Skip to main content

rucc_object/
file.rs

1//! Relocatable objects, in whichever of the formats the target wants.
2//!
3//! Design: `spec/11-asm-objects-debug.md` section 11.3, which says the three formats are written
4//! through the [`object`] crate's writer with our own layer above it for the parts it does not
5//! model. This is that layer, and what it holds is the part `object` cannot decide: which
6//! relocation an instruction wants, what a symbol's binding and type are, and the sections a
7//! linker expects to find whether or not anything was put in them.
8//!
9//! # One layout and two sets of answers
10//!
11//! Which sections a file has, what goes in each of them, which symbol says where each thing is and
12//! what each relocation is against are the same questions for ELF and for COFF, and they have the
13//! same answers, so they are asked once here. What differs is a short list: the number a relocation
14//! is, the field a visibility goes in, the note saying what the file was built to have checked, and
15//! the marker whose absence makes the stack executable. [`Flavour`] is that list, and the answers
16//! are in [`crate::elf`] and [`crate::coff`] beside each other where they can be read against one
17//! another.
18//!
19//! The alternative was two writers, and the reason against it is what a second copy of a layout
20//! decays into: a fix to one of them is a fix to one platform, and which platform got it is
21//! whichever the person who found the bug was building for.
22//!
23//! # What is not here
24//!
25//! Mach-O. The formats disagree about more than their headers: an Apple symbol carries an
26//! underscore in front of the C name and Mach-O has no way to say how long a function is, wanting
27//! `.subsections_via_symbols` instead. It is written when the target that needs it is.
28//!
29//! Thread-local storage. Reaching a thread-local variable is a different instruction sequence per
30//! model and the back end writes none of them, so a module carrying one is refused before it
31//! reaches here rather than written as an ordinary variable in the wrong section.
32
33use std::collections::{BTreeMap, HashMap, HashSet};
34
35use object::write::{
36    Object as Writer, Relocation, StandardSection, Symbol, SymbolId, SymbolSection,
37};
38use object::{
39    Architecture, BinaryFormat, Endianness, RelocationFlags, SectionFlags, SectionKind,
40    SymbolFlags, SymbolKind, SymbolScope,
41};
42use rucc_target::{ObjectFormat, TargetInfo};
43use rucc_tuple::Arch;
44
45use crate::section::{
46    Alias, Apart, Array, Binding, Data, Info, Object, Output, Place, Property, Reference, Reloc,
47    Sections, Text, Visibility,
48};
49use crate::{coff, elf};
50
51/// Which of the two formats is being written, and therefore which set of answers the questions this
52/// module cannot decide get.
53///
54/// A short list rather than a trait, because the list is short and closed: everything a format has
55/// an opinion about is a call to one of the methods below, so adding Mach-O is adding a third arm to
56/// each of them and the compiler names every one that was forgotten.
57#[derive(Debug, Clone, Copy, PartialEq, Eq)]
58pub(crate) enum Flavour {
59    /// Linux, the BSDs and the freestanding targets.
60    Elf,
61    /// Windows, under either of its two runtimes.
62    Coff,
63}
64
65impl Flavour {
66    /// Which one a target wants, and nothing for the two formats that are not written.
67    pub(crate) fn of(target: &TargetInfo) -> Option<Flavour> {
68        match target.object_format {
69            ObjectFormat::Elf => Some(Flavour::Elf),
70            ObjectFormat::Coff => Some(Flavour::Coff),
71            ObjectFormat::MachO | ObjectFormat::Wasm => None,
72        }
73    }
74
75    /// The format the writer underneath is asked for.
76    pub(crate) fn binary(self) -> BinaryFormat {
77        match self {
78            Flavour::Elf => BinaryFormat::Elf,
79            Flavour::Coff => BinaryFormat::Coff,
80        }
81    }
82
83    /// Which relocation this reference is, or `None` for one this format has none of.
84    ///
85    /// `after` is how many bytes of the instruction come after the four the linker writes over,
86    /// which ELF has already folded into the addend and COFF wants told apart. See [`crate::Reloc`].
87    pub(crate) fn reloc(self, reference: Reference, after: u8) -> Option<RelocationFlags> {
88        match self {
89            Flavour::Elf => elf::r_type(reference).map(|r_type| RelocationFlags::Elf { r_type }),
90            Flavour::Coff => coff::reloc(reference, after),
91        }
92    }
93
94    /// Say how far a name reaches beyond what its scope already said.
95    ///
96    /// Nothing on COFF, where a symbol has nowhere to keep it. A file built with
97    /// `-fvisibility=hidden` for Windows is a file where that flag changed nothing, which is what
98    /// gcc does there as well.
99    pub(crate) fn see(
100        self,
101        obj: &mut Writer<'_>,
102        id: SymbolId,
103        binding: Binding,
104        visibility: Visibility,
105    ) {
106        match self {
107            Flavour::Elf => elf::see(obj, id, binding, visibility),
108            Flavour::Coff => {}
109        }
110    }
111
112    /// The section a variable the loader writes into before anything reads it goes in, when the
113    /// program asked for the half of it the linker keeps apart, or nothing for a format that has no
114    /// such half and puts one in ordinary read only data with the rest.
115    fn rel_ro_local(self) -> Option<&'static str> {
116        match self {
117            Flavour::Elf => elf::REL_RO_LOCAL,
118            Flavour::Coff => coff::REL_RO_LOCAL,
119        }
120    }
121
122    /// The type and flags a section of function addresses the startup code calls has, where the
123    /// format has something to say about it.
124    ///
125    /// Nothing on COFF, where such a section is refused by [`beyond`] before it reaches here rather
126    /// than written under a name nothing on that platform gathers.
127    fn gathered(self, array: Array) -> Option<SectionFlags> {
128        match self {
129            Flavour::Elf => Some(elf::gathered(array)),
130            Flavour::Coff => None,
131        }
132    }
133
134    /// The header fields a file of assembly stated about one of its own sections, where the format
135    /// has fields to put them in.
136    ///
137    /// ELF has one for each of the letters, so what the source wrote is written down as it stands
138    /// and the section kind handed to the writer alongside is only a summary of it. COFF has no
139    /// field the letters map onto one for one, and the characteristics the writer works out from
140    /// that kind are the ones every other Windows assembler produces, so there is nothing to add and
141    /// saying so is [`None`] rather than a word built out of guesses.
142    pub(crate) fn stated(self, shape: crate::source::Shape) -> Option<SectionFlags> {
143        match self {
144            Flavour::Elf => {
145                Some(SectionFlags::Elf { sh_type: shape.sh_type(), sh_flags: shape.sh_flags() })
146            }
147            Flavour::Coff => None,
148        }
149    }
150
151    /// What kind of symbol a name out of a file of assembly is, given what `.type` said about it and
152    /// how far it reaches.
153    ///
154    /// The binding is a parameter because on COFF the two are not separable. ELF keeps the type and
155    /// the binding in different halves of a byte, so a name that nothing stated a type for is
156    /// `STT_NOTYPE` whether it is local or global, and that is what gas writes for a plain label.
157    /// COFF has no type field of that sort: what the writer underneath calls a label is storage
158    /// class `LABEL`, which is a name inside this file and nothing a linker will resolve against, so
159    /// a `.globl` with no `.type` under it would quietly stop being offered. The kind with no
160    /// function type on it and an external storage class is the data one, which is what gas for this
161    /// platform writes for the same input, so that is what an untyped global becomes here.
162    pub(crate) fn sort(self, sort: crate::source::Sort, binding: Binding) -> SymbolKind {
163        match sort {
164            crate::source::Sort::Func => SymbolKind::Text,
165            crate::source::Sort::Object => SymbolKind::Data,
166            crate::source::Sort::Thread => SymbolKind::Tls,
167            crate::source::Sort::File => SymbolKind::File,
168            crate::source::Sort::Untyped => match (self, binding) {
169                (Flavour::Coff, Binding::Global | Binding::Weak) => SymbolKind::Data,
170                _ => SymbolKind::Label,
171            },
172        }
173    }
174
175    /// The marker a linker looks for in every input, where there is one.
176    pub(crate) fn marker(self, obj: &mut Writer<'_>) {
177        match self {
178            Flavour::Elf => elf::marker(obj),
179            Flavour::Coff => coff::marker(obj),
180        }
181    }
182
183    /// What the file says it was built to have checked, where the format has a way to say it.
184    ///
185    /// ELF writes a note the linker keeps only the agreed part of. A PE image says the same thing in
186    /// the header of the finished image rather than in its inputs, so an object carries nothing and
187    /// the instructions the flag asked for are in the text either way.
188    fn property(self, obj: &mut Writer<'_>, property: Property) {
189        if !property.any() {
190            return;
191        }
192        match self {
193            Flavour::Elf => {
194                let note = obj.section_id(StandardSection::GnuProperty);
195                obj.append_section_data(note, &elf::record(property), 8);
196            }
197            Flavour::Coff => {}
198        }
199    }
200
201    /// Where the unwind table goes: the section the records are in and what it is aligned to, and
202    /// the second section holding what those records point at, on the format that keeps the two
203    /// apart.
204    fn tables(self) -> ((&'static str, u64), Option<(&'static str, u64)>) {
205        match self {
206            Flavour::Elf => (elf::FRAMES, None),
207            Flavour::Coff => (coff::FUNCTIONS, Some(coff::CODES)),
208        }
209    }
210
211    /// Anything that has to be written into the finished bytes rather than said to the writer.
212    fn finish(self, bytes: &mut [u8], ordered: &[String]) {
213        match self {
214            Flavour::Elf => elf::link(bytes, ordered),
215            Flavour::Coff => debug_assert!(ordered.is_empty(), "a record this format cannot write"),
216        }
217    }
218}
219
220/// Why an object file could not be written.
221#[derive(Debug, Clone, PartialEq, Eq)]
222pub enum Error {
223    /// A machine or a platform this does not write objects for.
224    Format {
225        /// The triple that was asked for.
226        triple: String,
227    },
228    /// The writer refused something it was given, which is a bug here rather than in a program.
229    Refused {
230        /// What it said, already formatted.
231        why: String,
232    },
233}
234
235impl std::fmt::Display for Error {
236    fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
237        match self {
238            Error::Format { triple } => {
239                write!(f, "there is no object writer for {triple} in this compiler yet")
240            }
241            Error::Refused { why } => {
242                write!(f, "the object writer refused what it was given: {why}")
243            }
244        }
245    }
246}
247
248impl std::error::Error for Error {}
249
250/// One text section and the variables beside it, as a relocatable object in the target's format.
251///
252/// `info` is the debug sections, already encoded, and is empty in a build that asked for none.
253/// What it holds is bytes and relocations for the same reason [`Text::unwind`] is bytes: the
254/// format's answer is the producer's to give and what is left here is where the sections go.
255///
256/// # Errors
257///
258/// [`Error::Format`] for a machine or a platform this does not write, and [`Error::Refused`] for
259/// anything the writer underneath objected to, which would be a bug here. An alias whose target
260/// this file does not define is refused the same way, since the front end is what reports that as
261/// a program's mistake and one reaching here means it did not. So is anything the target's format
262/// has no way to write, which for COFF is a thread-local variable, a reference through a table the
263/// platform does not have, a record of where a patcher's room is and a section the startup code is
264/// expected to gather. See [`Error`].
265pub fn write(
266    text: &Text,
267    data: &Data,
268    aliases: &[Alias],
269    target: &TargetInfo,
270    output: Output,
271    info: &Info,
272) -> Result<Vec<u8>, Error> {
273    let Output { sections, property } = output;
274    let flavour = Flavour::of(target).filter(|_| target.tuple.arch() == Arch::X86_64);
275    let Some(flavour) = flavour else {
276        return Err(Error::Format { triple: target.tuple.to_string() });
277    };
278    if flavour == Flavour::Coff {
279        beyond(text, data, info)?;
280    }
281    let mut obj = Writer::new(flavour.binary(), Architecture::X86_64, Endianness::Little);
282    // The one that holds every function when they are not being split up. Asked for even when it
283    // will stay empty, because it is the section the writer underneath starts a file with anyway
284    // and gcc writes an empty `.text` under `-ffunction-sections` too.
285    let whole = obj.section_id(StandardSection::Text);
286    if !sections.functions {
287        obj.append_section_data(whole, &text.bytes, u64::from(text.align));
288    }
289
290    // Every function defined here, then every variable, then every name either of them wanted that
291    // is not. A name is looked up rather than added twice, because two symbols with one name is
292    // not a file a linker accepts.
293    let mut symbols = BTreeMap::new();
294    // Where each function ended up, in the order they were written, so that a relocation inside
295    // one goes into the section that one is in and one that points at the start of one can be
296    // written against that section. The same list as `text.funcs` and in the same order, so the
297    // two are walked together below.
298    let mut split: Vec<(object::write::SectionId, u64)> = Vec::with_capacity(text.funcs.len());
299    // Which text section each record of where a patcher's room is belongs to, in the order the
300    // records were added, which is the order their headers come out in. See `link`.
301    let mut ordered: Vec<String> = Vec::new();
302    for func in &text.funcs {
303        // A section of its own, holding this function's bytes and nothing else, so the linker can
304        // drop it when nothing reaches it. The name is what gcc writes, and the leading `.text.`
305        // is not decoration: `--gc-sections` and the linker scripts that place code both match on
306        // it, and a section called something else would be placed by the catch all rule.
307        //
308        // The room a patcher was promised in front of the label goes in it too. Those bytes are
309        // the function's, they are just not under its name: the symbol is where the label was and
310        // the room is what came before, so a section holding one without the other would be a
311        // section a linker could place with the room missing.
312        let ahead = func.patch.map_or(0, |patch| patch.before);
313        let (section, at) = if sections.functions {
314            let name = format!(".text.{}", func.name).into_bytes();
315            let id = obj.add_section(Vec::new(), name, SectionKind::Text);
316            let bytes = &text.bytes[func.start - ahead..func.start + func.len];
317            obj.append_section_data(id, bytes, u64::from(func.align.max(1)));
318            (id, ahead as u64)
319        } else {
320            (whole, func.start as u64)
321        };
322        // Where the room is, in a section of its own that says nothing else. What reads it is a
323        // tracer patching every function in an image at once, and what it needs is every address
324        // in one place: a stripped kernel has no symbol table to walk instead, which is the whole
325        // reason the list is written rather than worked out later.
