumerators are non-negative is represented in
        // `unsigned int` by both compilers, and then it promotes to itself.
        let unsigned = types.declare_enum(None);
        types.complete_enum(unsigned, uint, false);
        let unsigned = types.enumeration(unsigned);
        assert_eq!(promote(&mut types, unsigned, &linux), uint);

        // An enumeration nobody has decided on yet answers `int`, so that an expression using
        // one is still checkable while the diagnostic about it is being written.
        let undecided = types.declare_enum(None);
        let undecided = types.enumeration(undecided);
        assert_eq!(promote(&mut types, undecided, &linux), int);
    }

    #[test]
    fn the_qualifiers_and_the_atomic_come_off_before_anything_else() {
        // By the time a value is being promoted the lvalue conversion has already happened, so
        // `_Atomic const int` and `int` are the same operand.
        let mut types = Types::new();
        let linux = linux();
        let int = types.int(IntKind::Int);
        let konst = types.qualified(int, Qualifiers::CONST);
        let atomic = types.atomic(konst);
        assert_eq!(promote(&mut types, atomic, &linux), int);
        assert_eq!(usual_arithmetic(&mut types, atomic, konst, &linux), Some(int));
    }

    #[test]
    fn the_usual_arithmetic_conversions_between_the_standard_integer_types() {
        // Every row measured with `_Generic` in gcc 13.3 and clang 18 on x86-64 Linux.
        let mut types = Types::new();
        let linux = linux();
        let cases = [
            (IntKind::Int, IntKind::UInt, IntKind::UInt),
            (IntKind::Int, IntKind::Long, IntKind::Long),
            (IntKind::UInt, IntKind::Long, IntKind::Long),
            (IntKind::UInt, IntKind::ULong, IntKind::ULong),
            (IntKind::Int, IntKind::LongLong, IntKind::LongLong),
            (IntKind::UInt, IntKind::LongLong, IntKind::LongLong),
            (IntKind::ULong, IntKind::LongLong, IntKind::ULongLong),
            (IntKind::Char, IntKind::Char, IntKind::Int),
            (IntKind::UChar, IntKind::UShort, IntKind::Int),
        ];
        for (left, right, want) in cases {
            let left = types.int(left);
            let right = types.int(right);
            let want = types.int(want);
            assert_eq!(usual_arithmetic(&mut types, left, right, &linux), Some(want));
            assert_eq!(usual_arithmetic(&mut types, right, left, &linux), Some(want), "either way");
        }
    }

    #[test]
    fn int128_is_sixteen_bytes_aligned_to_sixteen_and_outranks_long_long() {
        // Measured on gcc 13.3 on x86-64 Linux and clang on AArch64 Darwin, both of which
        // report the same size, the same alignment, and an offset of sixteen for a member
        // after a `char`.
        let mut types = Types::new();
        let linux = linux();
        let signed = types.int(IntKind::Int128);
        let unsigned = types.int(IntKind::UInt128);
        for id in [signed, unsigned] {
            let laid_out = layout(&types, id, &linux).expect("a complete type");
            assert_eq!(laid_out.size, 16);
            assert_eq!(laid_out.align, 16);
        }

        // `__int128 + unsigned long long` is `__int128`, because it wins on rank and is wide
        // enough to hold every value the other side had. Both compilers agree, and it is the
        // one pair that says the rank is above `long long` rather than beside it.
        let ull = types.int(IntKind::ULongLong);
        assert_eq!(usual_arithmetic(&mut types, signed, ull, &linux), Some(signed));
        // And it is its own promotion, the way every type at or above `int` is.
        assert_eq!(promote(&mut types, signed, &linux), signed);
    }

    #[test]
    fn a_bit_int_of_a_hundred_and_twenty_eight_bits_is_not_int128() {
        // Same width, different types. The alignment is the visible difference on x86-64,
        // where a `_BitInt` is aligned to its sixty four bit granule and `__int128` is not.
        let mut types = Types::new();
        let linux = linux();
        let int128 = types.int(IntKind::Int128);
        let bit_int = types.bit_int(true, 128);
        assert_ne!(int128, bit_int);
        assert!(!compatible(&types, int128, bit_int));
        assert_eq!(layout(&types, bit_int, &linux).expect("complete").align, 8);
        assert_eq!(layout(&types, int128, &linux).expect("complete").align, 16);
    }

    #[test]
    fn the_last_arm_takes_the_unsigned_type_of_the_wider_one() {
        // `unsigned long + long long` is `unsigned long long` on Linux: the `long long` wins on
        // rank and cannot hold every value of the `unsigned long`, so neither operand's own
        // type is the answer. This is the arm programs are surprised by.
        let mut types = Types::new();
        let linux = linux();
        let ulong = types.int(IntKind::ULong);
        let long_long = types.int(IntKind::LongLong);
        let want = types.int(IntKind::ULongLong);
        assert_eq!(usual_arithmetic(&mut types, ulong, long_long, &linux), Some(want));

