const Elf = @This();

const std = @import("std");
const builtin = @import("builtin");
const mem = std.mem;
const assert = std.debug.assert;
const Allocator = std.mem.Allocator;
const fs = std.fs;
const elf = std.elf;
const log = std.log.scoped(.link);
const DW = std.dwarf;
const leb128 = std.leb;

const Module = @import("../Module.zig");
const Compilation = @import("../Compilation.zig");
const codegen = @import("../codegen.zig");
const trace = @import("../tracy.zig").trace;
const Package = @import("../Package.zig");
const Value = @import("../value.zig").Value;
const Type = @import("../type.zig").Type;
const link = @import("../link.zig");
const File = link.File;
const build_options = @import("build_options");
const target_util = @import("../target.zig");
const glibc = @import("../glibc.zig");
const musl = @import("../musl.zig");
const Cache = @import("../Cache.zig");
const Air = @import("../Air.zig");
const Liveness = @import("../Liveness.zig");
const LlvmObject = @import("../codegen/llvm.zig").Object;

const default_entry_addr = 0x8000000;

pub const base_tag: File.Tag = .elf;

base: File,

ptr_width: PtrWidth,

/// If this is not null, an object file is created by LLVM and linked with LLD afterwards.
llvm_object: ?*LlvmObject = null,

/// Stored in native-endian format, depending on target endianness needs to be bswapped on read/write.
/// Same order as in the file.
sections: std.ArrayListUnmanaged(elf.Elf64_Shdr) = std.ArrayListUnmanaged(elf.Elf64_Shdr){},
shdr_table_offset: ?u64 = null,

/// Stored in native-endian format, depending on target endianness needs to be bswapped on read/write.
/// Same order as in the file.
program_headers: std.ArrayListUnmanaged(elf.Elf64_Phdr) = std.ArrayListUnmanaged(elf.Elf64_Phdr){},
phdr_table_offset: ?u64 = null,
/// The index into the program headers of a PT_LOAD program header with Read and Execute flags
phdr_load_re_index: ?u16 = null,
/// The index into the program headers of the global offset table.
/// It needs PT_LOAD and Read flags.
phdr_got_index: ?u16 = null,
entry_addr: ?u64 = null,

debug_strtab: std.ArrayListUnmanaged(u8) = std.ArrayListUnmanaged(u8){},
shstrtab: std.ArrayListUnmanaged(u8) = std.ArrayListUnmanaged(u8){},
shstrtab_index: ?u16 = null,

text_section_index: ?u16 = null,
symtab_section_index: ?u16 = null,
got_section_index: ?u16 = null,
debug_info_section_index: ?u16 = null,
debug_abbrev_section_index: ?u16 = null,
debug_str_section_index: ?u16 = null,
debug_aranges_section_index: ?u16 = null,
debug_line_section_index: ?u16 = null,

debug_abbrev_table_offset: ?u64 = null,

/// The same order as in the file. ELF requires global symbols to all be after the
/// local symbols, they cannot be mixed. So we must buffer all the global symbols and
/// write them at the end. These are only the local symbols. The length of this array
/// is the value used for sh_info in the .symtab section.
local_symbols: std.ArrayListUnmanaged(elf.Elf64_Sym) = .{},
global_symbols: std.ArrayListUnmanaged(elf.Elf64_Sym) = .{},

local_symbol_free_list: std.ArrayListUnmanaged(u32) = .{},
global_symbol_free_list: std.ArrayListUnmanaged(u32) = .{},
offset_table_free_list: std.ArrayListUnmanaged(u32) = .{},

/// Same order as in the file. The value is the absolute vaddr value.
/// If the vaddr of the executable program header changes, the entire
/// offset table needs to be rewritten.
offset_table: std.ArrayListUnmanaged(u64) = .{},

phdr_table_dirty: bool = false,
shdr_table_dirty: bool = false,
shstrtab_dirty: bool = false,
debug_strtab_dirty: bool = false,
offset_table_count_dirty: bool = false,
debug_abbrev_section_dirty: bool = false,
debug_aranges_section_dirty: bool = false,

debug_info_header_dirty: bool = false,
debug_line_header_dirty: bool = false,

error_flags: File.ErrorFlags = File.ErrorFlags{},

/// A list of text blocks that have surplus capacity. This list can have false
/// positives, as functions grow and shrink over time, only sometimes being added
/// or removed from the freelist.
///
/// A text block has surplus capacity when its overcapacity value is greater than
/// padToIdeal(minimum_text_block_size). That is, when it has so
/// much extra capacity, that we could fit a small new symbol in it, itself with
/// ideal_capacity or more.
///
/// Ideal capacity is defined by size + (size / ideal_factor)
///
/// Overcapacity is measured by actual_capacity - ideal_capacity. Note that
/// overcapacity can be negative. A simple way to have negative overcapacity is to
/// allocate a fresh text block, which will have ideal capacity, and then grow it
/// by 1 byte. It will then have -1 overcapacity.
text_block_free_list: std.ArrayListUnmanaged(*TextBlock) = .{},
last_text_block: ?*TextBlock = null,

/// A list of `SrcFn` whose Line Number Programs have surplus capacity.
/// This is the same concept as `text_block_free_list`; see those doc comments.
dbg_line_fn_free_list: std.AutoHashMapUnmanaged(*SrcFn, void) = .{},
dbg_line_fn_first: ?*SrcFn = null,
dbg_line_fn_last: ?*SrcFn = null,

/// A list of `TextBlock` whose corresponding .debug_info tags have surplus capacity.
/// This is the same concept as `text_block_free_list`; see those doc comments.
dbg_info_decl_free_list: std.AutoHashMapUnmanaged(*TextBlock, void) = .{},
dbg_info_decl_first: ?*TextBlock = null,
dbg_info_decl_last: ?*TextBlock = null,

/// When allocating, the ideal_capacity is calculated by
/// actual_capacity + (actual_capacity / ideal_factor)
const ideal_factor = 3;

/// In order for a slice of bytes to be considered eligible to keep metadata pointing at
/// it as a possible place to put new symbols, it must have enough room for this many bytes
/// (plus extra for reserved capacity).
const minimum_text_block_size = 64;
const min_text_capacity = padToIdeal(minimum_text_block_size);

pub const PtrWidth = enum { p32, p64 };

pub const TextBlock = struct {
    /// Each decl always gets a local symbol with the fully qualified name.
    /// The vaddr and size are found here directly.
    /// The file offset is found by computing the vaddr offset from the section vaddr
    /// the symbol references, and adding that to the file offset of the section.
    /// If this field is 0, it means the codegen size = 0 and there is no symbol or
    /// offset table entry.
    local_sym_index: u32,
    /// This field is undefined for symbols with size = 0.
    offset_table_index: u32,
    /// Points to the previous and next neighbors, based on the `text_offset`.
    /// This can be used to find, for example, the capacity of this `TextBlock`.
    prev: ?*TextBlock,
    next: ?*TextBlock,

    /// Previous/next linked list pointers.
    /// This is the linked list node for this Decl's corresponding .debug_info tag.
    dbg_info_prev: ?*TextBlock,
    dbg_info_next: ?*TextBlock,
    /// Offset into .debug_info pointing to the tag for this Decl.
    dbg_info_off: u32,
    /// Size of the .debug_info tag for this Decl, not including padding.
    dbg_info_len: u32,

    pub const empty = TextBlock{
        .local_sym_index = 0,
        .offset_table_index = undefined,
        .prev = null,
        .next = null,
        .dbg_info_prev = null,
        .dbg_info_next = null,
        .dbg_info_off = undefined,
        .dbg_info_len = undefined,
    };

    /// Returns how much room there is to grow in virtual address space.
    /// File offset relocation happens transparently, so it is not included in
    /// this calculation.
    fn capacity(self: TextBlock, elf_file: Elf) u64 {
        const self_sym = elf_file.local_symbols.items[self.local_sym_index];
        if (self.next) |next| {
            const next_sym = elf_file.local_symbols.items[next.local_sym_index];
            return next_sym.st_value - self_sym.st_value;
        } else {
            // We are the last block. The capacity is limited only by virtual address space.
            return std.math.maxInt(u32) - self_sym.st_value;
        }
    }

    fn freeListEligible(self: TextBlock, elf_file: Elf) bool {
        // No need to keep a free list node for the last block.
        const next = self.next orelse return false;
        const self_sym = elf_file.local_symbols.items[self.local_sym_index];
        const next_sym = elf_file.local_symbols.items[next.local_sym_index];
        const cap = next_sym.st_value - self_sym.st_value;
        const ideal_cap = padToIdeal(self_sym.st_size);
        if (cap <= ideal_cap) return false;
        const surplus = cap - ideal_cap;
        return surplus >= min_text_capacity;
    }
};

pub const Export = struct {
    sym_index: ?u32 = null,
};

pub const SrcFn = struct {
    /// Offset from the beginning of the Debug Line Program header that contains this function.
    off: u32,
    /// Size of the line number program component belonging to this function, not
    /// including padding.
    len: u32,

    /// Points to the previous and next neighbors, based on the offset from .debug_line.
    /// This can be used to find, for example, the capacity of this `SrcFn`.
    prev: ?*SrcFn,
    next: ?*SrcFn,

    pub const empty: SrcFn = .{
        .off = 0,
        .len = 0,
        .prev = null,
        .next = null,
    };
};

pub fn openPath(allocator: *Allocator, sub_path: []const u8, options: link.Options) !*Elf {
    assert(options.object_format == .elf);

    if (build_options.have_llvm and options.use_llvm) {
        const self = try createEmpty(allocator, options);
        errdefer self.base.destroy();

        self.llvm_object = try LlvmObject.create(allocator, sub_path, options);
        return self;
    }

    const file = try options.emit.?.directory.handle.createFile(sub_path, .{
        .truncate = false,
        .read = true,
        .mode = link.determineMode(options),
    });
    errdefer file.close();

    const self = try createEmpty(allocator, options);
    errdefer self.base.destroy();

    self.base.file = file;
    self.shdr_table_dirty = true;

    // Index 0 is always a null symbol.
    try self.local_symbols.append(allocator, .{
        .st_name = 0,
        .st_info = 0,
        .st_other = 0,
        .st_shndx = 0,
        .st_value = 0,
        .st_size = 0,
    });

    // There must always be a null section in index 0
    try self.sections.append(allocator, .{
        .sh_name = 0,
        .sh_type = elf.SHT_NULL,
        .sh_flags = 0,
        .sh_addr = 0,
        .sh_offset = 0,
        .sh_size = 0,
        .sh_link = 0,
        .sh_info = 0,
        .sh_addralign = 0,
        .sh_entsize = 0,
    });

    try self.populateMissingMetadata();

    return self;
}

pub fn createEmpty(gpa: *Allocator, options: link.Options) !*Elf {
    const ptr_width: PtrWidth = switch (options.target.cpu.arch.ptrBitWidth()) {
        0...32 => .p32,
        33...64 => .p64,
        else => return error.UnsupportedELFArchitecture,
    };
    const self = try gpa.create(Elf);
    self.* = .{
        .base = .{
            .tag = .elf,
            .options = options,
            .allocator = gpa,
            .file = null,
        },
        .ptr_width = ptr_width,
    };
    return self;
}

pub fn deinit(self: *Elf) void {
    if (build_options.have_llvm) {
        if (self.llvm_object) |llvm_object| llvm_object.destroy(self.base.allocator);
    }

    self.sections.deinit(self.base.allocator);
    self.program_headers.deinit(self.base.allocator);
    self.shstrtab.deinit(self.base.allocator);
    self.debug_strtab.deinit(self.base.allocator);
    self.local_symbols.deinit(self.base.allocator);
    self.global_symbols.deinit(self.base.allocator);
    self.global_symbol_free_list.deinit(self.base.allocator);
    self.local_symbol_free_list.deinit(self.base.allocator);
    self.offset_table_free_list.deinit(self.base.allocator);
    self.text_block_free_list.deinit(self.base.allocator);
    self.dbg_line_fn_free_list.deinit(self.base.allocator);
    self.dbg_info_decl_free_list.deinit(self.base.allocator);
    self.offset_table.deinit(self.base.allocator);
}

pub fn getDeclVAddr(self: *Elf, decl: *const Module.Decl) u64 {
    assert(self.llvm_object == null);
    assert(decl.link.elf.local_sym_index != 0);
    return self.local_symbols.items[decl.link.elf.local_sym_index].st_value;
}

fn getDebugLineProgramOff(self: Elf) u32 {
    return self.dbg_line_fn_first.?.off;
}

fn getDebugLineProgramEnd(self: Elf) u32 {
    return self.dbg_line_fn_last.?.off + self.dbg_line_fn_last.?.len;
}

/// Returns end pos of collision, if any.
fn detectAllocCollision(self: *Elf, start: u64, size: u64) ?u64 {
    const small_ptr = self.ptr_width == .p32;
    const ehdr_size: u64 = if (small_ptr) @sizeOf(elf.Elf32_Ehdr) else @sizeOf(elf.Elf64_Ehdr);
    if (start < ehdr_size)
        return ehdr_size;

    const end = start + padToIdeal(size);

    if (self.shdr_table_offset) |off| {
        const shdr_size: u64 = if (small_ptr) @sizeOf(elf.Elf32_Shdr) else @sizeOf(elf.Elf64_Shdr);
        const tight_size = self.sections.items.len * shdr_size;
        const increased_size = padToIdeal(tight_size);
        const test_end = off + increased_size;
        if (end > off and start < test_end) {
            return test_end;
        }
    }

    if (self.phdr_table_offset) |off| {
        const phdr_size: u64 = if (small_ptr) @sizeOf(elf.Elf32_Phdr) else @sizeOf(elf.Elf64_Phdr);
        const tight_size = self.sections.items.len * phdr_size;
        const increased_size = padToIdeal(tight_size);
        const test_end = off + increased_size;
        if (end > off and start < test_end) {
            return test_end;
        }
    }

    for (self.sections.items) |section| {
        const increased_size = padToIdeal(section.sh_size);
        const test_end = section.sh_offset + increased_size;
        if (end > section.sh_offset and start < test_end) {
            return test_end;
        }
    }
    for (self.program_headers.items) |program_header| {
        const increased_size = padToIdeal(program_header.p_filesz);
        const test_end = program_header.p_offset + increased_size;
        if (end > program_header.p_offset and start < test_end) {
            return test_end;
        }
    }
    return null;
}

fn allocatedSize(self: *Elf, start: u64) u64 {
    if (start == 0)
        return 0;
    var min_pos: u64 = std.math.maxInt(u64);
    if (self.shdr_table_offset) |off| {
        if (off > start and off < min_pos) min_pos = off;
    }
    if (self.phdr_table_offset) |off| {
        if (off > start and off < min_pos) min_pos = off;
    }
    for (self.sections.items) |section| {
        if (section.sh_offset <= start) continue;
        if (section.sh_offset < min_pos) min_pos = section.sh_offset;
    }
    for (self.program_headers.items) |program_header| {
        if (program_header.p_offset <= start) continue;
        if (program_header.p_offset < min_pos) min_pos = program_header.p_offset;
    }
    return min_pos - start;
}

fn findFreeSpace(self: *Elf, object_size: u64, min_alignment: u16) u64 {
    var start: u64 = 0;
    while (self.detectAllocCollision(start, object_size)) |item_end| {
        start = mem.alignForwardGeneric(u64, item_end, min_alignment);
    }
    return start;
}

/// TODO Improve this to use a table.
fn makeString(self: *Elf, bytes: []const u8) !u32 {
    try self.shstrtab.ensureUnusedCapacity(self.base.allocator, bytes.len + 1);
    const result = self.shstrtab.items.len;
    self.shstrtab.appendSliceAssumeCapacity(bytes);
    self.shstrtab.appendAssumeCapacity(0);
    return @intCast(u32, result);
}

