motiejus/zig
1
Fork 0
Code Releases Activity

elf: move logic for extracing atom's code into input files

Browse Source
This commit is contained in:
Jakub Konka
2023-09-28 18:35:26 +02:00
parent 91f2e66bf9
commit 89c2151a97
6 changed files with 237 additions and 50 deletions
Download Patch File Download Diff File Expand all files Collapse all files

16
src/link/Elf.zig
View File

@@ -1321,16 +1321,14 @@ pub fn flushModule(self: *Elf, comp: *Compilation, prog_node: *std.Progress.Node
// Beyond this point, everything has been allocated a virtual address and we can resolve
// the relocations, and commit objects to file.
if (self.zig_module_index) |index| {
for (self.file(index).?.zig_module.atoms.keys()) |atom_index| {
const zig_module = self.file(index).?.zig_module;
for (zig_module.atoms.keys()) |atom_index| {
const atom_ptr = self.atom(atom_index).?;
if (!atom_ptr.flags.alive) continue;
const code = try zig_module.codeAlloc(self, atom_index);
defer gpa.free(code);
const shdr = &self.shdrs.items[atom_ptr.outputShndx().?];
const file_offset = shdr.sh_offset + atom_ptr.value - shdr.sh_addr;
const size = math.cast(usize, atom_ptr.size) orelse return error.Overflow;
const code = try gpa.alloc(u8, size);
defer gpa.free(code);
const amt = try self.base.file.?.preadAll(code, file_offset);
if (amt != code.len) return error.InputOutput;
try atom_ptr.resolveRelocs(self, code);
try self.base.file.?.pwriteAll(code, file_offset);
}
@@ -1864,7 +1862,7 @@ fn writeObjects(self: *Elf) !void {
const file_offset = shdr.sh_offset + atom_ptr.value - shdr.sh_addr;
log.debug("writing atom({d}) at 0x{x}", .{ atom_ptr.atom_index, file_offset });
const code = try atom_ptr.codeInObjectUncompressAlloc(self);
const code = try object.codeDecompressAlloc(self, atom_ptr.atom_index);
defer gpa.free(code);
try atom_ptr.resolveRelocs(self, code);
@@ -3905,6 +3903,10 @@ pub fn calcImageBase(self: Elf) u64 {
};
}
pub fn isStatic(self: Elf) bool {
return self.base.options.link_mode == .Static;
}
pub fn isDynLib(self: Elf) bool {
return self.base.options.output_mode == .Lib and self.base.options.link_mode == .Dynamic;
}

