authorgravatar for kubkon@jakubkonka.comJakub Konka <kubkon@jakubkonka.com> 2022-12-09 19:18:06+01:00
committergravatar for noreply@github.comGitHub <noreply@github.com> 2022-12-09 19:18:06+01:00
logbd5a8f86a120093b126cfbde41e002dae161e883
treee15e4dfd7c82542bac7b472b9ff13c2b12369147
parentd0172488b2aa6490dcb983c24cac76e72fa972c0
parent182751ba27edf64304d2ae78043576b78c45fc1c
signature Signed by PGP key 4AEE18F83AFDEB23

Merge pull request #13843 from ziglang/dwarf-abs-paths

debug info: resolve relative paths to source files into absolute paths

8 files changed, 55 insertions(+), 82 deletions(-)

src/Compilation.zig-25
......@@ -1035,9 +1035,6 @@ pub const InitOptions = struct {
10351035 /// (Darwin) remove dylibs that are unreachable by the entry point or exported symbols
10361036 dead_strip_dylibs: bool = false,
10371037 libcxx_abi_version: libcxx.AbiVersion = libcxx.AbiVersion.default,
1038 /// (Windows) PDB source path prefix to instruct the linker how to resolve relative
1039 /// paths when consolidating CodeView streams into a single PDB file.
1040 pdb_source_path: ?[]const u8 = null,
10411038};
10421039
10431040fn addPackageTableToCacheHash(
......@@ -1740,27 +1737,6 @@ pub fn create(gpa: Allocator, options: InitOptions) !*Compilation {
17401737 };
17411738 };
17421739
1743 const pdb_source_path: ?[]const u8 = options.pdb_source_path orelse blk: {
1744 if (builtin.target.os.tag == .windows) {
1745 // PDB requires all file paths to be fully resolved, and it is really the
1746 // linker's responsibility to canonicalize any path extracted from the CodeView
1747 // in the object file. However, LLD-link has some very questionable defaults, and
1748 // in particular, it purposely bakes in path separator of the host system it was
1749 // built on rather than the targets, or just throw an error. Thankfully, they have
1750 // left a backdoor we can use via -PDBSOURCEPATH.
1751 const mod = module orelse break :blk null;
1752 var buffer: [std.fs.MAX_PATH_BYTES]u8 = undefined;
1753 const resolved_path = if (mod.main_pkg.root_src_directory.path) |base_path| p: {
1754 if (std.fs.path.isAbsolute(base_path)) break :blk base_path;
1755 const resolved_path = std.os.realpath(base_path, &buffer) catch break :blk null;
1756 const pos = std.mem.lastIndexOfLinear(u8, resolved_path, base_path) orelse resolved_path.len;
1757 break :p resolved_path[0..pos];
1758 } else std.os.realpath(".", &buffer) catch break :blk null;
1759 break :blk try arena.dupe(u8, resolved_path);
1760 }
1761 break :blk null;
1762 };
1763
17641740 const implib_emit: ?link.Emit = blk: {
17651741 const emit_implib = options.emit_implib orelse break :blk null;
17661742
......@@ -1906,7 +1882,6 @@ pub fn create(gpa: Allocator, options: InitOptions) !*Compilation {
19061882 .headerpad_max_install_names = options.headerpad_max_install_names,
19071883 .dead_strip_dylibs = options.dead_strip_dylibs,
19081884 .force_undefined_symbols = .{},
1909 .pdb_source_path = pdb_source_path,
19101885 });
19111886 errdefer bin_file.destroy();
19121887 comp.* = .{
src/codegen/llvm.zig+25-19
......@@ -443,20 +443,19 @@ pub const Object = struct {
443443 });
444444 defer gpa.free(producer);
445445
446 // For macOS stack traces, we want to avoid having to parse the compilation unit debug
447 // info. As long as each debug info file has a path independent of the compilation unit
448 // directory (DW_AT_comp_dir), then we never have to look at the compilation unit debug
449 // info. If we provide an absolute path to LLVM here for the compilation unit debug
450 // info, LLVM will emit DWARF info that depends on DW_AT_comp_dir. To avoid this, we
451 // pass "." for the compilation unit directory. This forces each debug file to have a
452 // directory rather than be relative to DW_AT_comp_dir. According to DWARF 5, debug
453 // files will no longer reference DW_AT_comp_dir, for the purpose of being able to
454 // support the common practice of stripping all but the line number sections from an
455 // executable.