/// TODO Improve this to use a table.
fn makeDebugString(self: *Elf, bytes: []const u8) !u32 {
    try self.debug_strtab.ensureUnusedCapacity(self.base.allocator, bytes.len + 1);
    const result = self.debug_strtab.items.len;
    self.debug_strtab.appendSliceAssumeCapacity(bytes);
    self.debug_strtab.appendAssumeCapacity(0);
    return @intCast(u32, result);
}

fn getString(self: *Elf, str_off: u32) []const u8 {
    assert(str_off < self.shstrtab.items.len);
    return mem.sliceTo(@ptrCast([*:0]const u8, self.shstrtab.items.ptr + str_off), 0);
}

fn updateString(self: *Elf, old_str_off: u32, new_name: []const u8) !u32 {
    const existing_name = self.getString(old_str_off);
    if (mem.eql(u8, existing_name, new_name)) {
        return old_str_off;
    }
    return self.makeString(new_name);
}

pub fn populateMissingMetadata(self: *Elf) !void {
    assert(self.llvm_object == null);

    const small_ptr = switch (self.ptr_width) {
        .p32 => true,
        .p64 => false,
    };
    const ptr_size: u8 = self.ptrWidthBytes();
    if (self.phdr_load_re_index == null) {
        self.phdr_load_re_index = @intCast(u16, self.program_headers.items.len);
        const file_size = self.base.options.program_code_size_hint;
        const p_align = 0x1000;
        const off = self.findFreeSpace(file_size, p_align);
        log.debug("found PT_LOAD free space 0x{x} to 0x{x}", .{ off, off + file_size });
        const entry_addr: u64 = self.entry_addr orelse if (self.base.options.target.cpu.arch == .spu_2) @as(u64, 0) else default_entry_addr;
        try self.program_headers.append(self.base.allocator, .{
            .p_type = elf.PT_LOAD,
            .p_offset = off,
            .p_filesz = file_size,
            .p_vaddr = entry_addr,
            .p_paddr = entry_addr,
            .p_memsz = file_size,
            .p_align = p_align,
            .p_flags = elf.PF_X | elf.PF_R,
        });
        self.entry_addr = null;
        self.phdr_table_dirty = true;
    }
    if (self.phdr_got_index == null) {
        self.phdr_got_index = @intCast(u16, self.program_headers.items.len);
        const file_size = @as(u64, ptr_size) * self.base.options.symbol_count_hint;
        // We really only need ptr alignment but since we are using PROGBITS, linux requires
        // page align.
        const p_align = if (self.base.options.target.os.tag == .linux) 0x1000 else @as(u16, ptr_size);
        const off = self.findFreeSpace(file_size, p_align);
        log.debug("found PT_LOAD free space 0x{x} to 0x{x}", .{ off, off + file_size });
        // TODO instead of hard coding the vaddr, make a function to find a vaddr to put things at.
        // we'll need to re-use that function anyway, in case the GOT grows and overlaps something
        // else in virtual memory.
        const got_addr: u32 = if (self.base.options.target.cpu.arch.ptrBitWidth() >= 32) 0x4000000 else 0x8000;
        try self.program_headers.append(self.base.allocator, .{
            .p_type = elf.PT_LOAD,
            .p_offset = off,
            .p_filesz = file_size,
            .p_vaddr = got_addr,
            .p_paddr = got_addr,
            .p_memsz = file_size,
            .p_align = p_align,
            .p_flags = elf.PF_R,
        });
        self.phdr_table_dirty = true;
    }
    if (self.shstrtab_index == null) {
        self.shstrtab_index = @intCast(u16, self.sections.items.len);
        assert(self.shstrtab.items.len == 0);
        try self.shstrtab.append(self.base.allocator, 0); // need a 0 at position 0
        const off = self.findFreeSpace(self.shstrtab.items.len, 1);
        log.debug("found shstrtab free space 0x{x} to 0x{x}", .{ off, off + self.shstrtab.items.len });
        try self.sections.append(self.base.allocator, .{
            .sh_name = try self.makeString(".shstrtab"),
            .sh_type = elf.SHT_STRTAB,
            .sh_flags = 0,
            .sh_addr = 0,
            .sh_offset = off,
            .sh_size = self.shstrtab.items.len,
            .sh_link = 0,
            .sh_info = 0,
            .sh_addralign = 1,
            .sh_entsize = 0,
        });
        self.shstrtab_dirty = true;
        self.shdr_table_dirty = true;
    }
    if (self.text_section_index == null) {
        self.text_section_index = @intCast(u16, self.sections.items.len);
        const phdr = &self.program_headers.items[self.phdr_load_re_index.?];

        try self.sections.append(self.base.allocator, .{
            .sh_name = try self.makeString(".text"),
            .sh_type = elf.SHT_PROGBITS,
            .sh_flags = elf.SHF_ALLOC | elf.SHF_EXECINSTR,
            .sh_addr = phdr.p_vaddr,
            .sh_offset = phdr.p_offset,
            .sh_size = phdr.p_filesz,
            .sh_link = 0,
            .sh_info = 0,
            .sh_addralign = phdr.p_align,
            .sh_entsize = 0,
        });
        self.shdr_table_dirty = true;
    }
    if (self.got_section_index == null) {
        self.got_section_index = @intCast(u16, self.sections.items.len);
        const phdr = &self.program_headers.items[self.phdr_got_index.?];

        try self.sections.append(self.base.allocator, .{
            .sh_name = try self.makeString(".got"),
            .sh_type = elf.SHT_PROGBITS,
            .sh_flags = elf.SHF_ALLOC,
            .sh_addr = phdr.p_vaddr,
            .sh_offset = phdr.p_offset,
            .sh_size = phdr.p_filesz,
            .sh_link = 0,
            .sh_info = 0,
            .sh_addralign = phdr.p_align,
            .sh_entsize = 0,
        });
        self.shdr_table_dirty = true;
    }
    if (self.symtab_section_index == null) {
        self.symtab_section_index = @intCast(u16, self.sections.items.len);
        const min_align: u16 = if (small_ptr) @alignOf(elf.Elf32_Sym) else @alignOf(elf.Elf64_Sym);
        const each_size: u64 = if (small_ptr) @sizeOf(elf.Elf32_Sym) else @sizeOf(elf.Elf64_Sym);
        const file_size = self.base.options.symbol_count_hint * each_size;
        const off = self.findFreeSpace(file_size, min_align);
        log.debug("found symtab free space 0x{x} to 0x{x}", .{ off, off + file_size });

        try self.sections.append(self.base.allocator, .{
            .sh_name = try self.makeString(".symtab"),
            .sh_type = elf.SHT_SYMTAB,
            .sh_flags = 0,
            .sh_addr = 0,
            .sh_offset = off,
            .sh_size = file_size,
            // The section header index of the associated string table.
            .sh_link = self.shstrtab_index.?,
            .sh_info = @intCast(u32, self.local_symbols.items.len),
            .sh_addralign = min_align,
            .sh_entsize = each_size,
        });
        self.shdr_table_dirty = true;
        try self.writeSymbol(0);
    }
    if (self.debug_str_section_index == null) {
        self.debug_str_section_index = @intCast(u16, self.sections.items.len);
        assert(self.debug_strtab.items.len == 0);
        try self.sections.append(self.base.allocator, .{
            .sh_name = try self.makeString(".debug_str"),
            .sh_type = elf.SHT_PROGBITS,
            .sh_flags = elf.SHF_MERGE | elf.SHF_STRINGS,
            .sh_addr = 0,
            .sh_offset = 0,
            .sh_size = self.debug_strtab.items.len,
            .sh_link = 0,
            .sh_info = 0,
            .sh_addralign = 1,
            .sh_entsize = 1,
        });
        self.debug_strtab_dirty = true;
        self.shdr_table_dirty = true;
    }
    if (self.debug_info_section_index == null) {
        self.debug_info_section_index = @intCast(u16, self.sections.items.len);

        const file_size_hint = 200;
        const p_align = 1;
        const off = self.findFreeSpace(file_size_hint, p_align);
        log.debug("found .debug_info free space 0x{x} to 0x{x}", .{
            off,
            off + file_size_hint,
        });
        try self.sections.append(self.base.allocator, .{
            .sh_name = try self.makeString(".debug_info"),
            .sh_type = elf.SHT_PROGBITS,
            .sh_flags = 0,
            .sh_addr = 0,
            .sh_offset = off,
            .sh_size = file_size_hint,
            .sh_link = 0,
            .sh_info = 0,
            .sh_addralign = p_align,
            .sh_entsize = 0,
        });
        self.shdr_table_dirty = true;
        self.debug_info_header_dirty = true;
    }
    if (self.debug_abbrev_section_index == null) {
        self.debug_abbrev_section_index = @intCast(u16, self.sections.items.len);

        const file_size_hint = 128;
        const p_align = 1;
        const off = self.findFreeSpace(file_size_hint, p_align);
        log.debug("found .debug_abbrev free space 0x{x} to 0x{x}", .{
            off,
            off + file_size_hint,
        });
        try self.sections.append(self.base.allocator, .{
            .sh_name = try self.makeString(".debug_abbrev"),
            .sh_type = elf.SHT_PROGBITS,
            .sh_flags = 0,
            .sh_addr = 0,
            .sh_offset = off,
            .sh_size = file_size_hint,
            .sh_link = 0,
            .sh_info = 0,
            .sh_addralign = p_align,
            .sh_entsize = 0,
        });
        self.shdr_table_dirty = true;
        self.debug_abbrev_section_dirty = true;
    }
    if (self.debug_aranges_section_index == null) {
        self.debug_aranges_section_index = @intCast(u16, self.sections.items.len);

        const file_size_hint = 160;
        const p_align = 16;
        const off = self.findFreeSpace(file_size_hint, p_align);
        log.debug("found .debug_aranges free space 0x{x} to 0x{x}", .{
            off,
            off + file_size_hint,
        });
        try self.sections.append(self.base.allocator, .{
            .sh_name = try self.makeString(".debug_aranges"),
            .sh_type = elf.SHT_PROGBITS,
            .sh_flags = 0,
            .sh_addr = 0,
            .sh_offset = off,
            .sh_size = file_size_hint,
            .sh_link = 0,
            .sh_info = 0,
            .sh_addralign = p_align,
            .sh_entsize = 0,
        });
        self.shdr_table_dirty = true;
        self.debug_aranges_section_dirty = true;
    }
    if (self.debug_line_section_index == null) {
        self.debug_line_section_index = @intCast(u16, self.sections.items.len);

        const file_size_hint = 250;
        const p_align = 1;
        const off = self.findFreeSpace(file_size_hint, p_align);
        log.debug("found .debug_line free space 0x{x} to 0x{x}", .{
            off,
            off + file_size_hint,
        });
        try self.sections.append(self.base.allocator, .{
            .sh_name = try self.makeString(".debug_line"),
            .sh_type = elf.SHT_PROGBITS,
            .sh_flags = 0,
            .sh_addr = 0,
            .sh_offset = off,
            .sh_size = file_size_hint,
            .sh_link = 0,
            .sh_info = 0,
            .sh_addralign = p_align,
            .sh_entsize = 0,
        });
        self.shdr_table_dirty = true;
        self.debug_line_header_dirty = true;
    }
    const shsize: u64 = switch (self.ptr_width) {
        .p32 => @sizeOf(elf.Elf32_Shdr),
        .p64 => @sizeOf(elf.Elf64_Shdr),
    };
    const shalign: u16 = switch (self.ptr_width) {
        .p32 => @alignOf(elf.Elf32_Shdr),
        .p64 => @alignOf(elf.Elf64_Shdr),
    };
    if (self.shdr_table_offset == null) {
        self.shdr_table_offset = self.findFreeSpace(self.sections.items.len * shsize, shalign);
        self.shdr_table_dirty = true;
    }
    const phsize: u64 = switch (self.ptr_width) {
        .p32 => @sizeOf(elf.Elf32_Phdr),
        .p64 => @sizeOf(elf.Elf64_Phdr),
    };
    const phalign: u16 = switch (self.ptr_width) {
        .p32 => @alignOf(elf.Elf32_Phdr),
        .p64 => @alignOf(elf.Elf64_Phdr),
    };
    if (self.phdr_table_offset == null) {
        self.phdr_table_offset = self.findFreeSpace(self.program_headers.items.len * phsize, phalign);
        self.phdr_table_dirty = true;
    }
    {
        // Iterate over symbols, populating free_list and last_text_block.
        if (self.local_symbols.items.len != 1) {
            @panic("TODO implement setting up free_list and last_text_block from existing ELF file");
        }
        // We are starting with an empty file. The default values are correct, null and empty list.
    }
}

pub const abbrev_compile_unit = 1;
pub const abbrev_subprogram = 2;
pub const abbrev_subprogram_retvoid = 3;
pub const abbrev_base_type = 4;
pub const abbrev_pad1 = 5;
pub const abbrev_parameter = 6;

pub fn flush(self: *Elf, comp: *Compilation) !void {
    if (build_options.have_llvm and self.base.options.use_lld) {
        return self.linkWithLLD(comp);
    } else {
        switch (self.base.options.effectiveOutputMode()) {
            .Exe, .Obj => {},
            .Lib => return error.TODOImplementWritingLibFiles,
        }
        return self.flushModule(comp);
    }
}

pub fn flushModule(self: *Elf, comp: *Compilation) !void {
    const tracy = trace(@src());
    defer tracy.end();

    if (build_options.have_llvm)
        if (self.llvm_object) |llvm_object| return try llvm_object.flushModule(comp);

    // TODO This linker code currently assumes there is only 1 compilation unit and it
    // corresponds to the Zig source code.
    const module = self.base.options.module orelse return error.LinkingWithoutZigSourceUnimplemented;

    const target_endian = self.base.options.target.cpu.arch.endian();
    const foreign_endian = target_endian != builtin.cpu.arch.endian();
    const ptr_width_bytes: u8 = self.ptrWidthBytes();
    const init_len_size: usize = switch (self.ptr_width) {
        .p32 => 4,
        .p64 => 12,
    };

    // Unfortunately these have to be buffered and done at the end because ELF does not allow
    // mixing local and global symbols within a symbol table.
    try self.writeAllGlobalSymbols();

    if (self.debug_abbrev_section_dirty) {
        const debug_abbrev_sect = &self.sections.items[self.debug_abbrev_section_index.?];

        // These are LEB encoded but since the values are all less than 127
        // we can simply append these bytes.
        const abbrev_buf = [_]u8{
            abbrev_compile_unit, DW.TAG.compile_unit, DW.CHILDREN.yes, // header
            DW.AT.stmt_list,     DW.FORM.sec_offset,  DW.AT.low_pc,
            DW.FORM.addr,        DW.AT.high_pc,       DW.FORM.addr,
            DW.AT.name,          DW.FORM.strp,        DW.AT.comp_dir,
            DW.FORM.strp,        DW.AT.producer,      DW.FORM.strp,
            DW.AT.language,      DW.FORM.data2,       0,
            0, // table sentinel
            abbrev_subprogram,
            DW.TAG.subprogram,
            DW.CHILDREN.yes, // header
            DW.AT.low_pc,
            DW.FORM.addr,
            DW.AT.high_pc,
            DW.FORM.data4,
            DW.AT.type,
            DW.FORM.ref4,
            DW.AT.name,
            DW.FORM.string,
            0,                         0, // table sentinel
            abbrev_subprogram_retvoid,
            DW.TAG.subprogram, DW.CHILDREN.yes, // header
            DW.AT.low_pc,      DW.FORM.addr,
            DW.AT.high_pc,     DW.FORM.data4,
            DW.AT.name,        DW.FORM.string,
            0,
            0, // table sentinel
            abbrev_base_type,
            DW.TAG.base_type,
            DW.CHILDREN.no, // header
            DW.AT.encoding,
            DW.FORM.data1,
            DW.AT.byte_size,
            DW.FORM.data1,
            DW.AT.name,
            DW.FORM.string, 0, 0, // table sentinel
            abbrev_pad1, DW.TAG.unspecified_type, DW.CHILDREN.no, // header
            0,                0, // table sentinel
            abbrev_parameter,
            DW.TAG.formal_parameter, DW.CHILDREN.no, // header
            DW.AT.location,          DW.FORM.exprloc,
            DW.AT.type,              DW.FORM.ref4,
            DW.AT.name,              DW.FORM.string,
            0,
            0, // table sentinel
            0,
            0,
            0, // section sentinel
        };

        const needed_size = abbrev_buf.len;
        const allocated_size = self.allocatedSize(debug_abbrev_sect.sh_offset);
        if (needed_size > allocated_size) {
            debug_abbrev_sect.sh_size = 0; // free the space
            debug_abbrev_sect.sh_offset = self.findFreeSpace(needed_size, 1);
        }
        debug_abbrev_sect.sh_size = needed_size;
        log.debug(".debug_abbrev start=0x{x} end=0x{x}", .{
            debug_abbrev_sect.sh_offset,
            debug_abbrev_sect.sh_offset + needed_size,
        });

        const abbrev_offset = 0;
        self.debug_abbrev_table_offset = abbrev_offset;
        try self.base.file.?.pwriteAll(&abbrev_buf, debug_abbrev_sect.sh_offset + abbrev_offset);
        if (!self.shdr_table_dirty) {
            // Then it won't get written with the others and we need to do it.
            try self.writeSectHeader(self.debug_abbrev_section_index.?);
        }

        self.debug_abbrev_section_dirty = false;
    }

    if (self.debug_info_header_dirty) debug_info: {
        // If this value is null it means there is an error in the module;
        // leave debug_info_header_dirty=true.
        const first_dbg_info_decl = self.dbg_info_decl_first orelse break :debug_info;
        const last_dbg_info_decl = self.dbg_info_decl_last.?;
        const debug_info_sect = &self.sections.items[self.debug_info_section_index.?];

        // We have a function to compute the upper bound size, because it's needed
        // for determining where to put the offset of the first `LinkBlock`.
        const needed_bytes = self.dbgInfoNeededHeaderBytes();
        var di_buf = try std.ArrayList(u8).initCapacity(self.base.allocator, needed_bytes);
        defer di_buf.deinit();

        // initial length - length of the .debug_info contribution for this compilation unit,
        // not including the initial length itself.
        // We have to come back and write it later after we know the size.
        const after_init_len = di_buf.items.len + init_len_size;
        // +1 for the final 0 that ends the compilation unit children.
        const dbg_info_end = last_dbg_info_decl.dbg_info_off + last_dbg_info_decl.dbg_info_len + 1;
        const init_len = dbg_info_end - after_init_len;
        switch (self.ptr_width) {
            .p32 => {
                mem.writeInt(u32, di_buf.addManyAsArrayAssumeCapacity(4), @intCast(u32, init_len), target_endian);
            },
            .p64 => {
                di_buf.appendNTimesAssumeCapacity(0xff, 4);
                mem.writeInt(u64, di_buf.addManyAsArrayAssumeCapacity(8), init_len, target_endian);
            },
        }
        mem.writeInt(u16, di_buf.addManyAsArrayAssumeCapacity(2), 4, target_endian); // DWARF version
        const abbrev_offset = self.debug_abbrev_table_offset.?;
        switch (self.ptr_width) {
            .p32 => {
                mem.writeInt(u32, di_buf.addManyAsArrayAssumeCapacity(4), @intCast(u32, abbrev_offset), target_endian);
                di_buf.appendAssumeCapacity(4); // address size
            },
            .p64 => {
                mem.writeInt(u64, di_buf.addManyAsArrayAssumeCapacity(8), abbrev_offset, target_endian);
                di_buf.appendAssumeCapacity(8); // address size
            },
        }
        // Write the form for the compile unit, which must match the abbrev table above.
        const name_strp = try self.makeDebugString(module.root_pkg.root_src_path);
        const comp_dir_strp = try self.makeDebugString(module.root_pkg.root_src_directory.path orelse ".");
        const producer_strp = try self.makeDebugString(link.producer_string);
        // Currently only one compilation unit is supported, so the address range is simply
        // identical to the main program header virtual address and memory size.
        const text_phdr = &self.program_headers.items[self.phdr_load_re_index.?];
        const low_pc = text_phdr.p_vaddr;
        const high_pc = text_phdr.p_vaddr + text_phdr.p_memsz;

        di_buf.appendAssumeCapacity(abbrev_compile_unit);
        self.writeDwarfAddrAssumeCapacity(&di_buf, 0); // DW.AT.stmt_list, DW.FORM.sec_offset
        self.writeDwarfAddrAssumeCapacity(&di_buf, low_pc);
        self.writeDwarfAddrAssumeCapacity(&di_buf, high_pc);
        self.writeDwarfAddrAssumeCapacity(&di_buf, name_strp);
        self.writeDwarfAddrAssumeCapacity(&di_buf, comp_dir_strp);
        self.writeDwarfAddrAssumeCapacity(&di_buf, producer_strp);
        // We are still waiting on dwarf-std.org to assign DW_LANG_Zig a number:
        // http://dwarfstd.org/ShowIssue.php?issue=171115.1
        // Until then we say it is C99.
        mem.writeInt(u16, di_buf.addManyAsArrayAssumeCapacity(2), DW.LANG.C99, target_endian);

        if (di_buf.items.len > first_dbg_info_decl.dbg_info_off) {
            // Move the first N decls to the end to make more padding for the header.
            @panic("TODO: handle .debug_info header exceeding its padding");
        }
        const jmp_amt = first_dbg_info_decl.dbg_info_off - di_buf.items.len;
        try self.pwriteDbgInfoNops(0, di_buf.items, jmp_amt, false, debug_info_sect.sh_offset);
        self.debug_info_header_dirty = false;
    }

    if (self.debug_aranges_section_dirty) {
        const debug_aranges_sect = &self.sections.items[self.debug_aranges_section_index.?];