        // The same pair on Windows, where `long` is thirty two bits, comes out as `long long`,
        // because there it does hold every value. A host-driven implementation gets one of
        // these two wrong.
        let windows = target("x86_64-pc-windows-msvc");
        assert_eq!(usual_arithmetic(&mut types, ulong, long_long, &windows), Some(long_long));
    }

    #[test]
    fn a_bit_int_is_ranked_by_its_width_against_the_standard_types() {
        // Measured with clang 18 on x86-64 Linux, which is the compiler that has `_BitInt`.
        let mut types = Types::new();
        let linux = linux();
        let b40 = types.bit_int(true, 40);
        let ub40 = types.bit_int(false, 40);
        let b8 = types.bit_int(true, 8);
        let b32 = types.bit_int(true, 32);
        let int = types.int(IntKind::Int);
        let uint = types.int(IntKind::UInt);
        let long = types.int(IntKind::Long);
        let char_ = types.int(IntKind::Char);

        // Wider than an `int`, so it outranks one.
        assert_eq!(usual_arithmetic(&mut types, b40, int, &linux), Some(b40));
        // Narrower than a `long`, so it loses to one.
        assert_eq!(usual_arithmetic(&mut types, b40, long, &linux), Some(long));
        // The same width as an `int`, and a standard type wins the tie.
        assert_eq!(usual_arithmetic(&mut types, b32, int, &linux), Some(int));
        assert_eq!(usual_arithmetic(&mut types, b32, uint, &linux), Some(uint));
        // The other side promotes first, so a `char` next to a narrow `_BitInt` is an `int`
        // and the `_BitInt` loses to it.
        assert_eq!(usual_arithmetic(&mut types, b8, char_, &linux), Some(int));
        // Unsigned and higher ranked wins outright, and unsigned and lower ranked loses to a
        // signed type wide enough to hold it.
        assert_eq!(usual_arithmetic(&mut types, ub40, int, &linux), Some(ub40));
        assert_eq!(usual_arithmetic(&mut types, ub40, long, &linux), Some(long));
        // Two bit-precise types of the same width and different signedness.
        assert_eq!(usual_arithmetic(&mut types, b40, ub40, &linux), Some(ub40));
    }

    #[test]
    fn a_floating_operand_decides_the_answer_whatever_the_other_side_is() {
        let mut types = Types::new();
        let linux = linux();
        let float = types.float(FloatKind::Float);
        let double = types.float(FloatKind::Double);
        let long_double = types.float(FloatKind::LongDouble);
        let ullong = types.int(IntKind::ULongLong);
        let int = types.int(IntKind::Int);

        assert_eq!(usual_arithmetic(&mut types, int, float, &linux), Some(float));
        assert_eq!(usual_arithmetic(&mut types, float, double, &linux), Some(double));
        assert_eq!(usual_arithmetic(&mut types, double, long_double, &linux), Some(long_double));
        // Sixty four bits of unsigned integer against a `float`, which is a `float` and loses
        // most of them. That is the rule rather than an oversight.
        assert_eq!(usual_arithmetic(&mut types, ullong, float, &linux), Some(float));
    }

    #[test]
    fn a_mask_is_the_signed_integers_of_the_lane_width() {
        let mut types = Types::new();
        let linux = linux();
        let int = types.int(IntKind::Int);
        let float = types.float(FloatKind::Float);
        let short = types.int(IntKind::Short);

        // A signed lane is already its own mask, so the answer is the vector it was given.
        let four_ints = types.vector(int, 4);
        assert_eq!(mask_of(&mut types, four_ints, &linux), Some(four_ints));

        // An unsigned lane answers as the signed type of the same width, which is what GCC
        // gives a comparison of two `unsigned int` vectors.
        let uint = types.int(IntKind::UInt);
        let four_uints = types.vector(uint, 4);
        assert_eq!(mask_of(&mut types, four_uints, &linux), Some(four_ints));

        // A float lane answers as an integer of the same width, since the mask is bits and not
        // a number and there is no float that is all ones.
        let four_floats = types.vector(float, 4);
        assert_eq!(mask_of(&mut types, four_floats, &linux), Some(four_ints));

        // The width is the lane's own and not a word, so a `short` lane keeps its two bytes.
        let two_shorts = types.vector(short, 2);
        assert_eq!(mask_of(&mut types, two_shorts, &linux), Some(two_shorts));

        // Not a vector, so there is no mask to give.
        assert_eq!(mask_of(&mut types, int, &linux), None);
    }

