macho: incrementally write dyld_private and stub_helper atoms

By incrementally I mean using the incremental linker machinery
and concepts. Currently, lots of repetition but already highlighted
a potential problem with resolving relocations for symbols that
weren't seen yet but wanting to write the atom to file (before
seeing the relevant atoms).
This commit is contained in:
Jakub Konka 2021-08-21 22:26:02 +02:00
parent d61d85abd2
commit ea499203fe

View File

@ -764,8 +764,14 @@ pub fn flush(self: *MachO, comp: *Compilation) !void {
try self.createStubHelperPreambleAtom();
if (!use_stage1) {
// TODO this should be made common when I figure out how to prealloc space for traditional linker path.
try self.writeStubHelperPreamble();
// TODO just a temp
const seg = self.load_commands.items[self.text_segment_cmd_index.?].Segment;
const sect = seg.sections.items[self.stub_helper_section_index.?];
self.stub_helper_stubs_start_off = sect.offset + switch (self.base.options.target.cpu.arch) {
.x86_64 => @intCast(u64, 15),
.aarch64 => @intCast(u64, 6 * @sizeOf(u32)),
else => unreachable,
};
}
// Apply pending updates
@ -842,7 +848,6 @@ pub fn flushModule(self: *MachO, comp: *Compilation) !void {
defer tracy.end();
try self.setEntryPoint();
try self.writeTextBlocks();
try self.writeRebaseInfoTable();
try self.writeBindInfoTable();
try self.writeLazyBindInfoTable();
@ -1989,7 +1994,29 @@ fn createDyldPrivateAtom(self: *MachO) !void {
.n_desc = 0,
.n_value = 0,
});
_ = try self.createEmptyAtom(match, local_sym_index, @sizeOf(u64), 3);
const last = self.blocks.get(match);
const atom = try self.createEmptyAtom(match, local_sym_index, @sizeOf(u64), 3);
if (!(build_options.is_stage1 and self.base.options.use_stage1)) {
const seg = self.load_commands.items[match.seg].Segment;
const sect = seg.sections.items[match.sect];
const base_addr = if (last) |last_atom| blk: {
const last_atom_sym = self.locals.items[last_atom.local_sym_index];
break :blk last_atom_sym.n_value;
} else sect.addr;
const n_sect = @intCast(u8, self.section_ordinals.getIndex(match).? + 1);
const atom_alignment = try math.powi(u32, 2, atom.alignment);
const vaddr = mem.alignForwardGeneric(u64, base_addr, atom_alignment);
const sym = &self.locals.items[local_sym_index];
sym.n_value = vaddr;
sym.n_sect = n_sect;
const file_offset = sect.offset + vaddr - sect.addr;
log.debug("writing code for symbol {s} at file offset 0x{x}", .{ self.getString(sym.n_strx), file_offset });
try self.base.file.?.pwriteAll(atom.code.items, file_offset);
try self.writeLocalSymbol(local_sym_index);
}
self.dyld_private_sym_index = local_sym_index;
}
@ -2019,6 +2046,7 @@ fn createStubHelperPreambleAtom(self: *MachO) !void {
.n_desc = 0,
.n_value = 0,
});
const last = self.blocks.get(match);
const atom = try self.createEmptyAtom(match, local_sym_index, size, alignment);
switch (arch) {
.x86_64 => {
@ -2130,6 +2158,29 @@ fn createStubHelperPreambleAtom(self: *MachO) !void {
}
self.stub_preamble_sym_index = local_sym_index;
if (!(build_options.is_stage1 and self.base.options.use_stage1)) {
const seg = self.load_commands.items[match.seg].Segment;
const sect = seg.sections.items[match.sect];
const base_addr = if (last) |last_atom| blk: {
const last_atom_sym = self.locals.items[last_atom.local_sym_index];
break :blk last_atom_sym.n_value;
} else sect.addr;
const n_sect = @intCast(u8, self.section_ordinals.getIndex(match).? + 1);
const atom_alignment = try math.powi(u32, 2, atom.alignment);
const vaddr = mem.alignForwardGeneric(u64, base_addr, atom_alignment);
const sym = &self.locals.items[local_sym_index];
sym.n_value = vaddr;
sym.n_sect = n_sect;
try atom.resolveRelocs(self);
const file_offset = sect.offset + vaddr - sect.addr;
log.debug("writing code for symbol {s} at file offset 0x{x}", .{ self.getString(sym.n_strx), file_offset });
try self.base.file.?.pwriteAll(atom.code.items, file_offset);
try self.writeLocalSymbol(local_sym_index);
}
// TODO this needs to be fixed
// We already prealloc stub helper size in populateMissingMetadata(), but
// perhaps it's not needed after all?
@ -2138,139 +2189,6 @@ fn createStubHelperPreambleAtom(self: *MachO) !void {
sect.size -= size;
}
fn writeStubHelperPreamble(self: *MachO) !void {
const text_segment = &self.load_commands.items[self.text_segment_cmd_index.?].Segment;
const stub_helper = &text_segment.sections.items[self.stub_helper_section_index.?];
const data_const_segment = &self.load_commands.items[self.data_const_segment_cmd_index.?].Segment;
const got = &data_const_segment.sections.items[self.got_section_index.?];
const data_segment = &self.load_commands.items[self.data_segment_cmd_index.?].Segment;
const data = &data_segment.sections.items[self.data_section_index.?];
self.stub_helper_stubs_start_off = blk: {
