Merge pull request #17319 from ziglang/elf-tls

elf: add basic TLS segment handling

5 files changed, 433 insertions, 138 deletions

diff --git a/src/link/Elf.zig b/src/link/Elf.zig
index bf5c232d5c..4e43a20433 100644
--- a/src/link/Elf.zig
+++ b/src/link/Elf.zig
@@ -43,6 +43,12 @@ phdr_load_ro_index: ?u16 = null,
 phdr_load_rw_index: ?u16 = null,
 /// The index into the program headers of a PT_LOAD program header with zerofill data.
 phdr_load_zerofill_index: ?u16 = null,
+/// The index into the program headers of the PT_TLS program header.
+phdr_tls_index: ?u16 = null,
+/// The index into the program headers of a PT_LOAD program header with TLS data.
+phdr_load_tls_data_index: ?u16 = null,
+/// The index into the program headers of a PT_LOAD program header with TLS zerofill data.
+phdr_load_tls_zerofill_index: ?u16 = null,
 
 entry_addr: ?u64 = null,
 page_size: u32,
@@ -56,10 +62,13 @@ strtab: StringTable(.strtab) = .{},
 /// Representation of the GOT table as committed to the file.
 got: GotSection = .{},
 
+/// Tracked section headers
 text_section_index: ?u16 = null,
 rodata_section_index: ?u16 = null,
 data_section_index: ?u16 = null,
 bss_section_index: ?u16 = null,
+tdata_section_index: ?u16 = null,
+tbss_section_index: ?u16 = null,
 eh_frame_section_index: ?u16 = null,
 eh_frame_hdr_section_index: ?u16 = null,
 dynamic_section_index: ?u16 = null,
@@ -238,7 +247,8 @@ pub fn createEmpty(gpa: Allocator, options: link.Options) !*Elf {
     else
         elf.VER_NDX_LOCAL;
 
-    var dwarf: ?Dwarf = if (!options.strip and options.module != null)
+    const use_llvm = options.use_llvm;
+    var dwarf: ?Dwarf = if (!options.strip and options.module != null and !use_llvm)
         Dwarf.init(gpa, &self.base, options.target)
     else
         null;
@@ -255,7 +265,6 @@ pub fn createEmpty(gpa: Allocator, options: link.Options) !*Elf {
         .page_size = page_size,
         .default_sym_version = default_sym_version,
     };
-    const use_llvm = options.use_llvm;
     if (use_llvm and options.module != null) {
         self.llvm_object = try LlvmObject.create(gpa, options);
     }
@@ -358,10 +367,12 @@ fn detectAllocCollision(self: *Elf, start: u64, size: u64) ?u64 {
         }
     }
 
-    for (self.shdrs.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) {
+    for (self.shdrs.items) |shdr| {
+        // SHT_NOBITS takes no physical space in the output file so set its size to 0.
+        const sh_size = if (shdr.sh_type == elf.SHT_NOBITS) 0 else shdr.sh_size;
+        const increased_size = padToIdeal(sh_size);
+        const test_end = shdr.sh_offset + increased_size;
+        if (end > shdr.sh_offset and start < test_end) {
             return test_end;
         }
     }
@@ -429,15 +440,15 @@ pub fn allocateSegment(self: *Elf, opts: AllocateSegmentOpts) error{OutOfMemory}
     const addr = opts.addr orelse blk: {
         const reserved_capacity = self.calcImageBase() * 4;
         // Calculate largest VM address
-        const count = self.phdrs.items.len;
         var addresses = std.ArrayList(u64).init(gpa);
         defer addresses.deinit();
-        try addresses.ensureTotalCapacityPrecise(count);
+        try addresses.ensureTotalCapacityPrecise(self.phdrs.items.len);
         for (self.phdrs.items) |phdr| {
+            if (phdr.p_type != elf.PT_LOAD) continue;
             addresses.appendAssumeCapacity(phdr.p_vaddr + reserved_capacity);
         }
         mem.sort(u64, addresses.items, {}, std.sort.asc(u64));
-        break :blk mem.alignForward(u64, addresses.items[count - 1], opts.alignment);
+        break :blk mem.alignForward(u64, addresses.pop(), opts.alignment);
     };
     log.debug("allocating phdr({d})({c}{c}{c}) from 0x{x} to 0x{x} (0x{x} - 0x{x})", .{
         index,
@@ -492,7 +503,7 @@ pub fn allocateAllocSection(self: *Elf, opts: AllocateAllocSectionOpts) error{Ou
         .sh_flags = opts.flags,
         .sh_addr = phdr.p_vaddr,
         .sh_offset = phdr.p_offset,
-        .sh_size = phdr.p_filesz,
+        .sh_size = phdr.p_memsz,
         .sh_link = 0,
         .sh_info = 0,
         .sh_addralign = opts.alignment,
@@ -543,7 +554,6 @@ pub fn populateMissingMetadata(self: *Elf) !void {
     };
     const ptr_size: u8 = self.ptrWidthBytes();
     const is_linux = self.base.options.target.os.tag == .linux;
-    const large_addrspace = self.base.options.target.ptrBitWidth() >= 32;
     const image_base = self.calcImageBase();
 