326        //
327        // The address is a relocation rather than a number, because a function is at a fixed
328        // offset in its own section and where that section lands is the linker's answer. It is
329        // written against the section rather than against the function's own name so that it still
330        // points at the room when the room is in front of the name.
331        //
332        // One section per function even when they all point at the same text, which is what gas
333        // produces and what lets a linker throw the record away with the function. `SHF_LINK_ORDER`
334        // is what ties the two together and it needs a section index the writer underneath does not
335        // set, so `link` fills it in afterwards. See `link`.
336        if let Some(patch) = func.patch {
337            let base = if sections.functions { func.start - ahead } else { 0 };
338            let name = elf::PATCHABLE.as_bytes().to_vec();
339            let id = obj.add_section(Vec::new(), name, SectionKind::Data);
340            obj.section_mut(id).flags = elf::ordered();
341            obj.append_section_data(id, &[0; 8], 8);
342            let symbol = obj.section_symbol(section);
343            let flags = flavour.reloc(Reference::Address { bytes: 8 }, 0).ok_or_else(|| {
344                Error::Refused { why: "no relocation holds an address here".to_owned() }
345            })?;
346            obj.add_relocation(
347                id,
348                Relocation { offset: 0, symbol, addend: (patch.at - base) as i64, flags },
349            )
350            .map_err(|why| Error::Refused { why: why.to_string() })?;
351            ordered.push(if sections.functions {
352                format!(".text.{}", func.name)
353            } else {
354                ".text".to_owned()
355            });
356        }
357        let id = obj.add_symbol(Symbol {
358            name: func.name.clone().into_bytes(),
359            value: at,
360            size: func.len as u64,
361            kind: SymbolKind::Text,
362            scope: scope_of(func.binding),
363            weak: func.binding == Binding::Weak,
364            section: SymbolSection::Section(section),
365            flags: SymbolFlags::None,
366        });
367        flavour.see(&mut obj, id, func.binding, func.visibility);
368        symbols.insert(func.name.clone(), id);
369        split.push((section, at));
370    }
371
372    // The places inside a function that have names of their own, which is where a label whose
373    // address an image holds is. After the functions, because the section one goes in is the
374    // section of the function it is inside and that is what the walk above worked out.
375    for label in &text.labels {
376        let after = text.funcs.partition_point(|func| func.start <= label.at);
377        let Some(index) = after.checked_sub(1) else {
378            let why = format!("'{}' is at {} and in front of every function", label.name, label.at);
379            return Err(Error::Refused { why });
380        };
381        let func = &text.funcs[index];
382        let (section, at) = if sections.functions {
383            // From the start of the section rather than from the symbol, which is the same
384            // correction a relocation inside a function gets below.
385            let base = func.start - func.patch.map_or(0, |patch| patch.before);
386            (split[index].0, (label.at - base) as u64)
387        } else {
388            (whole, label.at as u64)
389        };
390        let id = obj.add_symbol(Symbol {
391            name: label.name.clone().into_bytes(),
392            value: at,
393            // A label has no length. What is at it is the rest of the function, and a size here
394            // would be a claim that the bytes after it are a thing of their own.
395            size: 0,
396            kind: SymbolKind::Label,
397            // Never offered to another file. The name is one the compiler minted and what it
398            // points at is the middle of a function, so the only thing that resolves against it
399            // is the image in this same file that asked for it.
400            scope: SymbolScope::Compilation,
401            weak: false,
402            section: SymbolSection::Section(section),
403            flags: SymbolFlags::None,
404        });
405        symbols.insert(label.name.clone(), id);
406    }
407
408    // Where each variable's image landed in the section it went into, kept because a relocation in
409    // an image counts from the start of the image and one in a file counts from the start of the
410    // section. A variable that is not in a section has no entry, since nothing in a merged one can
411    // hold a relocation: the linker is being asked for zeroed space rather than for an image.
412    let mut placed = Vec::with_capacity(data.objects.len());
413    // The sections the writer has no name of its own for, remembered by name so that every variable
414    // that wants one lands in the same one. The rest come back from `section_id`, which already
415    // answers with the section it made the first time it was asked.
416    let mut named = HashMap::new();
417    for object in &data.objects {
418        let (section, offset) = put(&mut obj, object, &mut named, sections, flavour);
419        let id = obj.add_symbol(Symbol {
420            name: object.name.clone().into_bytes(),
421            // A common symbol says what it wants rather than where it is, and what it wants is
422            // recorded where an ordinary symbol records its address.
423            value: if object.place == Place::Merged { object.align } else { offset },
424            size: object.size,
425            // A thread-local variable is a different kind of symbol rather than a symbol in a
426            // different section, and it has to be both: the kind is what a linker checks a
427            // relocation against, so a `R_X86_64_PC32` aimed at one is refused rather than
428            // resolved to an address that would have been one thread's and is nobody's.
429            kind: match object.place {
430                Place::Thread { .. } => SymbolKind::Tls,
431                _ => SymbolKind::Data,
432            },
433            scope: scope_of(object.binding),
434            weak: object.binding == Binding::Weak,
435            section,
436            flags: SymbolFlags::None,
437        });
438        flavour.see(&mut obj, id, object.binding, object.visibility);
439        symbols.insert(object.name.clone(), id);
440        placed.push((section.id(), offset));
441    }
442
443    // The distances between two labels, written into the images just placed. Both labels were
444    // added above with the section they are in and where in it, so the distance is the one value
445    // less the other, and it is a number only when the section is the same one.
446    for apart in &data.apart {
447        let (Some(section), offset) = placed[apart.object] else { continue };
448        let value = distance(&obj, &symbols, apart)?;
449        let bytes = usize::from(apart.bytes);
450        let at = usize::try_from(offset).map_err(|why| Error::Refused { why: why.to_string() })?;
451        let at = at + apart.at;
452        let image = obj.section_mut(section).data_mut();
453        image[at..at + bytes].copy_from_slice(&value.to_le_bytes()[..bytes]);
454    }
455
456    // A second name for something already added, which is where the alias's own binding is the
457    // only thing it does not take from what it points at: the target of one may be a `static` and
458    // the alias of it may not be. Before the loop below rather than after it, because a reference
459    // to the new name is a reference to something this file defines and would otherwise be added
460    // as a name this file wants from somewhere else.
461    for alias in aliases {
462        let Some(&id) = symbols.get(&alias.target) else {
463            let why =
464                format!("'{}' is aliased to '{}', which is not here", alias.name, alias.target);
465            return Err(Error::Refused { why });
466        };
467        let (value, size) = (obj.symbol(id).value, obj.symbol(id).size);
468        let (kind, section) = (obj.symbol(id).kind, obj.symbol(id).section);
469        let id = obj.add_symbol(Symbol {
470            name: alias.name.clone().into_bytes(),
471            value,
472            size,
473            kind,
474            scope: scope_of(alias.binding),
475            weak: alias.binding == Binding::Weak,
476            section,
477            flags: SymbolFlags::None,
478        });
479        flavour.see(&mut obj, id, alias.binding, alias.visibility);
480        symbols.insert(alias.name.clone(), id);
481    }
482
483    // Not the unwind table's, which name functions this file defines and are written against the
484    // section rather than against the name. A record for anything else is refused below, so a name
485    // added here for one would be a name nothing goes on to use.
486    // The names a declaration wrote `weak` on, which the link is allowed to leave undefined and
487    // whose references then read a zero address. The listing writes a `.weak` for each of the same
488    // names, so the two paths put the same entries in whether or not anything refers to one.
489    let weak: HashSet<&str> = data.weak.iter().map(String::as_str).collect();
490    let relocs = || text.relocs.iter().chain(data.objects.iter().flat_map(|o| &o.relocs));
491    // The names something here reaches through the thread pointer, which is the one thing about an
492    // undefined name this file does know. A reference to a thread-local variable is a different kind
493    // of reference from a reference to an ordinary one and the code that makes it is already
494    // different, so the file has been told, and ELF wants the symbol to say so as well.
495    let thread: HashSet<&str> = relocs()
496        .filter(|reloc| reloc.kind == Reference::Thread)
497        .map(|reloc| reloc.symbol.as_str())
498        .collect();
499    let wanted: Vec<&String> =
500        relocs().map(|reloc| &reloc.symbol).chain(data.weak.iter()).collect();
501    for name in wanted {
502        if symbols.contains_key(name) {
503            continue;
504        }
505        let id = obj.add_symbol(Symbol {
506            name: name.clone().into_bytes(),
507            value: 0,
508            size: 0,
509            // What kind of thing an undefined name is is not known here and does not have to be:
510            // a linker resolves an undefined symbol by its name, and the type of one that is not
511            // defined anywhere in this file is nothing this file can say. A thread-local one is the
512            // exception, and the linker makes it one. A reference to a thread-local variable is
513            // satisfied by an offset into a block rather than by an address, so the linker has to
514            // know which of the two it is being asked for before it has found the definition, and it
515            // refuses a link where one file says `STT_TLS` and another does not rather than picking
516            // one. That is tamnd/rucc#1461: libmpfr writes `__gmpfr_flags` in one file and reads it
517            // in a hundred others, and `ld` stopped at the first reader with a mismatch.
518            kind: if thread.contains(name.as_str()) {
519                SymbolKind::Tls
520            } else {
521                SymbolKind::Unknown
522            },
523            scope: SymbolScope::Dynamic,
524            weak: weak.contains(name.as_str()),
525            section: SymbolSection::Undefined,
526            flags: SymbolFlags::None,
527        });
528        symbols.insert(name.clone(), id);
529    }
530
531    for reloc in &text.relocs {
532        // Which function's bytes this one is in, which is the question only the split path has to
533        // ask: when there is one text section every offset in it is already the offset in it.
534        // Every relocation is inside some function, since the padding between two of them is
535        // instructions that do nothing and holds nothing a linker fills in.
536        let (section, at) = if sections.functions {
537            let after = text.funcs.partition_point(|func| func.start <= reloc.at);
538            let Some(func) = after.checked_sub(1).map(|i| &text.funcs[i]) else {
539                let why = format!("a relocation at {} is in front of every function", reloc.at);
540                return Err(Error::Refused { why });
541            };
542            // From the start of the section rather than from the symbol, and the two are not the
543            // same byte in a function with room in front of its label.
544            let base = func.start - func.patch.map_or(0, |patch| patch.before);
545            (split[after - 1].0, (reloc.at - base) as u64)
546        } else {
547            (whole, reloc.at as u64)
548        };
549        add(&mut obj, section, at, reloc, &symbols, flavour)?;
550    }
551
552    // The unwind table, if there is one. Its own section rather than part of the text, because it
553    // is read rather than run: the loader maps it and the linker gathers every input's into one
554    // table and builds the index the unwinder searches.
555    if !text.unwind.bytes.is_empty() {
556        let ((name, align), second) = flavour.tables();
557        let frames = obj.add_section(Vec::new(), name.into(), SectionKind::ReadOnlyData);
558        obj.append_section_data(frames, &text.unwind.bytes, align);
559        // What the rows point at, on the format that keeps the descriptions in a section of their
560        // own, and a name for each of them, because a row reaches one through a relocation and a
561        // relocation names a symbol. The names are never offered to another file: what they point
562        // at is one function's prologue, described for the runtime of this program and nothing else.
563        let mut described = HashMap::new();
564        if !text.unwind.info.is_empty() {
565            let Some((name, align)) = second else {
566                let why = "an unwind table here is one section and it was given two".to_owned();
567                return Err(Error::Refused { why });
568            };
569            let codes = obj.add_section(Vec::new(), name.into(), SectionKind::ReadOnlyData);
570            obj.append_section_data(codes, &text.unwind.info, align);
571            for label in &text.unwind.labels {
572                let id = obj.add_symbol(Symbol {
573                    name: label.name.clone().into_bytes(),
574                    value: label.at as u64,
575                    size: 0,
576                    kind: SymbolKind::Label,
577                    scope: SymbolScope::Compilation,
578                    weak: false,
579                    section: SymbolSection::Section(codes),
580                    flags: SymbolFlags::None,
581                });
582                described.insert(label.name.clone(), id);
583            }
584        }
585        for reloc in &text.unwind.relocs {
586            let (symbol, addend) = match described.get(&reloc.symbol) {
587                // A description in the section above, reached by its own name and needing no
588                // correction, since the name is at the description rather than at the front of the
589                // section it is in.
590                Some(&id) => (id, reloc.addend),
591                // A function, and against the section it is in rather than against its own name,
592                // which is the same reason the record of a patcher's room is written that way and
593                // one more besides. The section is the only one of the two that is settled here: a
594                // global name is answered at load time by whichever object defines it first, so a
595                // distance measured to one is not a distance the linker can work out, and it says
596                // so and stops. The effect was that nothing this compiler wrote could go into a
597                // shared library at all, because every function has a record and every record
598                // pointed at a name.
599                //
600                // A function defined elsewhere has no record here, so the lookup failing means the
601                // record is for something that is not a function in this file, and that is a bug
602                // rather than a shape to handle: the writer says what it was given rather than
603                // guessing.
604                None => {
605                    let found = text.funcs.iter().position(|func| func.name == reloc.symbol);
606                    let Some((section, at)) = found.map(|i| split[i]) else {
607                        let why = format!(
608                            "'{}' has an unwind record and is not a function here",
609                            reloc.symbol
610                        );
611                        return Err(Error::Refused { why });
612                    };
613                    // Where the function starts inside its section, since the section symbol is
614                    // where the section starts and the two are the same byte only for the first
615                    // function in one.
616                    (obj.section_symbol(section), reloc.addend + at as i64)
617                }
618            };
619            let flags = flavour.reloc(reloc.kind, reloc.after).ok_or_else(|| Error::Refused {
620                why: format!("no relocation is {:?}", reloc.kind),
621            })?;
622            let record = Relocation { offset: reloc.at as u64, symbol, addend, flags };
623            obj.add_relocation(frames, record)
624                .map_err(|why| Error::Refused { why: why.to_string() })?;
625        }
626    }
627    // The debug information, if the build asked for any. One section per chunk under the name
628    // DWARF gives it, and none of them allocated: the loader does not map a debug section and
629    // nothing at run time reads one, which is what tells this apart from the unwind table, whose
630    // whole point is that a program walking its own stack can reach it.