        // Enough for all the data without resizing. When support for more compilation units
        // is added, the size of this section will become more variable.
        var di_buf = try std.ArrayList(u8).initCapacity(self.base.allocator, 100);
        defer di_buf.deinit();

        // initial length - length of the .debug_aranges contribution for this compilation unit,
        // not including the initial length itself.
        // We have to come back and write it later after we know the size.
        const init_len_index = di_buf.items.len;
        di_buf.items.len += init_len_size;
        const after_init_len = di_buf.items.len;
        mem.writeInt(u16, di_buf.addManyAsArrayAssumeCapacity(2), 2, target_endian); // version
        // When more than one compilation unit is supported, this will be the offset to it.
        // For now it is always at offset 0 in .debug_info.
        self.writeDwarfAddrAssumeCapacity(&di_buf, 0); // .debug_info offset
        di_buf.appendAssumeCapacity(ptr_width_bytes); // address_size
        di_buf.appendAssumeCapacity(0); // segment_selector_size

        const end_header_offset = di_buf.items.len;
        const begin_entries_offset = mem.alignForward(end_header_offset, ptr_width_bytes * 2);
        di_buf.appendNTimesAssumeCapacity(0, begin_entries_offset - end_header_offset);

        // Currently only one compilation unit is supported, so the address range is simply
        // identical to the main program header virtual address and memory size.
        const text_phdr = &self.program_headers.items[self.phdr_load_re_index.?];
        self.writeDwarfAddrAssumeCapacity(&di_buf, text_phdr.p_vaddr);
        self.writeDwarfAddrAssumeCapacity(&di_buf, text_phdr.p_memsz);

        // Sentinel.
        self.writeDwarfAddrAssumeCapacity(&di_buf, 0);
        self.writeDwarfAddrAssumeCapacity(&di_buf, 0);

        // Go back and populate the initial length.
        const init_len = di_buf.items.len - after_init_len;
        switch (self.ptr_width) {
            .p32 => {
                mem.writeInt(u32, di_buf.items[init_len_index..][0..4], @intCast(u32, init_len), target_endian);
            },
            .p64 => {
                // initial length - length of the .debug_aranges contribution for this compilation unit,
                // not including the initial length itself.
                di_buf.items[init_len_index..][0..4].* = [_]u8{ 0xff, 0xff, 0xff, 0xff };
                mem.writeInt(u64, di_buf.items[init_len_index + 4 ..][0..8], init_len, target_endian);
            },
        }

        const needed_size = di_buf.items.len;
        const allocated_size = self.allocatedSize(debug_aranges_sect.sh_offset);
        if (needed_size > allocated_size) {
            debug_aranges_sect.sh_size = 0; // free the space
            debug_aranges_sect.sh_offset = self.findFreeSpace(needed_size, 16);
        }
        debug_aranges_sect.sh_size = needed_size;
        log.debug(".debug_aranges start=0x{x} end=0x{x}", .{
            debug_aranges_sect.sh_offset,
            debug_aranges_sect.sh_offset + needed_size,
        });

        try self.base.file.?.pwriteAll(di_buf.items, debug_aranges_sect.sh_offset);
        if (!self.shdr_table_dirty) {
            // Then it won't get written with the others and we need to do it.
            try self.writeSectHeader(self.debug_aranges_section_index.?);
        }

        self.debug_aranges_section_dirty = false;
    }
    if (self.debug_line_header_dirty) debug_line: {
        if (self.dbg_line_fn_first == null) {
            break :debug_line; // Error in module; leave debug_line_header_dirty=true.
        }
        const dbg_line_prg_off = self.getDebugLineProgramOff();
        const dbg_line_prg_end = self.getDebugLineProgramEnd();
        assert(dbg_line_prg_end != 0);

        const debug_line_sect = &self.sections.items[self.debug_line_section_index.?];

        // The size of this header is variable, depending on the number of directories,
        // files, and padding. We have a function to compute the upper bound size, however,
        // because it's needed for determining where to put the offset of the first `SrcFn`.
        const needed_bytes = self.dbgLineNeededHeaderBytes();
        var di_buf = try std.ArrayList(u8).initCapacity(self.base.allocator, needed_bytes);
        defer di_buf.deinit();

        // initial length - length of the .debug_line contribution for this compilation unit,
        // not including the initial length itself.
        const after_init_len = di_buf.items.len + init_len_size;
        const init_len = dbg_line_prg_end - after_init_len;
        switch (self.ptr_width) {
            .p32 => {
                mem.writeInt(u32, di_buf.addManyAsArrayAssumeCapacity(4), @intCast(u32, init_len), target_endian);
            },
            .p64 => {
                di_buf.appendNTimesAssumeCapacity(0xff, 4);
                mem.writeInt(u64, di_buf.addManyAsArrayAssumeCapacity(8), init_len, target_endian);
            },
        }

        mem.writeInt(u16, di_buf.addManyAsArrayAssumeCapacity(2), 4, target_endian); // version

        // Empirically, debug info consumers do not respect this field, or otherwise
        // consider it to be an error when it does not point exactly to the end of the header.
        // Therefore we rely on the NOP jump at the beginning of the Line Number Program for
        // padding rather than this field.
        const before_header_len = di_buf.items.len;
        di_buf.items.len += ptr_width_bytes; // We will come back and write this.
        const after_header_len = di_buf.items.len;

        const opcode_base = DW.LNS.set_isa + 1;
        di_buf.appendSliceAssumeCapacity(&[_]u8{
            1, // minimum_instruction_length
            1, // maximum_operations_per_instruction
            1, // default_is_stmt
            1, // line_base (signed)
            1, // line_range
            opcode_base,

            // Standard opcode lengths. The number of items here is based on `opcode_base`.
            // The value is the number of LEB128 operands the instruction takes.
            0, // `DW.LNS.copy`
            1, // `DW.LNS.advance_pc`
            1, // `DW.LNS.advance_line`
            1, // `DW.LNS.set_file`
            1, // `DW.LNS.set_column`
            0, // `DW.LNS.negate_stmt`
            0, // `DW.LNS.set_basic_block`
            0, // `DW.LNS.const_add_pc`
            1, // `DW.LNS.fixed_advance_pc`
            0, // `DW.LNS.set_prologue_end`
            0, // `DW.LNS.set_epilogue_begin`
            1, // `DW.LNS.set_isa`
            0, // include_directories (none except the compilation unit cwd)
        });
        // file_names[0]
        di_buf.appendSliceAssumeCapacity(module.root_pkg.root_src_path); // relative path name
        di_buf.appendSliceAssumeCapacity(&[_]u8{
            0, // null byte for the relative path name
            0, // directory_index
            0, // mtime (TODO supply this)
            0, // file size bytes (TODO supply this)
            0, // file_names sentinel
        });

        const header_len = di_buf.items.len - after_header_len;
        switch (self.ptr_width) {
            .p32 => {
                mem.writeInt(u32, di_buf.items[before_header_len..][0..4], @intCast(u32, header_len), target_endian);
            },
            .p64 => {
                mem.writeInt(u64, di_buf.items[before_header_len..][0..8], header_len, target_endian);
            },
        }

        // We use NOPs because consumers empirically do not respect the header length field.
        if (di_buf.items.len > dbg_line_prg_off) {
            // Move the first N files to the end to make more padding for the header.
            @panic("TODO: handle .debug_line header exceeding its padding");
        }
        const jmp_amt = dbg_line_prg_off - di_buf.items.len;
        try self.pwriteDbgLineNops(0, di_buf.items, jmp_amt, debug_line_sect.sh_offset);
        self.debug_line_header_dirty = false;
    }

    if (self.phdr_table_dirty) {
        const phsize: u64 = switch (self.ptr_width) {
            .p32 => @sizeOf(elf.Elf32_Phdr),
            .p64 => @sizeOf(elf.Elf64_Phdr),
        };
        const phalign: u16 = switch (self.ptr_width) {
            .p32 => @alignOf(elf.Elf32_Phdr),
            .p64 => @alignOf(elf.Elf64_Phdr),
        };
        const allocated_size = self.allocatedSize(self.phdr_table_offset.?);
        const needed_size = self.program_headers.items.len * phsize;

        if (needed_size > allocated_size) {
            self.phdr_table_offset = null; // free the space
            self.phdr_table_offset = self.findFreeSpace(needed_size, phalign);
        }

        switch (self.ptr_width) {
            .p32 => {
                const buf = try self.base.allocator.alloc(elf.Elf32_Phdr, self.program_headers.items.len);
                defer self.base.allocator.free(buf);

                for (buf) |*phdr, i| {
                    phdr.* = progHeaderTo32(self.program_headers.items[i]);
                    if (foreign_endian) {
                        mem.bswapAllFields(elf.Elf32_Phdr, phdr);
                    }
                }
                try self.base.file.?.pwriteAll(mem.sliceAsBytes(buf), self.phdr_table_offset.?);
            },
            .p64 => {
                const buf = try self.base.allocator.alloc(elf.Elf64_Phdr, self.program_headers.items.len);
                defer self.base.allocator.free(buf);

                for (buf) |*phdr, i| {
                    phdr.* = self.program_headers.items[i];
                    if (foreign_endian) {
                        mem.bswapAllFields(elf.Elf64_Phdr, phdr);
                    }
                }
                try self.base.file.?.pwriteAll(mem.sliceAsBytes(buf), self.phdr_table_offset.?);
            },
        }
        self.phdr_table_dirty = false;
    }

    {
        const shstrtab_sect = &self.sections.items[self.shstrtab_index.?];
        if (self.shstrtab_dirty or self.shstrtab.items.len != shstrtab_sect.sh_size) {
            const allocated_size = self.allocatedSize(shstrtab_sect.sh_offset);
            const needed_size = self.shstrtab.items.len;

            if (needed_size > allocated_size) {
                shstrtab_sect.sh_size = 0; // free the space
                shstrtab_sect.sh_offset = self.findFreeSpace(needed_size, 1);
            }
            shstrtab_sect.sh_size = needed_size;
            log.debug("writing shstrtab start=0x{x} end=0x{x}", .{ shstrtab_sect.sh_offset, shstrtab_sect.sh_offset + needed_size });

            try self.base.file.?.pwriteAll(self.shstrtab.items, shstrtab_sect.sh_offset);
            if (!self.shdr_table_dirty) {
                // Then it won't get written with the others and we need to do it.
                try self.writeSectHeader(self.shstrtab_index.?);
            }
            self.shstrtab_dirty = false;
        }
    }
    {
        const debug_strtab_sect = &self.sections.items[self.debug_str_section_index.?];
        if (self.debug_strtab_dirty or self.debug_strtab.items.len != debug_strtab_sect.sh_size) {
            const allocated_size = self.allocatedSize(debug_strtab_sect.sh_offset);
            const needed_size = self.debug_strtab.items.len;

            if (needed_size > allocated_size) {
                debug_strtab_sect.sh_size = 0; // free the space
                debug_strtab_sect.sh_offset = self.findFreeSpace(needed_size, 1);
            }
            debug_strtab_sect.sh_size = needed_size;
            log.debug("debug_strtab start=0x{x} end=0x{x}", .{ debug_strtab_sect.sh_offset, debug_strtab_sect.sh_offset + needed_size });

            try self.base.file.?.pwriteAll(self.debug_strtab.items, debug_strtab_sect.sh_offset);
            if (!self.shdr_table_dirty) {
                // Then it won't get written with the others and we need to do it.
                try self.writeSectHeader(self.debug_str_section_index.?);
            }
            self.debug_strtab_dirty = false;
        }
    }
    if (self.shdr_table_dirty) {
        const shsize: u64 = switch (self.ptr_width) {
            .p32 => @sizeOf(elf.Elf32_Shdr),
            .p64 => @sizeOf(elf.Elf64_Shdr),
        };
        const shalign: u16 = switch (self.ptr_width) {
            .p32 => @alignOf(elf.Elf32_Shdr),
            .p64 => @alignOf(elf.Elf64_Shdr),
        };
        const allocated_size = self.allocatedSize(self.shdr_table_offset.?);
        const needed_size = self.sections.items.len * shsize;

        if (needed_size > allocated_size) {
            self.shdr_table_offset = null; // free the space
            self.shdr_table_offset = self.findFreeSpace(needed_size, shalign);
        }

        switch (self.ptr_width) {
            .p32 => {
                const buf = try self.base.allocator.alloc(elf.Elf32_Shdr, self.sections.items.len);
                defer self.base.allocator.free(buf);

                for (buf) |*shdr, i| {
                    shdr.* = sectHeaderTo32(self.sections.items[i]);
                    log.debug("writing section {}", .{shdr.*});
                    if (foreign_endian) {
                        mem.bswapAllFields(elf.Elf32_Shdr, shdr);
                    }
                }
                try self.base.file.?.pwriteAll(mem.sliceAsBytes(buf), self.shdr_table_offset.?);
            },
            .p64 => {
                const buf = try self.base.allocator.alloc(elf.Elf64_Shdr, self.sections.items.len);
                defer self.base.allocator.free(buf);

                for (buf) |*shdr, i| {
                    shdr.* = self.sections.items[i];
                    log.debug("writing section {}", .{shdr.*});
                    if (foreign_endian) {
                        mem.bswapAllFields(elf.Elf64_Shdr, shdr);
                    }
                }
                try self.base.file.?.pwriteAll(mem.sliceAsBytes(buf), self.shdr_table_offset.?);
            },
        }
        self.shdr_table_dirty = false;
    }
    if (self.entry_addr == null and self.base.options.effectiveOutputMode() == .Exe) {
        log.debug("flushing. no_entry_point_found = true", .{});
        self.error_flags.no_entry_point_found = true;
    } else {
        log.debug("flushing. no_entry_point_found = false", .{});
        self.error_flags.no_entry_point_found = false;
        try self.writeElfHeader();
    }

    // The point of flush() is to commit changes, so in theory, nothing should
    // be dirty after this. However, it is possible for some things to remain
    // dirty because they fail to be written in the event of compile errors,
    // such as debug_line_header_dirty and debug_info_header_dirty.
    assert(!self.debug_abbrev_section_dirty);
    assert(!self.debug_aranges_section_dirty);
    assert(!self.phdr_table_dirty);
    assert(!self.shdr_table_dirty);
    assert(!self.shstrtab_dirty);
    assert(!self.debug_strtab_dirty);
}

fn linkWithLLD(self: *Elf, comp: *Compilation) !void {
    const tracy = trace(@src());
    defer tracy.end();

    var arena_allocator = std.heap.ArenaAllocator.init(self.base.allocator);
    defer arena_allocator.deinit();
    const arena = &arena_allocator.allocator;

    const directory = self.base.options.emit.?.directory; // Just an alias to make it shorter to type.

    // If there is no Zig code to compile, then we should skip flushing the output file because it
    // will not be part of the linker line anyway.
    const module_obj_path: ?[]const u8 = if (self.base.options.module) |module| blk: {
        // stage1 puts the object file in the cache directory.
        if (self.base.options.use_stage1) {
            const obj_basename = try std.zig.binNameAlloc(arena, .{
                .root_name = self.base.options.root_name,
                .target = self.base.options.target,
                .output_mode = .Obj,
            });
            const o_directory = module.zig_cache_artifact_directory;
            const full_obj_path = try o_directory.join(arena, &[_][]const u8{obj_basename});
            break :blk full_obj_path;
        }

        try self.flushModule(comp);
        const obj_basename = self.base.intermediary_basename.?;
        const full_obj_path = try directory.join(arena, &[_][]const u8{obj_basename});
        break :blk full_obj_path;
    } else null;

    const is_obj = self.base.options.output_mode == .Obj;
    const is_lib = self.base.options.output_mode == .Lib;
    const is_dyn_lib = self.base.options.link_mode == .Dynamic and is_lib;
    const is_exe_or_dyn_lib = is_dyn_lib or self.base.options.output_mode == .Exe;
    const have_dynamic_linker = self.base.options.link_libc and
        self.base.options.link_mode == .Dynamic and is_exe_or_dyn_lib;
    const target = self.base.options.target;
    const gc_sections = self.base.options.gc_sections orelse !is_obj;
    const stack_size = self.base.options.stack_size_override orelse 16777216;
    const allow_shlib_undefined = self.base.options.allow_shlib_undefined orelse !self.base.options.is_native_os;
    const compiler_rt_path: ?[]const u8 = blk: {
        if (comp.compiler_rt_static_lib) |x| break :blk x.full_object_path;
        if (comp.compiler_rt_obj) |x| break :blk x.full_object_path;
        break :blk null;
    };

    // Here we want to determine whether we can save time by not invoking LLD when the
    // output is unchanged. None of the linker options or the object files that are being
    // linked are in the hash that namespaces the directory we are outputting to. Therefore,
    // we must hash those now, and the resulting digest will form the "id" of the linking
    // job we are about to perform.
    // After a successful link, we store the id in the metadata of a symlink named "id.txt" in
    // the artifact directory. So, now, we check if this symlink exists, and if it matches
    // our digest. If so, we can skip linking. Otherwise, we proceed with invoking LLD.
    const id_symlink_basename = "lld.id";

    var man: Cache.Manifest = undefined;
    defer if (!self.base.options.disable_lld_caching) man.deinit();

    var digest: [Cache.hex_digest_len]u8 = undefined;

    if (!self.base.options.disable_lld_caching) {
        man = comp.cache_parent.obtain();