456 const compile_unit_dir = d: {
457 if (options.target.isDarwin()) break :d ".";
458 const mod = options.module orelse break :d ".";
459 break :d mod.root_pkg.root_src_directory.path orelse ".";
446 // We fully resolve all paths at this point to avoid lack of source line info in stack
447 // traces or lack of debugging information which, if relative paths were used, would
448 // be very location dependent.
449 // TODO: the only concern I have with this is WASI as either host or target, should
450 // we leave the paths as relative then?
451 var buf: [std.fs.MAX_PATH_BYTES]u8 = undefined;
452 const compile_unit_dir = blk: {
453 const path = d: {
454 const mod = options.module orelse break :d ".";
455 break :d mod.root_pkg.root_src_directory.path orelse ".";
456 };
457 if (std.fs.path.isAbsolute(path)) break :blk path;
458 break :blk std.os.realpath(path, &buf) catch path; // If realpath fails, fallback to whatever path was
460459 };
461460 const compile_unit_dir_z = try gpa.dupeZ(u8, compile_unit_dir);
462461 defer gpa.free(compile_unit_dir_z);
......@@ -1389,13 +1388,20 @@ pub const Object = struct {
13891388 if (gop.found_existing) {
13901389 return @ptrCast(*llvm.DIFile, gop.value_ptr.*);
13911390 }
1392 const dir_path = file.pkg.root_src_directory.path orelse ".";
1391 const dir_path_z = d: {
1392 var buffer: [std.fs.MAX_PATH_BYTES]u8 = undefined;
1393 const dir_path = file.pkg.root_src_directory.path orelse ".";
1394 const resolved_dir_path = if (std.fs.path.isAbsolute(dir_path))
1395 dir_path
1396 else
1397 std.os.realpath(dir_path, &buffer) catch dir_path; // If realpath fails, fallback to whatever dir_path was
1398 break :d try std.fs.path.joinZ(gpa, &.{
1399 resolved_dir_path, std.fs.path.dirname(file.sub_file_path) orelse "",
1400 });
1401 };
1402 defer gpa.free(dir_path_z);
13931403 const sub_file_path_z = try gpa.dupeZ(u8, std.fs.path.basename(file.sub_file_path));
13941404 defer gpa.free(sub_file_path_z);
1395 const dir_path_z = try std.fs.path.joinZ(gpa, &.{
1396 dir_path, std.fs.path.dirname(file.sub_file_path) orelse "",
1397 });
1398 defer gpa.free(dir_path_z);
13991405 const di_file = o.di_builder.?.createFile(sub_file_path_z, dir_path_z);
14001406 gop.value_ptr.* = di_file.toNode();
14011407 return di_file;
src/link.zig-4
......@@ -217,10 +217,6 @@ pub const Options = struct {
217217 /// (Darwin) remove dylibs that are unreachable by the entry point or exported symbols
218218 dead_strip_dylibs: bool = false,
219219
220 /// (Windows) PDB source path prefix to instruct the linker how to resolve relative
221 /// paths when consolidating CodeView streams into a single PDB file.
222 pdb_source_path: ?[]const u8 = null,
223
224220 pub fn effectiveOutputMode(options: Options) std.builtin.OutputMode {
225221 return if (options.use_lld) .Obj else options.output_mode;
226222 }
src/link/Coff/lld.zig-4
......@@ -171,10 +171,6 @@ pub fn linkWithLLD(self: *Coff, comp: *Compilation, prog_node: *std.Progress.Nod
171171 });
172172 try argv.append(try allocPrint(arena, "-PDB:{s}", .{out_pdb}));
173173 try argv.append(try allocPrint(arena, "-PDBALTPATH:{s}", .{out_pdb}));
174
175 if (self.base.options.pdb_source_path) |path| {
176 try argv.append(try std.fmt.allocPrint(arena, "-PDBSOURCEPATH:{s}", .{path}));
177 }
178174 }
179175 if (self.base.options.lto) {
180176 switch (self.base.options.optimize_mode) {
src/link/Dwarf.zig+27-27
......@@ -1771,7 +1771,8 @@ pub fn writeDbgInfoHeader(self: *Dwarf, module: *Module, low_pc: u64, high_pc: u
17711771 }
17721772 // Write the form for the compile unit, which must match the abbrev table above.