631    //
632    // Every section is added before any relocation is, because a relocation in one of them names
633    // another as often as it names a function, and a name is resolved against the sections the
634    // file already has.
635    let mut named = HashMap::new();
636    for chunk in &info.chunks {
637        let id = obj.add_section(Vec::new(), chunk.name.clone().into_bytes(), SectionKind::Debug);
638        obj.append_section_data(id, &chunk.bytes, 1);
639        named.insert(chunk.name.as_str(), id);
640    }
641    for chunk in &info.chunks {
642        let section = named[chunk.name.as_str()];
643        for reloc in &chunk.relocs {
644            let (symbol, addend) = match named.get(reloc.symbol.as_str()) {
645                // Another debug section, reached by its own name. The distance is from the front
646                // of that section, which is what the section symbol is, so the addend stands.
647                Some(&id) => (obj.section_symbol(id), reloc.addend),
648                // A function, and against the section it is in rather than against its own name,
649                // for the reason the unwind table's records are written that way: a global name is
650                // answered at load time by whichever object defines it first, and a distance to
651                // one is not a distance a linker can work out.
652                None => match text.funcs.iter().position(|func| func.name == reloc.symbol) {
653                    Some(which) => {
654                        let (section, at) = split[which];
655                        (obj.section_symbol(section), reloc.addend + at as i64)
656                    }
657                    // Or a variable this file defines, which a `DW_TAG_variable` asks for the
658                    // address of. Against its section for the reason a function is, where it has
659                    // one. A variable the linker is being asked for zeroed space for has no
660                    // section to count from and nothing but its own name to ask by, which is the
661                    // one case here where the name goes in the relocation.
662                    None => {
663                        let found = data.objects.iter().position(|had| had.name == reloc.symbol);
664                        let Some(which) = found else {
665                            let why = format!(
666                                "'{}' is named by the debug information and is not defined here",
667                                reloc.symbol
668                            );
669                            return Err(Error::Refused { why });
670                        };
671                        match placed[which] {
672                            (Some(section), at) => {
673                                (obj.section_symbol(section), reloc.addend + at as i64)
674                            }
675                            (None, _) => (symbols[&reloc.symbol], reloc.addend),
676                        }
677                    }
678                },
679            };
680            let flags = flavour.reloc(reloc.kind, reloc.after).ok_or_else(|| Error::Refused {
681                why: format!("no relocation is {:?}", reloc.kind),
682            })?;
683            let record = Relocation { offset: reloc.at as u64, symbol, addend, flags };
684            obj.add_relocation(section, record)
685                .map_err(|why| Error::Refused { why: why.to_string() })?;
686        }
687    }
688    for (object, &(section, offset)) in data.objects.iter().zip(&placed) {
689        let Some(section) = section else { continue };
690        for reloc in &object.relocs {
691            add(&mut obj, section, offset + reloc.at as u64, reloc, &symbols, flavour)?;
692        }
693    }
694
695    // What the file was built to have checked, when it was built to have anything checked. Left
696    // out otherwise rather than written as a zero, because a linker treats a missing note and a
697    // note with no bits in it the same way and gcc writes nothing.
698    flavour.property(&mut obj, property);
699
700    // Written rather than left out, because a linker that does not find it in every input marks
701    // the stack executable, on the format that has one.
702    flavour.marker(&mut obj);
703
704    let mut bytes = obj.write().map_err(|why| Error::Refused { why: why.to_string() })?;
705    flavour.finish(&mut bytes, &ordered);
706    Ok(bytes)
707}
708
709/// How far one label is from another, from the symbols [`write()`] added for them.
710///
711/// # Errors
712///
713/// [`Error::Refused`] for a label that is not here, for two that are in different sections, and
714/// for a distance too far for the width it is written in.
715fn distance(
716    obj: &Writer<'_>,
717    symbols: &BTreeMap<String, SymbolId>,
718    apart: &Apart,
719) -> Result<i64, Error> {
720    let find = |name: &str| match symbols.get(name) {
721        Some(&id) => Ok(obj.symbol(id)),
722        None => Err(Error::Refused { why: format!("'{name}' is measured from and is not here") }),
723    };
724    let (to, from) = (find(&apart.to)?, find(&apart.from)?);
725    if to.section != from.section {
726        let why = format!("'{}' and '{}' are in different sections", apart.to, apart.from);
727        return Err(Error::Refused { why });
728    }
729    let value = (to.value as i64).wrapping_sub(from.value as i64).wrapping_add(apart.addend);
730    let bits = u32::from(apart.bytes) * 8;
731    if bits < 64 && (value >> (bits - 1)) != 0 && (value >> (bits - 1)) != -1 {
732        let why = format!("'{}' is too far from '{}' for {} bytes", apart.to, apart.from, bits / 8);
733        return Err(Error::Refused { why });
734    }
735    Ok(value)
736}
737
738/// Everything in this module the target's format has no way to write, refused by name.
739///
740/// Each of these is something ELF has and COFF does not, and each would otherwise be written as the
741/// nearest thing rather than refused, which is worse: a thread-local variable written as an ordinary
742/// one is a program where every thread shares what the source said each would have its own copy of,
743/// and a constructor list under a name the Windows runtime does not gather is a program whose
744/// constructors never run. A message naming the feature is what the caller turns into a diagnostic,
745/// and the front end refusing first is what stops one ever being seen.
746///
747/// # Errors
748///
749/// [`Error::Refused`], naming the one it found first.
750fn beyond(text: &Text, data: &Data, info: &Info) -> Result<(), Error> {
751    let why = |why: String| Err(Error::Refused { why });
752    if !info.chunks.is_empty() {
753        return why("debug information here goes in sections this writer does not name".to_owned());
754    }
755    if text.funcs.iter().any(|func| func.patch.is_some()) {
756        return why("a record of where a patcher's room is has no section flags here".to_owned());
757    }
758    for reloc in text.relocs.iter().chain(data.objects.iter().flat_map(|object| &object.relocs)) {
759        if matches!(reloc.kind, Reference::Got | Reference::Thread) {
760            return why(format!("nothing reaches '{}' through a table here", reloc.symbol));
761        }
762    }
763    for object in &data.objects {
764        if matches!(object.place, Place::Thread { .. }) {
765            return why(format!("'{}' is thread-local and this format is not", object.name));
766        }
767        let Place::Named(name) = &object.place else { continue };
768        if Array::of(name).is_some() {
769            return why(format!("'{name}' is not a list the startup code here gathers"));
770        }
771    }
772    Ok(())
773}
774
775/// Every name a linker can find in the object [`write()`] would write from the same input.
776///
777/// What asks for this is the archive writer. A static link resolves through the symbol index, so an
778/// index entry has to name a symbol the member really defines: an entry for a name that is not in
779/// the member is an archive the linker searches, pulls the member out of, and then still reports
780/// the name undefined. So the list comes from the writer rather than from the caller, because the
781/// writer is the only thing that knows what it wrote.
782///
783/// The names are as the C program spelled them, with nothing in front of them, which is what both
784/// the formats this writes have on this machine. Mach-O puts an underscore there and so does COFF on
785/// a 32-bit machine, and when either of those is written this is the function that has to say so,
786/// which is why it asks about the target it otherwise would not have to.
787///
788/// Order is the functions, then the variables, then the aliases, each in the order the module held
789/// them, which is the order [`write()`] adds the symbols in. A `static` is left out: it is a name the
790/// link has already finished with by the time an archive is searched, and an index entry for one
791/// would offer the linker a definition it is not allowed to use.
792///
793/// # Errors
794///
795/// [`Error::Format`] for a machine or a platform this does not write, which is the same refusal
796/// [`write()`] gives and is here for the same reason: a list of undecorated names for a format whose
797/// symbols carry an underscore is worse than no list at all.
798pub fn defines(
799    text: &Text,
800    data: &Data,
801    aliases: &[Alias],
802    target: &TargetInfo,
803) -> Result<Vec<String>, Error> {
804    if target.tuple.arch() != Arch::X86_64 || Flavour::of(target).is_none() {
805        return Err(Error::Format { triple: target.tuple.to_string() });
806    }
807    let names = text
808        .funcs
809        .iter()
810        .filter(|func| func.binding != Binding::Local)
811        .map(|func| func.name.clone())
812        .chain(
813            data.objects
814                .iter()
815                .filter(|object| object.binding != Binding::Local)
816                .map(|object| object.name.clone()),
817        )
818        .chain(
819            aliases
820                .iter()
821                .filter(|alias| alias.binding != Binding::Local)
822                .map(|alias| alias.name.clone()),
823        )
824        .collect();
825    Ok(names)
826}
827
828/// One variable's image into the section it belongs in, and where in that section it landed.
829///
830/// A zero filled variable takes as many bytes of the file as it is long on the way in and none on
831/// the way out, which is the whole point of the section it goes in. A merged one goes in no section
832/// at all: the linker is being asked for that much zeroed space under that name, and where it ends
833/// up is the linker's answer rather than this file's.
834fn put(
835    obj: &mut Writer<'_>,
836    object: &Object,
837    named: &mut HashMap<String, object::write::SectionId>,
838    sections: Sections,
839    flavour: Flavour,
840) -> (SymbolSection, u64) {
841    // A section of its own, named after the variable and after the section it would have gone in,
842    // which is what `-fdata-sections` asks for. A merged variable has no section to split and a
843    // named one was named by the program, so both are left where they are: the first is a request
844    // to the linker rather than an image, and the second would otherwise have the flag silently
845    // overrule what the source said.
846    if sections.data {
847        if let Some(name) = object.place.split(&object.name) {
848            let section = obj.add_section(Vec::new(), name.into_bytes(), kind_of(&object.place));
849            let offset = if carries_no_bytes(&object.place) {
850                obj.append_section_bss(section, object.size, object.align)
851            } else {
852                obj.append_section_data(section, &object.bytes, object.align)
853            };
854            return (SymbolSection::Section(section), offset);
855        }
856    }
857    let section = match &object.place {
858        Place::Written => obj.section_id(StandardSection::Data),
859        Place::ReadOnly => obj.section_id(StandardSection::ReadOnlyData),
860        // Read only after the loader has written it, which the writer knows as the relocatable
861        // read only data section and which is `.data.rel.ro` on ELF. The `.local` half is a layout
862        // hint the writer has no name for, so it is added by hand and remembered: asking again
863        // would make a second section with the same name, and a file with one of those per variable
864        // is a file whose section headers outweigh what they describe.
865        Place::RelocReadOnly { local } => match flavour.rel_ro_local().filter(|_| *local) {
866            Some(name) => made(obj, named, name, SectionKind::ReadOnlyDataWithRel),
867            None => obj.section_id(StandardSection::ReadOnlyDataWithRel),
868        },
869        Place::Zero => obj.section_id(StandardSection::UninitializedData),
870        Place::Thread { zero: false } => obj.section_id(StandardSection::Tls),
871        Place::Thread { zero: true } => obj.section_id(StandardSection::UninitializedTls),
872        Place::Merged => return (SymbolSection::Common, 0),
873        // A named section is the program's word for where this goes, and a program that names one
874        // wants what it named rather than what would have been chosen. It is written as ordinary
875        // data because nothing in the IR says otherwise, except for the three names the startup
876        // code calls what it finds in, which have a section type of their own and are gathered by
877        // the linker whether or not they carry it.
878        Place::Named(name) => {
879            let section = made(obj, named, name, SectionKind::Data);
880            if let Some(flags) = Array::of(name).and_then(|array| flavour.gathered(array)) {
881                obj.section_mut(section).flags = flags;
882            }
883            section
884        }
885    };
886    let offset = if carries_no_bytes(&object.place) {
887        obj.append_section_bss(section, object.size, object.align)
888    } else {
889        obj.append_section_data(section, &object.bytes, object.align)
890    };
891    (SymbolSection::Section(section), offset)
892}
893
894/// Whether the section this goes in says how big the variable is and holds none of its bytes.
895///
896/// Two of them, and they are the same answer twice: `.bss` is the image that is all zeros, and
897/// `.tbss` is a thread's own copy of one. A section like this costs its size in the section header
898/// and nothing in the file, which is what keeps a program with a large zeroed array small.
899fn carries_no_bytes(place: &Place) -> bool {
900    matches!(place, Place::Zero | Place::Thread { zero: true })
901}
902
903/// The section of this name, made the first time it is asked for and found afterwards.
904///
905/// Two variables the program put the same section name on belong in one section, the way two in
906/// `.data` do. Asking the writer for a new one each time would make a second header with the same
907/// name, which a linker takes and which makes a file with ten constructors in it carry ten section
908/// headers describing eight bytes each. `section_id` does this already for the sections it has
909/// names of its own for, and this is the same answer for the ones it does not.
910fn made(
911    obj: &mut Writer<'_>,
912    named: &mut HashMap<String, object::write::SectionId>,
913    name: &str,
914    kind: SectionKind,
915) -> object::write::SectionId {
916    if let Some(section) = named.get(name) {
917        return *section;
918    }
919    let section = obj.add_section(Vec::new(), name.as_bytes().to_vec(), kind);
920    named.insert(name.to_owned(), section);
921    section
922}
923
924/// What a section split off for one variable is, which is what the section it was split off from
925/// was.
926///
927/// Splitting changes the name and nothing else. A variable that was going to be in a page the
928/// loader maps read only is still in one, and a zero filled variable still costs the file nothing,
929/// so the flags a linker reads off the section header have to come out the same as they would
930/// have. The two kinds with no section of their own never reach here, and `Data` for them is a
931/// value that is never used rather than a claim about either.