    #[test]
    fn two_vectors_convert_between_each_other_when_the_bytes_line_up() {
        let mut types = Types::new();
        let linux = linux();
        let int = types.int(IntKind::Int);
        let uint = types.int(IntKind::UInt);
        let float = types.float(FloatKind::Float);
        let short = types.int(IntKind::Short);

        let four_ints = types.vector(int, 4);
        let four_uints = types.vector(uint, 4);
        let four_floats = types.vector(float, 4);
        let eight_shorts = types.vector(short, 8);
        let two_ints = types.vector(int, 2);

        // The case the whole thing exists for: a mask assigned to the unsigned vector it came
        // from, which GNU C converts and the standard rules would refuse.
        assert!(vectors_convertible(&types, four_uints, four_ints, &linux));
        // Both ways round, since assignment happens in both directions.
        assert!(vectors_convertible(&types, four_ints, four_uints, &linux));
        // The same sixteen bytes cut into eight lanes rather than four, which GCC also allows.
        assert!(vectors_convertible(&types, four_ints, eight_shorts, &linux));
        // Two floats of the same width, which is the other half of the rule.
        assert!(vectors_convertible(&types, four_floats, four_floats, &linux));

        // An integer lane against a float lane, which GCC refuses even at the same size,
        // because reading one as the other is a cast and not a conversion.
        assert!(!vectors_convertible(&types, four_ints, four_floats, &linux));
        // Different sizes, so there is nothing to reinterpret.
        assert!(!vectors_convertible(&types, four_ints, two_ints, &linux));
        // A scalar is not a vector, whichever side it is on.
        assert!(!vectors_convertible(&types, four_ints, int, &linux));
        assert!(!vectors_convertible(&types, int, four_ints, &linux));
    }

    /// Insists that `a + b` and `b + a` are both `expected` on this target.
    ///
    /// Both ways round, because the operands of `+` are not ordered and an implementation that
    /// keeps the left one when it cannot decide would pass half of these and be wrong.
    fn combines(target: &TargetInfo, a: FloatKind, b: FloatKind, expected: FloatKind) {
        let mut types = Types::new();
        let left = types.float(a);
        let right = types.float(b);
        let want = types.float(expected);
        assert_eq!(usual_arithmetic(&mut types, left, right, target), Some(want), "{a:?} + {b:?}");
        assert_eq!(usual_arithmetic(&mut types, right, left, target), Some(want), "{b:?} + {a:?}");
    }

    #[test]
    fn two_floating_types_of_the_same_format_are_still_two_types_and_one_of_them_wins() {
        // Every line here was read off gcc 16 with `_Generic` rather than off the standard, on
        // x86-64 Linux, where `long double` and `_Float64x` are both the x87 format and the
        // standard type is the one that comes out.
        let x86 = linux();
        combines(&x86, FloatKind::Double, FloatKind::Float64, FloatKind::Float64);
        combines(&x86, FloatKind::Float, FloatKind::Float32, FloatKind::Float32);
        combines(&x86, FloatKind::Double, FloatKind::Float32x, FloatKind::Double);
        combines(&x86, FloatKind::LongDouble, FloatKind::Float64x, FloatKind::LongDouble);
        combines(&x86, FloatKind::Float128, FloatKind::LongDouble, FloatKind::Float128);
        combines(&x86, FloatKind::Float64x, FloatKind::Float128, FloatKind::Float128);
        combines(&x86, FloatKind::Double, FloatKind::LongDouble, FloatKind::LongDouble);
        combines(&x86, FloatKind::Float32x, FloatKind::Float64, FloatKind::Float64);
        combines(&x86, FloatKind::Float64x, FloatKind::Float64, FloatKind::Float64x);
        combines(&x86, FloatKind::LongDouble, FloatKind::Float64, FloatKind::LongDouble);
    }

    #[test]
    fn the_widest_floating_type_is_a_question_about_the_target_and_not_about_the_names() {
        // The same reading against gcc 16 on aarch64-apple-darwin, where `long double` is a
        // `double` and loses to the `_Float64x` it beats on x86-64. The name says nothing about
        // which of the two is wider, which is why the ordering is worked out from the formats.
        let mac = target("aarch64-apple-darwin");
        combines(&mac, FloatKind::LongDouble, FloatKind::Float64x, FloatKind::Float64x);
        combines(&mac, FloatKind::Double, FloatKind::LongDouble, FloatKind::LongDouble);
        combines(&mac, FloatKind::Float128, FloatKind::LongDouble, FloatKind::Float128);
        combines(&mac, FloatKind::Float64x, FloatKind::Float64, FloatKind::Float64x);
        combines(&mac, FloatKind::Float32x, FloatKind::Float32, FloatKind::Float32x);
        combines(&mac, FloatKind::Float32x, FloatKind::Float64, FloatKind::Float64);
        combines(&mac, FloatKind::Double, FloatKind::Float64, FloatKind::Float64);
        // `_Float16` is the narrowest type there is and does not promote on the way in, so it
        // survives an operation only when nothing wider is there.
        combines(&mac, FloatKind::Float16, FloatKind::Float, FloatKind::Float);
        combines(&mac, FloatKind::Float16, FloatKind::Double, FloatKind::Double);
        combines(&mac, FloatKind::Float16, FloatKind::Float16, FloatKind::Float16);
    }