     if (self.phdr_table_index == null) {
@@ -566,23 +576,16 @@ pub fn populateMissingMetadata(self: *Elf) !void {
     }
 
     if (self.phdr_table_load_index == null) {
-        self.phdr_table_load_index = @intCast(self.phdrs.items.len);
-        try self.phdrs.append(gpa, .{
-            .p_type = elf.PT_LOAD,
-            .p_offset = 0,
-            .p_filesz = 0,
-            .p_vaddr = image_base,
-            .p_paddr = image_base,
-            .p_memsz = 0,
-            .p_align = self.page_size,
-            .p_flags = elf.PF_R,
+        self.phdr_table_load_index = try self.allocateSegment(.{
+            .addr = image_base,
+            .size = 0,
+            .alignment = self.page_size,
         });
         self.phdr_table_dirty = true;
     }
 
     if (self.phdr_load_re_index == null) {
         self.phdr_load_re_index = try self.allocateSegment(.{
-            .addr = self.defaultEntryAddress(),
             .size = self.base.options.program_code_size_hint,
             .alignment = self.page_size,
             .flags = elf.PF_X | elf.PF_R | elf.PF_W,
@@ -591,12 +594,10 @@ pub fn populateMissingMetadata(self: *Elf) !void {
     }
 
     if (self.phdr_got_index == null) {
-        const addr: u64 = if (large_addrspace) 0x4000000 else 0x8000;
         // We really only need ptr alignment but since we are using PROGBITS, linux requires
         // page align.
         const alignment = if (is_linux) self.page_size else @as(u16, ptr_size);
         self.phdr_got_index = try self.allocateSegment(.{
-            .addr = addr,
             .size = @as(u64, ptr_size) * self.base.options.symbol_count_hint,
             .alignment = alignment,
             .flags = elf.PF_R | elf.PF_W,
@@ -604,10 +605,8 @@ pub fn populateMissingMetadata(self: *Elf) !void {
     }
 
     if (self.phdr_load_ro_index == null) {
-        const addr: u64 = if (large_addrspace) 0xc000000 else 0xa000;
         const alignment = if (is_linux) self.page_size else @as(u16, ptr_size);
         self.phdr_load_ro_index = try self.allocateSegment(.{
-            .addr = addr,
             .size = 1024,
             .alignment = alignment,
             .flags = elf.PF_R | elf.PF_W,
@@ -615,10 +614,8 @@ pub fn populateMissingMetadata(self: *Elf) !void {
     }
 
     if (self.phdr_load_rw_index == null) {
-        const addr: u64 = if (large_addrspace) 0x10000000 else 0xc000;
         const alignment = if (is_linux) self.page_size else @as(u16, ptr_size);
         self.phdr_load_rw_index = try self.allocateSegment(.{
-            .addr = addr,
             .size = 1024,
             .alignment = alignment,
             .flags = elf.PF_R | elf.PF_W,
@@ -626,10 +623,8 @@ pub fn populateMissingMetadata(self: *Elf) !void {
     }
 