932fn kind_of(place: &Place) -> SectionKind {
933    match place {
934        Place::ReadOnly => SectionKind::ReadOnlyData,
935        Place::RelocReadOnly { .. } => SectionKind::ReadOnlyDataWithRel,
936        Place::Zero => SectionKind::UninitializedData,
937        Place::Thread { zero: false } => SectionKind::Tls,
938        Place::Thread { zero: true } => SectionKind::UninitializedTls,
939        Place::Written | Place::Merged | Place::Named(_) => SectionKind::Data,
940    }
941}
942
943/// One relocation, `at` bytes into the section it ended up in.
944///
945/// The offset is worked out by the caller rather than here, because the two callers count from
946/// different places: a relocation in an image counts from the start of that image and a relocation
947/// in a function counts from the start of that function, and neither of those is where the section
948/// begins once something else is in front of it.
949fn add(
950    obj: &mut Writer<'_>,
951    section: object::write::SectionId,
952    at: u64,
953    reloc: &Reloc,
954    symbols: &BTreeMap<String, SymbolId>,
955    flavour: Flavour,
956) -> Result<(), Error> {
957    let flags = flavour
958        .reloc(reloc.kind, reloc.after)
959        .ok_or_else(|| Error::Refused { why: format!("no relocation is {:?}", reloc.kind) })?;
960    obj.add_relocation(
961        section,
962        Relocation { offset: at, symbol: symbols[&reloc.symbol], addend: reloc.addend, flags },
963    )
964    .map_err(|why| Error::Refused { why: why.to_string() })
965}
966
967/// How far a name reaches, which is the one thing about a symbol ELF calls its binding.
968///
969/// `SymbolScope` is two facts in one word, and the trap is that the middle one is not the neutral
970/// answer it reads as. The writer turns `Compilation` into a local symbol, and it turns the choice
971/// between `Linkage` and `Dynamic` into `st_other`: `Linkage` is `STV_HIDDEN` and `Dynamic` is
972/// `STV_DEFAULT`. So there is no way to say global and decline to say anything about visibility,
973/// and picking the one whose name sounds like the smaller claim is picking hidden. That is what
974/// tamnd/rucc#733 was.
975///
976/// `Dynamic` is what every global asks for here, and the visibility is said afterwards by
977/// [`see`] rather than through this, so that nothing about `st_other` depends on reading one of
978/// these four names the way its author meant it.
979pub(crate) fn scope_of(binding: Binding) -> SymbolScope {
980    match binding {
981        Binding::Local => SymbolScope::Compilation,
982        Binding::Global | Binding::Weak => SymbolScope::Dynamic,
983    }
984}
985
986#[cfg(test)]
987mod tests {
988    use super::*;
989
990    use object::read::elf::Sym as _;
991    use object::read::{Object as _, ObjectSection as _, ObjectSymbol as _};
992    use object::{elf, pe};
993    use rucc_target::{Arch, Env, Os, Triple};
994
995    use crate::elf::PATCHABLE;
996    use crate::section::{Extent, Marker, Patch, Reloc};
997
998    /// A linux x86-64 target, which is the only one this writes.
999    fn target() -> TargetInfo {
1000        TargetInfo::new(Triple::new(Arch::X86_64, Os::Linux, Env::Gnu))
1001    }
1002
1003    /// One function of that name, at that offset, that many bytes long, and visible that far.
1004    ///
1005    /// Visibility is the field these cases mostly have no opinion about, so it is the one the
1006    /// helper fills in and the two that do have an opinion write for themselves.
1007    fn extent(name: String, start: usize, len: usize, binding: Binding) -> Extent {
1008        Extent {
1009            name,
1010            start,
1011            len,
1012            align: crate::FUNC_ALIGN,
1013            binding,
1014            visibility: Visibility::Default,
1015            patch: None,
1016        }
1017    }
1018
1019    /// A call to something outside the file, which is the shape every case here starts from.
1020    fn calling(name: &str) -> Text {
1021        Text {
1022            bytes: vec![0xe8, 0, 0, 0, 0, 0xc3],
1023            funcs: vec![extent("f".to_owned(), 0, 6, Binding::Global)],
1024            relocs: vec![Reloc {
1025                at: 1,
1026                symbol: name.to_owned(),
1027                kind: Reference::Call,
1028                addend: -4,
1029                after: 0,
1030            }],
1031            ..Text::default()
1032        }
1033    }
1034
1035    #[test]
1036    fn the_bytes_come_back_out_of_the_section_they_went_into() {
1037        let text = calling("puts");
1038        let bytes =
1039            write(&text, &Data::default(), &[], &target(), Output::default(), &Info::default())
1040                .expect("an object");
1041        let file = object::File::parse(&bytes[..]).expect("a readable object");
1042        let section = file.section_by_name(".text").expect("a text section");
1043        assert_eq!(section.data().expect("the bytes"), &text.bytes[..]);
1044    }
1045
1046    #[test]
1047    fn a_function_is_a_symbol_that_says_where_it_is_and_how_long_it_is() {
1048        let mut text = calling("puts");
1049        text.funcs.push(extent("g".to_owned(), 16, 1, Binding::Global));
1050        text.bytes.resize(17, 0x90);
1051        let bytes =
1052            write(&text, &Data::default(), &[], &target(), Output::default(), &Info::default())
1053                .expect("an object");
1054        let file = object::File::parse(&bytes[..]).expect("a readable object");
1055        let g = file.symbols().find(|s| s.name() == Ok("g")).expect("the second function");
1056        assert_eq!(g.address(), 16);
1057        assert_eq!(g.size(), 1);
1058        assert_eq!(g.kind(), SymbolKind::Text);
1059        assert!(g.is_global(), "nothing said otherwise about this one");
1060    }
1061
1062    #[test]
1063    fn a_function_no_other_file_can_see_is_a_local_symbol() {
1064        let mut text = calling("puts");
1065        text.funcs.push(extent("hidden".to_owned(), 16, 1, Binding::Local));
1066        text.funcs.push(extent("shared".to_owned(), 32, 1, Binding::Weak));
1067        text.bytes.resize(33, 0x90);
1068        let bytes =
1069            write(&text, &Data::default(), &[], &target(), Output::default(), &Info::default())
1070                .expect("an object");
1071        let file = object::File::parse(&bytes[..]).expect("a readable object");
1072        let hidden = file.symbols().find(|s| s.name() == Ok("hidden")).expect("the static one");
1073        // A symbol the linker keeps and does not let another file reach, which is the whole of
1074        // what `static` on a function means and what two files each defining their own need.
1075        assert!(hidden.is_local(), "a static function must not be offered to the linker");
1076        assert!(!hidden.is_weak());
1077        let shared = file.symbols().find(|s| s.name() == Ok("shared")).expect("the weak one");
1078        assert!(shared.is_weak(), "a weak function has to be able to lose");
1079        assert!(shared.is_global());
1080    }
1081
1082    /// A global is `STV_DEFAULT`, so a shared library built from these objects exports something.
1083    ///
1084    /// The bug in tamnd/rucc#733. Every global came out `STV_HIDDEN`, which a static link does not
1085    /// look at, so nothing here noticed and SQLite linked and ran and the whole test suite passed.
1086    /// What it costs is the dynamic symbol table: `gcc -shared` over one of these objects produced
1087    /// a library with an empty one, and `dlsym` could not find a function the file plainly defines.
1088    ///
1089    /// Written against `st_other` itself rather than against the reader's `scope`, because `scope`
1090    /// is the word that was misread in the first place and a test that asks it the same question
1091    /// would agree with whatever the writer did.
1092    /// The record of where a patcher's room is, and what it says about it.
1093    ///
1094    /// Four things have to be right at once for a linker to take it: the flags, the alignment, the
1095    /// relocation and the section it says it is ordered after. The last of those is the one the
1096    /// writer underneath cannot say, so a zero there would be a file `ld` refuses and a test that
1097    /// only looked at the bytes would not see it.
1098    #[test]
1099    fn where_a_patcher_may_write_is_recorded_in_a_section_tied_to_the_code_it_is_about() {
1100        let mut text = calling("puts");
1101        text.bytes.splice(0..0, [0x90, 0x90, 0x90]);
1102        text.funcs[0].start = 3;
1103        text.funcs[0].patch = Some(Patch { at: 0, before: 3 });
1104        text.relocs[0].at = 4;
1105        let bytes =
1106            write(&text, &Data::default(), &[], &target(), Output::default(), &Info::default())
1107                .expect("an object");
1108        let file = object::read::elf::ElfFile64::<Endianness>::parse(&bytes[..]).expect("readable");
1109        let section = file.section_by_name(PATCHABLE).expect("a record of the room");
1110        assert_eq!(section.size(), 8, "one address, and this file defines one function");
1111        assert_eq!(section.align(), 8);
1112        let header = section.elf_section_header();
1113        assert_eq!(
1114            header.sh_flags.get(Endianness::Little),
1115            elf::SHF_ALLOC | elf::SHF_WRITE | elf::SHF_LINK_ORDER
1116        );
1117        // Which is the whole point of the fixup: the index has to be the text section's own, and
1118        // the writer underneath had written a zero there.
1119        let index = file.section_by_name(".text").expect("a text section").index().0;
1120        assert_eq!(header.sh_link.get(Endianness::Little) as usize, index);
1121        assert_ne!(index, 0);
1122
1123        // And the address, which is the front of the room rather than the function's own symbol.
1124        let [(at, reloc)] = &section.relocations().collect::<Vec<_>>()[..] else {
1125            panic!("one address in the record")
1126        };
1127        assert_eq!(*at, 0);
1128        assert_eq!(reloc.addend(), 0);
1129        assert_eq!(reloc.flags(), RelocationFlags::Elf { r_type: elf::R_X86_64_64 });
1130    }
1131
1132    /// And a file that asked for none has no such section, which is nearly every file.
1133    #[test]
1134    fn a_file_that_promised_a_patcher_nothing_records_nothing() {
1135        let text = calling("puts");
1136        let bytes =
1137            write(&text, &Data::default(), &[], &target(), Output::default(), &Info::default())
1138                .expect("an object");
1139        let file = object::File::parse(&bytes[..]).expect("a readable object");
1140        assert!(file.section_by_name(PATCHABLE).is_none());
1141    }
1142
1143    /// The same when each function is a section of its own, which is what a kernel builds with.
1144    ///
1145    /// Each record then points at a different section, which is what makes the pairing worth
1146    /// asserting: getting it backwards would still produce a file every tool reads and every
1147    /// address in it would be about the wrong function.
1148    #[test]
1149    fn each_record_is_tied_to_its_own_function_when_they_are_split_up() {
1150        let mut text = calling("puts");
1151        text.funcs[0].patch = Some(Patch { at: 0, before: 0 });
1152        text.funcs.push(extent("g".to_owned(), 16, 1, Binding::Global));
1153        text.funcs[1].patch = Some(Patch { at: 16, before: 0 });
1154        text.bytes.resize(17, 0x90);
1155        let output =
1156            Output { sections: Sections { functions: true, data: false }, ..Output::default() };
1157        let bytes = write(&text, &Data::default(), &[], &target(), output, &Info::default())
1158            .expect("an object");
1159        let file = object::read::elf::ElfFile64::<Endianness>::parse(&bytes[..]).expect("readable");
1160        let links: Vec<usize> = file
1161            .sections()
1162            .filter(|section| section.name() == Ok(PATCHABLE))
1163            .map(|section| section.elf_section_header().sh_link.get(Endianness::Little) as usize)
1164            .collect();
1165        let index = |name: &str| file.section_by_name(name).expect("a text section").index().0;
1166        assert_eq!(links, [index(".text.f"), index(".text.g")]);
1167    }
1168
1169    #[test]
1170    fn a_global_is_visible_to_the_dynamic_linker_and_a_static_one_is_not_a_symbol_at_all() {
1171        let mut text = calling("puts");
1172        text.funcs.push(extent("g".to_owned(), 16, 1, Binding::Global));
1173        text.funcs.push(extent("w".to_owned(), 32, 1, Binding::Weak));
1174        text.funcs.push(extent("s".to_owned(), 48, 1, Binding::Local));
1175        text.bytes.resize(49, 0x90);
1176        let bytes =
1177            write(&text, &Data::default(), &[], &target(), Output::default(), &Info::default())
1178                .expect("an object");
1179        let file = object::read::elf::ElfFile64::<Endianness>::parse(&bytes[..]).expect("readable");
1180        let visibility = |name: &str| {
1181            file.symbols()
1182                .find(|s| s.name() == Ok(name))
1183                .expect("the function")
1184                .elf_symbol()
1185                .st_visibility()
1186        };
1187        // Nothing said hidden about either of these, so neither is.
1188        assert_eq!(visibility("g"), elf::STV_DEFAULT);
1189        assert_eq!(visibility("w"), elf::STV_DEFAULT, "a weak one is still a name others may use");
1190        // The `static` one is local, and a local symbol's visibility means nothing either way,
1191        // which is why the binding is what this asks about.
1192        assert_eq!(visibility("s"), elf::STV_DEFAULT);
1193    }
1194
1195    /// And the other direction: a name that did ask to be hidden is hidden, and a protected one is
1196    /// protected.
1197    ///
1198    /// The half of tamnd/rucc#733 that the fix above left open. Saying `STV_DEFAULT` for everything
1199    /// is right for everything nobody marked and wrong the moment something is marked, so the two
1200    /// tests together are what says the field carries an answer rather than a constant.
1201    ///
1202    /// Both are asked of a function and of a variable, because they are added by two different
1203    /// loops in `write` and a field one of them fills in is not a field the other one does.
