wasmtime 19.0.2 → 20.0.0

data/ext/cargo-vendor/object-0.33.0/src/build/elf.rs

@@ -0,0 +1,3033 @@
+//! This module provides a [`Builder`] for reading, modifying, and then writing ELF files.
+use alloc::vec::Vec;
+use core::convert::TryInto;
+use core::fmt;
+use core::marker::PhantomData;
+use hashbrown::HashMap;
+
+use crate::build::{ByteString, Bytes, Error, Id, IdPrivate, Item, Result, Table};
+use crate::elf;
+use crate::read::elf::{Dyn, FileHeader, ProgramHeader, Rela, SectionHeader, Sym};
+use crate::read::{self, FileKind, ReadRef};
+use crate::write;
+use crate::Endianness;
+
+/// A builder for reading, modifying, and then writing ELF files.
+///
+/// Public fields are available for modifying the values that will be written.
+/// Methods are available to add elements to tables, and elements can be deleted
+/// from tables by setting the `delete` field in the element.
+#[derive(Debug)]
+pub struct Builder<'data> {
+    /// The endianness.
+    ///
+    /// Used to set the data encoding when writing the ELF file.
+    pub endian: Endianness,
+    /// Whether file is 64-bit.
+    ///
+    /// Use to set the file class when writing the ELF file.
+    pub is_64: bool,
+    /// The alignment of [`elf::PT_LOAD`] segments.
+    pub load_align: u64,
+    /// The file header.
+    pub header: Header,
+    /// The segment table.
+    pub segments: Segments<'data>,
+    /// The section table.
+    pub sections: Sections<'data>,
+    /// The symbol table.
+    pub symbols: Symbols<'data>,
+    /// The dynamic symbol table.
+    pub dynamic_symbols: DynamicSymbols<'data>,
+    /// The base version for the GNU version definitions.
+    ///
+    /// This will be written as a version definition with index 1.
+    pub version_base: Option<ByteString<'data>>,
+    /// The GNU version definitions and dependencies.
+    pub versions: Versions<'data>,
+    /// The filenames used in the GNU version definitions.
+    pub version_files: VersionFiles<'data>,
+    /// The bucket count parameter for the hash table.
+    pub hash_bucket_count: u32,
+    /// The bloom shift parameter for the GNU hash table.
+    pub gnu_hash_bloom_shift: u32,
+    /// The bloom count parameter for the GNU hash table.
+    pub gnu_hash_bloom_count: u32,
+    /// The bucket count parameter for the GNU hash table.
+    pub gnu_hash_bucket_count: u32,
+    marker: PhantomData<()>,
+}
+
+impl<'data> Builder<'data> {
+    /// Create a new ELF builder.
+    pub fn new(endian: Endianness, is_64: bool) -> Self {
+        Self {
+            endian,
+            is_64,
+            load_align: 0,
+            header: Header::default(),
+            segments: Segments::new(),
+            sections: Sections::new(),
+            symbols: Symbols::new(),
+            dynamic_symbols: Symbols::new(),
+            version_base: None,
+            versions: Versions::new(),
+            version_files: VersionFiles::new(),
+            hash_bucket_count: 0,
+            gnu_hash_bloom_shift: 0,
+            gnu_hash_bloom_count: 0,
+            gnu_hash_bucket_count: 0,
+            marker: PhantomData,
+        }
+    }
+
+    /// Read the ELF file from file data.
+    pub fn read<R: ReadRef<'data>>(data: R) -> Result<Self> {
+        match FileKind::parse(data)? {
+            FileKind::Elf32 => Self::read32(data),
+            FileKind::Elf64 => Self::read64(data),
+            #[allow(unreachable_patterns)]
+            _ => Err(Error::new("Not an ELF file")),
+        }
+    }
+
+    /// Read a 32-bit ELF file from file data.
+    pub fn read32<R: ReadRef<'data>>(data: R) -> Result<Self> {
+        Self::read_file::<elf::FileHeader32<Endianness>, R>(data)
+    }
+
+    /// Read a 64-bit ELF file from file data.
+    pub fn read64<R: ReadRef<'data>>(data: R) -> Result<Self> {
+        Self::read_file::<elf::FileHeader64<Endianness>, R>(data)
+    }
+
+    fn read_file<Elf, R>(data: R) -> Result<Self>
+    where
+        Elf: FileHeader<Endian = Endianness>,
+        R: ReadRef<'data>,
+    {
+        let header = Elf::parse(data)?;
+        let endian = header.endian()?;
+        let is_mips64el = header.is_mips64el(endian);
+        let shstrndx = header.shstrndx(endian, data)? as usize;
+        let segments = header.program_headers(endian, data)?;
+        let sections = header.sections(endian, data)?;
+        let symbols = sections.symbols(endian, data, elf::SHT_SYMTAB)?;
+        let dynamic_symbols = sections.symbols(endian, data, elf::SHT_DYNSYM)?;
+
+        let mut builder = Builder {
+            endian,
+            is_64: header.is_type_64(),
+            load_align: 0,
+            header: Header {
+                os_abi: header.e_ident().os_abi,
+                abi_version: header.e_ident().abi_version,
+                e_type: header.e_type(endian),
+                e_machine: header.e_machine(endian),
+                e_entry: header.e_entry(endian).into(),
+                e_flags: header.e_flags(endian),
+                e_phoff: header.e_phoff(endian).into(),
+            },
+            segments: Segments::new(),
+            sections: Sections::new(),
+            symbols: Symbols::new(),
+            dynamic_symbols: Symbols::new(),
+            version_base: None,
+            versions: Versions::new(),
+            version_files: VersionFiles::new(),
+            hash_bucket_count: 0,
+            gnu_hash_bloom_shift: 0,
+            gnu_hash_bloom_count: 0,
+            gnu_hash_bucket_count: 0,
+            marker: PhantomData,
+        };
+
+        for segment in segments {
+            if segment.p_type(endian) == elf::PT_LOAD {
+                let p_align = segment.p_align(endian).into();
+                if builder.load_align != 0 && builder.load_align != p_align {
+                    return Err(Error::new("Unsupported alignments for PT_LOAD segments"));
+                }
+                builder.load_align = p_align;
+            }
+
+            let id = builder.segments.next_id();
+            builder.segments.push(Segment {
+                id,
+                delete: false,
+                p_type: segment.p_type(endian),
+                p_flags: segment.p_flags(endian),
+                p_offset: segment.p_offset(endian).into(),
+                p_vaddr: segment.p_vaddr(endian).into(),
+                p_paddr: segment.p_paddr(endian).into(),
+                p_filesz: segment.p_filesz(endian).into(),
+                p_memsz: segment.p_memsz(endian).into(),
+                p_align: segment.p_align(endian).into(),
+                sections: Vec::new(),
+                marker: PhantomData,
+            });
+        }
+        if !builder.segments.is_empty() && builder.load_align == 0 {
+            // There should be at least one PT_LOAD segment.
+            return Err(Error::new("Unsupported segments without a PT_LOAD segment"));
+        }
+
+        for (index, section) in sections.iter().enumerate().skip(1) {
+            let id = SectionId(index - 1);
+            let relocations = if let Some((rels, link)) = section.rel(endian, data)? {
+                Self::read_relocations(
+                    index,
+                    endian,
+                    is_mips64el,
+                    rels,
+                    link,
+                    &symbols,
+                    &dynamic_symbols,
+                )?
+            } else if let Some((rels, link)) = section.rela(endian, data)? {
+                Self::read_relocations(
+                    index,
+                    endian,
+                    is_mips64el,
+                    rels,
+                    link,
+                    &symbols,
+                    &dynamic_symbols,
+                )?
+            } else {
+                SectionData::Data(Bytes::default())
+            };
+            if let Some(hash) = section.hash_header(endian, data)? {
+                builder.hash_bucket_count = hash.bucket_count.get(endian);
+            }
+            if let Some(hash) = section.gnu_hash_header(endian, data)? {
+                builder.gnu_hash_bloom_shift = hash.bloom_shift.get(endian);
+                builder.gnu_hash_bloom_count = hash.bloom_count.get(endian);
+                builder.gnu_hash_bucket_count = hash.bucket_count.get(endian);
+            }
+            let data = match section.sh_type(endian) {
+                elf::SHT_NOBITS => SectionData::UninitializedData(section.sh_size(endian).into()),
+                elf::SHT_PROGBITS | elf::SHT_INIT_ARRAY | elf::SHT_FINI_ARRAY => {
+                    SectionData::Data(section.data(endian, data)?.into())
+                }
+                elf::SHT_REL | elf::SHT_RELA => relocations,
+                elf::SHT_SYMTAB => {
+                    if index == symbols.section().0 {
+                        SectionData::Symbol
+                    } else {
+                        return Err(Error(format!(
+                            "Unsupported SHT_SYMTAB section at index {}",
+                            index
+                        )));
+                    }
+                }
+                elf::SHT_SYMTAB_SHNDX => {
+                    if index == symbols.shndx_section().0 {
+                        SectionData::SymbolSectionIndex
+                    } else {
+                        return Err(Error(format!(
+                            "Unsupported SHT_SYMTAB_SHNDX section at index {}",
+                            index
+                        )));
+                    }
+                }
+                elf::SHT_DYNSYM => {
+                    if index == dynamic_symbols.section().0 {
+                        SectionData::DynamicSymbol
+                    } else {
+                        return Err(Error(format!(
+                            "Unsupported SHT_DYNSYM section at index {}",
+                            index
+                        )));
+                    }
+                }
+                elf::SHT_STRTAB => {
+                    if index == symbols.string_section().0 {
+                        SectionData::String
+                    } else if index == dynamic_symbols.string_section().0 {
+                        SectionData::DynamicString
+                    } else if index == shstrndx {
+                        SectionData::SectionString
+                    } else {
+                        return Err(Error(format!(
+                            "Unsupported SHT_STRTAB section at index {}",
+                            index
+                        )));
+                    }
+                }
+                elf::SHT_NOTE => SectionData::Note(section.data(endian, data)?.into()),
+                elf::SHT_DYNAMIC => {
+                    let (dyns, link) = section.dynamic(endian, data)?.unwrap();
+                    let dynamic_strings = sections.strings(endian, data, link)?;
+                    Self::read_dynamics::<Elf, _>(endian, dyns, dynamic_strings)?
+                }
+                elf::SHT_GNU_ATTRIBUTES => {
+                    let attributes = section.attributes(endian, data)?;
+                    Self::read_attributes(index, attributes, sections.len(), symbols.len())?
+                }
+                elf::SHT_HASH => SectionData::Hash,
+                elf::SHT_GNU_HASH => SectionData::GnuHash,
+                elf::SHT_GNU_VERSYM => SectionData::GnuVersym,
+                elf::SHT_GNU_VERDEF => SectionData::GnuVerdef,
+                elf::SHT_GNU_VERNEED => SectionData::GnuVerneed,
+                other => match (builder.header.e_machine, other) {
+                    (elf::EM_ARM, elf::SHT_ARM_ATTRIBUTES)
+                    | (elf::EM_AARCH64, elf::SHT_AARCH64_ATTRIBUTES) => {
+                        let attributes = section.attributes(endian, data)?;
+                        Self::read_attributes(index, attributes, sections.len(), symbols.len())?
+                    }
+                    // Some section types that we can't parse but that are safe to copy.
+                    // Lots of types missing, add as needed. We can't default to copying
+                    // everything because some types are not safe to copy.