        // We are about to obtain this lock, so here we give other processes a chance first.
        self.base.releaseLock();

        try man.addOptionalFile(self.base.options.linker_script);
        try man.addOptionalFile(self.base.options.version_script);
        try man.addListOfFiles(self.base.options.objects);
        for (comp.c_object_table.keys()) |key| {
            _ = try man.addFile(key.status.success.object_path, null);
        }
        try man.addOptionalFile(module_obj_path);
        try man.addOptionalFile(compiler_rt_path);

        // We can skip hashing libc and libc++ components that we are in charge of building from Zig
        // installation sources because they are always a product of the compiler version + target information.
        man.hash.add(stack_size);
        man.hash.addOptional(self.base.options.image_base_override);
        man.hash.add(gc_sections);
        man.hash.add(self.base.options.eh_frame_hdr);
        man.hash.add(self.base.options.emit_relocs);
        man.hash.add(self.base.options.rdynamic);
        man.hash.addListOfBytes(self.base.options.lib_dirs);
        man.hash.addListOfBytes(self.base.options.rpath_list);
        man.hash.add(self.base.options.each_lib_rpath);
        man.hash.add(self.base.options.skip_linker_dependencies);
        man.hash.add(self.base.options.z_nodelete);
        man.hash.add(self.base.options.z_notext);
        man.hash.add(self.base.options.z_defs);
        man.hash.add(self.base.options.z_origin);
        man.hash.add(self.base.options.z_noexecstack);
        man.hash.add(self.base.options.z_now);
        man.hash.add(self.base.options.z_relro);
        if (self.base.options.link_libc) {
            man.hash.add(self.base.options.libc_installation != null);
            if (self.base.options.libc_installation) |libc_installation| {
                man.hash.addBytes(libc_installation.crt_dir.?);
            }
            if (have_dynamic_linker) {
                man.hash.addOptionalBytes(self.base.options.dynamic_linker);
            }
        }
        man.hash.addOptionalBytes(self.base.options.soname);
        man.hash.addOptional(self.base.options.version);
        link.hashAddSystemLibs(&man.hash, self.base.options.system_libs);
        man.hash.add(allow_shlib_undefined);
        man.hash.add(self.base.options.bind_global_refs_locally);
        man.hash.add(self.base.options.tsan);
        man.hash.addOptionalBytes(self.base.options.sysroot);
        man.hash.add(self.base.options.linker_optimization);

        // We don't actually care whether it's a cache hit or miss; we just need the digest and the lock.
        _ = try man.hit();
        digest = man.final();

        var prev_digest_buf: [digest.len]u8 = undefined;
        const prev_digest: []u8 = Cache.readSmallFile(
            directory.handle,
            id_symlink_basename,
            &prev_digest_buf,
        ) catch |err| blk: {
            log.debug("ELF LLD new_digest={s} error: {s}", .{ std.fmt.fmtSliceHexLower(&digest), @errorName(err) });
            // Handle this as a cache miss.
            break :blk prev_digest_buf[0..0];
        };
        if (mem.eql(u8, prev_digest, &digest)) {
            log.debug("ELF LLD digest={s} match - skipping invocation", .{std.fmt.fmtSliceHexLower(&digest)});
            // Hot diggity dog! The output binary is already there.
            self.base.lock = man.toOwnedLock();
            return;
        }
        log.debug("ELF LLD prev_digest={s} new_digest={s}", .{ std.fmt.fmtSliceHexLower(prev_digest), std.fmt.fmtSliceHexLower(&digest) });

        // We are about to change the output file to be different, so we invalidate the build hash now.
        directory.handle.deleteFile(id_symlink_basename) catch |err| switch (err) {
            error.FileNotFound => {},
            else => |e| return e,
        };
    }

    const full_out_path = try directory.join(arena, &[_][]const u8{self.base.options.emit.?.sub_path});

    // Due to a deficiency in LLD, we need to special-case BPF to a simple file copy when generating
    // relocatables. Normally, we would expect `lld -r` to work. However, because LLD wants to resolve
    // BPF relocations which it shouldn't, it fails before even generating the relocatable.
    if (self.base.options.output_mode == .Obj and (self.base.options.lto or target.isBpfFreestanding())) {
        // In this case we must do a simple file copy
        // here. TODO: think carefully about how we can avoid this redundant operation when doing
        // build-obj. See also the corresponding TODO in linkAsArchive.
        const the_object_path = blk: {
            if (self.base.options.objects.len != 0)
                break :blk self.base.options.objects[0];

            if (comp.c_object_table.count() != 0)
                break :blk comp.c_object_table.keys()[0].status.success.object_path;

            if (module_obj_path) |p|
                break :blk p;

            // TODO I think this is unreachable. Audit this situation when solving the above TODO
            // regarding eliding redundant object -> object transformations.
            return error.NoObjectsToLink;
        };
        // This can happen when using --enable-cache and using the stage1 backend. In this case
        // we can skip the file copy.
        if (!mem.eql(u8, the_object_path, full_out_path)) {
            try fs.cwd().copyFile(the_object_path, fs.cwd(), full_out_path, .{});
        }
    } else {

        // Create an LLD command line and invoke it.
        var argv = std.ArrayList([]const u8).init(self.base.allocator);
        defer argv.deinit();
        // We will invoke ourselves as a child process to gain access to LLD.
        // This is necessary because LLD does not behave properly as a library -
        // it calls exit() and does not reset all global data between invocations.
        try argv.appendSlice(&[_][]const u8{ comp.self_exe_path.?, "ld.lld" });
        if (is_obj) {
            try argv.append("-r");
        }

        try argv.append("-error-limit=0");

        if (self.base.options.sysroot) |sysroot| {
            try argv.append(try std.fmt.allocPrint(arena, "--sysroot={s}", .{sysroot}));
        }

        if (self.base.options.lto) {
            switch (self.base.options.optimize_mode) {
                .Debug => {},
                .ReleaseSmall => try argv.append("--lto-O2"),
                .ReleaseFast, .ReleaseSafe => try argv.append("--lto-O3"),
            }
        }
        try argv.append(try std.fmt.allocPrint(arena, "-O{d}", .{
            self.base.options.linker_optimization,
        }));

        if (self.base.options.output_mode == .Exe) {
            try argv.append("-z");
            try argv.append(try std.fmt.allocPrint(arena, "stack-size={d}", .{stack_size}));
        }

        if (self.base.options.image_base_override) |image_base| {
            try argv.append(try std.fmt.allocPrint(arena, "--image-base={d}", .{image_base}));
        }

        if (self.base.options.linker_script) |linker_script| {
            try argv.append("-T");
            try argv.append(linker_script);
        }

        if (gc_sections) {
            try argv.append("--gc-sections");
        }

        if (self.base.options.eh_frame_hdr) {
            try argv.append("--eh-frame-hdr");
        }

        if (self.base.options.emit_relocs) {
            try argv.append("--emit-relocs");
        }

        if (self.base.options.rdynamic) {
            try argv.append("--export-dynamic");
        }

        if (self.base.options.z_nodelete) {
            try argv.append("-z");
            try argv.append("nodelete");
        }
        if (self.base.options.z_notext) {
            try argv.append("-z");
            try argv.append("notext");
        }
        if (self.base.options.z_defs) {
            try argv.append("-z");
            try argv.append("defs");
        }
        if (self.base.options.z_origin) {
            try argv.append("-z");
            try argv.append("origin");
        }
        if (self.base.options.z_noexecstack) {
            try argv.append("-z");
            try argv.append("noexecstack");
        }
        if (self.base.options.z_now) {
            try argv.append("-z");
            try argv.append("now");
        }
        if (self.base.options.z_relro) {
            try argv.append("-z");
            try argv.append("relro");
        }

        if (getLDMOption(target)) |ldm| {
            // Any target ELF will use the freebsd osabi if suffixed with "_fbsd".
            const arg = if (target.os.tag == .freebsd)
                try std.fmt.allocPrint(arena, "{s}_fbsd", .{ldm})
            else
                ldm;
            try argv.append("-m");
            try argv.append(arg);
        }

        if (self.base.options.link_mode == .Static) {
            if (target.cpu.arch.isARM() or target.cpu.arch.isThumb()) {
                try argv.append("-Bstatic");
            } else {
                try argv.append("-static");
            }
        } else if (is_dyn_lib) {
            try argv.append("-shared");
        }

        if (self.base.options.pie and self.base.options.output_mode == .Exe) {
            try argv.append("-pie");
        }

        if (self.base.options.link_mode == .Dynamic and target.os.tag == .netbsd) {
            // Add options to produce shared objects with only 2 PT_LOAD segments.
            // NetBSD expects 2 PT_LOAD segments in a shared object, otherwise
            // ld.elf_so fails to load, emitting a general "not found" error.
            // See https://github.com/ziglang/zig/issues/9109 .
            try argv.append("--no-rosegment");
            try argv.append("-znorelro");
        }

        try argv.append("-o");
        try argv.append(full_out_path);

        // csu prelude
        var csu = try CsuObjects.init(arena, self.base.options, comp);
        if (csu.crt0) |v| try argv.append(v);
        if (csu.crti) |v| try argv.append(v);
        if (csu.crtbegin) |v| try argv.append(v);

        // rpaths
        var rpath_table = std.StringHashMap(void).init(self.base.allocator);
        defer rpath_table.deinit();
        for (self.base.options.rpath_list) |rpath| {
            if ((try rpath_table.fetchPut(rpath, {})) == null) {
                try argv.append("-rpath");
                try argv.append(rpath);
            }
        }
        if (self.base.options.each_lib_rpath) {
            var test_path = std.ArrayList(u8).init(self.base.allocator);
            defer test_path.deinit();
            for (self.base.options.lib_dirs) |lib_dir_path| {
                for (self.base.options.system_libs.keys()) |link_lib| {
                    test_path.clearRetainingCapacity();
                    const sep = fs.path.sep_str;
                    try test_path.writer().print("{s}" ++ sep ++ "lib{s}.so", .{
                        lib_dir_path, link_lib,
                    });
                    fs.cwd().access(test_path.items, .{}) catch |err| switch (err) {
                        error.FileNotFound => continue,
                        else => |e| return e,
                    };
                    if ((try rpath_table.fetchPut(lib_dir_path, {})) == null) {
                        try argv.append("-rpath");
                        try argv.append(lib_dir_path);
                    }
                }
            }
        }

        for (self.base.options.lib_dirs) |lib_dir| {
            try argv.append("-L");
            try argv.append(lib_dir);
        }

        if (self.base.options.link_libc) {
            if (self.base.options.libc_installation) |libc_installation| {
                try argv.append("-L");
                try argv.append(libc_installation.crt_dir.?);
            }

            if (have_dynamic_linker) {
                if (self.base.options.dynamic_linker) |dynamic_linker| {
                    try argv.append("-dynamic-linker");
                    try argv.append(dynamic_linker);
                }
            }
        }

        if (is_dyn_lib) {
            if (self.base.options.soname) |soname| {
                try argv.append("-soname");
                try argv.append(soname);
            }
            if (self.base.options.version_script) |version_script| {
                try argv.append("-version-script");
                try argv.append(version_script);
            }
        }

        // Positional arguments to the linker such as object files.
        try argv.appendSlice(self.base.options.objects);

        for (comp.c_object_table.keys()) |key| {
            try argv.append(key.status.success.object_path);
        }

        if (module_obj_path) |p| {
            try argv.append(p);
        }

        // TSAN
        if (self.base.options.tsan) {
            try argv.append(comp.tsan_static_lib.?.full_object_path);
        }

        // libc
        if (is_exe_or_dyn_lib and
            !self.base.options.skip_linker_dependencies and
            !self.base.options.link_libc)
        {
            if (comp.libc_static_lib) |lib| {
                try argv.append(lib.full_object_path);
            }
        }

        // compiler-rt
        if (compiler_rt_path) |p| {
            try argv.append(p);
        }

        // Shared libraries.
        if (is_exe_or_dyn_lib) {
            const system_libs = self.base.options.system_libs.keys();
            const system_libs_values = self.base.options.system_libs.values();

            // Worst-case, we need an --as-needed argument for every lib, as well
            // as one before and one after.
            try argv.ensureUnusedCapacity(system_libs.len * 2 + 2);
            argv.appendAssumeCapacity("--as-needed");
            var as_needed = true;

            for (system_libs) |link_lib, i| {
                const lib_as_needed = !system_libs_values[i].needed;
                switch ((@as(u2, @boolToInt(lib_as_needed)) << 1) | @boolToInt(as_needed)) {
                    0b00, 0b11 => {},
                    0b01 => {
                        argv.appendAssumeCapacity("--no-as-needed");
                        as_needed = false;
                    },
                    0b10 => {
                        argv.appendAssumeCapacity("--as-needed");
                        as_needed = true;
                    },
                }

                // By this time, we depend on these libs being dynamically linked
                // libraries and not static libraries (the check for that needs to be earlier),
                // but they could be full paths to .so files, in which case we
                // want to avoid prepending "-l".
                const ext = Compilation.classifyFileExt(link_lib);
                const arg = if (ext == .shared_library) link_lib else try std.fmt.allocPrint(arena, "-l{s}", .{link_lib});
                argv.appendAssumeCapacity(arg);
            }

            if (!as_needed) {
                argv.appendAssumeCapacity("--as-needed");
                as_needed = true;
            }

            // libc++ dep
            if (self.base.options.link_libcpp) {
                try argv.append(comp.libcxxabi_static_lib.?.full_object_path);
                try argv.append(comp.libcxx_static_lib.?.full_object_path);
            }

            // libunwind dep
            if (self.base.options.link_libunwind) {
                try argv.append(comp.libunwind_static_lib.?.full_object_path);
            }

            // libc dep
            if (self.base.options.link_libc) {
                if (self.base.options.libc_installation != null) {
                    const needs_grouping = self.base.options.link_mode == .Static;
                    if (needs_grouping) try argv.append("--start-group");
                    try argv.appendSlice(target_util.libcFullLinkFlags(target));
                    if (needs_grouping) try argv.append("--end-group");
                } else if (target.isGnuLibC()) {
                    for (glibc.libs) |lib| {
                        const lib_path = try std.fmt.allocPrint(arena, "{s}{c}lib{s}.so.{d}", .{
                            comp.glibc_so_files.?.dir_path, fs.path.sep, lib.name, lib.sover,
                        });
                        try argv.append(lib_path);
                    }
                    try argv.append(try comp.get_libc_crt_file(arena, "libc_nonshared.a"));
                } else if (target.isMusl()) {
                    try argv.append(try comp.get_libc_crt_file(arena, switch (self.base.options.link_mode) {
                        .Static => "libc.a",
                        .Dynamic => "libc.so",
                    }));
                } else {
                    unreachable; // Compiler was supposed to emit an error for not being able to provide libc.
                }
            }
        }

        // crt postlude
        if (csu.crtend) |v| try argv.append(v);
        if (csu.crtn) |v| try argv.append(v);

        if (allow_shlib_undefined) {
            try argv.append("--allow-shlib-undefined");
        }

        if (self.base.options.bind_global_refs_locally) {
            try argv.append("-Bsymbolic");
        }

        if (self.base.options.verbose_link) {
            // Skip over our own name so that the LLD linker name is the first argv item.
            Compilation.dump_argv(argv.items[1..]);
        }

        // Sadly, we must run LLD as a child process because it does not behave
        // properly as a library.
        const child = try std.ChildProcess.init(argv.items, arena);
        defer child.deinit();

        if (comp.clang_passthrough_mode) {
            child.stdin_behavior = .Inherit;
            child.stdout_behavior = .Inherit;
            child.stderr_behavior = .Inherit;

            const term = child.spawnAndWait() catch |err| {
                log.err("unable to spawn {s}: {s}", .{ argv.items[0], @errorName(err) });
                return error.UnableToSpawnSelf;
            };
            switch (term) {
                .Exited => |code| {
                    if (code != 0) {
                        // TODO https://github.com/ziglang/zig/issues/6342
                        std.process.exit(1);
                    }
                },
                else => std.process.abort(),
            }
        } else {
            child.stdin_behavior = .Ignore;
            child.stdout_behavior = .Ignore;
            child.stderr_behavior = .Pipe;

            try child.spawn();

            const stderr = try child.stderr.?.reader().readAllAlloc(arena, 10 * 1024 * 1024);

            const term = child.wait() catch |err| {
                log.err("unable to spawn {s}: {s}", .{ argv.items[0], @errorName(err) });
                return error.UnableToSpawnSelf;
            };

            switch (term) {
                .Exited => |code| {
                    if (code != 0) {
                        // TODO parse this output and surface with the Compilation API rather than
                        // directly outputting to stderr here.
                        std.debug.print("{s}", .{stderr});
                        return error.LLDReportedFailure;
                    }
                },
                else => {
                    log.err("{s} terminated with stderr:\n{s}", .{ argv.items[0], stderr });
                    return error.LLDCrashed;
                },
            }

            if (stderr.len != 0) {
                log.warn("unexpected LLD stderr:\n{s}", .{stderr});
            }
        }
    }

    if (!self.base.options.disable_lld_caching) {
        // Update the file with the digest. If it fails we can continue; it only
        // means that the next invocation will have an unnecessary cache miss.
        Cache.writeSmallFile(directory.handle, id_symlink_basename, &digest) catch |err| {
            log.warn("failed to save linking hash digest file: {s}", .{@errorName(err)});
        };
        // Again failure here only means an unnecessary cache miss.
        man.writeManifest() catch |err| {
            log.warn("failed to write cache manifest when linking: {s}", .{@errorName(err)});
        };
        // We hang on to this lock so that the output file path can be used without
        // other processes clobbering it.
        self.base.lock = man.toOwnedLock();
    }
}

fn writeDwarfAddrAssumeCapacity(self: *Elf, buf: *std.ArrayList(u8), addr: u64) void {
    const target_endian = self.base.options.target.cpu.arch.endian();
    switch (self.ptr_width) {
        .p32 => mem.writeInt(u32, buf.addManyAsArrayAssumeCapacity(4), @intCast(u32, addr), target_endian),
        .p64 => mem.writeInt(u64, buf.addManyAsArrayAssumeCapacity(8), addr, target_endian),
    }
}

fn writeElfHeader(self: *Elf) !void {
    var hdr_buf: [@sizeOf(elf.Elf64_Ehdr)]u8 = undefined;

    var index: usize = 0;
    hdr_buf[0..4].* = "\x7fELF".*;
    index += 4;

    hdr_buf[index] = switch (self.ptr_width) {
        .p32 => elf.ELFCLASS32,
        .p64 => elf.ELFCLASS64,
    };
    index += 1;

    const endian = self.base.options.target.cpu.arch.endian();
    hdr_buf[index] = switch (endian) {
        .Little => elf.ELFDATA2LSB,
        .Big => elf.ELFDATA2MSB,
    };
    index += 1;

    hdr_buf[index] = 1; // ELF version
    index += 1;