1204    #[test]
1205    fn a_name_that_asked_to_be_hidden_is_hidden_and_a_protected_one_is_protected() {
1206        let mut text = calling("puts");
1207        for (index, (name, seen)) in
1208            [("h", Visibility::Hidden), ("p", Visibility::Protected)].into_iter().enumerate()
1209        {
1210            let mut func = extent(name.to_owned(), 16 + index * 16, 1, Binding::Global);
1211            func.visibility = seen;
1212            text.funcs.push(func);
1213        }
1214        text.bytes.resize(49, 0x90);
1215        let mut data = Data::default();
1216        for (name, seen) in [("vh", Visibility::Hidden), ("vp", Visibility::Protected)] {
1217            let mut object = variable(name, Place::Written);
1218            object.visibility = seen;
1219            data.objects.push(object);
1220        }
1221        let bytes = write(&text, &data, &[], &target(), Output::default(), &Info::default())
1222            .expect("an object");
1223        let file = object::read::elf::ElfFile64::<Endianness>::parse(&bytes[..]).expect("readable");
1224        let visibility = |name: &str| {
1225            file.symbols()
1226                .find(|s| s.name() == Ok(name))
1227                .expect("the symbol")
1228                .elf_symbol()
1229                .st_visibility()
1230        };
1231        assert_eq!(visibility("h"), elf::STV_HIDDEN);
1232        assert_eq!(visibility("p"), elf::STV_PROTECTED);
1233        assert_eq!(visibility("vh"), elf::STV_HIDDEN, "a variable goes through a second loop");
1234        assert_eq!(visibility("vp"), elf::STV_PROTECTED);
1235        // The one thing a visibility must not disturb, since `st_info` and `st_other` are written
1236        // in one go and the second was set after the first.
1237        let h = file.symbols().find(|s| s.name() == Ok("h")).expect("the function");
1238        assert!(h.is_global(), "hidden is about the dynamic linker and not about the binding");
1239        assert_eq!(h.size(), 1, "and it is still a function of the length it was");
1240    }
1241
1242    #[test]
1243    fn a_name_this_file_does_not_define_is_left_for_the_linker_to_find() {
1244        let bytes = write(
1245            &calling("puts"),
1246            &Data::default(),
1247            &[],
1248            &target(),
1249            Output::default(),
1250            &Info::default(),
1251        )
1252        .expect("an object");
1253        let file = object::File::parse(&bytes[..]).expect("a readable object");
1254        let puts = file.symbols().find(|s| s.name() == Ok("puts")).expect("the callee");
1255        assert!(puts.is_undefined(), "the file does not define it and must not claim to");
1256    }
1257
1258    #[test]
1259    fn a_call_asks_for_the_relocation_a_stub_may_answer_and_a_load_asks_for_the_one_that_may_not() {
1260        for (reference, wanted) in [
1261            (Reference::Call, elf::R_X86_64_PLT32),
1262            (Reference::Data, elf::R_X86_64_PC32),
1263            (Reference::Got, elf::R_X86_64_REX_GOTPCRELX),
1264            (Reference::Thread, elf::R_X86_64_GOTTPOFF),
1265        ] {
1266            let mut text = calling("puts");
1267            text.relocs[0].kind = reference;
1268            let bytes =
1269                write(&text, &Data::default(), &[], &target(), Output::default(), &Info::default())
1270                    .expect("an object");
1271            let file = object::File::parse(&bytes[..]).expect("a readable object");
1272            let section = file.section_by_name(".text").expect("a text section");
1273            let (offset, reloc) = section.relocations().next().expect("one relocation");
1274            assert_eq!(offset, 1);
1275            assert_eq!(reloc.addend(), -4);
1276            assert_eq!(reloc.flags(), RelocationFlags::Elf { r_type: wanted });
1277        }
1278    }
1279
1280    #[test]
1281    fn a_name_wanted_twice_is_one_symbol_rather_than_two() {
1282        let mut text = calling("puts");
1283        text.relocs.push(Reloc {
1284            at: 1,
1285            symbol: "puts".to_owned(),
1286            kind: Reference::Call,
1287            addend: -4,
1288            after: 0,
1289        });
1290        let bytes =
1291            write(&text, &Data::default(), &[], &target(), Output::default(), &Info::default())
1292                .expect("an object");
1293        let file = object::File::parse(&bytes[..]).expect("a readable object");
1294        assert_eq!(file.symbols().filter(|s| s.name() == Ok("puts")).count(), 1);
1295    }
1296
1297    #[test]
1298    fn a_function_that_is_also_called_is_not_a_second_symbol() {
1299        let text = calling("f");
1300        let bytes =
1301            write(&text, &Data::default(), &[], &target(), Output::default(), &Info::default())
1302                .expect("an object");
1303        let file = object::File::parse(&bytes[..]).expect("a readable object");
1304        let mut found = file.symbols().filter(|s| s.name() == Ok("f"));
1305        let f = found.next().expect("the function");
1306        assert!(!f.is_undefined(), "the file defines it");
1307        assert!(found.next().is_none(), "and defines it once");
1308    }
1309
1310    #[test]
1311    fn the_marker_that_says_the_stack_is_not_executable_is_written() {
1312        let bytes = write(
1313            &calling("puts"),
1314            &Data::default(),
1315            &[],
1316            &target(),
1317            Output::default(),
1318            &Info::default(),
1319        )
1320        .expect("an object");
1321        let file = object::File::parse(&bytes[..]).expect("a readable object");
1322        let note = file.section_by_name(".note.GNU-stack").expect("the marker");
1323        assert!(note.data().expect("no bytes").is_empty());
1324    }
1325
1326    /// What the file says it was built to have checked, byte for byte.
1327    ///
1328    /// Written against the bytes rather than against a reader, because the two lengths in the
1329    /// header count the padding after what they measure and a note whose lengths are one word out
1330    /// is one a linker drops without saying anything. What comes of that is a program the loader
1331    /// leaves the check turned off for, which is a build that looks like it worked.
1332    #[test]
1333    fn the_note_that_says_what_the_file_was_built_to_have_checked_is_written() {
1334        let property = Property { features: Property::IBT | Property::SHSTK };
1335        let output = Output { property, ..Output::default() };
1336        let bytes =
1337            write(&calling("puts"), &Data::default(), &[], &target(), output, &Info::default())
1338                .expect("an object");
1339        let file = object::File::parse(&bytes[..]).expect("a readable object");
1340        let note = file.section_by_name(".note.gnu.property").expect("the note");
1341        assert_eq!(note.align(), 8, "a note in a sixty four bit object is read a word at a time");
1342        let want: Vec<u8> = [
1343            4u32,
1344            16,
1345            5,
1346            u32::from_le_bytes(*b"GNU\0"),
1347            Property::X86_FEATURES,
1348            4,
1349            Property::IBT | Property::SHSTK,
1350            0,
1351        ]
1352        .iter()
1353        .flat_map(|word| word.to_le_bytes())
1354        .collect();
1355        assert_eq!(note.data().expect("the bytes"), &want[..]);
1356    }
1357
1358    /// And nothing at all when the file was built to have nothing checked.
1359    ///
1360    /// A note with an empty feature word and no note are the same thing to a linker, which drops
1361    /// the whole property when any input lacks it. gcc writes nothing, so a section header that
1362    /// describes nothing would be the one difference between the two compilers' objects.
1363    #[test]
1364    fn a_file_built_to_have_nothing_checked_says_nothing() {
1365        let bytes = write(
1366            &calling("puts"),
1367            &Data::default(),
1368            &[],
1369            &target(),
1370            Output::default(),
1371            &Info::default(),
1372        )
1373        .expect("an object");
1374        let file = object::File::parse(&bytes[..]).expect("a readable object");
1375        assert!(file.section_by_name(".note.gnu.property").is_none());
1376    }
1377
1378    /// Every unwind record names the function it is about, and each name goes where it is in the
1379    /// table rather than at the start of it.
1380    ///
1381    /// Written because working the offset out is the caller's job here, which is what the two text
1382    /// paths differ about, and a third caller that let it default to nothing would put every record
1383    /// in the table on the same function. Nothing else would notice: the section is the right
1384    /// length, the symbols are right, the link succeeds, and what comes of it is an unwinder that
1385    /// walks out of the wrong frame the first time something throws or a backtrace is taken.
1386    #[test]
1387    fn an_unwind_record_names_the_function_it_is_about_and_not_the_first_one() {
1388        let mut text = calling("puts");
1389        text.funcs.push(extent("g".to_owned(), 16, 1, Binding::Global));
1390        text.bytes.resize(17, 0x90);
1391        // A shared header and two records, whose contents nothing here reads: what is being asked
1392        // is where in them each name landed.
1393        text.unwind.bytes = vec![0; 64];
1394        for (at, name) in [(32usize, "f"), (48usize, "g")] {
1395            text.unwind.relocs.push(Reloc {
1396                at,
1397                symbol: name.to_owned(),
1398                kind: Reference::Address { bytes: 8 },
1399                addend: 0,
1400                after: 0,
1401            });
1402        }
1403        let bytes =
1404            write(&text, &Data::default(), &[], &target(), Output::default(), &Info::default())
1405                .expect("an object");
1406        let file = object::File::parse(&bytes[..]).expect("a readable object");
1407        let mut found = points_at(&file);
1408        found.sort_unstable();
1409        assert_eq!(found, [(32, ".text".to_owned(), 0), (48, ".text".to_owned(), 16)]);
1410    }
1411
1412    /// What each record in the unwind table points at: where it is, the section it reaches, and
1413    /// how far into that section the function it is about begins.
1414    fn points_at(file: &object::File<'_>) -> Vec<(u64, String, i64)> {
1415        let frames = file.section_by_name(".eh_frame").expect("the table");
1416        frames
1417            .relocations()
1418            .map(|(offset, reloc)| {
1419                let object::RelocationTarget::Symbol(index) = reloc.target() else {
1420                    panic!("a record points at something that is not a symbol");
1421                };
1422                let symbol = file.symbol_by_index(index).expect("a symbol that is in the table");
1423                assert_eq!(symbol.kind(), SymbolKind::Section, "a record names a section");
1424                let section = symbol.section_index().expect("a section symbol is in one");
1425                let name = file.section_by_index(section).expect("a readable section");
1426                (offset, name.name().expect("a named section").to_owned(), reloc.addend())
1427            })
1428            .collect()
1429    }
1430
1431    /// A record points at the section its function is in rather than at the function's name.
1432    ///
1433    /// Written for tamnd/rucc#1004, which was that nothing this compiler wrote could go into a
1434    /// shared library. A global name is answered at load time by whichever object defines it
1435    /// first, so the distance from a record to one of them is not a distance a static linker can
1436    /// work out, and `ld` says so and stops with advice to recompile with the flag that was
1437    /// already on the command line. A section is settled by then, which is why gcc measures to a
1438    /// local label and why this measures to the section.
1439    ///
1440    /// Both ways of splitting the text, because the offset is the part that differs: one section
1441    /// holding everything makes it the function's place in the whole text, and a section per
1442    /// function makes it whatever room a patcher was promised in front of the label.
1443    #[test]
1444    fn a_record_reaches_its_function_through_the_section_it_is_in() {
1445        let mut text = two();
1446        text.unwind.bytes = vec![0; 64];
1447        for (at, name) in [(32usize, "f"), (48usize, "g")] {
1448            text.unwind.relocs.push(Reloc {
1449                at,
1450                symbol: name.to_owned(),
1451                kind: Reference::Data,
1452                addend: 0,
1453                after: 0,
1454            });
1455        }
1456        let bytes =
1457            write(&text, &Data::default(), &[], &target(), Output::default(), &Info::default())
1458                .expect("an object");
1459        let file = object::File::parse(&bytes[..]).expect("a readable object");
1460        let mut whole = points_at(&file);
1461        whole.sort_unstable();
1462        assert_eq!(whole, [(32, ".text".to_owned(), 0), (48, ".text".to_owned(), 16)]);
1463
1464        let sections =
1465            Output { sections: Sections { functions: true, data: false }, ..Output::default() };
1466        let bytes = write(&text, &Data::default(), &[], &target(), sections, &Info::default())
1467            .expect("an object");
1468        let file = object::File::parse(&bytes[..]).expect("a readable object");
1469        let mut split = points_at(&file);
1470        split.sort_unstable();
1471        assert_eq!(split, [(32, ".text.f".to_owned(), 0), (48, ".text.g".to_owned(), 0)]);
1472    }
1473
1474    /// A record about a name this file does not define is refused rather than written.
1475    ///
1476    /// There is no such file today: the table is built beside the text out of the functions that
1477    /// were just compiled. It is refused rather than left to the linker because the alternative is
1478    /// the shape that was just fixed, a record measured to a name, and the writer saying what it
1479    /// was given is how that stays fixed.
1480    #[test]
1481    fn a_record_about_something_this_file_does_not_define_is_refused() {
1482        let mut text = calling("puts");
1483        text.unwind.bytes = vec![0; 64];
1484        text.unwind.relocs.push(Reloc {
1485            at: 32,
1486            symbol: "puts".to_owned(),
1487            kind: Reference::Data,
1488            addend: 0,
1489            after: 0,
1490        });
1491        let why =
1492            write(&text, &Data::default(), &[], &target(), Output::default(), &Info::default())
1493                .expect_err("a record about a name from somewhere else");
1494        assert!(why.to_string().contains("puts"), "{why}");
1495    }
1496
1497    /// The name of the section that symbol is defined in.
1498    fn lives_in<'a>(file: &'a object::File<'a>, name: &str) -> String {
1499        let symbol = file.symbols().find(|s| s.name() == Ok(name)).expect("the symbol");
1500        let index = symbol.section_index().expect("a section to be defined in");
1501        let section = file.section_by_index(index).expect("a readable section");
1502        section.name().expect("a named section").to_owned()
1503    }
1504
1505    /// Two functions, the second of them sixteen bytes in and calling something outside the file.
1506    fn two() -> Text {
1507        let mut text = calling("puts");
1508        // Padded to where the second one is aligned to, with the instruction that does nothing,
1509        // because the space in front of a function is reached by falling off the end of one.
1510        text.bytes.resize(16, 0x90);
1511        text.bytes.extend_from_slice(&[0xe8, 0, 0, 0, 0, 0xc3]);
1512        text.funcs.push(extent("g".to_owned(), 16, 6, Binding::Global));
1513        text.relocs.push(Reloc {
1514            at: 17,
1515            symbol: "puts".to_owned(),
1516            kind: Reference::Call,
1517            addend: -4,
1518            after: 0,
1519        });
1520        text
1521    }
1522
1523    /// What `-ffunction-sections` comes down to in an object file, which is the flag that makes
1524    /// `--gc-sections` able to drop anything: a linker can leave out a section nothing reaches and
1525    /// cannot leave out half of one.