+                    (elf::EM_ARM, elf::SHT_ARM_EXIDX)
+                    | (elf::EM_IA_64, elf::SHT_IA_64_UNWIND)
+                    | (elf::EM_MIPS, elf::SHT_MIPS_REGINFO)
+                    | (elf::EM_MIPS, elf::SHT_MIPS_DWARF) => {
+                        SectionData::Data(section.data(endian, data)?.into())
+                    }
+                    _ => return Err(Error(format!("Unsupported section type {:x}", other))),
+                },
+            };
+            let sh_flags = section.sh_flags(endian).into();
+            let sh_link = section.sh_link(endian);
+            let sh_link_section = if sh_link == 0 {
+                None
+            } else {
+                if sh_link as usize >= sections.len() {
+                    return Err(Error(format!(
+                        "Invalid sh_link {} in section at index {}",
+                        sh_link, index
+                    )));
+                }
+                Some(SectionId(sh_link as usize - 1))
+            };
+            let sh_info = section.sh_info(endian);
+            let sh_info_section = if sh_info == 0 || sh_flags & u64::from(elf::SHF_INFO_LINK) == 0 {
+                None
+            } else {
+                if sh_info as usize >= sections.len() {
+                    return Err(Error(format!(
+                        "Invalid sh_info link {} in section at index {}",
+                        sh_info, index
+                    )));
+                }
+                Some(SectionId(sh_info as usize - 1))
+            };
+            let sh_flags = section.sh_flags(endian).into();
+            let sh_addr = section.sh_addr(endian).into();
+            if sh_flags & u64::from(elf::SHF_ALLOC) != 0 {
+                for segment in &mut builder.segments {
+                    if segment.contains_address(sh_addr) {
+                        segment.sections.push(id);
+                    }
+                }
+            }
+            builder.sections.push(Section {
+                id,
+                delete: false,
+                name: sections.section_name(endian, section)?.into(),
+                sh_type: section.sh_type(endian),
+                sh_flags,
+                sh_addr,
+                sh_offset: section.sh_offset(endian).into(),
+                sh_size: section.sh_size(endian).into(),
+                sh_link_section,
+                sh_info,
+                sh_info_section,
+                sh_addralign: section.sh_addralign(endian).into(),
+                sh_entsize: section.sh_entsize(endian).into(),
+                data,
+            });
+        }
+
+        Self::read_symbols(
+            endian,
+            &symbols,
+            &mut builder.symbols,
+            builder.sections.len(),
+        )?;
+        Self::read_symbols(
+            endian,
+            &dynamic_symbols,
+            &mut builder.dynamic_symbols,
+            builder.sections.len(),
+        )?;
+        builder.read_gnu_versions(endian, data, &sections, &dynamic_symbols)?;
+
+        Ok(builder)
+    }
+
+    #[allow(clippy::too_many_arguments)]
+    fn read_relocations<Elf, Rel, R>(
+        index: usize,
+        endian: Elf::Endian,
+        is_mips64el: bool,
+        rels: &'data [Rel],
+        link: read::SectionIndex,
+        symbols: &read::elf::SymbolTable<'data, Elf, R>,
+        dynamic_symbols: &read::elf::SymbolTable<'data, Elf, R>,
+    ) -> Result<SectionData<'data>>
+    where
+        Elf: FileHeader<Endian = Endianness>,
+        Rel: Copy + Into<Elf::Rela>,
+        R: ReadRef<'data>,
+    {
+        if link.0 == 0 {
+            Self::read_relocations_impl::<Elf, Rel, true>(index, endian, is_mips64el, rels, 0)
+                .map(SectionData::DynamicRelocation)
+        } else if link == symbols.section() {
+            Self::read_relocations_impl::<Elf, Rel, false>(
+                index,
+                endian,
+                is_mips64el,
+                rels,
+                symbols.len(),
+            )
+            .map(SectionData::Relocation)
+        } else if link == dynamic_symbols.section() {
+            Self::read_relocations_impl::<Elf, Rel, true>(
+                index,
+                endian,
+                is_mips64el,
+                rels,
+                dynamic_symbols.len(),
+            )
+            .map(SectionData::DynamicRelocation)
+        } else {
+            return Err(Error(format!(
+                "Invalid sh_link {} in relocation section at index {}",
+                link.0, index,
+            )));
+        }
+    }
+
+    fn read_relocations_impl<Elf, Rel, const DYNAMIC: bool>(
+        index: usize,
+        endian: Elf::Endian,
+        is_mips64el: bool,
+        rels: &'data [Rel],
+        symbols_len: usize,
+    ) -> Result<Vec<Relocation<DYNAMIC>>>
+    where
+        Elf: FileHeader<Endian = Endianness>,
+        Rel: Copy + Into<Elf::Rela>,
+    {
+        let mut relocations = Vec::new();
+        for rel in rels {
+            let rel = (*rel).into();
+            let r_sym = rel.r_sym(endian, is_mips64el);
+            let symbol = if r_sym == 0 {
+                None
+            } else {
+                if r_sym as usize >= symbols_len {
+                    return Err(Error(format!(
+                        "Invalid symbol index {} in relocation section at index {}",
+                        r_sym, index,
+                    )));
+                }
+                Some(SymbolId(r_sym as usize - 1))
+            };
+            relocations.push(Relocation {
+                r_offset: rel.r_offset(endian).into(),
+                symbol,
+                r_type: rel.r_type(endian, is_mips64el),
+                r_addend: rel.r_addend(endian).into(),
+            });
+        }
+        Ok(relocations)
+    }
+
+    fn read_dynamics<Elf, R>(
+        endian: Elf::Endian,
+        dyns: &'data [Elf::Dyn],
+        strings: read::StringTable<'data, R>,
+    ) -> Result<SectionData<'data>>
+    where
+        Elf: FileHeader<Endian = Endianness>,
+        R: ReadRef<'data>,
+    {
+        let mut dynamics = Vec::with_capacity(dyns.len());
+        for d in dyns {
+            let tag = d.d_tag(endian).into().try_into().map_err(|_| {
+                Error(format!(
+                    "Unsupported dynamic tag 0x{:x}",
+                    d.d_tag(endian).into()
+                ))
+            })?;
+            if tag == elf::DT_NULL {
+                break;
+            }
+            let val = d.d_val(endian).into();
+            dynamics.push(if d.is_string(endian) {
+                let val =
+                    strings
+                        .get(val.try_into().map_err(|_| {
+                            Error(format!("Unsupported dynamic string 0x{:x}", val))
+                        })?)
+                        .map_err(|_| Error(format!("Invalid dynamic string 0x{:x}", val)))?;
+                Dynamic::String {
+                    tag,
+                    val: val.into(),
+                }
+            } else {
+                match tag {
+                    elf::DT_SYMTAB
+                    | elf::DT_STRTAB
+                    | elf::DT_STRSZ
+                    | elf::DT_HASH
+                    | elf::DT_GNU_HASH
+                    | elf::DT_VERSYM
+                    | elf::DT_VERDEF
+                    | elf::DT_VERDEFNUM
+                    | elf::DT_VERNEED
+                    | elf::DT_VERNEEDNUM => Dynamic::Auto { tag },
+                    _ => Dynamic::Integer { tag, val },
+                }
+            });
+        }
+        Ok(SectionData::Dynamic(dynamics))
+    }
+
+    fn read_symbols<Elf, R, const DYNAMIC: bool>(
+        endian: Elf::Endian,
+        symbols: &read::elf::SymbolTable<'data, Elf, R>,
+        builder_symbols: &mut Symbols<'data, DYNAMIC>,
+        sections_len: usize,
+    ) -> Result<()>
+    where
+        Elf: FileHeader<Endian = Endianness>,
+        R: ReadRef<'data>,
+    {
+        for (index, symbol) in symbols.iter().enumerate().skip(1) {
+            let id = SymbolId(index - 1);
+            let section =
+                if let Some(section_index) = symbols.symbol_section(endian, symbol, index)? {
+                    let section_id = section_index.0.wrapping_sub(1);
+                    if section_id >= sections_len {
+                        return Err(Error::new("Invalid symbol section index"));
+                    }
+                    Some(SectionId(section_id))
+                } else {
+                    None
+                };
+            builder_symbols.push(Symbol {
+                id,
+                delete: false,
+                name: symbols.symbol_name(endian, symbol)?.into(),
+                section,
+                st_info: symbol.st_info(),
+                st_other: symbol.st_other(),
+                st_shndx: symbol.st_shndx(endian),
+                st_value: symbol.st_value(endian).into(),
+                st_size: symbol.st_size(endian).into(),
+                version: VersionId::local(),
+                version_hidden: false,
+            });
+        }
+        Ok(())
+    }
+
+    fn read_attributes<Elf>(
+        index: usize,
+        attributes: read::elf::AttributesSection<'data, Elf>,
+        sections_len: usize,
+        symbols_len: usize,
+    ) -> Result<SectionData<'data>>
+    where
+        Elf: FileHeader<Endian = Endianness>,
+    {
+        let mut builder_attributes = AttributesSection::new();
+        let mut subsections = attributes.subsections()?;
+        while let Some(subsection) = subsections.next()? {
+            let mut builder_subsection = AttributesSubsection::new(subsection.vendor().into());
+            let mut subsubsections = subsection.subsubsections();
+            while let Some(subsubsection) = subsubsections.next()? {
+                let tag = match subsubsection.tag() {
+                    elf::Tag_File => AttributeTag::File,
+                    elf::Tag_Section => {
+                        let mut tag_sections = Vec::new();
+                        let mut indices = subsubsection.indices();
+                        while let Some(index) = indices.next()? {
+                            let index = index as usize;
+                            if index >= sections_len {
+                                return Err(Error(format!(
+                                    "Invalid section index {} in attribute",
+                                    index
+                                )));
+                            }
+                            tag_sections.push(SectionId(index - 1));
+                        }
+                        AttributeTag::Section(tag_sections)
+                    }
+                    elf::Tag_Symbol => {
+                        let mut tag_symbols = Vec::new();
+                        let mut indices = subsubsection.indices();
+                        while let Some(index) = indices.next()? {
+                            let index = index as usize;
+                            if index >= symbols_len {
+                                return Err(Error(format!(
+                                    "Invalid symbol index {} in attribute",
+                                    index
+                                )));
+                            }
+                            tag_symbols.push(SymbolId(index - 1));
+                        }
+                        AttributeTag::Symbol(tag_symbols)
+                    }
+                    tag => {
+                        return Err(Error(format!(
+                            "Unsupported attribute tag 0x{:x} in section at index {}",
+                            tag, index,
+                        )))
+                    }
+                };
+                let data = subsubsection.attributes_data().into();
+                builder_subsection
+                    .subsubsections
+                    .push(AttributesSubsubsection { tag, data });
+            }
+            builder_attributes.subsections.push(builder_subsection);
+        }
+        Ok(SectionData::Attributes(builder_attributes))
+    }
+
+    fn read_gnu_versions<Elf, R>(
+        &mut self,
+        endian: Elf::Endian,
+        data: R,
+        sections: &read::elf::SectionTable<'data, Elf, R>,
+        dynamic_symbols: &read::elf::SymbolTable<'data, Elf, R>,
+    ) -> Result<()>
+    where
+        Elf: FileHeader<Endian = Endianness>,
+        R: ReadRef<'data>,
+    {
+        let strings = dynamic_symbols.strings();
+        let mut ids = HashMap::new();
+        ids.insert(0, VersionId::local());
+        ids.insert(1, VersionId::global());
+
+        if let Some((mut verdefs, link)) = sections.gnu_verdef(endian, data)? {
+            if link != dynamic_symbols.string_section() {
+                return Err(Error::new("Invalid SHT_GNU_VERDEF section"));
+            }
+            while let Some((verdef, mut verdauxs)) = verdefs.next()? {
+                let flags = verdef.vd_flags.get(endian);
+                if flags & elf::VER_FLG_BASE != 0 {
+                    if flags != elf::VER_FLG_BASE
+                        || verdef.vd_ndx.get(endian) != 1
+                        || verdef.vd_cnt.get(endian) != 1
+                    {
+                        return Err(Error::new("Unsupported VER_FLG_BASE in SHT_GNU_VERDEF"));
+                    }
+                    if self.version_base.is_some() {
+                        return Err(Error::new("Duplicate VER_FLG_BASE in SHT_GNU_VERDEF"));
+                    }
+                    let verdaux = verdauxs.next()?.ok_or_else(|| {
+                        Error::new("Missing name for VER_FLG_BASE in SHT_GNU_VERDEF")
+                    })?;
+                    self.version_base = Some(verdaux.name(endian, strings)?.into());
+                    continue;
+                }
+
+                let index = verdef.vd_ndx.get(endian) & elf::VERSYM_VERSION;
+                let id = self.versions.next_id();
+                if ids.insert(index, id).is_some() {
+                    return Err(Error(format!("Duplicate SHT_GNU_VERDEF index {}", index)));
+                }
+
+                let mut names = Vec::new();
+                while let Some(verdaux) = verdauxs.next()? {
+                    names.push(verdaux.name(endian, strings)?.into());
+                }
+
+                let data = VersionData::Def(VersionDef { flags, names });
+                self.versions.push(Version {
+                    id,
+                    delete: false,
+                    data,
+                });
+            }
+        }
+
+        if let Some((mut verneeds, link)) = sections.gnu_verneed(endian, data)? {
+            if link != dynamic_symbols.string_section() {
+                return Err(Error::new("Invalid SHT_GNU_VERNEED section"));
+            }
+            while let Some((verneed, mut vernauxs)) = verneeds.next()? {
+                let file = VersionFileId(self.version_files.len());
+                self.version_files.push(VersionFile {
+                    id: file,
+                    delete: false,
+                    name: verneed.file(endian, strings)?.into(),
+                });
+                while let Some(vernaux) = vernauxs.next()? {
+                    let index = vernaux.vna_other.get(endian) & elf::VERSYM_VERSION;
+                    let id = self.versions.next_id();
+                    if ids.insert(index, id).is_some() {
+                        return Err(Error(format!("Duplicate SHT_GNU_VERNEED index {}", index)));
+                    }
+
+                    let data = VersionData::Need(VersionNeed {
+                        flags: vernaux.vna_flags.get(endian),
+                        name: vernaux.name(endian, strings)?.into(),
+                        file,
+                    });
+                    self.versions.push(Version {
+                        id,
+                        delete: false,
+                        data,
+                    });
+                }
+            }
+        }
+
+        if let Some((versyms, link)) = sections.gnu_versym(endian, data)? {
+            if versyms.len() != dynamic_symbols.len() || link != dynamic_symbols.section() {
+                return Err(Error::new("Invalid SHT_GNU_VERSYM section"));
+            }
+            for (id, versym) in versyms.iter().skip(1).enumerate() {
+                let index = versym.0.get(endian);
+                let symbol = self.dynamic_symbols.get_mut(SymbolId(id));
+                symbol.version = *ids
+                    .get(&(index & elf::VERSYM_VERSION))
+                    .ok_or_else(|| Error(format!("Invalid SHT_GNU_VERSYM index {:x}", index)))?;
+                symbol.version_hidden = index & elf::VERSYM_HIDDEN != 0;
+            }
+        }
+        Ok(())
+    }
+
+    /// Write the ELF file to the buffer.