switch (self.base.options.target.cpu.arch) {
.x86_64 => {
const code_size = 15;
var code: [code_size]u8 = undefined;
// lea %r11, [rip + disp]
code[0] = 0x4c;
code[1] = 0x8d;
code[2] = 0x1d;
{
const target_addr = data.addr + data.size - @sizeOf(u64);
const displacement = try math.cast(u32, target_addr - stub_helper.addr - 7);
mem.writeIntLittle(u32, code[3..7], displacement);
}
// push %r11
code[7] = 0x41;
code[8] = 0x53;
// jmp [rip + disp]
code[9] = 0xff;
code[10] = 0x25;
{
const got_index = self.got_entries_map.get(.{
.where = .undef,
.where_index = self.dyld_stub_binder_index.?,
}) orelse unreachable;
const addr = got.addr + got_index * @sizeOf(u64);
const displacement = try math.cast(u32, addr - stub_helper.addr - code_size);
mem.writeIntLittle(u32, code[11..], displacement);
}
try self.base.file.?.pwriteAll(&code, stub_helper.offset);
break :blk stub_helper.offset + code_size;
},
.aarch64 => {
var code: [6 * @sizeOf(u32)]u8 = undefined;
data_blk_outer: {
const this_addr = stub_helper.addr;
const target_addr = data.addr + data.size - @sizeOf(u64);
data_blk: {
const displacement = math.cast(i21, target_addr - this_addr) catch break :data_blk;
// adr x17, disp
mem.writeIntLittle(u32, code[0..4], aarch64.Instruction.adr(.x17, displacement).toU32());
// nop
mem.writeIntLittle(u32, code[4..8], aarch64.Instruction.nop().toU32());
break :data_blk_outer;
}
data_blk: {
const new_this_addr = this_addr + @sizeOf(u32);
const displacement = math.cast(i21, target_addr - new_this_addr) catch break :data_blk;
// nop
mem.writeIntLittle(u32, code[0..4], aarch64.Instruction.nop().toU32());
// adr x17, disp
mem.writeIntLittle(u32, code[4..8], aarch64.Instruction.adr(.x17, displacement).toU32());
break :data_blk_outer;
}
// Jump is too big, replace adr with adrp and add.
const this_page = @intCast(i32, this_addr >> 12);
const target_page = @intCast(i32, target_addr >> 12);
const pages = @intCast(i21, target_page - this_page);
mem.writeIntLittle(u32, code[0..4], aarch64.Instruction.adrp(.x17, pages).toU32());
const narrowed = @truncate(u12, target_addr);
mem.writeIntLittle(u32, code[4..8], aarch64.Instruction.add(.x17, .x17, narrowed, false).toU32());
}
// stp x16, x17, [sp, #-16]!
code[8] = 0xf0;
code[9] = 0x47;
code[10] = 0xbf;
code[11] = 0xa9;
binder_blk_outer: {
const got_index = self.got_entries_map.get(.{
.where = .undef,
.where_index = self.dyld_stub_binder_index.?,
}) orelse unreachable;
const this_addr = stub_helper.addr + 3 * @sizeOf(u32);
const target_addr = got.addr + got_index * @sizeOf(u64);
binder_blk: {
const displacement = math.divExact(u64, target_addr - this_addr, 4) catch break :binder_blk;
const literal = math.cast(u18, displacement) catch break :binder_blk;
// ldr x16, label
mem.writeIntLittle(u32, code[12..16], aarch64.Instruction.ldr(.x16, .{
.literal = literal,
}).toU32());
// nop
mem.writeIntLittle(u32, code[16..20], aarch64.Instruction.nop().toU32());
break :binder_blk_outer;
}
binder_blk: {
const new_this_addr = this_addr + @sizeOf(u32);
const displacement = math.divExact(u64, target_addr - new_this_addr, 4) catch break :binder_blk;
const literal = math.cast(u18, displacement) catch break :binder_blk;
// Pad with nop to please division.
// nop
mem.writeIntLittle(u32, code[12..16], aarch64.Instruction.nop().toU32());
// ldr x16, label
mem.writeIntLittle(u32, code[16..20], aarch64.Instruction.ldr(.x16, .{
.literal = literal,
}).toU32());
break :binder_blk_outer;
}
// Use adrp followed by ldr(immediate).
const this_page = @intCast(i32, this_addr >> 12);
const target_page = @intCast(i32, target_addr >> 12);
const pages = @intCast(i21, target_page - this_page);
mem.writeIntLittle(u32, code[12..16], aarch64.Instruction.adrp(.x16, pages).toU32());
const narrowed = @truncate(u12, target_addr);
const offset = try math.divExact(u12, narrowed, 8);
mem.writeIntLittle(u32, code[16..20], aarch64.Instruction.ldr(.x16, .{
.register = .{
.rn = .x16,
.offset = aarch64.Instruction.LoadStoreOffset.imm(offset),
},
}).toU32());
}
// br x16
code[20] = 0x00;
code[21] = 0x02;
code[22] = 0x1f;
code[23] = 0xd6;
try self.base.file.?.pwriteAll(&code, stub_helper.offset);
break :blk stub_helper.offset + 6 * @sizeOf(u32);
},
else => unreachable,
}
};
}
fn resolveSymbolsInObject(
self: *MachO,
object_id: u16,
@ -4873,7 +4791,7 @@ fn writeAllGlobalAndUndefSymbols(self: *MachO) !void {
const undefs_off = globals_off + globals_size;
const undefs_size = nundefs * @sizeOf(macho.nlist_64);
log.debug("writing extern symbols from 0x{x} to 0x{x}", .{ undefs_off, undefs_size + undefs_off });
log.debug("writing undef symbols from 0x{x} to 0x{x}", .{ undefs_off, undefs_size + undefs_off });
try self.base.file.?.pwriteAll(mem.sliceAsBytes(self.undefs.items), undefs_off);
// Update dynamic symbol table.