1526    ///
1527    /// The empty `.text` stays, because it is the section the writer underneath opens a file with
1528    /// and gcc 16 leaves an empty one behind under the flag too.
1529    #[test]
1530    fn every_function_gets_a_section_of_its_own_when_that_is_what_was_asked_for() {
1531        let sections =
1532            Output { sections: Sections { functions: true, data: false }, ..Output::default() };
1533        let bytes = write(&two(), &Data::default(), &[], &target(), sections, &Info::default())
1534            .expect("an object");
1535        let file = object::File::parse(&bytes[..]).expect("a readable object");
1536        assert_eq!(lives_in(&file, "f"), ".text.f");
1537        assert_eq!(lives_in(&file, "g"), ".text.g");
1538        assert!(file.section_by_name(".text").expect("the empty one").size() == 0);
1539        // Each one at nothing into its own section, and as long as it was: a function alone in a
1540        // section starts where the section does, whatever it started at when they shared one.
1541        for name in ["f", "g"] {
1542            let symbol = file.symbols().find(|s| s.name() == Ok(name)).expect("the function");
1543            assert_eq!(symbol.address(), 0, "{name}");
1544            assert_eq!(symbol.size(), 6, "{name}");
1545        }
1546        let section = file.section_by_name(".text.g").expect("the second function");
1547        assert_eq!(section.data().expect("the bytes"), &[0xe8, 0, 0, 0, 0, 0xc3]);
1548        // The padding between the two is gone with them, since it was there to align the second
1549        // one inside a section they shared and each section is aligned by the linker now.
1550        assert_eq!(section.align(), u64::from(crate::FUNC_ALIGN));
1551    }
1552
1553    /// A relocation counts from the start of whichever section its function ended up in, which is
1554    /// the arithmetic the split path has to do and the unsplit one never does.
1555    ///
1556    /// Getting it wrong is a call patched over the wrong bytes, which assembles, links, and jumps
1557    /// into the middle of an instruction at run time.
1558    #[test]
1559    fn a_relocation_moves_with_the_function_whose_bytes_it_is_in() {
1560        let sections =
1561            Output { sections: Sections { functions: true, data: false }, ..Output::default() };
1562        let bytes = write(&two(), &Data::default(), &[], &target(), sections, &Info::default())
1563            .expect("an object");
1564        let file = object::File::parse(&bytes[..]).expect("a readable object");
1565        for name in [".text.f", ".text.g"] {
1566            let section = file.section_by_name(name).expect("a function");
1567            let (offset, _) = section.relocations().next().expect("the call in it");
1568            // One byte in either way, because the call is the first instruction of both and the
1569            // opcode is one byte in front of the address the linker fills in.
1570            assert_eq!(offset, 1, "{name}");
1571            assert_eq!(section.relocations().count(), 1, "{name}");
1572        }
1573    }
1574
1575    /// One variable of four bytes, in whichever section its own answer puts it.
1576    fn variable(name: &str, place: Place) -> Object {
1577        Object {
1578            name: name.to_owned(),
1579            bytes: if carries_no_bytes(&place) { Vec::new() } else { vec![1, 0, 0, 0] },
1580            size: 4,
1581            align: 4,
1582            place,
1583            binding: Binding::Global,
1584            visibility: Visibility::Default,
1585            relocs: Vec::new(),
1586        }
1587    }
1588
1589    /// Two labels in `f` and an image holding the distance between them each way round.
1590    fn measured() -> (Text, Data) {
1591        let mut text = calling("puts");
1592        text.labels.push(Marker { name: ".L0".to_owned(), at: 1 });
1593        text.labels.push(Marker { name: ".L1".to_owned(), at: 5 });
1594        let mut table = variable("table", Place::ReadOnly);
1595        table.bytes = vec![0; 8];
1596        table.size = 8;
1597        let apart = |at, to: &str, from: &str| Apart {
1598            object: 0,
1599            at,
1600            to: to.to_owned(),
1601            from: from.to_owned(),
1602            addend: 0,
1603            bytes: 4,
1604        };
1605        let apart = vec![apart(0, ".L1", ".L0"), apart(4, ".L0", ".L1")];
1606        (text, Data { apart, weak: Vec::new(), objects: vec![table] })
1607    }
1608
1609    #[test]
1610    fn a_distance_between_two_labels_is_a_number_and_not_a_relocation() {
1611        let (text, data) = measured();
1612        let bytes = write(&text, &data, &[], &target(), Output::default(), &Info::default())
1613            .expect("an object");
1614        let file = object::File::parse(&bytes[..]).expect("a readable object");
1615        let section = file.section_by_name(".rodata").expect("a read only section");
1616        assert_eq!(section.relocations().count(), 0);
1617        let image = section.data().expect("the image");
1618        assert_eq!(image[..8], [4, 0, 0, 0, 0xfc, 0xff, 0xff, 0xff]);
1619    }
1620
1621    #[test]
1622    fn a_distance_between_labels_in_two_sections_is_refused() {
1623        // `.L1` moves to a second function, which `-ffunction-sections` puts in a section of its
1624        // own, and then no number is the distance.
1625        let (mut text, data) = measured();
1626        text.bytes.resize(22, 0x90);
1627        text.funcs.push(extent("g".to_owned(), 16, 6, Binding::Global));
1628        text.labels[1].at = 17;
1629        let output =
1630            Output { sections: Sections { functions: true, data: false }, ..Output::default() };
1631        let refused = write(&text, &data, &[], &target(), output, &Info::default());
1632        assert!(matches!(refused, Err(Error::Refused { .. })), "{refused:?}");
1633    }
1634
1635    /// A file of that one variable and nothing else.
1636    fn holding(object: Object) -> Vec<u8> {
1637        let data = Data { apart: Vec::new(), weak: Vec::new(), objects: vec![object] };
1638        write(&Text::default(), &data, &[], &target(), Output::default(), &Info::default())
1639            .expect("an object")
1640    }
1641
1642    #[test]
1643    fn what_a_variable_is_decides_which_section_it_goes_in() {
1644        for (place, wanted) in [
1645            (Place::Written, ".data"),
1646            (Place::ReadOnly, ".rodata"),
1647            (Place::RelocReadOnly { local: false }, ".data.rel.ro"),
1648            (Place::RelocReadOnly { local: true }, ".data.rel.ro.local"),
1649            (Place::Zero, ".bss"),
1650            (Place::Thread { zero: false }, ".tdata"),
1651            (Place::Thread { zero: true }, ".tbss"),
1652            (Place::Named(".init_array".to_owned()), ".init_array"),
1653        ] {
1654            let bytes = holding(variable("x", place.clone()));
1655            let file = object::File::parse(&bytes[..]).expect("a readable object");
1656            let section = file.section_by_name(wanted).unwrap_or_else(|| panic!("{place:?}"));
1657            assert_eq!(section.size(), 4, "{place:?}");
1658            // The zero filled one is as long as it says and carries none of it, which is the
1659            // whole reason the section exists.
1660            let carried = section.data().expect("the bytes").len();
1661            assert_eq!(carried, if carries_no_bytes(&place) { 0 } else { 4 }, "{place:?}");
1662        }
1663    }
1664
1665    /// The section is half of it and the symbol is the other half.
1666    ///
1667    /// A linker checks a relocation against the kind of the symbol it names, so a variable that is
1668    /// in `.tdata` and is an ordinary data symbol is one an ordinary reference resolves to an
1669    /// address that belongs to no thread. `STT_TLS` is what makes that reference an error instead.
1670    #[test]
1671    fn a_thread_local_variable_is_a_thread_local_symbol_and_not_only_a_thread_local_section() {
1672        for place in [Place::Thread { zero: false }, Place::Thread { zero: true }] {
1673            let bytes = holding(variable("counter", place.clone()));
1674            let file = object::File::parse(&bytes[..]).expect("a readable object");
1675            let symbol = file
1676                .symbols()
1677                .find(|symbol| symbol.name() == Ok("counter"))
1678                .unwrap_or_else(|| panic!("{place:?}"));
1679            assert_eq!(symbol.kind(), SymbolKind::Tls, "{place:?}");
1680        }
1681    }
1682
1683    /// The section type a startup list carries, which is what makes the CRT call what is in it.
1684    ///
1685    /// A section of the ordinary type with the right name is gathered by the linker in the same run
1686    /// and called by nobody, so the type is the whole of what this is about. The numbered name is
1687    /// the same kind of section as the plain one: the number is there so that the linker sorts it.
1688    #[test]
1689    fn a_section_of_function_addresses_carries_the_type_the_runtime_looks_for() {
1690        for (name, wanted) in [
1691            (".init_array", elf::SHT_INIT_ARRAY),
1692            (".init_array.00101", elf::SHT_INIT_ARRAY),
1693            (".fini_array", elf::SHT_FINI_ARRAY),
1694            (".preinit_array", elf::SHT_PREINIT_ARRAY),
1695            (".init_arrays", elf::SHT_PROGBITS),
1696        ] {
1697            let bytes = holding(variable("x", Place::Named(name.to_owned())));
1698            let file = object::File::parse(&bytes[..]).expect("a readable object");
1699            let section = file.section_by_name(name).unwrap_or_else(|| panic!("{name}"));
1700            let SectionFlags::Elf { sh_type, sh_flags } = section.flags() else {
1701                panic!("{name} is not an elf section");
1702            };
1703            assert_eq!(sh_type, wanted, "{name}");
1704            assert!(sh_flags.contains(elf::SHF_ALLOC | elf::SHF_WRITE), "{name}");
1705        }
1706    }
1707
1708    /// Two variables the program put one section name on, which belong in one section.
1709    ///
1710    /// A file with ten constructors in it would otherwise carry ten section headers describing eight
1711    /// bytes each, and the order the entries run in would be the order the linker happened to put
1712    /// the headers in rather than the order they were written.
1713    #[test]
1714    fn two_variables_in_one_named_section_share_it() {
1715        let objects = vec![
1716            variable("x", Place::Named(".init_array".to_owned())),
1717            variable("y", Place::Named(".init_array".to_owned())),
1718        ];
1719        let data = Data { apart: Vec::new(), weak: Vec::new(), objects };
1720        let bytes =
1721            write(&Text::default(), &data, &[], &target(), Output::default(), &Info::default())
1722                .expect("an object");
1723        let file = object::File::parse(&bytes[..]).expect("a readable object");
1724        let named: Vec<_> =
1725            file.sections().filter(|section| section.name() == Ok(".init_array")).collect();
1726        assert_eq!(named.len(), 1);
1727        assert_eq!(named[0].size(), 8);
1728    }
1729
1730    /// What `-fdata-sections` comes down to in an object file: the section a variable would have
1731    /// shared, with its own name after it. The names are gcc 16's, checked against it on a Linux
1732    /// host, and the part in front of the dot is what a linker script and `--gc-sections` match on.
1733    #[test]
1734    fn every_variable_gets_a_section_of_its_own_when_that_is_what_was_asked_for() {
1735        let sections =
1736            Output { sections: Sections { functions: false, data: true }, ..Output::default() };
1737        for (place, wanted) in [
1738            (Place::Written, ".data.x"),
1739            (Place::ReadOnly, ".rodata.x"),
1740            (Place::RelocReadOnly { local: false }, ".data.rel.ro.x"),
1741            (Place::RelocReadOnly { local: true }, ".data.rel.ro.local.x"),
1742            (Place::Zero, ".bss.x"),
1743            (Place::Thread { zero: false }, ".tdata.x"),
1744            (Place::Thread { zero: true }, ".tbss.x"),
1745        ] {
1746            let data = Data {
1747                apart: Vec::new(),
1748                weak: Vec::new(),
1749                objects: vec![variable("x", place.clone())],
1750            };
1751            let bytes = write(&Text::default(), &data, &[], &target(), sections, &Info::default())
1752                .expect("object");
1753            let file = object::File::parse(&bytes[..]).expect("a readable object");
1754            assert_eq!(lives_in(&file, "x"), wanted, "{place:?}");
1755            let section = file.section_by_name(wanted).expect("the section it named");
1756            assert_eq!(section.size(), 4, "{place:?}");
1757            // Which page it lands in is what the section it came out of decided, and splitting
1758            // must not quietly change it: the zero filled one still carries none of its bytes.
1759            let carried = section.data().expect("the bytes").len();
1760            assert_eq!(carried, if carries_no_bytes(&place) { 0 } else { 4 }, "{place:?}");
1761        }
1762    }
1763
1764    /// The two kinds of variable the flag leaves alone. A tentative definition is a request to the
1765    /// linker for that much zeroed space rather than an image, so there is no section to split off,
1766    /// and one the program named has the answer the source gave, which a flag must not overrule.
1767    #[test]
1768    fn a_variable_that_has_no_section_of_its_own_to_be_given_is_left_where_it_was() {
1769        let sections =
1770            Output { sections: Sections { functions: false, data: true }, ..Output::default() };
1771        let named = Place::Named(".init_array".to_owned());
1772        let objects = vec![variable("m", Place::Merged), variable("n", named)];
1773        let bytes = write(
1774            &Text::default(),
1775            &Data { apart: Vec::new(), weak: Vec::new(), objects },
1776            &[],
1777            &target(),
1778            sections,
1779            &Info::default(),
1780        )
1781        .expect("object");
1782        let file = object::File::parse(&bytes[..]).expect("a readable object");
1783        let m = file.symbols().find(|s| s.name() == Ok("m")).expect("the tentative one");
1784        assert!(m.is_common(), "still the linker's to merge and not in a section at all");
1785        assert_eq!(lives_in(&file, "n"), ".init_array");
1786        assert!(file.section_by_name(".init_array.n").is_none(), "the source already answered");
1787    }
1788
1789    /// A relocation in a variable's image counts from the start of the section it ended up in, the
1790    /// same question the split text has to answer and a shorter answer: a variable alone in a
1791    /// section starts where the section does.
