[ELF] Fix _TLS_MODULE_BASE_ relocations

Fixes #421

elf/main.cc

@@ -662,16 +662,6 @@ static int elf_main(int argc, char **argv) {
 
   // At this point, file layout is fixed.
 
-  // Some types of TLS relocations are defined relative to the beginning
-  // or the end of the TLS segment address. Find these addresses now.
-  for (ElfPhdr<E> phdr : create_phdr(ctx)) {
-    if (phdr.p_type == PT_TLS) {
-      ctx.tls_begin = phdr.p_vaddr;
-      ctx.tls_end = align_to(phdr.p_vaddr + phdr.p_memsz, phdr.p_align);
-      break;
-    }
-  }
-
   t_before_copy.stop();
 
   // Create an output file

elf/output-chunks.cc

@@ -268,6 +268,12 @@ std::vector<ElfPhdr<E>> create_phdr(Context<E> &ctx) {
     i++;
     while (i < ctx.chunks.size() && (ctx.chunks[i]->shdr.sh_flags & SHF_TLS))
       append(ctx.chunks[i++]);
+
+    // Some types of TLS relocations are defined relative to the TLS
+    // segment, so save its addresses for easy access.
+    ElfPhdr<E> &phdr = vec.back();
+    ctx.tls_begin = phdr.p_vaddr;
+    ctx.tls_end = align_to(phdr.p_vaddr + phdr.p_memsz, phdr.p_align);
   }
 
   // Add PT_DYNAMIC
@@ -1067,7 +1073,8 @@ std::vector<GotEntry<E>> GotSection<E>::get_entries(Context<E> &ctx) const {
 
   if constexpr (E::supports_tlsdesc)
     for (Symbol<E> *sym : tlsdesc_syms)
-      entries.push_back({sym->get_tlsdesc_idx(ctx), 0, E::R_TLSDESC, sym});
+      entries.push_back({sym->get_tlsdesc_idx(ctx), 0, E::R_TLSDESC,
+                         sym == ctx._TLS_MODULE_BASE_ ? nullptr : sym});
 
   for (Symbol<E> *sym : gottp_syms) {
     i64 idx = sym->get_gottp_idx(ctx);

elf/passes.cc

@@ -1473,8 +1473,20 @@ void fix_synthetic_symbols(Context<E> &ctx) {
     start(ctx._GLOBAL_OFFSET_TABLE_, ctx.got);
 
   // _TLS_MODULE_BASE_
-  if constexpr (E::supports_tlsdesc)
-    ctx._TLS_MODULE_BASE_->value = ctx.tls_begin;
+  if constexpr (E::supports_tlsdesc) {
+    u64 addr;
+    if constexpr (std::is_same_v<E, X86_64> || std::is_same_v<E, I386>) {
+      if (ctx.arg.relax && !ctx.arg.shared)
+        addr = ctx.tls_end;
+      else
+        addr = ctx.tls_begin;
+    } else {
+      addr = ctx.tls_begin;
+    }
+
+    ctx._TLS_MODULE_BASE_->shndx = -1;
+    ctx._TLS_MODULE_BASE_->value = addr;
+  }
 
   // __GNU_EH_FRAME_HDR
   start(ctx.__GNU_EH_FRAME_HDR, ctx.eh_frame_hdr);

test/elf/tls-module-base.sh

@@ -0,0 +1,57 @@
+#!/bin/bash
+export LC_ALL=C
+set -e
+CC="${CC:-cc}"
+CXX="${CXX:-c++}"
+GCC="${GCC:-gcc}"
+GXX="${GXX:-g++}"
+OBJDUMP="${OBJDUMP:-objdump}"
+MACHINE="${MACHINE:-$(uname -m)}"
+testname=$(basename "$0" .sh)
+echo -n "Testing $testname ... "
+cd "$(dirname "$0")"/../..
+mold="$(pwd)/mold"
+t=out/test/elf/$testname
+mkdir -p $t
+
+[ $MACHINE = x86_64 ] || { echo skipped; exit; }
+
+cat <<EOF | $CC -o $t/a.o -c -xassembler -
+.globl get_foo
+get_foo:
+  lea _TLS_MODULE_BASE_@TLSDESC(%rip), %rax
+  call *_TLS_MODULE_BASE_@TLSCALL(%rax)
+  lea foo@dtpoff(%rax), %rax
+  mov %fs:(%rax), %eax
+  ret
+.section .tdata, "awT", @progbits
+foo:
+.long 20
+EOF
+
+cat <<EOF | $CC -o $t/b.o -c -xc -
+_Thread_local int bar = 3;
+EOF
+
+cat <<EOF | $CC -o $t/c.o -c -xc -
+#include <stdio.h>
+
+int get_foo();
+extern _Thread_local int bar;
+
+int main() {
+  printf("%d %d\n", get_foo(), bar);
+}
+EOF
+
+$CC -B. -o $t/exe1 $t/a.o $t/b.o $t/c.o
+$QEMU $t/exe1 | grep -q '^20 3$'
+
+$CC -B. -o $t/exe2 $t/a.o $t/b.o $t/c.o -Wl,-no-relax
+$QEMU $t/exe2 | grep -q '^20 3$'
+
+$CC -B. -o $t/d.so $t/a.o -shared
+$CC -B. -o $t/exe3 $t/b.o $t/c.o $t/d.so
+$QEMU $t/exe3 | grep -q '^20 3$'
+
+echo OK