+    pub fn write(mut self, buffer: &mut dyn write::WritableBuffer) -> Result<()> {
+        struct SectionOut {
+            id: SectionId,
+            name: Option<write::StringId>,
+            offset: usize,
+            attributes: Vec<u8>,
+        }
+
+        struct SymbolOut {
+            id: SymbolId,
+            name: Option<write::StringId>,
+        }
+
+        struct DynamicSymbolOut {
+            id: DynamicSymbolId,
+            name: Option<write::StringId>,
+            hash: Option<u32>,
+            gnu_hash: Option<u32>,
+        }
+
+        #[derive(Default, Clone)]
+        struct VersionFileOut {
+            versions: Vec<VersionId>,
+        }
+
+        // TODO: require the caller to do this?
+        self.delete_orphans();
+        self.delete_unused_versions();
+
+        let mut writer = write::elf::Writer::new(self.endian, self.is_64, buffer);
+
+        // Find metadata sections, and assign section indices.
+        let mut shstrtab_id = None;
+        let mut symtab_id = None;
+        let mut symtab_shndx_id = None;
+        let mut strtab_id = None;
+        let mut dynsym_id = None;
+        let mut dynstr_id = None;
+        let mut hash_id = None;
+        let mut gnu_hash_id = None;
+        let mut gnu_versym_id = None;
+        let mut gnu_verdef_id = None;
+        let mut gnu_verneed_id = None;
+        let mut out_sections = Vec::with_capacity(self.sections.len());
+        let mut out_sections_index = vec![None; self.sections.len()];
+        if !self.sections.is_empty() {
+            writer.reserve_null_section_index();
+        }
+        for section in &self.sections {
+            let index = match &section.data {
+                SectionData::Data(_)
+                | SectionData::UninitializedData(_)
+                | SectionData::Relocation(_)
+                | SectionData::DynamicRelocation(_)
+                | SectionData::Note(_)
+                | SectionData::Dynamic(_)
+                | SectionData::Attributes(_) => writer.reserve_section_index(),
+                SectionData::SectionString => {
+                    if shstrtab_id.is_some() {
+                        return Err(Error::new("Multiple .shstrtab sections"));
+                    }
+                    shstrtab_id = Some(section.id);
+                    writer.reserve_shstrtab_section_index_with_name(&section.name)
+                }
+                SectionData::Symbol => {
+                    if symtab_id.is_some() {
+                        return Err(Error::new("Multiple .symtab sections"));
+                    }
+                    symtab_id = Some(section.id);
+                    writer.reserve_symtab_section_index_with_name(&section.name)
+                }
+                SectionData::SymbolSectionIndex => {
+                    if symtab_shndx_id.is_some() {
+                        return Err(Error::new("Multiple .symtab_shndx sections"));
+                    }
+                    symtab_shndx_id = Some(section.id);
+                    writer.reserve_symtab_shndx_section_index_with_name(&section.name)
+                }
+                SectionData::String => {
+                    if strtab_id.is_some() {
+                        return Err(Error::new("Multiple .strtab sections"));
+                    }
+                    strtab_id = Some(section.id);
+                    writer.reserve_strtab_section_index_with_name(&section.name)
+                }
+                SectionData::DynamicSymbol => {
+                    if dynsym_id.is_some() {
+                        return Err(Error::new("Multiple .dynsym sections"));
+                    }
+                    dynsym_id = Some(section.id);
+                    writer.reserve_dynsym_section_index_with_name(&section.name)
+                }
+                SectionData::DynamicString => {
+                    if dynstr_id.is_some() {
+                        return Err(Error::new("Multiple .dynstr sections"));
+                    }
+                    dynstr_id = Some(section.id);
+                    writer.reserve_dynstr_section_index_with_name(&section.name)
+                }
+                SectionData::Hash => {
+                    if hash_id.is_some() {
+                        return Err(Error::new("Multiple .hash sections"));
+                    }
+                    hash_id = Some(section.id);
+                    writer.reserve_hash_section_index_with_name(&section.name)
+                }
+                SectionData::GnuHash => {
+                    if gnu_hash_id.is_some() {
+                        return Err(Error::new("Multiple .gnu.hash sections"));
+                    }
+                    gnu_hash_id = Some(section.id);
+                    writer.reserve_gnu_hash_section_index_with_name(&section.name)
+                }
+                SectionData::GnuVersym => {
+                    if gnu_versym_id.is_some() {
+                        return Err(Error::new("Multiple .gnu.version sections"));
+                    }
+                    gnu_versym_id = Some(section.id);
+                    writer.reserve_gnu_versym_section_index_with_name(&section.name)
+                }
+                SectionData::GnuVerdef => {
+                    if gnu_verdef_id.is_some() {
+                        return Err(Error::new("Multiple .gnu.version_d sections"));
+                    }
+                    gnu_verdef_id = Some(section.id);
+                    writer.reserve_gnu_verdef_section_index_with_name(&section.name)
+                }
+                SectionData::GnuVerneed => {
+                    if gnu_verneed_id.is_some() {
+                        return Err(Error::new("Multiple .gnu.version_r sections"));
+                    }
+                    gnu_verneed_id = Some(section.id);
+                    writer.reserve_gnu_verneed_section_index_with_name(&section.name)
+                }
+            };
+            out_sections_index[section.id.0] = Some(index);
+
+            let name = if section.name.is_empty() {
+                None
+            } else {
+                Some(writer.add_section_name(&section.name))
+            };
+            out_sections.push(SectionOut {
+                id: section.id,
+                name,
+                offset: 0,
+                attributes: Vec::new(),
+            });
+        }
+
+        // Assign dynamic strings.
+        for section in &self.sections {
+            if let SectionData::Dynamic(dynamics) = &section.data {
+                for dynamic in dynamics {
+                    if let Dynamic::String { val, .. } = dynamic {
+                        writer.add_dynamic_string(val);
+                    }
+                }
+            }
+        }
+
+        // Assign dynamic symbol indices.
+        let mut out_dynsyms = Vec::with_capacity(self.dynamic_symbols.len());
+        let mut gnu_hash_symbol_count = 0;
+        for symbol in &self.dynamic_symbols {
+            let mut name = None;
+            let mut hash = None;
+            let mut gnu_hash = None;
+            if !symbol.name.is_empty() {
+                name = Some(writer.add_dynamic_string(&symbol.name));
+                if hash_id.is_some() {
+                    hash = Some(elf::hash(&symbol.name));
+                }
+                if gnu_hash_id.is_some() && symbol.st_shndx != elf::SHN_UNDEF {
+                    gnu_hash = Some(elf::gnu_hash(&symbol.name));
+                    gnu_hash_symbol_count += 1;
+                }
+            }
+            out_dynsyms.push(DynamicSymbolOut {
+                id: symbol.id,
+                name,
+                hash,
+                gnu_hash,
+            });
+        }
+        // We must sort for GNU hash before allocating symbol indices.
+        if gnu_hash_id.is_some() {
+            if self.gnu_hash_bucket_count == 0 {
+                return Err(Error::new(".gnu.hash bucket count is zero"));
+            }
+            // TODO: recalculate bucket_count?
+            out_dynsyms.sort_by_key(|sym| match sym.gnu_hash {
+                None => (0, 0),
+                Some(hash) => (1, hash % self.gnu_hash_bucket_count),
+            });
+        }
+        let mut out_dynsyms_index = vec![None; self.dynamic_symbols.len()];
+        if dynsym_id.is_some() {
+            writer.reserve_null_dynamic_symbol_index();
+        }
+        for out_dynsym in &mut out_dynsyms {
+            out_dynsyms_index[out_dynsym.id.0] = Some(writer.reserve_dynamic_symbol_index());
+        }
+
+        // Hash parameters.
+        let hash_index_base = 1; // Null symbol.
+        let hash_chain_count = hash_index_base + out_dynsyms.len() as u32;
+
+        // GNU hash parameters.
+        let gnu_hash_index_base = if gnu_hash_symbol_count == 0 {
+            0
+        } else {
+            out_dynsyms.len() as u32 - gnu_hash_symbol_count
+        };
+        let gnu_hash_symbol_base = gnu_hash_index_base + 1; // Null symbol.
+
+        // Assign symbol indices.
+        let mut out_syms = Vec::with_capacity(self.symbols.len());
+        // Local symbols must come before global.
+        let local_symbols = self
+            .symbols
+            .into_iter()
+            .filter(|symbol| symbol.st_bind() == elf::STB_LOCAL);
+        let global_symbols = self
+            .symbols
+            .into_iter()
+            .filter(|symbol| symbol.st_bind() != elf::STB_LOCAL);
+        for symbol in local_symbols.chain(global_symbols) {
+            let name = if symbol.name.is_empty() {
+                None
+            } else {
+                Some(writer.add_string(&symbol.name))
+            };
+
+            out_syms.push(SymbolOut {
+                id: symbol.id,
+                name,
+            });
+        }
+        let num_local = 1 + out_syms
+            .iter()
+            .take_while(|sym| self.symbols.get(sym.id).st_bind() == elf::STB_LOCAL)
+            .count() as u32;
+        let mut out_syms_index = vec![None; self.symbols.len()];
+        if symtab_id.is_some() {
+            writer.reserve_null_symbol_index();
+        }
+        for out_sym in out_syms.iter_mut() {
+            out_syms_index[out_sym.id.0] = Some(writer.reserve_symbol_index(None));
+        }
+
+        // Count the versions and add version strings.
+        let mut verdef_count = 0;
+        let mut verdaux_count = 0;
+        let mut verneed_count = 0;
+        let mut vernaux_count = 0;
+        let mut out_version_files = vec![VersionFileOut::default(); self.version_files.len()];
+        if let Some(version_base) = &self.version_base {
+            verdef_count += 1;
+            verdaux_count += 1;
+            writer.add_dynamic_string(version_base);
+        }
+        for version in &self.versions {
+            match &version.data {
+                VersionData::Def(def) => {
+                    verdef_count += 1;
+                    verdaux_count += def.names.len();
+                    for name in &def.names {
+                        writer.add_dynamic_string(name);
+                    }
+                }
+                VersionData::Need(need) => {
+                    vernaux_count += 1;
+                    writer.add_dynamic_string(&need.name);
+                    out_version_files[need.file.0].versions.push(version.id);
+                }
+            }
+        }
+        for file in &self.version_files {
+            verneed_count += 1;
+            writer.add_dynamic_string(&file.name);
+        }
+
+        // Build the attributes sections.
+        for out_section in &mut out_sections {
+            let SectionData::Attributes(attributes) = &self.sections.get(out_section.id).data
+            else {
+                continue;
+            };
+            if attributes.subsections.is_empty() {
+                continue;
+            }
+            let mut writer = writer.attributes_writer();
+            for subsection in &attributes.subsections {
+                writer.start_subsection(&subsection.vendor);
+                for subsubsection in &subsection.subsubsections {
+                    writer.start_subsubsection(subsubsection.tag.tag());
+                    match &subsubsection.tag {
+                        AttributeTag::File => {}
+                        AttributeTag::Section(sections) => {
+                            for id in sections {
+                                if let Some(index) = out_sections_index[id.0] {
+                                    writer.write_subsubsection_index(index.0);
+                                }
+                            }
+                            writer.write_subsubsection_index(0);
+                        }
+                        AttributeTag::Symbol(symbols) => {
+                            for id in symbols {
+                                if let Some(index) = out_syms_index[id.0] {
+                                    writer.write_subsubsection_index(index.0);
+                                }
+                            }
+                            writer.write_subsubsection_index(0);
+                        }
+                    }
+                    writer.write_subsubsection_attributes(&subsubsection.data);
+                    writer.end_subsubsection();
+                }
+                writer.end_subsection();
+            }
+            out_section.attributes = writer.data();
+        }
+
+        // TODO: support section headers in strtab
+        if shstrtab_id.is_none() && !out_sections.is_empty() {
+            return Err(Error::new(".shstrtab section is needed but not present"));
+        }
+        if symtab_id.is_none() && !out_syms.is_empty() {
+            return Err(Error::new(".symtab section is needed but not present"));
+        }
+        if symtab_shndx_id.is_none() && writer.symtab_shndx_needed() {
+            return Err(Error::new(
+                ".symtab.shndx section is needed but not present",
+            ));
+        }
+        if strtab_id.is_none() && writer.strtab_needed() {
+            return Err(Error::new(".strtab section is needed but not present"));
+        }
+        if dynsym_id.is_none() && !out_dynsyms.is_empty() {
+            return Err(Error::new(".dynsym section is needed but not present"));
+        }
+        if dynstr_id.is_none() && writer.dynstr_needed() {
+            return Err(Error::new(".dynstr section is needed but not present"));
+        }
+        if gnu_verdef_id.is_none() && verdef_count > 0 {
+            return Err(Error::new(
+                ".gnu.version_d section is needed but not present",
+            ));
+        }
+        if gnu_verneed_id.is_none() && verneed_count > 0 {
+            return Err(Error::new(
+                ".gnu.version_r section is needed but not present",
+            ));
+        }
+
+        // Start reserving file ranges.
+        writer.reserve_file_header();
+
+        let mut dynsym_addr = None;
+        let mut dynstr_addr = None;
+        let mut hash_addr = None;
+        let mut gnu_hash_addr = None;
+        let mut versym_addr = None;
+        let mut verdef_addr = None;
+        let mut verneed_addr = None;
+
+        if !self.segments.is_empty() {
+            // TODO: support program headers in other locations.
+            if self.header.e_phoff != writer.reserved_len() as u64 {
+                return Err(Error(format!(
+                    "Unsupported e_phoff value 0x{:x}",
+                    self.header.e_phoff
+                )));
+            }
+            writer.reserve_program_headers(self.segments.count() as u32);
+        }
+
+        let mut alloc_sections = Vec::new();
+        if !self.segments.is_empty() {
+            // Reserve alloc sections at original offsets.