1792    #[test]
1793    fn a_relocation_in_an_image_moves_with_the_variable_whose_image_it_is_in() {
1794        let sections =
1795            Output { sections: Sections { functions: false, data: true }, ..Output::default() };
1796        let pointer = Object {
1797            bytes: vec![0; 8],
1798            size: 8,
1799            align: 8,
1800            relocs: vec![Reloc {
1801                at: 0,
1802                symbol: "y".to_owned(),
1803                kind: Reference::Address { bytes: 8 },
1804                addend: 0,
1805                after: 0,
1806            }],
1807            ..variable("p", Place::Written)
1808        };
1809        let objects = vec![variable("first", Place::Written), pointer];
1810        let bytes = write(
1811            &Text::default(),
1812            &Data { apart: Vec::new(), weak: Vec::new(), objects },
1813            &[],
1814            &target(),
1815            sections,
1816            &Info::default(),
1817        )
1818        .expect("object");
1819        let file = object::File::parse(&bytes[..]).expect("a readable object");
1820        let section = file.section_by_name(".data.p").expect("the pointer's own section");
1821        let (offset, reloc) = section.relocations().next().expect("one relocation");
1822        // Nothing rather than the eight it would be if the variable in front of it were still
1823        // counted, which is what a section of its own means.
1824        assert_eq!(offset, 0);
1825        assert_eq!(reloc.flags(), RelocationFlags::Elf { r_type: elf::R_X86_64_64 });
1826    }
1827
1828    /// Two variables that want `.data.rel.ro.local` end up in one section, not two of one name.
1829    ///
1830    /// The writer has no name of its own for that section, so it is added by hand, and asking for
1831    /// it again makes a second section rather than handing back the first. SQLite has enough const
1832    /// tables of function pointers in it to turn that into eighty odd sections in one object, each
1833    /// with its own relocation section beside it, which is a pile of section headers describing
1834    /// eight bytes apiece.
1835    #[test]
1836    fn every_variable_that_wants_the_local_relocated_section_shares_one() {
1837        let place = Place::RelocReadOnly { local: true };
1838        let data = Data {
1839            apart: Vec::new(),
1840            weak: Vec::new(),
1841            objects: vec![variable("first", place.clone()), variable("second", place)],
1842        };
1843        let bytes =
1844            write(&Text::default(), &data, &[], &target(), Output::default(), &Info::default())
1845                .expect("an object");
1846        let file = object::File::parse(&bytes[..]).expect("a readable object");
1847        let named = file.sections().filter(|s| s.name() == Ok(".data.rel.ro.local")).count();
1848        assert_eq!(named, 1, "one section holding both, not one each");
1849    }
1850
1851    #[test]
1852    fn a_variable_is_a_symbol_that_says_where_it_is_and_how_long_it_is() {
1853        let mut data = Data {
1854            apart: Vec::new(),
1855            weak: Vec::new(),
1856            objects: vec![variable("first", Place::Written)],
1857        };
1858        data.objects.push(Object { align: 16, ..variable("second", Place::Written) });
1859        let bytes =
1860            write(&Text::default(), &data, &[], &target(), Output::default(), &Info::default())
1861                .expect("an object");
1862        let file = object::File::parse(&bytes[..]).expect("a readable object");
1863        let second = file.symbols().find(|s| s.name() == Ok("second")).expect("the second one");
1864        assert_eq!(second.kind(), SymbolKind::Data);
1865        assert_eq!(second.size(), 4);
1866        // Sixteen rather than four, because the second one asked for sixteen and the first one
1867        // had already used four. Getting this wrong is a variable at an address it said it would
1868        // never be at, which nothing downstream would notice until an aligned load faulted.
1869        assert_eq!(second.address(), 16);
1870    }
1871
1872    #[test]
1873    fn the_linkage_a_variable_had_is_the_binding_the_symbol_gets() {
1874        for (binding, global, weak) in [
1875            (Binding::Global, true, false),
1876            (Binding::Local, false, false),
1877            (Binding::Weak, true, true),
1878        ] {
1879            let bytes = holding(Object { binding, ..variable("x", Place::Written) });
1880            let file = object::File::parse(&bytes[..]).expect("a readable object");
1881            let x = file.symbols().find(|s| s.name() == Ok("x")).expect("the variable");
1882            assert_eq!(x.is_global(), global, "{binding:?}");
1883            assert_eq!(x.is_weak(), weak, "{binding:?}");
1884        }
1885    }
1886
1887    #[test]
1888    fn a_tentative_definition_asks_the_linker_for_space_rather_than_naming_any() {
1889        let bytes = holding(Object { align: 8, ..variable("x", Place::Merged) });
1890        let file = object::read::elf::ElfFile64::<Endianness>::parse(&bytes[..]).expect("readable");
1891        let x = file.symbols().find(|s| s.name() == Ok("x")).expect("the variable");
1892        assert!(x.is_common(), "the linker merges every definition of this name into one");
1893        assert_eq!(x.size(), 4);
1894        // What a common symbol records where an ordinary one records its address is what it wants
1895        // to be aligned to, because it has no address yet. The reader deliberately answers nothing
1896        // when asked for the address of one, so this is the field itself.
1897        assert_eq!(x.address(), 0);
1898        assert_eq!(x.elf_symbol().st_value(Endianness::Little), 8);
1899    }
1900
1901    #[test]
1902    fn an_address_in_an_image_is_the_address_and_not_a_distance_to_it() {
1903        let object = Object {
1904            bytes: vec![0; 8],
1905            size: 8,
1906            align: 8,
1907            relocs: vec![Reloc {
1908                at: 0,
1909                symbol: "y".to_owned(),
1910                kind: Reference::Address { bytes: 8 },
1911                addend: 16,
1912                after: 0,
1913            }],
1914            ..variable("p", Place::Written)
1915        };
1916        let bytes = holding(object);
1917        let file = object::File::parse(&bytes[..]).expect("a readable object");
1918        let section = file.section_by_name(".data").expect("a data section");
1919        let (offset, reloc) = section.relocations().next().expect("one relocation");
1920        assert_eq!(offset, 0);
1921        assert_eq!(reloc.addend(), 16);
1922        assert_eq!(reloc.flags(), RelocationFlags::Elf { r_type: elf::R_X86_64_64 });
1923        let y = file.symbols().find(|s| s.name() == Ok("y")).expect("what it points at");
1924        assert!(y.is_undefined(), "nothing here defines it and the linker is being asked for it");
1925    }
1926
1927    /// A name a declaration wrote `weak` on is undefined and may stay that way.
1928    ///
1929    /// The difference between this and the case above is one bit and the whole of what a link does
1930    /// about it: an ordinary undefined symbol is a name the linker has to find, and a weak one is a
1931    /// name it may fail to find, in which case every reference reads a zero address. That is what
1932    /// lets a library offer a hook a profiler may fill in, which is tamnd/rucc#1414.
1933    #[test]
1934    fn a_weak_undefined_name_is_one_the_link_may_leave_unfound() {
1935        let mut text = Text::default();
1936        text.funcs.push(extent("caller".to_owned(), 0, 8, Binding::Global));
1937        text.bytes.resize(8, 0x90);
1938        text.relocs.push(Reloc {
1939            at: 1,
1940            symbol: "hook".to_owned(),
1941            kind: Reference::Call,
1942            addend: -4,
1943            after: 0,
1944        });
1945        let data = Data {
1946            apart: Vec::new(),
1947            weak: vec!["hook".to_owned(), "never_called".to_owned()],
1948            objects: vec![],
1949        };
1950        let bytes = write(&text, &data, &[], &target(), Output::default(), &Info::default())
1951            .expect("an object");
1952        let file = object::File::parse(&bytes[..]).expect("a readable object");
1953
1954        let hook = file.symbols().find(|s| s.name() == Ok("hook")).expect("the one called");
1955        assert!(hook.is_undefined(), "nothing here defines it");
1956        assert!(hook.is_weak(), "so the link may leave it alone rather than fail");
1957
1958        // And one nothing refers to is still written down, because the listing writes a directive
1959        // for it and the two paths have to put the same entries in. A linker has nothing to do
1960        // about an undefined weak symbol no relocation names.
1961        let quiet = file.symbols().find(|s| s.name() == Ok("never_called")).expect("the other");
1962        assert!(quiet.is_undefined() && quiet.is_weak(), "{:?}", quiet.flags());
1963    }
1964
1965    /// A name this file reads through the thread pointer is undefined and is still known to be
1966    /// thread-local.
1967    ///
1968    /// The other undefined names here are written with no type at all, because a name this file does
1969    /// not define is a name this file has nothing to say about. A thread-local one is different in
1970    /// the one way that counts: a reference to it is satisfied by an offset into a block rather than
1971    /// by an address, so the linker has to know which of the two is wanted before it has found the
1972    /// definition, and rather than guess it refuses a link where one file says `STT_TLS` about a name
1973    /// and another does not. Writing the type is not extra information, it is the same information
1974    /// the relocation already carried, said where the linker looks for it.
1975    ///
1976    /// That is tamnd/rucc#1461. libmpfr defines `__gmpfr_flags` in `exceptions.c` and reads it in a
1977    /// hundred other files, and the link stopped at the first reader with `TLS definition in
1978    /// exceptions.o section .tdata mismatches non-TLS reference in add.o`.
1979    #[test]
1980    fn a_thread_local_name_this_file_only_reads_is_still_written_down_as_thread_local() {
1981        let mut text = Text::default();
1982        text.funcs.push(extent("reader".to_owned(), 0, 16, Binding::Global));
1983        text.bytes.resize(16, 0x90);
1984        text.relocs.push(Reloc {
1985            at: 3,
1986            symbol: "flags".to_owned(),
1987            kind: Reference::Thread,
1988            addend: -4,
1989            after: 0,
1990        });
1991        // One of them reached the ordinary way, so that what the type says is the relocation's doing
1992        // and not something every undefined name here would have got.
1993        text.relocs.push(Reloc {
1994            at: 10,
1995            symbol: "shared".to_owned(),
1996            kind: Reference::Got,
1997            addend: -4,
1998            after: 0,
1999        });
2000        let data = Data { apart: Vec::new(), weak: Vec::new(), objects: vec![] };
2001        let bytes = write(&text, &data, &[], &target(), Output::default(), &Info::default())
2002            .expect("an object");
2003        let file = object::File::parse(&bytes[..]).expect("a readable object");
2004
2005        let flags = file.symbols().find(|s| s.name() == Ok("flags")).expect("the thread-local one");
2006        assert!(flags.is_undefined(), "nothing here defines it");
2007        assert_eq!(flags.kind(), SymbolKind::Tls, "which is what the linker refuses to guess");
2008
2009        let shared = file.symbols().find(|s| s.name() == Ok("shared")).expect("the ordinary one");
2010        assert!(shared.is_undefined(), "nothing here defines this one either");
2011        assert_eq!(shared.kind(), SymbolKind::Unknown, "and there is nothing to say about it");
2012    }
2013
2014    /// Not a rewording of the case above: what is checked is the arithmetic between the two.
2015    #[test]
2016    fn a_relocation_counts_from_the_start_of_the_section_and_not_of_the_image_it_is_in() {
2017        let mut data = Data {
2018            apart: Vec::new(),
2019            weak: Vec::new(),
2020            objects: vec![variable("first", Place::Written)],
2021        };
2022        data.objects.push(Object {
2023            bytes: vec![0; 16],
2024            size: 16,
2025            align: 8,
2026            relocs: vec![Reloc {
2027                at: 8,
2028                symbol: "y".to_owned(),
2029                kind: Reference::Address { bytes: 8 },
2030                addend: 0,
2031                after: 0,
2032            }],
2033            ..variable("second", Place::Written)
2034        });
2035        let bytes =
2036            write(&Text::default(), &data, &[], &target(), Output::default(), &Info::default())
2037                .expect("an object");
2038        let file = object::File::parse(&bytes[..]).expect("a readable object");
2039        let section = file.section_by_name(".data").expect("a data section");
2040        let (offset, _) = section.relocations().next().expect("one relocation");
2041        // Eight into the second image, which starts eight in because the first one is four long
2042        // and the second is eight aligned.
2043        assert_eq!(offset, 16);
2044    }
2045
2046    #[test]
2047    fn a_second_name_is_a_second_symbol_at_the_first_one_s_address_and_no_second_image() {
2048        let data = Data {
2049            apart: Vec::new(),
2050            weak: Vec::new(),
2051            objects: vec![Object { binding: Binding::Local, ..variable("a", Place::Written) }],
2052        };
2053        let aliases = [Alias {
2054            name: "b".to_owned(),
2055            target: "a".to_owned(),
2056            binding: Binding::Global,
2057            visibility: Visibility::Default,
2058        }];
2059        let bytes = write(
2060            &Text::default(),
2061            &data,
2062            &aliases,
2063            &target(),
2064            Output::default(),
2065            &Info::default(),
2066        )
2067        .expect("an object");
2068        let file = object::File::parse(&bytes[..]).expect("a readable object");
2069        let a = file.symbols().find(|s| s.name() == Ok("a")).expect("the variable");
2070        let b = file.symbols().find(|s| s.name() == Ok("b")).expect("the second name");
2071        assert_eq!(b.address(), a.address(), "the same place");
2072        assert_eq!(b.size(), a.size());
2073        assert_eq!(b.section_index(), a.section_index());
2074        // The binding is the one thing the second name does not take from the first, which is
2075        // what `extern int b __attribute__((alias("a")))` on a `static a` asks for.
2076        assert!(a.is_local(), "the target was written `static`");
2077        assert!(b.is_global(), "and the name given to it was not");
2078        // Four bytes of image and not eight, since an alias is a name and not a copy.
2079        assert_eq!(file.section_by_name(".data").expect("a data section").size(), 4);
2080    }
2081
2082    #[test]
2083    fn a_function_can_be_given_a_second_name_the_same_way_a_variable_can() {
2084        let text = calling("puts");
2085        let aliases = [Alias {
2086            name: "g".to_owned(),
2087            target: "f".to_owned(),
2088            binding: Binding::Weak,
2089            visibility: Visibility::Default,
2090        }];
2091        let bytes = write(
2092            &text,
2093            &Data::default(),
2094            &aliases,
2095            &target(),
2096            Output::default(),
2097            &Info::default(),
2098        )
2099        .expect("an object");
2100        let file = object::File::parse(&bytes[..]).expect("a readable object");
2101        let f = file.symbols().find(|s| s.name() == Ok("f")).expect("the function");
2102        let g = file.symbols().find(|s| s.name() == Ok("g")).expect("the second name");
2103        assert_eq!(g.address(), f.address());
2104        assert_eq!(g.size(), f.size());
2105        assert_eq!(g.kind(), f.kind(), "a second name for a function is a function");
2106        assert!(g.is_weak(), "so that a program may define the name itself instead");
2107    }
2108
2109    /// The front end is what reports this as a program's mistake, so one arriving here is a bug
2110    /// in this compiler and is said so rather than written as an undefined symbol.