     if (self.phdr_load_zerofill_index == null) {
-        const addr: u64 = if (large_addrspace) 0x14000000 else 0xf000;
         const alignment = if (is_linux) self.page_size else @as(u16, ptr_size);
         self.phdr_load_zerofill_index = try self.allocateSegment(.{
-            .addr = addr,
             .size = 0,
             .alignment = alignment,
             .flags = elf.PF_R | elf.PF_W,
@@ -639,6 +634,53 @@ pub fn populateMissingMetadata(self: *Elf) !void {
         phdr.p_memsz = 1024;
     }
 
+    if (!self.base.options.single_threaded) {
+        if (self.phdr_load_tls_data_index == null) {
+            const alignment = if (is_linux) self.page_size else @as(u16, ptr_size);
+            self.phdr_load_tls_data_index = try self.allocateSegment(.{
+                .size = 1024,
+                .alignment = alignment,
+                .flags = elf.PF_R | elf.PF_W,
+            });
+        }
+
+        if (self.phdr_load_tls_zerofill_index == null) {
+            // TODO .tbss doesn't need any physical or memory representation (aka a loadable segment)
+            // since the loader only cares about the PT_TLS to work out TLS size. However, when
+            // relocating we need to have .tdata and .tbss contiguously laid out so that we can
+            // work out correct offsets to the start/end of the TLS segment. I am thinking that
+            // perhaps it's possible to completely spoof it by having an abstracted mechanism
+            // for this that wouldn't require us to explicitly track .tbss. Anyhow, for now,
+            // we go the savage route of treating .tbss like .bss.
+            const alignment = if (is_linux) self.page_size else @as(u16, ptr_size);
+            self.phdr_load_tls_zerofill_index = try self.allocateSegment(.{
+                .size = 0,
+                .alignment = alignment,
+                .flags = elf.PF_R | elf.PF_W,
+            });
+            const phdr = &self.phdrs.items[self.phdr_load_tls_zerofill_index.?];
+            phdr.p_offset = self.phdrs.items[self.phdr_load_tls_data_index.?].p_offset; // .tbss overlaps .tdata
+            phdr.p_memsz = 1024;
+        }
+
+        if (self.phdr_tls_index == null) {
+            self.phdr_tls_index = @intCast(self.phdrs.items.len);
+            const phdr_tdata = &self.phdrs.items[self.phdr_load_tls_data_index.?];
+            const phdr_tbss = &self.phdrs.items[self.phdr_load_tls_zerofill_index.?];
+            try self.phdrs.append(gpa, .{
+                .p_type = elf.PT_TLS,
+                .p_offset = phdr_tdata.p_offset,
+                .p_vaddr = phdr_tdata.p_vaddr,
+                .p_paddr = phdr_tdata.p_paddr,
+                .p_filesz = phdr_tdata.p_filesz,
+                .p_memsz = phdr_tbss.p_vaddr + phdr_tbss.p_memsz - phdr_tdata.p_vaddr,
+                .p_align = ptr_size,
+                .p_flags = elf.PF_R,
+            });
+            self.phdr_table_dirty = true;
+        }
+    }
+
     if (self.shstrtab_section_index == null) {
         assert(self.shstrtab.buffer.items.len == 0);
         try self.shstrtab.buffer.append(gpa, 0); // need a 0 at position 0
@@ -707,6 +749,31 @@ pub fn populateMissingMetadata(self: *Elf) !void {
         try self.last_atom_and_free_list_table.putNoClobber(gpa, self.bss_section_index.?, .{});
     }
 