17731773 const name_strp = try self.makeString(module.root_pkg.root_src_path);
1774 const compile_unit_dir = self.getCompDir(module);
1774 var compile_unit_dir_buffer: [std.fs.MAX_PATH_BYTES]u8 = undefined;
1775 const compile_unit_dir = resolveCompilationDir(module, &compile_unit_dir_buffer);
17751776 const comp_dir_strp = try self.makeString(compile_unit_dir);
17761777 const producer_strp = try self.makeString(link.producer_string);
17771778
......@@ -1823,19 +1824,15 @@ pub fn writeDbgInfoHeader(self: *Dwarf, module: *Module, low_pc: u64, high_pc: u
18231824 }
18241825}
18251826
1826fn getCompDir(self: Dwarf, module: *Module) []const u8 {
1827 // For macOS stack traces, we want to avoid having to parse the compilation unit debug
1828 // info. As long as each debug info file has a path independent of the compilation unit
1829 // directory (DW_AT_comp_dir), then we never have to look at the compilation unit debug
1830 // info. If we provide an absolute path to LLVM here for the compilation unit debug
1831 // info, LLVM will emit DWARF info that depends on DW_AT_comp_dir. To avoid this, we
1832 // pass "." for the compilation unit directory. This forces each debug file to have a
1833 // directory rather than be relative to DW_AT_comp_dir. According to DWARF 5, debug
1834 // files will no longer reference DW_AT_comp_dir, for the purpose of being able to
1835 // support the common practice of stripping all but the line number sections from an
1836 // executable.
1837 if (self.bin_file.tag == .macho) return ".";
1838 return module.root_pkg.root_src_directory.path orelse ".";
1827fn resolveCompilationDir(module: *Module, buffer: *[std.fs.MAX_PATH_BYTES]u8) []const u8 {
1828 // We fully resolve all paths at this point to avoid lack of source line info in stack
1829 // traces or lack of debugging information which, if relative paths were used, would
1830 // be very location dependent.
1831 // TODO: the only concern I have with this is WASI as either host or target, should
1832 // we leave the paths as relative then?
1833 const comp_dir_path = module.root_pkg.root_src_directory.path orelse ".";
1834 if (std.fs.path.isAbsolute(comp_dir_path)) return comp_dir_path;
1835 return std.os.realpath(comp_dir_path, buffer) catch comp_dir_path; // If realpath fails, fallback to whatever comp_dir_path was
18391836}
18401837
18411838fn writeAddrAssumeCapacity(self: *Dwarf, buf: *std.ArrayList(u8), addr: u64) void {
......@@ -2149,7 +2146,7 @@ pub fn writeDbgAranges(self: *Dwarf, addr: u64, size: u64) !void {
21492146 }
21502147}
21512148
2152pub fn writeDbgLineHeader(self: *Dwarf, module: *Module) !void {
2149pub fn writeDbgLineHeader(self: *Dwarf) !void {
21532150 const gpa = self.allocator;
21542151
21552152 const ptr_width_bytes: u8 = self.ptrWidthBytes();
......@@ -2167,7 +2164,7 @@ pub fn writeDbgLineHeader(self: *Dwarf, module: *Module) !void {
21672164 // Convert all input DI files into a set of include dirs and file names.