+            alloc_sections = out_sections
+                .iter()
+                .enumerate()
+                .filter_map(|(index, out_section)| {
+                    let section = self.sections.get(out_section.id);
+                    if section.is_alloc() {
+                        Some(index)
+                    } else {
+                        None
+                    }
+                })
+                .collect();
+            // The data for alloc sections may need to be written in a different order
+            // from their section headers.
+            alloc_sections.sort_by_key(|index| {
+                let section = &self.sections.get(out_sections[*index].id);
+                // SHT_NOBITS sections need to come before other sections at the same offset.
+                let file_size = if section.sh_type == elf::SHT_NOBITS {
+                    0
+                } else {
+                    section.sh_size
+                };
+                (section.sh_offset, file_size)
+            });
+            for index in &alloc_sections {
+                let out_section = &mut out_sections[*index];
+                let section = &self.sections.get(out_section.id);
+
+                if section.sh_offset < writer.reserved_len() as u64 {
+                    return Err(Error(format!(
+                        "Unsupported sh_offset value 0x{:x} for section '{}', expected at least 0x{:x}",
+                        section.sh_offset,
+                        section.name,
+                        writer.reserved_len(),
+                    )));
+                }
+                // The input sh_offset needs to be preserved so that offsets in program
+                // headers are correct.
+                writer.reserve_until(section.sh_offset as usize);
+                out_section.offset = match &section.data {
+                    SectionData::Data(data) => {
+                        writer.reserve(data.len(), section.sh_addralign as usize)
+                    }
+                    SectionData::UninitializedData(_) => {
+                        // Note: unaligned input sh_offset was observed in practice.
+                        writer.reserve(0, 1)
+                    }
+                    SectionData::DynamicRelocation(relocations) => writer
+                        .reserve_relocations(relocations.len(), section.sh_type == elf::SHT_RELA),
+                    SectionData::Note(data) => {
+                        writer.reserve(data.len(), section.sh_addralign as usize)
+                    }
+                    SectionData::Dynamic(dynamics) => writer.reserve_dynamics(1 + dynamics.len()),
+                    SectionData::DynamicSymbol => {
+                        dynsym_addr = Some(section.sh_addr);
+                        writer.reserve_dynsym()
+                    }
+                    SectionData::DynamicString => {
+                        dynstr_addr = Some(section.sh_addr);
+                        writer.reserve_dynstr()
+                    }
+                    SectionData::Hash => {
+                        hash_addr = Some(section.sh_addr);
+                        writer.reserve_hash(self.hash_bucket_count, hash_chain_count)
+                    }
+                    SectionData::GnuHash => {
+                        gnu_hash_addr = Some(section.sh_addr);
+                        writer.reserve_gnu_hash(
+                            self.gnu_hash_bloom_count,
+                            self.gnu_hash_bucket_count,
+                            gnu_hash_symbol_count,
+                        )
+                    }
+                    SectionData::GnuVersym => {
+                        versym_addr = Some(section.sh_addr);
+                        writer.reserve_gnu_versym()
+                    }
+                    SectionData::GnuVerdef => {
+                        verdef_addr = Some(section.sh_addr);
+                        writer.reserve_gnu_verdef(verdef_count, verdaux_count)
+                    }
+                    SectionData::GnuVerneed => {
+                        verneed_addr = Some(section.sh_addr);
+                        writer.reserve_gnu_verneed(verneed_count, vernaux_count)
+                    }
+                    _ => {
+                        return Err(Error(format!(
+                            "Unsupported alloc section type {:x} for section '{}'",
+                            section.sh_type, section.name,
+                        )));
+                    }
+                };
+                if out_section.offset as u64 != section.sh_offset {
+                    return Err(Error(format!(
+                        "Unaligned sh_offset value 0x{:x} for section '{}', expected 0x{:x}",
+                        section.sh_offset, section.name, out_section.offset,
+                    )));
+                }
+            }
+        }
+
+        // Reserve non-alloc sections at any offset.
+        for out_section in &mut out_sections {
+            let section = self.sections.get(out_section.id);
+            if !self.segments.is_empty() && section.is_alloc() {
+                continue;
+            }
+            out_section.offset = match &section.data {
+                SectionData::Data(data) => {
+                    writer.reserve(data.len(), section.sh_addralign as usize)
+                }
+                SectionData::UninitializedData(_) => {
+                    writer.reserve(0, section.sh_addralign as usize)
+                }
+                SectionData::Note(data) => {
+                    writer.reserve(data.len(), section.sh_addralign as usize)
+                }
+                SectionData::Attributes(_) => {
+                    writer.reserve(out_section.attributes.len(), section.sh_addralign as usize)
+                }
+                // These are handled elsewhere.
+                SectionData::Relocation(_)
+                | SectionData::SectionString
+                | SectionData::Symbol
+                | SectionData::SymbolSectionIndex
+                | SectionData::String => {
+                    continue;
+                }
+                _ => {
+                    return Err(Error(format!(
+                        "Unsupported non-alloc section type {:x}",
+                        section.sh_type
+                    )));
+                }
+            };
+        }
+
+        writer.reserve_symtab();
+        writer.reserve_symtab_shndx();
+        writer.reserve_strtab();
+
+        // Reserve non-alloc relocations.
+        for out_section in &mut out_sections {
+            let section = self.sections.get(out_section.id);
+            if !self.segments.is_empty() && section.is_alloc() {
+                continue;
+            }
+            let SectionData::Relocation(relocations) = &section.data else {
+                continue;
+            };
+            out_section.offset =
+                writer.reserve_relocations(relocations.len(), section.sh_type == elf::SHT_RELA);
+        }
+
+        writer.reserve_shstrtab();
+        writer.reserve_section_headers();
+
+        // Start writing.
+        writer.write_file_header(&write::elf::FileHeader {
+            os_abi: self.header.os_abi,
+            abi_version: self.header.abi_version,
+            e_type: self.header.e_type,
+            e_machine: self.header.e_machine,
+            e_entry: self.header.e_entry,
+            e_flags: self.header.e_flags,
+        })?;
+
+        if !self.segments.is_empty() {
+            writer.write_align_program_headers();
+            for segment in &self.segments {
+                writer.write_program_header(&write::elf::ProgramHeader {
+                    p_type: segment.p_type,
+                    p_flags: segment.p_flags,
+                    p_offset: segment.p_offset,
+                    p_vaddr: segment.p_vaddr,
+                    p_paddr: segment.p_paddr,
+                    p_filesz: segment.p_filesz,
+                    p_memsz: segment.p_memsz,
+                    p_align: segment.p_align,
+                });
+            }
+        }
+
+        // Write alloc sections.
+        if !self.segments.is_empty() {
+            for index in &alloc_sections {
+                let out_section = &mut out_sections[*index];
+                let section = self.sections.get(out_section.id);
+                writer.pad_until(out_section.offset);
+                match &section.data {
+                    SectionData::Data(data) => {
+                        writer.write(data);
+                    }
+                    SectionData::UninitializedData(_) => {}
+                    SectionData::DynamicRelocation(relocations) => {
+                        for rel in relocations {
+                            let r_sym = if let Some(symbol) = rel.symbol {
+                                out_dynsyms_index[symbol.0].unwrap().0
+                            } else {
+                                0
+                            };
+                            writer.write_relocation(
+                                section.sh_type == elf::SHT_RELA,
+                                &write::elf::Rel {
+                                    r_offset: rel.r_offset,
+                                    r_sym,
+                                    r_type: rel.r_type,
+                                    r_addend: rel.r_addend,
+                                },
+                            );
+                        }
+                    }
+                    SectionData::Note(data) => {
+                        writer.write(data);
+                    }
+                    SectionData::Dynamic(dynamics) => {
+                        for d in dynamics {
+                            match *d {
+                                Dynamic::Auto { tag } => {
+                                    // TODO: support more values
+                                    let val = match tag {
+                                        elf::DT_SYMTAB => dynsym_addr.ok_or(Error::new(
+                                            "Missing .dynsym section for DT_SYMTAB",
+                                        ))?,
+                                        elf::DT_STRTAB => dynstr_addr.ok_or(Error::new(
+                                            "Missing .dynstr section for DT_STRTAB",
+                                        ))?,
+                                        elf::DT_STRSZ => writer.dynstr_len() as u64,
+                                        elf::DT_HASH => hash_addr.ok_or(Error::new(
+                                            "Missing .hash section for DT_HASH",
+                                        ))?,
+                                        elf::DT_GNU_HASH => gnu_hash_addr.ok_or(Error::new(
+                                            "Missing .gnu.hash section for DT_GNU_HASH",
+                                        ))?,
+                                        elf::DT_VERSYM => versym_addr.ok_or(Error::new(
+                                            "Missing .gnu.version section for DT_VERSYM",
+                                        ))?,
+                                        elf::DT_VERDEF => verdef_addr.ok_or(Error::new(
+                                            "Missing .gnu.version_d section for DT_VERDEF",
+                                        ))?,
+                                        elf::DT_VERDEFNUM => verdef_count as u64,
+                                        elf::DT_VERNEED => verneed_addr.ok_or(Error::new(
+                                            "Missing .gnu.version_r section for DT_VERNEED",
+                                        ))?,
+                                        elf::DT_VERNEEDNUM => verneed_count as u64,
+                                        _ => {
+                                            return Err(Error(format!(
+                                                "Cannot generate value for dynamic tag 0x{:x}",
+                                                tag
+                                            )))
+                                        }
+                                    };
+                                    writer.write_dynamic(tag, val);
+                                }
+                                Dynamic::Integer { tag, val } => {
+                                    writer.write_dynamic(tag, val);
+                                }
+                                Dynamic::String { tag, ref val } => {
+                                    let val = writer.get_dynamic_string(val);
+                                    writer.write_dynamic_string(tag, val);
+                                }
+                            }
+                        }
+                        writer.write_dynamic(elf::DT_NULL, 0);
+                    }
+                    SectionData::DynamicSymbol => {
+                        writer.write_null_dynamic_symbol();
+                        for out_dynsym in &out_dynsyms {
+                            let symbol = self.dynamic_symbols.get(out_dynsym.id);
+                            let section =
+                                symbol.section.map(|id| out_sections_index[id.0].unwrap());
+                            writer.write_dynamic_symbol(&write::elf::Sym {
+                                name: out_dynsym.name,
+                                section,
+                                st_info: symbol.st_info,
+                                st_other: symbol.st_other,
+                                st_shndx: symbol.st_shndx,
+                                st_value: symbol.st_value,
+                                st_size: symbol.st_size,
+                            });
+                        }
+                    }
+                    SectionData::DynamicString => {
+                        writer.write_dynstr();
+                    }
+                    SectionData::Hash => {
+                        if self.hash_bucket_count == 0 {
+                            return Err(Error::new(".hash bucket count is zero"));
+                        }
+                        writer.write_hash(self.hash_bucket_count, hash_chain_count, |index| {
+                            out_dynsyms
+                                .get(index.checked_sub(hash_index_base)? as usize)?
+                                .hash
+                        });
+                    }
+                    SectionData::GnuHash => {
+                        if self.gnu_hash_bucket_count == 0 {
+                            return Err(Error::new(".gnu.hash bucket count is zero"));
+                        }
+                        writer.write_gnu_hash(
+                            gnu_hash_symbol_base,
+                            self.gnu_hash_bloom_shift,
+                            self.gnu_hash_bloom_count,
+                            self.gnu_hash_bucket_count,
+                            gnu_hash_symbol_count,
+                            |index| {
+                                out_dynsyms[(gnu_hash_index_base + index) as usize]
+                                    .gnu_hash
+                                    .unwrap()
+                            },
+                        );
+                    }
+                    SectionData::GnuVersym => {
+                        writer.write_null_gnu_versym();
+                        for out_dynsym in &out_dynsyms {
+                            let symbol = self.dynamic_symbols.get(out_dynsym.id);
+                            let mut index = symbol.version.0 as u16;
+                            if symbol.version_hidden {
+                                index |= elf::VERSYM_HIDDEN;
+                            }
+                            writer.write_gnu_versym(index);
+                        }
+                    }
+                    SectionData::GnuVerdef => {
+                        writer.write_align_gnu_verdef();
+                        if let Some(version_base) = &self.version_base {
+                            writer.write_gnu_verdef(&write::elf::Verdef {
+                                version: elf::VER_DEF_CURRENT,
+                                flags: elf::VER_FLG_BASE,
+                                index: 1,
+                                aux_count: 1,
+                                name: writer.get_dynamic_string(version_base),
+                            });
+                        }
+                        for version in &self.versions {
+                            if let VersionData::Def(def) = &version.data {
+                                let mut names = def.names.iter();
+                                let name = names.next().ok_or_else(|| {
+                                    Error(format!("Missing SHT_GNU_VERDEF name {}", version.id.0))
+                                })?;
+                                writer.write_gnu_verdef(&write::elf::Verdef {
+                                    version: elf::VER_DEF_CURRENT,
+                                    flags: def.flags,
+                                    index: version.id.0 as u16,
+                                    aux_count: def.names.len() as u16,
+                                    name: writer.get_dynamic_string(name),
+                                });
+                                for name in names {
+                                    writer.write_gnu_verdaux(writer.get_dynamic_string(name));
+                                }
+                            }
+                        }
+                    }
+                    SectionData::GnuVerneed => {
+                        writer.write_align_gnu_verneed();
+                        for file in &self.version_files {
+                            let out_file = &out_version_files[file.id.0];
+                            if out_file.versions.is_empty() {
+                                continue;
+                            }
+                            writer.write_gnu_verneed(&write::elf::Verneed {
+                                version: elf::VER_NEED_CURRENT,
+                                aux_count: out_file.versions.len() as u16,
+                                file: writer.get_dynamic_string(&file.name),
+                            });
+                            for id in &out_file.versions {
+                                let version = self.versions.get(*id);
+                                // This will always match.