    // OS ABI, often set to 0 regardless of target platform
    // ABI Version, possibly used by glibc but not by static executables
    // padding
    mem.set(u8, hdr_buf[index..][0..9], 0);
    index += 9;

    assert(index == 16);

    const elf_type = switch (self.base.options.effectiveOutputMode()) {
        .Exe => elf.ET.EXEC,
        .Obj => elf.ET.REL,
        .Lib => switch (self.base.options.link_mode) {
            .Static => elf.ET.REL,
            .Dynamic => elf.ET.DYN,
        },
    };
    mem.writeInt(u16, hdr_buf[index..][0..2], @enumToInt(elf_type), endian);
    index += 2;

    const machine = self.base.options.target.cpu.arch.toElfMachine();
    mem.writeInt(u16, hdr_buf[index..][0..2], @enumToInt(machine), endian);
    index += 2;

    // ELF Version, again
    mem.writeInt(u32, hdr_buf[index..][0..4], 1, endian);
    index += 4;

    const e_entry = if (elf_type == .REL) 0 else self.entry_addr.?;

    switch (self.ptr_width) {
        .p32 => {
            mem.writeInt(u32, hdr_buf[index..][0..4], @intCast(u32, e_entry), endian);
            index += 4;

            // e_phoff
            mem.writeInt(u32, hdr_buf[index..][0..4], @intCast(u32, self.phdr_table_offset.?), endian);
            index += 4;

            // e_shoff
            mem.writeInt(u32, hdr_buf[index..][0..4], @intCast(u32, self.shdr_table_offset.?), endian);
            index += 4;
        },
        .p64 => {
            // e_entry
            mem.writeInt(u64, hdr_buf[index..][0..8], e_entry, endian);
            index += 8;

            // e_phoff
            mem.writeInt(u64, hdr_buf[index..][0..8], self.phdr_table_offset.?, endian);
            index += 8;

            // e_shoff
            mem.writeInt(u64, hdr_buf[index..][0..8], self.shdr_table_offset.?, endian);
            index += 8;
        },
    }

    const e_flags = 0;
    mem.writeInt(u32, hdr_buf[index..][0..4], e_flags, endian);
    index += 4;

    const e_ehsize: u16 = switch (self.ptr_width) {
        .p32 => @sizeOf(elf.Elf32_Ehdr),
        .p64 => @sizeOf(elf.Elf64_Ehdr),
    };
    mem.writeInt(u16, hdr_buf[index..][0..2], e_ehsize, endian);
    index += 2;

    const e_phentsize: u16 = switch (self.ptr_width) {
        .p32 => @sizeOf(elf.Elf32_Phdr),
        .p64 => @sizeOf(elf.Elf64_Phdr),
    };
    mem.writeInt(u16, hdr_buf[index..][0..2], e_phentsize, endian);
    index += 2;

    const e_phnum = @intCast(u16, self.program_headers.items.len);
    mem.writeInt(u16, hdr_buf[index..][0..2], e_phnum, endian);
    index += 2;

    const e_shentsize: u16 = switch (self.ptr_width) {
        .p32 => @sizeOf(elf.Elf32_Shdr),
        .p64 => @sizeOf(elf.Elf64_Shdr),
    };
    mem.writeInt(u16, hdr_buf[index..][0..2], e_shentsize, endian);
    index += 2;

    const e_shnum = @intCast(u16, self.sections.items.len);
    mem.writeInt(u16, hdr_buf[index..][0..2], e_shnum, endian);
    index += 2;

    mem.writeInt(u16, hdr_buf[index..][0..2], self.shstrtab_index.?, endian);
    index += 2;

    assert(index == e_ehsize);

    try self.base.file.?.pwriteAll(hdr_buf[0..index], 0);
}

fn freeTextBlock(self: *Elf, text_block: *TextBlock) void {
    var already_have_free_list_node = false;
    {
        var i: usize = 0;
        // TODO turn text_block_free_list into a hash map
        while (i < self.text_block_free_list.items.len) {
            if (self.text_block_free_list.items[i] == text_block) {
                _ = self.text_block_free_list.swapRemove(i);
                continue;
            }
            if (self.text_block_free_list.items[i] == text_block.prev) {
                already_have_free_list_node = true;
            }
            i += 1;
        }
    }
    // TODO process free list for dbg info just like we do above for vaddrs

    if (self.last_text_block == text_block) {
        // TODO shrink the .text section size here
        self.last_text_block = text_block.prev;
    }
    if (self.dbg_info_decl_first == text_block) {
        self.dbg_info_decl_first = text_block.dbg_info_next;
    }
    if (self.dbg_info_decl_last == text_block) {
        // TODO shrink the .debug_info section size here
        self.dbg_info_decl_last = text_block.dbg_info_prev;
    }

    if (text_block.prev) |prev| {
        prev.next = text_block.next;

        if (!already_have_free_list_node and prev.freeListEligible(self.*)) {
            // The free list is heuristics, it doesn't have to be perfect, so we can
            // ignore the OOM here.
            self.text_block_free_list.append(self.base.allocator, prev) catch {};
        }
    } else {
        text_block.prev = null;
    }

    if (text_block.next) |next| {
        next.prev = text_block.prev;
    } else {
        text_block.next = null;
    }

    if (text_block.dbg_info_prev) |prev| {
        prev.dbg_info_next = text_block.dbg_info_next;

        // TODO the free list logic like we do for text blocks above
    } else {
        text_block.dbg_info_prev = null;
    }

    if (text_block.dbg_info_next) |next| {
        next.dbg_info_prev = text_block.dbg_info_prev;
    } else {
        text_block.dbg_info_next = null;
    }
}

fn shrinkTextBlock(self: *Elf, text_block: *TextBlock, new_block_size: u64) void {
    _ = self;
    _ = text_block;
    _ = new_block_size;
    // TODO check the new capacity, and if it crosses the size threshold into a big enough
    // capacity, insert a free list node for it.
}

fn growTextBlock(self: *Elf, text_block: *TextBlock, new_block_size: u64, alignment: u64) !u64 {
    const sym = self.local_symbols.items[text_block.local_sym_index];
    const align_ok = mem.alignBackwardGeneric(u64, sym.st_value, alignment) == sym.st_value;
    const need_realloc = !align_ok or new_block_size > text_block.capacity(self.*);
    if (!need_realloc) return sym.st_value;
    return self.allocateTextBlock(text_block, new_block_size, alignment);
}

fn allocateTextBlock(self: *Elf, text_block: *TextBlock, new_block_size: u64, alignment: u64) !u64 {
    const phdr = &self.program_headers.items[self.phdr_load_re_index.?];
    const shdr = &self.sections.items[self.text_section_index.?];
    const new_block_ideal_capacity = padToIdeal(new_block_size);

    // We use these to indicate our intention to update metadata, placing the new block,
    // and possibly removing a free list node.
    // It would be simpler to do it inside the for loop below, but that would cause a
    // problem if an error was returned later in the function. So this action
    // is actually carried out at the end of the function, when errors are no longer possible.
    var block_placement: ?*TextBlock = null;
    var free_list_removal: ?usize = null;

    // First we look for an appropriately sized free list node.
    // The list is unordered. We'll just take the first thing that works.
    const vaddr = blk: {
        var i: usize = 0;
        while (i < self.text_block_free_list.items.len) {
            const big_block = self.text_block_free_list.items[i];
            // We now have a pointer to a live text block that has too much capacity.
            // Is it enough that we could fit this new text block?
            const sym = self.local_symbols.items[big_block.local_sym_index];
            const capacity = big_block.capacity(self.*);
            const ideal_capacity = padToIdeal(capacity);
            const ideal_capacity_end_vaddr = sym.st_value + ideal_capacity;
            const capacity_end_vaddr = sym.st_value + capacity;
            const new_start_vaddr_unaligned = capacity_end_vaddr - new_block_ideal_capacity;
            const new_start_vaddr = mem.alignBackwardGeneric(u64, new_start_vaddr_unaligned, alignment);
            if (new_start_vaddr < ideal_capacity_end_vaddr) {
                // Additional bookkeeping here to notice if this free list node
                // should be deleted because the block that it points to has grown to take up
                // more of the extra capacity.
                if (!big_block.freeListEligible(self.*)) {
                    _ = self.text_block_free_list.swapRemove(i);
                } else {
                    i += 1;
                }
                continue;
            }
            // At this point we know that we will place the new block here. But the
            // remaining question is whether there is still yet enough capacity left
            // over for there to still be a free list node.
            const remaining_capacity = new_start_vaddr - ideal_capacity_end_vaddr;
            const keep_free_list_node = remaining_capacity >= min_text_capacity;

            // Set up the metadata to be updated, after errors are no longer possible.
            block_placement = big_block;
            if (!keep_free_list_node) {
                free_list_removal = i;
            }
            break :blk new_start_vaddr;
        } else if (self.last_text_block) |last| {
            const sym = self.local_symbols.items[last.local_sym_index];
            const ideal_capacity = padToIdeal(sym.st_size);
            const ideal_capacity_end_vaddr = sym.st_value + ideal_capacity;
            const new_start_vaddr = mem.alignForwardGeneric(u64, ideal_capacity_end_vaddr, alignment);
            // Set up the metadata to be updated, after errors are no longer possible.
            block_placement = last;
            break :blk new_start_vaddr;
        } else {
            break :blk phdr.p_vaddr;
        }
    };

    const expand_text_section = block_placement == null or block_placement.?.next == null;
    if (expand_text_section) {
        const text_capacity = self.allocatedSize(shdr.sh_offset);
        const needed_size = (vaddr + new_block_size) - phdr.p_vaddr;
        if (needed_size > text_capacity) {
            // Must move the entire text section.
            const new_offset = self.findFreeSpace(needed_size, 0x1000);
            const text_size = if (self.last_text_block) |last| blk: {
                const sym = self.local_symbols.items[last.local_sym_index];
                break :blk (sym.st_value + sym.st_size) - phdr.p_vaddr;
            } else 0;
            const amt = try self.base.file.?.copyRangeAll(shdr.sh_offset, self.base.file.?, new_offset, text_size);
            if (amt != text_size) return error.InputOutput;
            shdr.sh_offset = new_offset;
            phdr.p_offset = new_offset;
        }
        self.last_text_block = text_block;

        shdr.sh_size = needed_size;
        phdr.p_memsz = needed_size;
        phdr.p_filesz = needed_size;

        // The .debug_info section has `low_pc` and `high_pc` values which is the virtual address
        // range of the compilation unit. When we expand the text section, this range changes,
        // so the DW_TAG.compile_unit tag of the .debug_info section becomes dirty.
        self.debug_info_header_dirty = true;
        // This becomes dirty for the same reason. We could potentially make this more
        // fine-grained with the addition of support for more compilation units. It is planned to
        // model each package as a different compilation unit.
        self.debug_aranges_section_dirty = true;

        self.phdr_table_dirty = true; // TODO look into making only the one program header dirty
        self.shdr_table_dirty = true; // TODO look into making only the one section dirty
    }

    // This function can also reallocate a text block.
    // In this case we need to "unplug" it from its previous location before
    // plugging it in to its new location.
    if (text_block.prev) |prev| {
        prev.next = text_block.next;
    }
    if (text_block.next) |next| {
        next.prev = text_block.prev;
    }

    if (block_placement) |big_block| {
        text_block.prev = big_block;
        text_block.next = big_block.next;
        big_block.next = text_block;
    } else {
        text_block.prev = null;
        text_block.next = null;
    }
    if (free_list_removal) |i| {
        _ = self.text_block_free_list.swapRemove(i);
    }
    return vaddr;
}

pub fn allocateDeclIndexes(self: *Elf, decl: *Module.Decl) !void {
    if (self.llvm_object) |_| return;

    if (decl.link.elf.local_sym_index != 0) return;

    try self.local_symbols.ensureUnusedCapacity(self.base.allocator, 1);
    try self.offset_table.ensureUnusedCapacity(self.base.allocator, 1);

    if (self.local_symbol_free_list.popOrNull()) |i| {
        log.debug("reusing symbol index {d} for {s}", .{ i, decl.name });
        decl.link.elf.local_sym_index = i;
    } else {
        log.debug("allocating symbol index {d} for {s}", .{ self.local_symbols.items.len, decl.name });
        decl.link.elf.local_sym_index = @intCast(u32, self.local_symbols.items.len);
        _ = self.local_symbols.addOneAssumeCapacity();
    }

    if (self.offset_table_free_list.popOrNull()) |i| {
        decl.link.elf.offset_table_index = i;
    } else {
        decl.link.elf.offset_table_index = @intCast(u32, self.offset_table.items.len);
        _ = self.offset_table.addOneAssumeCapacity();
        self.offset_table_count_dirty = true;
    }

    const phdr = &self.program_headers.items[self.phdr_load_re_index.?];

    self.local_symbols.items[decl.link.elf.local_sym_index] = .{
        .st_name = 0,
        .st_info = 0,
        .st_other = 0,
        .st_shndx = 0,
        .st_value = phdr.p_vaddr,
        .st_size = 0,
    };
    self.offset_table.items[decl.link.elf.offset_table_index] = 0;
}

pub fn freeDecl(self: *Elf, decl: *Module.Decl) void {
    if (build_options.have_llvm) {
        if (self.llvm_object) |llvm_object| return llvm_object.freeDecl(decl);
    }

    // Appending to free lists is allowed to fail because the free lists are heuristics based anyway.
    self.freeTextBlock(&decl.link.elf);
    if (decl.link.elf.local_sym_index != 0) {
        self.local_symbol_free_list.append(self.base.allocator, decl.link.elf.local_sym_index) catch {};
        self.offset_table_free_list.append(self.base.allocator, decl.link.elf.offset_table_index) catch {};

        self.local_symbols.items[decl.link.elf.local_sym_index].st_info = 0;

        decl.link.elf.local_sym_index = 0;
    }
    // TODO make this logic match freeTextBlock. Maybe abstract the logic out since the same thing
    // is desired for both.
    _ = self.dbg_line_fn_free_list.remove(&decl.fn_link.elf);
    if (decl.fn_link.elf.prev) |prev| {
        self.dbg_line_fn_free_list.put(self.base.allocator, prev, {}) catch {};
        prev.next = decl.fn_link.elf.next;
        if (decl.fn_link.elf.next) |next| {
            next.prev = prev;
        } else {
            self.dbg_line_fn_last = prev;
        }
    } else if (decl.fn_link.elf.next) |next| {
        self.dbg_line_fn_first = next;
        next.prev = null;
    }
    if (self.dbg_line_fn_first == &decl.fn_link.elf) {
        self.dbg_line_fn_first = decl.fn_link.elf.next;
    }
    if (self.dbg_line_fn_last == &decl.fn_link.elf) {
        self.dbg_line_fn_last = decl.fn_link.elf.prev;
    }
}

fn deinitRelocs(gpa: *Allocator, table: *File.DbgInfoTypeRelocsTable) void {
    var it = table.valueIterator();
    while (it.next()) |value| {
        value.relocs.deinit(gpa);
    }
    table.deinit(gpa);
}

fn updateDeclCode(self: *Elf, decl: *Module.Decl, code: []const u8, stt_bits: u8) !*elf.Elf64_Sym {
    const required_alignment = decl.ty.abiAlignment(self.base.options.target);

    assert(decl.link.elf.local_sym_index != 0); // Caller forgot to allocateDeclIndexes()
    const local_sym = &self.local_symbols.items[decl.link.elf.local_sym_index];
    if (local_sym.st_size != 0) {
        const capacity = decl.link.elf.capacity(self.*);
        const need_realloc = code.len > capacity or
            !mem.isAlignedGeneric(u64, local_sym.st_value, required_alignment);
        if (need_realloc) {
            const vaddr = try self.growTextBlock(&decl.link.elf, code.len, required_alignment);
            log.debug("growing {s} from 0x{x} to 0x{x}", .{ decl.name, local_sym.st_value, vaddr });
            if (vaddr != local_sym.st_value) {
                local_sym.st_value = vaddr;

                log.debug("  (writing new offset table entry)", .{});
                self.offset_table.items[decl.link.elf.offset_table_index] = vaddr;
                try self.writeOffsetTableEntry(decl.link.elf.offset_table_index);
            }
        } else if (code.len < local_sym.st_size) {
            self.shrinkTextBlock(&decl.link.elf, code.len);
        }
        local_sym.st_size = code.len;
        local_sym.st_name = try self.updateString(local_sym.st_name, mem.sliceTo(decl.name, 0));
        local_sym.st_info = (elf.STB_LOCAL << 4) | stt_bits;
        local_sym.st_other = 0;
        local_sym.st_shndx = self.text_section_index.?;
        // TODO this write could be avoided if no fields of the symbol were changed.
        try self.writeSymbol(decl.link.elf.local_sym_index);
    } else {
        const decl_name = mem.sliceTo(decl.name, 0);
        const name_str_index = try self.makeString(decl_name);
        const vaddr = try self.allocateTextBlock(&decl.link.elf, code.len, required_alignment);
        log.debug("allocated text block for {s} at 0x{x}", .{ decl_name, vaddr });
        errdefer self.freeTextBlock(&decl.link.elf);

        local_sym.* = .{
            .st_name = name_str_index,
            .st_info = (elf.STB_LOCAL << 4) | stt_bits,
            .st_other = 0,
            .st_shndx = self.text_section_index.?,
            .st_value = vaddr,
            .st_size = code.len,
        };
        self.offset_table.items[decl.link.elf.offset_table_index] = vaddr;

        try self.writeSymbol(decl.link.elf.local_sym_index);
        try self.writeOffsetTableEntry(decl.link.elf.offset_table_index);
    }

    const section_offset = local_sym.st_value - self.program_headers.items[self.phdr_load_re_index.?].p_vaddr;
    const file_offset = self.sections.items[self.text_section_index.?].sh_offset + section_offset;
    try self.base.file.?.pwriteAll(code, file_offset);

    return local_sym;
}

fn finishUpdateDecl(
    self: *Elf,
    module: *Module,
    decl: *Module.Decl,
    dbg_info_type_relocs: *File.DbgInfoTypeRelocsTable,
    dbg_info_buffer: *std.ArrayList(u8),
) !void {
    // Now we emit the .debug_info types of the Decl. These will count towards the size of
    // the buffer, so we have to do it before computing the offset, and we can't perform the actual
    // relocations yet.
    {
        var it = dbg_info_type_relocs.iterator();
        while (it.next()) |entry| {
            entry.value_ptr.off = @intCast(u32, dbg_info_buffer.items.len);
            try self.addDbgInfoType(entry.key_ptr.*, dbg_info_buffer);
        }
    }

    const text_block = &decl.link.elf;
    try self.updateDeclDebugInfoAllocation(text_block, @intCast(u32, dbg_info_buffer.items.len));

    const target_endian = self.base.options.target.cpu.arch.endian();