+    if (self.phdr_load_tls_data_index) |phdr_index| {
+        if (self.tdata_section_index == null) {
+            self.tdata_section_index = try self.allocateAllocSection(.{
+                .name = ".tdata",
+                .phdr_index = phdr_index,
+                .alignment = ptr_size,
+                .flags = elf.SHF_ALLOC | elf.SHF_WRITE | elf.SHF_TLS,
+            });
+            try self.last_atom_and_free_list_table.putNoClobber(gpa, self.tdata_section_index.?, .{});
+        }
+    }
+
+    if (self.phdr_load_tls_zerofill_index) |phdr_index| {
+        if (self.tbss_section_index == null) {
+            self.tbss_section_index = try self.allocateAllocSection(.{
+                .name = ".tbss",
+                .phdr_index = phdr_index,
+                .alignment = ptr_size,
+                .flags = elf.SHF_ALLOC | elf.SHF_WRITE | elf.SHF_TLS,
+                .type = elf.SHT_NOBITS,
+            });
+            try self.last_atom_and_free_list_table.putNoClobber(gpa, self.tbss_section_index.?, .{});
+        }
+    }
+
     if (self.symtab_section_index == null) {
         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);
@@ -844,10 +911,7 @@ pub fn growAllocSection(self: *Elf, shdr_index: u16, needed_size: u64) !void {
     if (needed_size > self.allocatedSize(shdr.sh_offset) and !is_zerofill) {
         // Must move the entire section.
         const new_offset = self.findFreeSpace(needed_size, self.page_size);
-        const existing_size = if (self.last_atom_and_free_list_table.get(shdr_index)) |meta| blk: {
-            const last = self.atom(meta.last_atom_index) orelse break :blk 0;
-            break :blk (last.value + last.size) - phdr.p_vaddr;
-        } else shdr.sh_size;
+        const existing_size = shdr.sh_size;
         shdr.sh_size = 0;
 
         log.debug("new '{s}' file offset 0x{x} to 0x{x}", .{
@@ -857,12 +921,18 @@ pub fn growAllocSection(self: *Elf, shdr_index: u16, needed_size: u64) !void {
         });
 
         const amt = try self.base.file.?.copyRangeAll(shdr.sh_offset, self.base.file.?, new_offset, existing_size);
+        // TODO figure out what to about this error condition - how to communicate it up.
         if (amt != existing_size) return error.InputOutput;
 
         shdr.sh_offset = new_offset;
         phdr.p_offset = new_offset;
     }
 
+    shdr.sh_size = needed_size;
+    if (!is_zerofill) {
+        phdr.p_filesz = needed_size;
+    }
+
     const mem_capacity = self.allocatedVirtualSize(phdr.p_vaddr);
     if (needed_size > mem_capacity) {
         // We are exceeding our allocated VM capacity so we need to shift everything in memory
@@ -889,13 +959,8 @@ pub fn growAllocSection(self: *Elf, shdr_index: u16, needed_size: u64) !void {
         }
     }
 
-    shdr.sh_size = needed_size;
     phdr.p_memsz = needed_size;
 
-    if (!is_zerofill) {
-        phdr.p_filesz = needed_size;
-    }
-
     self.markDirty(shdr_index, phdr_index);
 }
 
@@ -965,21 +1030,15 @@ pub fn growNonAllocSection(
     const shdr = &self.shdrs.items[shdr_index];
 
     if (needed_size > self.allocatedSize(shdr.sh_offset)) {
-        const existing_size = if (self.symtab_section_index.? == shdr_index) blk: {
-            const sym_size: u64 = switch (self.ptr_width) {
-                .p32 => @sizeOf(elf.Elf32_Sym),
-                .p64 => @sizeOf(elf.Elf64_Sym),
-            };
-            break :blk @as(u64, shdr.sh_info) * sym_size;
-        } else shdr.sh_size;
+        const existing_size = shdr.sh_size;
         shdr.sh_size = 0;
         // Move all the symbols to a new file location.
         const new_offset = self.findFreeSpace(needed_size, min_alignment);
 