+                                if let VersionData::Need(need) = &version.data {
+                                    debug_assert_eq!(*id, version.id);
+                                    writer.write_gnu_vernaux(&write::elf::Vernaux {
+                                        flags: need.flags,
+                                        index: version.id.0 as u16,
+                                        name: writer.get_dynamic_string(&need.name),
+                                    });
+                                }
+                            }
+                        }
+                    }
+                    _ => {
+                        return Err(Error(format!(
+                            "Unsupported alloc section type {:x}",
+                            section.sh_type
+                        )));
+                    }
+                }
+            }
+        }
+
+        // Write non-alloc sections.
+        for out_section in &mut out_sections {
+            let section = self.sections.get(out_section.id);
+            if !self.segments.is_empty() && section.is_alloc() {
+                continue;
+            }
+            match &section.data {
+                SectionData::Data(data) => {
+                    writer.write_align(section.sh_addralign as usize);
+                    debug_assert_eq!(out_section.offset, writer.len());
+                    writer.write(data);
+                }
+                SectionData::UninitializedData(_) => {
+                    // Nothing to do.
+                }
+                SectionData::Note(data) => {
+                    writer.write_align(section.sh_addralign as usize);
+                    debug_assert_eq!(out_section.offset, writer.len());
+                    writer.write(data);
+                }
+                SectionData::Attributes(_) => {
+                    writer.write_align(section.sh_addralign as usize);
+                    debug_assert_eq!(out_section.offset, writer.len());
+                    writer.write(&out_section.attributes);
+                }
+                // These are handled elsewhere.
+                SectionData::Relocation(_)
+                | SectionData::SectionString
+                | SectionData::Symbol
+                | SectionData::SymbolSectionIndex
+                | SectionData::String => {}
+                _ => {
+                    return Err(Error(format!(
+                        "Unsupported non-alloc section type {:x}",
+                        section.sh_type
+                    )));
+                }
+            }
+        }
+
+        writer.write_null_symbol();
+        for out_sym in &out_syms {
+            let symbol = self.symbols.get(out_sym.id);
+            let section = symbol.section.map(|id| out_sections_index[id.0].unwrap());
+            writer.write_symbol(&write::elf::Sym {
+                name: out_sym.name,
+                section,
+                st_info: symbol.st_info,
+                st_other: symbol.st_other,
+                st_shndx: symbol.st_shndx,
+                st_value: symbol.st_value,
+                st_size: symbol.st_size,
+            });
+        }
+        writer.write_symtab_shndx();
+        writer.write_strtab();
+
+        // Write non-alloc relocations.
+        for section in &self.sections {
+            if !self.segments.is_empty() && section.is_alloc() {
+                continue;
+            }
+            let SectionData::Relocation(relocations) = &section.data else {
+                continue;
+            };
+            writer.write_align_relocation();
+            for rel in relocations {
+                let r_sym = if let Some(id) = rel.symbol {
+                    out_syms_index[id.0].unwrap().0
+                } else {
+                    0
+                };
+                writer.write_relocation(
+                    section.sh_type == elf::SHT_RELA,
+                    &write::elf::Rel {
+                        r_offset: rel.r_offset,
+                        r_sym,
+                        r_type: rel.r_type,
+                        r_addend: rel.r_addend,
+                    },
+                );
+            }
+        }
+
+        writer.write_shstrtab();
+
+        writer.write_null_section_header();
+        for out_section in &out_sections {
+            let section = self.sections.get(out_section.id);
+            match &section.data {
+                SectionData::Data(_)
+                | SectionData::UninitializedData(_)
+                | SectionData::Relocation(_)
+                | SectionData::DynamicRelocation(_)
+                | SectionData::Note(_)
+                | SectionData::Dynamic(_)
+                | SectionData::Attributes(_) => {
+                    let sh_size = match &section.data {
+                        SectionData::Data(data) => data.len() as u64,
+                        SectionData::UninitializedData(len) => *len,
+                        SectionData::Relocation(relocations) => {
+                            (relocations.len()
+                                * self.class().rel_size(section.sh_type == elf::SHT_RELA))
+                                as u64
+                        }
+                        SectionData::DynamicRelocation(relocations) => {
+                            (relocations.len()
+                                * self.class().rel_size(section.sh_type == elf::SHT_RELA))
+                                as u64
+                        }
+                        SectionData::Note(data) => data.len() as u64,
+                        SectionData::Dynamic(dynamics) => {
+                            ((1 + dynamics.len()) * self.class().dyn_size()) as u64
+                        }
+                        _ => 0,
+                    };
+                    let sh_link = if let Some(id) = section.sh_link_section {
+                        if let Some(index) = out_sections_index[id.0] {
+                            index.0
+                        } else {
+                            return Err(Error(format!(
+                                "Invalid sh_link from section '{}' to deleted section '{}'",
+                                section.name,
+                                self.sections.get(id).name,
+                            )));
+                        }
+                    } else {
+                        0
+                    };
+                    let sh_info = if let Some(id) = section.sh_info_section {
+                        if let Some(index) = out_sections_index[id.0] {
+                            index.0
+                        } else {
+                            return Err(Error(format!(
+                                "Invalid sh_info link from section '{}' to deleted section '{}'",
+                                section.name,
+                                self.sections.get(id).name,
+                            )));
+                        }
+                    } else {
+                        section.sh_info
+                    };
+                    writer.write_section_header(&write::elf::SectionHeader {
+                        name: out_section.name,
+                        sh_type: section.sh_type,
+                        sh_flags: section.sh_flags,
+                        sh_addr: section.sh_addr,
+                        sh_offset: out_section.offset as u64,
+                        sh_size,
+                        sh_link,
+                        sh_info,
+                        sh_addralign: section.sh_addralign,
+                        sh_entsize: section.sh_entsize,
+                    });
+                }
+                SectionData::SectionString => {
+                    writer.write_shstrtab_section_header();
+                }
+                SectionData::Symbol => {
+                    writer.write_symtab_section_header(num_local);
+                }
+                SectionData::SymbolSectionIndex => {
+                    writer.write_symtab_shndx_section_header();
+                }
+                SectionData::String => {
+                    writer.write_strtab_section_header();
+                }
+                SectionData::DynamicString => {
+                    writer.write_dynstr_section_header(section.sh_addr);
+                }
+                SectionData::DynamicSymbol => {
+                    writer.write_dynsym_section_header(section.sh_addr, 1);
+                }
+                SectionData::Hash => {
+                    writer.write_hash_section_header(section.sh_addr);
+                }
+                SectionData::GnuHash => {
+                    writer.write_gnu_hash_section_header(section.sh_addr);
+                }
+                SectionData::GnuVersym => {
+                    writer.write_gnu_versym_section_header(section.sh_addr);
+                }
+                SectionData::GnuVerdef => {
+                    writer.write_gnu_verdef_section_header(section.sh_addr);
+                }
+                SectionData::GnuVerneed => {
+                    writer.write_gnu_verneed_section_header(section.sh_addr);
+                }
+            }
+        }
+        debug_assert_eq!(writer.reserved_len(), writer.len());
+        Ok(())
+    }
+
+    /// Delete segments, symbols, relocations, and dynamics that refer
+    /// to deleted items.
+    ///
+    /// This calls `delete_orphan_segments`, `delete_orphan_symbols`,
+    /// `delete_orphan_relocations`, and `delete_orphan_dynamics`.
+    pub fn delete_orphans(&mut self) {
+        self.delete_orphan_segments();
+        self.delete_orphan_symbols();
+        self.delete_orphan_relocations();
+        self.delete_orphan_dynamics();
+    }
+
+    /// Set the delete flag for segments that only refer to deleted sections.
+    pub fn delete_orphan_segments(&mut self) {
+        let sections = &self.sections;
+        for segment in &mut self.segments {
+            // We only delete segments that have become empty due to section deletions.
+            if segment.sections.is_empty() {
+                continue;
+            }
+            segment.sections.retain(|id| !sections.get(*id).delete);
+            segment.delete = segment.sections.is_empty();
+        }
+    }
+
+    /// Set the delete flag for symbols that refer to deleted sections.
+    pub fn delete_orphan_symbols(&mut self) {
+        for symbol in &mut self.symbols {
+            if let Some(section) = symbol.section {
+                if self.sections.get_mut(section).delete {
+                    symbol.delete = true;
+                }
+            }
+        }
+        for symbol in &mut self.dynamic_symbols {
+            if let Some(section) = symbol.section {
+                if self.sections.get_mut(section).delete {
+                    symbol.delete = true;
+                }
+            }
+        }
+    }
+
+    /// Delete relocations that refer to deleted symbols.
+    pub fn delete_orphan_relocations(&mut self) {
+        let symbols = &self.symbols;
+        let dynamic_symbols = &self.dynamic_symbols;
+        for section in &mut self.sections {
+            match &mut section.data {
+                SectionData::Relocation(relocations) => {
+                    relocations.retain(|relocation| match relocation.symbol {
+                        None => true,
+                        Some(id) => !symbols.get(id).delete,
+                    });
+                }
+                SectionData::DynamicRelocation(relocations) => {
+                    relocations.retain(|relocation| match relocation.symbol {
+                        None => true,
+                        Some(id) => !dynamic_symbols.get(id).delete,
+                    });
+                }
+                _ => {}
+            }
+        }
+    }
+
+    /// Delete dynamic entries that refer to deleted sections.
+    pub fn delete_orphan_dynamics(&mut self) {
+        let mut have_dynsym = false;
+        let mut have_dynstr = false;
+        let mut have_hash = false;
+        let mut have_gnu_hash = false;
+        let mut have_versym = false;
+        let mut have_verdef = false;
+        let mut have_verneed = false;
+        for section in &self.sections {
+            match &section.data {
+                SectionData::DynamicSymbol => have_dynsym = true,
+                SectionData::DynamicString => have_dynstr = true,
+                SectionData::Hash => have_hash = true,
+                SectionData::GnuHash => have_gnu_hash = true,
+                SectionData::GnuVersym => have_versym = true,
+                SectionData::GnuVerdef => have_verdef = true,
+                SectionData::GnuVerneed => have_verneed = true,
+                _ => {}
+            }
+        }
+        for section in &mut self.sections {
+            if let SectionData::Dynamic(dynamics) = &mut section.data {
+                dynamics.retain(|dynamic| match dynamic {
+                    Dynamic::Auto {
+                        tag: elf::DT_SYMTAB,
+                    } => have_dynsym,
+                    Dynamic::Auto {
+                        tag: elf::DT_STRTAB,
+                    }
+                    | Dynamic::Auto { tag: elf::DT_STRSZ } => have_dynstr,
+                    Dynamic::Auto { tag: elf::DT_HASH } => have_hash,
+                    Dynamic::Auto {
+                        tag: elf::DT_GNU_HASH,
+                    } => have_gnu_hash,
+                    Dynamic::Auto {
+                        tag: elf::DT_VERSYM,
+                    } => have_versym,
+                    Dynamic::Auto {
+                        tag: elf::DT_VERNEED,
+                    }
+                    | Dynamic::Auto {
+                        tag: elf::DT_VERNEEDNUM,
+                    } => have_verneed,
+                    Dynamic::Auto {
+                        tag: elf::DT_VERDEF,
+                    }
+                    | Dynamic::Auto {
+                        tag: elf::DT_VERDEFNUM,
+                    } => have_verdef,
+                    _ => true,
+                });
+            }
+        }
+    }
+
+    /// Delete unused GNU version entries.
+    pub fn delete_unused_versions(&mut self) {
+        let mut version_used = vec![false; self.versions.len() + VERSION_ID_BASE];
+        for symbol in &self.dynamic_symbols {
+            version_used[symbol.version.0] = true;
+        }
+        let mut version_file_used = vec![false; self.version_files.len()];
+        for version in &mut self.versions {
+            if !version_used[version.id.0] {
+                version.delete = true;
+                continue;
+            }
+            if let VersionData::Need(need) = &version.data {
+                version_file_used[need.file.0] = true;
+            }
+        }
+        for file in &mut self.version_files {
+            if !version_file_used[file.id.0] {
+                file.delete = true;
+            }
+        }
+    }
+
+    /// Return the ELF file class that will be written.
+    ///
+    /// This can be useful for calculating sizes.
+    pub fn class(&self) -> write::elf::Class {
+        write::elf::Class { is_64: self.is_64 }
+    }
+
+    /// Calculate the size of the file header.
+    pub fn file_header_size(&self) -> usize {
+        self.class().file_header_size()
+    }
+
+    /// Calculate the size of the program headers.
+    pub fn program_headers_size(&self) -> usize {
+        self.segments.count() * self.class().program_header_size()
+    }
+
+    /// Calculate the size of the dynamic symbol table.
+    ///
+    /// To get an accurate result, you may need to first call
+    /// [`Self::delete_orphan_symbols`].
+    pub fn dynamic_symbol_size(&self) -> usize {
+        (1 + self.dynamic_symbols.count()) * self.class().sym_size()
+    }
+
+    /// Calculate the size of the dynamic string table.