    {
        // Now that we have the offset assigned we can finally perform type relocations.
        var it = dbg_info_type_relocs.valueIterator();
        while (it.next()) |value| {
            for (value.relocs.items) |off| {
                mem.writeInt(
                    u32,
                    dbg_info_buffer.items[off..][0..4],
                    text_block.dbg_info_off + value.off,
                    target_endian,
                );
            }
        }
    }

    try self.writeDeclDebugInfo(text_block, dbg_info_buffer.items);

    // Since we updated the vaddr and the size, each corresponding export symbol also needs to be updated.
    const decl_exports = module.decl_exports.get(decl) orelse &[0]*Module.Export{};
    return self.updateDeclExports(module, decl, decl_exports);
}

pub fn updateFunc(self: *Elf, module: *Module, func: *Module.Fn, air: Air, liveness: Liveness) !void {
    if (build_options.skip_non_native and builtin.object_format != .elf) {
        @panic("Attempted to compile for object format that was disabled by build configuration");
    }
    if (build_options.have_llvm) {
        if (self.llvm_object) |llvm_object| return llvm_object.updateFunc(module, func, air, liveness);
    }

    const tracy = trace(@src());
    defer tracy.end();

    var code_buffer = std.ArrayList(u8).init(self.base.allocator);
    defer code_buffer.deinit();

    // For functions we need to add a prologue to the debug line program.
    var dbg_line_buffer = try std.ArrayList(u8).initCapacity(self.base.allocator, 26);
    defer dbg_line_buffer.deinit();

    var dbg_info_buffer = std.ArrayList(u8).init(self.base.allocator);
    defer dbg_info_buffer.deinit();

    var dbg_info_type_relocs: File.DbgInfoTypeRelocsTable = .{};
    defer deinitRelocs(self.base.allocator, &dbg_info_type_relocs);

    const decl = func.owner_decl;
    const line_off = @intCast(u28, decl.src_line + func.lbrace_line);

    const ptr_width_bytes = self.ptrWidthBytes();
    dbg_line_buffer.appendSliceAssumeCapacity(&[_]u8{
        DW.LNS.extended_op,
        ptr_width_bytes + 1,
        DW.LNE.set_address,
    });
    // This is the "relocatable" vaddr, corresponding to `code_buffer` index `0`.
    assert(dbg_line_vaddr_reloc_index == dbg_line_buffer.items.len);
    dbg_line_buffer.items.len += ptr_width_bytes;

    dbg_line_buffer.appendAssumeCapacity(DW.LNS.advance_line);
    // This is the "relocatable" relative line offset from the previous function's end curly
    // to this function's begin curly.
    assert(self.getRelocDbgLineOff() == dbg_line_buffer.items.len);
    // Here we use a ULEB128-fixed-4 to make sure this field can be overwritten later.
    leb128.writeUnsignedFixed(4, dbg_line_buffer.addManyAsArrayAssumeCapacity(4), line_off);

    dbg_line_buffer.appendAssumeCapacity(DW.LNS.set_file);
    assert(self.getRelocDbgFileIndex() == dbg_line_buffer.items.len);
    // Once we support more than one source file, this will have the ability to be more
    // than one possible value.
    const file_index = 1;
    leb128.writeUnsignedFixed(4, dbg_line_buffer.addManyAsArrayAssumeCapacity(4), file_index);

    // Emit a line for the begin curly with prologue_end=false. The codegen will
    // do the work of setting prologue_end=true and epilogue_begin=true.
    dbg_line_buffer.appendAssumeCapacity(DW.LNS.copy);

    // .debug_info subprogram
    const decl_name_with_null = decl.name[0 .. mem.sliceTo(decl.name, 0).len + 1];
    try dbg_info_buffer.ensureUnusedCapacity(25 + decl_name_with_null.len);

    const fn_ret_type = decl.ty.fnReturnType();
    const fn_ret_has_bits = fn_ret_type.hasCodeGenBits();
    if (fn_ret_has_bits) {
        dbg_info_buffer.appendAssumeCapacity(abbrev_subprogram);
    } else {
        dbg_info_buffer.appendAssumeCapacity(abbrev_subprogram_retvoid);
    }
    // These get overwritten after generating the machine code. These values are
    // "relocations" and have to be in this fixed place so that functions can be
    // moved in virtual address space.
    assert(dbg_info_low_pc_reloc_index == dbg_info_buffer.items.len);
    dbg_info_buffer.items.len += ptr_width_bytes; // DW.AT.low_pc,  DW.FORM.addr
    assert(self.getRelocDbgInfoSubprogramHighPC() == dbg_info_buffer.items.len);
    dbg_info_buffer.items.len += 4; // DW.AT.high_pc,  DW.FORM.data4
    if (fn_ret_has_bits) {
        const gop = try dbg_info_type_relocs.getOrPut(self.base.allocator, fn_ret_type);
        if (!gop.found_existing) {
            gop.value_ptr.* = .{
                .off = undefined,
                .relocs = .{},
            };
        }
        try gop.value_ptr.relocs.append(self.base.allocator, @intCast(u32, dbg_info_buffer.items.len));
        dbg_info_buffer.items.len += 4; // DW.AT.type,  DW.FORM.ref4
    }
    dbg_info_buffer.appendSliceAssumeCapacity(decl_name_with_null); // DW.AT.name, DW.FORM.string

    const res = try codegen.generateFunction(&self.base, decl.srcLoc(), func, air, liveness, &code_buffer, .{
        .dwarf = .{
            .dbg_line = &dbg_line_buffer,
            .dbg_info = &dbg_info_buffer,
            .dbg_info_type_relocs = &dbg_info_type_relocs,
        },
    });
    const code = switch (res) {
        .appended => code_buffer.items,
        .fail => |em| {
            decl.analysis = .codegen_failure;
            try module.failed_decls.put(module.gpa, decl, em);
            return;
        },
    };

    const local_sym = try self.updateDeclCode(decl, code, elf.STT_FUNC);

    const target_endian = self.base.options.target.cpu.arch.endian();

    // Since the Decl is a function, we need to update the .debug_line program.
    // Perform the relocations based on vaddr.
    switch (self.ptr_width) {
        .p32 => {
            {
                const ptr = dbg_line_buffer.items[dbg_line_vaddr_reloc_index..][0..4];
                mem.writeInt(u32, ptr, @intCast(u32, local_sym.st_value), target_endian);
            }
            {
                const ptr = dbg_info_buffer.items[dbg_info_low_pc_reloc_index..][0..4];
                mem.writeInt(u32, ptr, @intCast(u32, local_sym.st_value), target_endian);
            }
        },
        .p64 => {
            {
                const ptr = dbg_line_buffer.items[dbg_line_vaddr_reloc_index..][0..8];
                mem.writeInt(u64, ptr, local_sym.st_value, target_endian);
            }
            {
                const ptr = dbg_info_buffer.items[dbg_info_low_pc_reloc_index..][0..8];
                mem.writeInt(u64, ptr, local_sym.st_value, target_endian);
            }
        },
    }
    {
        const ptr = dbg_info_buffer.items[self.getRelocDbgInfoSubprogramHighPC()..][0..4];
        mem.writeInt(u32, ptr, @intCast(u32, local_sym.st_size), target_endian);
    }

    try dbg_line_buffer.appendSlice(&[_]u8{ DW.LNS.extended_op, 1, DW.LNE.end_sequence });

    // Now we have the full contents and may allocate a region to store it.

    // This logic is nearly identical to the logic below in `updateDeclDebugInfoAllocation` for
    // `TextBlock` and the .debug_info. If you are editing this logic, you
    // probably need to edit that logic too.

    const debug_line_sect = &self.sections.items[self.debug_line_section_index.?];
    const src_fn = &decl.fn_link.elf;
    src_fn.len = @intCast(u32, dbg_line_buffer.items.len);
    if (self.dbg_line_fn_last) |last| not_first: {
        if (src_fn.next) |next| {
            // Update existing function - non-last item.
            if (src_fn.off + src_fn.len + min_nop_size > next.off) {
                // It grew too big, so we move it to a new location.
                if (src_fn.prev) |prev| {
                    self.dbg_line_fn_free_list.put(self.base.allocator, prev, {}) catch {};
                    prev.next = src_fn.next;
                }
                assert(src_fn.prev != next);
                next.prev = src_fn.prev;
                src_fn.next = null;
                // Populate where it used to be with NOPs.
                const file_pos = debug_line_sect.sh_offset + src_fn.off;
                try self.pwriteDbgLineNops(0, &[0]u8{}, src_fn.len, file_pos);
                // TODO Look at the free list before appending at the end.
                src_fn.prev = last;
                last.next = src_fn;
                self.dbg_line_fn_last = src_fn;

                src_fn.off = last.off + padToIdeal(last.len);
            }
        } else if (src_fn.prev == null) {
            if (src_fn == last) {
                // Special case: there is only 1 function and it is being updated.
                // In this case there is nothing to do. The function's length has
                // already been updated, and the logic below takes care of
                // resizing the .debug_line section.
                break :not_first;
            }
            // Append new function.
            // TODO Look at the free list before appending at the end.
            src_fn.prev = last;
            last.next = src_fn;
            self.dbg_line_fn_last = src_fn;

            src_fn.off = last.off + padToIdeal(last.len);
        }
    } else {
        // This is the first function of the Line Number Program.
        self.dbg_line_fn_first = src_fn;
        self.dbg_line_fn_last = src_fn;

        src_fn.off = padToIdeal(self.dbgLineNeededHeaderBytes());
    }

    const last_src_fn = self.dbg_line_fn_last.?;
    const needed_size = last_src_fn.off + last_src_fn.len;
    if (needed_size != debug_line_sect.sh_size) {
        if (needed_size > self.allocatedSize(debug_line_sect.sh_offset)) {
            const new_offset = self.findFreeSpace(needed_size, 1);
            const existing_size = last_src_fn.off;
            log.debug("moving .debug_line section: {d} bytes from 0x{x} to 0x{x}", .{
                existing_size,
                debug_line_sect.sh_offset,
                new_offset,
            });
            const amt = try self.base.file.?.copyRangeAll(debug_line_sect.sh_offset, self.base.file.?, new_offset, existing_size);
            if (amt != existing_size) return error.InputOutput;
            debug_line_sect.sh_offset = new_offset;
        }
        debug_line_sect.sh_size = needed_size;
        self.shdr_table_dirty = true; // TODO look into making only the one section dirty
        self.debug_line_header_dirty = true;
    }
    const prev_padding_size: u32 = if (src_fn.prev) |prev| src_fn.off - (prev.off + prev.len) else 0;
    const next_padding_size: u32 = if (src_fn.next) |next| next.off - (src_fn.off + src_fn.len) else 0;

    // We only have support for one compilation unit so far, so the offsets are directly
    // from the .debug_line section.
    const file_pos = debug_line_sect.sh_offset + src_fn.off;
    try self.pwriteDbgLineNops(prev_padding_size, dbg_line_buffer.items, next_padding_size, file_pos);

    // .debug_info - End the TAG.subprogram children.
    try dbg_info_buffer.append(0);

    return self.finishUpdateDecl(module, decl, &dbg_info_type_relocs, &dbg_info_buffer);
}

pub fn updateDecl(self: *Elf, module: *Module, decl: *Module.Decl) !void {
    if (build_options.skip_non_native and builtin.object_format != .elf) {
        @panic("Attempted to compile for object format that was disabled by build configuration");
    }
    if (build_options.have_llvm) {
        if (self.llvm_object) |llvm_object| return llvm_object.updateDecl(module, decl);
    }

    const tracy = trace(@src());
    defer tracy.end();

    if (decl.val.tag() == .extern_fn) {
        return; // TODO Should we do more when front-end analyzed extern decl?
    }
    if (decl.val.castTag(.variable)) |payload| {
        const variable = payload.data;
        if (variable.is_extern) {
            return; // TODO Should we do more when front-end analyzed extern decl?
        }
    }

    var code_buffer = std.ArrayList(u8).init(self.base.allocator);
    defer code_buffer.deinit();

    var dbg_line_buffer = std.ArrayList(u8).init(self.base.allocator);
    defer dbg_line_buffer.deinit();

    var dbg_info_buffer = std.ArrayList(u8).init(self.base.allocator);
    defer dbg_info_buffer.deinit();

    var dbg_info_type_relocs: File.DbgInfoTypeRelocsTable = .{};
    defer deinitRelocs(self.base.allocator, &dbg_info_type_relocs);

    // TODO implement .debug_info for global variables
    const decl_val = if (decl.val.castTag(.variable)) |payload| payload.data.init else decl.val;
    const res = try codegen.generateSymbol(&self.base, decl.srcLoc(), .{
        .ty = decl.ty,
        .val = decl_val,
    }, &code_buffer, .{
        .dwarf = .{
            .dbg_line = &dbg_line_buffer,
            .dbg_info = &dbg_info_buffer,
            .dbg_info_type_relocs = &dbg_info_type_relocs,
        },
    });
    const code = switch (res) {
        .externally_managed => |x| x,
        .appended => code_buffer.items,
        .fail => |em| {
            decl.analysis = .codegen_failure;
            try module.failed_decls.put(module.gpa, decl, em);
            return;
        },
    };

    _ = try self.updateDeclCode(decl, code, elf.STT_OBJECT);
    return self.finishUpdateDecl(module, decl, &dbg_info_type_relocs, &dbg_info_buffer);
}

/// Asserts the type has codegen bits.
fn addDbgInfoType(self: *Elf, ty: Type, dbg_info_buffer: *std.ArrayList(u8)) !void {
    switch (ty.zigTypeTag()) {
        .Void => unreachable,
        .NoReturn => unreachable,
        .Bool => {
            try dbg_info_buffer.appendSlice(&[_]u8{
                abbrev_base_type,
                DW.ATE.boolean, // DW.AT.encoding ,  DW.FORM.data1
                1, // DW.AT.byte_size,  DW.FORM.data1
                'b', 'o', 'o', 'l', 0, // DW.AT.name,  DW.FORM.string
            });
        },
        .Int => {
            const info = ty.intInfo(self.base.options.target);
            try dbg_info_buffer.ensureUnusedCapacity(12);
            dbg_info_buffer.appendAssumeCapacity(abbrev_base_type);
            // DW.AT.encoding, DW.FORM.data1
            dbg_info_buffer.appendAssumeCapacity(switch (info.signedness) {
                .signed => DW.ATE.signed,
                .unsigned => DW.ATE.unsigned,
            });
            // DW.AT.byte_size,  DW.FORM.data1
            dbg_info_buffer.appendAssumeCapacity(@intCast(u8, ty.abiSize(self.base.options.target)));
            // DW.AT.name,  DW.FORM.string
            try dbg_info_buffer.writer().print("{}\x00", .{ty});
        },
        .Optional => {
            if (ty.isPtrLikeOptional()) {
                try dbg_info_buffer.ensureUnusedCapacity(12);
                dbg_info_buffer.appendAssumeCapacity(abbrev_base_type);
                // DW.AT.encoding, DW.FORM.data1
                dbg_info_buffer.appendAssumeCapacity(DW.ATE.address);
                // DW.AT.byte_size,  DW.FORM.data1
                dbg_info_buffer.appendAssumeCapacity(@intCast(u8, ty.abiSize(self.base.options.target)));
                // DW.AT.name,  DW.FORM.string
                try dbg_info_buffer.writer().print("{}\x00", .{ty});
            } else {
                log.debug("TODO implement .debug_info for type '{}'", .{ty});
                try dbg_info_buffer.append(abbrev_pad1);
            }
        },
        else => {
            log.debug("TODO implement .debug_info for type '{}'", .{ty});
            try dbg_info_buffer.append(abbrev_pad1);
        },
    }
}

fn updateDeclDebugInfoAllocation(self: *Elf, text_block: *TextBlock, len: u32) !void {
    const tracy = trace(@src());
    defer tracy.end();

    // This logic is nearly identical to the logic above in `updateDecl` for
    // `SrcFn` and the line number programs. If you are editing this logic, you
    // probably need to edit that logic too.

    const debug_info_sect = &self.sections.items[self.debug_info_section_index.?];
    text_block.dbg_info_len = len;
    if (self.dbg_info_decl_last) |last| not_first: {
        if (text_block.dbg_info_next) |next| {
            // Update existing Decl - non-last item.
            if (text_block.dbg_info_off + text_block.dbg_info_len + min_nop_size > next.dbg_info_off) {
                // It grew too big, so we move it to a new location.
                if (text_block.dbg_info_prev) |prev| {
                    self.dbg_info_decl_free_list.put(self.base.allocator, prev, {}) catch {};
                    prev.dbg_info_next = text_block.dbg_info_next;
                }
                next.dbg_info_prev = text_block.dbg_info_prev;
                text_block.dbg_info_next = null;
                // Populate where it used to be with NOPs.
                const file_pos = debug_info_sect.sh_offset + text_block.dbg_info_off;
                try self.pwriteDbgInfoNops(0, &[0]u8{}, text_block.dbg_info_len, false, file_pos);
                // TODO Look at the free list before appending at the end.
                text_block.dbg_info_prev = last;
                last.dbg_info_next = text_block;
                self.dbg_info_decl_last = text_block;

                text_block.dbg_info_off = last.dbg_info_off + padToIdeal(last.dbg_info_len);
            }
        } else if (text_block.dbg_info_prev == null) {
            if (text_block == last) {
                // Special case: there is only 1 .debug_info block and it is being updated.
                // In this case there is nothing to do. The block's length has
                // already been updated, and logic in writeDeclDebugInfo takes care of
                // resizing the .debug_info section.
                break :not_first;
            }
            // Append new Decl.
            // TODO Look at the free list before appending at the end.
            text_block.dbg_info_prev = last;
            last.dbg_info_next = text_block;
            self.dbg_info_decl_last = text_block;

            text_block.dbg_info_off = last.dbg_info_off + padToIdeal(last.dbg_info_len);
        }
    } else {
        // This is the first Decl of the .debug_info
        self.dbg_info_decl_first = text_block;
        self.dbg_info_decl_last = text_block;

        text_block.dbg_info_off = padToIdeal(self.dbgInfoNeededHeaderBytes());
    }
}

fn writeDeclDebugInfo(self: *Elf, text_block: *TextBlock, dbg_info_buf: []const u8) !void {
    const tracy = trace(@src());
    defer tracy.end();