2111    #[test]
2112    fn a_second_name_for_something_this_file_does_not_define_is_refused() {
2113        let aliases = [Alias {
2114            name: "b".to_owned(),
2115            target: "a".to_owned(),
2116            binding: Binding::Global,
2117            visibility: Visibility::Default,
2118        }];
2119        let error = write(
2120            &Text::default(),
2121            &Data::default(),
2122            &aliases,
2123            &target(),
2124            Output::default(),
2125            &Info::default(),
2126        )
2127        .expect_err("nothing to point at");
2128        assert!(matches!(error, Error::Refused { .. }), "{error:?}");
2129    }
2130
2131    #[test]
2132    fn a_platform_this_does_not_write_is_said_so_rather_than_written_as_elf() {
2133        let text = calling("puts");
2134        for triple in [
2135            Triple::new(Arch::Aarch64, Os::Linux, Env::Gnu),
2136            Triple::new(Arch::X86_64, Os::Darwin, Env::Gnu),
2137        ] {
2138            let error = write(
2139                &text,
2140                &Data::default(),
2141                &[],
2142                &TargetInfo::new(triple),
2143                Output::default(),
2144                &Info::default(),
2145            )
2146            .expect_err("no writer");
2147            assert!(matches!(error, Error::Format { .. }), "{error:?}");
2148        }
2149    }
2150
2151    /// What the archive's symbol index is built from is what the linker can find in the member.
2152    ///
2153    /// Written against the object rather than against the list, because the two agreeing is the
2154    /// whole point: a list that says more than the file does is an archive that promises a
2155    /// definition it does not have, and a list that says less is a member nothing pulls out.
2156    #[test]
2157    fn the_names_a_linker_can_find_are_the_names_the_list_gives() {
2158        let mut text = calling("puts");
2159        text.funcs.push(extent("hidden".to_owned(), 16, 1, Binding::Local));
2160        text.funcs.push(extent("shared".to_owned(), 32, 1, Binding::Weak));
2161        text.bytes.resize(33, 0x90);
2162        let data = Data {
2163            apart: Vec::new(),
2164            weak: Vec::new(),
2165            objects: vec![variable("seen", Place::Written), {
2166                let mut quiet = variable("quiet", Place::Zero);
2167                quiet.binding = Binding::Local;
2168                quiet
2169            }],
2170        };
2171        let aliases = [Alias {
2172            name: "second".to_owned(),
2173            target: "f".to_owned(),
2174            binding: Binding::Global,
2175            visibility: Visibility::Default,
2176        }];
2177
2178        let names = defines(&text, &data, &aliases, &target()).expect("a list");
2179        assert_eq!(names, ["f", "shared", "seen", "second"]);
2180
2181        let bytes = write(&text, &data, &aliases, &target(), Output::default(), &Info::default())
2182            .expect("an object");
2183        let file = object::File::parse(&bytes[..]).expect("a readable object");
2184        let found: Vec<String> = file
2185            .symbols()
2186            .filter(|symbol| symbol.is_global() && symbol.is_definition())
2187            .map(|symbol| symbol.name().unwrap_or_default().to_owned())
2188            .collect();
2189        let mut sorted = names.clone();
2190        sorted.sort();
2191        let mut theirs = found;
2192        theirs.sort();
2193        assert_eq!(sorted, theirs, "the list and the file have to say the same thing");
2194    }
2195
2196    /// A windows x86-64 target, which is the other format this writes.
2197    fn windows() -> TargetInfo {
2198        TargetInfo::new(Triple::new(Arch::X86_64, Os::Windows, Env::Gnu))
2199    }
2200
2201    /// What the four bytes a relocation covers hold, which is where COFF keeps its addend.
2202    fn inline(bytes: &[u8], section: &str, at: usize) -> i32 {
2203        let file = object::File::parse(bytes).expect("a readable object");
2204        let found = file.section_by_name(section).expect("the section").data().expect("the bytes");
2205        i32::from_le_bytes(found[at..at + 4].try_into().expect("four bytes"))
2206    }
2207
2208    #[test]
2209    fn a_windows_target_is_written_rather_than_refused() {
2210        let text = calling("puts");
2211        let bytes =
2212            write(&text, &Data::default(), &[], &windows(), Output::default(), &Info::default())
2213                .expect("an object");
2214        let file = object::File::parse(&bytes[..]).expect("a readable object");
2215        assert_eq!(file.format(), BinaryFormat::Coff);
2216        let section = file.section_by_name(".text").expect("a text section");
2217        assert_eq!(section.data().expect("the bytes"), &text.bytes[..]);
2218        let names: Vec<&str> = file.symbols().filter_map(|symbol| symbol.name().ok()).collect();
2219        assert!(names.contains(&"f"), "{names:?}");
2220        assert!(names.contains(&"puts"), "{names:?}");
2221    }
2222
2223    /// The whole reason a relocation carries where the instruction ended as well as the addend.
2224    ///
2225    /// A call ends at the four bytes the linker writes over, and a store of a constant through an
2226    /// address counted from the instruction pointer has the constant after them, and ELF tells the
2227    /// two apart by the addend alone. COFF cannot: it says how far the end is in the relocation type
2228    /// and works the addend out from that, so the same four bytes come out of two different types
2229    /// and both have to end up meaning the same distance.
2230    #[test]
2231    fn how_far_the_instruction_runs_past_the_hole_is_in_the_relocation_type() {
2232        for (after, typ) in [
2233            (0, pe::IMAGE_REL_AMD64_REL32),
2234            (1, pe::IMAGE_REL_AMD64_REL32_1),
2235            (4, pe::IMAGE_REL_AMD64_REL32_4),
2236            (5, pe::IMAGE_REL_AMD64_REL32_5),
2237        ] {
2238            let mut text = calling("puts");
2239            // The same distance every time, said the way ELF says it: from where the four bytes
2240            // start, with everything else folded in.
2241            text.relocs[0].addend = -4 - i64::from(after);
2242            text.relocs[0].after = after;
2243            text.bytes.resize(6 + after as usize, 0x90);
2244            text.funcs[0].len = text.bytes.len();
2245            let bytes = write(
2246                &text,
2247                &Data::default(),
2248                &[],
2249                &windows(),
2250                Output::default(),
2251                &Info::default(),
2252            )
2253            .expect("an object");
2254            let file = object::File::parse(&bytes[..]).expect("a readable object");
2255            let section = file.section_by_name(".text").expect("a text section");
2256            let (_, reloc) = section.relocations().next().expect("the relocation");
2257            assert_eq!(reloc.flags(), RelocationFlags::Coff { typ }, "{after}");
2258            // And the bytes come out holding nothing, because the distance the instruction wants
2259            // and the distance the type already says are the same one.
2260            assert_eq!(inline(&bytes, ".text", 1), 0, "{after}");
2261        }
2262    }
2263
2264    /// The addend a COFF object keeps is in the bytes rather than in the relocation, so the number
2265    /// the caller handed over has to survive the trip through the type.
2266    #[test]
2267    fn a_distance_the_instruction_did_not_ask_for_stays_in_the_bytes() {
2268        let mut text = calling("puts");
2269        text.relocs[0].addend = 12;
2270        let bytes =
2271            write(&text, &Data::default(), &[], &windows(), Output::default(), &Info::default())
2272                .expect("an object");
2273        assert_eq!(inline(&bytes, ".text", 1), 16, "twelve past the end, which is four past here");
2274    }
2275
2276    #[test]
2277    fn an_address_written_into_an_image_is_the_wide_relocation_here_too() {
2278        let object = Object {
2279            bytes: vec![0; 8],
2280            size: 8,
2281            align: 8,
2282            relocs: vec![Reloc {
2283                at: 0,
2284                symbol: "y".to_owned(),
2285                kind: Reference::Address { bytes: 8 },
2286                addend: 0,
2287                after: 0,
2288            }],
2289            ..variable("p", Place::Written)
2290        };
2291        let data = Data { apart: Vec::new(), weak: Vec::new(), objects: vec![object] };
2292        let bytes =
2293            write(&Text::default(), &data, &[], &windows(), Output::default(), &Info::default())
2294                .expect("an object");
2295        let file = object::File::parse(&bytes[..]).expect("a readable object");
2296        let section = file.section_by_name(".data").expect("a data section");
2297        let (_, reloc) = section.relocations().next().expect("the relocation");
2298        let typ = pe::IMAGE_REL_AMD64_ADDR64;
2299        assert_eq!(reloc.flags(), RelocationFlags::Coff { typ });
2300    }
2301
2302    /// `.data.rel.ro` is an ELF answer to a problem this format solves elsewhere, so both halves of
2303    /// it land in ordinary read only data, which is where the platform's own linker puts them.
2304    #[test]
2305    fn a_variable_the_loader_writes_into_is_read_only_data_here() {
2306        for local in [false, true] {
2307            let data = Data {
2308                apart: Vec::new(),
2309                weak: Vec::new(),
2310                objects: vec![variable("p", Place::RelocReadOnly { local })],
2311            };
2312            let bytes = write(
2313                &Text::default(),
2314                &data,
2315                &[],
2316                &windows(),
2317                Output::default(),
2318                &Info::default(),
2319            )
2320            .expect("an object");
2321            let file = object::File::parse(&bytes[..]).expect("a readable object");
2322            assert!(file.section_by_name(".rdata").is_some(), "{local}");
2323            assert!(file.section_by_name(".data.rel.ro.local").is_none(), "{local}");
2324        }
2325    }
2326
2327    /// No marker and no note, because a PE image says both of those things in the header of the
2328    /// finished image rather than in each of its inputs.
2329    #[test]
2330    fn the_sections_only_elf_reads_are_left_out_rather_than_written_empty() {
2331        let text = calling("puts");
2332        let output = Output { property: Property { features: 3 }, ..Output::default() };
2333        let bytes = write(&text, &Data::default(), &[], &windows(), output, &Info::default())
2334            .expect("an object");
2335        let file = object::File::parse(&bytes[..]).expect("a readable object");
2336        assert!(file.section_by_name(".note.GNU-stack").is_none());
2337        assert!(file.section_by_name(".note.gnu.property").is_none());
2338    }
2339
2340    /// Each of these is something this format has no way to write, and writing the nearest thing
2341    /// would be worse than refusing: a thread-local variable written as an ordinary one is one copy
2342    /// where the program asked for one per thread, and a constructor list under a name nothing
2343    /// gathers is a program whose constructors never run.
2344    #[test]
2345    fn what_this_format_cannot_say_is_refused_by_name() {
2346        let ordinary = Text::default();
2347        let empty = Data::default();
2348
2349        let mut thread = Data::default();
2350        thread.objects.push(variable("t", Place::Thread { zero: false }));
2351
2352        let mut gathered = Data::default();
2353        gathered.objects.push(variable("c", Place::Named(".init_array".to_owned())));
2354
2355        let mut table = calling("puts");
2356        table.relocs[0].kind = Reference::Got;
2357
2358        let mut room = calling("puts");
2359        room.funcs[0].patch = Some(Patch { at: 0, before: 0 });
2360
2361        let cases: [(&str, &Text, &Data); 4] = [
2362            ("thread-local", &ordinary, &thread),
2363            ("startup", &ordinary, &gathered),
2364            ("table", &table, &empty),
2365            ("patcher", &room, &empty),
2366        ];
2367        for (what, text, data) in cases {
2368            let error = write(text, data, &[], &windows(), Output::default(), &Info::default())
2369                .expect_err("something this format cannot write");
2370            assert!(matches!(error, Error::Refused { .. }), "{what}: {error:?}");
2371        }
2372    }
2373
2374    /// A visibility is not refused, because there is nothing to refuse: it is a fact about a dynamic
2375    /// symbol table and a COFF symbol has nowhere to keep one, which is what gcc does on the
2376    /// platform as well.
2377    #[test]
2378    fn a_visibility_this_format_cannot_keep_changes_nothing_rather_than_failing() {
2379        let mut text = calling("puts");
2380        text.funcs[0].visibility = Visibility::Hidden;
2381        let bytes =
2382            write(&text, &Data::default(), &[], &windows(), Output::default(), &Info::default())
2383                .expect("an object");
2384        let file = object::File::parse(&bytes[..]).expect("a readable object");
2385        let symbol = file.symbols().find(|symbol| symbol.name() == Ok("f")).expect("the function");
2386        assert!(symbol.is_global(), "a name others may use either way");
2387    }
2388
2389    #[test]
2390    fn the_names_a_linker_can_find_are_the_same_list_on_either_format() {
2391        let text = calling("puts");
2392        let data = Data {
2393            apart: Vec::new(),
2394            weak: Vec::new(),
2395            objects: vec![variable("shared", Place::Written)],
2396        };
2397        let theirs = defines(&text, &data, &[], &windows()).expect("a list");
2398        assert_eq!(theirs, defines(&text, &data, &[], &target()).expect("a list"));
2399    }
2400
2401    /// The same refusal the writer gives, for the reason the function says: an undecorated name is
2402    /// the wrong answer for a format whose symbols carry an underscore, and a wrong index entry is
2403    /// worse than no archive.
2404    #[test]
2405    fn a_platform_this_does_not_write_has_no_list_of_names_either() {
2406        let text = calling("puts");
2407        for triple in [
2408            Triple::new(Arch::Aarch64, Os::Linux, Env::Gnu),
2409            Triple::new(Arch::X86_64, Os::Darwin, Env::Gnu),
2410        ] {
2411            let error = defines(&text, &Data::default(), &[], &TargetInfo::new(triple))
2412                .expect_err("no writer");
2413            assert!(matches!(error, Error::Format { .. }), "{error:?}");
2414        }
2415    }
2416}