+    ///
+    /// This adds all of the currently used dynamic strings to a string table,
+    /// calculates the size of the string table, and discards the string table.
+    ///
+    /// To get an accurate result, you may need to first call
+    /// [`Self::delete_orphan_symbols`] and [`Self::delete_unused_versions`].
+    pub fn dynamic_string_size(&self) -> usize {
+        let mut dynstr = write::string::StringTable::default();
+        for section in &self.sections {
+            if let SectionData::Dynamic(dynamics) = &section.data {
+                for dynamic in dynamics {
+                    if let Dynamic::String { val, .. } = dynamic {
+                        dynstr.add(val);
+                    }
+                }
+            }
+        }
+        for symbol in &self.dynamic_symbols {
+            dynstr.add(&symbol.name);
+        }
+        if let Some(version_base) = &self.version_base {
+            dynstr.add(version_base);
+        }
+        for version in &self.versions {
+            match &version.data {
+                VersionData::Def(def) => {
+                    for name in &def.names {
+                        dynstr.add(name);
+                    }
+                }
+                VersionData::Need(need) => {
+                    dynstr.add(&need.name);
+                }
+            }
+        }
+        for file in &self.version_files {
+            dynstr.add(&file.name);
+        }
+        dynstr.size(1)
+    }
+
+    /// Calculate the size of the hash table.
+    ///
+    /// To get an accurate result, you may need to first call
+    /// [`Self::delete_orphan_symbols`].
+    pub fn hash_size(&self) -> usize {
+        let chain_count = 1 + self.dynamic_symbols.count();
+        self.class()
+            .hash_size(self.hash_bucket_count, chain_count as u32)
+    }
+
+    /// Calculate the size of the GNU hash table.
+    ///
+    /// To get an accurate result, you may need to first call
+    /// [`Self::delete_orphan_symbols`].
+    pub fn gnu_hash_size(&self) -> usize {
+        let symbol_count = self.dynamic_symbols.count_defined();
+        self.class().gnu_hash_size(
+            self.gnu_hash_bloom_count,
+            self.gnu_hash_bucket_count,
+            symbol_count as u32,
+        )
+    }
+
+    /// Calculate the size of the GNU symbol version section.
+    ///
+    /// To get an accurate result, you may need to first call
+    /// [`Self::delete_orphan_symbols`] and [`Self::delete_unused_versions`].
+    pub fn gnu_versym_size(&self) -> usize {
+        let symbol_count = 1 + self.dynamic_symbols.count();
+        self.class().gnu_versym_size(symbol_count)
+    }
+
+    /// Calculate the size of the GNU version definition section.
+    ///
+    /// To get an accurate result, you may need to first call
+    /// [`Self::delete_orphan_symbols`] and [`Self::delete_unused_versions`].
+    pub fn gnu_verdef_size(&self) -> usize {
+        let mut verdef_count = 0;
+        let mut verdaux_count = 0;
+        if self.version_base.is_some() {
+            verdef_count += 1;
+            verdaux_count += 1;
+        }
+        for version in &self.versions {
+            if let VersionData::Def(def) = &version.data {
+                verdef_count += 1;
+                verdaux_count += def.names.len();
+            }
+        }
+        self.class().gnu_verdef_size(verdef_count, verdaux_count)
+    }
+
+    /// Calculate the size of the GNU version dependency section.
+    ///
+    /// To get an accurate result, you may need to first call
+    /// [`Self::delete_orphan_symbols`] and [`Self::delete_unused_versions`].
+    pub fn gnu_verneed_size(&self) -> usize {
+        let verneed_count = self.version_files.count();
+        let mut vernaux_count = 0;
+        for version in &self.versions {
+            if let VersionData::Need(_) = &version.data {
+                vernaux_count += 1;
+            }
+        }
+        self.class().gnu_verneed_size(verneed_count, vernaux_count)
+    }
+
+    /// Calculate the memory size of a section.
+    ///
+    /// Returns 0 for sections that are deleted or aren't allocated.
+    ///
+    /// To get an accurate result, you may need to first call
+    /// [`Self::delete_orphan_symbols`] and [`Self::delete_unused_versions`].
+    pub fn section_size(&self, section: &Section<'_>) -> usize {
+        if section.delete || !section.is_alloc() {
+            return 0;
+        }
+        match &section.data {
+            SectionData::Data(data) => data.len(),
+            SectionData::UninitializedData(len) => *len as usize,
+            SectionData::Relocation(relocations) => {
+                relocations.len() * self.class().rel_size(section.sh_type == elf::SHT_RELA)
+            }
+            SectionData::DynamicRelocation(relocations) => {
+                relocations.len() * self.class().rel_size(section.sh_type == elf::SHT_RELA)
+            }
+            SectionData::Note(data) => data.len(),
+            SectionData::Dynamic(dynamics) => (1 + dynamics.len()) * self.class().dyn_size(),
+            SectionData::DynamicString => self.dynamic_string_size(),
+            SectionData::DynamicSymbol => self.dynamic_symbol_size(),
+            SectionData::Hash => self.hash_size(),
+            SectionData::GnuHash => self.gnu_hash_size(),
+            SectionData::GnuVersym => self.gnu_versym_size(),
+            SectionData::GnuVerdef => self.gnu_verdef_size(),
+            SectionData::GnuVerneed => self.gnu_verneed_size(),
+            // None of these should be allocated.
+            SectionData::SectionString
+            | SectionData::Symbol
+            | SectionData::SymbolSectionIndex
+            | SectionData::String
+            | SectionData::Attributes(_) => 0,
+        }
+    }
+
+    /// Set the `sh_size` field for every allocated section.
+    ///
+    /// This is useful to call prior to doing memory layout.
+    ///
+    /// To get an accurate result, you may need to first call
+    /// [`Self::delete_orphan_symbols`] and [`Self::delete_unused_versions`].
+    pub fn set_section_sizes(&mut self) {
+        for id in (0..self.sections.len()).map(SectionId) {
+            let section = self.sections.get(id);
+            if section.delete || !section.is_alloc() {
+                continue;
+            }
+            self.sections.get_mut(id).sh_size = self.section_size(section) as u64;
+        }
+    }
+
+    /// Find the section containing the dynamic table.
+    ///
+    /// This uses the `PT_DYNAMIC` program header to find the dynamic section.
+    pub fn dynamic_section(&self) -> Option<SectionId> {
+        let segment = self
+            .segments
+            .iter()
+            .find(|segment| segment.p_type == elf::PT_DYNAMIC)?;
+        // TODO: handle multiple sections in the segment?
+        segment.sections.iter().copied().next()
+    }
+
+    /// Find the dynamic table entries.
+    ///
+    /// This uses the `PT_DYNAMIC` program header to find the dynamic section,
+    pub fn dynamic_data(&self) -> Option<&[Dynamic<'data>]> {
+        let section = self.dynamic_section()?;
+        match &self.sections.get(section).data {
+            SectionData::Dynamic(dynamics) => Some(dynamics),
+            _ => None,
+        }
+    }
+
+    /// Find the dynamic table entries.
+    ///
+    /// This uses the `PT_DYNAMIC` program header to find the dynamic section,
+    pub fn dynamic_data_mut(&mut self) -> Option<&mut Vec<Dynamic<'data>>> {
+        let section = self.dynamic_section()?;
+        match &mut self.sections.get_mut(section).data {
+            SectionData::Dynamic(dynamics) => Some(dynamics),
+            _ => None,
+        }
+    }
+
+    /// Find the section containing the interpreter path.
+    ///
+    /// This uses the `PT_INTERP` program header to find the interp section.
+    pub fn interp_section(&self) -> Option<SectionId> {
+        let segment = self
+            .segments
+            .iter()
+            .find(|segment| segment.p_type == elf::PT_INTERP)?;
+        // TODO: handle multiple sections in the segment?
+        segment.sections.iter().copied().next()
+    }
+
+    /// Find the interpreter path.
+    ///
+    /// This uses the `PT_INTERP` program header to find the interp section.
+    pub fn interp_data(&self) -> Option<&[u8]> {
+        let section = self.interp_section()?;
+        match &self.sections.get(section).data {
+            SectionData::Data(data) => Some(data),
+            _ => None,
+        }
+    }
+
+    /// Find the interpreter path.
+    ///
+    /// This uses the `PT_INTERP` program header to find the interp section.
+    pub fn interp_data_mut(&mut self) -> Option<&mut Bytes<'data>> {
+        let section = self.interp_section()?;
+        match &mut self.sections.get_mut(section).data {
+            SectionData::Data(data) => Some(data),
+            _ => None,
+        }
+    }
+}
+
+/// ELF file header.
+///
+/// This corresponds to fields in [`elf::FileHeader32`] or [`elf::FileHeader64`].
+/// This only contains the ELF file header fields that can be modified.
+/// The other fields are automatically calculated.
+#[derive(Debug, Default)]
+pub struct Header {
+    /// The OS ABI field in the file header.
+    ///
+    /// One of the `ELFOSABI*` constants.
+    pub os_abi: u8,
+    /// The ABI version field in the file header.
+    ///
+    /// The meaning of this field depends on the `os_abi` value.
+    pub abi_version: u8,
+    /// The object file type in the file header.
+    ///
+    /// One of the `ET_*` constants.
+    pub e_type: u16,
+    /// The architecture in the file header.
+    ///
+    /// One of the `EM_*` constants.
+    pub e_machine: u16,
+    /// Entry point virtual address in the file header.
+    pub e_entry: u64,
+    /// The processor-specific flags in the file header.
+    ///
+    /// A combination of the `EF_*` constants.
+    pub e_flags: u32,
+    /// The file offset of the program header table.
+    ///
+    /// Writing will fail if the program header table cannot be placed at this offset.
+    pub e_phoff: u64,
+}
+
+/// An ID for referring to a segment in [`Segments`].
+#[derive(Clone, Copy, PartialEq, Eq)]
+pub struct SegmentId(usize);
+
+impl fmt::Debug for SegmentId {
+    fn fmt(&self, f: &mut fmt::Formatter<'_>) -> fmt::Result {
+        write!(f, "SegmentId({})", self.0)
+    }
+}
+
+impl Id for SegmentId {
+    fn index(&self) -> usize {
+        self.0
+    }
+}
+
+impl IdPrivate for SegmentId {
+    fn new(id: usize) -> Self {
+        SegmentId(id)
+    }
+}
+
+/// A segment in [`Segments`].
+///
+/// This corresponds to [`elf::ProgramHeader32`] or [`elf::ProgramHeader64`].
+#[derive(Debug)]
+pub struct Segment<'data> {
+    id: SegmentId,
+    /// Ignore this segment when writing the ELF file.
+    pub delete: bool,
+    /// The `p_type` field in the ELF program header.
+    ///
+    /// One of the `PT_*` constants.
+    pub p_type: u32,
+    /// The `p_flags` field in the ELF program header.
+    ///
+    /// A combination of the `PF_*` constants.
+    pub p_flags: u32,
+    /// The `p_offset` field in the ELF program header.
+    ///
+    /// This is the file offset of the data in the segment. This should
+    /// correspond to the file offset of the sections that are placed in
+    /// this segment. Currently there is no support for section data
+    /// that is not contained in sections.
+    pub p_offset: u64,
+    /// The `p_vaddr` field in the ELF program header.
+    pub p_vaddr: u64,
+    /// The `p_paddr` field in the ELF program header.
+    pub p_paddr: u64,
+    /// The `p_filesz` field in the ELF program header.
+    pub p_filesz: u64,
+    /// The `p_memsz` field in the ELF program header.
+    pub p_memsz: u64,
+    /// The `p_align` field in the ELF program header.
+    pub p_align: u64,
+    /// The sections contained in this segment.
+    pub sections: Vec<SectionId>,
+    // Might need to add reference to data if no sections.
+    marker: PhantomData<&'data ()>,
+}
+
+impl<'data> Item for Segment<'data> {
+    type Id = SegmentId;
+
+    fn is_deleted(&self) -> bool {
+        self.delete
+    }
+}
+
+impl<'data> Segment<'data> {
+    /// The ID used for referring to this segment.
+    pub fn id(&self) -> SegmentId {
+        self.id
+    }
+
+    /// Returns true if the segment type is `PT_LOAD`.
+    pub fn is_load(&self) -> bool {
+        self.p_type == elf::PT_LOAD
+    }
+
+    /// Returns true if the segment contains the given file offset.
+    pub fn contains_offset(&self, offset: u64) -> bool {
+        offset >= self.p_offset && offset - self.p_offset < self.p_filesz
+    }
+
+    /// Return the address corresponding to the given file offset.
+    ///
+    /// This will return a meaningless value if `contains_offset` is false.
+    pub fn address_from_offset(&self, offset: u64) -> u64 {
+        self.p_vaddr
+            .wrapping_add(offset.wrapping_sub(self.p_offset))
+    }
+
+    /// Returns true if the segment contains the given address.
+    pub fn contains_address(&self, address: u64) -> bool {
+        address >= self.p_vaddr && address - self.p_vaddr < self.p_memsz
+    }
+
+    /// Remove all sections from the segment, and set its size to zero.
+    pub fn remove_sections(&mut self) {
+        self.p_filesz = 0;
+        self.p_memsz = 0;
+        self.sections.clear();
+    }
+
+    /// Add a section to the segment.
+    ///
+    /// If this is a [`elf::PT_LOAD`] segment, then the file offset and address of the
+    /// section is changed to be at the end of the segment.
+    ///
+    /// The segment's file and address ranges are extended to include the section.