    #[test]
    fn a_complex_operand_makes_the_answer_complex_after_the_real_types_have_combined() {
        let mut types = Types::new();
        let linux = linux();
        let cfloat = types.complex_float(FloatKind::Float);
        let cdouble = types.complex_float(FloatKind::Double);
        let cldouble = types.complex_float(FloatKind::LongDouble);
        let double = types.float(FloatKind::Double);
        let long_double = types.float(FloatKind::LongDouble);
        let float = types.float(FloatKind::Float);
        let int = types.int(IntKind::Int);

        assert_eq!(usual_arithmetic(&mut types, cfloat, double, &linux), Some(cdouble));
        assert_eq!(usual_arithmetic(&mut types, cfloat, int, &linux), Some(cfloat));
        assert_eq!(usual_arithmetic(&mut types, cdouble, long_double, &linux), Some(cldouble));
        assert_eq!(usual_arithmetic(&mut types, cfloat, float, &linux), Some(cfloat));
    }

    #[test]
    fn an_operand_that_is_not_arithmetic_has_no_common_type() {
        // The caller is the one holding the span, so this says no rather than guessing.
        let mut types = Types::new();
        let linux = linux();
        let int = types.int(IntKind::Int);
        let pointer = types.pointer(int);
        assert_eq!(usual_arithmetic(&mut types, pointer, int, &linux), None);
        assert_eq!(usual_arithmetic(&mut types, pointer, pointer, &linux), None);
        let void = types.void();
        assert_eq!(usual_arithmetic(&mut types, void, int, &linux), None);
        // And a type that is not arithmetic is still its own promotion, so a caller may promote
        // first and ask questions afterwards.
        assert_eq!(promote(&mut types, pointer, &linux), pointer);
    }

    #[test]
    fn the_conversions_read_through_sugar() {
        let mut interner = Interner::new();
        let mut types = Types::new();
        let linux = linux();
        let char_ = types.int(IntKind::Char);
        let name = types.typedef(interner.intern("byte"), char_);
        let int = types.int(IntKind::Int);
        assert_eq!(promote(&mut types, name, &linux), int);
    }

    /// A prototype returning `void`.
    fn prototype(types: &mut Types, params: Vec<TypeId>, variadic: bool) -> TypeId {
        let ret = types.void();
        types.function(FunctionType { ret, params, variadic, prototyped: true })
    }

    /// `void f()` as it means before C23: a declaration that says nothing about the parameters.
    fn old_style(types: &mut Types) -> TypeId {
        let ret = types.void();
        types.function(FunctionType { ret, params: Vec::new(), variadic: false, prototyped: false })
    }

    /// A complete record with the given tag and members.
    fn tagged(types: &mut Types, tag: Symbol, fields: &[FieldDecl]) -> RecordId {
        let id = types.declare_record(RecordKind::Struct, Some(tag));
        let laid_out = lay_out(types, RecordKind::Struct, fields);
        types.complete_record(id, laid_out);
        id
    }

    #[test]
    fn a_type_is_compatible_with_itself_however_it_was_written() {
        let mut interner = Interner::new();
        let mut types = Types::new();
        let int = types.int(IntKind::Int);
        let name = types.typedef(interner.intern("int32_t"), int);
        assert!(compatible(&types, name, int), "the sugar is the same type underneath");
        assert_eq!(composite(&mut types, name, int), Some(name), "and it keeps its name");

        // The qualifiers have to match exactly, which is what keeps `const int *` and `int *`
        // apart as parameter types.
        let konst = types.qualified(int, Qualifiers::CONST);
        assert!(!compatible(&types, konst, int));
        let konst_pointer = types.pointer(konst);
        let pointer = types.pointer(int);
        assert!(!compatible(&types, konst_pointer, pointer));
        assert_eq!(composite(&mut types, konst_pointer, pointer), None);

        // And a different type is a different type. `char` is not `signed char` even on a target
        // where the two have the same range, which is why they are separate kinds here.
        let char_ = types.int(IntKind::Char);
        let schar = types.int(IntKind::SChar);
        assert!(!compatible(&types, char_, schar));
        // `_Atomic int` is not `int` either, since it is a type and not a qualifier.
        let atomic = types.atomic(int);
        assert!(!compatible(&types, atomic, int));
    }