    // This logic is nearly identical to the logic above in `updateDecl` for
    // `SrcFn` and the line number programs. If you are editing this logic, you
    // probably need to edit that logic too.

    const debug_info_sect = &self.sections.items[self.debug_info_section_index.?];

    const last_decl = self.dbg_info_decl_last.?;
    // +1 for a trailing zero to end the children of the decl tag.
    const needed_size = last_decl.dbg_info_off + last_decl.dbg_info_len + 1;
    if (needed_size != debug_info_sect.sh_size) {
        if (needed_size > self.allocatedSize(debug_info_sect.sh_offset)) {
            const new_offset = self.findFreeSpace(needed_size, 1);
            const existing_size = last_decl.dbg_info_off;
            log.debug("moving .debug_info section: {} bytes from 0x{x} to 0x{x}", .{
                existing_size,
                debug_info_sect.sh_offset,
                new_offset,
            });
            const amt = try self.base.file.?.copyRangeAll(debug_info_sect.sh_offset, self.base.file.?, new_offset, existing_size);
            if (amt != existing_size) return error.InputOutput;
            debug_info_sect.sh_offset = new_offset;
        }
        debug_info_sect.sh_size = needed_size;
        self.shdr_table_dirty = true; // TODO look into making only the one section dirty
        self.debug_info_header_dirty = true;
    }
    const prev_padding_size: u32 = if (text_block.dbg_info_prev) |prev|
        text_block.dbg_info_off - (prev.dbg_info_off + prev.dbg_info_len)
    else
        0;
    const next_padding_size: u32 = if (text_block.dbg_info_next) |next|
        next.dbg_info_off - (text_block.dbg_info_off + text_block.dbg_info_len)
    else
        0;

    // To end the children of the decl tag.
    const trailing_zero = text_block.dbg_info_next == null;

    // We only have support for one compilation unit so far, so the offsets are directly
    // from the .debug_info section.
    const file_pos = debug_info_sect.sh_offset + text_block.dbg_info_off;
    try self.pwriteDbgInfoNops(prev_padding_size, dbg_info_buf, next_padding_size, trailing_zero, file_pos);
}

pub fn updateDeclExports(
    self: *Elf,
    module: *Module,
    decl: *Module.Decl,
    exports: []const *Module.Export,
) !void {
    if (build_options.skip_non_native and builtin.object_format != .elf) {
        @panic("Attempted to compile for object format that was disabled by build configuration");
    }
    if (build_options.have_llvm) {
        if (self.llvm_object) |llvm_object| return llvm_object.updateDeclExports(module, decl, exports);
    }

    const tracy = trace(@src());
    defer tracy.end();

    try self.global_symbols.ensureUnusedCapacity(self.base.allocator, exports.len);
    if (decl.link.elf.local_sym_index == 0) return;
    const decl_sym = self.local_symbols.items[decl.link.elf.local_sym_index];

    for (exports) |exp| {
        if (exp.options.section) |section_name| {
            if (!mem.eql(u8, section_name, ".text")) {
                try module.failed_exports.ensureUnusedCapacity(module.gpa, 1);
                module.failed_exports.putAssumeCapacityNoClobber(
                    exp,
                    try Module.ErrorMsg.create(self.base.allocator, decl.srcLoc(), "Unimplemented: ExportOptions.section", .{}),
                );
                continue;
            }
        }
        const stb_bits: u8 = switch (exp.options.linkage) {
            .Internal => elf.STB_LOCAL,
            .Strong => blk: {
                if (mem.eql(u8, exp.options.name, "_start")) {
                    self.entry_addr = decl_sym.st_value;
                }
                break :blk elf.STB_GLOBAL;
            },
            .Weak => elf.STB_WEAK,
            .LinkOnce => {
                try module.failed_exports.ensureUnusedCapacity(module.gpa, 1);
                module.failed_exports.putAssumeCapacityNoClobber(
                    exp,
                    try Module.ErrorMsg.create(self.base.allocator, decl.srcLoc(), "Unimplemented: GlobalLinkage.LinkOnce", .{}),
                );
                continue;
            },
        };
        const stt_bits: u8 = @truncate(u4, decl_sym.st_info);
        if (exp.link.elf.sym_index) |i| {
            const sym = &self.global_symbols.items[i];
            sym.* = .{
                .st_name = try self.updateString(sym.st_name, exp.options.name),
                .st_info = (stb_bits << 4) | stt_bits,
                .st_other = 0,
                .st_shndx = self.text_section_index.?,
                .st_value = decl_sym.st_value,
                .st_size = decl_sym.st_size,
            };
        } else {
            const name = try self.makeString(exp.options.name);
            const i = if (self.global_symbol_free_list.popOrNull()) |i| i else blk: {
                _ = self.global_symbols.addOneAssumeCapacity();
                break :blk self.global_symbols.items.len - 1;
            };
            self.global_symbols.items[i] = .{
                .st_name = name,
                .st_info = (stb_bits << 4) | stt_bits,
                .st_other = 0,
                .st_shndx = self.text_section_index.?,
                .st_value = decl_sym.st_value,
                .st_size = decl_sym.st_size,
            };

            exp.link.elf.sym_index = @intCast(u32, i);
        }
    }
}

/// Must be called only after a successful call to `updateDecl`.
pub fn updateDeclLineNumber(self: *Elf, module: *Module, decl: *const Module.Decl) !void {
    _ = module;
    const tracy = trace(@src());
    defer tracy.end();

    if (self.llvm_object) |_| return;

    const func = decl.val.castTag(.function).?.data;
    const casted_line_off = @intCast(u28, decl.src_line + func.lbrace_line);

    const shdr = &self.sections.items[self.debug_line_section_index.?];
    const file_pos = shdr.sh_offset + decl.fn_link.elf.off + self.getRelocDbgLineOff();
    var data: [4]u8 = undefined;
    leb128.writeUnsignedFixed(4, &data, casted_line_off);
    try self.base.file.?.pwriteAll(&data, file_pos);
}

pub fn deleteExport(self: *Elf, exp: Export) void {
    if (self.llvm_object) |_| return;

    const sym_index = exp.sym_index orelse return;
    self.global_symbol_free_list.append(self.base.allocator, sym_index) catch {};
    self.global_symbols.items[sym_index].st_info = 0;
}

fn writeProgHeader(self: *Elf, index: usize) !void {
    const foreign_endian = self.base.options.target.cpu.arch.endian() != builtin.cpu.arch.endian();
    const offset = self.program_headers.items[index].p_offset;
    switch (self.ptr_width) {
        .p32 => {
            var phdr = [1]elf.Elf32_Phdr{progHeaderTo32(self.program_headers.items[index])};
            if (foreign_endian) {
                mem.bswapAllFields(elf.Elf32_Phdr, &phdr[0]);
            }
            return self.base.file.?.pwriteAll(mem.sliceAsBytes(&phdr), offset);
        },
        .p64 => {
            var phdr = [1]elf.Elf64_Phdr{self.program_headers.items[index]};
            if (foreign_endian) {
                mem.bswapAllFields(elf.Elf64_Phdr, &phdr[0]);
            }
            return self.base.file.?.pwriteAll(mem.sliceAsBytes(&phdr), offset);
        },
    }
}

fn writeSectHeader(self: *Elf, index: usize) !void {
    const foreign_endian = self.base.options.target.cpu.arch.endian() != builtin.cpu.arch.endian();
    switch (self.ptr_width) {
        .p32 => {
            var shdr: [1]elf.Elf32_Shdr = undefined;
            shdr[0] = sectHeaderTo32(self.sections.items[index]);
            if (foreign_endian) {
                mem.bswapAllFields(elf.Elf32_Shdr, &shdr[0]);
            }
            const offset = self.shdr_table_offset.? + index * @sizeOf(elf.Elf32_Shdr);
            return self.base.file.?.pwriteAll(mem.sliceAsBytes(&shdr), offset);
        },
        .p64 => {
            var shdr = [1]elf.Elf64_Shdr{self.sections.items[index]};
            if (foreign_endian) {
                mem.bswapAllFields(elf.Elf64_Shdr, &shdr[0]);
            }
            const offset = self.shdr_table_offset.? + index * @sizeOf(elf.Elf64_Shdr);
            return self.base.file.?.pwriteAll(mem.sliceAsBytes(&shdr), offset);
        },
    }
}

fn writeOffsetTableEntry(self: *Elf, index: usize) !void {
    const shdr = &self.sections.items[self.got_section_index.?];
    const phdr = &self.program_headers.items[self.phdr_got_index.?];
    const entry_size: u16 = self.archPtrWidthBytes();
    if (self.offset_table_count_dirty) {
        // TODO Also detect virtual address collisions.
        const allocated_size = self.allocatedSize(shdr.sh_offset);
        const needed_size = self.offset_table.items.len * entry_size;
        if (needed_size > allocated_size) {
            // Must move the entire got section.
            const new_offset = self.findFreeSpace(needed_size, entry_size);
            const amt = try self.base.file.?.copyRangeAll(shdr.sh_offset, self.base.file.?, new_offset, shdr.sh_size);
            if (amt != shdr.sh_size) return error.InputOutput;
            shdr.sh_offset = new_offset;
            phdr.p_offset = new_offset;
        }
        shdr.sh_size = needed_size;
        phdr.p_memsz = needed_size;
        phdr.p_filesz = needed_size;

        self.shdr_table_dirty = true; // TODO look into making only the one section dirty
        self.phdr_table_dirty = true; // TODO look into making only the one program header dirty

        self.offset_table_count_dirty = false;
    }
    const endian = self.base.options.target.cpu.arch.endian();
    const off = shdr.sh_offset + @as(u64, entry_size) * index;
    switch (entry_size) {
        2 => {
            var buf: [2]u8 = undefined;
            mem.writeInt(u16, &buf, @intCast(u16, self.offset_table.items[index]), endian);
            try self.base.file.?.pwriteAll(&buf, off);
        },
        4 => {
            var buf: [4]u8 = undefined;
            mem.writeInt(u32, &buf, @intCast(u32, self.offset_table.items[index]), endian);
            try self.base.file.?.pwriteAll(&buf, off);
        },
        8 => {
            var buf: [8]u8 = undefined;
            mem.writeInt(u64, &buf, self.offset_table.items[index], endian);
            try self.base.file.?.pwriteAll(&buf, off);
        },
        else => unreachable,
    }
}

fn writeSymbol(self: *Elf, index: usize) !void {
    const tracy = trace(@src());
    defer tracy.end();

    const syms_sect = &self.sections.items[self.symtab_section_index.?];
    // Make sure we are not pointlessly writing symbol data that will have to get relocated
    // due to running out of space.
    if (self.local_symbols.items.len != syms_sect.sh_info) {
        const sym_size: u64 = switch (self.ptr_width) {
            .p32 => @sizeOf(elf.Elf32_Sym),
            .p64 => @sizeOf(elf.Elf64_Sym),
        };
        const sym_align: u16 = switch (self.ptr_width) {
            .p32 => @alignOf(elf.Elf32_Sym),
            .p64 => @alignOf(elf.Elf64_Sym),
        };
        const needed_size = (self.local_symbols.items.len + self.global_symbols.items.len) * sym_size;
        if (needed_size > self.allocatedSize(syms_sect.sh_offset)) {
            // Move all the symbols to a new file location.
            const new_offset = self.findFreeSpace(needed_size, sym_align);
            const existing_size = @as(u64, syms_sect.sh_info) * sym_size;
            const amt = try self.base.file.?.copyRangeAll(syms_sect.sh_offset, self.base.file.?, new_offset, existing_size);
            if (amt != existing_size) return error.InputOutput;
            syms_sect.sh_offset = new_offset;
        }
        syms_sect.sh_info = @intCast(u32, self.local_symbols.items.len);
        syms_sect.sh_size = needed_size; // anticipating adding the global symbols later
        self.shdr_table_dirty = true; // TODO look into only writing one section
    }
    const foreign_endian = self.base.options.target.cpu.arch.endian() != builtin.cpu.arch.endian();
    switch (self.ptr_width) {
        .p32 => {
            var sym = [1]elf.Elf32_Sym{
                .{
                    .st_name = self.local_symbols.items[index].st_name,
                    .st_value = @intCast(u32, self.local_symbols.items[index].st_value),
                    .st_size = @intCast(u32, self.local_symbols.items[index].st_size),
                    .st_info = self.local_symbols.items[index].st_info,
                    .st_other = self.local_symbols.items[index].st_other,
                    .st_shndx = self.local_symbols.items[index].st_shndx,
                },
            };
            if (foreign_endian) {
                mem.bswapAllFields(elf.Elf32_Sym, &sym[0]);
            }
            const off = syms_sect.sh_offset + @sizeOf(elf.Elf32_Sym) * index;
            try self.base.file.?.pwriteAll(mem.sliceAsBytes(sym[0..1]), off);
        },
        .p64 => {
            var sym = [1]elf.Elf64_Sym{self.local_symbols.items[index]};
            if (foreign_endian) {
                mem.bswapAllFields(elf.Elf64_Sym, &sym[0]);
            }
            const off = syms_sect.sh_offset + @sizeOf(elf.Elf64_Sym) * index;
            try self.base.file.?.pwriteAll(mem.sliceAsBytes(sym[0..1]), off);
        },
    }
}

fn writeAllGlobalSymbols(self: *Elf) !void {
    const syms_sect = &self.sections.items[self.symtab_section_index.?];
    const sym_size: u64 = switch (self.ptr_width) {
        .p32 => @sizeOf(elf.Elf32_Sym),
        .p64 => @sizeOf(elf.Elf64_Sym),
    };
    const foreign_endian = self.base.options.target.cpu.arch.endian() != builtin.cpu.arch.endian();
    const global_syms_off = syms_sect.sh_offset + self.local_symbols.items.len * sym_size;
    switch (self.ptr_width) {
        .p32 => {
            const buf = try self.base.allocator.alloc(elf.Elf32_Sym, self.global_symbols.items.len);
            defer self.base.allocator.free(buf);

            for (buf) |*sym, i| {
                sym.* = .{
                    .st_name = self.global_symbols.items[i].st_name,
                    .st_value = @intCast(u32, self.global_symbols.items[i].st_value),
                    .st_size = @intCast(u32, self.global_symbols.items[i].st_size),
                    .st_info = self.global_symbols.items[i].st_info,
                    .st_other = self.global_symbols.items[i].st_other,
                    .st_shndx = self.global_symbols.items[i].st_shndx,
                };
                if (foreign_endian) {
                    mem.bswapAllFields(elf.Elf32_Sym, sym);
                }
            }
            try self.base.file.?.pwriteAll(mem.sliceAsBytes(buf), global_syms_off);
        },
        .p64 => {
            const buf = try self.base.allocator.alloc(elf.Elf64_Sym, self.global_symbols.items.len);
            defer self.base.allocator.free(buf);

            for (buf) |*sym, i| {
                sym.* = .{
                    .st_name = self.global_symbols.items[i].st_name,
                    .st_value = self.global_symbols.items[i].st_value,
                    .st_size = self.global_symbols.items[i].st_size,
                    .st_info = self.global_symbols.items[i].st_info,
                    .st_other = self.global_symbols.items[i].st_other,
                    .st_shndx = self.global_symbols.items[i].st_shndx,
                };
                if (foreign_endian) {
                    mem.bswapAllFields(elf.Elf64_Sym, sym);
                }
            }
            try self.base.file.?.pwriteAll(mem.sliceAsBytes(buf), global_syms_off);
        },
    }
}

/// Always 4 or 8 depending on whether this is 32-bit ELF or 64-bit ELF.
fn ptrWidthBytes(self: Elf) u8 {
    return switch (self.ptr_width) {
        .p32 => 4,
        .p64 => 8,
    };
}

/// Does not necessarily match `ptrWidthBytes` for example can be 2 bytes
/// in a 32-bit ELF file.
fn archPtrWidthBytes(self: Elf) u8 {
    return @intCast(u8, self.base.options.target.cpu.arch.ptrBitWidth() / 8);
}

/// The reloc offset for the virtual address of a function in its Line Number Program.
/// Size is a virtual address integer.
const dbg_line_vaddr_reloc_index = 3;
/// The reloc offset for the virtual address of a function in its .debug_info TAG.subprogram.
/// Size is a virtual address integer.
const dbg_info_low_pc_reloc_index = 1;

/// The reloc offset for the line offset of a function from the previous function's line.
/// It's a fixed-size 4-byte ULEB128.
fn getRelocDbgLineOff(self: Elf) usize {
    return dbg_line_vaddr_reloc_index + self.ptrWidthBytes() + 1;
}

fn getRelocDbgFileIndex(self: Elf) usize {
    return self.getRelocDbgLineOff() + 5;
}

fn getRelocDbgInfoSubprogramHighPC(self: Elf) u32 {
    return dbg_info_low_pc_reloc_index + self.ptrWidthBytes();
}

fn dbgLineNeededHeaderBytes(self: Elf) u32 {
    const directory_entry_format_count = 1;
    const file_name_entry_format_count = 1;
    const directory_count = 1;
    const file_name_count = 1;
    const root_src_dir_path_len = if (self.base.options.module.?.root_pkg.root_src_directory.path) |p| p.len else 1; // "."
    return @intCast(u32, 53 + directory_entry_format_count * 2 + file_name_entry_format_count * 2 +
        directory_count * 8 + file_name_count * 8 +
        // These are encoded as DW.FORM.string rather than DW.FORM.strp as we would like
        // because of a workaround for readelf and gdb failing to understand DWARFv5 correctly.
        root_src_dir_path_len +
        self.base.options.module.?.root_pkg.root_src_path.len);
}

fn dbgInfoNeededHeaderBytes(self: Elf) u32 {
    _ = self;
    return 120;
}

const min_nop_size = 2;