21682165 var arena = std.heap.ArenaAllocator.init(gpa);
21692166 defer arena.deinit();
2170 const paths = try self.genIncludeDirsAndFileNames(arena.allocator(), module);
2167 const paths = try self.genIncludeDirsAndFileNames(arena.allocator());
21712168
21722169 // The size of this header is variable, depending on the number of directories,
21732170 // files, and padding. We have a function to compute the upper bound size, however,
......@@ -2551,7 +2548,7 @@ fn addDIFile(self: *Dwarf, mod: *Module, decl_index: Module.Decl.Index) !u28 {
25512548 return @intCast(u28, gop.index + 1);
25522549}
25532550
2554fn genIncludeDirsAndFileNames(self: *Dwarf, arena: Allocator, module: *Module) !struct {
2551fn genIncludeDirsAndFileNames(self: *Dwarf, arena: Allocator) !struct {
25552552 dirs: []const []const u8,
25562553 files: []const []const u8,
25572554 files_dirs_indexes: []u28,
......@@ -2565,19 +2562,22 @@ fn genIncludeDirsAndFileNames(self: *Dwarf, arena: Allocator, module: *Module) !
25652562 var files_dir_indexes = std.ArrayList(u28).init(arena);
25662563 try files_dir_indexes.ensureTotalCapacity(self.di_files.count());
25672564
2568 const comp_dir = self.getCompDir(module);
2569
25702565 for (self.di_files.keys()) |dif| {
2571 const full_path = try dif.fullPath(arena);
2572 const dir_path = std.fs.path.dirname(full_path) orelse ".";
2573 const sub_file_path = std.fs.path.basename(full_path);
2566 const dir_path = d: {
2567 var buffer: [std.fs.MAX_PATH_BYTES]u8 = undefined;
2568 const dir_path = dif.pkg.root_src_directory.path orelse ".";
2569 const abs_dir_path = if (std.fs.path.isAbsolute(dir_path))
2570 dir_path
2571 else
2572 std.os.realpath(dir_path, &buffer) catch dir_path; // If realpath fails, fallback to whatever dir_path was
2573 break :d try std.fs.path.join(arena, &.{
2574 abs_dir_path, std.fs.path.dirname(dif.sub_file_path) orelse "",
2575 });
2576 };
2577 const sub_file_path = try arena.dupe(u8, std.fs.path.basename(dif.sub_file_path));
25742578
25752579 const dir_index: u28 = blk: {
2576 const actual_dir_path = if (mem.indexOf(u8, dir_path, comp_dir)) |_| inner: {
2577 if (comp_dir.len == dir_path.len) break :blk 0;
2578 break :inner dir_path[comp_dir.len + 1 ..];
2579 } else dir_path;
2580 const dirs_gop = dirs.getOrPutAssumeCapacity(actual_dir_path);
2580 const dirs_gop = dirs.getOrPutAssumeCapacity(dir_path);
25812581 break :blk @intCast(u28, dirs_gop.index + 1);
25822582 };
25832583
src/link/Elf.zig+1-1
......@@ -1149,7 +1149,7 @@ pub fn flushModule(self: *Elf, comp: *Compilation, prog_node: *std.Progress.Node
11491149 }
11501150
11511151 if (self.debug_line_header_dirty) {
1152 try dw.writeDbgLineHeader(module);
1152 try dw.writeDbgLineHeader();
11531153 self.debug_line_header_dirty = false;
11541154 }
11551155 }
src/link/MachO/DebugSymbols.zig+1-1
......@@ -281,7 +281,7 @@ pub fn flushModule(self: *DebugSymbols, macho_file: *MachO) !void {
281281 }
282282
283283 if (self.debug_line_header_dirty) {
284 try self.dwarf.writeDbgLineHeader(module);
284 try self.dwarf.writeDbgLineHeader();
285285 self.debug_line_header_dirty = false;
286286 }
287287
src/link/Wasm.zig+1-1
......@@ -2672,7 +2672,7 @@ pub fn flushModule(wasm: *Wasm, comp: *Compilation, prog_node: *std.Progress.Nod
26722672 // as locations are always offsets relative to 'code' section.
26732673 try dwarf.writeDbgInfoHeader(mod, 0, code_section_size);
26742674 try dwarf.writeDbgAranges(0, code_section_size);
2675 try dwarf.writeDbgLineHeader(mod);
2675 try dwarf.writeDbgLineHeader();
26762676 }
26772677
26782678 var debug_bytes = std.ArrayList(u8).init(wasm.base.allocator);