    #[test]
    fn an_enumeration_is_compatible_with_the_type_it_is_represented_in() {
        // gcc 13.3 and clang 18 both represent `enum E { A, B }` in `unsigned int`, and both
        // accept a redeclaration that writes the representation instead of the tag.
        let mut types = Types::new();
        let uint = types.int(IntKind::UInt);
        let int = types.int(IntKind::Int);
        let id = types.declare_enum(None);
        types.complete_enum(id, uint, false);
        let e = types.enumeration(id);
        assert!(compatible(&types, e, uint));
        assert!(compatible(&types, uint, e), "and the relation is symmetric");
        assert!(!compatible(&types, e, int));

        // Two enumeration declarations are two types. Each is compatible with what it is
        // represented in, and that does not make them compatible with each other.
        let other = types.declare_enum(None);
        types.complete_enum(other, uint, false);
        let other = types.enumeration(other);
        assert!(!compatible(&types, e, other));

        // One nobody has decided on yet is compatible with nothing but itself, because the
        // answer is not known rather than no.
        let undecided = types.declare_enum(None);
        let undecided = types.enumeration(undecided);
        assert!(!compatible(&types, undecided, uint));
        assert!(compatible(&types, undecided, undecided));
    }

    #[test]
    fn an_array_without_a_size_is_compatible_with_one_that_has_it() {
        // `extern int a[]; int a[4];` is a complete array of four afterwards, which gcc reports
        // as a `sizeof` of sixteen. A compiler that keeps the first type has lost the size.
        let mut types = Types::new();
        let int = types.int(IntKind::Int);
        let unknown = types.array(int, ArrayLen::Unknown);
        let four = types.array(int, ArrayLen::Fixed(4));
        let five = types.array(int, ArrayLen::Fixed(5));
        assert!(compatible(&types, unknown, four));
        assert!(!compatible(&types, four, five));
        assert_eq!(composite(&mut types, unknown, four), Some(four));
        assert_eq!(composite(&mut types, four, unknown), Some(four), "either way round");
        assert_eq!(composite(&mut types, four, five), None);

        // A variable length array is compatible with both, because its size is not something a
        // declaration can be checked against.
        let vla = types.array(int, ArrayLen::Variable(VlaId(0)));
        assert!(compatible(&types, vla, four));
        assert_eq!(composite(&mut types, vla, four), Some(four));

        // The element types have to be compatible too, and the composite reaches into them.
        let long = types.int(IntKind::Long);
        let longs = types.array(long, ArrayLen::Fixed(4));
        assert!(!compatible(&types, four, longs));
    }

    #[test]
    fn a_parameter_declared_as_an_array_is_a_pointer() {
        // `int fn(int p[3])` and `int fn(int *p)` are one declaration and one definition, which
        // both compilers accept. The adjustment is part of forming the parameter type, so two
        // functions written either way are not merely compatible but identical.
        let mut types = Types::new();
        let int = types.int(IntKind::Int);
        let three = types.array(int, ArrayLen::Fixed(3));
        let pointer = types.pointer(int);
        assert_eq!(adjust_parameter(&mut types, three), pointer);

        // A function parameter becomes a pointer to the function the same way.
        let function = prototype(&mut types, vec![int], false);
        let function_pointer = types.pointer(function);
        assert_eq!(adjust_parameter(&mut types, function), function_pointer);

        // And the qualifiers on the outermost node go, so `void f(const int)` and `void f(int)`
        // declare the same function. The pointee of a `const int *` keeps its own.
        let konst = types.qualified(int, Qualifiers::CONST);
        assert_eq!(adjust_parameter(&mut types, konst), int);
        let to_konst = types.pointer(konst);
        assert_eq!(adjust_parameter(&mut types, to_konst), to_konst);
    }

    #[test]
    fn an_old_style_declaration_is_compatible_with_the_prototypes_a_call_could_not_tell_from_it() {
        // Measured with gcc 13.3 in C17 mode, which is the compiler that still has the old
        // meaning of `()`. It names the rule in its own diagnostic: an argument type that has a
        // default promotion cannot match an empty parameter name list declaration.
        let mut types = Types::new();
        let old = old_style(&mut types);
        let int = types.int(IntKind::Int);
        let long = types.int(IntKind::Long);
        let char_ = types.int(IntKind::Char);
        let float = types.float(FloatKind::Float);
        let double = types.float(FloatKind::Double);

        let takes_int = prototype(&mut types, vec![int], false);
        assert!(compatible(&types, old, takes_int));
        assert!(compatible(&types, takes_int, old), "and the relation is symmetric");
        // The composite is the prototype, so the calls written before it can still be checked.
        assert_eq!(composite(&mut types, old, takes_int), Some(takes_int));

        let pointer = types.pointer(int);
        for params in [vec![long], vec![double], vec![pointer], vec![int, long]] {
            let ty = prototype(&mut types, params, false);
            assert!(compatible(&types, old, ty), "nothing here is touched by a promotion");
        }