/// Writes to the file a buffer, prefixed and suffixed by the specified number of
/// bytes of NOPs. Asserts each padding size is at least `min_nop_size` and total padding bytes
/// are less than 1044480 bytes (if this limit is ever reached, this function can be
/// improved to make more than one pwritev call, or the limit can be raised by a fixed
/// amount by increasing the length of `vecs`).
fn pwriteDbgLineNops(
    self: *Elf,
    prev_padding_size: usize,
    buf: []const u8,
    next_padding_size: usize,
    offset: u64,
) !void {
    const tracy = trace(@src());
    defer tracy.end();

    const page_of_nops = [1]u8{DW.LNS.negate_stmt} ** 4096;
    const three_byte_nop = [3]u8{ DW.LNS.advance_pc, 0b1000_0000, 0 };
    var vecs: [512]std.os.iovec_const = undefined;
    var vec_index: usize = 0;
    {
        var padding_left = prev_padding_size;
        if (padding_left % 2 != 0) {
            vecs[vec_index] = .{
                .iov_base = &three_byte_nop,
                .iov_len = three_byte_nop.len,
            };
            vec_index += 1;
            padding_left -= three_byte_nop.len;
        }
        while (padding_left > page_of_nops.len) {
            vecs[vec_index] = .{
                .iov_base = &page_of_nops,
                .iov_len = page_of_nops.len,
            };
            vec_index += 1;
            padding_left -= page_of_nops.len;
        }
        if (padding_left > 0) {
            vecs[vec_index] = .{
                .iov_base = &page_of_nops,
                .iov_len = padding_left,
            };
            vec_index += 1;
        }
    }

    vecs[vec_index] = .{
        .iov_base = buf.ptr,
        .iov_len = buf.len,
    };
    vec_index += 1;

    {
        var padding_left = next_padding_size;
        if (padding_left % 2 != 0) {
            vecs[vec_index] = .{
                .iov_base = &three_byte_nop,
                .iov_len = three_byte_nop.len,
            };
            vec_index += 1;
            padding_left -= three_byte_nop.len;
        }
        while (padding_left > page_of_nops.len) {
            vecs[vec_index] = .{
                .iov_base = &page_of_nops,
                .iov_len = page_of_nops.len,
            };
            vec_index += 1;
            padding_left -= page_of_nops.len;
        }
        if (padding_left > 0) {
            vecs[vec_index] = .{
                .iov_base = &page_of_nops,
                .iov_len = padding_left,
            };
            vec_index += 1;
        }
    }
    try self.base.file.?.pwritevAll(vecs[0..vec_index], offset - prev_padding_size);
}

/// Writes to the file a buffer, prefixed and suffixed by the specified number of
/// bytes of padding.
fn pwriteDbgInfoNops(
    self: *Elf,
    prev_padding_size: usize,
    buf: []const u8,
    next_padding_size: usize,
    trailing_zero: bool,
    offset: u64,
) !void {
    const tracy = trace(@src());
    defer tracy.end();

    const page_of_nops = [1]u8{abbrev_pad1} ** 4096;
    var vecs: [32]std.os.iovec_const = undefined;
    var vec_index: usize = 0;
    {
        var padding_left = prev_padding_size;
        while (padding_left > page_of_nops.len) {
            vecs[vec_index] = .{
                .iov_base = &page_of_nops,
                .iov_len = page_of_nops.len,
            };
            vec_index += 1;
            padding_left -= page_of_nops.len;
        }
        if (padding_left > 0) {
            vecs[vec_index] = .{
                .iov_base = &page_of_nops,
                .iov_len = padding_left,
            };
            vec_index += 1;
        }
    }

    vecs[vec_index] = .{
        .iov_base = buf.ptr,
        .iov_len = buf.len,
    };
    vec_index += 1;

    {
        var padding_left = next_padding_size;
        while (padding_left > page_of_nops.len) {
            vecs[vec_index] = .{
                .iov_base = &page_of_nops,
                .iov_len = page_of_nops.len,
            };
            vec_index += 1;
            padding_left -= page_of_nops.len;
        }
        if (padding_left > 0) {
            vecs[vec_index] = .{
                .iov_base = &page_of_nops,
                .iov_len = padding_left,
            };
            vec_index += 1;
        }
    }

    if (trailing_zero) {
        var zbuf = [1]u8{0};
        vecs[vec_index] = .{
            .iov_base = &zbuf,
            .iov_len = zbuf.len,
        };
        vec_index += 1;
    }

    try self.base.file.?.pwritevAll(vecs[0..vec_index], offset - prev_padding_size);
}

fn progHeaderTo32(phdr: elf.Elf64_Phdr) elf.Elf32_Phdr {
    return .{
        .p_type = phdr.p_type,
        .p_flags = phdr.p_flags,
        .p_offset = @intCast(u32, phdr.p_offset),
        .p_vaddr = @intCast(u32, phdr.p_vaddr),
        .p_paddr = @intCast(u32, phdr.p_paddr),
        .p_filesz = @intCast(u32, phdr.p_filesz),
        .p_memsz = @intCast(u32, phdr.p_memsz),
        .p_align = @intCast(u32, phdr.p_align),
    };
}

fn sectHeaderTo32(shdr: elf.Elf64_Shdr) elf.Elf32_Shdr {
    return .{
        .sh_name = shdr.sh_name,
        .sh_type = shdr.sh_type,
        .sh_flags = @intCast(u32, shdr.sh_flags),
        .sh_addr = @intCast(u32, shdr.sh_addr),
        .sh_offset = @intCast(u32, shdr.sh_offset),
        .sh_size = @intCast(u32, shdr.sh_size),
        .sh_link = shdr.sh_link,
        .sh_info = shdr.sh_info,
        .sh_addralign = @intCast(u32, shdr.sh_addralign),
        .sh_entsize = @intCast(u32, shdr.sh_entsize),
    };
}

fn getLDMOption(target: std.Target) ?[]const u8 {
    switch (target.cpu.arch) {
        .i386 => return "elf_i386",
        .aarch64 => return "aarch64linux",
        .aarch64_be => return "aarch64_be_linux",
        .arm, .thumb => return "armelf_linux_eabi",
        .armeb, .thumbeb => return "armebelf_linux_eabi",
        .powerpc => return "elf32ppclinux",
        .powerpc64 => return "elf64ppc",
        .powerpc64le => return "elf64lppc",
        .sparc, .sparcel => return "elf32_sparc",
        .sparcv9 => return "elf64_sparc",
        .mips => return "elf32btsmip",
        .mipsel => return "elf32ltsmip",
        .mips64 => {
            if (target.abi == .gnuabin32) {
                return "elf32btsmipn32";
            } else {
                return "elf64btsmip";
            }
        },
        .mips64el => {
            if (target.abi == .gnuabin32) {
                return "elf32ltsmipn32";
            } else {
                return "elf64ltsmip";
            }
        },
        .s390x => return "elf64_s390",
        .x86_64 => {
            if (target.abi == .gnux32) {
                return "elf32_x86_64";
            } else {
                return "elf_x86_64";
            }
        },
        .riscv32 => return "elf32lriscv",
        .riscv64 => return "elf64lriscv",
        else => return null,
    }
}

fn padToIdeal(actual_size: anytype) @TypeOf(actual_size) {
    // TODO https://github.com/ziglang/zig/issues/1284
    return std.math.add(@TypeOf(actual_size), actual_size, actual_size / ideal_factor) catch
        std.math.maxInt(@TypeOf(actual_size));
}

// Provide a blueprint of csu (c-runtime startup) objects for supported
// link modes.
//
// This is for cross-mode targets only. For host-mode targets the system
// compiler can be probed to produce a robust blueprint.
//
// Targets requiring a libc for which zig does not bundle a libc are
// host-mode targets. Unfortunately, host-mode probes are not yet
// implemented. For now the data is hard-coded here. Such targets are
// { freebsd, netbsd, openbsd, dragonfly }.
const CsuObjects = struct {
    crt0: ?[]const u8 = null,
    crti: ?[]const u8 = null,
    crtbegin: ?[]const u8 = null,
    crtend: ?[]const u8 = null,
    crtn: ?[]const u8 = null,

    fn init(arena: *mem.Allocator, link_options: link.Options, comp: *const Compilation) !CsuObjects {
        // crt objects are only required for libc.
        if (!link_options.link_libc) return CsuObjects{};

        var result: CsuObjects = .{};

        // TODO: https://github.com/ziglang/zig/issues/4629
        // - use inline enum type
        // - reduce to enum-literals for values
        const Mode = enum {
            dynamic_lib,
            dynamic_exe,
            dynamic_pie,
            static_exe,
            static_pie,
        };

        // Flatten crt case types.
        const mode: Mode = switch (link_options.output_mode) {
            .Obj => return CsuObjects{},
            .Lib => switch (link_options.link_mode) {
                .Dynamic => Mode.dynamic_lib,
                .Static => return CsuObjects{},
            },
            .Exe => switch (link_options.link_mode) {
                .Dynamic => if (link_options.pie) Mode.dynamic_pie else Mode.dynamic_exe,
                .Static => if (link_options.pie) Mode.static_pie else Mode.static_exe,
            },
        };

        if (link_options.target.isAndroid()) {
            switch (mode) {
                // zig fmt: off
                .dynamic_lib => result.set( null, null, "crtbegin_so.o",      "crtend_so.o",      null ),
                .dynamic_exe,
                .dynamic_pie => result.set( null, null, "crtbegin_dynamic.o", "crtend_android.o", null ),
                .static_exe,
                .static_pie  => result.set( null, null, "crtbegin_static.o",  "crtend_android.o", null ),
                // zig fmt: on
            }
        } else {
            switch (link_options.target.os.tag) {
                .linux => {
                    switch (mode) {
                        // zig fmt: off
                        .dynamic_lib => result.set( null,      "crti.o", "crtbeginS.o", "crtendS.o", "crtn.o" ),
                        .dynamic_exe => result.set( "crt1.o",  "crti.o", "crtbegin.o",  "crtend.o",  "crtn.o" ),
                        .dynamic_pie => result.set( "Scrt1.o", "crti.o", "crtbeginS.o", "crtendS.o", "crtn.o" ),
                        .static_exe  => result.set( "crt1.o",  "crti.o", "crtbeginT.o", "crtend.o",  "crtn.o" ),
                        .static_pie  => result.set( "rcrt1.o", "crti.o", "crtbeginS.o", "crtendS.o", "crtn.o" ),
                        // zig fmt: on
                    }
                    if (link_options.libc_installation) |_| {
                        // hosted-glibc provides crtbegin/end objects in platform/compiler-specific dirs
                        // and they are not known at comptime. For now null-out crtbegin/end objects;
                        // there is no feature loss, zig has never linked those objects in before.
                        // TODO: probe for paths, ie. `cc -print-file-name`
                        result.crtbegin = null;
                        result.crtend = null;
                    } else {
                        // Bundled glibc only has Scrt1.o .
                        if (result.crt0 != null and link_options.target.isGnuLibC()) result.crt0 = "Scrt1.o";
                    }
                },
                .dragonfly => switch (mode) {
                    // zig fmt: off
                    .dynamic_lib => result.set( null,      "crti.o", "crtbeginS.o",  "crtendS.o", "crtn.o" ),
                    .dynamic_exe => result.set( "crt1.o",  "crti.o", "crtbegin.o",   "crtend.o",  "crtn.o" ),
                    .dynamic_pie => result.set( "Scrt1.o", "crti.o", "crtbeginS.o",  "crtendS.o", "crtn.o" ),
                    .static_exe  => result.set( "crt1.o",  "crti.o", "crtbegin.o",   "crtend.o",  "crtn.o" ),
                    .static_pie  => result.set( "Scrt1.o", "crti.o", "crtbeginS.o",  "crtendS.o", "crtn.o" ),
                    // zig fmt: on
                },
                .freebsd => switch (mode) {
                    // zig fmt: off
                    .dynamic_lib => result.set( null,      "crti.o", "crtbeginS.o", "crtendS.o", "crtn.o" ),
                    .dynamic_exe => result.set( "crt1.o",  "crti.o", "crtbegin.o",  "crtend.o",  "crtn.o" ),
                    .dynamic_pie => result.set( "Scrt1.o", "crti.o", "crtbeginS.o", "crtendS.o", "crtn.o" ),
                    .static_exe  => result.set( "crt1.o",  "crti.o", "crtbeginT.o", "crtend.o",  "crtn.o" ),
                    .static_pie  => result.set( "Scrt1.o", "crti.o", "crtbeginS.o", "crtendS.o", "crtn.o" ),
                    // zig fmt: on
                },
                .netbsd => switch (mode) {
                    // zig fmt: off
                    .dynamic_lib => result.set( null,     "crti.o", "crtbeginS.o", "crtendS.o", "crtn.o" ),
                    .dynamic_exe => result.set( "crt0.o", "crti.o", "crtbegin.o",  "crtend.o",  "crtn.o" ),
                    .dynamic_pie => result.set( "crt0.o", "crti.o", "crtbeginS.o", "crtendS.o", "crtn.o" ),
                    .static_exe  => result.set( "crt0.o", "crti.o", "crtbeginT.o", "crtend.o",  "crtn.o" ),
                    .static_pie  => result.set( "crt0.o", "crti.o", "crtbeginT.o", "crtendS.o", "crtn.o" ),
                    // zig fmt: on
                },
                .openbsd => switch (mode) {
                    // zig fmt: off
                    .dynamic_lib => result.set( null,      null, "crtbeginS.o", "crtendS.o", null ),
                    .dynamic_exe,
                    .dynamic_pie => result.set( "crt0.o",  null, "crtbegin.o",  "crtend.o",  null ),
                    .static_exe,
                    .static_pie  => result.set( "rcrt0.o", null, "crtbegin.o",  "crtend.o",  null ),
                    // zig fmt: on
                },
                .haiku => switch (mode) {
                    // zig fmt: off
                    .dynamic_lib => result.set( null,          "crti.o", "crtbeginS.o", "crtendS.o", "crtn.o" ),
                    .dynamic_exe => result.set( "start_dyn.o", "crti.o", "crtbegin.o",  "crtend.o",  "crtn.o" ),
                    .dynamic_pie => result.set( "start_dyn.o", "crti.o", "crtbeginS.o", "crtendS.o", "crtn.o" ),
                    .static_exe  => result.set( "start_dyn.o", "crti.o", "crtbegin.o",  "crtend.o",  "crtn.o" ),
                    .static_pie  => result.set( "start_dyn.o", "crti.o", "crtbeginS.o", "crtendS.o", "crtn.o" ),
                    // zig fmt: on
                },
                .solaris => switch (mode) {
                    // zig fmt: off
                    .dynamic_lib => result.set( null,     "crti.o", null, null, "crtn.o" ),
                    .dynamic_exe,
                    .dynamic_pie => result.set( "crt1.o", "crti.o", null, null, "crtn.o" ),
                    .static_exe,
                    .static_pie  => result.set( null,     null,     null, null, null     ),
                    // zig fmt: on
                },
                else => {},
            }
        }

        // Convert each object to a full pathname.
        if (link_options.libc_installation) |lci| {
            const crt_dir_path = lci.crt_dir orelse return error.LibCInstallationMissingCRTDir;
            switch (link_options.target.os.tag) {
                .dragonfly => {
                    if (result.crt0) |*obj| obj.* = try fs.path.join(arena, &[_][]const u8{ crt_dir_path, obj.* });
                    if (result.crti) |*obj| obj.* = try fs.path.join(arena, &[_][]const u8{ crt_dir_path, obj.* });
                    if (result.crtn) |*obj| obj.* = try fs.path.join(arena, &[_][]const u8{ crt_dir_path, obj.* });

                    var gccv: []const u8 = undefined;
                    if (link_options.target.os.version_range.semver.isAtLeast(.{ .major = 5, .minor = 4 }) orelse true) {
                        gccv = "gcc80";
                    } else {
                        gccv = "gcc54";
                    }

                    if (result.crtbegin) |*obj| obj.* = try fs.path.join(arena, &[_][]const u8{ crt_dir_path, gccv, obj.* });
                    if (result.crtend) |*obj| obj.* = try fs.path.join(arena, &[_][]const u8{ crt_dir_path, gccv, obj.* });
                },
                .haiku => {
                    const gcc_dir_path = lci.gcc_dir orelse return error.LibCInstallationMissingCRTDir;
                    if (result.crt0) |*obj| obj.* = try fs.path.join(arena, &[_][]const u8{ crt_dir_path, obj.* });
                    if (result.crti) |*obj| obj.* = try fs.path.join(arena, &[_][]const u8{ crt_dir_path, obj.* });
                    if (result.crtn) |*obj| obj.* = try fs.path.join(arena, &[_][]const u8{ crt_dir_path, obj.* });

                    if (result.crtbegin) |*obj| obj.* = try fs.path.join(arena, &[_][]const u8{ gcc_dir_path, obj.* });
                    if (result.crtend) |*obj| obj.* = try fs.path.join(arena, &[_][]const u8{ gcc_dir_path, obj.* });
                },
                else => {
                    inline for (std.meta.fields(@TypeOf(result))) |f| {
                        if (@field(result, f.name)) |*obj| {
                            obj.* = try fs.path.join(arena, &[_][]const u8{ crt_dir_path, obj.* });
                        }
                    }
                },
            }
        } else {
            inline for (std.meta.fields(@TypeOf(result))) |f| {
                if (@field(result, f.name)) |*obj| {
                    if (comp.crt_files.get(obj.*)) |crtf| {
                        obj.* = crtf.full_object_path;
                    } else {
                        @field(result, f.name) = null;
                    }
                }
            }
        }

        return result;
    }

    fn set(
        self: *CsuObjects,
        crt0: ?[]const u8,
        crti: ?[]const u8,
        crtbegin: ?[]const u8,
        crtend: ?[]const u8,
        crtn: ?[]const u8,
    ) void {
        self.crt0 = crt0;
        self.crti = crti;
        self.crtbegin = crtbegin;
        self.crtend = crtend;
        self.crtn = crtn;
    }
};