-        log.debug("moving '{?s}' from 0x{x} to 0x{x}", .{
-            self.shstrtab.get(shdr.sh_name),
-            shdr.sh_offset,
+        log.debug("new '{s}' file offset 0x{x} to 0x{x}", .{
+            self.shstrtab.getAssumeExists(shdr.sh_name),
             new_offset,
+            new_offset + existing_size,
         });
 
         if (requires_file_copy) {
@@ -1223,19 +1282,48 @@ pub fn flushModule(self: *Elf, comp: *Compilation, prog_node: *std.Progress.Node
     try self.allocateObjects();
     self.allocateLinkerDefinedSymbols();
 
+    // .bss always overlaps .data in file offset, but is zero-sized in file so it doesn't
+    // get mapped by the loader
+    if (self.data_section_index) |data_shndx| blk: {
+        const bss_shndx = self.bss_section_index orelse break :blk;
+        const data_phndx = self.phdr_to_shdr_table.get(data_shndx).?;
+        const bss_phndx = self.phdr_to_shdr_table.get(bss_shndx).?;
+        self.shdrs.items[bss_shndx].sh_offset = self.shdrs.items[data_shndx].sh_offset;
+        self.phdrs.items[bss_phndx].p_offset = self.phdrs.items[data_phndx].p_offset;
+    }
+
+    // Same treatment for .tbss section.
+    if (self.tdata_section_index) |tdata_shndx| blk: {
+        const tbss_shndx = self.tbss_section_index orelse break :blk;
+        const tdata_phndx = self.phdr_to_shdr_table.get(tdata_shndx).?;
+        const tbss_phndx = self.phdr_to_shdr_table.get(tbss_shndx).?;
+        self.shdrs.items[tbss_shndx].sh_offset = self.shdrs.items[tdata_shndx].sh_offset;
+        self.phdrs.items[tbss_phndx].p_offset = self.phdrs.items[tdata_phndx].p_offset;
+    }
+
+    if (self.phdr_tls_index) |tls_index| {
+        const tdata_phdr = &self.phdrs.items[self.phdr_load_tls_data_index.?];
+        const tbss_phdr = &self.phdrs.items[self.phdr_load_tls_zerofill_index.?];
+        const phdr = &self.phdrs.items[tls_index];
+        phdr.p_offset = tdata_phdr.p_offset;
+        phdr.p_filesz = tdata_phdr.p_filesz;
+        phdr.p_vaddr = tdata_phdr.p_vaddr;
+        phdr.p_paddr = tdata_phdr.p_vaddr;
+        phdr.p_memsz = tbss_phdr.p_vaddr + tbss_phdr.p_memsz - tdata_phdr.p_vaddr;
+    }
+
     // 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 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);
+            if (shdr.sh_type == elf.SHT_NOBITS) continue;
+            const code = try zig_module.codeAlloc(self, atom_index);
             defer gpa.free(code);
-            const amt = try self.base.file.?.preadAll(code, file_offset);
-            if (amt != code.len) return error.InputOutput;
+            const file_offset = shdr.sh_offset + atom_ptr.value - shdr.sh_addr;
             try atom_ptr.resolveRelocs(self, code);
             try self.base.file.?.pwriteAll(code, file_offset);
         }
@@ -1268,22 +1356,6 @@ pub fn flushModule(self: *Elf, comp: *Compilation, prog_node: *std.Progress.Node
     try self.updateSymtabSize();
     try self.writeSymtab();
 
-    // .bss always overlaps .data in file offset, but is zero-sized in file so it doesn't
-    // get mapped by the loader
-    if (self.data_section_index) |data_shndx| blk: {
-        const bss_shndx = self.bss_section_index orelse break :blk;
-        const data_phndx = self.phdr_to_shdr_table.get(data_shndx).?;
-        const bss_phndx = self.phdr_to_shdr_table.get(bss_shndx).?;
-        self.shdrs.items[bss_shndx].sh_offset = self.shdrs.items[data_shndx].sh_offset;
-        self.phdrs.items[bss_phndx].p_offset = self.phdrs.items[data_phndx].p_offset;
-    }
-
-    // Dump the state for easy debugging.
-    // State can be dumped via `--debug-log link_state`.
-    if (build_options.enable_logging) {
-        state_log.debug("{}", .{self.dumpState()});
-    }
-
     if (self.dwarf) |*dw| {
         if (self.debug_abbrev_section_dirty) {
             try dw.writeDbgAbbrev();
@@ -1470,6 +1542,12 @@ pub fn flushModule(self: *Elf, comp: *Compilation, prog_node: *std.Progress.Node
         try self.writeElfHeader();
     }
 