        // A `char` promotes to `int` and a `float` to `double`, so a call through the old style
        // declaration would have passed something else and the two conflict.
        for params in [vec![char_], vec![float], vec![int, char_]] {
            let ty = prototype(&mut types, params, false);
            assert!(!compatible(&types, old, ty));
            assert_eq!(composite(&mut types, old, ty), None);
        }

        // An ellipsis conflicts too, which gcc also says in as many words.
        let variadic = prototype(&mut types, vec![int], true);
        assert!(!compatible(&types, old, variadic));

        // An enumeration parameter comes through when what it is represented in does.
        let uint = types.int(IntKind::UInt);
        let id = types.declare_enum(None);
        types.complete_enum(id, uint, false);
        let e = types.enumeration(id);
        let takes_enum = prototype(&mut types, vec![e], false);
        assert!(compatible(&types, old, takes_enum));

        // Two old style declarations agree about nothing and so cannot disagree.
        assert!(compatible(&types, old, old));

        // The return type still has to match, which is the one part `()` does say.
        let returns_int = types.function(FunctionType {
            ret: int,
            params: Vec::new(),
            variadic: false,
            prototyped: false,
        });
        assert!(!compatible(&types, returns_int, takes_int));
    }

    #[test]
    fn from_c23_an_empty_parameter_list_is_a_prototype_and_conflicts_where_it_used_to_merge() {
        // The dialect decides what `()` means and the parser records the decision, so the same
        // pair of declarations is a redeclaration in C17 and a conflict in C23. Both compilers
        // report exactly that.
        let mut types = Types::new();
        let int = types.int(IntKind::Int);
        let takes_int = prototype(&mut types, vec![int], false);
        let takes_nothing = prototype(&mut types, Vec::new(), false);
        let old = old_style(&mut types);
        assert!(!compatible(&types, takes_nothing, takes_int));
        assert!(compatible(&types, old, takes_int), "the C17 reading of the same source");
    }

    #[test]
    fn two_prototypes_have_to_agree_about_everything() {
        let mut types = Types::new();
        let int = types.int(IntKind::Int);
        let long = types.int(IntKind::Long);
        let base = prototype(&mut types, vec![int, int], false);
        for other in [vec![int], vec![int, long], vec![int, int, int], Vec::new()] {
            let other = prototype(&mut types, other, false);
            assert!(!compatible(&types, base, other));
        }
        let variadic = prototype(&mut types, vec![int, int], true);
        assert!(!compatible(&types, base, variadic), "`...` is part of the type");

        // The parameters are compared with the same rules as anything else, so an array size
        // inside a parameter's type is compared and an unknown one is not.
        let four = types.array(int, ArrayLen::Fixed(4));
        let unknown = types.array(int, ArrayLen::Unknown);
        let to_four = types.pointer(four);
        let to_unknown = types.pointer(unknown);
        let a = prototype(&mut types, vec![to_four], false);
        let b = prototype(&mut types, vec![to_unknown], false);
        assert!(compatible(&types, a, b));
        // And the composite takes the size, which is the whole reason it exists.
        assert_eq!(composite(&mut types, a, b), Some(a));
    }

    #[test]
    fn a_pointer_composite_reaches_through_to_what_is_pointed_at() {
        let mut types = Types::new();
        let int = types.int(IntKind::Int);
        let four = types.array(int, ArrayLen::Fixed(4));
        let unknown = types.array(int, ArrayLen::Unknown);
        let to_four = types.pointer(four);
        let to_unknown = types.pointer(unknown);
        assert_eq!(composite(&mut types, to_unknown, to_four), Some(to_four));

        // The pointer's own qualifiers survive, since a compatible pair has the same ones.
        let konst_to_unknown = types.qualified(to_unknown, Qualifiers::CONST);
        let konst_to_four = types.qualified(to_four, Qualifiers::CONST);
        assert_eq!(composite(&mut types, konst_to_unknown, konst_to_four), Some(konst_to_four));
    }

    #[test]
    fn two_record_declarations_with_the_same_tag_and_the_same_members_are_compatible() {
        // C23 6.2.7p1, which is what lets one header be included twice. clang 18 implements it
        // and gcc 13.3 still rejects the redefinition, so this is a divergence rather than a
        // reading; in the older dialects the redefinition never gets as far as being compared.
        let mut interner = Interner::new();
        let mut types = Types::new();
        let tag = interner.intern("point");
        let x = interner.intern("x");
        let y = interner.intern("y");
        let int = types.int(IntKind::Int);
        let members = [FieldDecl::new(Some(x), int), FieldDecl::new(Some(y), int)];

        let first = tagged(&mut types, tag, &members);
        let second = tagged(&mut types, tag, &members);
        let first = types.record(first);
        let second = types.record(second);
        assert_ne!(first, second, "still two declarations and two types");
        assert!(compatible(&types, first, second));