+    /// This uses the `sh_size` field of the section, not the size of the section data.
+    ///
+    /// The section's id is added to the segment's list of sections.
+    pub fn append_section(&mut self, section: &mut Section<'_>) {
+        debug_assert_eq!(self.p_filesz, self.p_memsz);
+        if self.p_type == elf::PT_LOAD {
+            let align = section.sh_addralign;
+            let offset = (self.p_offset + self.p_filesz + (align - 1)) & !(align - 1);
+            let addr = (self.p_paddr + self.p_memsz + (align - 1)) & !(align - 1);
+            section.sh_offset = offset;
+            section.sh_addr = addr;
+        }
+        self.append_section_range(section);
+        self.sections.push(section.id);
+    }
+
+    /// Extend this segment's file and address ranges to include the given section.
+    ///
+    /// If the segment's `p_memsz` is zero, then this signifies that the segment
+    /// has no file or address range yet. In this case, the segment's file and address
+    /// ranges are set equal to the section. Otherwise, the segment's file and address
+    /// ranges are extended to include the section.
+    ///
+    /// This uses the `sh_size` field of the section, not the size of the section data.
+    pub fn append_section_range(&mut self, section: &Section<'_>) {
+        let section_filesize = if section.sh_type == elf::SHT_NOBITS {
+            0
+        } else {
+            section.sh_size
+        };
+        if self.p_memsz == 0 {
+            self.p_offset = section.sh_offset;
+            self.p_filesz = section_filesize;
+            self.p_vaddr = section.sh_addr;
+            self.p_paddr = section.sh_addr;
+            self.p_memsz = section.sh_size;
+        } else {
+            if self.p_offset > section.sh_offset {
+                self.p_offset = section.sh_offset;
+            }
+            let filesz = section.sh_offset + section_filesize - self.p_offset;
+            if self.p_filesz < filesz {
+                self.p_filesz = filesz;
+            }
+            if self.p_vaddr > section.sh_addr {
+                self.p_vaddr = section.sh_addr;
+                self.p_paddr = section.sh_addr;
+            }
+            let memsz = section.sh_addr + section.sh_size - self.p_vaddr;
+            if self.p_memsz < memsz {
+                self.p_memsz = memsz;
+            }
+        }
+    }
+
+    /// Recalculate the file and address ranges of the segment.
+    ///
+    /// Resets the segment's file and address ranges to zero, and then
+    /// calls `append_section_range` for each section in the segment.
+    pub fn recalculate_ranges(&mut self, sections: &Sections<'data>) {
+        self.p_offset = 0;
+        self.p_filesz = 0;
+        self.p_vaddr = 0;
+        self.p_paddr = 0;
+        self.p_memsz = 0;
+        let ids = core::mem::take(&mut self.sections);
+        for id in &ids {
+            let section = sections.get(*id);
+            self.append_section_range(section);
+        }
+        self.sections = ids;
+    }
+}
+
+/// A segment table.
+pub type Segments<'data> = Table<Segment<'data>>;
+
+impl<'data> Segments<'data> {
+    /// Add a new segment to the table.
+    pub fn add(&mut self) -> &mut Segment<'data> {
+        let id = self.next_id();
+        self.push(Segment {
+            id,
+            delete: false,
+            p_type: 0,
+            p_flags: 0,
+            p_offset: 0,
+            p_vaddr: 0,
+            p_paddr: 0,
+            p_filesz: 0,
+            p_memsz: 0,
+            p_align: 0,
+            sections: Vec::new(),
+            marker: PhantomData,
+        });
+        self.get_mut(id)
+    }
+
+    /// Find a `PT_LOAD` segment containing the given offset.
+    pub fn find_load_segment_from_offset(&self, offset: u64) -> Option<&Segment<'data>> {
+        self.iter()
+            .find(|segment| segment.is_load() && segment.contains_offset(offset))
+    }
+
+    /// Add a new `PT_LOAD` segment to the table.
+    ///
+    /// The file offset and address will be derived from the current maximum for any segment.
+    pub fn add_load_segment(&mut self, flags: u32, align: u64) -> &mut Segment<'data> {
+        let mut max_offset = 0;
+        let mut max_addr = 0;
+        for segment in &*self {
+            let offset = segment.p_offset + segment.p_filesz;
+            if max_offset < offset {
+                max_offset = offset;
+            }
+            let addr = segment.p_vaddr + segment.p_memsz;
+            if max_addr < addr {
+                max_addr = addr;
+            }
+        }
+        // No alignment is required for the segment file offset because sections
+        // will add their alignment to the file offset when they are added.
+        let offset = max_offset;
+        // The address must be chosen so that addr % align == offset % align.
+        let addr = ((max_addr + (align - 1)) & !(align - 1)) + (offset & (align - 1));
+
+        let segment = self.add();
+        segment.p_type = elf::PT_LOAD;
+        segment.p_flags = flags;
+        segment.p_offset = offset;
+        segment.p_vaddr = addr;
+        segment.p_paddr = addr;
+        segment.p_align = align;
+        segment
+    }
+
+    /// Add a copy of a segment to the table.
+    ///
+    /// This will copy the segment type, flags and alignment.
+    ///
+    /// Additionally, if the segment type is `PT_LOAD`, then the file offset and address
+    /// will be set as in `add_load_segment`.
+    pub fn copy(&mut self, id: SegmentId) -> &mut Segment<'data> {
+        let segment = self.get(id);
+        let p_type = segment.p_type;
+        let p_flags = segment.p_flags;
+        let p_align = segment.p_align;
+        if p_type == elf::PT_LOAD {
+            self.add_load_segment(p_flags, p_align)
+        } else {
+            let segment = self.add();
+            segment.p_type = p_type;
+            segment.p_flags = p_flags;
+            segment.p_align = p_align;
+            segment
+        }
+    }
+}
+
+/// An ID for referring to a section in [`Sections`].
+#[derive(Clone, Copy, PartialEq, Eq)]
+pub struct SectionId(usize);
+
+impl fmt::Debug for SectionId {
+    fn fmt(&self, f: &mut fmt::Formatter<'_>) -> fmt::Result {
+        write!(f, "SectionId({})", self.0)
+    }
+}
+
+impl Id for SectionId {
+    fn index(&self) -> usize {
+        self.0
+    }
+}
+
+impl IdPrivate for SectionId {
+    fn new(id: usize) -> Self {
+        SectionId(id)
+    }
+}
+
+/// A section in [`Sections`].
+///
+/// This corresponds to [`elf::SectionHeader32`] or [`elf::SectionHeader64`].
+#[derive(Debug)]
+pub struct Section<'data> {
+    id: SectionId,
+    /// Ignore this section when writing the ELF file.
+    pub delete: bool,
+    /// The name of the section.
+    ///
+    /// This is automatically added to the section header string table,
+    /// and the resulting string table offset is used to set the `sh_name`
+    /// field in the ELF section header.
+    pub name: ByteString<'data>,
+    /// The `sh_type` field in the ELF section header.
+    ///
+    /// One of the `SHT_*` constants.
+    pub sh_type: u32,
+    /// The `sh_flags` field in the ELF section header.
+    ///
+    /// A combination of the `SHF_*` constants.
+    pub sh_flags: u64,
+    /// The `sh_addr` field in the ELF section header.
+    pub sh_addr: u64,
+    /// The `sh_offset` field in the ELF section header.
+    ///
+    /// This is the file offset of the data in the section.
+    /// Writing will fail if the data cannot be placed at this offset.
+    ///
+    /// This is only used for sections that have `SHF_ALLOC` set.
+    /// For other sections, the section data is written at the next available
+    /// offset.
+    pub sh_offset: u64,
+    /// The `sh_size` field in the ELF section header.
+    ///
+    /// This size is not used when writing. The size of the `data` field is
+    /// used instead.
+    pub sh_size: u64,
+    /// The ID of the section linked to by the `sh_link` field in the ELF section header.
+    pub sh_link_section: Option<SectionId>,
+    /// The `sh_info` field in the ELF section header.
+    ///
+    /// Only used if `sh_info_section` is `None`.
+    pub sh_info: u32,
+    /// The ID of the section linked to by the `sh_info` field in the ELF section header.
+    pub sh_info_section: Option<SectionId>,
+    /// The `sh_addralign` field in the ELF section header.
+    pub sh_addralign: u64,
+    /// The `sh_entsize` field in the ELF section header.
+    pub sh_entsize: u64,
+    /// The section data.
+    pub data: SectionData<'data>,
+}
+
+impl<'data> Item for Section<'data> {
+    type Id = SectionId;
+
+    fn is_deleted(&self) -> bool {
+        self.delete
+    }
+}
+
+impl<'data> Section<'data> {
+    /// The ID used for referring to this section.
+    pub fn id(&self) -> SectionId {
+        self.id
+    }
+
+    /// Returns true if the section flags include `SHF_ALLOC`.
+    pub fn is_alloc(&self) -> bool {
+        self.sh_flags & u64::from(elf::SHF_ALLOC) != 0
+    }
+
+    /// Return the segment permission flags that are equivalent to the section flags.
+    pub fn p_flags(&self) -> u32 {
+        let mut p_flags = elf::PF_R;
+        if self.sh_flags & u64::from(elf::SHF_WRITE) != 0 {
+            p_flags |= elf::PF_W;
+        }
+        if self.sh_flags & u64::from(elf::SHF_EXECINSTR) != 0 {
+            p_flags |= elf::PF_X;
+        }
+        p_flags
+    }
+}
+
+/// The data for a [`Section`].
+#[derive(Debug, Clone)]
+pub enum SectionData<'data> {
+    /// The section contains the given raw data bytes.
+    Data(Bytes<'data>),
+    /// The section contains uninitialised data bytes of the given length.
+    UninitializedData(u64),
+    /// The section contains relocations.
+    Relocation(Vec<Relocation>),
+    /// The section contains dynamic relocations.
+    DynamicRelocation(Vec<DynamicRelocation>),
+    /// The section contains notes.
+    // TODO: parse notes
+    Note(Bytes<'data>),
+    /// The section contains dynamic entries.
+    Dynamic(Vec<Dynamic<'data>>),
+    /// The section contains attributes.
+    ///
+    /// This may be GNU attributes or other vendor-specific attributes.
+    Attributes(AttributesSection<'data>),
+    /// The section contains the strings for the section headers.
+    SectionString,
+    /// The section contains the symbol table.
+    Symbol,
+    /// The section contains the extended section index for the symbol table.
+    SymbolSectionIndex,
+    /// The section contains the strings for symbol table.
+    String,
+    /// The section contains the dynamic symbol table.
+    DynamicSymbol,
+    /// The section contains the dynamic string table.
+    DynamicString,
+    /// The section contains the hash table.
+    Hash,
+    /// The section contains the GNU hash table.
+    GnuHash,
+    /// The section contains the GNU symbol versions.
+    GnuVersym,
+    /// The section contains the GNU version definitions.
+    GnuVerdef,
+    /// The section contains the GNU version dependencies.
+    GnuVerneed,
+}
+
+/// A section table.
+pub type Sections<'data> = Table<Section<'data>>;
+
+impl<'data> Sections<'data> {
+    /// Add a new section to the table.
+    pub fn add(&mut self) -> &mut Section<'data> {
+        let id = self.next_id();
+        self.push(Section {
+            id,
+            delete: false,
+            name: ByteString::default(),
+            sh_type: 0,
+            sh_flags: 0,
+            sh_addr: 0,
+            sh_offset: 0,
+            sh_size: 0,
+            sh_link_section: None,
+            sh_info: 0,
+            sh_info_section: None,
+            sh_addralign: 0,
+            sh_entsize: 0,
+            data: SectionData::Data(Bytes::default()),
+        })
+    }
+
+    /// Add a copy of a section to the table.
+    ///
+    /// This will set the file offset of the copy to zero.
+    /// [`Segment::append_section`] can be used to assign a valid file offset and a new address.
+    pub fn copy(&mut self, id: SectionId) -> &mut Section<'data> {
+        let section = self.get(id);
+        let id = self.next_id();
+        let name = section.name.clone();
+        let sh_type = section.sh_type;
+        let sh_flags = section.sh_flags;
+        let sh_addr = section.sh_addr;
+        let sh_size = section.sh_size;
+        let sh_link_section = section.sh_link_section;
+        let sh_info = section.sh_info;
+        let sh_info_section = section.sh_info_section;
+        let sh_addralign = section.sh_addralign;
+        let sh_entsize = section.sh_entsize;
+        let data = section.data.clone();
+        self.push(Section {
+            id,
+            delete: false,
+            name,
+            sh_type,
+            sh_flags,
+            sh_addr,
+            sh_offset: 0,
+            sh_size,
+            sh_link_section,
+            sh_info,
+            sh_info_section,
+            sh_addralign,
+            sh_entsize,
+            data,
+        })
+    }
+}
+
+/// An ID for referring to a symbol in [`Symbols`].
+#[derive(Clone, Copy, PartialEq, Eq)]
+pub struct SymbolId<const DYNAMIC: bool = false>(usize);
+
+impl<const DYNAMIC: bool> fmt::Debug for SymbolId<DYNAMIC> {
+    fn fmt(&self, f: &mut fmt::Formatter<'_>) -> fmt::Result {
+        self.0.fmt(f)
+    }
+}
+
+impl<const DYNAMIC: bool> Id for SymbolId<DYNAMIC> {
+    fn index(&self) -> usize {
+        self.0
+    }
+}
+
+impl<const DYNAMIC: bool> IdPrivate for SymbolId<DYNAMIC> {
+    fn new(id: usize) -> Self {
+        SymbolId(id)
+    }
+}
+
+/// A symbol in [`Symbols`].