+    // Dump the state for easy debugging.
+    // State can be dumped via `--debug-log link_state`.
+    if (build_options.enable_logging) {
+        state_log.debug("{}", .{self.dumpState()});
+    }
+
     // 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,
@@ -1779,7 +1857,7 @@ 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);
@@ -2785,10 +2863,6 @@ fn updateDeclCode(
         try self.got.writeEntry(self, gop.index);
     }
 
-    const phdr_index = self.phdr_to_shdr_table.get(shdr_index).?;
-    const section_offset = sym.value - self.phdrs.items[phdr_index].p_vaddr;
-    const file_offset = self.shdrs.items[shdr_index].sh_offset + section_offset;
-
     if (self.base.child_pid) |pid| {
         switch (builtin.os.tag) {
             .linux => {
@@ -2810,7 +2884,13 @@ fn updateDeclCode(
         }
     }
 
-    try self.base.file.?.pwriteAll(code, file_offset);
+    const shdr = self.shdrs.items[shdr_index];
+    if (shdr.sh_type != elf.SHT_NOBITS) {
+        const phdr_index = self.phdr_to_shdr_table.get(shdr_index).?;
+        const section_offset = sym.value - self.phdrs.items[phdr_index].p_vaddr;
+        const file_offset = shdr.sh_offset + section_offset;
+        try self.base.file.?.pwriteAll(code, file_offset);
+    }
 }
 
 pub fn updateFunc(self: *Elf, mod: *Module, func_index: InternPool.Index, air: Air, liveness: Liveness) !void {
@@ -3358,9 +3438,14 @@ fn allocateLinkerDefinedSymbols(self: *Elf) void {
     // _end
     {
         const end_symbol = self.symbol(self.end_index.?);
+        end_symbol.value = 0;
         for (self.shdrs.items, 0..) |*shdr, shndx| {
-            if (shdr.sh_flags & elf.SHF_ALLOC != 0) {
-                end_symbol.value = shdr.sh_addr + shdr.sh_size;
+            if (shdr.sh_flags & elf.SHF_ALLOC == 0) continue;
+            const phdr_index = self.phdr_to_shdr_table.get(@intCast(shndx)).?;
+            const phdr = self.phdrs.items[phdr_index];
+            const value = phdr.p_vaddr + phdr.p_memsz;
+            if (end_symbol.value < value) {
+                end_symbol.value = value;
                 end_symbol.output_section_index = @intCast(shndx);
             }
         }
@@ -3424,6 +3509,7 @@ fn updateSymtabSize(self: *Elf) !void {
         .p64 => @alignOf(elf.Elf64_Sym),
     };
     const needed_size = (sizes.nlocals + sizes.nglobals + 1) * sym_size;
+    shdr.sh_size = needed_size;
     try self.growNonAllocSection(self.symtab_section_index.?, needed_size, sym_align, true);
 }
 
@@ -3820,12 +3906,8 @@ pub fn calcImageBase(self: Elf) u64 {
     };
 }
 
-pub fn defaultEntryAddress(self: Elf) u64 {
-    if (self.entry_addr) |addr| return addr;
-    return switch (self.base.options.target.cpu.arch) {
-        .spu_2 => 0,
-        else => default_entry_addr,
-    };
+pub fn isStatic(self: Elf) bool {
+    return self.base.options.link_mode == .Static;
 }
 
 pub fn isDynLib(self: Elf) bool {
@@ -4011,6 +4093,22 @@ pub fn comdatGroupOwner(self: *Elf, index: ComdatGroupOwner.Index) *ComdatGroupO
     return &self.comdat_groups_owners.items[index];
 }
 