187
src/link/Elf/Atom.zig
View File

@@ -59,38 +59,6 @@ pub fn outputShndx(self: Atom) ?u16 {
return self.output_section_index;
}
pub fn codeInObject(self: Atom, elf_file: *Elf) error{Overflow}![]const u8 {
const object = self.file(elf_file).?.object;
return object.shdrContents(self.input_section_index);
}
/// Returns atom's code and optionally uncompresses data if required (for compressed sections).
/// Caller owns the memory.
pub fn codeInObjectUncompressAlloc(self: Atom, elf_file: *Elf) ![]u8 {
const gpa = elf_file.base.allocator;
const data = try self.codeInObject(elf_file);
const shdr = self.inputShdr(elf_file);
if (shdr.sh_flags & elf.SHF_COMPRESSED != 0) {
const chdr = @as(*align(1) const elf.Elf64_Chdr, @ptrCast(data.ptr)).*;
switch (chdr.ch_type) {
.ZLIB => {
var stream = std.io.fixedBufferStream(data[@sizeOf(elf.Elf64_Chdr)..]);
var zlib_stream = std.compress.zlib.decompressStream(gpa, stream.reader()) catch
return error.InputOutput;
defer zlib_stream.deinit();
const size = std.math.cast(usize, chdr.ch_size) orelse return error.Overflow;
const decomp = try gpa.alloc(u8, size);
const nread = zlib_stream.reader().readAll(decomp) catch return error.InputOutput;
if (nread != decomp.len) {
return error.InputOutput;
}
return decomp;
},
else => @panic("TODO unhandled compression scheme"),
}
} else return gpa.dupe(u8, data);
}
pub fn priority(self: Atom, elf_file: *Elf) u64 {
const index = self.file(elf_file).?.index();
return (@as(u64, @intCast(index)) << 32) | @as(u64, @intCast(self.input_section_index));
@@ -327,7 +295,15 @@ pub fn freeRelocs(self: Atom, elf_file: *Elf) void {
zig_module.relocs.items[self.relocs_section_index].clearRetainingCapacity();
}
pub fn scanRelocs(self: Atom, elf_file: *Elf, undefs: anytype) !void {
pub fn scanRelocsRequiresCode(self: Atom, elf_file: *Elf) error{Overflow}!bool {
for (try self.relocs(elf_file)) |rel| {
if (rel.r_type() == elf.R_X86_64_GOTTPOFF) return true;
}
return false;
}
pub fn scanRelocs(self: Atom, elf_file: *Elf, code: ?[]const u8, undefs: anytype) !void {
const is_dyn_lib = elf_file.isDynLib();
const file_ptr = self.file(elf_file).?;
const rels = try self.relocs(elf_file);
var i: usize = 0;
@@ -391,7 +367,38 @@ pub fn scanRelocs(self: Atom, elf_file: *Elf, undefs: anytype) !void {
elf.R_X86_64_TPOFF32,
elf.R_X86_64_TPOFF64,
=> {
// if (is_shared) self.picError(symbol, rel, elf_file);
if (is_dyn_lib) {
// TODO
// self.picError(symbol, rel, elf_file);
}
},
elf.R_X86_64_TLSGD => {
// TODO verify followed by appropriate relocation such as PLT32 __tls_get_addr
if (elf_file.isStatic() or
(!symbol.flags.import and !is_dyn_lib))
{
// Relax if building with -static flag as __tls_get_addr() will not be present in libc.a
// We skip the next relocation.
i += 1;
} else if (!symbol.flags.import and is_dyn_lib) {
symbol.flags.needs_gottp = true;
i += 1;
} else {
symbol.flags.needs_tlsgd = true;
}
},
elf.R_X86_64_GOTTPOFF => {
const should_relax = blk: {
// if (!elf_file.options.relax or is_shared or symbol.flags.import) break :blk false;
if (!x86_64.canRelaxGotTpOff(code.?[rel.r_offset - 3 ..])) break :blk false;
break :blk true;
};
if (!should_relax) {
symbol.flags.needs_gottp = true;
}
},
else => {
@@ -446,7 +453,10 @@ pub fn resolveRelocs(self: Atom, elf_file: *Elf, code: []u8) !void {
var stream = std.io.fixedBufferStream(code);
const cwriter = stream.writer();
for (try self.relocs(elf_file)) |rel| {
const rels = try self.relocs(elf_file);
var i: usize = 0;
while (i < rels.len) : (i += 1) {
const rel = rels[i];
const r_type = rel.r_type();
if (r_type == elf.R_X86_64_NONE) continue;
@@ -531,6 +541,39 @@ pub fn resolveRelocs(self: Atom, elf_file: *Elf, code: []u8) !void {
elf.R_X86_64_TPOFF32 => try cwriter.writeIntLittle(i32, @as(i32, @truncate(S + A - TP))),
elf.R_X86_64_TPOFF64 => try cwriter.writeIntLittle(i64, S + A - TP),
elf.R_X86_64_TLSGD => {
if (target.flags.has_tlsgd) {
// TODO
// const S_ = @as(i64, @intCast(target.tlsGdAddress(elf_file)));
// try cwriter.writeIntLittle(i32, @as(i32, @intCast(S_ + A - P)));
} else if (target.flags.has_gottp) {
// TODO
// const S_ = @as(i64, @intCast(target.getGotTpAddress(elf_file)));
// try relaxTlsGdToIe(relocs[i .. i + 2], @intCast(S_ - P), elf_file, &stream);
i += 1;
} else {
try x86_64.relaxTlsGdToLe(
self,
rels[i .. i + 2],
@as(i32, @intCast(S - TP)),
elf_file,
&stream,
);
i += 1;
}
},
elf.R_X86_64_GOTTPOFF => {
if (target.flags.has_gottp) {
// TODO
// const S_ = @as(i64, @intCast(target.gotTpAddress(elf_file)));
// try cwriter.writeIntLittle(i32, @as(i32, @intCast(S_ + A - P)));
} else {
x86_64.relaxGotTpOff(code[rel.r_offset - 3 ..]) catch unreachable;
try cwriter.writeIntLittle(i32, @as(i32, @intCast(S - TP)));
}
},
else => {},
}
}
@@ -697,6 +740,80 @@ const x86_64 = struct {
}
}
pub fn canRelaxGotTpOff(code: []const u8) bool {
const old_inst = disassemble(code) orelse return false;
switch (old_inst.encoding.mnemonic) {
.mov => if (Instruction.new(old_inst.prefix, .mov, &.{
old_inst.ops[0],
// TODO: hack to force imm32s in the assembler
.{ .imm = Immediate.s(-129) },
})) |inst| {
inst.encode(std.io.null_writer, .{}) catch return false;
return true;
} else |_| return false,
else => return false,
}
}
pub fn relaxGotTpOff(code: []u8) !void {
const old_inst = disassemble(code) orelse return error.RelaxFail;
switch (old_inst.encoding.mnemonic) {
.mov => {
const inst = try Instruction.new(old_inst.prefix, .mov, &.{
old_inst.ops[0],
// TODO: hack to force imm32s in the assembler
.{ .imm = Immediate.s(-129) },
});
relocs_log.debug(" relaxing {} => {}", .{ old_inst.encoding, inst.encoding });
encode(&.{inst}, code) catch return error.RelaxFail;
},
else => return error.RelaxFail,
}
}
pub fn relaxTlsGdToLe(
self: Atom,
rels: []align(1) const elf.Elf64_Rela,
value: i32,
elf_file: *Elf,
stream: anytype,
) !void {
assert(rels.len == 2);
const writer = stream.writer();
switch (rels[1].r_type()) {
elf.R_X86_64_PC32,
elf.R_X86_64_PLT32,
elf.R_X86_64_GOTPCREL,
elf.R_X86_64_GOTPCRELX,
=> {
var insts = [_]u8{
0x64, 0x48, 0x8b, 0x04, 0x25, 0, 0, 0, 0, // movq %fs:0,%rax
0x48, 0x81, 0xc0, 0, 0, 0, 0, // add $tp_offset, %rax
};
std.mem.writeIntLittle(i32, insts[12..][0..4], value);
try stream.seekBy(-4);
try writer.writeAll(&insts);
relocs_log.debug(" relaxing {} and {}", .{
fmtRelocType(rels[0].r_type()),
fmtRelocType(rels[1].r_type()),
});
},
else => {
var err = try elf_file.addErrorWithNotes(1);
try err.addMsg(elf_file, "fatal linker error: rewrite {} when followed by {}", .{
fmtRelocType(rels[0].r_type()),
fmtRelocType(rels[1].r_type()),
});
try err.addNote(elf_file, "in {}:{s} at offset 0x{x}", .{
self.file(elf_file).?.fmtPath(),
self.name(elf_file),
rels[0].r_offset,
});
},
}
}
fn disassemble(code: []const u8) ?Instruction {
var disas = Disassembler.init(code);
const inst = disas.next() catch return null;