        // A different member name, a different member type, a different count, a different tag
        // and a different keyword are each enough to make them different types.
        let z = interner.intern("z");
        let long = types.int(IntKind::Long);
        let renamed = [FieldDecl::new(Some(x), int), FieldDecl::new(Some(z), int)];
        let retyped = [FieldDecl::new(Some(x), int), FieldDecl::new(Some(y), long)];
        for other in [&renamed[..], &retyped[..], &members[..1]] {
            let other = tagged(&mut types, tag, other);
            let other = types.record(other);
            assert!(!compatible(&types, first, other));
        }
        let elsewhere = tagged(&mut types, interner.intern("pair"), &members);
        let elsewhere = types.record(elsewhere);
        assert!(!compatible(&types, first, elsewhere));

        // An anonymous record is compatible with nothing but itself: there is no name by which
        // a second declaration could be claiming to be the same type.
        let anonymous = record(&mut types, RecordKind::Struct, &members);
        let also_anonymous = record(&mut types, RecordKind::Struct, &members);
        assert!(!compatible(&types, anonymous, also_anonymous));

        // Nor is an incomplete declaration, which has no members to compare.
        let incomplete = types.declare_record(RecordKind::Struct, Some(tag));
        let incomplete = types.record(incomplete);
        assert!(!compatible(&types, first, incomplete));
        assert!(compatible(&types, incomplete, incomplete));
    }

    #[test]
    fn a_self_referential_record_is_compared_without_going_round_forever() {
        // `struct node { int value; struct node *next; }` declared twice. Comparing the two
        // reaches the same pair again through the pointer, and the second time it is an
        // assumption rather than a question.
        let mut interner = Interner::new();
        let mut types = Types::new();
        let tag = interner.intern("node");
        let value = interner.intern("value");
        let next = interner.intern("next");
        let int = types.int(IntKind::Int);

        let node = |types: &mut Types| {
            let id = types.declare_record(RecordKind::Struct, Some(tag));
            let ty = types.record(id);
            let pointer = types.pointer(ty);
            let members = [FieldDecl::new(Some(value), int), FieldDecl::new(Some(next), pointer)];
            let laid_out = lay_out(types, RecordKind::Struct, &members);
            types.complete_record(id, laid_out);
            ty
        };
        let first = node(&mut types);
        let second = node(&mut types);
        assert_ne!(first, second);
        assert!(compatible(&types, first, second));

        // The guard is an assumption and not an answer, so a difference below the cycle is still
        // found: the same structure with the two members the other way round is a different one.
        let id = types.declare_record(RecordKind::Struct, Some(tag));
        let ty = types.record(id);
        let pointer = types.pointer(ty);
        let members = [FieldDecl::new(Some(next), pointer), FieldDecl::new(Some(value), int)];
        let laid_out = lay_out(&types, RecordKind::Struct, &members);
        types.complete_record(id, laid_out);
        assert!(!compatible(&types, first, ty));
    }

    #[test]
    fn layout_reads_through_sugar() {
        let mut interner = Interner::new();
        let mut types = Types::new();
        let long = types.int(IntKind::Long);
        let name = types.typedef(interner.intern("word"), long);
        let array = types.array(name, ArrayLen::Fixed(4));
        assert_eq!(layout(&types, array, &linux()).unwrap(), Layout::new(32, 8));
    }

    /// A recipe worked out, with `sizes` standing for how large each member turned out to be.
    ///
    /// The lowering does this with instructions and this does it with numbers, which is what
    /// makes a recipe testable here: the tree is the whole answer, and what a member is as long
    /// as is the only thing either side has to be told.
    fn work_out(recipe: &Extent, sizes: &[u64]) -> u64 {
        match recipe {
            Extent::Bytes(count) => *count,
            Extent::Member(index) => sizes[*index as usize],
            Extent::Sum(parts) => parts.iter().map(|part| work_out(part, sizes)).sum(),
            Extent::RoundUp(inner, to) => work_out(inner, sizes).next_multiple_of(*to),
            Extent::Max(parts) => parts.iter().map(|part| work_out(part, sizes)).max().unwrap_or(0),
        }
    }

    /// An array of `int` whose length the program computes.
    fn measured(types: &mut Types, which: u32) -> TypeId {
        let int = types.int(IntKind::Int);
        types.array(int, ArrayLen::Variable(VlaId(which)))
    }