+pub fn tpAddress(self: *Elf) u64 {
+    const index = self.phdr_tls_index orelse return 0;
+    const phdr = self.phdrs.items[index];
+    return mem.alignForward(u64, phdr.p_vaddr + phdr.p_memsz, phdr.p_align);
+}
+
+pub fn dtpAddress(self: *Elf) u64 {
+    return self.tlsAddress();
+}
+
+pub fn tlsAddress(self: *Elf) u64 {
+    const index = self.phdr_tls_index orelse return 0;
+    const phdr = self.phdrs.items[index];
+    return phdr.p_vaddr;
+}
+
 const ErrorWithNotes = struct {
     /// Allocated index in misc_errors array.
     index: usize,
@@ -4043,7 +4141,7 @@ const ErrorWithNotes = struct {
     }
 };
 
-fn addErrorWithNotes(self: *Elf, note_count: usize) error{OutOfMemory}!ErrorWithNotes {
+pub fn addErrorWithNotes(self: *Elf, note_count: usize) error{OutOfMemory}!ErrorWithNotes {
     try self.misc_errors.ensureUnusedCapacity(self.base.allocator, 1);
     return self.addErrorWithNotesAssumeCapacity(note_count);
 }
diff --git a/src/link/Elf/Atom.zig b/src/link/Elf/Atom.zig
index d044ef40a2..5b91dddff4 100644
--- a/src/link/Elf/Atom.zig
+++ b/src/link/Elf/Atom.zig
@@ -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;
@@ -336,6 +312,8 @@ pub fn scanRelocs(self: Atom, elf_file: *Elf, undefs: anytype) !void {
 
         if (rel.r_type() == elf.R_X86_64_NONE) continue;
 
+        const r_offset = std.math.cast(usize, rel.r_offset) orelse return error.Overflow;
+
         const symbol_index = switch (file_ptr) {
             .zig_module => |x| x.symbol(rel.r_sym()),
             .object => |x| x.symbols.items[rel.r_sym()],
@@ -388,7 +366,54 @@ pub fn scanRelocs(self: Atom, elf_file: *Elf, undefs: anytype) !void {
 
             elf.R_X86_64_PC32 => {},
 
-            else => @panic("TODO"),
+            elf.R_X86_64_TPOFF32,
+            elf.R_X86_64_TPOFF64,
+            => {
+                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.?[r_offset - 3 ..])) break :blk false;
+                    break :blk true;
+                };
+                if (!should_relax) {
+                    symbol.flags.needs_gottp = true;
+                }
+            },
+
+            else => {
+                var err = try elf_file.addErrorWithNotes(1);
+                try err.addMsg(elf_file, "fatal linker error: unhandled relocation type {}", .{
+                    fmtRelocType(rel.r_type()),
+                });
+                try err.addNote(elf_file, "in {}:{s} at offset 0x{x}", .{
+                    self.file(elf_file).?.fmtPath(),
+                    self.name(elf_file),
+                    r_offset,
+                });
+            },
         }
     }
 }
@@ -430,7 +455,10 @@ 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;
 
@@ -463,9 +491,9 @@ pub fn resolveRelocs(self: Atom, elf_file: *Elf, code: []u8) !void {
         // Relative offset to the start of the global offset table.
         const G = @as(i64, @intCast(target.gotAddress(elf_file))) - GOT;
         // // Address of the thread pointer.
-        // const TP = @as(i64, @intCast(elf_file.getTpAddress()));
+        const TP = @as(i64, @intCast(elf_file.tpAddress()));
         // // Address of the dynamic thread pointer.
-        // const DTP = @as(i64, @intCast(elf_file.getDtpAddress()));
+        // const DTP = @as(i64, @intCast(elf_file.dtpAddress()));
 
         relocs_log.debug("  {s}: {x}: [{x} => {x}] G({x}) ({s})", .{
             fmtRelocType(r_type),
@@ -512,10 +540,43 @@ pub fn resolveRelocs(self: Atom, elf_file: *Elf, code: []u8) !void {
                 try cwriter.writeIntLittle(i32, @as(i32, @intCast(G + GOT + A - P)));
             },
 