43
src/link/Elf/Object.zig
View File

@@ -208,6 +208,8 @@ fn getOutputSectionIndex(self: *Object, elf_file: *Elf, shdr: elf.Elf64_Shdr) er
break :blk prefix;
}
}
if (std.mem.eql(u8, name, ".tcommon")) break :blk ".tbss";
if (std.mem.eql(u8, name, ".common")) break :blk ".bss";
break :blk name;
};
const @"type" = switch (shdr.sh_type) {
@@ -427,7 +429,13 @@ pub fn scanRelocs(self: *Object, elf_file: *Elf, undefs: anytype) !void {
const shdr = atom.inputShdr(elf_file);
if (shdr.sh_flags & elf.SHF_ALLOC == 0) continue;
if (shdr.sh_type == elf.SHT_NOBITS) continue;
try atom.scanRelocs(elf_file, undefs);
if (try atom.scanRelocsRequiresCode(elf_file)) {
// TODO ideally, we don't have to decompress at this stage (should already be done)
// and we just fetch the code slice.
const code = try self.codeDecompressAlloc(elf_file, atom_index);
defer elf_file.base.allocator.free(code);
try atom.scanRelocs(elf_file, code, undefs);
} else try atom.scanRelocs(elf_file, null, undefs);
}
for (self.cies.items) |cie| {
@@ -590,7 +598,7 @@ pub fn convertCommonSymbols(self: *Object, elf_file: *Elf) !void {
try self.atoms.append(gpa, atom_index);
const is_tls = global.getType(elf_file) == elf.STT_TLS;
const name = if (is_tls) ".tls_common" else ".common";
const name = if (is_tls) ".tbss" else ".bss";
const atom = elf_file.atom(atom_index).?;
atom.atom_index = atom_index;
@@ -684,7 +692,7 @@ pub fn globals(self: *Object) []const Symbol.Index {
return self.symbols.items[start..];
}
pub fn shdrContents(self: *Object, index: u32) error{Overflow}![]const u8 {
fn shdrContents(self: Object, index: u32) error{Overflow}![]const u8 {
assert(index < self.shdrs.items.len);
const shdr = self.shdrs.items[index];
const offset = math.cast(usize, shdr.sh_offset) orelse return error.Overflow;
@@ -692,6 +700,35 @@ pub fn shdrContents(self: *Object, index: u32) error{Overflow}![]const u8 {
return self.data[offset..][0..size];
}
/// Returns atom's code and optionally uncompresses data if required (for compressed sections).
/// Caller owns the memory.
pub fn codeDecompressAlloc(self: Object, elf_file: *Elf, atom_index: Atom.Index) ![]u8 {
const gpa = elf_file.base.allocator;
const atom_ptr = elf_file.atom(atom_index).?;
assert(atom_ptr.file_index == self.index);
const data = try self.shdrContents(atom_ptr.input_section_index);
const shdr = atom_ptr.inputShdr(elf_file);
if (shdr.sh_flags & elf.SHF_COMPRESSED != 0) {
const chdr = @as(*align(1) const elf.Elf64_Chdr, @ptrCast(data.ptr)).*;
switch (chdr.ch_type) {
.ZLIB => {
var stream = std.io.fixedBufferStream(data[@sizeOf(elf.Elf64_Chdr)..]);
var zlib_stream = std.compress.zlib.decompressStream(gpa, stream.reader()) catch
return error.InputOutput;
defer zlib_stream.deinit();
const size = std.math.cast(usize, chdr.ch_size) orelse return error.Overflow;
const decomp = try gpa.alloc(u8, size);
const nread = zlib_stream.reader().readAll(decomp) catch return error.InputOutput;
if (nread != decomp.len) {
return error.InputOutput;
}
return decomp;
},
else => @panic("TODO unhandled compression scheme"),
}
} else return gpa.dupe(u8, data);
}
fn getString(self: *Object, off: u32) [:0]const u8 {
assert(off < self.strtab.len);
return mem.sliceTo(@as([*:0]const u8, @ptrCast(self.strtab.ptr + off)), 0);