+///
+/// This corresponds to [`elf::Sym32`] or [`elf::Sym64`].
+#[derive(Debug)]
+pub struct Symbol<'data, const DYNAMIC: bool = false> {
+    id: SymbolId<DYNAMIC>,
+    /// Ignore this symbol when writing the ELF file.
+    pub delete: bool,
+    /// The name of the symbol.
+    pub name: ByteString<'data>,
+    /// The section referenced by the symbol.
+    ///
+    /// Used to set the `st_shndx` field in the ELF symbol.
+    pub section: Option<SectionId>,
+    /// The `st_info` field in the ELF symbol.
+    pub st_info: u8,
+    /// The `st_other` field in the ELF symbol.
+    pub st_other: u8,
+    /// The `st_shndx` field in the ELF symbol.
+    ///
+    /// Only used if `Self::section` is `None`.
+    pub st_shndx: u16,
+    /// The `st_value` field in the ELF symbol.
+    pub st_value: u64,
+    /// The `st_size` field in the ELF symbol.
+    pub st_size: u64,
+    /// GNU version for dynamic symbols.
+    pub version: VersionId,
+    /// Set the [`elf::VERSYM_HIDDEN`] flag for this symbol.
+    pub version_hidden: bool,
+}
+
+impl<'data, const DYNAMIC: bool> Item for Symbol<'data, DYNAMIC> {
+    type Id = SymbolId<DYNAMIC>;
+
+    fn is_deleted(&self) -> bool {
+        self.delete
+    }
+}
+
+impl<'data, const DYNAMIC: bool> Symbol<'data, DYNAMIC> {
+    /// The ID used for referring to this symbol.
+    pub fn id(&self) -> SymbolId<DYNAMIC> {
+        self.id
+    }
+
+    /// Get the `st_bind` component of the `st_info` field.
+    #[inline]
+    pub fn st_bind(&self) -> u8 {
+        self.st_info >> 4
+    }
+
+    /// Get the `st_type` component of the `st_info` field.
+    #[inline]
+    pub fn st_type(&self) -> u8 {
+        self.st_info & 0xf
+    }
+
+    /// Set the `st_info` field given the `st_bind` and `st_type` components.
+    #[inline]
+    pub fn set_st_info(&mut self, st_bind: u8, st_type: u8) {
+        self.st_info = (st_bind << 4) + (st_type & 0xf);
+    }
+}
+
+/// A symbol table.
+pub type Symbols<'data, const DYNAMIC: bool = false> = Table<Symbol<'data, DYNAMIC>>;
+
+impl<'data, const DYNAMIC: bool> Symbols<'data, DYNAMIC> {
+    /// Number of defined symbols.
+    pub fn count_defined(&self) -> usize {
+        self.into_iter()
+            .filter(|symbol| symbol.st_shndx != elf::SHN_UNDEF)
+            .count()
+    }
+
+    /// Add a new symbol to the table.
+    pub fn add(&mut self) -> &mut Symbol<'data, DYNAMIC> {
+        let id = self.next_id();
+        self.push(Symbol {
+            id,
+            delete: false,
+            name: ByteString::default(),
+            section: None,
+            st_info: 0,
+            st_other: 0,
+            st_shndx: 0,
+            st_value: 0,
+            st_size: 0,
+            version: VersionId::local(),
+            version_hidden: false,
+        })
+    }
+}
+
+/// A relocation stored in a [`Section`].
+///
+/// This corresponds to [`elf::Rel32`], [`elf::Rela32`], [`elf::Rel64`] or [`elf::Rela64`].
+#[derive(Debug, Clone, Copy, PartialEq, Eq)]
+pub struct Relocation<const DYNAMIC: bool = false> {
+    /// The `r_offset` field in the ELF relocation.
+    pub r_offset: u64,
+    /// The symbol referenced by the ELF relocation.
+    pub symbol: Option<SymbolId<DYNAMIC>>,
+    /// The `r_type` field in the ELF relocation.
+    pub r_type: u32,
+    /// The `r_addend` field in the ELF relocation.
+    ///
+    /// Only used if the section type is `SHT_RELA`.
+    pub r_addend: i64,
+}
+
+/// A dynamic symbol ID.
+pub type DynamicSymbolId = SymbolId<true>;
+
+/// A dynamic symbol.
+pub type DynamicSymbol<'data> = Symbol<'data, true>;
+
+/// A dynamic symbol table.
+pub type DynamicSymbols<'data> = Symbols<'data, true>;
+
+/// A dynamic relocation.
+pub type DynamicRelocation = Relocation<true>;
+
+/// An entry in the dynamic section.
+///
+/// This corresponds to [`elf::Dyn32`] or [`elf::Dyn64`].
+#[derive(Debug, Clone, PartialEq, Eq)]
+pub enum Dynamic<'data> {
+    /// The value is an automatically generated integer.
+    ///
+    /// Writing will fail if the value cannot be automatically generated.
+    Auto {
+        /// The `d_tag` field in the dynamic entry.
+        ///
+        /// One of the `DT_*` values.
+        tag: u32,
+    },
+    /// The value is an integer.
+    Integer {
+        /// The `d_tag` field in the dynamic entry.
+        ///
+        /// One of the `DT_*` values.
+        tag: u32,
+        /// The `d_val` field in the dynamic entry.
+        val: u64,
+    },
+    /// The value is a string.
+    String {
+        /// The `d_tag` field in the dynamic entry.
+        ///
+        /// One of the `DT_*` values.
+        tag: u32,
+        /// The string value.
+        ///
+        /// This will be stored in the dynamic string section.
+        val: ByteString<'data>,
+    },
+}
+
+impl<'data> Dynamic<'data> {
+    /// The `d_tag` field in the dynamic entry.
+    ///
+    /// One of the `DT_*` values.
+    pub fn tag(&self) -> u32 {
+        match self {
+            Dynamic::Auto { tag } => *tag,
+            Dynamic::Integer { tag, .. } => *tag,
+            Dynamic::String { tag, .. } => *tag,
+        }
+    }
+}
+
+/// An ID for referring to a filename in [`VersionFiles`].
+#[derive(Clone, Copy, PartialEq, Eq)]
+pub struct VersionFileId(usize);
+
+impl fmt::Debug for VersionFileId {
+    fn fmt(&self, f: &mut fmt::Formatter<'_>) -> fmt::Result {
+        write!(f, "VersionFileId({})", self.0)
+    }
+}
+
+impl Id for VersionFileId {
+    fn index(&self) -> usize {
+        self.0
+    }
+}
+
+impl IdPrivate for VersionFileId {
+    fn new(id: usize) -> Self {
+        VersionFileId(id)
+    }
+}
+
+/// A filename used for GNU versioning.
+///
+/// Stored in [`VersionFiles`].
+#[derive(Debug)]
+pub struct VersionFile<'data> {
+    id: VersionFileId,
+    /// Ignore this file when writing the ELF file.
+    pub delete: bool,
+    /// The filename.
+    pub name: ByteString<'data>,
+}
+
+impl<'data> Item for VersionFile<'data> {
+    type Id = VersionFileId;
+
+    fn is_deleted(&self) -> bool {
+        self.delete
+    }
+}
+
+impl<'data> VersionFile<'data> {
+    /// The ID used for referring to this filename.
+    pub fn id(&self) -> VersionFileId {
+        self.id
+    }
+}
+
+/// A table of filenames used for GNU versioning.
+pub type VersionFiles<'data> = Table<VersionFile<'data>>;
+
+impl<'data> VersionFiles<'data> {
+    /// Add a new filename to the table.
+    pub fn add(&mut self, name: ByteString<'data>) -> VersionFileId {
+        let id = self.next_id();
+        self.push(VersionFile {
+            id,
+            name,
+            delete: false,
+        });
+        id
+    }
+}
+
+const VERSION_ID_BASE: usize = 2;
+
+/// An ID for referring to a version in [`Versions`].
+#[derive(Clone, Copy, PartialEq, Eq)]
+pub struct VersionId(usize);
+
+impl fmt::Debug for VersionId {
+    fn fmt(&self, f: &mut fmt::Formatter<'_>) -> fmt::Result {
+        write!(f, "VersionId({})", self.0)
+    }
+}
+
+impl Id for VersionId {
+    fn index(&self) -> usize {
+        self.0 - VERSION_ID_BASE
+    }
+}
+
+impl IdPrivate for VersionId {
+    fn new(id: usize) -> Self {
+        VersionId(VERSION_ID_BASE + id)
+    }
+}
+
+impl VersionId {
+    /// Return `True` if this is a special version that does not exist in the version table.
+    pub fn is_special(&self) -> bool {
+        self.0 < VERSION_ID_BASE
+    }
+
+    /// Return the ID for a version index of [`elf::VER_NDX_LOCAL`].
+    pub fn local() -> Self {
+        VersionId(elf::VER_NDX_LOCAL as usize)
+    }
+
+    /// Return the ID for a version index of [`elf::VER_NDX_GLOBAL`].
+    pub fn global() -> Self {
+        VersionId(elf::VER_NDX_GLOBAL as usize)
+    }
+}
+
+/// A version for a symbol.
+#[derive(Debug)]
+pub struct Version<'data> {
+    id: VersionId,
+    /// The data for this version.
+    pub data: VersionData<'data>,
+    /// Ignore this version when writing the ELF file.
+    pub delete: bool,
+}
+
+impl<'data> Item for Version<'data> {
+    type Id = VersionId;
+
+    fn is_deleted(&self) -> bool {
+        self.delete
+    }
+}
+
+impl<'data> Version<'data> {
+    /// The ID used for referring to this version.
+    pub fn id(&self) -> VersionId {
+        self.id
+    }
+}
+
+/// The data for a version for a symbol.
+#[derive(Debug)]
+pub enum VersionData<'data> {
+    /// The version for a defined symbol.
+    Def(VersionDef<'data>),
+    /// The version for an undefined symbol.
+    Need(VersionNeed<'data>),
+}
+
+/// A GNU version definition.
+#[derive(Debug)]
+pub struct VersionDef<'data> {
+    /// The names for the version.
+    ///
+    /// This usually has two elements. The first element is the name of this
+    /// version, and the second element is the name of the previous version
+    /// in the tree of versions.
+    pub names: Vec<ByteString<'data>>,
+    /// The version flags.
+    ///
+    /// A combination of the `VER_FLG_*` constants.
+    pub flags: u16,
+}
+
+/// A GNU version dependency.
+#[derive(Debug)]
+pub struct VersionNeed<'data> {
+    /// The filename of the library providing this version.
+    pub file: VersionFileId,
+    /// The name of the version.
+    pub name: ByteString<'data>,
+    /// The version flags.
+    ///
+    /// A combination of the `VER_FLG_*` constants.
+    pub flags: u16,
+}
+
+/// A table of versions that are referenced by symbols.
+pub type Versions<'data> = Table<Version<'data>>;
+
+impl<'data> Versions<'data> {
+    /// Add a version.
+    pub fn add(&mut self, data: VersionData<'data>) -> VersionId {
+        let id = self.next_id();
+        self.push(Version {
+            id,
+            data,
+            delete: false,
+        });
+        id
+    }
+}
+
+/// The contents of an attributes section.
+#[derive(Debug, Default, Clone)]
+pub struct AttributesSection<'data> {
+    /// The subsections.
+    pub subsections: Vec<AttributesSubsection<'data>>,
+}
+
+impl<'data> AttributesSection<'data> {
+    /// Create a new attributes section.
+    pub fn new() -> Self {
+        Self::default()
+    }
+}
+
+/// A subsection of an attributes section.
+#[derive(Debug, Clone)]
+pub struct AttributesSubsection<'data> {
+    /// The vendor namespace for these attributes.
+    pub vendor: ByteString<'data>,
+    /// The sub-subsections.
+    pub subsubsections: Vec<AttributesSubsubsection<'data>>,
+}
+
+impl<'data> AttributesSubsection<'data> {
+    /// Create a new subsection.
+    pub fn new(vendor: ByteString<'data>) -> Self {
+        AttributesSubsection {
+            vendor,
+            subsubsections: Vec::new(),
+        }
+    }
+}
+
+/// A sub-subsection in an attributes section.
+#[derive(Debug, Clone)]
+pub struct AttributesSubsubsection<'data> {
+    /// The sub-subsection tag.
+    pub tag: AttributeTag,
+    /// The data containing the attributes.
+    pub data: Bytes<'data>,
+}
+
+/// The tag for a sub-subsection in an attributes section.
+#[derive(Debug, Clone)]
+pub enum AttributeTag {
+    /// The attributes apply to the whole file.
+    ///
+    /// Correspeonds to [`elf::Tag_File`].
+    File,
+    /// The attributes apply to the given sections.
+    ///
+    /// Correspeonds to [`elf::Tag_Section`].
+    Section(Vec<SectionId>),
+    /// The attributes apply to the given symbols.
+    ///
+    /// Correspeonds to [`elf::Tag_Symbol`].
+    Symbol(Vec<SymbolId>),
+}
+
+impl AttributeTag {
+    /// Return the corresponding `elf::Tag_*` value for this tag.
+    pub fn tag(&self) -> u8 {
+        match self {
+            AttributeTag::File => elf::Tag_File,
+            AttributeTag::Section(_) => elf::Tag_Section,
+            AttributeTag::Symbol(_) => elf::Tag_Symbol,
+        }
+    }
+}