-            else => {
-                log.err("TODO: unhandled relocation type {}", .{fmtRelocType(rel.r_type())});
-                @panic("TODO unhandled relocation type");
+            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[r_offset - 3 ..]) catch unreachable;
+                    try cwriter.writeIntLittle(i32, @as(i32, @intCast(S - TP)));
+                }
             },
+
+            else => {},
         }
     }
 }
@@ -681,6 +742,80 @@ 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;
diff --git a/src/link/Elf/Object.zig b/src/link/Elf/Object.zig
index fe14831d1a..191f6774a5 100644
--- a/src/link/Elf/Object.zig
+++ b/src/link/Elf/Object.zig
@@ -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) {
@@ -233,8 +235,7 @@ fn getOutputSectionIndex(self: *Object, elf_file: *Elf, shdr: elf.Elf64_Shdr) er
         const is_alloc = flags & elf.SHF_ALLOC != 0;
         const is_write = flags & elf.SHF_WRITE != 0;
         const is_exec = flags & elf.SHF_EXECINSTR != 0;
-        const is_tls = flags & elf.SHF_TLS != 0;
-        if (!is_alloc or is_tls) {
+        if (!is_alloc) {
             log.err("{}: output section {s} not found", .{ self.fmtPath(), name });
             @panic("TODO: missing output section!");
         }
@@ -243,7 +244,7 @@ fn getOutputSectionIndex(self: *Object, elf_file: *Elf, shdr: elf.Elf64_Shdr) er
         if (is_exec) phdr_flags |= elf.PF_X;
         const phdr_index = try elf_file.allocateSegment(.{
             .size = Elf.padToIdeal(shdr.sh_size),
-            .alignment = if (is_tls) shdr.sh_addralign else elf_file.page_size,
+            .alignment = elf_file.page_size,
             .flags = phdr_flags,
         });
         const shndx = try elf_file.allocateAllocSection(.{
@@ -428,7 +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| {
@@ -591,7 +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;
@@ -685,7 +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;
@@ -693,6 +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);
diff --git a/src/link/Elf/Symbol.zig b/src/link/Elf/Symbol.zig
index 1be66c33e2..a0236f4c87 100644
--- a/src/link/Elf/Symbol.zig
+++ b/src/link/Elf/Symbol.zig
@@ -196,9 +196,10 @@ pub fn setOutputSym(symbol: Symbol, elf_file: *Elf, out: *elf.Elf64_Sym) void {
         //     if (symbol.flags.is_canonical) break :blk symbol.address(.{}, elf_file);
         //     break :blk 0;
         // }
-        // if (st_shndx == elf.SHN_ABS) break :blk symbol.value;
-        // const shdr = &elf_file.sections.items(.shdr)[st_shndx];
-        // if (Elf.shdrIsTls(shdr)) break :blk symbol.value - elf_file.getTlsAddress();
+        if (st_shndx == elf.SHN_ABS) break :blk symbol.value;
+        const shdr = &elf_file.shdrs.items[st_shndx];
+        if (shdr.sh_flags & elf.SHF_TLS != 0 and file_ptr != .linker_defined)
+            break :blk symbol.value - elf_file.tlsAddress();
         break :blk symbol.value;
     };
     out.* = .{
@@ -327,10 +328,12 @@ 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,
diff --git a/src/link/Elf/ZigModule.zig b/src/link/Elf/ZigModule.zig
index 93908e5f1f..c79680dbf5 100644
--- a/src/link/Elf/ZigModule.zig
+++ b/src/link/Elf/ZigModule.zig
@@ -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,