    #[test]
    fn a_member_of_no_fixed_size_leaves_the_size_and_what_follows_it_to_the_program() {
        let mut types = Types::new();
        let int = types.int(IntKind::Int);
        let rows = measured(&mut types, 0);
        let laid_out = lay_out(&types, RecordKind::Struct, &[member(rows), member(int)]);

        // Nothing is known here but the alignment, which never varies: it is decided by the
        // members rather than by where they land.
        assert_eq!(laid_out.layout, Layout::new(0, 4));
        let variable = laid_out.variable.expect("a record with a member of no fixed size");
        // The member in front of the variable one sits where its number says, and the one after
        // it does not. There is no rounding in between, because an array of `int` ends on a four
        // byte boundary however long it is.
        assert_eq!(variable.offsets[0], None);
        let after = variable.offsets[1].as_ref().expect("an offset the program works out");
        for count in 0..6u64 {
            let sizes = [4 * count, 4];
            assert_eq!(work_out(after, &sizes), 4 * count);
            assert_eq!(work_out(&variable.size, &sizes), 4 * count + 4);
        }
        assert_eq!(laid_out.fields[1].align, 4);
    }

    #[test]
    fn a_member_after_one_of_no_fixed_size_is_rounded_up_where_its_alignment_asks_for_it() {
        let mut types = Types::new();
        let char_ty = types.int(IntKind::Char);
        let rows = measured(&mut types, 0);
        let double = types.float(FloatKind::Double);
        let fields = [member(char_ty), member(rows), member(double)];
        let laid_out = lay_out(&types, RecordKind::Struct, &fields);

        assert_eq!(laid_out.layout, Layout::new(0, 8));
        assert_eq!(offsets(&laid_out)[..2], [0, 32]);
        let variable = laid_out.variable.expect("a record with a member of no fixed size");
        assert_eq!(variable.offsets[..2], [None, None]);
        let after = variable.offsets[2].as_ref().expect("an offset the program works out");
        for count in 0..6u64 {
            let sizes = [1, 4 * count, 8];
            let at = (4 + 4 * count).next_multiple_of(8);
            assert_eq!(work_out(after, &sizes), at);
            assert_eq!(work_out(&variable.size, &sizes), at + 8);
        }
    }

    #[test]
    fn a_union_with_a_member_of_no_fixed_size_is_as_long_as_the_longest_of_them() {
        let mut types = Types::new();
        let rows = measured(&mut types, 0);
        let double = types.float(FloatKind::Double);
        let laid_out = lay_out(&types, RecordKind::Union, &[member(rows), member(double)]);

        assert_eq!(laid_out.layout, Layout::new(0, 8));
        let variable = laid_out.variable.expect("a union with a member of no fixed size");
        // Every member of a union starts where the union does, so none of them has an offset the
        // program has to work out.
        assert!(variable.offsets.iter().all(Option::is_none));
        for count in 0..6u64 {
            let sizes = [4 * count, 8];
            let want = (4 * count).max(8).next_multiple_of(8);
            assert_eq!(work_out(&variable.size, &sizes), want);
        }
    }

    #[test]
    fn packed_takes_the_rounding_out_of_a_record_the_program_measures() {
        let mut types = Types::new();
        let char_ty = types.int(IntKind::Char);
        let int = types.int(IntKind::Int);
        let rows = measured(&mut types, 0);
        let fields = [member(char_ty), member(rows), member(int)];
        let options = RecordOptions { packed: true, ..RecordOptions::default() };
        let laid_out = layout_record(&types, RecordKind::Struct, &fields, &options, &linux())
            .expect("a packed record with a member of no fixed size");

        assert_eq!(laid_out.layout, Layout::new(0, 1));
        assert_eq!(laid_out.fields[2].align, 1);
        let variable = laid_out.variable.expect("a record with a member of no fixed size");
        let after = variable.offsets[2].as_ref().expect("an offset the program works out");
        for count in 0..6u64 {
            let sizes = [1, 4 * count, 4];
            assert_eq!(work_out(after, &sizes), 1 + 4 * count);
            assert_eq!(work_out(&variable.size, &sizes), 1 + 4 * count + 4);
        }
    }

    #[test]
    fn a_bit_field_after_a_member_of_no_fixed_size_is_turned_down() {
        let mut interner = Interner::new();
        let mut types = Types::new();
        let int = types.int(IntKind::Int);
        let rows = measured(&mut types, 0);
        let fields = [member(rows), bits(&mut interner, "b", int, 3)];
        let options = RecordOptions::default();
        let failed = layout_record(&types, RecordKind::Struct, &fields, &options, &linux());

        // Which unit such a bit-field lands in is a question about an address the program has
        // not worked out yet, and gcc answers it while the program runs. Nothing here does, so
        // the layout says so rather than putting the member somewhere plausible.
        assert_eq!(failed, Err(RecordError::VariableBitField { index: 1 }));
    }
}