6
src/link/Elf/Symbol.zig
View File

@@ -327,10 +327,12 @@ pub const Flags = packed struct {
has_dynamic: bool = false,
/// Whether the symbol contains TLSGD indirection.
tlsgd: bool = false,
needs_tlsgd: bool = false,
has_tlsgd: bool = false,
/// Whether the symbol contains GOTTP indirection.
gottp: bool = false,
needs_gottp: bool = false,
has_gottp: bool = false,
/// Whether the symbol contains TLSDESC indirection.
tlsdesc: bool = false,

25
src/link/Elf/ZigModule.zig
View File

@@ -144,7 +144,14 @@ pub fn scanRelocs(self: *ZigModule, elf_file: *Elf, undefs: anytype) !void {
for (self.atoms.keys()) |atom_index| {
const atom = elf_file.atom(atom_index) orelse continue;
if (!atom.flags.alive) continue;
try atom.scanRelocs(elf_file, undefs);
if (try atom.scanRelocsRequiresCode(elf_file)) {
// TODO ideally we don't have to fetch the code here.
// Perhaps it would make sense to save the code until flushModule where we
// would free all of generated code?
const code = try self.codeAlloc(elf_file, atom_index);
defer elf_file.base.allocator.free(code);
try atom.scanRelocs(elf_file, code, undefs);
} else try atom.scanRelocs(elf_file, null, undefs);
}
}
@@ -253,6 +260,22 @@ pub fn asFile(self: *ZigModule) File {
return .{ .zig_module = self };
}
/// Returns atom's code.
/// Caller owns the memory.
pub fn codeAlloc(self: ZigModule, elf_file: *Elf, atom_index: Atom.Index) ![]u8 {
const gpa = elf_file.base.allocator;
const atom = elf_file.atom(atom_index).?;
assert(atom.file_index == self.index);
const shdr = &elf_file.shdrs.items[atom.outputShndx().?];
const file_offset = shdr.sh_offset + atom.value - shdr.sh_addr;
const size = std.math.cast(usize, atom.size) orelse return error.Overflow;
const code = try gpa.alloc(u8, size);
errdefer gpa.free(code);
const amt = try elf_file.base.file.?.preadAll(code, file_offset);
if (amt != code.len) return error.InputOutput;
return code;
}
pub fn fmtSymtab(self: *ZigModule, elf_file: *Elf) std.fmt.Formatter(formatSymtab) {
return .{ .data = .{
.self = self,

10
test/link/elf.zig
View File

@@ -158,19 +158,25 @@ fn addRunArtifact(comp: *Compile) *Run {
return run;
}
fn addZigSourceBytes(comp: *Compile, bytes: []const u8) void {
fn addZigSourceBytes(comp: *Compile, comptime bytes: []const u8) void {
const b = comp.step.owner;
const file = WriteFile.create(b).add("a.zig", bytes);
file.addStepDependencies(&comp.step);
comp.root_src = file;
}
fn addCSourceBytes(comp: *Compile, bytes: []const u8) void {
fn addCSourceBytes(comp: *Compile, comptime bytes: []const u8) void {
const b = comp.step.owner;
const file = WriteFile.create(b).add("a.c", bytes);
comp.addCSourceFile(.{ .file = file, .flags = &.{} });
}
fn addAsmSourceBytes(comp: *Compile, comptime bytes: []const u8) void {
const b = comp.step.owner;
const file = WriteFile.create(b).add("a.s", bytes ++ "\n");
comp.addAssemblyFile(file);
}
const std = @import("std");
const Build = std.Build;